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.pgen.1007093 | MITF – A controls branching morphogenesis and nephron endowment | Congenital nephron number varies widely in the human population and individuals with low nephron number are at risk of developing hypertension and chronic kidney disease. The development of the kidney occurs via an orchestrated morphogenetic process where metanephric mesenchyme and ureteric bud reciprocally interact to induce nephron formation. The genetic networks that modulate the extent of this process and set the final nephron number are mostly unknown. Here, we identified a specific isoform of MITF (MITF-A), a bHLH-Zip transcription factor, as a novel regulator of the final nephron number. We showed that overexpression of MITF-A leads to a substantial increase of nephron number and bigger kidneys, whereas Mitfa deficiency results in reduced nephron number. Furthermore, we demonstrated that MITF-A triggers ureteric bud branching, a phenotype that is associated with increased ureteric bud cell proliferation. Molecular studies associated with an in silico analyses revealed that amongst the putative MITF-A targets, Ret was significantly modulated by MITF-A. Consistent with the key role of this network in kidney morphogenesis, Ret heterozygosis prevented the increase of nephron number in mice overexpressing MITF-A. Collectively, these results uncover a novel transcriptional network that controls branching morphogenesis during kidney development and identifies one of the first modifier genes of nephron endowment.
| The number of nephrons, the functional unit of kidney, varies widely among humans. Indeed, it has been shown that kidneys may contain from 0.3 to more than 2 million of nephrons. Nephrons are formed during development via a coordinated morphogenetic program in which the metanephric mesenchyme reciprocally and recursively interacts with the ureteric bud. The fine-tuning of this cross-talk determines the final number of nephrons. Strong evidence indicates that suboptimal nephron endowment is associated with an increased risk of hypertension and chronic kidney disease, a major healthcare burden. Indeed, chronic kidney disease is characterized by the progressive decline of renal function towards end stage renal disease, which occurs once a critical number of nephrons has been lost. Elucidating the molecular mechanisms that control nephron endowment is, therefore, a critical issue for public health. However, little is known about the factors that determine the final number of nephrons in the healthy population. Our data showed that nephron endowment is genetically predetermined and identified Mitfa, a bHLH transcription factor, as one of the first modifiers of nephron formation during kidney development. By generating an allelic series of transgenic mice expressing different levels of MITF-A, we discovered that MITF-A promotes final nephron endowment. In addition, we elucidated the molecular mechanisms by which MITF-A promotes nephron formation and identified RET as one of the critical effectors.
| For decades, it was believed that the number of nephrons in the kidneys does not vary among normal individuals. However, several studies performed over the last 30 years have clearly demonstrated that the number of nephrons varies widely among human populations and even among healthy individuals of the same ethnicity [1]. In fact, kidneys may contain from 0.3 to more than 2 million nephrons. Strong evidence indicates that suboptimal nephron endowment is associated with an increased risk for developing essential hypertension and chronic kidney disease (CKD) [1,2]. CKD is characterized by a progressive decline in renal function eventually leading to end stage renal disease which occurs when a critical number of nephrons has been lost, irrespective of the cause of the renal damage. Over 10% of the adult population is estimated to suffer from CKD [3]. Elucidating the cellular events and genetic networks that control nephron endowment is, therefore, a critical issue for understanding the predisposition to CKD and consequently for public health.
Little is known about the factors that determine nephron number in the normal population. Individual nephrons are formed during fetal development via an orchestrated morphogenetic process that is characterized by the reciprocal interaction between the metanephric mesenchyme (MM) and the ureteric bud (UB) [4,5] and that critically determines the final nephron number [6]. In fact, subtle modifications in the efficiency and/or accuracy of this process can lead to renal hypoplasia [7]. Epidemiological studies have indicated that both environmental and genetic factors play critical roles in nephron development [6,8,9]. Global caloric and protein restriction as well as iron or vitamin A deficiency result in reduced nephron number [9]. Similarly, reduced birth weight either due to premature birth or intrauterine growth retardation correlates with low nephron number [10]. On the other hand, it has been shown that hypomorphic variants of PAX2 and RET, two genes encoding critical players of branching morphogenesis, are associated with a 10% decrease in renal volume at birth [11,12]. Conversely, it has been reported that a relatively common variant of ALDH1A2, a gene encoding an enzyme involved in retinoic acid metabolism, is associated with a 22% increase in kidney size at birth [13].
Microphthalmia-associated transcription factor (MITF) is a member of the basic/helix-loop-helix/leucine zipper (b-HLH-Zip) family of transcription factors [14]. To date, at least nine isoforms of MITF have been identified that arise from the use of alternative promoters transcribing different (protein coding) first exons, leading to the expression of proteins carrying distinct N terminal portions [15,16]. Some isoforms, including MITF-A, are widely expressed, while others are cell-type-specific. MITF is known to be essential for the development and function of different cell lineages such as pigmented cells, mast cells and osteoclasts [15,17]. In addition, MITF plays a role in pathophysiological events, such as in cardiac growth and hypertrophy [18], B cell homeostasis [19], and melanoma proliferation and invasiveness [20]. Recently, MITF has also been shown to be crucial for renal pathology since MITF variants were shown to affect the progression of renal disease [21] or influence the occurrence of renal carcinoma [22]. In this context, our group has discovered that a hypomorphic variant in the 5' UTR of MITF-A predisposes FVB/N mice to develop renal lesions after experimental nephron reduction [21].
Here, in order to improve our knowledge on MITF-A function in renal pathophysiology, we generated several lines of transgenic mice expressing different levels of MITF-A. We discovered that MITF-A regulates the final nephron number by modulating UB branching. Enhanced UB branching occurs via an increase of cell proliferation involving key developmental targets, including RET. Collectively, our results reveal a novel actor of kidney development and uncover a novel function of MITF-A.
The Ksp-cadherin, FLAG-MITF-A fusion gene was generated by cloning the full-length mouse MITF-A tagged at the 5′end by a FLAG epitope into the Ksp/BGH/link plasmid (kindly provided by Peter Igarashi) containing the minimal Ksp-cadherin promoter followed by the beta-globin enhancer element [23]. The poly(A) sequence for the human beta-globin gene was isolated and cloned into the Ksp-cadherin-FLAG-MITF-A plasmid. Transgenic mouse lines were generated by microinjecting the purified DNA construct into fertilized mouse oocytes derived from a pure genetic background FVB/N. Twenty FLAG-MITF-A transgenic founder lines were identified by PCR analysis of tail DNA using primers specific to the MITF-A transgene (S3 Table).
To generate the targeting construct, the potential Mitfa promoter/Mitfa exon and its flanking regions (15,942 bp) were cloned using plasmid rescue from BAC RP23-77E9. A floxed neomycin resistance expression cassette flanked by 200 bp of sequence flanking the Mitfa promoter/Mitfa exon was used to replace 5,956 bp of the putative Mitfa promoter/Mitfa exon from the above plasmid and used for standard targeting of LC3 ES cells (genotype [C57BL/6Nx129S6]F1). A correctly targeted ES cell colony was used to generate chimeric animals. Of several germline-transmitting lines, one was selected and crossed with C57BL/6JN/129S4-Prm1-Cre deleter mice (Jackson Laboratories, stock 003328, backcrossed twice to C57BL/6J) to remove the floxed neomycin cassette. Offspring lacking the neo-cassette were backcrossed to C57BL/6J mice for nine generations and then bred to homozygosity. Wild-type littermates were always used as controls. The primers used for genotyping are indicated in S3 Table.
To generate MITF-Awt/tgMITF-A;Retwt/- mice, heterozygous FVB/N mice overexpressing MITF-A (see above) were crossed with heterozygous C57BL/6 Ret knockout mice [38]. Double heterozygous offsprings seemed normal and were born in the expected proportion.
Animals were fed ad libitum and housed at constant ambient temperature in a 12-hour light cycle. Animal procedures were approved by the Departmental Director of “Services Vétérinaires de la Préfecture de Police de Paris” and by the ethical committee of the Paris Descartes University as well as by the NIH/NINDS intramural program.
For the post-natal characterization of the MITF-A transgenic line, transgenic mice and wild-type littermates were sacrificed 2, 4, 6 and 12 months after birth (n = 5 to 12 for each genotype and time point). For renal function studies, urine samples were collected one week before sacrifice using metabolic cages over the course of 24 hours and blood samples were obtained at time of sacrifice (n = 4 to 6 for each genotype and time point) and kidneys dissected for appropriate studies. In addition, liver, spleen, heart, lung and brain were removed at 2 months for mRNA studies. For developmental studies, except where stated otherwise, the number of kidneys removed from E13.5 to P0 ranged between 14 and 25 for each genotype. For genetic interaction between Mitfa and Ret studies, mice were sacrificed at 3 weeks, kidneys were removed and the number of glomeruli was counted in females, exclusively (n = 3 to 7 for each genotype, i.e. MITF-Awt/wt;Retwt/wt, MITF-Awt/tgMITF-A;Retwt/wt, MITF-Awt/wt;Retwt/-, and MITF-Awt/tgMITF-A;Retwt/-).
Plasma urea levels and urinary albumin, protein and creatinine levels were measured using an Olympus multiparametric autoanalyser (Instrumentation Laboratory).
For morphological analysis, kidneys were fixed in 4% paraformaldehyde, paraffin embedded, and 4-μm sections were stained with periodic acid Schiff (PAS) or Hematoxylin & Eosin (HE). A pathologist, blinded to the nature of the group, examined and evaluated all the sections. For morphometric analysis, glomerular and tubular surfaces were measured on PAS-stained sections, using a Nikon digital camera Dx/m/1200 and Lucia software (Laboratory Imaging Ltd., Prague, Czech Republic). Twenty randomly selected microscopic fields of the cortex were studied (X200). At least 20 glomeruli were analyzed for each animal. Tubular morphometric analysis was performed on sagittal kidney sections. The area of external profile of proximal tubules and the area of the lumen were measured, and the epithelial surface was calculated as the difference between these two areas. At least 35 tubular cross sections were analyzed for each animal. Tubular cell size was evaluated on the same tubular cross sections and was calculated as the ratio of epithelial surface to nuclei number for each tubular section.
Kidneys of two month-old mice were de-capsulated and macerated in 5N HCL for 30 minutes at 37°C. After rinsing with distilled water, the suspension was stored overnight at 4°C. The flask was shaken on the following day and the tubules and glomeruli were suspended in 25 ml of distilled water. Glomeruli were counted on five different 250 μl aliquots using a Nikon Eclipse E800 microscope.
Kidneys from E13.5 mice were dissected and fixed in cold methanol for 15 minutes. After washing in PBS for 15 minutes, the entire kidneys were treated with 0.1% Triton in PBS for 2 hours. To visualize the branches and tips of the ureteric tree, kidneys were incubated overnight with primary mouse antibody against Calbindin-D28K (Sigma-Aldrich, 1:200), a calcium-binding protein expressed in the ureteric epithelium. Metanephroi were then washed in PBS and incubated with the secondary antibody AlexaFluor 555 goat anti-mouse IgG (Invitrogen, Molecular Probes, 1:100). The number of UB tips was counted using a Nikon Eclipse E800 microscope.
For cell proliferation, 4-μm sections of paraffin-embedded kidneys were incubated with a rabbit anti-phospho-histone H3 antibody (Millipore) at 1:200 and a mouse anti-PCNA antibody (DAKO) at 1:50, followed by the appropriate secondary antibody. The number of proliferating cells was determined in UB and expressed as the number of pH3- or PCNA- positive cells per UB structure on whole kidney sections.
Apoptosis was detected in 4-μm sections of paraffin-embedded kidneys by TUNEL assay using the In Situ Cell Death Detection kit (Roche) according to the manufacturer’s protocol. The number of apoptotic cells was determined as the number of TUNEL–positive nuclei per microscopic field in whole kidney sections. Ten microscopic fields were scored for each kidney in 5 embryos of each genotype.
Four μm sections were retrieved with Tris-EDTA (TE) buffer (pH = 9.0) at 90°C for 20 minutes, then incubated first overnight at 4°C with a rabbit anti-MITF-A antibody [21] 1/200, then with a secondary biotinylated anti-rabbit antibody (Vector) 1/500, followed by streptavidine-peroxidase (Dako) 1/500. Staining was revealed by 3–3’-diaminobenzidine (DAB).
In situ hybridization was carried out on 10-μm cryosections from OCT embedded kidneys frozen in isopentane solution under liquid nitrogen. Digoxigenin-dUTP–labeled sense and antisense RNA probes were synthesized using Roche reagents. The following cDNA templates were used: Ret, Wnt11, Pax2, BMP7, Wnt9b and Spry1 (a kind gift from Isabelle Gross, INSERM U682). An RNA probe against MITF-A was obtained from cDNA encompassing both exon 1A and exon 1B. The resulting RNA probe could hybridize with the endogenous MITF-A and the MITF-A transgene and potentially with other MITF isoforms containing exon 1B (MITF-H, MITF-C, MITF-J, MITF-B, MITF-Mc). We designed this probe, since we failed to get a specific staining with a shorter probe (177 nucleotides) that overlaps only with MITF-A sequence. Cryosections were fixed with 4% paraformaldehyde for 10 min, acetylated for 10 min and then pre-hybridized for 3 hours. Hybridization was performed overnight at 60°C in the presence of 1 μg/ml for each RNA probe. After successive washings at 60°C in 2X and 0.2X SCC buffer, the sections were washed in Tris buffer pH 7.6 (0.1 M Tris, 0.15 M NaCl) incubated in 2% blocking reagent solution (Roche) for 3 hours, and then incubated overnight with anti-digoxigenin antibody (alkaline phosphatase-conjugated Fab, Roche, 1:1,500). Sections were successively washed in Tris buffer, pH 7.6 and in Tris buffer, pH 9.5 (0.1 M Tris, 0.1M NaCl, 50 mM Mgcl2) and finally incubated with NBT/BCIP AP substrate solution (Roche).
For colocalization staining, the sections were successively fixed in 4% paraformaldehyde and incubated overnight with a rabbit antibody anti-laminin (Sigma-Aldrich), at 1:200, followed by an Alexa-Fluor 555 donkey anti-rabbit antibody (Invitrogen, Molecular Probes), at 1/200.
MITF-A immunoblotting was performed on nuclear extracts obtained after a 100,000 g ultracentrifugation of kidneys homogenized in 10 mM Hepes /1.7 M sucrose buffer (pH 7.9) supplemented with DTT (0.5 mM), spermine (0.15 mM), spermidine (0.5 mM), benzamidine (2 mM), anti-protease and anti-phosphatase inhibitors. The nuclear pellet was suspended in 8 M urea buffer and centrifuged. The supernatant was assayed for MITF-A immunoblotting using a rabbit antibody raised against a recombinant GST-MITF-A protein [21]. A polyclonal antibody against Lamin-A/C (Epitomics) was used as a control for nuclear proteins.
Kidneys were incubated in RNA later (Ambion) overnight and stored at -80°C until extraction. Total RNA was extracted from embryonic kidneys using RNeasy Micro Kit from Qiagen. RNAs were DNase treated (DNase I RNase-free, Qiagen) and reverse transcribed according to the manufacturer's protocols (Superscript II, Invitrogen). Quantitative RT-PCR was carried out using an ABI PRISM 7700 Sequence Detection system (Applied Biosystems). HPRT was used as the normalization control. The primers used (Eurogentec) are described in S3 Table.
Positional weight matric (PWM) was generated from 47 published functional MITF binding sites [27] and by using the program MEME 3.5.0 (http://meme.nbcr.net/meme/) to search both strands for putative motifs. The informative portion of the aligned sequences was converted into a PWM by assigning to each position/nucleotide the frequency of that given nucleotide in that given position. PWM sequence logo representation was created by submitting MEME informative sequences to WebLogo (http://weblogo.berkeley.edu/) The identification of the conserved predicted MITF-A binding sites among the 102 kidney developmental genes (MGI abnormal kidney development data base) was then carried out as described previously [26].
Data are expressed as means ± SEM. Differences between the experimental groups were evaluated using ANOVA, and, when significant (P < 0.05), followed by the Tukey-Kramer test. When only two groups were compared, the Mann-Whitney test was used. The Pearson’s correlation coefficient was used to test correlation between variables. The statistical analysis was performed using Graph Prism Software.
In order to elucidate the role of MITF-A in renal pathophysiology, we generated a set of FVB/N transgenic mouse lines (here indicated as MITF-A mice) overexpressing MITF-A in the kidney under the control of the Ksp-cadherin promoter (Fig 1A). Twenty founders were identified, and lines were established from each of them. Among them, only three (line 14, 42 and 47) overexpressed MITF-A selectively in the kidney, and not in liver, spleen, heart, lung, or brain (S1 Fig). Quantitative RT-PCR showed that Mitf-A expression levels varied in the kidneys of the three transgenic lines. The highest level was detected in the kidneys of line 42 heterozygotes and homozygotes, in which Mitf-A mRNA was increased 10- and 20-fold, respectively, as compared to kidneys of wild-type littermates (Fig 1B). In contrast, the increase was only 5-fold and 3-fold in heterozygotes of line 14 and 47, respectively (S2A and S2D Fig). Western-blot analysis showed that the levels of the nuclear protein paralleled that of the corresponding mRNA (Fig 1C).
Transgenic mice overexpressing MITF-A were viable, fertile, and appeared phenotypically normal. Their kidneys, however, were significantly larger than those of wild-type littermates when analyzed two months after birth (Fig 2A and 2B). This difference was already detectable at birth, was maintained at least until 12 months of age (S3 Fig), and was kidney specific as no differences were detected in body, heart, liver and spleen weights (Table 1). Renal histology revealed no gross abnormalities of the cortex/medulla architecture as well as of the glomerular and tubular morphology in MITF-A transgenic mice from 2 to 12 months after birth (S4A Fig). Furthermore, no differences in blood urea, albuminuria and urinary protein excretion were found at 2, 4, 6 and 12 months after birth (S4B Fig), suggesting that renal function was not affected.
Kidney overgrowth may result from an increase of either nephron size (hypertrophy) or nephron number. To determine which of these events might account for the increased kidney size in MITF-A transgenic mice, we first performed a morphometric analysis and measured the average surface of tubular and glomerular sections. Our results showed that the average tubular epithelial cell surface (external surface minus lumen surface for individual tubules) was comparable between MITF-A transgenic mice and wild-type littermates two months after birth (S1 Table), and the average glomerular surface area was even significantly decreased in MITF-A transgenic mice as compared to wild-type littermates (S1 Table). These data argue against a role of hypertrophy in MITF-A induced renal overgrowth. Therefore, we next evaluated the average number of nephrons contained in each kidney. Remarkably, in both males and females, the number of glomeruli was significantly higher (30%) in MITF-A homozygous transgenic kidneys as compared to their wild-type counterparts (Fig 2C). A highly significant correlation (r = 0.96; P < 0.0001) was found between kidney weight and glomeruli number (Fig 2D). Moreover, the number of glomeruli was positively correlated with the different Mitf-A expression levels found in wild type, heterozygous and homozygous mice of line 42 (Fig 2E), suggesting a role for MITF-A in final nephron endowment. Nevertheless, we could not formally rule out that the biological effect observed in these mice was due to adverse effects of transgene integration in a crucial genomic site. In order to rule out this possibility, we studied the two additional MITF-A overexpressing transgenic mouse lines. Consistent with the fact that these lines expressed MITF-A at lower levels than line 42, the increase in nephron number in heterozygotes was less pronounced (S2B and S2E Fig). Of note, as in line 42, renal morphology appeared normal in two month-old heterozygous mice from both line 14 and 47 (S2C and S2F Fig). This finding further supported the notion that transgenic MITF-A levels were correlated with nephron number and excluded the possibility that a specific integration of the transgene accounted for kidney overgrowth.
To further assess the role of MITF-A in nephron endowment, we produced a mouse line carrying a specific deletion of the Mitfa promoter by homologous recombination (Fig 3A). As expected, Mitf-A mRNA expression was negligible in kidneys of Mitfa-/- mice (Fig 3B) and total Mitf mRNA expression levels (that include all the isoforms) were significantly decreased (Fig 3C), indicating that the loss of the MITF-A isoform was not compensated by an increase of any of the other isoforms. Homozygous null Mitfa mice were fertile, had a normal phenotype and were indistinguishable from their wild-type littermates. Consistent with the above results obtained with MITF-A overexpressing transgenic mice, however, the number of their glomeruli was significantly reduced (20%) as compared to wild-type littermates (Fig 3D).
The final number of nephron is determined by the fine-tuning of mesenchyme/ureteric bud (UB) crosstalk. Since the Ksp-cadherin promoter drives the expression of the transgene in the UB [23], we hypothesized that MITF-A may affect branching morphogenesis. To assess this, we monitored branching morphogenesis in mice of line 42 during embryogenesis. Using calbindin-1 staining, we first showed that the overall branching pattern appeared normal. However, the number of branchings was significantly increased in transgenic embryos (Fig 4A), while the global architecture of developing kidneys was similar between MITF-A transgenic and wild-type embryos from E13.5 to P14 (Fig 4B). Notably, at E13.5, metanephroi of MITF-A transgenic embryos exhibited the typical un-induced metanephric mesenchyme and growing branch tips in the external portion of the cortex (Fig 4B). Quantitative analysis confirmed that the number of UB tips was significantly higher in MITF-A transgenic kidneys of embryos from line 42 as compared to wild-type controls (Fig 4C and S5A Fig). These results were confirmed in transgenic line 47, in which the number of UB tips was 36.3 ± 2.1 in heterozygous transgenics and 26.3 ± 1.7 in wild-type littermates, respectively (P < 0.001). On the other hand, in Mitfa-/- embryos, the number of UB tips was mildly but significantly reduced at E13.5 as compared to that in wild-type littermates (Fig 4D and S5B Fig). Quantitative RT-PCR confirmed the increase of MITF-A mRNA in metanephroi of E13.5 MITF-A overexpressing embryos as compared to wild-type controls (Fig 4E). On the contrary, the expression of Mitf-A was negligible in Mitfa-/- embryonic kidneys (Fig 4F). Taken together, these data point to an unexpected role for MITF-A in branching morphogenesis.
Since our data revealed MITF-A as a novel actor in nephrogenesis, we next tried to define its expression pattern and its effect on other MITF isoforms. In situ hybridization revealed that in wild type embryos at E13.5, Mitf-A is expressed in the UB, the mesenchyme and in S-shaped bodies (Fig 5A). In transgenic metanephroi, Mitf-A staining was markedly increased, particularly in UB branches (Fig 5B) as expected from the pattern of expression driven by the Ksp-cadherin promoter [23]. In addition, we observed an increase of Mitf-A expression in S-bodies and ureteric tips (Fig 5C). A further increase in Mitf-A expression in transgenic metanephroi was found at E15.5 (S6 Fig). Immunohistochemical analysis corroborated these observations. In fact, it showed that endogenous MITF-A protein was expressed in both UB stalks and tips and to a much lower extent in mesenchyme (Fig 5D). Moreover, we observed a marked increase of MITF-A staining in UB stalks, tips and S-bodies of MITF-A transgenic embryos (Fig 5D). Since MITF is composed of several distinct isoforms, it was important to analyze the pattern of expression of all isoforms during nephrogenesis. Quantitative RT-PCR showed that, amongst the nine MITF isoforms, only Mitf-A, Mitf-H, Mitf-C, Mitf-J and Mitf-Mc, but not Mitf-M, Mitf-B, Mitf-D and Mitf-E, were detectably expressed in kidneys of E13.5 embryos, (S7A Fig). Interestingly, MITF-A overexpression was associated with an increase in Mitf-C and Mitf-J mRNA expression (S7B Fig), suggesting positive regulation by MITF-A.
It has been well documented that both cell proliferation [7,24] and apoptosis [25] play a role in kidney development. Interestingly, MITF has been implicated in the control of both of these events [20] and so we evaluated them in our experimental model. Using phospho-histone-H3 staining, we observed a strong increase in cell proliferation in UB of E13.5 MITF-A transgenic kidneys as compared to their age-matched wild-type counterparts (Fig 6A). These results were confirmed using an antibody directed against PCNA, a protein selectively expressed in proliferative S phase cells (Fig 6B). Conversely, apoptosis, as judged by TUNEL assay, was dramatically decreased in transgenic mice compared to wild-type mice. This was, however, mostly seen in the metanephric mesenchyme and thus suggested a non cell-autonomous effect of MITF-A (Fig 6C).
Since MITF-A is a transcription factor, the observed effects are likely due to changes in the expression of one or more renal developmental genes potentially under the control of MITF-A. To identify potential target genes of MITF-A during kidney development, we took advantage of a method that we recently developed to predict STAT3 functional binding sites through comparative genomics [26] and that we adapted to MITF on the basis of the 47 known genomic MITF binding sites [27]. Using this approach, we limited the analysis to genes known to be involved in kidney development (MGI abnormal kidney development data base). Of the 102 genes analyzed, we identified 81 with a slight enrichment of conserved binding sites (CBS) for MITF (S2 Table). By comparison, the percentage of genes showing enrichment of MITF-A CBS was lower among those specifically affecting liver development. To refine our analysis, we systematically retrieved data from the GUDMAP database (http://www.gudmap.org/) to select genes that are expressed in UB stalks and tips where the MITF-A transgene was expressed. For genes non-included in the GUDMAP database, we scanned the literature for UB localization. These analyses identified 28 potential MITF-A targets expressed in UB (S2 Table). Among these, we first focused our attention on Bmp7 [28] and Pax2 [29], both of which are known to be involved in branching morphogenesis and to cooperate with MITF in other contexts [30,31]. However, in situ hybridization and quantitative RT-PCR revealed that the expression of neither Bmp7 nor Pax2 was modified by MITF-A overexpression (Fig 7A and 7B). Similarly, the expression of Wnt9b, another potential MITF-A target which acts as a paracrine factor in the metanephric mesenchyme induction and in UB branching [32], was unaffected by MITF-A overexpression (Fig 7A and 7B). Likewise, RARα, a receptor for retinoic acid, the active form of vitamin A, did not show any difference. This was in a way surprising since Vitamin A deficiency during pregnancy can lead to reduced nephron number [9] similar to what we here report for Mitfa deficiency in mice. Hence, we next turned our attention to our MITF-A enriched genes that are known to be critically involved in branching morphogenesis, i.e. Ret, Wnt11 and Spry1 [33–35]. Interestingly, while Wnt11 expression was unchanged in MITF-A transgenic metanephroi at E13.5, the expression of Ret was significantly increased (Fig 7C and 7D). In particular, in situ hybridization revealed that Ret mRNA expression was significantly enhanced in ureteric tips of MITF-A transgenic kidneys as compared to kidneys of wild-type littermates (Fig 7C). Quantitative RT-PCR showed that the increase of Ret mRNA paralleled MITF-A expression levels when comparing kidneys of wild-type, heterozygous and homozygous transgenic embryos (Fig 7D). Conversely, we observed that the expression of Ret was decreased in Mitfa-/- embryos as compared to wild-type littermates (S8 Fig). Consistent with the increase of Ret, the expression of Spry1, a downstream target of Ret, was also found increased in kidneys of MITF-A transgenic embryos as compared to wild-type controls (Fig 7D). To corroborate this observation, we decided to measure the expression of other known targets of RET [36,37]. Interestingly, our results showed that several of these targets, i.e. the transcriptional regulator Etv5, the chemoreceptor Cxcr4, and the transcriptional factor Myb were significantly up-regulated in MITF-A transgenic embryos as compared to controls at E13.5 (S9 Fig). Altogether these results suggest that MITF-A might promote increased nephron endowment by modulating RET signaling.
Finally, to investigate if RET is a critical effector of MITF-A, we crossed transgenic mice overexpressing MITF-A (MITF-Awt/tgMITF-A) with heterozygous Ret knockout mice (Retwt/-) [38]. Remarkably, the results strongly support the genetic interaction between Ret and Mitfa. In fact, in basal conditions the heterozygosis for Ret, as expected, did not interfere with nephron endowment in absence of MITF-A overexpression. However, in mice overexpressing MITF-A, the heterozygosis for Ret dramatically reduced nephron endowment (Fig 8A). Consistently, the kidney weight was also significantly decreased in the transgenic MITF-A mice lacking one copy of Ret (Fig 8B). Remarkably, both the kidney weight and the number of glomeruli were similar between MITF-Awt/tgMITF-A;Retwt/- mice and double wild-type mice (Fig 8). Collectively, these results indicate that RET is the key target of MITF-A in branching morphogenesis.
During kidney development, ureteric bud branching morphogenesis is a fundamental process that defines the normal architecture of the kidney and the final nephron number. However, the molecular networks that orchestrate this complex biological process have been only partially elucidated. In particular, the genetic programs that regulate nephron endowment in normal kidneys are still unknown. By generating lines of transgenic and knockout mice that express different levels of MITF-A, we have uncovered an important role for this transcription factor in controlling branching morphogenesis during kidney development. Notably, we showed that MITF-A overexpression increased UB branching, which led to a substantial increase of nephron number. Conversely, Mitfa deficiency resulted in reduced branching and nephron number. Mechanistically, MITF-A triggered UB cell proliferation, while it inhibited mesenchyme metanephric apoptosis. In addition, by coupling an in silico analysis with molecular studies, we showed that amongst the putative MITF-A target genes, Ret, a key factor of branching morphogenesis, and its downstream signaling pathway were significantly modified by MITF-A. Consistent with the idea that Ret is a critical target, we showed that Ret heterozygosis reverted the MITF-A induced phenotype. Collectively, these results provide novel insights into the genetic networks that control branching morphogenesis in kidney development and identified one of the first modifiers of physiological nephron endowment.
Nephrogenesis starts when the metanephric mesenchyme induces the nearby Wolffian duct to produce UB outgrowth, which then elongates, invades the mesenchyme, and undergoes a process of branching morphogenesis to give rise to the renal collecting duct system. At the same time, in a reciprocal inductive process, the ureteric tips induce surrounding mesenchyme to condense, epithelialize, and differentiate into mature nephrons, the functional unit of the kidney [5]. Abnormalities of this highly coordinated morphogenetic process can lead to defects ranging from severe congenital abnormalities of kidney and urinary tract (CAKUT) [39,40] to reduced nephron number [6]. Despite the fact that the morphological events leading to nephron formation are well characterized, the genes and genetic networks that orchestrate these events are still poorly elucidated. Our study identified Mitfa as a novel candidate modifier of nephron endowment. To our knowledge, this is the first gene whose dosage affects exclusively the final number of nephrons without affecting overall renal morphology or function. Previous studies have shown that both the overexpression of a dominant negative isoform of the type II activin receptor [41] or Tgfβ2 haploinsufficiency [42] lead to higher nephron number. Intriguingly, however, Tgfβ2 homozygous mice display renal agenesis [43], whereas mice lacking activin B die at birth [44]. Similarly, while common hypomorphic variants of RET or PAX2 have been associated with subtle renal hypoplasia [11,12], their inactivation has been shown to result in severe renal malformations [40]. Therefore, is seems that in contrast to Mitfa, these other genes have a more complex function in kidney morphogenesis and cannot be considered simply as modifiers of nephron number.
Intriguingly, we observed that in Mitfa-/- mice, Mitfa deficiency affected more the final number of nephrons than the number of UB at E13.5. Since nephrogenesis is a continuous process, it is tempting to speculate that the increased nephron number may result from a mild but continuous effect of MITF-A on UB branching and nephrogenesis. Indeed, we cannot formally exclude that the overexpression of MITF-A in UB might lead to the recruitment of more mesenchymal kidney cells.
MITF-A belongs to a family of transcription factors containing seven distinct isoforms differing at their amino termini [16] among which five (MITF-A, MITF-H, MITF-C, MITF-J and MITF-Mc) we showed to be expressed in E13.5 metanephroi and in adult kidney. These isoforms, which arise from the utilization of different promoters, may differ in expression patterns but share the main functional domains, including the dimerization and DNA-binding bHLH-Zip domain [15]. Mutations in the bHLH and basic domain of MITF have been reported in patients with Waardenburg and Tietz syndromes [45,46]. In both syndromes, melanocyte function is severely impaired resulting in deafness, lack of pigmentation and eye abnormalities. It might be surprising that no gross abnormalities in other tissues, including kidney, have been reported in these syndromes, despite the fact that the mutations affect the common functional sequence of all MITF isoforms. Whether, however, these patients have a reduced nephron number is an interesting hypothesis that has not been investigated.
We previously showed that a hypomorphic variant of Mitfa predisposes FVB/N mice to CKD progression in an experimental model of nephron reduction [21]. In this context, we observed that MITF-A acts by interacting with histone deacetylases to repress the transcription of Tgfa, a ligand of EGFR and a critical mediator of CKD progression. Remarkably, the number of nephrons is identical in Tgfa-/- and Tgfa+/+ mice (12,056 ± 561 and 11,250 ± 328, respectively), suggesting that a different genetic program is triggered by MITF-A during branching morphogenesis. To start to identify this program we combined comparative genomics with mRNA transcripts quantification. The results showed that several of the potential targets are normally expressed in MITF-A transgenic embryos. However, the expression of Ret, a tyrosine kinase receptor critically involved in renal branching morphogenesis [36], was significantly increased in MITF-A transgenic mice. In addition, we observed that Ret overexpression resulted in the activation of its signaling pathway, as judged by the up-regulation of several downstream critical targets, i.e. Etv5, Spry1, Cxcr4 or Myb [36,37]. More importantly, we confirmed the genetic interaction between Mitfa and Ret and showed that Ret heterozygosis prevents the effect of MITF-A overexpression on nephron endowment. In fact, the number of nephrons was similar between MITF-Awt/tgMITF-A;Retwt/- mice and wild-type littermates. Interestingly, although Ret inactivation has been shown to induce a very severe kidney phenotype (renal agenesis) [40], a common single polymorphism within an exonic splicing enhancer was associated with reduced kidney size at birth [12], supporting the idea that subtle changes in RET expression levels might account for nephron number variability. Hence, all these data together point to Ret as the critical target of MITF-A during branching morphogenesis. Although we cannot ascertain that Ret is a direct target of MITF-A, several lines of evidence support this idea. First, our in silico analysis revealed that the Ret regulatory region contains 12 MITF binding sites that are conserved in at least 5 species. Second, we demonstrated that MITF-A is expressed in the same structures than RET, i.e. the ureteric tips, and that RET expression levels in transgenic ureteric tips increased proportionally to those of MITF-A. Third, we observed that increased UB branching in E13.5 MITF-A metanephroi was associated with a marked increase of UB cell proliferation, an event known to participate in branching morphogenesis [24] and to be modulated by ERK1/2 [47] and PI3K [48], two effectors of RET. Whether other targets of MITF-A are involved in MITF-A promoting nephrogenesis is a possibility that we cannot formally exclude.
It has been previously shown that the overexpression of the anti-apoptotic protein BCL2 in the developing kidneys of Pax2+/- mice prevents the decrease of UB branching and nephron formation by suppressing cell apoptosis [49]. In addition, in the same model, it has been observed that BCL2 overexpression alone leads to increased kidney weight and increased nephron number. Conversely, Bcl2 gene inactivation results in decreased UB branching and renal hypoplasia, but also in severe cystic dysplasia [50]. Our study showed that MITF-A overexpression in UB lead to reduced apoptosis, but mainly in metanephric mesenchyme. Interestingly, BCL2 has been shown to be a direct target of MITF-M in melanocytes [51]. However, previous studies showed that BCL2 is exclusively expressed in metanephric mesenchyme, indicating that the inhibition of apoptosis in MITF-A transgenic embryos is mainly non-cell autonomous and, therefore, not involve BCL2 as a direct target.
In conclusion, our study has revealed a novel function of MITF-A and highlighted its crucial role in kidney morphogenesis and nephron endowment. Hence, it is conceivable that polymorphisms in the MITF gene might influence inter-individual differences in nephron number that one can observe in the human population. The fact that suboptimal nephron number has been shown to predispose people to hypertension and CKD points to MITF-A as a potential prognostic marker for identifying patients at risk of renal disease.
|
10.1371/journal.pgen.1007168 | Genetic variants in pachyonychia congenita-associated keratins increase susceptibility to tooth decay | Pachyonychia congenita (PC) is a cutaneous disorder primarily characterized by nail dystrophy and painful palmoplantar keratoderma. PC is caused by mutations in KRT6A, KRT6B, KRT6C, KRT16, and KRT17, a set of keratin genes expressed in the nail bed, palmoplantar epidermis, oral mucosal epithelium, hair follicle and sweat gland. RNA-seq analysis revealed that all PC-associated keratins (except for Krt6c that does exist in the mouse genome) are expressed in the mouse enamel organ. We further demonstrated that these keratins are produced by ameloblasts and are incorporated into mature human enamel. Using genetic and intraoral examination data from 573 adults and 449 children, we identified several missense polymorphisms in KRT6A, KRT6B and KRT6C that lead to a higher risk for dental caries. Structural analysis of teeth from a PC patient carrying a p.Asn171Lys substitution in keratin-6a (K6a) revealed disruption of enamel rod sheaths resulting in altered rod shape and distribution. Finally, this PC-associated substitution as well as more frequent caries-associated SNPs, found in two of the KRT6 genes, that result in p.Ser143Asn substitution (rs28538343 in KRT6B and rs151117600 in KRT6C), alter the assembly of K6 filaments in ameloblast-like cells. These results identify a new set of keratins involved in tooth enamel formation, distinguish novel susceptibility loci for tooth decay and reveal additional clinical features of pachyonychia congenita.
| Tooth decay, more commonly known as dental cavities, is the most common chronic disease worldwide, both in children and in adults. It consists in the destruction of tooth enamel, the outer layer of the teeth, by acid-producing bacteria. Enamel is the hardest tissue in the body, comprised of 96% minerals. However, it contains a small fraction of proteins that is important for its resistance to mechanical stress and decay. Here we show that this protein fraction contains a set of structural proteins (K6a, K6b, K6c, K16 and K17) that belong to the keratin family and are present specifically in the skin of the palms and soles, as well as in nails. We further show that common genetic mutations that affect the composition of these proteins lead to an increased number of cavities. Rare mutations in these keratins lead to a human disease called pachyonychia congenita (PC) and characterized by severe nail malformations and lesions in the skin of the palms and soles. Analysis of wisdom teeth from one of these patients showed that their enamel exhibited structural defects. These results demonstrate that these keratins are important components of tooth enamel and that common genetic variants in the genes that encode them influence tooth decay risk in the general population.
| Tooth enamel is made of 96% hydroxyapatite minerals, which makes it the hardest tissue in the human body. Enamel is also the first compartment of the tooth to be attacked by dental caries, a chronic disease that affects 42% of children and 92% of adults, with various degrees of severity (number of teeth and tooth surfaces affected) in the general population. Dental caries is initiated at the surface of the tooth by bacteria metabolizing food residues and releasing acids that dissolve enamel minerals [1]. Even though dental caries is influenced by environmental and behavioral factors, there is clear evidence that susceptibility to caries is also driven by host genetic factors [1–3], and genome-wide association studies (GWASs) have revealed genetic variants associated with increased susceptibility to tooth decay [4–8]. These genetic factors may influence the quality of dental tissues and ability to resist carious attacks, may impact other aspects of the oral environment such as the quality of the saliva, enamel pellicle and oral microbiome, and may differ between the primary and permanent dentitions [9].
Tooth enamel is produced in two phases [10]: first, a secretion phase during which enamel matrix proteins are secreted and deposited in a highly structured manner to form enamel rods; and second, a maturation phase during which most enamel matrix proteins are degraded to make space for the full expansion of hydroxyapatite minerals. After maturation, the enamel is left with only 1% of proteins that are abundant near the dentin-enamel junction (DEJ) and expand throughout the enamel as thin layers of enamel rod sheaths located at the interrod region [11]. This organic material has been shown to play a crucial role in the biomechanical properties of enamel [12, 13] and in the resistance to caries [14, 15]. Until recently, the exact composition of the insoluble fraction in this organic material had been uncharacterized, even though there was strong evidence that the highly crosslinked proteins present in mature enamel had biochemical properties similar to those of keratins [11, 16–18]. We showed previously that the organic material in mature enamel is in part composed of epithelial hair keratins, and that missense mutations in KRT75, previously linked to common hair disorders, were associated with increased susceptibility to dental caries [19].
In the present study, we focus on the presence in enamel of another set of keratins encoded by genes mutated in pachyonychia congenita (PC), a cutaneous disorder characterized by nail dystrophy and painful palmoplantar keratoderma [20, 21]. Additional features of this disease may include oral leukokeratosis, follicular keratosis, cysts, hyperhidrosis, and natal teeth. Some of these phenotypic traits are consistent with the expression pattern of the keratins involved. Herein we present novel findings that relate this set of keratins to the development of tooth enamel and to the susceptibility to tooth decay.
In a previous study, we determined through RNA-seq analysis that subsets of epithelial keratins were expressed in the enamel organ in mouse [19]. Of particular interest was the expression of Krt6a, Krt6b, Krt16 and Krt17 (Fig 1A), a set of keratin genes encoding keratin-6a (K6a), keratin-6b (K6b), keratin-16 (K16) and keratin-17 (K17), respectively, and in which mutations in humans lead to pachyonychia congenita (PC-K6a, OMIM #615726; PC-K6b, OMIM #615728; PC-K16, OMIM #167200; PC-K17, OMIM #167210), characterized by nail dystrophy and painful palmoplantar keratoderma [20, 21]. In humans, the KRT6 family includes a third member (KRT6C, encoding K6c), mutations in which have been associated with a milder form of PC with no/minor nail defects (PC-K6c) that was initially reported as palmoplantar keratoderma, non-epidermolytic, focal or diffuse (PPKNEFD, OMIM #615735).
Immunohistochemical analysis revealed that K6 (using an antibody recognizing all K6 proteins) and K17 are produced by rodent ameloblasts but exhibit very distinct distributions (Fig 1B and S1A Fig). K6 distribution is relatively diffuse in secretory ameloblasts (Fig 1B) while K17 forms characteristic keratin filaments that run throughout the ameloblasts and underlying tissues (stratum intermedium and papillary layer) (S1A Fig). At the apex of the ameloblasts and outside the Tomes’ processes, highly specialized structures where the deposition of enamel is coordinated, K6 was detected primarily at the interrod region (Fig 1B, magnification top panel). In the same area, K17 staining resulted in parallel transverse bands within the rods in addition to a longitudinal interrod distribution similar to that obtained with K6 staining (S1A Fig, magnification top panel). These distributions indicate that K6 and K17 are both incorporated into the enamel matrix but with distinct patterns. To determine whether these keratins were part of the organic material present in mature human enamel, we performed immunohistochemical staining on polished sections of human third molars (Fig 1C and S1B Fig). Consistent with its distribution near the apex of rodent ameloblasts, K6 was detected primarily where the enamel rod sheaths are located, at the periphery of the enamel rods (Fig 1C). More intense staining was detected at regular intervals near the DEJ, along structures that are likely to correspond to enamel tufts, areas of higher accumulation of organic material (Fig 1C, left panel). In addition to intense staining near the DEJ, K6 was detected throughout the thickness of enamel and restricted to the periphery of the rods (Fig 1C, right panel). K16 and K17 were detected near the DEJ where they were not restricted to the interrod regions but also present in the core of the rods (S1B Fig), a pattern consistent with the distribution of K17 near the apex of rodent ameloblasts. The restricted pattern of K6 distribution at the enamel rod sheaths was confirmed with two different antibodies, a polyclonal antibody raised in guinea-pig against the C-terminus of the protein and a monoclonal antibody raised in mouse against the N-terminus (S2 Fig). These results indicate that PC-associated keratins are part of the organic material present in mature enamel but exhibit distinct distributions.
In order to determine if the presence of K6a, K6b, K6c, K16 and K17 in mature human enamel had an impact on the susceptibility to tooth decay, we tested the association between SNPs in the genes encoding these keratins and three measures of dental caries experience assessed in the primary dentition of 449 children (mixed European descent, 6–12 years) and permanent dentition of 573 adults (mixed European descent, 25–50 years). We focused our attention to common missense SNPs that occur at a sufficient frequency (minor allele frequency > 1%) allowing for statistical testing in our unselected population-based cohort. Three missense SNPs in KRT6A, eight in KRT6C, and seven in KRT6B responded to these criteria (Table 1). Across all 18 missense SNPs, seven SNPs showed nominal evidence of association (p < 0.05) with at least one measure of dental caries experience in either adults or children, and the following five SNPs exhibited associations with all three measures of dental caries experience (Table 1):
The SNPs identified in KRT6A and KRT6C were associated with increased caries experience in adults only. Among the missense polymorphisms identified in KRT6B, rs61746354 (K6bY497C) was associated with higher caries experience in children, while rs144860693 (K6bG97R) and rs28538343 (K6bS143N) were associated with higher caries experience in adults (Table 1). These results indicate that the effect of specific polymorphisms in keratin genes may differ across dentition (primary vs. permanent).
Only one common missense SNP in KRT16, rs111383277 (KRT16:c.121C>T; K16p.Arg41Cys), was at a frequency higher than 1%, while none were found in KRT17. rs111383277 did not show significant association with dental caries experience in the cohorts tested. Due to this limited number of common SNPs in KRT16 and KRT17, we were not able to conclude on the potential implication of these two keratins in caries risk.
Keratins are structured into three major domains with a central “rod” domain, directly involved in the dimerization and further assembly of keratin filaments, flanked by a “head” domain and a “tail” domain on the N-terminal side and C-terminal side, respectively (Fig 1D). Interestingly, all the missense polymorphisms that showed significant association with higher caries experience in KRT6A, KRT6B and KRT6C result in amino acid substitutions in the head or tail domains, while all the mutations that have been associated so far with PC are located at the beginning or at the end of the rod domain (Fig 1D). The KRT6B polymorphism associated with higher caries experience in children (rs61746354, K6bY497C) is the only SNP that results in an amino acid substitution in the tail domain (Fig 1D). The missense SNPs in KRT6 genes were present at various frequencies in the cohorts studied (S1 Table). Moderate to high linkage disequilibrium (R2 between 0.69 and 0.89) was observed between rs17845411 (K6aN21S), rs151117600 (K6cS143N), rs144860693 (K6bG97R), and rs28538343 (K6bS143N) (Fig 1E).
Genotype frequencies and quantifications of caries experience per genotype group for the three missense SNPs identified in KRT6B are shown in Fig 2. The frequencies of rs144860693 (K6bG97R) and rs28538343 (K6bS143N), the two variants that exhibited the most significant association with dental caries risk in adults, are high in the cohorts studied (Fig 2A and 2B; S1 Table). These two SNPs have a major impact on caries experience, with an estimated increase in the average number of carious tooth surfaces of 1.6 and 2.4 surfaces per copy of the risk allele, respectively (Table 1). These variants did not demonstrate a statistically significant effect on average caries experience in children (Fig 2A and 2B). rs61746354 (K6bY497C), the missense SNP in KRT6B that was associated with higher caries risk in children and occurs at a frequency higher that 4% in our cohorts (S1 Table), was associated with an estimated 1-surface increase in the average number of carious tooth surfaces (Table 1 and Fig 2C). Genotype frequencies and quantifications of caries experience per genotype for the other missense SNPs identified in KRT6A and KRT6C are shown in S3 Fig.
Given that KRT6B harbors three missense SNPs showing significant association with caries experience, we wanted to further quantify the genetic relationship of missense variants in this gene on dental caries. To do so, we examined pairwise interactions between rs144860693 (K6bG97R), rs28538343 (K6bS143N) and rs61746354 (K6bY497C). Statistically significant interaction effects were observed between rs144860693 (K6bG97R) and rs28538343 (K6bS143N) on the number of surfaces with untreated decay (DS) model, between rs144860693 (K6bG97R) and rs61746354 (K6bY497C), and rs28538343 (K6bS143N) and rs61746354 (K6bY497C) on the number of decayed, missing due to decay, and filled surfaces (DMFS), adjusting for age, sex, and all the other SNPs in KRT6B (S2 Table). Even though rs61746354 (K6bY497C) was associated with higher caries risk in children only, this SNP exhibited a significant statistical interaction effect with rs144860693 (K6bG97R) and rs28538343 (K6bS143N) in adults (S2 Table). Therefore, the effect of the two SNPs that result in amino acid substitutions in the head domain of K6b on caries risk in adults may be influenced by the presence or absence of the p.Tyr497Cys substitution in the tail domain of the same keratin, a SNP that by itself is associated to higher caries risk only in children. The two SNPs resulting in p.Ser143Asn substitution in KRT6B and KRT6C (rs28538343 and rs151117600, respectively) also exhibited statistically significant interaction effect on the number of surfaces with untreated decay (DS) in adults (p-value = 0.044). When focusing on the 4 SNPs that lead to higher caries risk in adults, we found a significant cumulative effect of the number of risk alleles on caries experience (S4 Fig).
Given that KRT75 is adjacent and phylogenetically related to the KRT6 genes in the human genome, we explored potential linkage disequilibrium and interaction effects between the KRT75 SNP previously shown to increase caries experience in adults [19] and the newly identified SNPs in KRT6 genes. The previously reported SNP rs2232387 (K75A161T) was not in linkage disequilibrium with any of the KRT6 SNPs (S5 Fig) and there was no statistical interaction (all p-values >0.05) between the same SNPs.
Altogether, our data support genetic association between SNPs in KRT6A, KRT6B and KRT6C and tooth decay risk, in a way that is dentition-specific, and with statistical interaction between various loci in these three genes.
In order to assess how mutations in KRT6 genes may affect enamel structure, we analyzed third molars that were extracted from a PC patient who is heterozygous for the KRT6A:c.513C>A transversion that results in p.Asn171Lys amino acid substitution (K6aN171K) (Fig 3A). This patient is a white male who was 18 years old at the time his third molars were extracted, and is the member of a family in which the mutation in KRT6A was previously reported [22]. The patient experienced 20/20 nail dystrophy, very painful palmoplantar keratoderma, oral leukokeratosis, follicular keratosis, but did not have natal teeth. The overall shape and structure of the third molar enamel did not appear defective based on micro-computed tomography analysis (Fig 3B). However, scanning electron microscopy analysis of polished sections of the teeth (section plane transverse to the enamel rods) revealed alteration of the distribution and shape of enamel rods when compared to third molars extracted from healthy patients (Fig 3C).
The insoluble organic material present in mature enamel can be isolated after full demineralization of a tooth in EDTA. When isolated from molars extracted from this PC patient, the insoluble material exhibited uneven alignment of the enamel rod sheaths that tended to form curls (Fig 3D). To assess the effects of the K6aN171K mutant protein on K6 distribution in enamel, we performed immunohistochemical staining using anti-K6 antibody on polished sections of the patient’s tooth. K6 staining was still found to be stronger in the tuft areas near the DEJ (Fig 3E). However, K6 distribution was no longer restricted to the interrod but could also be found as smaller rings or clumps within rods (Fig 3E). These results indicate that this PC-causing mutation in KRT6A leads to improper incorporation of the K6a protein into enamel rod sheaths, which results in altered shape and arrangement of enamel rods.
Missense mutations in keratins may affect their assembly, modify their subcellular localization and/or affect their interaction with keratin-associated proteins. Phosphorylation and glycosylation of the head and tail domain of intermediate filaments proteins have been shown to influence their interaction with other proteins and their subcellular localization [23, 24]. When comparing the position of the SNPs we determined to be associated with increased caries experience and potential sites for post-translational modifications in K6 proteins, we observed that the p.Ser143Asn substitution (rs28538343 in KRT6B and rs151117600 in KRT6C) is immediately adjacent to an LLS/TPL consensus phosphorylation site that is highly conserved in type II keratins [24, 25], and within a potential N-linked glycosylation site (Fig 4A and S6 Fig). Although it remains to be determined how K6 proteins interact with and are deposited into the enamel matrix in the context of a secretory stage ameloblast in vivo, we assessed the effect of the p.Ser143Asn substitution in K6 in a context in which the mutant protein is overexpressed in ameloblast-like cells (ALC) [26]. In this assay, we also analyzed the behavior of K6N171K mutant protein carried by the PC patient included in this study (Fig 3), a mutation located in the rod domain and known to have a severe effect on keratin filament assembly [27, 28]. Given the high degree of sequence identity between K6 proteins (S7 Fig) and the fact that the mutations of interest are located in highly conserved regions (Fig 4A), we used K6a as a model protein for this assay. We used site-directed mutagenesis to introduce the c.428G>A transition (results in p.Ser143Asn substitution) and the c.513C>A transversion (results in p.Asn171Lys substitution) into the KRT6A cDNA, and cloned the different isoforms (KRT6AWT, KRT6Ac.428G>A and KRT6Ac.513C>A) into a vector that allows for tetracycline inducible co-expression of KRT6A isoforms and GFP (Fig 4B).
These constructs were used to transfect ALC-TetON cells in which expression of KRT6A isoforms and GFP can be induced by addition of doxycycline to the culture medium (Fig 4C). Immunohistochemical analysis using anti-K6 antibody was used to determine the distribution of K6a isoforms in ALC-TetON cells (Fig 4D). While K6aWT formed thick and relatively short bundles of keratin filaments in ALC-TetON cells, K6aS143N tended to form a web of thinner filaments together with large aggregates (Fig 4D). These large aggregates were not seen with K6a proteins harboring the p.Asn21Ser and p.Gly97Arg substitutions (S8A and S8B Fig) caused by the other SNPs that are associated with higher caries in adults (rs17845411 and rs144860693, respectively) and are in partial linkage disequilibrium with the SNPs leading to p.Ser143Asn substitution (rs28538343 in KRT6B and rs151117600 in KRT6C). These results suggest that the p.Ser143Asn substitution may contribute most significantly to the caries-prone phenotype in adults. The behavior of the K6aN171K mutant protein fused to a YFP tag has been previously studied in the context of human hepatoma PLC cells in which the mutant protein was shown to form aggregates primarily located in the cytoplasm [27, 28]. In the context of ALC-TetON cells, K6aN171K formed aggregates that showed heightened accumulation in the nucleus (Fig 4D). These results confirm a severe impairment of K6a assembly in PC patients with p.Asn171Lys substitution. The fact that the aggregates in PLC cells were mostly in the cytoplasm may reflect a cell-specific behavior of the mutant protein or may be due to the YFP tag that was fused to K6a in these experiments [27, 28].
Given that the SNP leading to p.Tyr497Cys substitution in the tail domain of K6b (rs61746354) is the only one we found associated with higher caries risk in children, we tested its effect on K6a assembly. Similarly to K6aS143N, the K6aY497C isoform tended to form large aggregates in ALC-TetON cells (S8C Fig), which suggests that this SNP may be the cause of the caries-prone phenotype in children. Even though this substitution is not found near a potential posttranslational modification site, the presence of a new cysteine in the tail domain may result in the formation of disruptive disulfide bonds. Given the interaction effects measured between the SNPs that lead to the p.Ser143Asn and p.Tyr497Cys substitutions, we generated a DNA construct for the expression of a K6a protein that harbors both substitutions (K6aS143N-Y497C). This double mutant tends to form aggregates to the same extent as the single mutant proteins (S8D Fig).
Taken together, these results confirm that the PC-associated p.Asn171Lys substitution results in profound impairment of K6a protein assembly, and reveal that the caries-associated p.Ser143Asn and p.Tyr497Cys substitutions in K6 proteins also affect the behavior of the proteins, when overexpressed in an ameloblast cell line.
The present report highlights the contribution of specific sets of keratins to the organic fraction of mature tooth enamel and demonstrates through genetic and analytical studies their crucial function in the formation of enamel and its resistance to decay. K75, an epithelial hair keratin in which mutations have been associated with hair disorders, was the first keratin we investigated in this context [19]. K6 proteins, that are the focus of the present study, are expressed in epithelia that withstand particularly high levels of mechanical strain (palmoplantar skin, oral epithelium) as well as in the supporting layers of the hair follicle where their function is similar to that of K75 which is not expressed in palmoplantar epidermis and oral epithelium. Our findings demonstrate that, as K75, K6 proteins play a crucial role in the enamel rod sheath and that mutations in the genes encoding these keratins may impair the stability of the organic structural component of mature enamel. We propose that, with their unique biochemical properties, K75 and K6 contribute to the toughness, elasticity and resistance to degradation of the enamel rod sheaths, which contributes to establishing proper shape and arrangement of enamel rods and enhances the biomechanical properties of tooth enamel [12, 13]. Moreover, since it has long been suggested that the stability of the proteins in mature enamel influences the resistance of enamel to carious attack [14, 15], we propose that keratins contribute to the stability of enamel rod sheaths and therefore to the resistance of enamel to decay.
Since K6 proteins are also expressed in the oral epithelium, and patients with PC may exhibit oral leukokeratosis, there could be a partial involvement of the oral cavity in the increased susceptibility to caries measured in this study. However, the fact that we found SNPs that lead to a higher number of caries in children and not in adults (same oral cavity but different set of teeth) strongly suggests that defects in the dental tissue itself are the major factor leading to this effect. The structure and chemical composition of tooth enamel is known to be different between primary and permanent teeth. Primary teeth exhibit thinner and whiter enamel with a smoother surface and higher content in calcium and phosphate when compared to permanent teeth [29, 30]. Enamel from primary teeth also has a greater susceptibility to demineralization [31]. Moreover, it has been proposed that the genetic factors influencing dental caries differ between primary and permanent dentition [9]. However, there has been no study comparing the composition of the organic material present in the enamel from primary and permanent teeth. The dentition-specific effect we report here for SNPs in KRT6 genes, which we previously observed for two SNPs in KRT75, with one affecting adults and the other one affecting children [19], suggests that the combination and/or the mode of incorporation of these keratins in the enamel rod sheaths is different in primary and permanent teeth.
Even though this is the first evidence of K6 proteins being incorporated into the enamel matrix, a previous yeast-two-hybrid study determined that K6 could interact with enamel matrix proteins such as amelogenin and tuftelin [32]. It is therefore likely that K6 proteins interact with enamel matrix proteins during the process of enamel secretion. However, the mode of incorporation of keratins into the enamel matrix remains to be elucidated. The interaction of keratins with other proteins is known to involve their head and tail domains rather than the rod domain through which heterodimerization of acidic and basic keratins is established. These interactions are regulated by posttranslational modifications such as phosphorylation and glycosylation [23, 24], and mutations impairing such modifications have been linked to skin diseases [33], as well as diseases related to liver and pancreatic injury [34]. Interestingly, all the caries-associated missense SNPs we identified in KRT6 genes result in substitutions in the head and tail domains of the proteins, which suggests that they may affect their interaction with other proteins rather than their heterodimerization. We further demonstrate that the p.Ser143Asn substitution that may affect phosphorylation and/or glycosylation of the head domain of K6 proteins affects the behavior of K6A in the context of an ameloblast cell line, which suggests that this substitution found in both K6b and K6c may contribute most significantly to the caries-prone phenotype in adults. Functional studies will be required to elucidate the effects of the p.Ser143Asn substitution on the biochemical properties of K6 proteins, in particular on its ability to undergo posttranslational modifications that may affect interaction with enamel matrix proteins and incorporation into the enamel in vivo. Interestingly, in a recent clinical report, an isolated case of PC was proposed to be caused by de novo c.428G>A mutation in KRT6A that leads to the p.Ser143Asn substitution [35]. This is so far the only report of PC-causing mutation outside of the rod domain. Based on its location in the head domain and on the high frequency of the same substitution in K6b and K6c, the p.Ser143Asn substitution in K6a is unlikely to be the sole cause for the PC phenotype in this patient.
In epidermal tissues, K6 proteins (Type II, basic or neutral) form heterodimers with K16 or K17 (Type I, acidic) to assemble in larger polymeric structures. Interestingly, the subcellular distribution of K16 and K17 proteins is distinct from the distribution of K6 in the enamel organ, which suggests that they do not follow their canonical mode of assembly in this tissue. Due to the low number of frequent missense SNPs in KRT16 and KRT17 in our cohorts, the present study did not allow us to make any conclusion on the potential genetic association between variants in these two genes and dental caries experience. However, the striking difference in the way K16 and K17 proteins are incorporated into enamel suggests that their function in this tissue is distinct from that of K6 proteins. Based on the restricted localization of K16 and K17 near the DEJ and in the core of the enamel rods, these keratins may be involved in shock absorption and protection against fracture [13] rather than in the resistance to caries. Structural analysis of enamel from PC patients with mutations in KRT16 and KRT17 will help address this question.
In conclusion, we show for the first time that (i) K6 proteins are incorporated into mature tooth enamel at the rod sheaths, (ii) SNPs in KRT6 genes are associated with increased susceptibility to dental caries, (iii) a PC patient with a mutation in KRT6A exhibits defects in enamel structure, and (iv) caries-associated p.Ser143Asn substitution in K6 proteins impairs proper protein interactions.
We thank the Pachyonychia Congenita Project and Ms. Holly Evans for providing us with clinical information and extracted third molars from a PC patient (20040468–1057496), and the NIDCR dental clinic for providing extracted third molars from healthy patients (NCT01805869).
For the COHRA study, written informed consent was provided by all adult participants, and verbal assent with parental written consent was provided by all child participants. All procedures, forms and protocols were approved by the Institutional Review Boards of the University of Pittsburgh and West Virginia University.
Written informed consent was obtained from the pachyonychia congenita patient, as part of a research registry approved by an institutional review board that complies with all principles of the Helsinki Accord (Western IRB study no. 20040468).
All animal work was approved by the NIAMS Animal Care and Use Committee.
The Center for Oral Health Research in Appalachia (COHRA) study was initiated to investigate the community-, family-, and individual-level contributors to oral health [36]. Participants from rural counties of Pennsylvania and West Virginia were enrolled via a household-based recruitment strategy, with eligible households required to include at least one biological parent-child pair. All other members of eligible households were invited to participate without regard to biological or legal relationships, or oral health status. Written informed consent was provided by all adult participants, and verbal assent with parental written consent was provided by all child participants. All procedures, forms and protocols were approved by the Institutional Review Boards of the University of Pittsburgh and West Virginia University.
Intra-oral examinations of all participants were performed by licensed dentists and/or research dental hygienists. Each surface of each tooth (excluding third molars) was examined for evidence of decay, from which dental caries indices were generated. Three measurements of caries experience were considered: (1) the number of surfaces with untreated decay (DS/ds); (2) the traditional DMFS/dfs indices which represent the number of decayed (D/d), missing due to decay (M), and filled (F/f) tooth surfaces (S/s) in the permanent (DMFS) and primary (dfs) dentitions; and (3) the partial DMFS and dfs indices limited to the molars and premolar pit and fissure surfaces which are at high risk of decay. DNA samples were collected via blood, saliva or buccal swab. Genotyping for approximately 600,000 polymorphisms was performed by the Center for Inherited Disease Research at Johns Hopkins University using the Illumina Human610-Quadv1_B BeadChip (Illumina). Extensive data cleaning and quality assurance analyses were performed as previously reported [4]. Imputation of approximately 16 million unobserved genetic variants was performed using the 1000 Genome Project (phase 1 June 2011 release) as reference. In brief, pre-phasing of haplotypes was performed via SHAPEIT2 [37] and imputation was performed via IMPUTE2 [38].
Linear regression was used to test the association of dental caries experience with genetic polymorphisms under the additive genetic model while adjusting for age and sex. Pairwise SNP-by-SNP interaction effects were tested in the same modeling framework by including main effects of each SNP and their product term for selected variants within the same gene region, along with age and sex. Analyses were performed separately for dental caries in the permanent dentition of adults (ages 25–50 years) and the primary dentition of children (ages 6–12 years). All analyses were limited to self-reported non-Hispanic whites (mixed European descent); self-reported race was consistent with genetically-determined ancestry. Analyses were performed using PLINK (v1.9) (http://www.cog-genomics.org/plink/1.9/)[39] and R (R Foundation for Statistical Computing).
Third molars were obtained from a patient involved in the International Pachyonychia Congenita Research Registry (IPCRR), under the IRB Protocol number 20040468 and IRB Study number 1057496. Third molars from healthy patients were obtained from the NIDCR OP-1 Dental Clinic that were collected under the IRB Protocol number NCT01805869.
RNA-seq analysis was performed as described previously [40]. Briefly, mandibles were dissected from P10 mice, transferred to RNAlater solution (Life Technologies) and stored at 4°C. Enamel organs were dissected from mandibles and homogenized in Trizol reagent (Invitrogen) using the Precellys 24 (Bertin Technologies). Total RNA was extracted and further purified using the RNAeasy mini kit (Qiagen). RNA-seq analysis was performed using the Mondrian SP kit (Illumina) and the Illumina HiSeq 2000 sequencing system.
Rat mandibles at postnatal day 10 were fixed overnight at 4°C in 4% paraformaldehyde in 1X PBS, dehydrated and embedded in paraffin and 10 μm-thick sections were prepared. Immunohistochemical analysis was performed using a blocking solution containing 5% goat serum and 7.5% BlockHen II (Aves Labs, Tigard, OR) in 1X PBS. Enzymatic antigen retrieval was performed using Ultravision Trypsin (Thermo Fisher Scientific, Waltham MA). Primary antibodies used: Guinea-pig anti-K6 (Progen Biotechnik GmbH, Germany), guinea-pig anti-K17 (Progen Biotechnik GmbH, Germany). Alexa-488 anti-guinea-pig (Thermo Fisher Scientific, Waltham MA) was used as secondary antibody. Images were acquired on a Leica LS5 confocal microscope (Leica Microsystems Inc., Buffalo Grove, IL).
Ground, polished and etched human teeth were stained with guinea-pig anti-K6 (Progen Biotechnik GmbH, Germany), mouse anti-K6 (Abcam, Cambridge MA), guinea-pig anti-K16 (Progen Biotechnik GmbH, Germany) or guinea-pig anti-K17 (Progen Biotechnik GmbH, Germany) antibody. Alexa-488 anti-guinea-pig and Alexa-555 anti-mouse (Thermo Fisher Scientific, Waltham MA) were used as secondary antibodies.
Ground, polished and etched human teeth were prepared for SEM as described previously [19]. Samples were fixed overnight at 4°C in 2% glutaraldehyde, 2% PFA in 0.1M phosphate buffer pH 7.4 and dehydrated through a series of 50%, 70%, 95% and 100% ethanol solutions. They were incubated for 10 min in hexamethyldisilazane, air-dried for 30 min, mounted on aluminum specimen mount stubs covered with conductive carbon adhesive tabs (Electron Microscopy Sciences, Hatfield, PA), sputter-coated with gold and analyzed under a Field Emission Scanning Electron Microscope S4800 (Hitashi, Toronto, Canada) at 10 kV.
Micro-CT analysis of fixed molars was performed as described previously [19] using the Skyscan 1172 desktop X-ray microfocus CT scanner and the following parameters: 0.5mm AI + 0.1mm Cu filters, 100 kV, 100 uA, 8.00 micron resolution, 0.4 degrees rotation step over 180 degrees). CTvox software (Bruker microCT) was used for 3D reconstruction.
The cDNAs encoding the, K6aN21S, K6aG97R, K6aS143N, K6aY497C and K6aN171K mutant proteins were generated by site-directed mutagenesis using the QuikChange Site-directed Mutagenesis Kit (Stratagene). The following primers were used: N21S-forward: GGGGTTTCAGTGCCAgCTCAGCCAGGC, N21S-reverse: GCCTGGCTGAGcTGGCACTGAAACCCC, G97R-forward: GGCTTTGGTGGCGCCaGGAGTGGATTGG, G97R-reverse: CCAATCCACTCCtGGCGCCACCAAAGCC, S143N-forward: GTCAACCAGAaTCTCCTGACTCCCCTC, S143N-reverse: GAGGGGAGTCAGGAGAtTCTGGTTGAC, Y497C-forward: CCGTCTCCAGTGGCTgTGGCGGTGCCAG, Y497C-reverse: CTGGCACCGCCAcAGCCACTGGAGACGG, N171K-forward: GATCAAGACCCTCAAaAACAAGTTTGCC, N171K-reverse: GGCAAACTTGTTtTTGAGGGTCTTGATC (lower cases indicate the position of the mutations). The pBi4-GFP vector was used to simultaneously express the reporter protein EGFP with K6aWT, K6aN21S, K6aG97R, K6aS143N, K6aY497C, K6aS143N-Y497C, or K6aN171K under control of a unique tetracycline responsive element.
Murine Ameloblast-like cells (ALC) were kindly provided by Dr. Sugiyama [26] and used to produce a tetracycline inducible ameloblast cell line. These cells were stably transfected with a prtTA2-M2/IRES-Neo plamid obtained after subcloning of the rtTA2-M2 cassette [41] into the pCMV-IRES-Neo (Clontech). rtTA-M2 is a mutagenized form of rtTA that shows a lower basal activity and a higher sensitivity to doxycycline (Dox) than the original rtTA [41]. The presence of the IRES cassette before the neomycin (Neo) resistance gene allowed coexpression of the rtTA-M2 transactivator and the Neo resistance gene, increasing the chance to select clones that express sufficient amounts of the transactivator in Neo-resistant cells. Clones were isolated and functionality of the tet system was screened by transient transfection with pTRE2-luc (expression of luciferase under the control of tetracycline responsive element). Cells were grown in the presence or absence of 2 ug/ml Dox and luciferase activity was estimated. One clone exhibiting a low basal activity of the transactivator in the absence of Dox and a strong induction in the presence of Dox (+Dox/-Dox ratio) was selected for subsequent experiments and named ALC-TetON. ALC-TetON cells were grown in Dulbecco’s modified Eagle’s medium (10% fetal bovine serum, 1% penicillin/streptomycin and 1 ug/ml G418). For transfections, the cells were grown to at least 70% confluence. 2 million cells were used per transfection with either the pBi4-GFP/K6aWT, pBi4-GFP/K6aN21S, pBi4-GFP/K6aG97R, pBi4-GFP/K6aS143N, pBi4-GFP/K6aY497C, pBi4-GFP/K6aS143N-Y497C, or pBi4-GFP/K6aN171K plasmid and the pCMV-K16WT plasmid (Amaxa Nucleofactor).
Transfected cells were seeded on glass coverslips coated with 0.1% gelatin and immediately induced with 2ug/ml doxycycline. Twenty-four hours after induction, cells were washed three times in 1X PBS and fixed with 4% paraformaldehyde in PBS for 15 min at room temperature. A 5 min incubation in 0.2% Triton in 1X PBS was used to permeabilize the cells before blocking unspecific sites using 3% BSA in PBS for 1 h. Primary antibodies diluted in blocking solution were applied for 1 h. Primary antibody used: guinea-pig anti-K6 (Progen). Secondary antibodies diluted in blocking solution were applied for 30 min. Secondary antibodies used: Alexa Fluor 555 anti-guinea pig IgG (Life Technologies). Nuclei were stained using DAPI and coverslips were mounted on glass slides using Mowiol (Calbiochem). Images were acquired using a LEICA Sp5 confocal microscope.
|
10.1371/journal.ppat.1002498 | The Intracellular DNA Sensor IFI16 Gene Acts as Restriction Factor for Human Cytomegalovirus Replication | Human interferon (IFN)-inducible IFI16 protein, an innate immune sensor of intracellular DNA, modulates various cell functions, however, its role in regulating virus growth remains unresolved. Here, we adopt two approaches to investigate whether IFI16 exerts pro- and/or anti-viral actions. First, the IFI16 gene was silenced using specific small interfering RNAs (siRNA) in human embryo lung fibroblasts (HELF) and replication of DNA and RNA viruses evaluated. IFI16-knockdown resulted in enhanced replication of Herpesviruses, in particular, Human Cytomegalovirus (HCMV). Consistent with this, HELF transduction with a dominant negative form of IFI16 lacking the PYRIN domain (PYD) enhanced the replication of HCMV. Second, HCMV replication was compared between HELFs overexpressing either the IFI16 gene or the LacZ gene. IFI16 overexpression decreased both virus yield and viral DNA copy number. Early and late, but not immediate-early, mRNAs and proteins were strongly down-regulated, thus IFI16 may exert its antiviral effect by impairing viral DNA synthesis. Constructs with the luciferase reporter gene driven by deleted or site-specific mutated forms of the HCMV DNA polymerase (UL54) promoter demonstrated that the inverted repeat element 1 (IR-1), located between −54 and −43 relative to the transcription start site, is the target of IFI16 suppression. Indeed, electrophoretic mobility shift assays and chromatin immunoprecipitation demonstrated that suppression of the UL54 promoter is mediated by IFI16-induced blocking of Sp1-like factors. Consistent with these results, deletion of the putative Sp1 responsive element from the HCMV UL44 promoter also relieved IFI16 suppression. Together, these data implicate IFI16 as a novel restriction factor against HCMV replication and provide new insight into the physiological functions of the IFN-inducible gene IFI16 as a viral restriction factor.
| Only recently, intrinsic cellular-based defense mechanisms which give cells the capacity to resist pathogens have been discovered as an essential component of immunity. However, unlike the innate and adaptive branches of the immune system, intrinsic immune defenses are mediated by cellular restriction factors that are constitutively expressed and active even before a pathogen enters the cell. The protein family HIN-200 may act as sensors of foreign DNA and modulate various functions such as growth, apoptosis, and senescence. Here we show that, in the absence of functional IFI16, the replication of some Herpesviruses and in particular of Human Cytomegalovirus (HCMV) is significantly enhanced. Accordingly, IFI16 overexpression strongly inhibited HCMV replication. Accumulation of viral DNA copies was down-regulated along with expression of early and late viral gene expression suggesting that IFI16 inhibits viral DNA synthesis. Using transient transfection, luciferase, gel shift assay, and chromatin immunoprecipitation, we demonstrate that IFI16 suppresses the transcriptional activity of the viral DNA polymerase gene (UL54) and the UL44 gene, also required for viral DNA synthesis. The finding that the nuclear DNA sensor IFI16 controls virus growth represents an important step forward in understanding the intrinsic mechanisms that drive viral infections sustained by DNA viruses such as Herpesviruses.
| Human Cytomegalovirus (HCMV) is a β-Herpesvirus that commonly and persistently infects humans [1]–[2]. HCMV does not constitute a serious threat to immunocompetent individuals, but causes life-threatening complications in individuals with suppressed immune systems, such as patients with AIDS, cancer patients undergoing chemotherapy, and organ transplant recipients treated with immunosuppressants [3]. Viral gene expression in HCMV undergoes sequential regulation, which leads to the occurrence of induction and repression cycles in the immediate early (IE), early (E), and late (L) phases of viral replication. IE1 and IE2 induce the expression of early protein, mediate G1/S cell cycle arrest and host replication shut-off [4]–[5].
Many mammals, including humans, are equipped with genes encoding so-called “restriction factors” that provide considerable resistance to viral infection [6]. Such intrinsic immune mechanisms are highly important as they provide an antiviral frontline defense mediated by constitutively expressed proteins, already present and active before a virus enters a cell [7]–[8]. These intrinsic immune mechanisms were initially discovered as being active against retroviruses and involve the APOBEC3 class of cytidine deaminases, a large family of proteins termed the TRIM family, and tetherin, an interferon-inducible protein whose expression blocks the release of HIV-1. It has recently emerged, however, that such intrinsic immune mechanisms are also active against other viruses, such as Vesicular Stomatitis Virus, Filoviruses, Influenza Virus, and Hepatitis C Virus [9]. Moreover, four cellular proteins – promyelocitic leukemia protein (PML), hDaxx, Sp100 [10], and viperin – have been identified as restriction factors involved in mediating intrinsic immunity against HCMV infection [11]. PML and hDaxx are components of a subnuclear structure known as nuclear domain 10 (ND10) or PML nuclear bodies. Direct evidence for their antiviral role has been obtained from infection studies using cells devoid of intact ND10. Primary human fibroblasts depleted of PML using small interfering RNA (siRNA) significantly increased the plaque forming efficiency of HCMV as a result of an augmented immediate-early (IE) gene expression [12]. hDaxx represses HCMV IE gene expression [11], [13], [14] and replication [12] through the action of a histone deacetylase (HDAC), thereby inducing a transcriptionally inactive chromatin state around the major IE enhancer/promoter (MIEP) of HCMV [13]–[14]. Together, these findings revealed that ND10 proteins, PML and hDaxx, act as cellular restriction factors that are able to induce silencing of HCMV gene expression, thus controlling virus replication.
Of the various interferon-inducible proteins, the p200 family of proteins, now designated the PYHIN family, consists of a group of homologous human and mouse proteins that have an N-terminal PYRIN domain (PYD) and one or two partially conserved 200 amino acid long C-terminal domains (HIN domain). These proteins display multifaceted activity due to their ability to bind to various target proteins (e.g. transcription factors, signaling proteins, and tumor suppressor proteins) and modulate different cell functions. Increasing evidence supports a role for these proteins as regulators of various cell functions, including proliferation, differentiation, apoptosis, senescence, and inflammasome assembly, as well as in the control of organ transplants (reviewed in [15]–[16]). More recently, two members of the PYHIN family, namely AIM2 and IFI16, have been shown to bind to and function as pattern recognition receptors (PRR) of virus-derived intracellular DNA [17]–[20]. In particular, IFI16 has been shown to interact with the adaptor molecule ASC and procaspase-1 to form a functional inflammasome during Kaposi Sarcoma-Associated Herpesvirus (KSHV) infection. KSHV gene expression is required for inflammasome activation and IFI16 colocalizes with the KSHV genome in infected cell nuclei [20]. Moreover, the induction of IRF3 and NF-κB-dependent genes by HSV-1 infection of RAW264.7 cells is strongly impaired by siRNA specific for p204, the mouse ortholog of IFI16 [19]. However, although many different functions have been ascribed to these proteins, their roles as antiviral restriction factors have yet to be investigated, whilst such roles have long been established for other IFN-inducible proteins, such as, PKR, 2′-5′ oligoadenylate synthetase, and MxA [21]–[22].
In the present study, by either silencing or overexpressing the IFI16 protein, we demonstrate for the first time that IFI16 acts as a restriction factor for Human Cytomegalovirus (HCMV) replication. Transfection and electrophoretic mobility shift assay (EMSA) experiments, performed using nuclear extracts from HCMV infected cells, showed that the promoter of the DNA polymerase gene (UL54) is the target of IFI16-induced viral suppression. Finally, biochemical and immunochemical analyses reveal that IFI16 inhibits DNA polymerase gene activity by preventing the binding of transcription factor Sp1 to the −54/−43 IR-1 promoter element.
To determine the effects of IFI16 silencing on virus replication, IFI16-knockdown HELFs were infected with HCMV (AD169 strain), HSV-1, HSV-2, ADV, or VSV at an MOI of 0.05. As shown in Figure 1A, the replication of HSV at 48 hpi and of HCMV at 144 hpi was increased in IFI16-silenced HELFs. In contrast, the replication of a clinical isolate of Adenovirus and that of a VSV laboratory strain was not significantly affected (Figure 1A). Consistent with previous data, Western blot analysis confirmed that the electroporation of HELFs with four IFI16 specific siRNAs knocked IFI16 expression down by more than 90% for at least 14 days (Figure 1B).
To provide further evidence supporting the physiological relevance of IFI16 in the control of Herpesvirus replication, we generated HELFs overexpressing wild type (wt) or mutated IFI16 proteins using recombinant lentiviral vectors, or the LacZ gene as a control. These V5-tagged IFI16 proteins bear deletions between residues 1–83 (ΔPYDIFI16, indicated as ΔDIFI16) or between residues 515–710 (ΔHIN-BIFI16, indicated as ΔBIFI16), respectively, and thus held the potential to inhibit the activity of the endogenous counterpart. Expression of the exogenous V5-tagged IFI16 was confirmed using anti-V5 antibodies that recognized proteins of about 82 kDa (wtIFI16), 70 kDa (ΔDIFI16) or 60 kDa (ΔBIFI16), and 121 kDa in the control (LacZ) (Figure 2A).When overexpressed in stably-transfected cell lines, both ΔDIFI16 and ΔBIFI16 decreased the ability of full length IFI16 (AdV IFI16) to induce proinflammatory molecules and to trigger caspase-3 and 7 activity (Figure S1 panel A and B respectively). These results suggest that the mutant IFI16 proteins may behave as dominant negative (dn) towards the endogenous counterpart.
The effects of ΔDIFI16 and ΔBIFI16 deleted mutants on viral growth were then examined by comparing the ability of HELFs transfected with either the mutated forms of IFI16 or the full length protein to support HCMV complete replication. In line with the results of the previous experiments that used siRNAs to silence IFI16, expression of the ΔDIFI16 protein substantially increased the extent of HCMV replication (Figure 2B) at both the MOIs used. Expression of ΔBIFI16 showed a level of virus yield similar to that observed in HELFs transduced with the LacZ gene, whereas overexpression of full length IFI16 reduced virus replication. Altogether, these results demonstrate that reducing IFI16 activity with ΔDIFI16 consistently increases the rate of HCMV replication, whereas overexpression of the full length protein down-regulates its replication, implying IFI16 as a novel restriction factor in the replicative cycle of Herpesviruses.
However, these above results are in apparent contrast with those of Cristea et al. [23] and those previously published by Rolle et al. [24] with the IFI16 mouse homolog Ifi204, who found that knocking down IFI16 or Ifi204 expression caused a delay in the accumulation of infectious CMV progeny, but no net reduction in its replication. To explain these discrepancies, we performed kinetic experiments in which IFI16-knockdown HELFs (i.e. IFI16-silenced using siRNA) were infected with HCMV at two different MOI (0.05 and 1, respectively) and the viral yield evaluated at different time points post infection (hpi). At 96 hpi, the number of infectious particles measured in the IFI16-knockdown cells (siRNA IFI16) infected at an MOI of 0.05 was actually decreased compared to that of HELFs electroporated with control siRNA (siRNA ctrl) or left not electroporated (NE) (Figure 1C, left panel). At 120 hpi, the number of plaque forming units (PFU) in the absence of IFI16 expression started to be higher than that observed in the control HELF cultures electroporated with siRNA ctrl or sham-electroporated. In contrast, at 144 hpi, the accumulation of HCMV progeny exhibited an ∼3- fold increased yield in the absence of IFI16 expression. At the higher MOI, no differences in virus yield were observed irrespective of the level of IFI16 expression (Figure 1C, right panel). Thus, in line with the results of Cristea et al. [23] and Rolle et al. [24], our findings indicate that Herpesvirus replication, and in particular that of HCMV, may be impaired by IFI16 silencing in the first hours after infection at the lower MOI, but as virus replication progresses this impairment becomes lost as shown by increased viral yields compared to controls at the later time points p.i. At higher MOI, however, the relevance of IFI16 in the control of HCMV replication appears to be less.
HCMV is able to replicate in vivo and in vitro in many different host cells including vascular endothelial cells, epithelial cells, connective tissue cells, hepatocytes, and various leukocyte populations (reviewed in [25]). Since fibroblasts are not the most important target cells in vivo, we evaluated the effect of IFI16 silencing on HCMV replication in human umbilical vein endothelial cells (HUVEC), a cell system considered to be more pertinent for HCMV replication in vivo. For this purpose, HUVEC were electroporated with siRNA IFI16 or siRNA ctrl or left not electroporated (NE) and then infected with the endotheliotropic HCMV strain VR1814 at MOI of 0.1 or 1. As shown in Figure S2, and in line with the results obtained using fibroblasts, replication of the VR1814 strain was significantly increased at an MOI of 1 compared to that observed in cells treated with siRNA ctrl (∼7 fold increase). In contrast to our previously observations in fibroblasts, at the lower MOI the level of virus replication in endothelial cells in the absence of IFI16 was still different but at lower degree from that observed in controls. This difference may be explained by the fact that the efficiency of infection and replication of the VR1814 strain in endothelial cells is much lower than that of the AD169 virus strain in fibroblasts [26]. Taken together, these results demonstrate that IFI16 may also restrict HCMV replication in physiologically relevant target cells in vivo such as endothelial cells.
To gain more insight into the antiviral activity of IFI16, we focused on the HCMV model. HELFs were infected with AdV IFI16 or AdV LacZ at an MOI of 200 and the IFI16 protein content monitored by Western blot analysis. As shown in Figure 3A, IFI16 protein levels started to increase as early as 24 hpi and continued to increase until 48 hpi. When HELFs overexpressing wt IFI16 for 24 h were infected with HCMV at an MOI of 0.1, a ∼2.5 log decrease in viral production was observed on day 4, compared to HELFs left untransduced or transduced with the control AdV LacZ gene (Figure 3B, left panel). A similar pattern was observed at the higher MOI of 1, but with lower levels of virus growth suppression (∼1 log reduction) (Figure 3B, right panel). Thus, consistent with the previous findings, which showed increased virus replication in the absence of IFI16, the overexpression of IFI16 strongly inhibits HCMV replication.
To investigate the molecular basis of the antiviral activity of IFI16, we examined the effects of its overexpression on different phases of the HCMV replication cycle. To this purpose, HELFs were infected with AdV IFI16 or AdV LacZ (MOI of 200) or left uninfected (mock), and infected 24 h later with HCMV (MOI of 1) for a further 24 h. The amounts of IE (IE1 and IE2), UL44, UL54, and UL83 transcripts were then assessed by quantitative real-time PCR as markers of IE (IE1 and IE2), E (UL44 and UL54), and L (UL83) mRNAs (Figure 4A). According to Cristea et al [23] results, we saw no difference in expression of the products of immediate early genes (IE1 and IE2) between IFI16-overexpressing and LacZ- or mock-infected cells. In contrast, mRNA synthesis of early (UL44 and UL54) and early-late (UL83) genes was significantly reduced (2-, 6- and 6- fold respectively) in the cells expressing IFI16. Total protein extracts were then analyzed for their content of immediate early (IE), early (UL44), and early-late (UL83) proteins by immunoblotting with specific antibodies. As shown in Figure 4B, the expression of the HCMV UL44 and UL83 proteins was strongly impaired in AdV IFI16-infected HELFs compared to that seen in control cells, mirroring the results obtained at the mRNA level. The expression of these proteins is indispensable for productive HCMV infection (reviewed in [1]). A possible explanation for these findings is that IFI16 may exert its antiviral effect by inhibiting the synthesis or function of an HCMV-encoded component critical for viral DNA synthesis and/or maturation. Further support to this hypothesis comes from the experiments in which viral DNA synthesis was measured in HELFs transduced with IFI16 or LacZ genes for 24 h and then infected with HCMV. As shown in Figure 4C, starting at 48 hpi and continuing at later time points (72 and 96 hpi), a significant decrease in the number of viral DNA copies was observed in HELFs transduced with the IFI16 gene compared to cells left untransduced or transduced with the LacZ gene.
Taken together, these results indicate that IFI16 reduces HCMV replication by inhibiting the expression of E and L genes required for the viral DNA synthesis and completion of the viral productive cycle.
Previous studies have shown that cotransfection of IFI16 and a CAT reporter gene containing the wild type UL54 promoter results in a dose-dependent decrease in reporter activity [27] suggesting an interplay between IFI16 and transcription factors responsible for UL54 promoter activation. These observations were made, however, in uninfected cells using the UL54 essential promoter as a target of cellular transcription factors.
To investigate whether IFI16 overexpression may affect UL54 gene promoter activity in HCMV-infected cells, reporter plasmids containing UL54 promoter segments containing progressive 5′ deletions were transfected into HELFs left uninfected or subsequently infected with AdV IFI16 or AdV LacZ as control. Twenty-four hours later, the cells were infected with HCMV to transactivate the UL54 promoter and luciferase activity assessed following an additional 24 h. As shown in Figure 5A, HCMV infection significantly increased luciferase activity of all the reporter constructs when examined in mock- (data not shown) or LacZ-overexpressing HELFs compared to HCMV-uninfected cells. In contrast, when IFI16 was overexpressed prior to HCMV infection, luciferase activity decreased by more than 75% or 90% in cells transfected with the pUL54 0.4 or pUL54 0.3 indicator plasmids respectively, compared to AdV LacZ control cells. Similarly reduced levels (70%) were also observed with the minimal promoter construct pUL54 0.15 that contains nucleotide sequences up to −150 relative to the transcription start site, including the DR-ATF and IR-1 responsive elements. Together, these results indicate that inhibition of HCMV replication by IFI16 is consequent to DNA polymerase activity downregulation.
The HCMV DNA polymerase promoter contains a DNA element located between −54 and −43 relative to the transcription start site that has been shown to be required for both basal transcriptional activity and transactivation by IE2 [28], [29]. Mutations of the 8-bp inverted repeat element 1 (IR-1) diminishes transactivation by IE2 and abrogates the binding of cellular transcription factors, such as Sp1 [30], [31]. To investigate the involvement of IR-1 in IFI16-mediated UL54 suppression, transient transfection assays were performed using luciferase reporter constructs driven by versions of the minimal UL54 promoter (pUL54 0.15) mutated in either the DR-ATF or IR-1 element. Mutations in the DR-ATF element of the UL54 promoter (positions -82- 95) (pUL54 0.15 mut DR-ATF) slightly affected transactivation of the promoter by HCMV (141660 RLU vs. 88331 RLU) but did not impact on IFI16-mediated suppression (42930 RLU vs 34180 RLU), suggesting that the target of IFI16 might reside in the IR-1 element (Figure 5A). In accordance with the results previously reported by Luu et al. [30] and Wu et al. [31], the IR-1 mutation in the construct pUL54 0.15 (pUL54 0.15 IR-1 mut) caused more than a 95% decrease in HCMV-induced UL54 promoter activity compared to the wild type pUL54 0.15, consistent with its prominent role in the control of UL54 promoter transactivation by HCMV. IFI16 overexpression before HCMV infection did not further reduce the HCMV-driven transactivation of the IR-1 mutant construct in terms of luciferase activity, indicating that the DNA target of IFI16 suppressor activity might reside in the IR-1 element.
Suppression of UL54 promoter transactivation by IFI16 is apparently at variance with the findings of Cristea et al. [23] who showed that HCMV pUL83 stimulates activity of the major immediate-early promoter (MIEP) through its interaction with cellular IFI16 protein. Therefore, it would seem possible that positive or negative modulation of promoter activity by IFI16 might depend on the type of promoter itself and the specific transcription factors interacting. To investigate this hypothesis, the pUL54 0.15 construct and the pMIEP-Gl3 plasmid, a reporter plasmid in which the luciferase gene is driven by the HCMV major immediate-early promoter (MIEP), were transfected into HELFs that were then infected with either AdV IFI16 or AdV LacZ followed by HCMV infection. As expected, luciferase expression driven by the pUL54 0.15 promoter was inhibited (Figure 5B). In contrast, IFI16 overexpression significantly increased MIEP activity, confirming that modulation of transcription by IFI16 is highly specific and largely dependent on the type of target promoter (e.g. UL54 promoter vs. the MIEP).
To investigate whether the IFI16/Sp1 interaction affects the expression of other viral genes, the expression of UL44, a component of the HCMV DNA polymerase complex, was also investigated. UL44 protein expression was affected by IFI16 overexpression (Figure 4B), thus strengthening the evidence indicating an antiviral role of IFI16. Transfection experiments were performed using constructs containing the luciferase gene driven by the UL44 promoter containing progressive deletions in its DNA element responsive to different transcription factors [32]–[34]. Consistent with the results obtained with the UL54 promoter, IFI16 overexpression significantly reduced luciferase activity by more than 70% in cells transfected with the pUL44-600-3T plasmid (containing the UL44 promoter from −613 nt to +67 relative to the proximal transcription start site and all three TATA elements) or the pUL44-600-1T indicator plasmid (containing the UL44 promoter from −613 nt to −92 relative to the proximal transcription start site and only the distal TATA element), compared to AdV LacZ control cells. Interestingly, in the absence of the Sp1 responsive element upstream of the three transcription starting sites (from −613 nt to −164 relative to the proximal transcription start site), IFI16 stimulated the activity of the pUL44-160-3T indicator plasmid, as seen with the MIEP promoter. These results suggest that the IFI16/Sp1 interaction is important for modulating the activity of target viral genes.
Previous mutagenesis scanning and EMSA analyses of the UL54 −54/−43 sequence (IR-1 element) indicated that Sp1 is the cellular factor responsible for abetting the action of HCMV IE proteins at the UL54 promoter [30]. To investigate whether Sp1 could be the target of IFI16, causing the modulation of IR-1 activity in the context of HCMV infection, nuclear extracts from AdV IFI16- or AdV LacZ-transduced HELFs infected at an MOI of 200 for 24 h and then with HCMV at an MOI of 2 for 24 h were analyzed by EMSA using an IR-1 oligonucleotide probe. Consistent with the results of Luu and Flores [30], the oligonucleotide spanning the IR-1 element formed two major complexes (1 and 2) with nuclear extracts from cells infected with HCMV 24 h earlier (Figure 6A, lane 2) that were reduced in the mock-infected cells (lane 1). These complexes could be specifically competed by a 100-fold excess of unlabeled IR-1wt (lane 3), but not by the same concentration of a mutated oligonucleotide (lane 4). In HELFs overexpressing IFI16 (lane 5), subsequently infected with HCMV (lane 6), and incubated with the labeled IR-1wt oligonucleotide, the generation of both complex 1 and 2 was significantly reduced at 24 hpi, suggesting that IFI16 inhibits the binding of both cellular and viral-induced transcription factors to the IR-1 sequence. The suppression of complex 1 and 2 must be due to IFI16 since infection of HELFs with AdV LacZ (used as control) does not impair the induction of the two complexes by HCMV (lanes 7 and 8). To confirm that the complexes 1 and 2 impaired by IFI16 overexpression contained Sp1, antibodies specific for Sp1 protein or unrelated antibodies (ctrl) were added to EMSA reactions for supershift analysis. As shown in Figure 6B, addition of anti-Sp1 (lanes 3 and 4), but not unrelated antibodies (lanes 5 and 6), supershifted the protein complexes of both mock and HCMV-infected cells, indicating that Sp1 is a component in both IFI16-suppressed complexes 1 and 2. In line with the results reported by Luu and Flores [30], anti-IFI16 antibodies did not alter the mobility of the two complexes, demonstrating that IFI16 does not form part of either protein complex 1 or 2 (data not shown). Altogether, these results demonstrate that suppression of UL54 promoter activity by IFI16 is associated with inhibition of the formation of Sp1-containing complexes with the IR-1 element.
To identify the mechanisms underlying the suppression of Sp1-induced transcription of the UL54 promoter by IFI16, immunoprecipitation (IP) experiments and in vivo chromatin IP (ChIP) assays were performed. IP was carried out on infected and uninfected cell lysates using polyclonal Abs recognizing IFI16 or a control antibodies (ctrl). Immunoprecipitated proteins were examined by Western blotting using antibodies recognizing Sp1 (Figure 7A). A species of the same size as Sp1 that reacted with the polyclonal anti-Sp1 antibody could be observed in the nuclear protein extracts (Figure 7A, lanes 1, 2, 3) or and in the proteins immunoprecipitated from the infected lysates using the anti-IFI16 Abs (Figure 7A, lanes 8, 9). It is also worth noting that a band of greater intensity was observed in lysates derived from HCMV-infected cells overexpressing IFI16 (Fig. 7A, lane 9). In the uninfected cell lysates (mock), no migrating band of the same size as Sp1 could be detected (Figure 7A, lane 7). Similarly, no bands were detected when infected cell lysates were immunoprecipitated with control antibodies (ctrl) and reacted with anti-Sp1 antibodies (Figure 7A, lanes 4–6). Finally, the presence of IFI16 protein was also confirmed in the very same nuclear cell extracts (Fig. 7A, bottom panel). Thus, these results indicate that in IFI16-overexpressing and HCMV infected HELFs, IFI16 and Sp1 physically interact. It has been demonstrated elsewhere that IFI16 binds to DNA [35]. DNA-binding proteins can associate during IP due to their adjacent binding on DNA rather that due to protein-protein interactions [36]. To determine whether nucleic acid is required for the IFI16/Sp1 association, IP was performed using an anti-IFI16 Ab or a control Ab (ctrl) in the presence or absence of benzonase (a non specific nuclease), and the IPs were probed by Western blot using anti-Sp1 Ab. No Sp1 was detected in IPs using the control Ab (Figure 7B, lanes 2 and 3). In contrast, Sp1 was detected in IPs from infected lysates in both the absence (lane 4) and presence (lane 5) of benzonase. To confirm the action of benzonase on nucleic acid, the cell lysates analyzed in Figure 7B were examined on an ethidium bromide-stained agarose gel. In the absence of benzonase, a robust staining of nucleic acid could be seen, whereas staining in the presence of benzonase was no more intense than that of a no-sample control (data not shown). Altogether, these results demonstrate that suppression of UL54 transcription by IFI16 is accompanied by the displacement of Sp1 from its promoter due to its direct association with IFI16.
To corroborate these results further, the binding of Sp1 to the UL54 promoter in the presence of IFI16 was then analyzed in vivo by ChIP assay. Formaldehyde cross-linked sonicated chromatin fragments from HELFs infected with AdV IFI16 or AdV LacZ 24 h earlier and then infected with HCMV for 24 h were immunoprecipitated using an anti-Sp1 polyclonal antibody. The DNA released from the immunocomplexes was then analyzed by quantitative real-time PCR to detect the enrichment of the IR-1 sequence of the UL54 promoter in the immunoprecipitates. The rate of amplification was verified using cross-linked non-immunoprecipitated chromatin (input). Consistent with the EMSA and IP results, real-time PCR analysis of the purified ChIPed DNA showed that the Sp1 antibody pulled approximately 80% more Sp1-bound UL54 promoter DNA down in extracts from AdV LacZ-infected HELFs compared with those from AdV IFI16-infected HELFs (Figure 7C). Unrelated, affinity-purified polyclonal antibodies, used as negative control, did not immunoprecipitate the complex containing the UL54 promoter (data not shown). Altogether, these findings indicate that the suppression of UL54 gene transcription in the presence of IFI16 is due to the inhibition of Sp1 recruitment to its promoter.
To learn more about the basis of the IFI16/Sp1 interaction and its effects on HCMV replication, cell lines stably expressing the V5-tagged mutant forms of IFI16, namely ΔDIFI16, ΔBIFI16, or the full-length IFI16 (wtIFI16), were infected with HCMV at an MOI of 2 PFU/cell. Twenty-four hours later, IP was carried out using monoclonal Ab recognizing the V5 tag, or control antibody (ctrl). Immunoprecipitated proteins were examined by Western blotting using antibodies recognizing Sp1. As shown in Figure 7D, a species of the same size as Sp1 could be observed in the total nuclear extracts (Figure 7D, top panel, lane 1–3). A band migrating at a similar molecular weight could also be observed in the proteins immunoprecipitated from HELFs overexpressing the full length IFI16 form (wtIFI16) (Figure 7D, lane 9). In contrast, no bands corresponding to Sp1 could be observed in the immunoprecipitates from the ΔBIFI16 cell line when blotted with anti-Sp1 antibodies (Figure 7D, lane 8). A detectable, although weaker band corresponding to Sp1 could be observed in the immunoprecipitates from the ΔDIFI16 cell line (Fig. 7D, lane 7). As expected, protein extracts immunoprecipitated from infected cell lines with control antibodies (ctrl) did not display any migrating band (Figure 7D, lane 4–6). Finally, the presence of V5-tagged proteins was also confirmed in the very same nuclear cell extracts (Fig. 7D, bottom panel). Altogether, these results demonstrate that the interaction of IFI16 with Sp1 most likely depends on the integrity of the HIN domains present on the full length protein. In support of this hypothesis, the ΔDIFI16 mutant containing the two HIN domains partially maintained its ability to bind Sp1, although at levels much lower than those of the wtIFI16. By contrast, the ΔBIFI16 mutant lacking the HIN-B domain completely lost its ability to interact with Sp1.
IFI16 was recently identified to be an intracellular sensor of HSV-1 DNA, which stimulates the expression of IFN-β and pro-inflammatory genes through activation of IRF3 and NF-κB transcription factors [8]. To clarify whether IFI16 could suppress HCMV replication through IFN-β induction, HELFs were electroporated with siRNA specific for IFN-β or siRNA ctrl or left not electroporated (NE), infected with AdV IFI16 at an MOI of 200, and 24 h later with HCMV at an MOI of 0.1. At different time points post infection, HCMV yield was measured. As shown in Figure 8A, this treatment led to the inhibition of IFN-β induction in response to HCMV infection. By contrast, suppression of IFN-β production did not prevent the inhibition of HCMV replication by IFI16 overexpression, demonstrating that IFI16 does not require functional IFN-β and that it directly inhibits HCMV replication (Figure 8B). IFN-β production was wiped out by siRNA specific for IFN-β, as proven by the finding that an inducer of IFN-β (poly I:C) failed to block HCMV replication upon HELF treatment with siRNA IFN-β, while it was blocked in cells treated with siRNA ctrl (data not shown).
This study demonstrates that IFI16 acts as restriction factor for HCMV replication. Inhibition of HCMV growth was observed by adopting two different experimental approaches. In the first approach, primary HELFs were generated whereby the IFI16 protein underwent a long-lasting knockdown through the use of either siRNAs or the overexpression of dominant negative forms of the IFI16 protein that lacked the death domain (PYD) at the N-terminus or the HIN-B domain at the C-terminus. HELFs were chosen for two main reasons: i) they are fully permissive to HCMV; and ii) they are characterized by a very low basal level of IFI16 expression that can be adequately up-regulated following transduction with the AdV IFI16 vector. When HCMV replication was analyzed in these cells, in either the absence of IFI16 or in the presence of an inactivated form of IFI16, viral yield at low MOI was significantly increased. The finding that IFI16 counteracts HCMV replication is further supported by the outcome of the experiments using HELFs overexpressing IFI16. In these cell cultures, the HCMV yield at low MOI was severely impaired, confirming that IFI16 is indeed endowed with antiviral activity. Viral gene expression analysis on both the mRNA and protein level showed that IFI16 did not affect IE expression, but rather a viral replication step down-stream of IE expression. In line with this finding, and by measuring the DNA viral load, we demonstrated that IFI16 down-regulates viral DNA synthesis by affecting the bona fide activity of the UL54 DNA polymerase gene and the UL44 gene. Definitive support for this came from transfection experiments which showed IFI16 overexpression to significantly impair the activity of both UL54 and UL44 gene promoters, responsible for viral DNA synthesis. Previous studies have indeed shown that IFI16 can be purified by DNA affinity chromatography using a region of the UL54 [30]. Moreover, cotransfection of IFI16 and a CAT reporter gene containing the wild type UL54 promoter results in a dose-dependent decrease in reporter activity [27], suggesting an interplay between IFI16 and transcription factors responsible for UL54 promoter activation. All these observations were made, however, in uninfected cells using the UL54 essential promoter as a simple target of cellular transcription factors and the mechanisms responsible for the inhibition of UL54 transactivation by IFI16 in infected cells remained to be elucidated. Therefore, to provide new evidence about the action of IFI16 in the context of HCMV infection, promoter scan analyses, EMSA and ChIP, were each performed in IFI16-overexpressing HELFs infected with HCMV. The results obtained demonstrate that in order to exert its antiviral activity IFI16 binds and displaces the Sp1 transcription factor interacting with the responsive IR-1 element present in the UL54 promoter. Sp1 detachment from its DNA cognate element leads to a decrease in HCMV DNA synthesis and, as a consequence, the inhibition of virus replication. Consistent with this, we have previously demonstrated that activation of the NF-κB response is mediated by an IFI16-induced blockade of Sp1-like factor recruitment to the promoter of the IκBα gene, which encodes the main NF-κB inhibitor [37].
In order to provide new insights into the mechanisms of the interaction of IFI16 with Sp1 and its effects on HCMV replication, we produced cell lines overexpressing mutated forms of IFI16 lacking the HIN-B domain (ΔBIFI16) or the PYD domain (ΔDIFI16). Virus yield experiments demonstrated that HCMV replication in the ΔDIFI16 cell line was enhanced compared to that in the cell lines expressing the wild type IFI16 or IFI16 lacking the HIN-B domain (ΔBIFI16). A conceivable explanation for these results could be the following. The mutated form of IFI16 lacking the HIN-B domain (ΔBIFI16) is unable to physically interact with Sp1 and therefore can no longer relieve the suppressive activity of endogenous IFI16 on viral promoters, such as UL54 and UL44. In contrast, in cells lacking the PYD domain (ΔDIFI16), IFI16 to some extent maintains its capability to interact with Sp1 and compete with the endogenous form on the viral promoter. As a consequence, the suppressive activity of the endogenous IFI16 protein is retained, stimulating HCMV replication. The finding that the HIN-B domain is responsible for the Sp1 interaction is in line with results recently reported by Liao et al. [38], which show that the HIN-B and HIN-A domains are together responsible for the IFI16/p53 interaction. These results corroborate the notion that IFI16 is a modular protein and that its different functions correspond to its different domains.
A different role of IFI16 in HCMV replication has been demonstrated by Cristea et al. [23], who identified the interaction of pUL83 (pp65) with IFI16 throughout the course of HCMV infection and showed that pUL83 recruits IFI16 to the major immediate-early promoter (MIEP) and stimulates, rather than inhibiting, MIEP activity. Consistent with these observations, when we compared UL54 promoter activity with MIEP activity in IFI16 overexpressing cells, stimulation of only the latter promoter was observed. The differential sensitivity of the two promoters to IFI16 activity may be explained by the presence of four NF-κB responsive elements on the MIEP. Functional analysis of the ICAM-1 promoter by deletion- or site-specific mutagenesis has indeed demonstrated that NF-κB is the main mediator of IFI16-driven gene induction [37]. NF-κB activation appears to be triggered by IFI16 through a novel mechanism involving suppression of IκBα mRNA and protein expression. Furthermore, to study the activity of IFI16 in p53-mediated gene expression, Fujiuchi et al. [39] examined BAX promoter (a p53 target gene) activation by coexpressing p53 and IFI16. When the proteins were coexpressed, promoter activity was enhanced up to 17-fold. Consistent with the results showing the collaboration of p53 and IFI16 in transcription, endogenous levels of BAX, p21WAF1, and HDM2 were synergistically induced by expressing both proteins, as shown by Western blot analysis. Taken together, these results demonstrate that depending on the factors and the type of promoter it is interacting with, IFI16 may act either as a positive or negative transcription regulatory factor. Moreover, we have previously observed that murine CMV (MCMV) replication was delayed in mouse embryo fibroblasts (MEF) following inactivation of the IFI16 mouse homolog Ifi204 with a p204-dominant-negative mutant. These results suggested that the activity of this protein was required for efficient MCMV replication [24]. These discrepancies may be thus explained by the findings that IFI16 requirement varies during HCMV replication and depends on the MOI employed.
IFI16 has been shown to act as an innate immune sensor of intracellular dsDNA [35]. Upon sensing dsDNA, the IFI16 protein triggers the induction of IFN-β. IFI16 directly associated with IFN-β-inducing viral DNA motifs recruits STING a critical mediator of the IFN-β response to DNA. It is therefore possible that IFI16 inhibits HCMV replication through the induction of IFN-β. This is quite unlikely, however, in the light of the following observations. First of all, the outcome of the transfection experiments and EMSA indicates that IFI16 directly interacts with and down-regulates Sp1, which is responsible for UL54 promoter activation. Secondly, knockdown of the IFN-β gene by specific siRNAs does not impair the ability of IFI16 to down-regulate HCMV replication. Finally, the addition of anti-IFN-type I antibodies does not impair the capability of IFI16 to suppress HCMV replication (unpublished). Thus, IFI16 appears to directly inhibit virus HCMV replication rather than down-regulating viral growth through activation of an IFN pathway.
Overall, although the detailed mechanisms of IFI16-mediated repression of viral E and L gene expression remain to be fully determined, the results presented in this study congruently demonstrate that the actions of IFI16 contribute to a cell's intrinsic repression mechanism of HCMV gene expression. It remains to be determined, however, how the virus counteracts IFI16 activity and shifts the balance toward viral evasion and its consequent growth.
Low-passage human embryonic lung fibroblasts (HELFs), human embryo kidney 293 cells (HEK 293) (Microbix Biosystems Inc.), African green monkey kidney cells (Vero) and mouse connective tissue fibroblasts (L929) were grown in Eagle's minimal essential medium (Gibco-BRL) supplemented with 10% fetal bovine serum (FBS; Gibco-BRL). Human umbilical vein endothelial cells (HUVECs) were isolated by trypsin treatment of umbilical cord veins cultured in Endothelial Growth Medium (EGM) corresponding to Endothelial Basal Medium (EBM) (Clonetics, San Diego, CA) containing 2% FCS, human recombinant vascular endothelial growth factor (hrVEGF), basic fibroblast growth factor (bFGF), human epidermal growth factor (hEGF), insulin growth factor (IGF-1), hydrocortisone, ascorbic acid, heparin, gentamycin and amphotericin B (1 mg/ml each). Experiments were carried out with cells at passages 4–8. HCMV strain AD169 (ATCC-VR538) and a clinical isolate of Adenovirus were propagated on HELF cells, clinical isolates of HSV-1 and HSV-2 on Vero cells and Vesicular Stomatitis Virus (VSV) serotype Indiana on L929 cells and titrated by standard plaque assay, as previously described [40]. HCMV VR1814 is a derivative of a clinical isolate and grows efficiently on HUVECs [41].
HELF cells were transiently transfected with a MicroPorator (Digital Bio) according to the manufacturer's instructions (1200 V, 30 ms pulse width, one impulse) with a pool of IFI16 small interfering RNAs (siRNA IFI16), siRNA IFN-β, or control siRNA (siRNA ctrl) as negative control (final concentration: 300 nM; Qiagen). The IFI16 and IFN-β siRNA sequences are reported in Table S1. IFI16 or IFN-β siRNA-induced blockade was checked by immunoblotting with rabbit anti-IFI16 antibodies or by real-time PCR with IFN-β specific primers (Table S1) respectively, at the time points indicated.
Lentiviral vectors carrying the full-length IFI16 ORF (wt IFI16) or IFI16 ORF lacking the PYD domain (ΔPYDIFI16) or the HIN-B domain (ΔHIN-BIFI16) or the LacZ gene as a control, were generated as described by Azzimonti et al. [42]. To obtain lentiviral lines, HELFs were transduced with the recombinant Lentivirus and successfully transduced cells selected using blasticidin (4 µg/ml; for a maximum of 10 days). Transduction efficiency was assessed by immunoblotting for the V5-epitope. The adenovirus transfer vector pAC-CMV IFI16 was constructed as described in Gugliesi et al. [43].
The HCMV UL54 (pol) promoter sequences (positions −425 to +15 relative to the UL54 transcription start site, GenBank NC_006273) were amplified by PCR using purified HCMV AD169 DNA as the template and the primer sets reported in Table S1. The 5′- and 3′-primers were engineered with HindIII and KpnI restriction sites. The PCR fragments were subsequently digested and directionally cloned into the corresponding sites of the pGL3-basic vector (Promega) to obtain the pUL54 0.4 construct. The pUL54 0.3 and pUL54 0.15 constructs were derived from the pUL54 0.4 construct and contain UL54 promoter sequences −285 to +15 and −150 to +15, respectively. These constructs were generated by PCR using the UL54 appropriate primers (Table S1). The fragments were then ligated into the HindIII and KpnI sites of the pGL3-basic vector. To obtain UL54-0.15 promoter sequence with inactivated IR-1 and DR-ATF binding sites, the sequences of these sites were modified by site-directed mutagenesis (Quick Change XL Site-Direct Mutagenesis Kit, Stratagene). The IR-1 (−54 to −42) and DR-ATF (−97 to −79) recognition sites of pUL54 0.15 were changed into unique restriction sites (−41 to −47, XbaI, and −89 to −95, EcoRI, respectively) using the IR-1 mutant and DR-ATF mutant oligonucleotides (Table S1) and their complementary oligonucleotides. The HCMV IE promoter-enhancer sequence (position −666 to +19 relative to the IE1/IE2 transcription start site, GenBank K03104.1) was amplified out of the purified HCMV AD169 genome by PCR using the primer sets shown in Table S1. The 5′- and 3′-primers were engineered using NheI and HindIII restriction sites (underlined). The PCR fragments were subsequently digested and directionally cloned into the corresponding sites of the pGL3-basic vector (Promega) to obtain the pMIEP construct. The correctness of all the amplified viral sequences was confirmed by sequencing. The HCMV UL44 promoter sequences (positions −613 to +67 relative to the proximal UL44 transcription start site, GenBank NC_006273) were amplified by PCR using purified HCMV AD169 DNA as the template and the primer sets reported in Table S1. The 5′- and 3′-primers were engineered with HindIII and BglII restriction sites. The PCR fragments were subsequently digested and directionally cloned into the corresponding sites of the pGL3-basic vector (Promega) to obtain the pUL44-600-3T construct. The pUL44-660-1T and pUL44-160-3T constructs were derived from the pUL44-600-3T construct and contain the UL44 promoter sequences −613 to −92 and −164 to +67 relative to the proximal UL44 transcription start site, respectively. These constructs were generated by PCR using the UL44 appropriate primers (Table S1). The fragments were then ligated into the HindIII and BglII sites of the pGL3-basic vector. The correctness of all the amplified viral sequences was confirmed by sequencing.
Cells were electroporated using a Micro-Porator MP-100 (Digital BioTechnology), according to the manufacturer's instructions (a single 1400 V pulse, 20 ms pulse width). After 24 h, cells were infected with AdV IFI16 or the control indicator plasmid AdV LacZ (MOI of 200 PFU/ml) and 24 h later the cells were infected with HCMV AD169 (MOI of 0.5). Following a further 24 h, luciferase activity was measured using the Dual Luciferase Reporter Assay System Kit (Promega) on a Lumino luminometer (Stratec Biomedical Systems, Birkenfeld, Germany), as previously described by Baggetta et al. [44].
Whole-cell protein extracts were prepared and subjected to immunoblot analysis as described in Gugliesi et al.[43]. The following antibodies were used: rabbit polyclonal anti- C-terminal IFI16 antibodies (diluted 1∶1000) or mouse monoclonal antibodies (MAb) anti- IEA (IE1 plus IE2, 11-003; Argene, diluted 1∶250), UL44 (P1202-2, Virusys, clone CH16, diluted 1∶500), UL83 (CA003-100, Virusys, clone 3A12, diluted 1∶1000), V5 (R960-25, Invitrogen, diluted 1∶5000); MAb against β-actin (MAB1501R; Chemicon, Temecula, CA, diluted 1∶2000) were used as a control for protein loading. Immunocomplexes were detected with sheep anti-mouse or donkey anti-rabbit immunoglobulin antibodies conjugated to horseradish peroxidase (Amersham) and visualized by enhanced chemiluminescence (Super Signal; Pierce).
HELFs were infected with AdV IFI16 or control plasmid AdV LacZ (MOI of 200 PFU/cell) for 24 h and subsequently infected with HCMV strain AD169 (MOI of 2 PFU/cell) for 24 h. Nuclear extracts were collected using the Nuclear Extract Kit (Active Motif) according to the manufacturer's instructions. Electrophoretic mobility shift assays were carried out as previously described [37]. Briefly, nuclear extracts (15 µg of protein) were incubated in a binding buffer (10 mM Tris-HCl [pH 7.5], 50 mM NaCl, 0.5 mM EDTA, 0.5 mM dithiothreitol, 1 mM MgCl2, 5% glycerol) containing 2 µg of poly (dI-dC) (GE Healthcare) and the 32P oligonucleotide probe representing the wild-type (wt) or mutated (mut) HCMV UL54 promoter IR-1 motif (− 65 to −35 respect to the transcription start site). Sequences are reported in Table S1. For supershift experiments, 2 µg of polyclonal antibody recognizing Sp1 (Millipore) or 2 µg of rabbit polyclonal anti-human C-terminal IFI16 antibodies were used. Unlabeled 30-bp annealed oligonucleotide was added as the competitor DNA in 100-fold molar excess above the level of the probe.
For immunoprecipitation experiments, nuclear cell proteins were obtained as described for EMSA analysis. 30 µg of proteins were incubated with antibodies of interest (2 µg) for 1 h at room temperature with rotation. The immune complexes were collected using protein G–Sepharose beads (Sigma-Aldrich) for an additional 1 h at room temperature with rotation. The Sepharose beads were pelleted and washed three times with RIPA buffer. Finally, the proteins were eluted using Laemmli sample buffer and resolved on 8% SDS-PAGE gel to assess the protein binding by Western blotting. Where indicated, 400 U benzonase (Novagen) was added for 30 minutes at 4°C after clarification of the lysate by centrifugation, as described in Strang et al. [45].
The ChIP assay was performed as previously described [37]. Briefly, HELFs were infected with AdV IFI16 at an MOI of 200 for 24 h and then with HCMV at an MOI of 2 for 24h. DNA-protein complexes were cross-linked with PBS containing 1% formaldehyde for 10 min, and the reaction was stopped via the addition of glycine to a final concentration of 125 mM for 5 min at room temperature. Nuclear extracts were collected by using the nuclear Nuclear Extract Kit (Active Motif) according to the manufacturer's instructions. Nuclear extract were sonicated to shear chromatin to a final size of 500–2000 bp and the supernatant recovered and used directly for immunoprecipitation experiments by incubation with appropriate antibodies (2 µg) for 1 h at room temperature. The immune complexes were collected as described above with protein G–Sepharose beads (SIGMA). After immunoprecipitation, beads were collected and sequentially washed as described in Caposio et al. [37]. DNA-protein cross-links were reversed by incubation in 1% SDS/TE buffer at 65°C overnight. The samples were digested with proteinase K and DNA was extracted by phenol/chloroform/isoamyl alcohol and then incubated for 30 min at 37°C TE/RNase A buffer. The input lysates were processed as above. DNA was analyzed by quantitative real-time Sybr green PCR using primers for the IR-1 sequence; the primer sequences used are: forward: 5′- GGTCCTTTGCGACCAGAAT- 3′; reverse: 5′- TATACTCGACAGCGGCGTCT- 3′. The amount of the DNA precipitated by the antibody was normalized to the total input DNA that was not subjected to immunoprecipitation. The value of 1 was assigned to the normalized level of IR-1 immunoprecipitated with unrelated antibody.
Real-time quantitative reverse transcription-PCR (RT-PCR) analysis was performed on an Mx 3000 P apparatus (Stratagene). Total RNA was extracted with the NucleoSpin RNA kit (Macherey-Nagel) and 1 µg was retrotranscribed using the Revert-Aid H-Minus FirstStrand cDNA Synthesis Kit (Fermentas). Reverse-transcribed cDNAs were amplified in duplicate using Brilliant Sybr green QPCR master mix (Fermentas), as described in Luganini et al. [40] for viral genes or for cellular cytokines [44].
To determine the number of viral DNA genomes per nanogram of cellular reference DNA (18S rRNA gene), viral DNA levels were measured by quantitative real-time PCR, as described in Luganini et al. [40], using the previously reported probe and primers to amplify a segment of the IE1 gene [46]. HCMV DNA copy numbers were normalized by dividing by the amount of human 18S rRNA gene (Assay-on- Demand, 18S, assay no. HS99999901_s1; Applied Biosystems) amplified per reaction mixture. A standard curve of serially diluted genomic DNA mixed with an IE1-encoding plasmid (from 107 to 1 copy) was created in parallel with each analysis [47].
Caspase 3–7 protease activity was assessed by measuring the extent of cleavage of a fluorometric peptide substrate using the SensoLyte AFC Caspase Sampler Kit “Fluorimetric” (Anaspec). Doxorubicin treatment (0.5 µM for 18 hours) was used for the positive control. Experiments were performed according to the manufacturer's instructions. Fluorescence was measured at an excitation wavelength of 405 nm and an emission wavelength of 500 nm using the VICTOR3 1420 multilabel counter (Perkin–Elmer). Protease activity was expressed as fold induction relative to the basal level measured in each uninfected cell line.
All statistical tests were performed using GraphPad Prism version 5.00 for Windows (GraphPad Software, San Diego California USA, www.graphpad.com). The data were presented as the means ± standard deviations (SD). Means between two groups were compared by using a two-tailed t-test.
Means between three groups were compared by using a one-way or two-way analysis of variance with Bonferroni's post-test. Differences were considered statistically significant at p<0.05.
|
10.1371/journal.pntd.0000550 | Role of the Endogenous Antioxidant System in the Protection of Schistosoma mansoni Primary Sporocysts against Exogenous Oxidative Stress | Antioxidants produced by the parasite Schistosoma mansoni are believed to be involved in the maintenance of cellular redox balance, thus contributing to larval survival in their intermediate snail host, Biomphalaria glabrata. Here, we focused on specific antioxidant enzymes, including glutathione-S-transferases 26 and 28 (GST26 and 28), glutathione peroxidase (GPx), peroxiredoxin 1 and 2 (Prx1 and 2) and Cu/Zn superoxide dismutase (SOD), known to be involved in cellular redox reactions, in an attempt to evaluate their endogenous antioxidant function in the early-developing primary sporocyst stage of S. mansoni. Previously we demonstrated a specific and consistent RNA interference (RNAi)-mediated knockdown of GST26 and 28, Prx1 and 2, and GPx transcripts, and an unexpected elevation of SOD transcripts in sporocysts treated with gene-specific double-stranded (ds)RNA. In the present followup study, in vitro transforming sporocysts were exposed to dsRNAs for GST26 and 28, combined Prx1/2, GPx, SOD or green-fluorescent protein (GFP, control) for 7 days in culture, followed by assessment of the effects of specific dsRNA treatments on protein levels using semi-quantitative Western blot analysis (GST26, Prx1/2 only), and larval susceptibility to exogenous oxidative stress in in vitro killing assays. Significant decreases (80% and 50%) in immunoreactive GST26 and Prx1/2, respectively, were observed in sporocysts treated with specific dsRNA, compared to control larvae treated with GFP dsRNA. Sporocysts cultured with dsRNAs for GST26, GST28, Prx1/2 and GPx, but not SOD dsRNA, were significantly increased in their susceptibility to H2O2 oxidative stress (60–80% mortalities at 48 hr) compared to GFP dsRNA controls (∼18% mortality). H2O2-mediated killing was abrogated by bovine catalase, further supporting a protective role for endogenous sporocyst antioxidants. Finally, in vitro killing of S. mansoni sporocysts by hemocytes of susceptible NMRI B. glabrata snails was increased in larvae treated with Prx1/2, GST26 and GST28 dsRNA, compared to those treated with GFP or SOD dsRNAs. Results of these experiments strongly support the hypothesis that endogenous expression and regulation of larval antioxidant enzymes serve a direct role in protection against external oxidative stress, including immune-mediated cytotoxic reactions. Moreover, these findings illustrate the efficacy of a RNAi-type approach in investigating gene function in larval schistosomes.
| Species of the human blood fluke Schistosoma are estimated to infect approximately 200 million people worldwide, resulting in loss of health, vitality and productivity mainly among the world's poorest inhabitants. Since snail intermediate hosts represent an essential part of the flukes' life cycle, an understanding of the strategies used by the intramolluscan schistosome larvae to survive within this host may provide novel approaches for disrupting larval development and thus transmission to humans. Anti-oxidant enzymes produced by the parasite Schistosoma mansoni are believed to play a critical role in the maintenance of cellular redox balance, contributing to larval survival in their snail host, Biomphalaria glabrata. In this study, we have incorporated a RNA interference approach attempting to knock down specific anti-oxidant enzymes, including gluthatione-S-transferases 26 and 28 (GST26 and 28), gluthatione peroxidase (GPx), peroxiredoxins 1 and 2 (Prx1/2) and superoxide dismutase (SOD), and to evaluate their endogenous anti-oxidant function in the sporocyst stage of S. mansoni. Results clearly demonstrated a significantly higher susceptibility of antioxidant double-stranded (ds)RNA-treated larvae to in vitro H2O2 treatment or hemocytic encapsulation compared to GFP dsRNA controls. Taken together, our findings support the hypothesis that endogenous expression and regulation of larval antioxidant enzymes serve a direct role in protection against external oxidative stress, including immune-mediated cytotoxic reactions.
| Miracidial penetration and entry into the molluscan intermediate host represent a critical transition period in which the previously free-living larval stage is now confronted with a potentially hostile environment as it attempts to establish a viable infection [1],[2]. Miracidia of the human blood fluke Schistosoma mansoni shed their ciliated epidermal plates soon after entry into the host snail Biomphalaria spp., transforming to primary or mother sporocysts. It is during this time of transition and early sporocyst development that larvae are especially vulnerable to oxidative stress generated from products of oxidized plasma hemoglobin [3], or reactive oxygen or nitrogen species (ROS and RNS, respectively) resulting from hemocyte-mediated immune responses [4]–[7]. In such a potentially damaging environment, it is vital that parasites possess the capability of maintaining a redox equilibrium in order to counteract the effects of ROS/RNS generated both internally (products of endogenous metabolic oxidative reactions) and externally (environmental insults) [1],[8].
Recent studies have shown that S. mansoni larvae possess numerous enzymes involved in ROS metabolism and detoxification of oxidative products [9]–[14], and, like their adult stage counterparts [15]–[18], appear to complement each other to maintain the redox balance in the parasite. Included among these enzymes are the following: (i) glutathione-S-transferases 26 and 28 (GST26 and GST28) that function to neutralize potential membrane damage by the linked catalysis of glutathione (GSH) reduction with detoxification reactions involving thiol-conjugation to xenobiotics [19], (ii) peroxiredoxin (Prx1 and Prx2) that are involved in maintaining redox balance, by reducing hydrogen peroxide (H2O2) using a thioredoxin as an electron donor [20], (iii) superoxide dismutases (SOD), metalloenzymes responsible for catalyzing the dismutation of the superoxide radical to hydrogen peroxide as a defense mechanism against oxygen toxicity [21], and (iv) glutathione peroxidase (GPx), an H2O2-metabolizing enzyme that protects membranes from damage by phospholipid peroxidation [20],[22]. It is noteworthy that unlike most organisms, catalase, an enzyme responsible for H2O2 metabolism, is absent in S. mansoni [18],[23],[24], but is functionally replaced by Prx and GPx [16]. Interestingly, for schistosome GPx, whose H2O2-reactivity is typically very low in adult worms [8], exposure to the mammalian host environment induces enzyme activity and appears to be positively correlated to the parasite's resistance to oxidative stress [22]. In contrast to GPx, high levels of Prx activity are found in adult S. mansoni worms, and these enzymes are believed to be key components in maintaining redox balance, as well as are major contributors to antioxidant activity [16].
Previous findings have demonstrated that in vitro cultured S. mansoni sporocysts are highly sensitive to H2O2 toxicity [5], and that sublethal exposure of sporocysts in vitro to ROS, in particular H2O2, elicits an upregulation of genes encoding various antioxidant proteins [7],[11]. These data support the hypothesis that the primary sporocyst is capable of interfering with, or deactivating ROS-mediated damage, through activity of an endogenous antioxidant system [1]. However, to date, a functional role of specific antioxidant enzymes within intact larvae in providing protection against external ROS insults has not been demonstrated. Recently Mourão et al. [25] demonstrated consistent transcript knockdown for various antioxidant/redox-active detoxicant mRNA species in S. mansoni sporocysts using RNA interference as originally described [26]. These included transcripts for GST26 and 28, Prx1 and 2, and GPx. As a followup to these findings, the present study was conducted to determine the functional consequences of these induced antioxidant gene changes, especially their relevance to S. mansoni sporocyst interactions with the intermediate snail host B. glabrata.
Research procedures involving mice used in the course of this study were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) at the University of Wisconsin-Madison under assurance no. A3368-01.
The NMRI strain of S. mansoni was used for all experiments. S. mansoni eggs were isolated from livers obtained from mice harboring 7-week old infections, and miracidia hatched in an artificial “pond water” supplemented with antibiotics (50 µg/mL streptomycin and 60 µg/mL penicillin) [27]. Larvae were washed twice in ice-cold, sterile pond water by centrifugation, before being resuspended in Chernin's Balanced Saline Solution (CBSS; [28]), containing glucose and trehalose (1 g/L each) streptomycin and penicillin (50 µg/mL and 60 µg/mL, respectively). Miracidia were then counted and distributed into 48- or 96-well polystyrene tissue culture plates (Costar, Corning Incorporated, NY), at concentrations of ∼500, 1000 or 8000 miracidia/well for oxidative stress experiments, immunocytochemistry or Western blot analyses, respectively. Finally, double-stranded RNAs were synthesized from isolated sporocyst cDNA using T7 RiboMAX Express RNAi Kit (Promega, Madison, WI), according to manufacturer protocol. Briefly, dsRNAs synthesis reactions were allowed to incubate for 16 hr at 37°C prior to DNAse treatment. DsRNA products were then extracted by phenol/chloroform and purified by precipitation with isopropanol. DsRNAs (50 nM final concentration) for specific antioxidant genes or green-fluorescent protein (GFP; specificity control dsRNA) were added to cultures containing 100 µL of CBSS for the oxidative stress assays and immunocytochemistry and 400 µL for the Western blot experiments. Because of the sequence and functional similarities of Prx1 and 2, dsRNAs for these transcripts were combined as a single treatment, designated hereafter as Prx1/2. Larvae were incubated for 7 days as previously detailed [25], after which time the functional consequences of dsRNA treatments were determined in functional assays described below. It should be noted that in a previous series of RNAi experiments conducted in parallel with the present study [25], a consistent, significant knockdown of steady-state transcript levels for each of the antioxidant genes currently under study was well documented. The only exception was the Cu/Zn superoxide dismutase (SOD) gene, in which larval exposure to SOD dsRNA resulted in a consistent increase, not knockdown, of SOD transcripts.
To assess the effects of antioxidant dsRNA on the expression of specific proteins in sporocysts, we analyzed protein extracts of dsRNA-exposed sporocysts by Western immunoblot analysis [29] incorporating specific antibodies to two antioxidant species; namely SmGST26 (Cell Signaling Technology, Danvers, MA) and SmPrx1/2 (gift from Dr. D. Williams). Briefly, protein samples (∼8 µg) and Precision Plus Dual Color Marker (Bio-Rad, Bio-Rad Laboratories, Inc., Hercules, CA) were separated on 12.5% SDS-PAGE gels and transferred by semi-dry electroblotting (Amersham Biosciences) to nitrocellulose membranes (Bio-Rad). After blocking overnight in TBS (2.42 g Tris base, 8 g NaCl, pH 7.6) containing 5% bovine serum albumin (BSA), membranes were incubated in specific antibodies or a mouse anti-α tubulin antibody (serving as loading control, 1∶1000 dilution; Upstate Biotechnology Inc., Lake Placid, NY) for 16 hr at 4°C with gentle rocking. Membranes were then washed for 30 min in TBS-Tween (0.1%), and incubated for 1 hr in TBS-BSA (5%) containing either alkaline phosphatase (AP)-conjugated goat anti-rabbit IgG or AP-rabbit anti-mouse IgG at dilutions of 1∶104 and 1∶5000, respectively (Promega, Madison, WI). The colorimetric immunoreactivity was detected with the chromogen 5-bromo-4-chloro-3-indolyl phosphate (BCIP) and nitro-blue tetrazolium (NBT) diluted in AP buffer (0.1 M Tris, 0.1 M NaCl, 0.05 M MgCl2, pH 9.5).
To quantify the observed immunoreactivity of each target protein in sporocysts treated with specific dsRNA and control GFP dsRNA, the intensities of reactive target bands were measured using Ultraviolet Transilluminator BioImaging Systems (UVP, Inc., Upland, CA) and normalized to the αtubulin band with LabWorks Image Acquisition and Analysis Software (version 4.6) in order to quantitatively evaluate the effects of antioxidant dsRNA treatment on specific protein levels. Three independent experimental replicates were performed and analyzed by Student's t-test, with significance set at P≤0.05.
In order to compare in situ GST26 and Prx protein levels in antioxidant dsRNA-treated parasites, we prepared whole, intact sporocysts for immunofluorescent observations. All washing steps, in eppendorf tubes, were performed by centrifugation at 1600 rpm for 2 min and repeated 5 times, or as otherwise mentioned. Following transformation and in vitro cultivation (24 hr), sporocysts were washed 3 times in CBSS, to remove detached ciliated plates, prior transfer to siliconized-tubes containing 2% paraformaldehyde and 1% Triton-X100/sPBS. Larvae were fixed overnight at 4°C under gentle agitation, then washed in snail phosphate-buffered saline (sPBS; [30]) and resuspended in blocking buffer (5% normal goat serum + 0.02% sodium azide in sPBS) for 16 hr at 4°C. Rabbit-anti-GST26 or mouse anti-Prx1/2 primary antibodies, diluted at 1∶2000, and 1∶200, respectively, were then added to the larvae in fresh blocking buffer for 16 hr at 4°C under gentle agitation. This was followed by 5 washes, 10 min each, in sPBS, and resuspension in blocking buffer containing 4 µg/mL AlexaFluor 488-conjugated anti-rabbit/mouse antibody, 7 units/mL phalloidin-Alexa 546 and 10 µg/mL Hoechst 33258 dye (Invitrogen). Tubes containing samples were incubated for 16 hr at 4°C under constant rotation, followed by washing in sPBS, resuspension in 40 µl of sPBS and mounting on coverslips. A Nikon Eclipse TE2000 (Nikon Instrument Inc., Melville, NY) inverted epifluorescence microscope equipped with a Bio-Rad Radiance 2100 MP Rainbow Confocal/Multiphoton Imaging System (W.M. Keck Laboratory for Biological Imaging, Instrumentation, UW-Medical School) was used for specimen imaging and evaluation.
Previous work in our lab has established a consistent and specific pattern of altered antioxidant transcript expression in primary sporocysts after 7 days of double-stranded (ds) RNA exposure [25]. Specifically, statistically significant knockdown of S. mansoni GST26, GST28, GPx, and Prx1/Prx2 transcript levels, and an unexpected robust increase in those of SOD were observed in dsRNA-treated larval populations. To further explore the functional relevance of these enzymes in this parasite model, we conducted experiments to determine how antioxidant dsRNA exposure affected gene expression at the protein level (for selected enzymes), and whether a functional association could be established between antioxidant gene knockdown and parasite survival in presence of stressors such as reactive oxygen species (H2O2) or encapsulating hemocytes.
To verify that specific dsRNA treatments had a predicted downregulating effect on sporocysts at the protein levels, Western blot analyses were performed on sporocysts treated with dsRNA for GST26, Prx1/2 and GFP (control) using antibodies specifically against S. mansoni GST26 and Prx1/2 [20]. In all experiments a crossreactive anti-α tubulin antibody served as a loading control. As shown in Figure 1, proteins extracted from GFP dsRNA-treated sporocysts (specificity control) presented two distinctive bands at ∼26 and 55 kDa, corresponding to GST26 and α tubulin, respectively. However, although larvae treated with GST26 dsRNA also exhibited the 55 kDa α tubulin protein, little immunoreactivity was observed at 26 kDa, suggesting an RNAi-induced GST26 protein knockdown (Fig. 1A). Quantification of band intensities by scanning densitometry, using anti-α tubulin reactivity to normalize protein loads in both treatment samples, confirmed that GST26 protein levels were significantly reduced (by ∼80%) in GST26 dsRNA-treated sporocysts compared to the nonspecific GFP dsRNA control group (Fig. 1C). Similarly, although not as dramatic, larval exposure to Prx1/2 dsRNA also exhibited a significant ∼50% decrease in protein level compared to the GFP control treatment by semi-quantitative Western blot analysis (Figs. 1B and D).
Consistent with Western blot analyses, in situ confocal observations of anti-GST26 localization in GST26 dsRNA-exposed and control GFP dsRNA-treated sporocysts revealed contrasting expressions of immunoreactivities. Anti-GST26 antibodies strongly reacted with endogeneous S. mansoni GST26 in sporocyst controls (Fig. 2A), but was much reduced in those treated with GST26 dsRNA (Fig. 2B), indicating a RNAi-mediated GST26 protein knockdown. Immunolocalization of anti-Prx1/2, however, revealed little difference in observed staining intensities between the GFP and Prx dsRNA-treated groups (Figs. 2C and 2D, respectively), except for a slight decrease in surface immunoreactivity in Prx-treated sporocysts. This also is consistent with the smaller knockdown effect of dsRNA exposure on Prx protein expression seen in immunoblot analysis (Fig. 1B).
In order to evaluate the effects of a potential loss of antioxidant activity in sporocysts due to dsRNA-induced antioxidant knockdown, we exposed groups of treated parasites to a range of hydrogen peroxide (H2O2) concentrations. In these preliminary tests 50 µM H2O2 was determined to represent a sublethal dosage under our experimental conditions (% larval death was not significantly different from control groups), whereas mortality rates significantly increased at 100 µM and higher H2O2 concentrations (data not shown). As shown in Figure 3, none of the dsRNA-treated sporocysts exhibited significant increases in H2O2-mediated mortality when compared to the GFP control treatments after 4 hr of exposure. However, at 24 and 48 hr sporocysts in all dsRNA-treatments, except the SOD dsRNA-exposed group, displayed significant increases in mortality with an average of 35% sporocyst death compared to 8% in control treatments after 24 hr, and 60 to 80% mortalities, compared to ∼18% in control treatments, at 48 hr post treatments (FdsRNA = 28.21, P≤0.0001; FTime = 84.71, P≤0.0001, N = 4). In contrast to other antioxidant treatments, sporocysts exposed to SOD dsRNA exhibited a H2O2-mediated mortality rate similar to that of control treatments at all time points (Fig. 3). See Figure 3 legend for means comparisons using Bonferroni's post-test.
To confirm that sporocyst death was specifically due to H2O2 as an exogeneous oxidative stressor, we exposed dsRNA-treated sporocysts to 50 µM H2O2 in presence or absence of bovine catalase or to catalase only (no H2O2 control), and evaluated sporocysts mortality in all treatments after 48 hr. Overall ANOVA indicated a significant effect of dsRNA treatment and H2O2-exposure (FdsRNA = 7.44, P≤0.001; FOxid = 15.33, P≤0.0001, N = 6). Within each treatment group, the percent mortalities for sporocysts exposed to GPx, GST26, GST28 and Prx1/2 dsRNAs were very similar when incubated in H2O2+catalase or catalase only (t values ranging from 0.23–1.74; all nonsignificant) (Fig. 4). These results are in contrast to the effects of exposure to H2O2 alone (positive killing control), in which mortality rates for sporocysts treated with the same antioxidant dsRNAs were significantly higher (ranging from 50–75%) when compared to 25% average sporocyst death in the catalase treatment groups (see Fig. 4 for Bonferroni's post-test comparisons). As previously observed, SOD dsRNA-treated larvae, again showed no difference in mortality rates between the different treatments, nor when compared to the control GFP dsRNA group.
Finally, in order to evaluate the effect of dsRNA antioxidant knockdown on snail hemocyte-sporocyst interactions in vitro, dsRNA-treated sporocysts were co-cultured with isolated hemocytes from the susceptible NMRI strain of Biomphalaria glabrata. After 24 hr of sporocyst-hemocytes incubation in an in vitro cell-mediated cytotoxicity assay [5], we observed that dsRNA knockdown of GST26 (t = 2.50, P≤0.01), GST28 (P≤0.0461) and Prx1/2 (t = 3.17, P≤0.04) resulted in small, but statistically significant increases in larval death, averaging ∼20% compared to ∼8% mortality in the GFP dsRNA control group (Fig. 5). Note that sporocysts treated with GPx dsRNA also showed an increase in mean mortality rate, but was not statistically significant when compared to the GFP control parasites. As observed in previous experiments, sporocysts treated with SOD dsRNA exhibited no difference in mortality compared to the GFP-treated control sample.
Enzymes involved in cellular redox pathways, which include proteins with antioxidant activities, are believed to be essential components regulating B. glabrata/S. mansoni molecular interaction [1],[2]. It is now well recognized that certain strains of B. glabrata snail immune cells or hemocytes produce substantial amounts of reactive oxygen [4],[5] and nitrogen [6] species as a consequence of stimulation by known activators of ROS/RNS or when encountering S. mansoni sporocysts, and that sporocysts are exquisitely sensitive to ROS-mediated killing, especially to H2O2. Moreover, in a series of followup studies, Bayne and co-workers have implicated a Cu/Zn-superoxide dismutase (SOD1) as a key enzyme involved in oxidative killing activity by hemocytes of resistant (R) strains of B. glabrata snails. Their studies demonstrated that (1) SOD transcript expression and enzyme activity are higher in certain R vs. susceptible (S) snail hemocytes [32] and this correlates with greater H2O2 production in the R strain [33], (2) B. glabrata SOD1 is comprised of 3 alleles, of which one (B allele) is significantly associated with R snails [34], and (3) SOD1 B allelelic expression is higher in R hemocytes than those of the S strain [35]. Based on their findings it is suggested that snail strain differences in SOD hemocyte expression may be causally linked to the observed S and R strain phenotypes. Because SOD catalyzes the conversion of superoxide to cytotoxic H2O2 it is reasoned that upregulation of the SOD1 gene and its resultant heightening of SOD enzymatic activity in R hemocytes may represent a possible mechanism for the differential larval killing response by R vs. S snail hemocytes [2].
While snail hemocytes produce H2O2 as an anti-parasite effector molecule, evidence also strongly supports the presence of an active antioxidant system in early developing S. mansoni sporocysts [11]–[13]. Catalase gene homologues were not found in recent searches of the S. mansoni genomic and EST databases, and this is consistent with earlier findings [16],[23],[24] indicating that these parasites must possess alternative means for neutralizing H2O2 and other ROS. As clearly demonstrated in mammalian stages of S. mansoni, this is accomplished by a thiol-dependent redox system involving thioredoxin glutathione reductase (TGR) as the central enzyme driving redox reactions [36]. Similarly, early intramolluscan larval stages also express redox genes, including TGR, thioredoxin, Cu/Zn SOD, GPx, Prx and GST [7], [11]–[13],[37], and in the case of GPx [7] and Prx1 and 2 [11], sporocyst expression levels are dramatically increased in response to ROS exposure. In addition, Cu/Zn SOD, GST26 and 28 and Prx were recently identified in larval transformation proteins (LTP) released during in vitro transformation of miracidia to sporocysts, demonstrating not only the synthesis of these antioxidants by miracidia, but also their active release during larval infection [9],[14]. Implied in these findings is the notion that antioxidant LTPs may be playing a potential protective role during early parasite development. This prospect of larval-protective antioxidants was given further credence by Vermeire and Yoshino [11] who demonstrated that Prx1/2 in LTP can function as scavengers of exogenous H2O2 suggesting the potential importance of excreted antioxidants as a sporocyst defense mechanisms.
In this study, we provide the first evidence for a functional role of the endogenous antioxidants GPx, Prx and GSTs in the survival of S. mansoni sporocysts confronted with exogenous oxidative stress. By successfully knocking down antioxidant transcript/protein levels using an RNAi-type approach, we were able to characterize the impact of introduced molecular H2O2 and presumed ROS produced during hemocyte encapsulation reactions on survival of intact primary sporocysts of S. mansoni. In a previous companion study that was run in parallel with the current experiments [25] we showed that larval treatment with double-stranded RNA (dsRNA) for all of the antioxidants, except SOD, produced a consistent, significant and specific transcript knockdown in sporocysts. In the present study, consistent with the transcript knockdown seen earlier, we demonstrated a dsRNA-associated decrease in GST 26 and Prx1/2 protein levels using specific antibodies in a semi-quantitative Western blot assay. This protein knockdown effect was supported by immunocytochemistry (ICC) in the case of GST26, but not as clearly for Prx. Importantly, the dsRNA-mediated decrease in GST26 and Prx protein content correlated well with significant increases in sporocyst mortality at 24 and 48 hr post-H2O2 exposure compared to the dsRNA control groups, implying a functional role for endogenous GST26 and Prx in the protection of primary sporocysts against external oxidative stress. Although lack of specific antibodies to the other antioxidants precluded a complete analysis of the other RNAi targeted genes used in this study, we continued to see a consistent correlation between dsRNA-induced decrease in transcript levels [25] and sporocyst survival patterns for larvae treated with GST28 and GPx dsRNA that were similar to those treated with GST26 and Prx1/2 dsRNAs. Indeed, compared to the untreated and GFP dsRNA controls, exposure of antioxidant dsRNA-treated sporocysts to a sublethal concentration of H2O2 in vitro resulted in dramatic decreases in parasite survival in all treatment groups except SOD, supporting the notion that GST28 and GPx, similar to Prx and GST26, also are capable of enhancing sporocyst survival in an oxidative environment.
These new findings are consistent with the extensive and ongoing work on the redox mechanism in the adult stage of S. mansoni, in which an active thiol-dependent redox maintenance system revolves around a thioredoxin glutathione reductase (TGR; [36]), a single enzyme that combines the activities of two enzymes, thioredoxin reductase and glutathione reductase, present in mammals [17]. Schistosome TGR is responsible for maintaining the reduced and active states of both thioredoxin (TR) and glutathione (GSH), allowing them to activate several Prxs and GPx, which in turn are capable of reducing H2O2 and other hydroperoxides [8]. Furthermore, in a more recent study, Sayed and coworkers [16] showed that Prx activity was essential to S. mansoni adult worm survival in vitro, further supporting the importance of maintaining a steady supply of this, and other antioxidant enzymes by S. mansoni adults. It appears that, like adult worms, early intramolluscan stages also must rely on robust endogenous system of antioxidant production that allows the parasite to overcome oxidative stress from both internal and external sources.
In addition to the antioxidant protective role of S. mansoni sporocysts in the presence of exogeneously introduced oxidative stress, we observed a similar survival pattern in dsRNA antioxidant-treated sporocysts that have come in contact with hemocytes from the susceptible NMRI B. glabrata strain. Our rationale for incorporating susceptible snail hemocytes in these experiments was to test the hypothesis that reducing the antioxidant capacity of sporocysts would increase their vulnerability to sublethal levels of ROS normally produced by NMRI snail hemocytes in in vitro cell-mediated cytotoxicity (CMC) assays [5],[38]. In this in vitro biologically-relevant context, we demonstrated a significant protective role of Prx and GSTs in sporocysts during hemocyte interactions. Co-culture of plasma-free hemocytes from susceptible NMRI snails with Prx, GST26, and GST28 dsRNA-treated sporocysts induced an increase in sporocyst mortality (to ∼20%) within 24 h of initial contact, when compared to GFP dsRNA-treated control group (8%). GPx dsRNA-treated sporocysts also showed a comparable increase in hemocyte-mediated killing, but high variance in replicate values rendered the increase nonsignificant. Thus the protective role of GPx against hemocyte-mediated ROS attack still remains to be proven. Taken together, however, our overall results suggest that ROS production in susceptible snail hemocytes is capable of overpowering antioxidant-deficient parasites. Zelck and Janowsky [7] hypothesized that susceptible snails generate relatively small amount of ROS, which in turn may induce antioxidant production in schistosomes, effectively neutralizing snail-generated ROS. In this study, we have demonstrated that effectively reducing their antioxidant enzyme capacity, sporocyst survival, when confronted by a usually benign hemocyte challenge, is significantly reduced, thus supporting the critical importance of the endogenous antioxidant system in establishing viable larval infections within the susceptible snail host.
Finally, a major exception to our present finding of enhanced larval susceptibility to oxidative stress by redox proteins was signal peptide (SP) Cu/Zn SOD [39]. In this case Cu/Zn SOD dsRNA treatment consistently had no effect on parasite survival whether in the presence of sublethal H2O2 or encapsulating hemocytes. These differing effects of SOD dsRNA exposure may have been predicted as treated S. mansoni sporocysts consistently displayed extreme elevations, rather than knockdown in transcript levels [25], indicating a strong induction of SOD gene expression in these larval stages. At present, the signaling mechanisms involved in this response are not known although, as suggested by Zelck and Von Janowsky [7] and Vermeire and Yoshino [11], sporocysts may be sensing ROS levels (including H2O2) and responding by upregulating protective antioxidant proteins. It is speculated that larval treatement with SOD dsRNA may have caused an initial downregulation of SOD transcripts that then led to a compensatory triggering of SOD over-expression. However, as shown in other systems, small interfering dsRNA also can trigger activation of transcription [40] and, therefore, could also represent a likely mechanism [25]. Its unusual expression pattern not withstanding, results indicate that hyperexpression of the SOD gene in S. mansoni sporocysts appeared to have a “neutral” effect on dsRNA-treated larvae (i.e., an effect similar to control dsRNA treatment) (present study). This does not necessarily imply that SOD has no role to play in maintaining redox balance within sporocysts both internally or in response to exogenous ROS sources. However, the mechanisms by which this is accomplished are currently unknown and represent the subject of further followup investigations in our lab.
|
10.1371/journal.pcbi.1007208 | Where did you come from, where did you go: Refining metagenomic analysis tools for horizontal gene transfer characterisation | Horizontal gene transfer (HGT) has changed the way we regard evolution. Instead of waiting for the next generation to establish new traits, especially bacteria are able to take a shortcut via HGT that enables them to pass on genes from one individual to another, even across species boundaries. The tool Daisy offers the first HGT detection approach based on read mapping that provides complementary evidence compared to existing methods. However, Daisy relies on the acceptor and donor organism involved in the HGT being known. We introduce DaisyGPS, a mapping-based pipeline that is able to identify acceptor and donor reference candidates of an HGT event based on sequencing reads. Acceptor and donor identification is akin to species identification in metagenomic samples based on sequencing reads, a problem addressed by metagenomic profiling tools. However, acceptor and donor references have certain properties such that these methods cannot be directly applied. DaisyGPS uses MicrobeGPS, a metagenomic profiling tool tailored towards estimating the genomic distance between organisms in the sample and the reference database. We enhance the underlying scoring system of MicrobeGPS to account for the sequence patterns in terms of mapping coverage of an acceptor and donor involved in an HGT event, and report a ranked list of reference candidates. These candidates can then be further evaluated by tools like Daisy to establish HGT regions. We successfully validated our approach on both simulated and real data, and show its benefits in an investigation of an outbreak involving Methicillin-resistant Staphylococcus aureus data.
| Evolution is traditionally viewed as a process where changes are only vertically inherited from parent to offspring across generations. Many principles such as phylogenetic trees and even the “tree of life” are based on that doctrine. The concept of horizontal gene transfer changed the way we regard evolution completely. Horizontal gene transfer is the movement of genetic information between distantly related organisms of the same generation. Genome sequencing not only provided further evidence complementing experimental evidence but also shed light onto the frequency and prominence of this concept. Especially the rapid spread of antimicrobial resistance genes is a prominent example for the impact that horizontal gene transfer can have for public health. Next generation sequencing brought means for quick and relatively cheap analysis of even complex metagenomic samples where horizontal gene transfer is bound to happen frequently. Methods to directly detect and characterise horizontal gene transfer from such sequencing data, however, are still lacking. We here provide a method to identify organisms potentially involved in horizontal gene transfer events to be used in downstream analysis that enables a characterisation of a horizontal gene transfer event in terms of impact and prevalence.
| For a long time, evolution in terms of gene transfer was thought to happen only along the tree of life, i.e. from parent to offspring generation. The discovery of horizontal gene transfer (HGT) [1–4] has revolutionised this dogma, and revealed the mechanism that enables bacteria to quickly adapt to environmental pressure [5–7]. Via HGT, bacteria can directly transfer one or multiple genes from one individual to another across species boundaries. The known and prominent mechanisms of HGT are transformation (uptake of nascent DNA from the environment), conjugation (direct transfer from cell to cell), and transduction (transfer via bacteriophages) [7]. In all cases, a piece of DNA sequence is—directly or indirectly—transferred from the so called donor organism to the acceptor organism and integrated into the genome (see also Fig 1).
Especially conjugation and transduction facilitate the transfer of pathogenicity islands and mobile genetic elements involving antimicrobial resistance (AMR) genes [8–10]. Today, we are facing the rise of so called “superbugs” [10, 11] as a result of bacterial adaptation and gain of resistance to antibiotic treatment, showing the need for methods to identify, characterise and trace HGT events.
The discrepancy between vertical, phylogenetic evolution and evidence for horizontal exchange and evolution across branches of a phylogenetic tree inspired existing genome-based HGT methods. For a fixed set of species and a potential horizontally transferred gene, these methods detect HGT events by looking at inconsistencies between the gene tree and a phylogenetic tree built for the set of species [12]. As a prerequisite, a candidate gene for which to run the calculation and comparison has to be known. Sequence content based methods aim to identify genes of foreign origin in a given genome by exploiting sequence pattern such as k-mer frequencies or GC content which vary between different species [13], [14]. All methods are based on an assembled genomes, meaning they are also prone to the problems of misassemblies. Although AMRs are a prominent example for horizontally transferred genes, methods to directly identify antimicrobial resistance (AMR) genes do not necessarily connect the presence of an AMR gene to an HGT event (e.g., KmerResistance [15]).
In previous work, we developed an approach that aims to call HGT events directly from next-generation sequencing (NGS) data [16] in a tool called Daisy. Instead of focusing on the sequence content or rather inconsistencies in the sequence content of the organism that acquired genes through HGT, Daisy examines the origin of the transfer, namely the prespecified acceptor and the donor organisms, and directly maps the NGS reads to these references. By facilitating structural variant detection methods, we can thereby identify the transferred region from the donor and the insertion site within the acceptor. A prerequisite for Daisy is therefore that both acceptor and donor references are known. This, however, is not always the case, and hence requires methods that are able to infer acceptor and donor reference candidates from the NGS reads of the organism assumed to be the result of an HGT event. Such methods are not yet available.
However, the problem of acceptor and donor identification directly from NGS data is akin to the problem tackled by metagenomic profiling studies that aim to unravel metagenomic samples. Here, so called metagenomic classification approaches aim at identifying all organisms present in a sample by directly analysing sequencing data with a complex mixture of various organisms [17]. While in this classical scenario all reads of a single organism in the sample can theoretically be assigned to one reference organism during identification, this is not the case for an organism that carries foreign genes acquired via HGT. Most reads will be assigned to the acceptor genome but only a fraction can map to the donor genome (see mapped reads in Fig 1). Hence, we have to account for this two mapping properties of the reads during analysis. Another requirement is the resolution of classification on strain level, if possible, since two strains of the same species can already significantly differ in their sequence content.
Metagenomic classification approaches follow either a taxonomy dependent or taxonomy independent approach [18, 19]. The general procedure for both approaches is to assign sequencing reads stemming from the same organism in the sample into the same group, a process also referred to as binning. Taxonomic dependent binning approaches assign the reads to specific taxonomic groups, and hereby infer the presence of these taxa in the sample. These methods either also make use of sequence composition patterns, e.g., Kraken [20], or they determine mapping-based sequence similarities for the read assignment, e.g., MEGAN [21], Clinical PathoScope [22] or DUDes [23]. Both approaches will most likely identify the acceptor reference of an HGT organism due to the homogeneous coverage and comparatively high number of reads. The drawback of all read assignment approaches is the limitation in the presence of mobile genetic elements, e.g., integrated via HGT or of hitherto unknown—or unsequenced—organisms in the sample. Reads belonging to these genes or unknown organisms are either assigned to a similar but incorrect taxa or not assigned at all, leading to wrong identifications and biases in abundance estimation. To ensure robustness, many approaches deliberately discard taxonomic candidates with only low and local coverage. Hence these approaches will likely discard any donor candidate references. Composition-based methods such as Kraken would also perform poorly pinpointing the correct donor based on evidence of only few reads given the fairly large number of usually detected species.
In our group, we developed MicrobeGPS [24], a metagenomics approach that accounts for sequences not yet present in the database. Instead of reporting fixed taxa with assigned reads, MicrobeGPS in turn uses the candidate taxa to describe the organisms in the sample in terms of a genomic distance measure. That is, it uses available references to model the composition of the organisms present in the sample in terms of coverage profiles and continuity, instead of directly assigning reference organisms to characterize the sample. If the organism in the sample is present in the database and covered homogeneously then the distance approximates to zero. If not, MicrobeGPS identifies the closest relatives by positioning the organism among references with the lowest genomic distance. Hence, the tool considers scores and metrics that reflect a donor-like, in-homogeneous coverage but filters out false positive candidates with inhomogeneous coverage for the purpose of species assignment. From the perspective of HGT detection, these may be highly relevant and should not be excluded.
Here we present DaisyGPS, a pipeline building on concepts of MicrobeGPS and tailored to the identification of acceptor and donor candidates from sequencing reads of an organism that may be involved in an HGT event. DaisyGPS uses genome distance metrics to define a score that allows the classification into acceptor and donor among the reported organisms. Owing to the properties of these scores, we still find the closest relatives of acceptor and donor in case these references are not present in the database. DaisyGPS further offers optional blacklists and a species filter to refine the search space for acceptor and donor candidates. DaisyGPS and Daisy are integrated into one pipeline called DaisySuite to offer a comprehensive HGT detection. We validate DaisySuite on a large-scale simulation where we show sensitivity and specificity of our approach and the robustness when applied to non-HGT samples. By simulating evolutionary distances, we demonstrate in another experiment that DaisySuite can detect HGTs in organisms that diverge from the original acceptor and donor. In addition, we used the simulated metagenomic data sets from the CAMI challenge [25] in combination with our simulated HGT reads to show that DaisySuite is able to detect HGTs in metagenomic samples. On a real data set from an Methicillin-resistant Staphylococcus aureus (MRSA) outbreak, we demonstrate the ability of the DaisySuite to distinguish between the outbreak associated and unassociated samples in terms of sequenced content potentially acquired through HGT events.
The problem of mapping-based HGT detection from NGS data is twofold: First, the acceptor reference (organism that receives genetic information) and donor reference (organism that the information is transferred from) that are involved in the HGT event have to be identified. In the following, we refer to the organism that derived from the acceptor and acquired genes from the donor in an HGT event as an HGT organism. Based on that, the precise HGT region from the donor and its insertion site within the acceptor can be characterised. We presented a method to solve the second task in [16]. Here, we propose the tool DaisyGPS (see also Fig 2) with the objective to identify possible acceptor and donor genome candidates given reads of a—pure or metagenomic—sample containing a potential HGT organism. We provide Daisy and DaisyGPS in an integrated pipeline that we call DaisySuite. DaisySuite is publically available at https://gitlab.com/rki_bioinformatics/DaisySuite, an extended documentation can be found at https://daisysuite.readthedocs.io/en/latest/index.html.
The genome of the HGT organism consists mainly of the acceptor genome (see Fig 1). When the reads of the HGT organism are mapped against the acceptor reference, most reads should map properly. Therefore a high and continuous mapping coverage pattern of the acceptor genome can be expected. In contrast to that, only a small part of the donor genome is present within the genome of the HGT organism, hence only a small fraction of the reads should map against the donor reference and then only within a zoned part (i.e. the part that has been transferred). This results in a discontinuous mapping coverage pattern where only a small part of the reference shows a high mapping coverage (see Fig 1).
In a first step, we need to define metrics that represent the expectations we have, i.e. how much of the genome is covered by reads (mapping coverage) and how uniformly these reads are distributed across the genome (discontinuous vs. continuous patterns). Given only the reads of the HGT organism, the acceptor and donor candidate identification problem is similar to aspects of metagenomic profiling. A standard problem in metagenomics is the identification of organisms in a sample using a read data set of this sample. At first glance, it may appear that the methods designed to solve this problem can also be applied to our identification objective, i.e. we have the read data set of the HGT organism and we are looking for two organisms (acceptor and donor) that are in the sample. However, because the HGT organism consists mainly of the acceptor genome, such an approach works only well for the identification of the acceptor. For the donor, additional information is needed to guarantee a reliable identification because references with only local or discontinuous coverage are usually dismissed by the profiler. We use the metagenomic profiling tool MicrobeGPS to obtain a coverage profile of our given HGT organism from mapping coverage metrics. MicrobeGPS fits our requirements as it can be configured to not filter any organisms and reports additional metrics that we use to represent acceptor and donor attributes. We evaluate the gathered metrics and establish a score that reflects our defined acceptor or donor coverage properties. The candidates are ranked by this score and a list of acceptor and donor candidates is generated. These acceptor and donor candidates can then be further analysed with tools such as Daisy.
For the purpose of HGT detection, we aim to define a scoring that reflects the mapping coverage properties of the acceptor and donor references: The acceptor has a continuous, homogeneous coverage over the complete length of the genome. The donor has a local, but still homogeneous coverage in the area where the transferred genes are originated but should have nearly no coverage at all otherwise. The score should further allow a clear distinction between acceptor and donor candidates and provide a meaningful ranking according to the likelihood of being the most suitable candidate.
As a basis for our scoring, we use the Genome Dataset Validity defined in [26] and homogeneity metric defined in [24]. The Genome Dataset Validity, or short validity, describes the fraction of the reference genome for which there is read evidence. In contrast, the homogeneity reflects how evenly the reads are distributed. Both have a range ∈ [0, 1]. The validity is defined such that a genome that is covered—either low or high—over the full length has a high validity (≈ 1). The validity can be interpreted as a measure of sequence similarity between the sequenced sample and a reference genome. Analogous to the homogeneity metric, we define a heterogeneity metric based on the Kolmogorov-Smirnov test statistic defined in [24] such that an evenly covered genome has a low heterogeneity (≈ 0) and a genome with local, high coverage a high heterogeneity (≈ 1). Note that the heterogeneity is a vertical translation of the homogeneity defined in [24], i.e. heterogeneity = 1 − homogeneity.
An acceptor is a genome with a continuous, high coverage that then has a high validity (≈ 1) and a low heterogeneity (≈ 0) score whereas a distantly related donor genome with only local, discontinuous coverage has a low validity (≈ 0) and a high heterogeneity (≈ 1) score.
As can be seen above, both validity and heterogeneity are complementary for acceptors and donors, and hence the relation of both metrics infers the property of a candidate between being an acceptor or a donor candidate.
We define:
s c o r e = v a l i d i t y - h e t e r o g e n e i t y with s c o r e ∈ [ −1 , 1 ] (1)
Acceptor candidates have a homogeneous coverage and hence high validity and low heterogeneity, i.e. validity > heterogeneity. Therefore, the value for a completely covered acceptor with uniform read distribution would approach +1. Likewise, the value for a donor that is only covered in a small region would approach −1. In addition to the coverage profile, there is a high evidence by sheer read numbers for acceptors:
a c c e p t o r - s c o r e = w * s c o r e with w = # m a p p e d r e a d s # t o t a l r e a d s (2)
where w is the fraction of all mapped reads that mapped to the specific acceptor candidate. For the donor, however, the size of the transferred region is not known in advance. Hence, we do not expect a specific read number evidence and therefore omit the weighting and define
d o n o r - s c o r e = s c o r e (3)
Both acceptor-score and donor-score are determined for every candidate and they have a codomain of [-1, 1]. Hence, we classify the candidates with acceptor-score ≥ 0 as acceptor and rank them from highest to lowest score. Donor candidates have a high heterogeneity and low validity, i.e. validity < heterogeneity. Therefore, we classify candidates with donor-score < 0 as donor candidates and rank them from lowest to highest score.
There is a special case if acceptor and donor are very similar. Here, the donor might not express the attributes we are looking for. In particular, the donor might have a significant read number evidence arising from acceptor reads also mapping to the donor. These shared reads lead to more regions of the donor genome being covered (higher validity) and to a less local, more homogeneous coverage pattern across the donor genome (lower heterogeneity), hence validity ≈ heterogeneity and donor-score ≈ 0. For such an event to occur, the true acceptor itself must be covered well (and evenly) enough to exhibit the hallmarks of an acceptor. Given that the donor is highly similar to the acceptor, a prime example being E. coli and Shigella, the validity of the donor strongly increases while the heterogeneity still takes the highly covered parts originating from the transferred region into account, allowing a positive donor-score. In contrast to this, a negative donor-score may easily occur due to spurious reads mapping to a reference genome without high similarity to the acceptor. Hence, we introduce a third classification and classify candidates with a donor-score > 0 as acceptor-like donors and rank them from lowest to highest.
A user definable number of the highest ranked candidates of each class (default: two acceptors, three donors and two acceptor-like donors) is then used to report all possible acceptor-donor candidate pairs, i.e. the cartesian product acceptors × (donors ∩ acceptor-like donors). For all these pairs, a follow-up Daisy run is triggered.
There are scenarios where it is necessary to exclude certain candidates from being reported. For example, in a reanalysis case, the assembled sequence from the sample reads might already been added to the reference set of your choice. For HGT detection from such reads, however, there is no information gain if DaisyGPS reports this entry as a suitable acceptor. Other examples include cases, where one can exclude certain species or taxa due to preanalysis information that nevertheless could be reported by DaisyGPS due to their high sequence similarity to the sampled organism or the presumed acceptor or donor candidates. To make the search for acceptor and donor candidates adaptable for such cases, DaisyGPS features the blacklisting of certain taxa. It is possible to exclude single taxa, a complete species taxon or a complete subtree below a specified taxon. For a default run, the filter is turned off.
DaisyGPS generally considers candidates on different taxonomic levels, e.g. species and strain level, and reports the candidate level with the best scores. Often the strain references contain additional sequences compared to the species level reference representative, and hence, the species reference will mostly have a homogeneous coverage that will then lead to a high acceptor score. Usually identification on species level is sufficient. There are however species such as, e.g., E.coli, where a high number of strains have been sequenced already and differ in their properties such as pathogenicity among the strains (e.g. E.coli K12 versus EHEC strain O157:H7). In these cases, a mere detection of the acceptor or donor on a species level might not be precise enough. For these situations, we implemented a species filter. If this filter is activated, only candidates below species level are reported. In case no candidate would be reported with an active species filter, the filter is disabled and the user informed that for further analysis also candidates on species level are used. For a default run, this filter is also turned off.
By default, DaisyGPS reports multiple acceptor candidates within the same species, given that they have equally high scores. If such a candidate organism is within an overrepresented group of the database, e.g., E. coli, they are often also overrepresented in the reported candidates due to the high similarity between strains of the same species. In this case, it can be beneficial to allow a broader view over the possible candidates by restricting the number of reported species representatives. Another use case can emerge when a priori knowledge about a donor exists and, optionally in combination with other filters, a more verbose overview of suitable species is prefered. For such occasions, we implemented a filter that allows to specify how many candidates per species are reported. We recommend to use this filter for metagenomic samples to reflect the high diversity of the sample among the acceptor and donor candidates.
Snakemake is a common workflow management system [27] which we used to implement the different steps of DaisyGPS. We generated the alignment file required for MicrobeGPS by mapping the reads of the HGT organism against the NCBI RefSeq (complete RefSeq, no plasmids, downloaded March 15th 2017) [28] using Yara [29, 30] in all-mapper mode, i.e. all suitable hits are reported for each hit. To ensure compatibility, we reimplemented the Daisy workflow in Snakemake as well, and integrated both into a combined suite (called DaisySuite, see also Fig 2). DaisyGPS yields a configurable number of acceptors, donors and acceptor-like donors (default: 2, 3, 2). For each possible pair of acceptor and donor, a Daisy call is inferred. Daisy then tries to identify HGT regions for each acceptor-donor pair and reports them as candidates if the regions pass the thresholds defined in [16] for mapping coverage, number of split-reads and number of read pairs between acceptor and donor. Both pipelines can still be run independently. To unburden installation, we provide a setup script and provide DaisySuite components as Conda [31] packages. The simulations are also integrated into the DaisySuite pipeline (see DaisySuite documentation for details).
The setup of the validation is according to the types of data sets. In a first phase, we want to show a proof of concept given data with sufficient ground truth. The aim is to predict the correct acceptor and donor candidates with DaisyGPS and at the same time to reproduce the results obtained from Daisy. We therefore use the data sets already shown in the Daisy paper for sake of consistency. We set DaisyGPS to report a total of two acceptor candidates, four donor candidates, and two acceptor-like donor candidates for every data set and we evaluate if the correct acceptor and donor candidates are among them. For incorrect candidates of acceptor and donor, Daisy should not report HGT candidates unless the transferred region is present in multiple strains or there are multiple possible acceptors present with high sequence similarities as, e.g., among E.coli strains. For the EHEC data set, we activate the species filter since we are interested in strain candidates, and further blacklist taxa from the HGT organism to be analysed (E.coli O157:H7, taxon 83334) and the complete O157 lineage (parent taxon 1045010). For the KOFL11 data set, the HGT organism is blacklisted as well (E.coli KOFL11, taxon 595495). In a second part, we want to estimate the rate of sensitivity and specificity of the DaisySuite. We designed a large-scale simulation analogous to the H.pylori data set with positive and negative simulations (100 simulations each). From the positive simulations, we calculate the sensitivity for both DaisyGPS and Daisy (see below for definitions on metrics). DaisyGPS is designed with high sensitivity in mind and always reports the closest fitting candidates given sequencing data, even for non-HGT organisms. Hence, also for the negative simulations, DaisyGPS will report candidates and we expect a low specificity here. Daisy, however, should then report only few—if any—HGT candidates from the acceptor-donor pairs. Furthermore, we want to inspect how much the HGT-organism can mutate before the true acceptor and donor cannot be detected anymore. We use the H. pylori data set and insert SNPs and small indels at varying rates. We repeat this procedure two times for ten different mutation rates, resulting in a total of 20 data sets. We then check for each sample if DaisyGPS is still able to detect true acceptor and donor and if so, whether Daisy is able to detect the true HGT region. In addition, we want to estimate the applicability for metagenomic samples by using three simulated metagenomic samples with varying complexity that include reads from the H. pylori data set. DaisySuite should still report the correct acceptor and donor candidates for the H. pylori data set. MicrobeGPS is a metagenomic profiling tool and will hence report all organisms in the sample alongside the true acceptor and donor candidates. Hence, we have to adjust our settings and procedure for this analysis: To report more distinct candidates for downstream analysis, we increase the number of reported acceptor and donor candidates to 30, respectively, but set the maximal number of candidates per species to one. We only perform a follow up Daisy analysis for the true acceptor and donor—if the pair is reported. For metagenomic samples, we would generally recommend this procedure of separated DaisyGPS and Daisy runs while adjusting and trying different filter settings for DaisyGPS, and then only run Daisy on the most likely candidates.
In the last evaluation part, we test the DaisySuite on real data with unknown or uncertain ground truth. The MRSA outbreak data set consists of 14 samples, seven outbreak related and seven unrelated. Here we want to test if DaisySuite is able to distinguish between the outbreak and non-outbreak samples according to their reported acceptor, donor and HGT region candidates.
The interpretation of various statistics depends on the hypothesis to be tested. In our analysis in the large-scale simulations, we differentiate between two scenarios: in the first one, we expect to detect an HGT event (positive test), while in the other one we assume the absence of an HGT event (negative test). For each simulation or run, a DaisyGPS call will lead to multiple pairs to be evaluated by Daisy. We therefore distinguish between statistics on runs and statistics on pairs that we will explain in the following.
For DaisyGPS, we consider during a positive test a single run as a true positive (TP) if the correct acceptor/donor pair is reported. Accordingly, a false negative (FN) occurs when the correct pair is not reported. Since the number of reported pairs is set by our settings, we will almost always have a fixed number of downstream verifications (except if there are not enough candidates to report) and thus we report the number of runs instead of pairs. Consequently, we can define the sensitivity as TP / #Runs. In a negative test setting, we deem those runs as true negatives (TNs) where either no pairs are reported or acceptor and donor of the pair are the very same organism. Note that if no other suitable candidates are available, the same organism may be reported as both acceptor and donor due to sorting by the respective scores, e.g. even an organism already reported as acceptor with a donor-score > 0 can be reported as donor if there is no candidate with a lower donor-score. All other pairs are regarded as FP that will each trigger an unnecessary verification in the downstream tools. Since we are interested in how many runs did not cause verifications, we can characterize the specificity by TN / #Runs. While it is obvious in both settings to rely on an exact match of the reported results and the ground truth, a reported organism still may be very close to the ground truth organism in terms of sequence similarity (negative and positive settings) and even include the very regions involved in the HGT event (positive setting). To account for this, we also use BLASTN in the case that no TP was reported and compare the FP to the ground truth. If the Blast identity of the FP to the ground truth is above 80% we change the classification from FP to BLAST-supported TP (Blast TP) since Daisy might still be able to infer the correct HGT region from these Blast TPs given the sufficient sequence similarity.
In Daisy, we evaluate acceptor/donor pairs and therefore the statistics are defined based on the condition of a pair reported by DaisyGPS. In a positive simulation, Daisy TP pairs are those that represent the correct pair and are detected by Daisy. It directly follows that each correct pair that is not supported by Daisy can be seen as a false negative (FN). Given that the pair is incorrect, i.e. a FP from DaisyGPS where the acceptor or donor is wrong, we count a rightly not supported pair as true negative (TN) and an erroneously detected pair as FP. To measure how many pairs are correctly identified, we define the sensitivity as (TP + TN) / #Pairs. Considering a negative test setting, we are mainly interested in the pairs that are wrongly reported as being involved in an HGT event. We declare those pairs as FP and describe the specificity as (#Pairs—FP) / #Pairs. It also follows that all the pairs that are not detected are TN. For a comprehensive summary of the classifications, refer to S1 Table.
Lastly, in the context of the complete DaisySuite pipeline, we evaluate the combined results of DaisyGPS and Daisy. Each pair reported by DaisyGPS for a single simulation induces an evaluation by Daisy. Since the overall result of the pipeline should indicate whether a simulation contains an HGT event or not, the classification of a DaisySuite run depends exclusively on the consolidated results of each Daisy evaluation for a single simulation. In a positive test setting, we want to find exactly the one pair that represents the HGT event. From that follows that a complete DaisySuite run can be classified as TP if Daisy supports solely the correct pair, i.e. Daisy reports the TP and no FP. This also implies that DaisyGPS needs to detect the TP. Similarly, in a negative test setting, a TN occurs if Daisy reports no HGT candidates at all.
DaisySuite is run with default parameters as of version 1.2.1 unless stated otherwise. The option to limit the maximum amount of candidates reported per species was introduced in version 1.3.0. The new version, however, did not introduce any changes to the used software versions, default parameters or other algorithmic aspects of DaisySuite. The parameter to combine potentially overlapping HGT candidates within Daisy is set to 20 bp, hence, overlapping regions with start and end positions differing by more than 20 bp are reported as separate candidates. For the comparison of the number and content of HGT sequences, we clustered overlapping HGT candidates with the tool usearch9 (v9.1.13_i86linux32) with identity 1.0 [37].
For validation, we determine the true presence of an HGT region in the samples by mapping the sample reads to all suggested, clustered regions with Bowtie2 (version 2.2.4). For comparison, we take the mean coverage of every region and apply a sigmoidal function to map all mean coverages to the [0.5,1] space for displaying a meaningful heatmap. The application of a sigmoidal function and the heatmap is computed in R (Rscript version 3.3.3). The heatmap function in R uses a hierarchical clustering with complete linkage as default, and we turned of the dendrogram for the columns. In addition, we perform a whole-genome alignment using the Mauve plugin (version 2.3.1) as part of the Geneious software (version 10.0.5) to establish shared HGT regions among the samples. To do this, we concatenate all HGT regions of a sample and separate the regions with segments of 1000*’N’ to avoid fragmented regions or overlapping local collinear blocks (LCBs).
In the first part of the validation, we test DaisyGPS on three data sets from simulated and real data with sufficient ground truth and already previously evaluated with Daisy. Since DaisySuite combines both tools, DaisyGPS and Daisy, the aim is to reproduce our previous results even without donor and acceptor being prespecified.
The H.pylori data set was simulated from E.coli K12 substr. DH10B as acceptor and H. pylori strain M1 as donor. DaisyGPS successfully reports both as such (see S2 and S3 Tables for complete candidate and HGT reports), and the subsequent Daisy run also reports the true HGT site. In addition to the only true HGT candidate previously already reported in the Daisy paper, DaisySuite reports another, FP HGT site for a region from Haemophilus ducreyi. The HGT region reported for H. ducreyi strain GHA9 has no continuous similarity with the HGT region from H.pylori (no blast hits longer than 15 bp, see S4 Table). However, the region on H. ducreyi shares the first 1200 bp and the last 1300 bp with the acceptor E.coli K12 substr. DH10B on multiple sites, and since beginning and end of the region are covered, almost six times as many split-reads are found as for the true acceptor site. The total coverage of the region is relatively low with 30x compared to 95x of the H.pylori but obviously high enough to pass the coverage filter.
The EHEC E.coli O157:H7 Sakai is supposedly derived by an HGT event where a defective prophage has been transferred from Shigella dysenteriae to E.coli O55:H7. Both are reported by DaisyGPS as candidates (see S5 Table). In line with its strong sequence similarity to the E.coli species, S.dysenteriae is labeled as an acceptor-like donor candidate. The proposed alternative HGT insertion site from our previous Daisy paper is still reported (see S6 Table).
The KO11FL data set comprises a transgenic E.coli W variant with transferred genes from Zymomonas mobilis and a plasmid that was not analysed here. DaisyGPS successfully reports E.coli W and Zymomonas mobilis as acceptor and donor candidates (see S7 Table). Daisy does not report any FP HGT candidates.
After validating DaisyGPS on data previously evaluated with Daisy as a proof of principle, we analyse DaisySuite in terms of robustness and sensitivity by performing a large-scale simulation. We perform the simulation for the H.pylori data set in a randomised and automated fashion generating 100 simulations with a transferred HGT region. To evaluate robustness, we also perform 100 negative simulations where an acceptor genome is simulated but no HGT region is inserted. With the positive simulations, we can estimate the sensitivity of the complete DaisySuite. For DaisyGPS, we evaluate how many from the 100 simulations have the correct acceptor and donor genome identified. Since DaisyGPS reports more than one potential acceptor-donor pair, we count a TP hit if the true pair is among them, and only count a FN if the true pair was not reported at all. In case the correct pair is not reported (acceptor or donor or both), we consider pairs with Blast sequence identity > 80% also as a potential HGT candidate pair, and also count them as a TP. To evaluate Daisy, we consider all pairs proposed by DaisyGPS.
For a true pair reported by DaisyGPS, Daisy can either report a TP HGT region or a FN if the region could not be identified. For an acceptor-donor pair wrongly proposed by DaisyGPS, Daisy can either report no HGT candidate region (TN) or a FP hit. When we summarise the DaisySuite results over all pairs of one simulation, we only count a TP for that simulation if Daisy did not report any FPs (despite any TPs or TNs).
Table 1 states the resulting counts for DaisyGPS and for the complete DaisySuite summarised over the 100 simulations. DaisyGPS yields a sensitivity of 79%. From the 79 TPs, 22 are based on either a wrong acceptor, or donor, or both but have still sufficient Blast similarity to the original acceptor or donor to be counted as TP according to our scoring. 69% of the TPs and FPs resulted in a TP or TN call from Daisy. It is noticeable that all DaisySuite FPs are Blast FPs.
Table 2 states the number of reported pairs proposed by DaisyGPS and a detailed count based on each pair for Daisy. From the resulting 818 pairs, Daisy then reports the correct HGT region, or correctly no HGT region from a DaisyGPS FP, with a sensitivity of 89%.
In addition to the positive simulations, we performed another 100 negative simulations where we randomly selected and variated an acceptor genome but did not insert any foreign region from a donor. DaisyGPS can now either produce a TN hit, i.e. report no candidates at all, or FP candidates. Since DaisyGPS is very sensitive by design, we expect it to generally report candidates and, hence, we want to estimate if these negative HGTs trigger reports by a Daisy follow-up call. As expected, the specificity for DaisyGPS is very low with 6% (see Table 3). However, Daisy reports only six FPs out of 743 pairs, i.e. three simulations produced a FP HGT report.
From these results we can infer that DaisySuite is able to distinguish HGT from non-HGT organisms and is very robust if no HGT is present.
To determine how robust our method is if the true acceptors and donors divert from the representative genome in the database, we performed a simulation over evolutionary distances by introducing increasing SNP and small indel rates into the H. pylori data set. We used the H. pylori data set to generate 20 simulations with varying mutation rates. We introduced both SNPs and indels starting with a rate of 0.01 and 0.001, respectively. We then incremented the rates by 0.01 (SNPs) and 0.001 (indels) for a total of 10 steps, yielding a maximum SNP rate of 0.1 and a maximum indel rate of 0.01. Each step was repeated twice to account for the randomness of mutations and read simulation.
Table 4 shows the results for the candidate detection by DaisyGPS. For this experiment, we used default settings, in particular, we report up to two acceptors and three donors. For up to 0.03 SNP rate and 0.003 indel rate, we can reliably determine the correct acceptor and donor as the top ranked candidates on strain level. Higher mutation rates obscur true acceptor by making other representatives of the Enterobacteriaceae family more similar to the HGT-organism, such that the true acceptor (on strain level) is not within the two highest ranking candidates anymore. For SNP rates 0.03-0.04 and indel rates 0.003-0.004, family representatives for Enterobacteriaceae are reported. For higher mutation rates, species representatives for E. coli are reported. For the ranks of the true acceptors and donors, please see S8 Table.
In general, the donor can be detected on strain level even for higher rates. For SNP rates ranging from 0.01 to 0.09, we detect the true donor at least once among the three best candidates within two repetitions. This may be attributed to the fact that only a small part of the HGT organism stems from the donor and hence is less heavily altered by randomly distributed mutation events. For a SNP rate of 0.1, solely representatives of the species E. coli are reported, hence the true donor is not detectable.
To further investigate whether the reported candidates lead to an HGT region detection, we continued to run Daisy. For all data sets for which the true positive acceptor and donor were reported at strain level, Daisy could identify the correct location of the HGT event. Other E. coli strains likewise passed the thresholds and subsequently were also reported, although the true site was always the—or among the—highest scoring locations. The number of reported HGT sites increases the higher the mutation rates grow, and starting at a mutation rate of 0.04 (SNP) and 0.004 (indel), it can also be observed that the number of reported locations increases tremendously, making a practical evaluation infeasible. This clearly shows the limitations of the mapping-based approach with regards to genetic divergence, especially in such a highly represented and highly similar species as E. coli.
To evaluate the applicability for metagenomic samples, we use three simulated metagenomic data sets with spiked in reads from the H.pylori data set. The metagenomic data sets are from the CAMI challenge and have a varying complexity in terms of the number of contained organisms, classified as low, medium, and high. To account for the metagenomic context, we set the number of reported acceptors and donors to 30, respectively, and only report one candidate per species. The true E.coli K12 acceptor is among the top 20 ranked candidates (low rank 7, medium rank 8, high rank 18, see S9–S11 Tables for full lists of reported candidates), so a maximal number of 20 acceptor candidates would have been sufficient for identification even for the high complexity sample. Donor identification is more challenging due to the less amount of reads that can be assigned. Still, the true H.pylori donor is among the top 30 ranked candidates (low rank 12, medium rank 7, high rank 24). A follow-up Daisy run on the true acceptor-donor pair successfully reports the correct HGT region for all three complexities.
MRSA strains are generally assumed to undergo HGT events frequently [38, 39]. The MRSA data set considered here consists of 14 samples with seven of them related to an MRSA outbreak (O1-O7) and seven MRSA samples not associated with the outbreak (N1-N7) but that occurred in the same time frame [36]. [36] analysed all 14 samples and compared them to the EMRSA-15 representative HO 5096 0412 as the supposedly closest relative of the outbreak strains. We first evaluate acceptor and donor candidates reported by DaisyGPS in relation to the proposed HO 5096 0412 reference and then investigate HGT region candidates reported by Daisy regarding a possible distinction of outbreak vs. non-outbreak samples. We activate the species filter as we are again interested in strain level candidates.
For all outbreak samples O1-O7, S.aureus HO 5096 0412 was reported as acceptor candidate by DaisyGPS (see Table 5 and S12–S39 Tables for individual results for each of the 14 MRSA data sets analysed). The same acceptor was also reported for non-outbreak samples N2, N6 and N7. Acceptor candidates for sample N1 are S.aureus ECT-R-2 and N315, for N3 and N4 S.aureus MSSA476 and MW2, and for N5 S.aureus MRSA252. Although not associated with the outbreak, samples N3 and N4 are from patients that shared the same room in the hospital where the outbreak occurred and hence are possibly related [36].
The reported donors are largely the same for both outbreak and non-outbreak samples (see Table 6). No donor was reported exclusively for the outbreak samples but three donors only for non-outbreak strains N1, N4 and N6. These are S.epidermidis strains ATCC 12228 and PM221 as well as Enterococcus faecium Aus0004. Although S.aureus HO 5096 0412 was reported for all outbreak samples, there is no clear distinction in acceptor and donor candidates reported by DaisyGPS apart from the non-outbreak only donors.
Table 5 states the total number of clustered HGT regions and the number of the clustered regions where HO 5096 0412 is the acceptor that are found by DaisySuite. Most HGT regions hence have the EMRSA-15 representative as acceptor.
Fig 3 shows a Mauve alignment of the concatenated HGT regions of all 14 samples. There is a clear connection between the HGT regions from the lower seven samples O1-O7 that are the outbreak related samples. Samples N1-N7 also share some regions but do not have a clear connection as among the outbreak related strains. The overlap between outbreak and non-outbreak HGT regions is also low.
Fig 4 shows the presence of the 41 HGT regions determined by mapping coverage called by Daisy among all samples. The purpose of the coverage analysis is to evaluate again if the HGT regions differ between the outbreak and non-outbreak strains but also to estimate if there are regions shared by all outbreak strains that are FN candidates of Daisy, or regions not covered at all that are likely FP candidates.
The clustering of samples according to the dendrogram shown in Fig 4 was done automatically (see settings part), and hence reflects the relation of the samples according to the mapping coverage of the proposed HGT regions.
All outbreak strains are clustered together and share most of their HGT regions. All non-outbreak strains for which DaisyGPS did not report EMRSA-15 as an acceptor candidate are clustered away furthest from the outbreak strains (N1, N3—N5). The likely related samples N3 and N4 are clustered together. Regarding a distinction of outbreak and non-outbreak strains, DaisySuite is able to determine the outbreak-related HGT regions which differ from the HGT candidates for the non-outbreak strains. Hence, a distinction is possible. Although DaisySuite only called one HGT region for O6, we can deduce from the coverage profile that more HGT regions called for the other outbreak samples are present as well but were missed by DaisySuite. As can be seen in the heatmap, clusters 34 and 37 are not covered by any sample and hence likely FPs. We detected the AMR gene mecA on Cluster 0, however, resistance is shared among all 14 samples according to [36]. No further AMR genes tested by [36] are detected on the other clusters. However, most of these AMR genes are on plasmids that were not analysed here.
We presented DaisyGPS, a pipeline that utilises metagenomic profiling strategies to identify acceptor and donor candidates from NGS reads of a potential HGT organism. DaisyGPS, together with Daisy, is part of the comprehensive HGT detection suite DaisySuite. We successfully validated DaisyGPS on simulated and real data previously analysed in [16]. We further demonstrated robustness of the DaisySuite on a large-scale simulation with 100 negative HGT tests, showing that DaisySuite correctly reports no HGT events with a specificity of 97%. On a large-scale simulation with 100 positive HGT simulations, DaisySuite reports the correct HGT event with a total sensitivity of 69%. From the 818 pairs reported by DaisyGPS among the 100 simulations, Daisy called the TP and TN regions with a sensitivity of 89%. Lastly, we evaluated DaisySuite on an MRSA outbreak data set with seven outbreak associated samples and seven not associated with the outbreak but that occurred during the same time frame. Here we could show that DaisySuite successfully distinguishes between associated and not associated samples regarding their suggested HGT regions, i.e. the outbreak samples show a distinct number and content of reported HGT regions.
One has to acknowledge that all outbreak strains have a high sequence similarity to the EMRSA-15 strain, which is not necessarily the case for the non-outbreak strains. This is also reflected in the results from DaisyGPS where S.aureus HO 5096 0412 is the best acceptor candidate for all outbreak strains but not reported at all for some non-outbreak strains. It directly follows that a sequence comparison based analysis as done with DaisySuite will likely find different patterns for the outbreak and non-outbreak strains, and a difference in HGT region candidates might seem obvious. However, starting from having established such a difference, there is value in then analysing the shared HGT region candidates among the outbreak-related strains. For this proof of concept, we performed a relatively simple evaluation by performing a coverage analysis of all HGT regions across all samples and investigating the presence of AMR genes within the HGT regions. But a future thorough follow-up analysis of the origin and functionality provided by the potential HGT sites could benefit our understanding of the risk and pathogenicity of these outbreak strains.
The observed FP and FN candidates, however, also reveal weaknesses of the sequence comparison approach. DaisyGPS is designed with a focus on sensitivity and hence inevitably leads to FP acceptor and donor candidate pairs to be examined by Daisy. Since these FPs are still due to a sufficient degree of mapping coverage, spurious split-reads and spanning reads can cause downstream FP calls as observed for the simulated data set from E.coli K12 DH10 and H.pylori. The reported HGT site from H.ducreyi has only similarities in the start and end part of the proposed region compared to the transferred H.pylori region though. Insertion sites can also lie within repeat regions which enhances the negative impact of ambiguous mappings. This emphasises that a critical evaluation of HGT predictions is always crucial. To help interpret the HGT predictions from DaisySuite, the reported acceptor and donor candidates are ranked according to their respective score, and only the HGT sites passing the user defined thresholds (listed in the complete TSV results file) are reported in the final VCF results. In the supplementary results tables, we stated the parameters used for filtering or adjusting to the requirements of the data set. We also provide a documentation on usage at https://daisysuite.readthedocs.io/en/latest/tutorial/example.html.
From the missing HGT region calls for sample O6 that could be inferred from the coverage analysis, we can deduce that DaisySuite does not detect all HGT regions due to insufficient evidence. A potential cause could be that DaisyGPS did not report the correct donor reference. Even if DaisyGPS could find an appropriate donor genome, it is still likely that the genome content differs between the region present in the donor and the region actually present in the HGT organism. An alternative, complementary approach to cope with this problem of a lack of a suitable donor candidate could be to facilitate local, insertion sequence assembly. By offering identified insertion sequences, we can still provide the content of a potential HGT sequence and thereby enable downstream analysis. This approach would also support the detection of novel HGT sequences not present in current reference databases, and therefore also the detection of, e.g., novel antimicrobial resistance genes. Popins [40] is a tool for population-based insertion calling developed for human sequencing data (see, e.g., [41]). Popins only locally assembles unmapped reads (same input as for Daisy) with Velvet guided by a reference, thereby minimising the risk of potential misassemblies. On top of the assembly, Popins first uses spanning pairs (see red read pairs in Fig 1) to place an insertion in the (acceptor) reference, and then performs a local split-read alignment around the potential breakpoint. If multiple samples are provided, Popins merges contigs across samples into supercontigs, assuming that the same insertion is present in multiple samples. Although different bacterial samples do not represent a population as given for human populations, outbreak related samples still resemble a population such that one could use Popins for this purpose and gain valuable information. However, local insertion assembly only gives evidence for an insertion compared to the chosen acceptor reference, that does not necessarily mean that the insertion resulted from an HGT event. Hence, means to sophistically include insertion assembly results into the HGT context need to be defined first. Despite the evidence for an HGT event that DaisySuite can provide, the results should always be tested for alternative causations such as gene loss.
Our metagenomic analyses show that DaisySuite is able to detect HGTs not only from pure samples. However, the automatic detection of HGT events with DaisySuite in metagenomic samples has limitations if the diversity within the sample gets more complex. DaisyGPS uses the metagenomic classification tool MicrobeGPS, and hence, identifies organisms in the sample as part of the pipeline. All identified organisms with a homogeneous coverage are—per se—possible acceptor candidates. We increased the thresholds for the reported acceptor and also donor candidates to 30 entries, respectively, and limited the number of candidates per species to one so that the ground truth acceptor and donor of the simulated H. pylori are still listed. Note that this number not only depends on the number of organisms in the sample but also on their sequence similarity—especially to the expected acceptor and donor candidates.
The resulting 400 Daisy runs would require too much compute time and space for a systematic and automatic follow up. In general for metagenomic samples, we would recommend to only run DaisyGPS first and then define a confined set of likely candidates for follow up analysis. For future developments, we would suggest to integrate another mapping-based filtering for this definition where we would search for likely pairs via paired-end reads with one read mapping to an acceptor and the other to a donor candidate. We use this criterion also in the Daisy follow up as evidence but in our opinion it would also serve well for candidate (pair) filtering.
[42] applied a method that is similar to Daisy to detect mobile genetic elements (MGEs) in the human gut microbiome. Although this study shows the general applicability of our approach in a large scale metagenomic study, the focus here can only be the collection of now present or absent MGEs in the microbiome (rather than particular strains). [42] also point out that such a MGE characterisation is more meaningful in a time series analysis rather than from a single sample snapshot. Daisy has also been applied to infer horizontally transferred genes in the Daphnia iridescent virus 1 [43] which shows that our approach can be further applied in other contexts than bacteria.
DaisySuite uses mapping-based similarity to determine candidates. This can lead to biases if the true candidates are missing in the database or for historic events that are obscured through amelioration. DaisyGPS will still report the next best candidates (i.e. with the most sequence similarity) but the FPs in our large scale simulation arising from Blast hits already show the potential for downstream errors. Further, our simulation over evolutionary distances clearly show the limitations for acceptor and donor identification above a certain distance. This limitation also goes hand in hand with a sufficient sequencing coverage to avoid further bias by random sequencing errors, and also to allow a reliable Daisy follow-up analysis. From our experiments, we would recommend to provide at least a 10x sequencing coverage.
DaisySuite facilitates the capabilities of programs designed for different tasks, including mapping, metagenomic profiling and structural variant detection. Although this allows us to combine the strength of each tool to tackle the problem of HGT detection, we are also vulnerable to bottlenecks regarding the runtime of single steps. In particular, data sets that create big mapping results and/or contain many split reads may increase the runtime significantly. In general, the overall runtime ranges around one to two hours on a standard machine to process a standard sample, e.g., the H. pylori data set. However, very big or diverse data sets, such as created in our genetic divergence experiment, will increase the runtime manifold and in extreme cases render them infeasible to run. The main bottleneck for DaisyGPS is the metagenomic profiling via MicrobeGPS, whereas for Daisy the split read detection by Gustaf and—if Gustaf detects enough split reads—the HGT detection itself. In the future, we hope to alleviate this problem by modernising or helping to modernise the respective tools.
As with all computational methods, they cannot fully replace critical human thinking and should be cross validated by other means. In an HGT detection study, we would recommend to use other HGT detection methods (computational and/or wet lab) to support findings by individual methods. Although we see this as crucial, we think it lies outside the scope of DaisySuite to provide such a cross validation.
With DaisyGPS, we present a tool for acceptor and donor identification from NGS reads of an HGT organism. To do that, DaisyGPS refines metrics already defined and used for metagenomic profiling purposes to account for the acceptor and donor specific coverage profiles. We integrated DaisyGPS with Daisy into a comprehensive HGT detection suite, called DaisySuite, that provides an automatic workflow to first determine acceptor and donor candidates and then identify and characterise HGT regions from the suggested acceptor-donor pairs. We successfully evaluated DaisyGPS on data previously analysed with Daisy, and demonstrated sensitivity and robustness of the DaisySuite in a large-scale simulation with 100 simulated positive and negative HGT events. We could further show the benefits of an HGT analysis with DaisySuite on an MRSA outbreak data set where DaisySuite reported HGT candidates that help to distinguish between outbreak associated and unassociated samples and therefore also provide information for outbreak strain characterisation.
|
10.1371/journal.pntd.0001553 | Mapping of Schistosomiasis and Soil-Transmitted Helminthiasis in the Regions of Centre, East and West Cameroon | Schistosomiasis and soil-transmitted helminthiasis (STH) are widely distributed in Cameroon. Although mass drug administration (MDA) of mebendazole is implemented nationwide, treatment with praziquantel was so far limited to the three northern regions and few health districts in the southern part of Cameroon, based on previous mapping conducted 25 years ago. To update the disease distribution map and determine where treatment with praziquantel should be extended, mapping surveys were conducted in three of the seven southern regions of Cameroon, i.e. Centre, East and West.
Parasitological surveys were conducted in April–May 2010 in selected schools in all 63 health districts of the three targeted regions, using appropriate research methodologies, i.e. Kato-Katz and urine filtration.
The results showed significant variation of schistosomiasis and STH prevalence between schools, villages, districts and regions. Schistosoma mansoni was the most prevalent schistosome species, with an overall prevalence of 5.53%, followed by S. haematobium (1.72%) and S. guineensis (0.14%). The overall prevalence of schistosomiasis across the three regions was 7.31% (95% CI: 6.86–7.77%). The prevalence for Ascaris lumbricoides was 11.48 (95% CI: 10.93–12.04%), Trichuris trichiura 18.22% (95% CI: 17.56–18.90%) and hookworms 1.55% (95% CI: 1.35–1.78%), with an overall STH prevalence of 24.10% (95% CI: 23.36–24.85%) across the three regions. STH was more prevalent in the East region (46.57%; 95% CI: 44.41–48.75%) in comparison to the Centre (25.12; 95% CI: 24.10–26.17%) and West (10.49%; 95% CI: 9.57–11.51%) regions.
In comparison to previous data, the results showed an increase of schistosomiasis transmission in several health districts, whereas there was a significant decline of STH infections. Based on the prevalence data, the continuation of annual or bi-annual MDA for STH is recommended, as well as an extension of praziquantel in identified moderate and high risk communities for schistosomiasis.
| Schistosomiasis and soil-transmitted helminthiasis (STH) are a major public health problem in Cameroon. The national control strategy of these diseases was based on historical data collected 25 years ago, which might be outdated in some situations due to several factors including control activities, improved or degraded sanitation and hygiene, socio-economic improvement and disease transmission dynamics. To help planning, improving control strategies and evaluation of control activities, there was a need to update the distribution of schistosomiasis and STH. We conducted parasitological surveys in three regions of Cameroon, i.e. Centre, East and West. Our results showed a significant decrease of STH infection prevalence and intensities in all these three regions, in comparison to previous mapping data, with an overall decline of prevalence from 81.1–93% to 10.5–46.6%. These results show the positive impact of annual deworming campaigns, and illustrate the progressive success of the national programme for the control of schistosomiasis and STH in Cameroon. Furthermore, our results showed an increase of the number of high transmission foci of schistosomiasis, and allowed identifying new health districts requiring mass treatment with praziquantel, and those where deworming should be reinforced.
| Recent years have witnessed an increased interest in the control of neglected tropical diseases (NTDs), and today there exists a global momentum for the control of these diseases, as well as an unprecedented opportunity for cost-effective action, through an integrated control [1]–[5]. Interest in the integrated control of NTDs is currently at an all-time high, due in part to new funding committed by a number of governmental and non-governmental donors, high-level political commitment in the endemic countries, and the existence of donated anthelminthic drugs which can be safely co-administrated and used in a coordinated way to address these scourges [6]–[8]. Four of these diseases are mainly controlled through the ‘preventive chemotherapy’ intervention, i.e. schistosomiasis, soil-transmitted helminthiasis (STH), onchocerciasis and lymphatic filariasis, according to the World Health Organization (WHO) recommendations [4]. Schistosomiasis and STH occur throughout the developing world and remain a major public health problem in the poorest communities with enormous consequences for development. Praziquantel is the sole drug for treatment and morbidity control of schistosomiasis in sub-Saharan Africa. Control of STH uses two main drugs, i.e. albendazole or mebendazole. Based on infection prevalence, communities can be classified into low-risk (<10% for schistosomiasis and <20% for STH), moderate-risk (≥10% but <50% for schistosomiasis and ≥20% but <50% for STH) and high-risk (≥50% for both) categories according to the WHO disease specific thresholds, and this classification is used to determine the appropriate treatment regimen as specified in the WHO guidelines [4].
In Cameroon, it is estimated that more than 5 million people are at risk of infection with schistosomiasis, and 2 million persons are currently infected [9]. STHs are widely distributed all over the country, and it is estimated that more than 10 million people are infected with intestinal worms [9]. The national epidemiological survey conducted in 1985–1987 showed the occurrence of three species of schistosomes: Schistosoma haematobium, S. mansoni and S. guineensis (formerly S. intercalatum Lower Guinea strain [10], [11]); and three major species of STH: Ascaris lumbricoides, Trichuris trichiura and Necator americanus. The highest transmission levels of schistosomiasis occurred in the savannah areas of the northern Cameroon, whereas STHs were more prevalent in the southern forest part of the country [12]–[14]. School-aged children are the most infected, and polyparasitism is very frequent; with a high proportion of children carrying at least 2 species of parasites [15].
Cameroon adopted a strategic plan for the control of schistosomiasis and STH in 2004. Starting with very limited budget, the control programme gradually mobilized national and international partners to enable a rapid scaling-up of activities to encompass all ten regions in 2007. Since then, national deworming campaigns were implemented annually. School-aged children were treated with mebendazole nationwide, whereas praziquantel was distributed only in high endemic areas for schistosomiasis [16]. Interestingly, the Government of Cameroon recently moved into an integrated approach for the control of NTDs, including co-implementation of different control interventions and co-administration of several drugs, i.e. praziquantel, ivermectin, mebendazole and albendazole. This integrated approach is the basis for cost-effectiveness and streamlined efficiency. Since 2009, Cameroon receives assistance from the United States Agency for International Development (USAID) through its NTD Control Program to facilitate integration of national programs and support mass drug administration (MDA) [17].
Because knowing the distribution of the targeted NTDs is essential for developing an adequate implementation strategy and types of drug co-administrations, one of the efforts of the USAID's NTD control program in Cameroon was focused on updating the disease-distribution information. Hence, efforts were made to support on-the-ground activities to map the disease distribution where sufficient information was not available. Indeed, the baseline data for schistosomiasis and STH in Cameroon were collected 25 years ago [12], [13]. It is well known that the transmission of these diseases is dynamic over time, particularly after years of treatment and other health interventions [18]. Therefore, epidemiological surveys were scheduled in the different regions of Cameroon in order to update the distribution and the level of endemicity of schistosomiasis and STH to facilitate the planning of implementation strategies in these regions. The first study phase targeted three of the ten regions of Cameroon, i.e. Centre, East and West. The present paper reports the outcome of the mapping exercises, compares the current situation with the baseline data from 1980s, and provides recommendations for the control of schistosomiasis and STH in these regions.
The study was approved by the National Ethics Committee of Cameroon (Nr 082/CNE/DNM/09), and was a public health exercise through the Ministry of Public Health and the Ministry of Basic Education. Parasitological surveys were conducted in schools with the approval of the administrative authorities, school inspectors, directors and teachers. Information about the national programme for the control of schistosomiasis and STH, and the objectives of the study were explained to the schoolchildren and to their parents or guardians from whom written informed consent was obtained. Children willing to participate were registered. Each child was assigned an identification number and data collected were entered in a database. No identification of any children can be revealed upon publication. Children were treated during the MDA campaign implemented by the national control programme.
Cameroon is divided up into a three-tiered system including 10 regions at the first level, 58 divisions (departments) at the second level, and 360 sub-districts (arrondissements) at the third level. The population of Cameroon is estimated to be 19,406,100 inhabitants in 2010. Population density shows marked variation across the country, ranging from a mean of 7.4 inhabitants/km2 in the East region to 141.5 inhabitants/km2 in the Littoral region. School-aged children account for 28% of the country population and are estimated at 5,433,708 [19]. The health system in Cameroon is decentralized and organized into central, regional and district levels. There are 179 health districts. The three regions targeted for mapping, i.e. Centre, West and East are located in the southern forest area of the country. These regions are subdivided in 29, 14 and 20 health districts, respectively.
A stratified random-cluster sampling procedure, with the 5th grade as the basic sampling unit, was used in the previous mapping of schistosomiasis and STH in Cameroon, conducted in 1985–1987 [12], [20]. In order to assess the current levels of infections and to compare the data with previous ones, the schools were selected using the list of villages and schools previously investigated, the ecological zones and the risk factors for schistosomiasis transmission [21], [22]. Selection was made so that all health districts in the three targeted regions of Cameroon were covered spatially. Due to financial limitations, an average of four primary schools (proportional to the district's size and population density) was selected per health district. The geographical co-ordinates of each of the sampled schools were recorded with global positioning system (GPS) devices. The study was conducted in April–May 2010.
In the 1985–1987 study, a 10 ml urine sample and a single Kato-Katz slide were examined for schistosome and STH infections [12], [20]. In the current study, in each school, urine and stool samples were collected from 50 children selected randomly in the upper classes, approximately half boys and half girls. Children were preferentially selected from the 5th grade, and then in other grades where the number of children in the 5th grade was fewer than 50. The samples were collected in 60 mL plastic screw-cap vials, between 10.00 and 14.00 hours. The samples were preserved with sodium azide [12], [20] and transported to the Centre for Schistosomiasis & Parasitology in Yaoundé for examination. In the laboratory, each urine sample was agitated to ensure adequate dispersal of eggs, 10 mL of urine were filtered through a Nucleopore® filter, and the filters were examined by microscopy for the presence of schistosome eggs. Stool samples were examined by a single thick smear technique using a 41.7 mg Kato-Katz template. Each Kato slide was read twice; immediately after slide preparation for hookworm eggs, and the following day for schistosome and other STH eggs. Parasitic infections were recorded; number of eggs for each parasite was counted; and intensity of infection was calculated and expressed as eggs per gram of feces (epg) or eggs per 10 ml of urine (egg/10 ml).
The different parasitological data were analyzed by the epidemiological unit of the Centre for Schistosomiasis & Parasitology using appropriate statistical tests and methods. The data were subsequently exported into SPSS (IBM, Version 19) for statistical analysis. The Complex Samples Crosstabs procedure was used for calculating the prevalence and the Descriptives procedure was used for calculating the intensity of infections, taking into account the cluster nature of schools with districts as strata and schools as clusters and including the finite population correction assuming equal probability sampling without replacement. Sample weighting was applied for each district according to the ratio of the proportionally expected number of schools to be surveyed and the number of actually surveyed schools in each district assuming similar number of children in each school [23]. The 95% confidence intervals (CIs) for prevalence were calculated using the Wilson score method without the continuity correction after adjusting for sample weighting [24]. Arithmetic mean intensities of infection with 95% CIs for different parasite species were calculated including all children examined [25]–[27]. The Chi-square test using the Complex Samples Crosstabs procedure was used to investigate the relationship between prevalence of infections and sex, age groups, districts and regions, and the Complex Samples Logistic regression procedure was used to compare the differences in prevalence between 1985–1987 and 2010. The Kruskal-Wallis test was used to compare the differences in intensities of infections. The levels of endemicity of schistosomiasis and STH and the degrees of intensity of individual infections were categorized according to the WHO recommendations [4], [28]. A geographical information system (GIS) software ArcGIS (ESRI Inc., Version 9.2) was used to plot the point prevalence of the infections for each surveyed school on a map.
A total of 244 schools were surveyed: 118 in the Centre region, 67 in the East region and 59 in the West region. A total of 12 594 pupils aged 2–23 years old (6251 males and 6343 females) from these 244 schools were registered and included in the study. Of these children registered, 12 486 (99.14%) provided urine samples and 12 243 (97.21%) provided stool samples. The mean age (± standard deviation) of children examined was 11.30±1.98 years (male: 11.45±2.0 and female: 11.15±1.94).
The arithmetic mean intensity of infection in the three regions for each species of schistosomiasis is shown in Table 1. The egg counts for intestinal schistosomiasis ranged from 0 to 13,818 epg, and from 0 to 2,600 eggs/10 ml for urinary schistosomiasis. The overall arithmetic mean infection intensity was 33.24 epg for S. mansoni, 2.46 eggs/10 ml for S. haematobium, and 0.23 epg for S. guineensis. The Centre region was most heavily infected with S. mansoni (61.04 epg) and the West region with S.haematobium (6.86 eggs/10 ml). It appears that infections were light (<100 epg) in the majority of schools, with only 2.5% moderate or heavy S. mansoni infections and 0.72% heavy S. haematobium infections across the three regions (Table 1). Boys were more heavily infected with S. mansoni or S. haematobium than girls (p<0.01). The age distribution of intensity of infection for individual schistosome species is shown in Figure 3. Intensity of infection increased with age for S. haematobium in children examined while children of 9–14 years old were more heavily infected with S. mansoni (p<0.001).
The current distribution of schistosomiasis and STH in 2010 was compared with the distribution in 1985–1987 [12]–[14], using the overall schistosomiasis and STH prevalence. The prevalence distribution of schistosomiasis in 1985–1987 and in 2010 is shown in Figures 5A and 5B, respectively, with the prevalence categorized according to the WHO prevalence thresholds [4]. It shows that the overall endemic areas of schistosomiasis did not change significantly. However, there was an increase in the number of high transmission foci of schistosomiasis in several health districts; e.g. health district of Malantouen in the West region where prevalence was up to 95.92% in the village of Matta, and health districts of Mbalmayo and Bafia in the Centre region with prevalence up to 71.43% and 52.78% in the villages of Dzeng and Yorro, respectively. Statistical comparison was carried out taking into account the geographical location of districts, age and sex. The results are shown in Table 2. Compared with the 1985–1987 data, the overall schistosomiasis prevalence in 2010 across the three regions and that in the Centre region did not change significantly (p>0.05), while prevalence in the East region decreased and prevalence in the West region increased, both significantly (p<0.01), though the level of infection in these two regions were relatively lower. Among the three schistosome species, the overall S. haematobium prevalence remained unchanged (p>0.05), while the overall S. mansoni prevalence significantly increased from 4.3% to 5.53% (p<0.05), and that of S. guineensis, though low, decreased significantly (p<0.001).
The prevalence distribution of STH in 1985–1987 and in 2010 is shown in Figure 6. There was a clear and significant decrease of STH prevalence in all three regions. Indeed, statistical comparison showed that the overall STH prevalence declined significantly from 93.02%, 92.34% and 81.14% to 25.12%, 46.56% and 10.51% in the Centre, East and West regions, respectively (all p<0.001) (Table 2). However, the decrease of STH was significantly lower in the East region in comparison to the two other regions. Detailed analysis of individual STH species showed significant reductions of 86.99% for hookworms, 82.27% for A. lumbricoides and 78% for T. trichiura (all p<0.001).
Analysis of polyparasitic infections showed that in 1985–1987, 61.93% of school-aged children examined were infected with more than one and up to five parasite species, but this proportion decreased significantly to 10.19% in 2010 (p<0.001) (Table 2).
The present study showed that schistosomiasis was moderately endemic (prevalence between 10–49%) in 23 of the 244 schools investigated, and highly endemic (prevalence ≥50%) in 4 schools. These moderate and high-risk communities are distributed in 13 of the 63 health districts investigated. The results confirmed the typical focal distribution of schistosomiasis in these regions. When comparing our results with the previous nationwide data collected in 1985–1987 by Ratard et al. [12], it appears a slight increase of the number of high transmission foci of schistosomiasis and an overall increase of S. mansoni infections – the most prevalent schistosome species in the three regions. This is not surprising given the fact that no MDA with praziquantel had been implemented in these health districts since the last mapping survey, apart from the health district of Ndikinimeki in the Centre region. The national control programme for schistosomiasis and intestinal helminthiasis was officially launched in 2004 in Cameroon [16]. Since 2007, school-aged children had been dewormed annually with mebendazole nationwide in all 179 health districts, whereas praziquantel were distributed only in schistosomiasis highly endemic health districts, including all 51 health districts of the three northern regions of Cameroon, where the highest transmission level of schistosomiasis were found [12], [29], and only in one of the 63 health districts of the three investigated regions, i.e. the district of Ndikinimeki, Centre region. The comparison of 1985 and 2010 data showed a significant decrease of schistosomiasis prevalence within the health district of Ndikinimeki, with a decline from 81.60% to 41% in the town of Makenene for example. Changing situation of schistosomiasis varied in the three regions and among the three different species, and this may reflect the differences in transmission dynamics in these different regions. The main factors influencing schistosomiasis transmission may include the changing demographic situation, socioeconomic development, water and sanitation, snail population dynamics etc. However, such information was not collected in the current mapping survey, which may be a topic for future studies.
One of the key outcomes and recommendations from this study is that in future deworming campaigns, the distribution of praziquantel should be undertaken in all 13 health districts in these three regions where schistosomiasis prevalence were ≥10%, according to WHO preventive chemotherapy guidelines [4]. Considering the overall low endemicity of schistosomiasis in the majority of these health districts, treatment will be conducted at district level in rural zones, whereas in urban settings treatment will be focused in those sub-districts with high prevalence spots of schistosomiasis. It should be noted that in both 1985 and 2010 surveys, single Kato-Katz slides were conducted as commonly used for mapping studies. Therefore, the prevalence and abundance of S. mansoni and STH may have been underestimated, due to the low sensitivity of Kato-Katz technique and day-to-day variation in egg excretions, particularly in light infections.
For STH, our study showed an overall significant decrease of infection prevalence in all three regions investigated, in comparison to previous mapping data collected in 1985–1987 [13], [14], [29]. Indeed, the STH prevalence declined from 93% to 25.1% in the Centre region, from 81% to 10.5% in the West region, and from 92.3% to 46.6% in the East region. These results clearly illustrate the positive impact of the school-based deworming campaigns with mebendazole implemented annually by the Ministry of Public Health, through the National Programme for the Control of Schistosomiasis and Intestinal Helminthiasis. The decline was lower in the East region compared to the two other regions. The previous mapping data showed that the three regions surveyed were among the higher STH prevalence areas within the country [13], [14]. Apart from the ivermectin MDA implemented in onchocerchiasis endemic communities, these regions have not been subjected to albendazole distribution which is used for lymphatic filariasis control. It is therefore interesting to see that the overall STH prevalence has been reduced so much by mainly mebendazole distribution. Though it has been shown that mebendazole is not as efficient as albendazole in deworming, particularly for hookworms [30]–[32], the present data show that mebendazole still has a significant role to play in the current effort to control NTDs. Several other factors, such as socio-economic development, improved sanitation and hygiene, environmental changes and collateral effect of other drugs, may have also contributed to the reduction of STH transmission. However, as discussed above this may be a topic for future studies.
Despite the observed significant reduction of STH infections, the prevalence and intensities of A. lumbricoides and T. trichiura infections were still relatively high, particularly in the East region. Several factors may explain the lower reduction of STH infections in this region, including the low socio-economic status and poor sanitation in most of the rural settings, which favor high parasite transmission and frequent human re-infections. The East region is the largest and the most sparsely populated region in Cameroon. The vast majority of its inhabitants being subsistence farmers, the low level of development in the region, and its thick forests and equatorial climate are favorable factors for STH and other NTDs. Also, the lower school attendance rates in villages, in comparison to towns, may have affected the treatment coverage of all school-aged children through a single school-based deworming campaign approach. It is well known that the epidemiology of STH infections is influenced by several determinants, including environment, population heterogeneity, age, household clustering, genetics and polyparasitism [33]. STHs affect the poor and infections are particularly abundant among people living in rural or deprived urban settings with low socio-economic status and poor sanitation [34]. Further investigations should be conducted to identify the major factors affecting the deworming effect in order to improve the impact of the current integrated NTD control programme.
The mapping results showed that the majority of health districts (34 over the total of 63, i.e. 53.97%) were still within the STH infection categories requiring large-scale preventive chemotherapy interventions, i.e. infection prevalence ≥20%. In communities with prevalence ≥50%, WHO recommends treatment of all school-aged children – enrolled and not enrolled – twice per year, and even three times if resources are available; whereas in communities where prevalence is ≥20% but <50%, school-aged children should be treated once a year. Therefore, the government of Cameroon should continue implementing annual deworming of school-aged children in all districts of the Centre, East and West regions. In addition, preschool children, women of childbearing age and adults at high-risk in certain occupations should also be treated, according to WHO recommendations [4]. In particular, in the East region where STH infection prevalence and intensities remain very high, it should be envisaged to deworm school-aged children at least twice a year. Furthermore, the alternating use of mebendazole and albendazole from one deworming round to another should be envisaged to optimize treatment efficacy against STHs [35].
Finally, the results of the present study highlight the new health districts where the MDA of praziquantel should be implemented for the treatment of schistosomiasis. For future deworming campaigns, all school-aged children should be treated with praziquantel in moderate (i.e. prevalence ≥10% but <50%) and high-risk communities (i.e. prevalence ≥50%). Also, praziquantel should be made available in dispensaries and clinics for treatment of suspected cases, in accordance with WHO recommendations [4]. Interestingly, this study provided data for accurate estimation of increased praziquantel needs, and the results will contribute to update global information on the distribution of schistosomiasis and STH, recently developed as an open-access database [36], [37].
|
10.1371/journal.pntd.0005884 | Feasibility of utilizing the SD BIOLINE Onchocerciasis IgG4 rapid test in onchocerciasis surveillance in Senegal | As effective onchocerciasis control efforts in Africa transition to elimination efforts, different diagnostic tools are required to support country programs. Senegal, with its long standing, successful control program, is transitioning to using the SD BIOLINE Onchocerciasis IgG4 (Ov16) rapid test over traditional skin snip microscopy. The aim of this study is to demonstrate the feasibility of integrating the Ov16 rapid test into onchocerciasis surveillance activities in Senegal, based on the following attributes of acceptability, usability, and cost. A cross-sectional study was conducted in 13 villages in southeastern Senegal in May 2016. Individuals 5 years and older were invited to participate in a demographic questionnaire, an Ov16 rapid test, a skin snip biopsy, and an acceptability interview. Rapid test technicians were interviewed and a costing analysis was conducted. Of 1,173 participants, 1,169 (99.7%) agreed to the rapid test while 383 (32.7%) agreed to skin snip microscopy. The sero-positivity rate of the rapid test among those tested was 2.6% with zero positives 10 years and younger. None of the 383 skin snips were positive for Ov microfilaria. Community members appreciated that the rapid test was performed quickly, was not painful, and provided reliable results. The total costs for this surveillance activity was $22,272.83, with a cost per test conducted at $3.14 for rapid test, $7.58 for skin snip microscopy, and $13.43 for shared costs. If no participants had refused skin snip microscopy, the total cost per method with shared costs would have been around $16 per person tested. In this area with low onchocerciasis sero-positivity, there was high acceptability and perceived value of the rapid test by community members and technicians. This study provides evidence of the feasibility of implementing the Ov16 rapid test in Senegal and may be informative to other country programs transitioning to Ov16 serologic tools.
| As onchocerciasis control programs succeed and transition to elimination efforts, different diagnostic tools are needed. The goal of this study was to determine if integrating the Ov16 rapid test is feasible based on acceptability, usability, and cost. A study was conducted in 13 villages in southeastern Senegal in May 2016. Community members were invited to participate in a demographic questionnaire, a rapid test, a skin snip biopsy, and an acceptability interview. Technicians were also interviewed and a costing analysis was conducted. Out of 1,173 participants, 1,169 (99.7%) agreed to the rapid test while 383 (32.7%) agreed to skin snip microscopy. The rapid test result was reactive in 2.6% of those tested, while none of the skin snips were positive. Community members thought the rapid test was performed quickly, was not painful, and provided reliable results. If no one had refused skin snip microscopy, the total cost would have been around $16 per person tested for either method. In this area with little if any remaining onchocerciasis, there was high acceptability and perceived value of the rapid test. This study suggests that implementing the Ov16 rapid test in Senegal is feasible and these findings may be informative to other country programs.
| Onchocerciasis, commonly known as river blindness, is caused by the filarial parasite O. volvulus (Ov) that affects an estimated 37 million people, with an estimated 187 million living in areas at risk of infection, primarily in Africa.[1,2] An estimated 1.1 million disability-adjusted life years (DALYs) were lost in 2015 due to onchocerciasis, as it can lead to severe and disfiguring skin disease, visual impairment, and eventually blindness.[3] Onchocerciasis especially affects poor rural communities and the risk of infection is substantially higher among socioeconomically disadvantaged groups.[4] In Africa, efforts to date have focused primarily on disease control through mass drug administration (MDA) with ivermectin, an antiparasitic drug donated by Merck.[5,6]. Recent evidence from Sudan, Senegal, Mali and Uganda suggests elimination is possible in Africa as it is in the Americas.[6–11]
In response to this success, the global strategy has shifted from disease control to disease elimination.[6,12] The 2016 World Health Organization (WHO) guidelines on stopping MDA and verifying elimination describe three phases of onchocerciasis elimination programs that require different diagnostic tools: transmission suppression, transmission interruption, and transmission elimination.[13] The standard method is direct observation of the Ov microfilaria in a skin snip biopsy using microscopy. Skin snip microscopy is highly specific and able to detect active infections, but has diminished sensitivity in low-prevalence settings. As the prevalence of onchocerciasis in endemic communities decreases, more sensitive diagnostic tests are needed.[14]
Ov16 serology, used to detect IgG antibodies to the Ov16 antigen in a sentinel population of children under ten years, is now recommended to determine if interruption of transmission of Ov has occurred.[13] Laboratory-based Ov16 ELISA (enzyme-linked immunosorbent assay) is one method to measure these markers, though it requires collecting samples in the field to transport to a laboratory setting for analysis. Currently, there is no standardized commercially available Ov16 ELISA so variations in protocols and procedures exist across labs.[15,16] In 2014, a field deployable, rapid diagnostic tool that could be more easily integrated into current onchocerciases surveillance programs in endemic countries was developed and made commercially available (SD BIOLINE Onchocerciasis IgG4 rapid test, referred to here as Ov16 rapid test).[17,18] Performance of the Ov16 rapid test continues to be evaluated in the field and current global research priorities focus on operational and implementation research to demonstrate utility and increase access of the Ov16 rapid test, particularly in low prevalence settings which have undergone multiple rounds of MDA.[14,19]
In Senegal, MDA and surveillance has been ongoing since 1988 and has resulted in the successful control of onchocerciasis.[7–9] Additionally, MDA for onchocerciasis and lymphatic filariasis (LF) are now integrated. After over 25 years of control efforts, program managers require more clarity around whether transmission has been interrupted. Though skin snip microscopy has been used to this point, it is a painful and invasive procedure that may result in decreased participation in surveillance activities in communities where decades of testing have occurred. Implementation research on the Ov16 rapid test in Senegal was desired to support the program transition to elimination. In 2015, a workshop was held with representatives from the Senegal Ministry of Health and Social Action (MoH) to discuss the current process for using skin snip microscopy and evaluate the potential process if they were to use the rapid test, to streamline introduction of the new test. Acceptability, usability and costing data was also needed to inform decisions on use of the tests in surveillance activities.
The aim of this study is to demonstrate the feasibility of integrating the Ov16 rapid test into onchocerciasis surveillance activities in Senegal, based on the attributes of acceptability, usability, and cost. Quantitative and qualitative methods are used to evaluate the following outcomes: 1) the diagnostic results of the Ov16 rapid test compared to skin snip microscopy; 2) an assessment of the acceptability and usability of the rapid test among community members and health workers; and 3) an estimation of the economic costs to conduct a surveillance activity by diagnostic method from the government’s point of view. A recently developed comprehensive quality assurance (QA) program was also piloted to support proper use of the rapid tests. This implementation research is intended to build evidence to support the introduction of the Ov16 rapid test in Senegal, as well as to inform other settings that may be at a comparable phase of elimination programming.
A cross-sectional study using qualitative and quantitative methods was conducted to assess the feasibility of integrating the Ov16 rapid test into ongoing surveillance activities. The study was conducted along with the Senegal MoH, which currently utilizes skin snip microscopy for onchocerciasis diagnosis. The study was performed in 13 villages in the Kédougou and Saraya districts of southeastern Senegal in May 2016. Villages were representative of the region endemic for onchocerciasis in Senegal, and are co-endemic for LF. These communities started MDA with ivermectin (IV) in 1988. Albendazole was added to the MDA in 2015 and was last administered in these villages in March 2015. Individuals 5 years and older were invited to participate in any or all components of the study, including a demographic and health history questionnaire, the Ov16 rapid test, two skin snip biopsies for microscopy, and an exit interview. Community sensitization was conducted in each village 2–3 days prior to the surveillance activity.
This study was approved by the PATH Research Ethics Committee and the Senegal National Ethics Committee for Health Research. Informed consent or assent was obtained from all participants. All participants 18 years and older provided written informed consent, and all participants under 18 years provided assent in addition to their parent or guardian providing written informed consent.
Prior to study start, a comprehensive quality assurance program was introduced and a training on proper use of the Ov16 rapid test was conducted. The QA program includes training resources such as videos and PowerPoint slides, as well as a quality assurance panel to verify a quality product was received, and daily quality controls to ensure proper functioning of the test throughout data collection. For more information: http://sites.path.org/dx/ntd/training-and-qaqc-materials/. The Ov16 rapid test was performed per the product instructions, which involves transferring 10 μL of finger stick capillary blood to the cassette using a disposable capillary tube that is included with the test. After buffer is added to the cassette, the test runs for 20 minutes and then results are recorded. All rapid test results were read a second time the next day as a research activity to compare 20 minute and overnight results. Skin snip microscopy was performed by taking two skin snips from the iliac crests with a sterile 2 mm corneoscleral punch biopsy tool. The skin snips were incubated in distilled water for 30 minutes, then examined under a light microscope to detect the presence of Ov microfilaria. Skin snips that were negative at 30 minutes were incubated in saline for 24 hours and examined again by microscope to confirm the negative result.
A demographic and health history questionnaire was completed for all participants. Data were entered directly into a mobile-phone based data collection application developed using the Open Data Kit (ODK) 2.0 software, and captured village-specific GPS coordinates as well.[20] Rapid test and skin snip microscopy data was also recorded in the data collection application, including any refusals to perform a test and reasons for that refusal. As rapid test results were not available for at least 20 minutes, individuals who refused the skin snip or rapid test did so prior to knowledge of their test results. Characteristics of participants were reported as proportions for dichotomous variables, and median (interquartile range) or mean (standard deviation) for continuous variables. Sero-positivity of Ov16 rapid test was evaluated using equally distributed age categories. Age was also evaluated as a confounder for continuous and dichotomous variables using linear and logistic regression, respectively. Logistic regression was used to determine associations between exposure characteristics and rapid test result, adjusted for age (Table 1). Participation rates for the two diagnostic methods were compared by McNemar test. Questionnaire data was analyzed using StataSE version 13.1.
Targeted members of the surveillance team and community members participating in surveillance activities were interviewed to provide feedback on the acceptability and usability of the rapid test. All rapid test technicians, were asked questions regarding their experience in using the tests. Community members were sampled purposively based on the diagnostic testing they participated in as well as their willingness to participate in an exit interview. A semi-structured interview guide was used to gather data on the user experience, how the test was received by participants, and how the test compared to experiences with skin snip microscopy. Interviews with community members and technicians were recorded as audio files, then transcribed and translated from local languages into French and then into English. Interview data were coded using content analysis based on key themes from the semi-structured interviews.[21] Refinements were made to the codebook in an iterative fashion during the analysis process and reviewed by two researchers who reached consensus on the findings. Interview data was analyzed using NVivo version 10.
A costing analysis was conducted to assess the costs related to the implementation of the onchocerciasis surveillance activity in Senegal by diagnostic test. The study focuses on the economic costs from the government’s perspective, therefore valuing volunteers’ time. Data was mainly gathered from secondary sources such as financial reports, consolidated budgets, and other secondary sources of financial information from PATH and from the Senegalese onchocerciasis surveillance team. A simple structured questionnaire was also used to identify resources used during surveillance activities that had been purchased in previous years. Where needed, costs were calculated using the ingredient approach, multiplying the input prices by the number of inputs used.[22,23] Key input prices for this analysis are: Ov16 rapid test ($1.20), skin snip tool ($225), and microscope ($2,490). Costs were captured for all activities conducted during the surveillance activity, including training, field work, and data reporting. Field work cost categories were further split into labor, supplies, devices and instruments, transport and lodging, and data reporting. Drugs costs were zero since none of the study participants tested positive by skin snip microscopy, which would indicate treatment according to standard of care. The identified resources used were then allocated to the rapid test, the skin snip microscopy, or to shared costs, which were costs that were incurred independent of the type of test used, such as data entry and analysis, or transport cost to the villages.
Total costs were first calculated by cost category and then aggregated across categories. The costs per test performed were calculated by dividing the total costs by the number of participants evaluated with each test. We also estimated the costs assuming the same population size (n = 1,169) for both tests by proportionally scaling up the variable costs for skin snip microscopy, while keeping the fixed costs constant. This was done because of the difference in the number of participants tested with the rapid test (n = 1,169) compared to skin snip microscopy (n = 383) and the presence of high fixed costs of devices and instruments, allowing a comparison of the costs without the volume effect. All cost estimates are presented in $US using an exchange rate of 591.45 XOF per $US (World Bank Development Indicators, World Bank). Further details on the cost analysis is available as supporting information (S1 File)
Of the 1,173 participants who agreed to participate in the study, the median age was 12 years and ranged from 5 to 92 years. The most common professions were farmer, student and housewife. Participation rates for the two diagnostic tests differed with a total of 1,169 participants (99.7%) agreeing to be tested by the rapid test and 383 participants (32.7%) agreeing to be tested by skin snip microscopy (p<0.0001). (Table 2)
The sero-positivity rate of the rapid test among those who performed the test was 2.6% (30/1,169) among all ages (age range of positives, 11–81), 0.4% (3/775) among 20 years and under, and 6.9% (27/394) among those over 20 years. The 20 year breakpoint was evaluated based on the distribution of the data. (Fig 1) All results were either positive (sero-positive) or negative, as no invalid results were detected. There were zero positive skin snip results. Age was associated with the rapid test result (p<0.001) and 13/17 exposure characteristics, though it was not associated with refusal to participate in either diagnostic test. In an age adjusted analysis, odds of having a rapid test positive result decreased the more recently IV was distributed in their village (0.83, 95% CI: 0.74–0.93). The 3 participants under 20 years who had a positive rapid test result had all lived in their villages their whole lives, frequently went to the stream near their village, and did not report experiencing any of the onchocerciasis symptoms that were included in the questionnaire. Two of the 3 individuals reported having taken IV in the last year. (Table 1)
Interviews were conducted with 4–5 community members from all 13 villages (n = 61). Over 90% of participants (57/61) reported that they valued or appreciated the rapid test. Community members liked that the test was performed quickly and was not painful, and they perceived it to provide reliable results. Community members noted that the test brought health knowledge to the community, enabled them to access follow-on care if needed, and could effectively be used to test children. Most participants (55/61) reported that they had no concerns about the rapid test while 10 percent of participants (6/61) disliked the finger stick component of the rapid test procedure. Moreover, many participants indicated that they would be more likely to participate in future surveillance activities if the rapid test was used and suggested that its use would spur broader participation within the community. The more common reasons for refusing the skin snip biopsy were that they “did not like the idea” and “thought it would be too painful”. Some community members discussed historical experiences with the biopsy procedure like it being painful and suggested that now, they are less willing to undergo the biopsy if the results are consistently negative. With regards to preferences for either diagnostic tool, 50% of respondents preferred the rapid test to skin snip microscopy, 39% expressed liking both tests, and 10% preferred the skin snip microscopy. The primary reasons were that the rapid test was less painful, quicker, and could provide individual test results. Some respondents, particularly those who were sensitized to the differences between the principles of each test, wanted to continue to use both tests or preferred the skin snip microscopy because it could provide a confirmation of infection. (Fig 2)
Community members also noted that the role of the surveillance team and the information they provided influenced their experience with the test. Nearly all of the participants indicated that the surveillance team was skilled and they trusted their abilities. For some, the skills of the surveillance team translated to credibility of the test itself and trust in the test result. The information that the surveillance team provided to community members regarding the test procedure and the test results varied among participants. Some felt like the test and their result were well-explained while others reported not learning their test results. Community members expressed a strong preference for understanding the test procedure and purpose as well as their results. Participants overwhelmingly preferred to receive their individual results though a minority of respondents also wanted to receive sero-prevalence results to understand the health status of the entire community. Community members also noted appreciation for the health services being made available in their village.
Interviews were conducted with all rapid test technicians (n = 7). All technicians commented that they liked using the rapid test and most preferred it to skin snip microscopy as they perceived it to be more reliable and quicker to complete. They noted that it was less painful for participants, and thus made their jobs easier as community members were more willing to participate in surveillance. When prompted, one technician reported that the disposable capillary pipette was difficult to use; no other challenges were reported. Most of the rapid test technicians trusted the results of the test, in part due to emphasis on the quality assurance program throughout the study. However, these technicians noted some of the limitations of an antibody test and one technician stated a preference for skin snip microscopy to confirm infections. All technicians indicated that they would be willing to use the rapid test in future surveillance activities.
The total costs for the onchocerciasis surveillance activities in the 13 villages was estimated at $22,272.83. Costs were allocated to rapid test, skin snip microscopy, or shared costs. Shared costs were those incurred independent of the diagnostic test used and accounted for 70% of the total study costs ($ 15,697.48). Total test-specific costs were $3,671.76 for rapid test and $2,903.59 for skin snip microscopy, though the number of tests performed with each method differed (1169 and 383, respectively). Most of the total study costs (87%) were related to the field work activities, while training costs accounted for 10% of the total costs and data entry and reporting was 3% of the total costs. Of the field work costs, the main cost driver was transport costs (57%), followed by supplies, instruments, and devices (27%). In this surveillance activity, the total cost per test performed was $3.14 for rapid test, $7.58 for skin snip microscopy, and $13.43 for shared cost, giving a total cost per person tested of $16.57 for the rapid test, $21.01 for the skin snip microscopy, and $24.15 if both diagnostic tests were performed on the same participant. If no participants had refused the skin snip microscopy, so the same number of participants were tested with both diagnostic tests, the cost per person tested by skin snip microscopy would have decreased to $2.91. Adding in shared costs, the total cost per participant in the surveillance activity would have been $16.57 for the rapid test and $16.34 for skin snip microscopy, a difference of $ 0.23 per method. (Fig 3)
This study demonstrates that the inclusion of the rapid test in surveillance activities is feasible based on acceptability, usability, and costs. The sero-positivity rate of the rapid test among those who performed the test was 2.6% among all ages with no positives detected under 10 years of age, and no invalid results. The 3 individuals under 20 years who had positive rapid test results may be positive due to exposure outside their community, or from residual transmission occurring over 10 years ago, however the possibility that these are false positives cannot be ruled out. While early prototypes of the Ov16 rapid test demonstrated a 97–98% specificity[18,24], the product insert states the performance of the commercially available Ov16 rapid test compared to skin snip microscopy in a laboratory setting to be 81.1% (95% CI: 70.7–88.4%) sensitive and 99.0% (95% CI: 94.8–99.8%) specific using whole blood, or 85.3% (95% CI: 75.6–91.6%) sensitive and 99.0% (95% CI: 94.7–99.8%) specific using serum and plasma (http://www.standardia.com/en/home/product/Rapid_Diagnostic_Test/Anti-Onchocerciasis_IgG4.html). Evaluation of the performance of the tool in the field is ongoing. Currently more data evaluating the sensitivity and specificity of the Ov16 rapid test to skin snip microscopy and ELISA in field settings is needed.
The difference in participation rates for Ov16 rapid test and skin snip microscopy suggests a greater willingness in these communities to undergo a rapid test with a finger prick compared to a more invasive skin snip procedure (99.7% and 32.7% respectively, p<0.0001). Some individuals who initially refused the skin snip biopsy later changed their mind and had the skin snip biopsy performed after learning their rapid test was positive. This may have resulted in an increased participation rate for skin snip than would have been seen otherwise, though the difference was likely small as there were relatively few positives. The 2016 WHO guidelines call out a need to further investigate the acceptability of skin snip microscopy in low prevalence settings. These findings align with others that observed high refusal rates for skin snip microscopy in similar settings.[9,13] The refusal of the skin snip biopsy in our study was largely due to not liking the idea of the test and considering the test to be too painful. Moreover, some community members suggested that they are less willing to undergo the biopsy if the results are consistently negative. The value of the different diagnostic tests during the distinct phases of elimination is important, and community members and rapid test technicians noted that skin snip microscopy remains the primary method for assessing infection status and recommending treatment, while the rapid test is a screening tool to inform decisions regarding the continuation or culmination of MDA.
Community members reported high levels of acceptability and willingness to participate in surveillance activities that included the rapid test. The role of the surveillance team and the information they provided influenced community members’ experience with the test. Clear communication about the test purpose, procedure, and result was appreciated by community members and increased their trust in the result and motivation to participate in other onchocerciasis control activities. The influence of the surveillance team should not be overlooked as they may be a valuable tool to encourage participation in future surveillance activities and greater compliance with mass drug administration. Community member feedback also showed that in areas endemic for onchocerciases and where consistent access to quality health services may be lacking, there is an appreciation for the delivery of health services through NTD control programs. Community members expressed a desire for greater knowledge of the health of their community, potential risk factors, and their achieved progress towards program goals.
The Ov16 rapid test is intended for use in populations nearing elimination. In this setting, the population is predominately healthy and unaffected by onchocerciasis. Attributes such as invasiveness of the test may be more important in these settings, particularly when testing is focused on children.
The costing analysis showed that the cost per person tested in this activity was $16.57 for the rapid test, $21.01 for the skin snip microscopy, and $24.15 for both methods. If no participants had refused the skin snip microscopy, the costs per participant using either method would have been comparable at around $16. The labor and instrument costs for skin snip microscopy were largely fixed and independent of the number of participants tested. The skin snip microscopy team had to remain with the study team for the duration of the surveillance activity regardless of how many people they were testing. Multiple skin snip microscopy instruments were used for this activity due to the need to sterilize equipment after each use, and these instruments were assumed to not be shared with other programs. Additionally, the rapid test had slightly lower training costs due to shorter training, but higher costs for devices and instruments that were dependent on the number of people tested. Performing skin snip microscopy on a subset of participants who are all receiving rapid testing is costly, due to the high costs of instruments and the need for two teams of technicians (rapid test technicians and skin snip technicians). This costing information may be useful when considering how to transition programs from skin snip to rapid testing. However, while comparable conclusions may arise from similar studies, the cost estimates from this study are specific to the Senegalese context. For example, costing results would vary based on country salary and per diem policies, the surveillance activity approach used such as number of days spent in the field and number of people tested per village. Similarly, different assumptions regarding the useful life of instruments for skin snipping would also affect the results. Additional implementation research is important in other locations to evaluate how results vary by setting.
A comprehensive quality assurance (QA) program including training videos and materials, quality assurance panels and daily quality control standards, was implemented along with the Ov16 rapid test to facilitate proper use of the test in this study. The QA program resources, which are freely available to implementing programs, minimize improper handling of the tests and user errors, while ensuring consistent product quality. Findings from this study also suggest this QA program may provide surveillance teams with greater confidence in technician skills and validity of the results, which benefits the community members by supporting the surveillance team in their dissemination of information and results. More research is needed to understand the role QA practices have on influencing user and participant confidence, identify QA best practices, and drive adoption of these practices through integration into global guidelines.[25]
A process map illustrating the use of one or both tools in surveillance activities was generated from the 2015 workshop, this study, and the 2016 WHO guidelines. (Fig 4) The “current practice” uses skin snip microscopy only, a more relevant strategy in higher prevalence areas. The “parallel method” uses skin snip microscopy and Ov16 rapid test, which may be more appropriate when programs are transitioning to stopping MDA and implementing Ov16 serology such as in this study. However, in these transition areas, high refusal rates for skin snip may prevent programs from attaining a sufficient sample size to determine with certainty if program goals have been reached.[13] An alternative to the parallel method may be testing only Ov16 rapid test positives with skin snip microscopy, though in low prevalence settings few if any skin snip positives would likely be detected. The “final method” uses the rapid test only and “should be used in children under 10 years to demonstrate interruption of transmission”.[13] This study was not designed with a sampling methodology sufficient to determine prevalence or if transmission had been broken. According to guidelines, roughly 2000 children under 10 years of age would need to be tested to detect a prevalence of less than 0.1% with sufficient confidence, and only 368 children under 10 years were included in this study, all of whom were rapid test negative.[13]
Finally, diagnostic tools play an important role in influencing health outcomes, usually through the intended benefits of enabling timely diagnosis, accurate disease surveillance, and proper treatment. These tools may also have the ability to influence individual and community behaviors, such as participation in surveillance activities and confidence in control program activities. Taking a broader perspective, the true value of diagnostic tools may go beyond the intended utility to include extended benefits such as increased utilization of health care services, individual agency over the health care experience, and confidence in provider abilities. These broader benefits should be identified and measured in future implementation research to better understand how to move technologies beyond innovation and validation, and into adoption and scale-up.[19] Ov16 as a biomarker has successfully moved from discovery and development at the bench, to evidence of effectiveness in the field. The remaining barriers are optimized and context-specific integration into systems and programs. As global focus shifts to the integration of onchocerciasis and lymphatic filariasis (LF) programs to reach elimination in Africa faster, a rapid assessment of Ov transmission through LF transmission assessment surveys (TAS) will be required.[26] A more appropriate tool for this work may be the SD BIOLINE Oncho/LF IgG4 biplex rapid test to detect ongoing onchocerciasis and LF transmission simultaneously (http://www.standardia.com/en/home/product/Rapid_Diagnostic_Test/Oncho-LF_IgG4_biplex.html). [24] As more settings achieve success in control of either disease, integrated surveillance for transmission interruption may be a best-buy, and implementation research to support successful adoption and scale up is essential.
In this area of Senegal with low onchocerciasis sero-positivity, there was high participation with the rapid test, while participation with skin snip microscopy was significantly lower. Acceptability and perceived value of the rapid test was high among community members and rapid test technicians. The role of the surveillance team and the information they provided influenced community members’ trust in the result and motivation to participate. This may be a valuable tool to encourage participation in future surveillance activities and greater compliance with mass drug administration. This study provides evidence of the feasibility of implementing the Ov16 rapid test and the associate costs, which may be informative to other country programs interested in adopting this new tool as they move from control to elimination of onchocerciasis.
|
10.1371/journal.ppat.1002674 | A Single Polar Residue and Distinct Membrane Topologies Impact the Function of the Infectious Bronchitis Coronavirus E Protein | The coronavirus E protein is a small membrane protein with a single predicted hydrophobic domain (HD), and has a poorly defined role in infection. The E protein is thought to promote virion assembly, which occurs in the Golgi region of infected cells. It has also been implicated in the release of infectious particles after budding. The E protein has ion channel activity in vitro, although a role for channel activity in infection has not been established. Furthermore, the membrane topology of the E protein is of considerable debate, and the protein may adopt more than one topology during infection. We previously showed that the HD of the infectious bronchitis virus (IBV) E protein is required for the efficient release of infectious virus, an activity that correlated with disruption of the secretory pathway. Here we report that a single residue within the hydrophobic domain, Thr16, is required for secretory pathway disruption. Substitutions of other residues for Thr16 were not tolerated. Mutations of Thr16 did not impact virus assembly as judged by virus-like particle production, suggesting that alteration of secretory pathway and assembly are independent activities. We also examined how the membrane topology of IBV E affected its function by generating mutant versions that adopted either a transmembrane or membrane hairpin topology. We found that a transmembrane topology was required for disrupting the secretory pathway, but was less efficient for virus-like particle production. The hairpin version of E was unable to disrupt the secretory pathway or produce particles. The findings reported here identify properties of the E protein that are important for its function, and provide insight into how the E protein may perform multiple roles during infection.
| Coronaviruses are enveloped viruses that bud and assemble intracellularly, and therefore must use the host secretory pathway for release. Coronavirus E is a small protein that contains a single predicted hydrophobic domain and is targeted to the Golgi region. The E protein has been implicated in the assembly of coronavirus particles, as well as in virus release after assembly. The mechanism of action is not understood, but may involve ion channel activity. The membrane topology of the E protein is also unclear, and the protein may adopt distinct topologies that have different functions. We previously showed that the E protein from the infectious bronchitis virus could disrupt the secretory pathway to the apparent advantage of the virus. Here we have mapped this activity to a single, essential residue within the hydrophobic domain. Additionally, we developed mutant versions of IBV E that adopt a single membrane topology, and showed that a transmembrane topology is required for disruption of the secretory pathway. Our results broaden the understanding of E protein function and will impact the development of antiviral strategies.
| Coronaviruses (CoVs) are enveloped, positive strand RNA viruses that infect a variety of mammalian and avian species. In humans, CoVs are responsible for nearly 20% of common cold cases. CoVs can also lead to more serious disease as seen during the outbreak of the severe acute respiratory syndrome coronavirus (SARS-CoV) in 2003. To better prepare for the emergence of another highly pathogenic CoV it is important to increase our understanding of CoV biology.
The CoV virion consists of a helical nucleocapsid, made up of the CoV N protein and the genome, surrounded by a lipid envelope. Three structural proteins are embedded in the virion envelope. The CoV S protein is a type I transmembrane protein and is responsible for the attachment and fusion of the virion during entry. The CoV M protein has three transmembrane domains and drives the organization of the virion through its interactions with the other structural proteins [1]. The CoV E protein is small (76–108aa), is predicted to contain a single hydrophobic domain (HD), and is a minor component of the virion envelope. CoV E and CoV M drive the assembly of the virion [2]. CoV assembly occurs intracellularly at the endoplasmic reticulum-Golgi intermediate compartment (ERGIC) [3]. This results in fully assembled infectious particles within the lumen of the Golgi complex and downstream secretory organelles. Thus, virions must use the host secretory pathway in order to reach the plasma membrane and be released from infected cells.
In addition to its role in assembly, CoV E may have other functions during infection. Studies in planar lipid bilayers have shown that CoV E has ion channel activity [4], [5]. These studies also showed that the small molecule hexamethylene amiloride (HMA) inhibits the ion channel activity of mouse hepatitis virus (MHV) E and human coronavirus 229E (HCoV 229E) E. While there is no direct evidence that CoV E acts as an ion channel during infection, addition of HMA to either MHV or HCoV 229E infected cells inhibits viral replication, and mutations introduced into the HD of MHV E impair virus production suggesting that the putative ion channel activity may play a role during infection [5], [6]. If CoV E acts as an ion channel, it must form higher order structures because it contains only one predicted transmembrane domain. Indeed, structural and computational studies have suggested that CoV E forms a homo-pentamer in the membrane with a pore in the middle [7]–[9]. Understanding the role of a pentameric E ion channel is an important question in the field.
The membrane topology of CoV E is of considerable debate. CoV E has a short (∼10aa) hydrophilic N-terminus followed by a long hydrophobic domain (∼25aa) and a hydrophilic C-terminus. The N-terminus does not contain a canonical ER signal sequence [10]. The hydrophobic domain is unusually long for a protein targeted to the ERGIC/Golgi complex, but does not appear to be long enough to span the lipid bilayer twice [11]. These properties make it difficult to predict the topology based on the primary sequence. Complicating matters is the fact that multiple topologies have been reported in the literature for different CoV E proteins. Both IBV E and SARS-CoV E have been reported to exist as a type III transmembrane protein (Nexo, Ccyto) [12], [13]. Other investigators have reported the opposite topology for SARS-CoV E and transmissible gastroenteritis virus (TGEV) E (Ncyto, Cexo) [14], [15]. Yet another topology reported for CoV E is a membrane hairpin, where the hydrophobic domain bends into the cytoplasmic leaflet of the membrane with the N- and C-termini in the cytoplasm. The hairpin topology has been reported for MHV E and SARS-CoV E [15]–[17]. These discrepancies suggest that CoV E may adopt more than one membrane topology. If this is the case, CoV E may perform distinct functions depending on how it is inserted into the membrane. For example, a transmembrane version of CoV E could oligomerize and act as an ion channel, whereas a membrane hairpin could drive virion budding.
Since CoVs assemble intracellularly, their virions must pass through the host secretory pathway for egress. How or if the secretory pathway is modified in infected cells is not well understood, but may involve the E protein. A version of TGEV lacking the E protein was unable to produce infectious particles, but electron microscopy revealed that immature virions were present in secretory organelles of infected cells [18]. Alanine insertion scanning mutagenesis of the HD of MHV E produced mutant viruses that showed a defect in the release of infectious particles [6]. These results demonstrate that CoV E is important for virion trafficking, but did not identify the mechanism. It has long been appreciated that CoV infection drives a rearrangement of host cell membranes including the Golgi complex [19]. More recently it was shown that during CoV infection virions appear in large virion-containing vacuoles derived from Golgi/ERGIC membranes [20]. Recently we showed that the E protein of IBV promotes the release of infectious particles. We also observed that expression of IBV E results in the disruption of anterograde protein traffic and causes the Golgi complex to disassemble, and that all of these effects were dependent on the HD of IBV E [21]. This finding linked the efficient release of particles to the alteration of the host secretory pathway, and demonstrated that IBV E has a role during infection beyond assembly.
In the present study we set out to determine what properties of the HD of IBV E were important for disrupting the secretory pathway. We performed alanine scanning mutagenesis on the HD and identified a key residue required for disrupting the secretory pathway. We also addressed the role of topology in disrupting the secretory pathway by designing mutant versions of IBV E that adopted either a transmembrane or a membrane hairpin topology. This allowed, for the first time, functional analysis of the two specific forms.
When the HD of IBV E (GenBank ID: CAC39117) is modeled as an alpha helix and viewed in a helical wheel projection, polar uncharged amino acids cluster on one side (Figure 1A). If the cluster of polar uncharged residues is important for the disruption of the secretory pathway, mutating them to alanine should inhibit their effect, while mutations on the opposite side of the helix should have no effect. To test this hypothesis, single alanine mutations of the polar uncharged residues as well as residues on the opposite side of the helix were made. The mutant proteins were transiently expressed in HeLa cells and their expression was confirmed by immunoblot (Figure 1B). Next, we determined whether the mutants disrupted protein trafficking. The mutant proteins were expressed along with the model cargo protein vesicular stomatitis virus glycoprotein (VSV G). Trafficking of VSV G was measured using metabolic labeling in a pulse-chase assay coupled with endoglycosidase H (endo H) digestion. Since glycoproteins become resistant to digestion with endo H in the medial-Golgi, this assay monitors the rate at which a glycoprotein moves through the Golgi complex. All of the alanine mutants disrupted trafficking with the exception of IBV E T16A (Figure 1 C and D). Thus, a single polar uncharged residue within the HD of IBV E is necessary for disrupting protein trafficking.
In addition to disrupting protein trafficking, IBV E expression disrupts Golgi morphology [21]. As with the trafficking defect, the disruption of the Golgi complex is dependent on the HD of IBV E. We reasoned that if the trafficking defect and Golgi complex disruption were occurring by the same process, T16 would be necessary for both effects. Indirect immunofluorescence microscopy was performed on cells transiently expressing IBV E or the mutant proteins. Cells were stained for IBV E and GM130, a marker of the Golgi complex. All of the mutant proteins disrupted the Golgi complex like IBV E with the exception of T16A, which had no effect on Golgi complex morphology (Figure 2A and B). These results, along with the data shown in Figure 1, demonstrate that a single polar uncharged residue within the HD of IBV E (T16) is necessary for the disruption of the secretory pathway.
We next determined if there was any flexibility in the amino acid required at position 16 in IBV E. A multiple sequence alignment of several different CoV E proteins showed that a polar uncharged residue is conserved at position 16 (Figure 3A). We introduced mutations at position 16 in IBV E that replaced the threonine with serine, asparagine or glutamine. The mutant proteins were transiently expressed along with VSV G to determine their effect on protein trafficking. None of the proteins disrupted trafficking of VSV G, showing that these residues could not substitute for threonine (Figure 3B). We examined the morphology of the Golgi complex in cells expressing the mutant E proteins using indirect immunofluorescence microscopy. Corroborating the trafficking results, none of the conserved mutations disrupted Golgi complex morphology as judged by GM130 staining (Figure 3C). Thus, there is a strict requirement for threonine at position 16 in IBV E.
Previously we reported that replacing the sequence of the HD of IBV E does not affect virus-like particle (VLP) production [22]. However, since these earlier experiments were carried out using a different cell type and expression system, we wanted to confirm that mutating T16 did not impair VLP production. We co-expressed IBV E and the T16 mutants along with plasmids encoding IBV M and IBV N in HeLa cells. The supernatant and cells were collected separately, and VLPs were purified from the supernatant via centrifugation over a sucrose cushion. The level of VLPs produced was measured by immunoblotting and comparing the signal for M in the VLP fraction to the cell fraction. We found that none of the mutations had a significant impact on steady-state VLP production as judged by the amount of M released (Figure 4A and B). Thus, T16 is required for altering the secretory pathway, but is not required for VLP production.
CoVs fall into three distinct groups based on genome similarities, alpha, beta and gamma. IBV is a gamma-CoV, and we wanted to determine if the effect on the secretory pathway was a property of other CoV E proteins. We transiently expressed the E proteins from the beta-CoVs SARS-CoV (GenBank ID: NP_828854.1) and MHV (GenBank ID: ACO72886) as well as the alpha-CoV TGEV (GenBank ID: ABG89321) in HeLa cells. Using antibodies directed against the various CoV E proteins or GM130, the morphology of the Golgi complex was examined. Somewhat surprisingly, none of the other CoV E proteins caused the Golgi complex to disassemble (Figure S1). Other markers for the Golgi complex were also distributed normally (data not shown). We determined if any of the other CoV E proteins impacted protein traffic through the Golgi complex using the pulse-chase endo H assay described above. VSV G trafficking was unaffected by expression of any of these E proteins (data not shown). Taken together these data indicate that the effect of SARS-CoV E, MHV E, and TGEV E on the host secretory pathway may be different than that of IBV E, and potentially point to an important difference in the function of the proteins. However, we found that the half-life of IBV E was longer (3.6 h) than that of MHV E (2 h), SARS-CoV E (2.1 h), or TGEV E (2.6 h) (data not shown). Additionally, we could not compare the absolute expression level of each protein (since the antibodies to detect each one are different). Thus, it is possible that MHV E, SARS-CoV E and TGEV E do not accumulate to as high a level as IBV E in this expression system, and therefore do not demonstrate the disruption in the secretory pathway observed for IBV E.
Multiple groups have proposed different membrane topologies for the CoV E protein, either as a transmembrane protein or as a membrane hairpin (Figure 5B, cartoons) [10], [12]–[16]. It is possible that CoV E may adopt multiple membrane topologies, each with distinct function(s). To test the role of topology in IBV E function, mutant versions of IBV E were created with either a transmembrane or membrane hairpin topology. To promote a transmembrane topology we added a canonical cleavable N-terminal signal sequence onto the N-terminus of IBV E (ssIBV E), which will force the cleaved N-terminus into the ER lumen [23], [24]. To produce a potential membrane hairpin we added a FLAG tag onto the N-terminus (FLAG-IBVE). The rationale for this was that other N-terminally FLAG tagged CoV E proteins adopt a membrane hairpin topology [15], [16]. We transiently expressed IBV E, ssIBV E, and FLAG-IBV E in HeLa cells and probed their membrane topology using selective permeabilization of the plasma membrane with digitonin, followed by indirect immunofluorescence microscopy. As a control we co-expressed a luminal ER protein (CFP-KDEL) along with a protein present on the cytoplasmic side of the Golgi complex (golgin160-Myc). As expected, the cytoplasmic epitope of golgin160-Myc was accessible after either Triton X-100 or digitonin permeabilization, whereas the luminal epitope of CFP-KDEL was not accessible when cells were permeabilized with digitonin (Figure 5A). For IBV E and ssIBV E we stained for either the N- or C-terminus using antibodies directed to either end of the protein. We found that both IBV E and ssIBV E largely existed as transmembrane proteins with the N-terminus in the lumen and the C-terminus in the cytoplasm (Figure 5B). For FLAG-IBV E we used a similar approach but stained for the N-terminus with an anti-FLAG antibody because our anti-IBV E N-terminal antibody was unable to recognize the modified N-terminus. The results showed that FLAG-IBV E had both the N- and C-termini in the cytoplasm (Figure 5B). We also found that the mutations did not affect the targeting of either construct, as both colocalized with Golgi complex markers (Figure 6A and data not shown). We quantified the difference in the staining intensity under the different permeabilization conditions by measuring the fluorescence signal for the N- and C-termini in the same cell. After subtracting the background signal, the N∶C ratio was calculated, and normalized to the ratio from the Triton X-100 samples for ease of comparison (Figure 5C). As expected the ratio dropped dramatically for both IBV E and ssIBV E, but not for FLAG-IBV E. Thus, mutant IBV E proteins predominantly adopt either a transmembrane or membrane hairpin topology. It is worth noting that the N∶C ratio was lower for ssIBV E than for IBV E. While this difference was not statistically significant, we speculate that there may be a small population of IBV E that is inserted as a membrane hairpin.
Having developed versions of IBV E that adopt a unique orientation in the membrane, we determined if topology was important for disrupting the Golgi complex. IBV E, ssIBV E, and FLAG-IBV E were transiently expressed in HeLa cells and subjected to indirect immunofluorescence microscopy. Staining for IBV E and GM130 revealed that ssIBV E disrupted Golgi complex morphology to a similar degree as IBV E (Figure 6A and B). However, FLAG-IBV E had no effect on Golgi complex morphology (Figure 6A and B). This result suggests that the transmembrane topology is necessary for inducing Golgi complex disassembly. Since IBV E with mutations at T16 did not disrupt the secretory pathway, it was important to confirm that these mutations did not disrupt topology. Indeed, the selective permeabilization assay demonstrated that the topology of IBV E-T16A was identical to IBV E (Figure S2).
Next we tested whether expression of the topology constructs affected protein trafficking. We found that ssIBV E disrupted protein trafficking similar to IBV E (Figure 7A and B). FLAG-IBV E did not disrupt trafficking to the same extent as ssIBV E or IBV E but still had some effect (Figure 7A and B). This could indicate the IBV E hairpin does have some effect on trafficking, albeit to a smaller degree. It is also possible that a portion of FLAG-IBV E is inserted in a transmembrane topology and this small pool of protein is sufficient to alter trafficking, but insufficient to disrupt the morphology of the Golgi complex.
The addition of a FLAG tag to the N-terminus of IBV E could result in a number of effects beyond changing the topology of IBV E. Thus, we generated a version of IBV E that had a canonical signal sequence, followed by a FLAG tag on the N-terminus (ssFLAG-IBV E). When transiently expressed in HeLa cells, ssFLAG-IBV E was not targeted as well to the Golgi complex as the other constructs (Figure S3A). Also, selective permeabilization showed that a larger portion of the N-terminus was in the cytoplasm compared to ssIBV E (Figure S3B). These observations suggest that the FLAG tag may alter insertion and targeting when added behind a cleaved signal sequence. However, even with these caveats, ssFLAG-IBV E still disrupted trafficking similarly to IBV E and ssIBV E (Figure S3D). ssFLAG-IBV E also disrupted the Golgi complex, but to a lower degree than IBV E or ssIBV E, possibly due to less efficient targeting (Figure S3C). These results strongly support our interpretation of the importance of topology and IBV E function.
To test how the membrane topology of IBV E affects particle assembly, we assayed IBV E, ssIBV E and FLAG-IBV E in a VLP assay. We co-expressed the E constructs along with plasmids encoding IBV M and IBV N in HeLa cells and determined the amount of VLPs released into the supernatant by immunoblotting (Figure 8A and B). Cells expressing ssIBV E produced less VLPs than those expressing wild-type IBV E, suggesting that the transmembrane topology can at least partially drive assembly, possibly by inducing membrane curvature in a lattice of IBV M. In support of this result, ssFLAG-IBV E also produced reduced levels of VLPs (Figure S3E). Cells expressing FLAG-IBV E produced almost no VLPs, indicating that the hairpin topology alone may not support the production of particles. This result is harder to interpret. It is not clear if the membrane hairpin is unable to drive assembly, release, or both.
We reported previously that replacing the entire HD of IBV E with a heterologous sequence eliminated disruption of the secretory pathway in transfected cells, and dramatically reduced the release of infectious virus from infected cells [21]. Total particle release was only modestly affected, however, suggesting that the HD of IBV E is important for preventing damage to virions during egress. Here we have shown that a single amino acid in the HD of IBV E (T16) is critical for disruption of the secretory pathway in cells expressing IBV E, but was not required for VLP production. This result suggests that the alteration to the secretory pathway is uncoupled from the role of E in assembly. Additionally, we generated versions of IBV E that adopted either a transmembrane or membrane hairpin topology. Using these mutants, we showed that a transmembrane topology was required for secretory pathway disruption. The residue equivalent to T16 in SARS-CoV E, N15, is predicted to lie in the pore region of a homo-pentamer [7]. Studies on a lysine-flanked peptide of the SARS-CoV E HD showed that N15 was important for the ion channel activity of the peptide in planar lipid bilayers [25]. Since we found that a transmembrane topology and T16 are required for disrupting the secretory pathway, and both are predicted to be important for ion channel activity, it is certainly possible that the disruption of the secretory pathway is due to the putative channel activity of IBV E. Alteration of Golgi complex structure and disruption of protein traffic occur when the ion balance at the Golgi complex is disrupted [26]–[29]. While an active ion channel at the Golgi complex could explain our observations, how altering the ion balance of secretory organelles might facilitate release of infectious particles remains unknown. We speculate that the demands of trafficking large virion cargo require the expansion of the Golgi complex cisternae, which may be achieved by changing the luminal ion concentration. Alternatively, a change in luminal environment may inactivate proteases present in the secretory pathway, thus protecting the virions from degradation that could render them non-infectious. The membrane rearrangements observed in CoV-infected cells are likely due at least partially to a disruption in the luminal microenvironment, although syncytia formation also contributes [19], [20]. Expression of the E protein in the absence of infection allowed us to assess its contribution to membrane rearrangements directly.
Many viruses encode small membrane proteins that have ion channel activity [30]. As a group these proteins are referred to as viroporins. The best studied viroporin is influenza M2, which forms a tetrameric pH-activated proton channel [31]. The M2 channel acidifies the interior of the virion during entry to aid in unpacking the genome [32]. For some strains of influenza virus, M2 also plays an important role in the secretory pathway where it raises the pH of the trans-Golgi to prevent the premature activation of the fusion protein [29], [33], [34]. Hepatitis C virus (HCV), like CoVs, assembles intracellularly and must navigate the secretory pathway for release. Interestingly, HCV encodes a proton selective viroporin, p7 [35]–[37]. While the exact role p7 is not fully understood, it is important in the assembly and release of HCV virions, and expression of p7 leads to the alkalinization of secretory organelles [37]–[39]. It is possible that HCV-p7 and CoV E have analogous roles during infection for altering the secretory pathway to promote the release of virions. Viroporins appear to play important roles in the assembly and trafficking of many viruses; understanding their exact role(s) is important as they represent good targets for therapeutic intervention via small molecule inhibitors.
While T16 in IBV E is required for disrupting the secretory pathway, it is not important for virus assembly as judged by VLP production. Our VLP results also suggest that disruption of the secretory pathway is not required for virus egress, since the T16A mutant produced the same level of VLPs as the wild-type E protein. However, the VLP assay does not allow measurement of infectivity, which was greatly reduced for particles released from cells infected with IBV carrying an E protein with a heterologous HD [21]. Another difference between infection and the VLP assay is that more particles are produced in a shorter time during infection, it is likely then that the stress on the secretory pathway is much more robust during infection. Thus, the VLP assay may not accurately reflect virion trafficking during infection. To measure the effect of T16 on virion trafficking, assays that measure both the amount, rate, and route of infectious particle trafficking are necessary. A future goal will be to analyze recombinant viruses carrying mutations at T16 with quantitative trafficking assays.
If CoV E is important for the release of infectious particles, why do some CoVs show only a modest reduction in infectivity when E is deleted [40]? Moreover, why do we only observe a measurable disruption in the secretory pathway with IBV E and not the E proteins from other CoVs? The answer to these questions may lie in the exact role(s) that the CoV E protein plays for each virus. While the E proteins from different CoVs share a similar domain structure, there is large variation in their primary sequence. Additionally, the requirement of CoV E for the production of infectious virus is not consistent between different CoVs. The E protein of the TGEV is essential for the production of infectious virus [18]. However, a version of MHV lacking the E gene can replicate, albeit at a greatly reduced titer [41]. Finally, a recombinant version of SARS-CoV with E deleted shows only a modest reduction in infectivity when passaged in cultured cell lines [40]. These results suggest that CoV E may have evolved to perform divergent functions in different CoVs. Somewhat surprisingly then, it was reported that the E protein from several different CoVs, including IBV E, could substitute for MHV E during infection [42]. Even more striking, when MHV ΔE was passaged, revertants were recovered with a partial duplication of the M gene (consisting of the N terminus and three transmembrane domains but lacking the C-terminal tail) that were able to largely compensate for the lack of E [43]. Taken together, these results show that at least some function(s) of the E protein are conserved among CoVs. However, the requirement for its function(s) may vary significantly due to the compensatory action of other viral proteins or differences in cell and tissue types infected. Of all the CoVs whose E proteins were tested here, IBV is the only one with an avian host. The requirements for assembly and release in avian species may be slightly different than in mammals. We tested whether the disruption of the secretory pathway caused by IBV E occurred in DF-1 chicken fibroblasts (cultured at 39°C), and found that the secretory pathway was disrupted similar to HeLa cells (unpublished data). Another potential difference is the cell type in which each virus replicates. Certainly the requirements for virus egress in different tissues could be an important factor. Another possibility is that the compartmental localization of the E proteins may vary in the absence of the other viral proteins and the impact of each CoV E on the secretory pathway could depend on the Golgi subcompartment in which it is localized. This possibility could be addressed by immunoelectron microscopy on cells expressing the various E proteins. There is a notable difference in the ion specificity and channel behavior among the different E protein channels in planar lipid bilayers [4], [5]. Unlike the other CoV E channels characterized, the IBV E channel demonstrated rectification, where ions are moved predominately in one direction [5]. Additionally, the IBV E channel is insensitive to the small molecule HMA, unlike the other CoV E proteins tested [5], [9]. If the ion specificity or activity varies between the CoV E proteins, it could certainly explain the differences in behavior reported here. The best way to study these differences would require electrophysiological measurements using patch clamp analysis on purified Golgi membranes. This approach would allow the direct measurement of the CoV E protein in its natural membrane with the proper post-translational modifications, but will be very technically challenging. One last point is that the sequences of the CoV E proteins are highly variable. Of note, IBV E is significantly larger and contains more polar residues in its HD than the other CoV E proteins (see Figure 3A). It will be important to determine how these differences relate to the function of the proteins. This could be addressed by determining how chimeric proteins affect the secretory pathway and virus replication.
Previous reports on CoV E protein topology have suggested that it may exist either as a transmembrane protein or as a membrane hairpin with both the N- and C-termini in the cytoplasm. The ability to adopt multiple membrane topologies could be a mechanism to increase the number of protein functions within the constrictions of genome size. Here, we generated mutant versions of IBV E that adopted either a membrane hairpin or transmembrane topology. We found that the transmembrane version of the protein behaved largely like IBV E, with the exception that it was unable to drive VLP production to the same degree. The membrane hairpin version of IBV E was unable to disrupt the secretory pathway or drive VLP production. These data suggest that IBV E largely functions as a transmembrane protein, with no apparent role for the membrane hairpin. However, such conclusions should be drawn with caution. While we determined that ssFLAG-IBV E behaved largely like ssIBV E, addition of the FLAG tag onto the N-terminus of IBV E could have any number of off-target effects, especially when considering the interaction of the E protein with M. We attempted to generate a membrane hairpin using several different strategies, including altering the charge distribution on either end of the HD, extending the N terminus with different tags, and shortening the C-terminus. Our only successful strategy was adding the FLAG tag onto the N-terminus. It should be noted that all reports of CoV E demonstrating that it adopts a membrane hairpin upon expression have been carried out using N-terminally tagged proteins [15], [16]. In fact the most recent data on the topology of SARS-CoV E using the untagged protein and antibodies directed to either terminus show that the predominant topology is Nexo, Ccyto [13]. What remains unclear is if a membrane hairpin plays a role during infection. It is possible that a portion of the E protein adopts a membrane hairpin topology. We did observe a small difference between ssIBV E and IBV E when we quantified the signal from our selective permeabilization experiment. A small amount of CoV E in the membrane hairpin conformation could play a catalytic role during assembly, and while not necessarily required for assembly, it may increase the efficiency of assembly. This would explain why FLAG-IBV E could not support VLP production on its own. This idea could be addressed by developing infectious clones of IBV carrying the topology mutants of IBV E and examining particle production biochemically and by electron microscopy of infected cells. Also of interest is the mechanism for generation of multiple topologies. A transmembrane topology is likely generated through the canonical signal recognition particle pathway like other type III membrane proteins [23], but the generation of a hairpin could involve a different mechanism. One could speculate that a hairpin could be generated through post-translational insertion, possibly directly into the target membrane [44].
The IBV E protein is a multifunctional viral protein that plays a role in both the assembly and release of infectious virus. The exact mechanism by which the protein alters the secretory pathway to facilitate infectious particle release is still unknown, but may depend on a single amino acid in the HD. Identification of the mechanism will be a big step in understanding the interplay between the secretory pathway and CoV trafficking. Also of interest is how E protein function varies among CoVs and what underlies any difference(s). Understanding these questions will provide insight into both therapeutic approaches to CoV infection and increase our understanding of how CoVs use the host secretory pathway to their advantage.
HeLa cells were cultured in Dulbecco's Modified Eagle Medium (DMEM) (Invitrogen) with 10% Fetal Bovine Serum (FBS) (Atlanta Biologicals), and 0.1 mg/ml Normocin (InvivoGen) at 37°C under 5% CO2. Transient transfection of HeLa cells was performed using Fugene6 or XtremeGene 9 (Roche) according to the manufacturer's protocol. Experiments were performed 18–22 hours post transfection unless noted otherwise.
The expression plasmids for IBV E, VSV G and IBV M have previously been described [12], [21], [45]. The sequence for IBV N was amplified by RT-PCR of RNA from IBV infected cells. The sequence was inserted into pcDNA3.1 using BamHI and EcoRI sites, and subcloned into pCAGGS using KpnI and XhoI. Mutations of the HD of IBV E were introduced via Quikchange (Stratagene) site directed mutagenesis. ssIBV E was generated by inserting a BglII site directly upstream of the start codon of IBV E using Quikchange mutagenesis. The vector was digested with EcoRI and BglII and synthetic oligonucleotides encoding the signal sequence of VSV G (MKCLLYLAFLFIGVNCRS) with flanking EcoRI and BlgII sites was ligated upstream of the start of IBV E to generate ssIBV E. The FLAG-IBV E construct was made in a similar way. A sequence encoding an initiation codon and the FLAG epitope (MDYKDDDDK) with flanking BglII sites was ligated directly upstream of the start codon of IBV E. ssFLAG E was generated by ligating the same FLAG epitope (MDYKDDDDK) into the ssIBV E construct after digestion with BglII. pCAGGS SARS E (Urbani) has been previously described [46]. Plasmids containing the coding sequences for MHV E (A59) and TGEV E (Purdue p115) were kindly provided by Paul Masters (Wadsworth Center, Albany, NY). The coding sequence of MHV E and TGEV E were PCR amplified and inserted into pCAGGS using EcoRI and KpnI or EcoRI and XhoI respectively. The CFP-KDEL expression vector was from clontech. The construct consists of a signal sequence followed by the cyan fluorescent protein and a KDEL ER retrieval sequence in the C-terminal tail of the protein. Golgin160-myc has been previously described [47].
The following antibodies have been previously described: Rabbit and rat antibodies recognizing the C termini of IBV E, rabbit antibody recognizing the N terminal portion of IBV E [12], rabbit anti-IBV M used for immunoblotting [48], and the rabbit anti-VSV polyclonal antibody used for immunoprecipitation [49]. The rabbit anti-MHV E and rabbit anti-TGEV E used for immunofluorescence were kind gifts from Paul Masters, and have been previously described [42]. The rabbit anti-IBV N antibody was a kind gift from Ellen Collisson and has been previously described [50]. Mouse anti-GM130 was from BD Biosciences, rabbit anti-GFP was from Molecular Probes, mouse anti-FLAG M2 was from Sigma, and the monoclonal mouse anti-Myc antibody (clone 9E10) was from Roche Molecular Biochemicals. The Alexa Fluor 488 conjugated donkey anti-rabbit IgG, Alexa Fluor 488 conjugated donkey anti-mouse IgG, Alexa Fluor 568 conjugated donkey anti-rabbit IgG and Alexa Fluor 568 conjugated anti-mouse IgG were from Molecular Probes. The Texas Red conjugated donkey anti-rat was from Jackson ImmunoResearch Laboratories. The horseradish peroxidase conjugated donkey anti-rabbit antibody was from Amersham.
Multiple sequence alignment of CoV E proteins was carried out using ClustalW2 at the European Bioinformatics Institutes server [51]. The figure was generated using jalview version 2 [52]. GenBank accession numbers of the sequences used in the alignment are as follows: TGEV E (ABG89321), IBV E (CAC39117), SARS E (NP_828854.1), MHV E (ACO72886), FIPV E (AAY16378), HCoV HKU1 (YP_173240), PEDV (NP_598312), PHEV (YP_459955), Bovine CoV (NP_150081), HCoV OC43 (NP_937952), and HCoV 229E (NP_073554).
HeLa cells were transfected with pCAGGS VSV G (1 µg) along with either a control plasmid (0.5 µg pCAGGS IBV M) or a pCAGGS E construct (0.5 µg). Cells were incubated in cysteine-methionine free DMEM for 15 min, labeled with 50 µCi of Expre35S35S [35S]-methionine-cysteine (Perkin Elmer) in cysteine-methionine free DMEM for 20 min, and chased in normal growth medium. Prior to collection, labeled cells were washed with PBS. Samples were lysed in detergent solution with protease inhibitor cocktail and clarified at 20,000×g. SDS was added to 0.2% and the samples were pre-cleared with Staphylococcus aureus Pansorbin cells. Rabbit anti-VSV antibody was added to each sample and incubated for 20′. Immune complexes were collected with 20 µl of washed Staphylococcus aureus Pansorbin cells and washed two times in RIPA buffer (10 mM Tris [pH 7.4], 0.1% SDS, 1% deoxycholic acid, 1% NP40, 150 mM NaCl). Immune complexes were eluted in 1% SDS [pH 6.8] at 100°C and digested in 75 mM Na-citrate [pH 5.5] with 0.2 µl endo H (100 units) (New England Biolabs) at 37°C overnight. Concentrated sample buffer (200 mM Tris-HCl [pH 6.8], 8% SDS, 60% glycerol, 0.2% bromophenol blue) was added to each sample prior to separation on 10% SDS-PAGE. Labeled proteins were visualized by using a Molecular Imager FX phosphorimager (Bio-Rad) and quantified using Quantity One software (BioRad).
HeLa cells plated on glass coverslips were processed for immunofluorescence 18–22 h after transfection. For assaying Golgi disruption in cells expressing IBV E or the HD mutants cells were fixed in 3% paraformaldehyde for 10 min. The fixative was quenched with PBS containing 10 mM glycine (PBS/Gly). The cells were permeabilized in 0.5% TX-100 for 3 min and washed in PBS/Gly. Cells were stained with rabbit anti-IBV E (1∶1000) and mouse anti-GM130 (1∶1000). Secondary antibodies were Alexa Fluor 488 conjugated anti-rabbit IgG (1∶1000), Alexa Fluor 568 conjugated anti-mouse IgG (1∶1000). DNA was stained prior to imaging with Hoechst 33285 (0.1 µg/ml). All images were collected using an Axioscop microscope (Zeiss) equipped for epifluorescence using an ORCA-03G charge-coupled-device camera (Hamamatsu, Japan). Data analysis was done using iVision software (BioVision Technologies) and Microsoft Excel. To determine if the Golgi complex was disrupted in cells expressing IBV E or its mutants, the staining for the Golgi complex (as judged by the GM130 staining) was outlined. The area encompassing the Golgi complex was measured for cells expressing IBV E, the HD mutants, or non-transfected cells. A normal Golgi was determined to be the average area of non-transfected cells +/−1.5 standard deviations. Cells with a staining area larger than this were scored as disrupted. The percent disrupted was calculated by dividing the number of cells scored as disrupted by the total number of cells measured.
HeLa cells were transfected with CFP-KDEL (0.2 µg) and golgin160-Myc (1 µg) for control samples or with IBV E (0.5 µg), pCAGGS-ssIBVE (1.5 µg), pCAGGS-FLAG-IBV E (0.5 µg), pCAGGS ssFLAG-IBV E (0.5 µg), pCAGGS IBV E T16A (0.5 µg). For the Triton samples the protocol listed above was followed. For selective permeabilization of the plasma membrane, cells were washed with a cold KHM (20 mM HEPES [pH 7.4], 110 mM KOOCH3, 2 mM Mg(OOCH3)2) and kept on ice. The cells were permeabilized with 75 µg/ml digitonin (EM Sciences) for 10 min. The digitonin solution was removed and the cells were rinsed twice with cold KHM. The cells were moved to room temperature and fixed with 3% paraformaldehyde for 10 min. The control cells were incubated with mouse anti-Myc (1∶2) and rabbit anti-GFP (1∶500). The C-terminus of IBV E, ssIBV E, and FLAG-IBV E were detected using a C-terminal rat anti-IBV E antibody (1∶500). The N-terminus of IBV E and ssIBV E was detected using a rabbit anti N-terminal IBV E antibody (1∶100). The N-terminus of FLAG-IBV E was detected using a mouse anti-FLAG antibody (1∶500). Secondary antibodies were The Alexa Fluor 488 conjugated donkey anti-rabbit IgG (1∶1000), Alexa Fluor 568 conjugated anti-mouse IgG (1∶1000), and Texas Red conjugated donkey anti-rat (1∶500). DNA was stained prior to imaging with Hoechst 33285 (0.1 µg/ml).
For quantitation, images of equal exposure time were taken of both the Triton X-100 and digitonin samples for each antibody. To obtain the staining intensity, an initial background measurement was obtained on the C-terminal staining by drawing a region of interest (ROI) around an untransfected cell and measuring the fluorescence intensity. The exact same ROI and measurement was then made on the corresponding N-terminal image. Next, an ROI was drawn around a cell showing C-terminal staining and the mean fluorescence intensity was measured. Again, the exact same ROI was overlaid onto the corresponding N-terminal image and the mean fluorescence intensity was measured. The ratio of N- to C-terminal staining was calculated by first subtracting the background from each measurement, and then dividing the N-terminal value by the C-terminal value. For the data shown, the final ratios were normalized so that the signal ratio in the Triton X-100 samples was equal to 1.
HeLa cells were plated in 6 cm dishes and transfected with a combination of plasmids encoding IBV M (2 µg), IBV N (1.5 µg), IBV E (0.1 µg), ssIBV E (0.4 µg), FLAG-IBV E (0.2 µg), ssFLAG IBV E (0.2 µg), T16A (0.1 µg), T16S (0.1 µg), T16N (0.1 µg) and T16N (0.1 µg). Samples were prepared 42–48 hours post transfection. The medium was clarified via centrifugation at 4500×g for 20 min. The supernatant was loaded onto a 20% sucrose cushion and centrifuged at 234,000×g in a TLA-110 rotor for 60 min. The supernatant was discarded and the pellet containing the VLPs was resuspended in 1× glycoprotein denaturation buffer (New England Biolabs) containing 100-fold concentrated protease inhibitor cocktail (Sigma). To collect the cell fraction, dishes were washed with cold PBS. The cells were scraped off the dish in 1 ml PBS and pelleted at 4000×g for 2 min. The pellet was resuspended in detergent solution and insoluble material was pelleted at 20,000×g for 1 min. 10× glycoprotein denaturation buffer was added to 1×. Both the VLP and cell fractions were heated at 100°C for 1 min. Both samples were digested with PNGase F (New England Biolabs) according to the manufacturer's protocol. After digestion concentrated sample buffer was added to a final concentration of 50 mM Tris [pH 6.8], 2% SDS, 0.05% bromophenol blue, 15% glycerol. Samples were separated on 15% PAGE gels (10% of cell fraction, 100% of VLP fraction) and transferred to polyvinylidene fluoride Immobilon membranes (Millipore). Proteins were detected using rabbit anti-IBV N (1∶10,000), rabbit anti-IBV M (1∶5000) and rabbit anti-IBV E (1∶10,000) primary antibodies and horseradish peroxidase conjugated donkey anti-rabbit IgG (1∶10,000) secondary antibody. After incubation in secondary antibody, the membrane was incubated with HyGlo Quick Spray chemiluminescent detection reagent (Denville Scientific Inc.). Images were collected using a Versa Doc model 5000 (Bio-Rad) and Quantity One software.
HeLa cells were treated with 100 µg/ml cycloheximide (Sigma) diluted into culture media at 18 hours-post transfection. Cells were fixed and prepared for immunofluorescence as described above at 0, 3, and 6 hrs after cycloheximide treatment. Images were collected from each time point at the same exposure time, and the mean fluorescence intensity was determined for cells expressing the E protein. The half-lives of each E protein were calculated by plotting the signal intensity versus time on a semi-log graph.
|
10.1371/journal.pcbi.1003151 | Improving Pharmacokinetic-Pharmacodynamic Modeling to Investigate Anti-Infective Chemotherapy with Application to the Current Generation of Antimalarial Drugs | Mechanism-based pharmacokinetic-pharmacodynamic (PK/PD) modelling is the standard computational technique for simulating drug treatment of infectious diseases with the potential to enhance our understanding of drug treatment outcomes, drug deployment strategies, and dosing regimens. Standard methodologies assume only a single drug is used, it acts only in its unconverted form, and that oral drugs are instantaneously absorbed across the gut wall to their site of action. For drugs with short half-lives, this absorption period accounts for a significant period of their time in the body. Treatment of infectious diseases often uses combination therapies, so we refined and substantially extended the PK/PD methodologies to incorporate (i) time lags and drug concentration profiles resulting from absorption across the gut wall and, if required, conversion to another active form; (ii) multiple drugs within a treatment combination; (iii) differing modes of action of drugs in the combination: additive, synergistic, antagonistic; (iv) drugs converted to an active metabolite with a similar mode of action. This methodology was applied to a case study of two first-line malaria treatments based on artemisinin combination therapies (ACTs, artemether-lumefantrine and artesunate-mefloquine) where the likelihood of increased artemisinin tolerance/resistance has led to speculation on their continued long-term effectiveness. We note previous estimates of artemisinin kill rate were underestimated by a factor of seven, both the unconverted and converted form of the artemisinins kill parasites and the extended PK/PD methodology produced results consistent with field observations. The simulations predict that a potentially rapid decline in ACT effectiveness is likely to occur as artemisinin resistance spreads, emphasising the importance of containing the spread of artemisinin resistance before it results in widespread drug failure. We found that PK/PD data is generally very poorly reported in the malaria literature, severely reducing its value for subsequent re-application, and we make specific recommendations to improve this situation.
| Pharmacokinetic-pharmacodynamic (PK/PD) models of infectious diseases provide vital insights into the effectiveness of drug treatments (including the optimal dosage level, frequency and duration) by explicitly relating drug concentration after treatment to a pathogen kill rate, and ultimately the models describe whether an infection is likely to be cleared. Furthermore, they can address issues such as poor patient compliance and the spread of drug resistance that are too expensive and/or unethical to determine in the field. Despite their potential, the methodologies used in previous PK/PD models have been based upon the assumptions that only one drug is used in treatment, that the drug is immediately available in its active form at the site of action, and that the parent drug is not further converted to active metabolites. These assumptions severely limit the application of such models. We therefore extend the methodology to remove these assumptions and use this model to investigate two first-line treatments of malaria. The model accurately replicated field data and was then used to predict the impact of increasing drug tolerance and resistance on treatment outcome. We identified key PK/PD data that can, and should, be measured and reported in future field studies to maximise the predictive ability of mathematical models.
| Most human infections are currently treatable by drugs. Clinical trials remain the gold standard, empirical approach guiding drug deployment policy and practical issues such as dosing regimes. However in silico simulations based on computational predictions of drug treatment outcome have the potential to play a vital ancillary role in designing and guiding these deployment practices. Accurate simulations can rapidly investigate the consequences of putative changes in deployment practices such as changes in regimen (dosage level, frequency and duration of treatment) and can investigate and potentially quantify the threat posed by the evolution of drug resistance. The methodology used to investigate such factors in silico is mechanism-based PK/PD modelling, whose basic methodology and range of applications was recently reviewed by Czock and Keller [1]. In essence, this approach incorporates existing PK and PD parameters estimates into differential equations to calculate the decline in drug concentration after treatment, then converts this into a pathogen killing rate to find how pathogen number declines after treatment and whether the infection is eventually cleared. Note the distinction between PK/PD mechanism based modelling (the subject of this manuscript) which uses existing PK estimates to simulate drug treatment, and PK parameter estimation models (usually using non-linear analysis) which are applied to human clinical data to actually produce the PK estimates; a recurring theme of this manuscript is that the former fails to fully utilise all the data produced by the latter and we describe the computational extensions required to achieve this.
PK/PD mechanism-based modelling assumes a single drug is instantaneously present in the patient after treatment (the drug absorption and conversion processes often reported in PK estimation models of human data are ignored) and that pathogens are killed by the drug in its unaltered form [1]. In practice, drug combinations are now mandatory for the treatment of many infections, including the ‘big three’ infective killers HIV, TB and malaria so the single-drug PK/PD methodology needs to be updated to reflect these policies. Many drugs also have short half-lives so the time taken for their absorption (across the gut in the case of oral regimens) may be a significant period relative to half-life and needs to be incorporated into the methodology. Finally, many drugs undergo conversion in the human (often in the liver) to other active forms that also kill the pathogens. This manuscript describes the computational extensions required to update the standard mechanistic-based modelling approach to allow for multiple drugs within a combination, and their absorption/conversion phases. We then illustrate their application to the current batch of first line antimalarial drugs, the artemisinin-based combination therapies (ACTs).
Malaria caused by Plasmodium falciparum, is one of the top three infective killers of humans with an estimated 0.75 to 1.5 million deaths per annum [2]. ACTs are now the WHO recommended first-line treatment for uncomplicated malaria [3]. The deployment of these combination therapies was designed to slow or even prevent the evolution of drug resistance which has, historically, been a potent threat to successful malaria treatment; delays in changing policy led to the widespread retention of ineffective drugs and acrimonious accusations of ‘medical malpractice’ aimed at such august institutions as the World Health Organisation [4] and the malaria community must prevent any similar situation arising. However, the policy of deploying ACTs worldwide has lead to increasing levels of artemisinin-tolerance and possibly artemisinin-resistance in Plasmodium falciparum being reported on the Cambodia-Thailand border [5], [6], [7], [8], [9] leading to intense speculation about how this will affect the current and future effectiveness of ACTs (e.g. [10], [11]). It is not possible to directly observe the consequences of antimalarial drug resistance until it is too late, so the best approach is to develop the best possible in silico models to help guide deployment policies aimed at maintaining long-term effectiveness of these key anti-infective drugs. We therefore apply our updated in silico PK/PD modelling methodology to explicitly investigate two front-line ACTs and the public health consequences of increasing tolerance and resistance. Accurate PK/PD modelling has two further important applications. Firstly, it can generate accurate simulations of field data upon which methods of analysis can be developed and refined [12]; the underlying parameters of interest are often unknown in field data but are easily recovered from simulated data enabling the performance of statistical tests to be gauged. Secondly, they can be used to investigate real-life situation that cannot be ethically addressed in the field, an obvious example being poor adherence to a treatment regimen.
We use mechanistic PK/PD modelling [1] as previously described in Winter & Hastings [13] with the four key extensions outlined below.
Standard PK/PD models [1] and their subsequent application to malaria [13], [14], [15], [16], [17] have previously assumed the drugs are instantaneously present in the serum at time t = 0, are not converted to any other form and decay at a rate Ct = C0e-kt, where Ct is the drug concentration at time t and k is the terminal elimination rate. This assumption is questionable for ACTs as their absorption and subsequent conversion to its active metabolite dihydroartemisinin (DHA) occur over a time period of 1–2 hours, roughly equivalent to their half-life (Figure S1). To address this assumption we track the time course of artemisinin absorption and conversion as illustrated in Figure 1 i.e. absorption across the gut (component A) into the serum (component B) at rate x, its elimination from the body at rate y or its conversion to the active metabolite (DHA) (component C) at rate z and the subsequent elimination of DHA from the body at rate k.
The drug-dependent killing function, f(C), was described using the standard Michaelis-Menton equation(1)where C is the drug concentration (mg/l) which decays over time, Vmax is the maximal drug-killing rate (per day), IC50 is the concentration at which 50% of the maximal killing rate occurs (mg/l) and n is the slope of the dose response curve. The problem is therefore to find how C varies over time following treatment so that it can be incorporated into Equation 1.
We use a standard one-compartmental model (Figure 1) that appears appropriate for constituents of current ACTs (Text S1), to track the changes in concentration over time. To avoid confusion, we note that “one compartment” is used in the standard PK sense i.e. only one body compartment (in this case, serum) is investigated besides the gut. The change in drug concentration occurring for each component over time (allowing for complications caused by the presence of the drug/metabolite from previous dosages) can be described by three differential equations(2)(3)(4)
To find the amount of converted and unconverted drug in the serum at time t, Equations 3 and 4 were integrated using laplace transformations [18] (Text S1). Integrating Equation 3 gives(5)where B(t) is the amount (mg) of unconverted drug in the serum at time t, A′ is the amount (mg) of drug in the gut at the immediate end of the previous time step (time steps correspond to the time between dosages, described in Text S1) i.e. at t = 0 (A′ = 0 if this is the first dose of a multi-dose regimen), D is the drug dosage (mg) given and B′ is the amount (mg) of unconverted drug in the serum at the immediate end of the previous time period i.e. at t = 0 (B′ = 0 if it is the first dose). Inclusion of any drug left over from the previous time step (denoted A′, B′ and C′) is essential when including repeat dosages.
Integrating Equation 4 (Text S1) gives(6)where C(t) is the amount of converted drug present in the serum, k is the elimination rate of the converted drug, C′ is the amount (mg) of converted drug in the serum at the immediate end of the previous time step (C′ = 0 for the first dose) and M represents the molecular weight of both the unconverted drug (MB) and converted drug (MC). We are tracking drugs in mg so the ratio of the molecular weights of species B and C, MB and MC respectively, are required to account for the changes in molecular weight that occur during conversion.
The drug-dependent killing described in Equation 1 required the amount of drug to be converted to a concentration (mg/l). This was found by dividing the amount of drug by the volume of distribution (l) which is the weight of the patient W, multiplied by the volume of distribution Vd per kg. The value of Vd differs between the drugs so VdB and Vdc represent volumes of distribution for drug forms B and C respectively.
The concentration of component B at time t, CB(t), is therefore(7)and the concentration of component C at time t, CC(t) is(8)
The use of Laplace transformations in PK is relatively well established [18] so it would be straightforward to extend the calculations for increasing numbers of compartments, drug forms and conversion elimination routes.
The existence of additional compartments in PK estimation models can be taken as an example. To recap, PK/PD mechanism based modelling of malaria requires drug concentrations in the ‘blood’ compartment but all PK estimation models try to include additional compartments where drugs can go; for example a drug may go into a “fat” compartment with fluxes between the blood and fat compartments. PK estimation models decide whether additional compartments are justified by using an information criterion (usually AIC). The problem is that PK estimation modelling is not straightforward and a fair amount of subjective judgement may be required. This subjectivity, combined with different datasets, may result in different analyses of the same drug fitting 1 or 2-compartment models [19]. When using the model it is important that researches maintain consistency in the PK model structures (i.e. assuming one or two compartments). For example, PK parameters derived from a two-compartment model should be incorporated into a PK/PD model that also uses a two-compartment structure. The use of Laplace transforms to incorporate 2 compartmental models is illustrated in the Text S1; users wishing to use a 2 compartmental model can therefore replace equations 5 and 6 obtained above for a one compartment model with Text S1 equations 1.20 and 1.21 obtained from a 2-compartment model.
The PK/PD modelling now allows for artemisinin absorption and conversion (described above), so the ability to track more than two drug concentrations simultaneously and convert them into a drug-killing rate is crucial. This feature is absent from previous pharmacological models of malaria, which track only a single drug [1] although we previously extended the methodology to track up to two drugs [13]. Existing pharmacological models typically use a standard differential equation [1] to find a mathematical description for the rate of change in total parasite growth and death rates(9)where P is the number of parasites in the infection, t is time after treatment (days), a is the parasite growth rate (per day), f(C) represents the drug-dependent rate of parasite killing which depends on the drug concentration C, and f(I) the killing resulting from the hosts background immunity.
As antimalarial drugs are now typically deployed as combination therapies and as each drug may affect parasites in its unconverted and/or converted forms, predicting the changing numbers of parasites requires an expansion of Equation 9(10)where r is the number of drugs, the drug effect f(Cd) is the effect of each drug, d. Note that we regard each active entity as a distinct “drug”. For example artemether-lumefantrine (AR-LF) includes three drug forms lumefantrine (LF), artemether (AR) (unconverted) and its active metabolite DHA (dihydroartemisinin). Note that Equation 10 assumes drugs kill independently; this is discussed further below.
Integrating Equation 10 allows us to predict the number of parasites at any time, t, after treatment with any number of drugs. This was done by first integrating Equation 9 using the separation-of-variables technique(11)
Integrating both sides of Equation 11 givesso
Taking the exponential of both sides (and noting that a times 0 = 0) givesso(12)
The problem is now to integrate f(C). Assuming there are r separate drugs/metabolites with antimalarial activity. In this case, f(C) becomes(13)
So for each drug/metabolite d we need to calculate its concentration over time Cd using the compartment model Equations (7 and 8) and the substitute Cd into the killing rate equation(14)
Note in Equation 14, is the maximum drug killing Vmax for drug d.
Substituting Equation 13 into 12 gives(15)or, equivalently,(16)
Note that Cd may be a complicated expression (including Equations 7 and 8) and so has to be integrated numerically. As before [13], if the predicted parasite number (Pt) falls below 1 we assume the infection has been cleared and the patient cured, immunity is currently ignored (see Winter & Hastings [13] for further discussion).
These computational extensions to the mechanistic PK/PD modelling allow for the presence of two or more drug forms simultaneously present in the human host, and active against the infection. It therefore becomes necessary to consider and specify how these drug forms interact in their effect against the parasites. There appears to be four main computational choices.
Pharmacological ‘mechanism-based’ modelling [1] has been used previously to investigate key features of antimalarial drug treatment either as monotherapies [14], [15], [16], [17] or with recent emphasis on the current generation of ACTs [13]. We have previously touched upon the potential consequences of increasing artemisinin resistance using standard pharmacokinetic-pharmacodynamic (PK/PD) modelling techniques [13] however, as mentioned in the paper, the model relied heavily on two main assumptions built in to the existing methodology. First, that all drugs are instantaneously absorbed and, if appropriate, converted to their active metabolites. Whilst this may be reasonable for drugs with a long half-life it is not practical for drugs like the artemisinins where absorption and conversion times are almost equal to their short half-lives. The second assumption, that no more than two drugs could be present simultaneously, was reasonable when modelling the ACTs if both drugs were instantaneously absorbed and converted. However, conversion of the artemisinins requires that the artemisinins be modelled as two separate component drugs i.e. the parent drug and the DHA metabolite together with the partner drug and so modelling the ACTs requires a minimum of three drugs be tracked simultaneously. Here we have addressed the methodological challenges of incorporating the absorption and conversion phases of drugs into PK/PD modelling while simultaneously tracking the concentration of more than two drugs, a feature absent in previous pharmacological models [14], [15], [16].
The PK/PD model parameters required to simulate treatment are given in Table S1 and described in the Text S1. The PK extensions for the artemisinins required additional parameters describing the drug absorption rate across the gut, the conversion rate to DHA and the elimination of DHA from the body (Figure 1). These parameters and their associated distributions can be found in Table S1 with details of model calibration and validation included in the Text S1. Variation in model parameters was previously [13] added assuming a coefficient of variation of 30% in all parameters. In reality, some parameters are much more variable [22] while others maybe less so. We now incorporate more appropriate levels of variation into the PK/PD parameters using drug specific distributions thus making results more compelling for specific ACTs. To validate the model's predictive ability, the maximum serum concentration (Cmax) and time to achieve Cmax (Tmax) were compared to field data (Text S1).
The methodology described above now allows for the action of both the unconverted and converted forms of the artemisinins. However, given that they have similar modes of action their effect on parasite numbers is unlikely to be additive (as is assumed in Equation 11). As such, the drug effect, f(C), for each of the artemisinin forms was calculated each time-step but only the dominant form (i.e. parent drug or active metabolite) with the greater drug killing effect was used to compute the number of parasites in the next time step. Activity, and hence killing, of artemisinins and the partner drug were assumed to be independent.
A major change was made to the artemisinin maximal drug kill rate (Vmax). Previous estimates of the Vmax [13], [23], [24] have been based upon the assumption that drug killing is maximal immediately after treatment and remains so for 48 hours after treatment. This is quantified by the parasite reduction ratio (PRR); a ratio of the number of parasites at time of treatment scaled by their number 48 hours after treatment. So, assuming the decline in parasitaemia is first order, the parasite count (Pt) at any given time (t) is given by(18)where P0 is the number of parasites present at the start of treatment.
This appears to be reasonable for drugs given at relatively high doses with a long half-life because the maximal killing will extend over the 48 hours after treatment. However, it is unrealistic for the artemisinins whose short half-lives mean parasites are typically only exposed to high concentrations of artemisinins during the first 6–8 hours following treatment (Figures S1 and S2). The steady decline in parasite numbers after this period presumably reflects dead or dying parasites being cleared by host mechanisms. PK/PD modelling of drug effect assumes deaths only occur in the presence of the drug (i.e. 6–8 hours post-treatment) hence the need for this increased kill rate. So, given PRR = P0/Pt [23] (where Pt is usually assumed to be 48 hours), the relationship between PRR and parasite killing rate Vmax is(19)When t is assumed to be 48 hours and PRR is 104 then the maximal artemisinin drug kill rate (Vmax) is 4.6 as used previously by ourselves and others [13], [23]; we now consider that value inappropriate because a 6 hour burst acting at a kill rate of 4.6 would achieve a PRR of well below 104 . Consequently, we assume artemisinin maximal drug killing occurs only during the 6 hours when the drugs are actually present at therapeutic concentration (Figures S1 and S2), so achieving a PRR of 1000 (White [23] gives a range of 103 to 105 for the artemisinins) requires Vmax to be 27.6. Note, if the maximal drug killing is assumed to occur over 8 hours and the PRR is assumed to be 10,000 (within the range reported in White [23]) Vmax again equals 27.6. Consequently our artemisinin maximum killing rate is approximately 7-fold higher than in previous simulations.
Two treatment combinations were investigated, artesunate-mefloquine (AS-MQ) and AR-LF, both are highly effective ACTs currently used to treat malaria. Variation in how humans metabolise the drug and parasite drug sensitivity was added to the model parameters (Table S1) using parameter specific estimates of co-efficient of variation, CV. The technical details regarding parameter variability are included in the Text S1.
The extended pharmacokinetic-pharmacodynamic (PK/PD) model can then be implemented to address a critical feature of current ACT deployment: how is the observed increase in artemisinin tolerance likely to affect the long-term effectiveness of ACTs? The crucial operational question is whether there is likely to be a sudden catastrophic decrease in ACT effectiveness, a gradual decline or, a best case scenario, a margin of safety such that we can have relatively large increases in artemisinin tolerance/resistance before ACT failures start to increase?
The partner drugs, LF and MQ, are currently largely effective monotherapies if administered correctly (although MQ in south east Asia may be problematic) so increasing artemisinin resistance would, by definition, have little or no impact on therapeutic outcome. To avoid this trivial case, we investigated how increasing levels of artemisinin resistance impacted treatment failure rates if resistance to the partner drug was already present or spreading. When modelling MQ treatments the MQ IC50 values were either 1-, 2-, 5-, 10-, 15-, 20- or 25-fold greater than the current default value (Table S1) and when modelling LF treatments LF IC50 values were either 1-, 2-, 5-, 10-, 20-, 25- or 50-fold greater than the current default value (Table S1). Resistance to artemisinins was investigated in two ways. First by increasing the IC50 of the AS, AR or DHA (the active metabolite) independently and then by assuming the IC50s of the parent species and DHA were completely correlated i.e. the IC50s were increased simultaneously by the same amount. This was necessary because it is not clear whether parasites will evolve resistance independently to the artemisinin entities or whether there will be substantial cross-resistance to different entities (see later discussion) The IC50 range of both artemisinin forms included one value 10-fold smaller than the mean and values 1-, 20-, 40-, 80- or 100-fold greater than the mean.
Details of implementation are in the Text S1. For each of the 10,000 patients simulated the model recorded whether an infection (with one clone) was cleared and, if so, the parasite clearance time (PCT; defined as the time taken for an infection to fall below the limit of microscopic detection, which was assumed to be 108). This was done first for the partner drugs without the artemisinin component, i.e. as monotherapies, to give a baseline failure rate. Then, by comparing the results of the monotherapy with those of the ACTs we were able to quantify the ability of the artemisinin component to reduce failure rates and PCTs.
The artemisinin drug concentration profiles of the model are consistent with those measured in the field (discussed in Text S1 and Figure S1). Analysis of both ACTs showed that adding an artemisinin to a partner drug reduced failure rates below that of the monotherapy regardless of the initial levels of partner drug resistance, the latter being achieved through varying the partner drug IC50 value (Figure 2); the only exception was the trivial case when partner drugs were fully effective as monotherapies. For AS-MQ, the exact proportion of failures prevented by the artemisinin component was dependent on the initial level of resistance to the partner drug. Regardless of whether the IC50s of the artemisinins were correlated, adding an artemisinin at its default IC50 value to a partner drug reduced failure rates by between 70 and 90%. This is a relative reduction, for example, a 50% reduction is equivalent to fall in failure rates from 40% to 20% or from 12% to 6% (Figure 2, panels A, C and E). This is consistent with field observations that adding AS to MQ reduced the absolute risk of failing treatment but did not result in a fully effective ACT [25]; this has also been observed for other failing monotherapies not modelled here (chloroquine, amodiaquine, sulfadoxine-pyrimethamine) [25]. The results also show that the addition of AR to LF monotherapies reduced failure rates to zero when modelling the mean parameter values (Figure 2, panels B, D and F).
Figure 2 shows the failure rates of the ACTs when the IC50s of the two artemisinin drug forms were either varied independently (Figure 2, panels A to D) or varied simultaneously (Figure 2, panels E and F). When the IC50s of the artemisinin drug forms were varied independently increasing the IC50 of either had very little effect in the failure rates (Figure 2, panels A, B, C and D). This was particularly clear for AR-LF treatments where increasing either AR or DHA IC50 caused no measurable increase in drug failure rates (Figure 2, panels B and D). This occurs because resistance to one form is compensated by continued sensitivity to the other form because both forms are potentially capable of high rates of parasite killing (Figure S2). Increasing AS IC50 alone also had little effect on the AS-MQ failure rates (Figure 2, panel A), again highlighting the importance of its active metabolite on parasite survival. When DHA IC50 was increased by 20-fold in AS-MQ treatment (Figure 2, panel C), treatment failures increased by 25 to 65% (relative increase) depending on the level of resistance to the partner drug. This is the only time increasing either the artemisinin drug forms alone affected treatment outcome and further DHA IC50 increases (above 20-fold) had little further effect on treatment outcome (Figure 2, panel C). Failure rates to AS-MQ assuming the artemisinin drug forms were uncorrelated (Figure 2, panels A and C) remained lower than those seen when assuming they were correlated (Figure 2, panel E) thus implying both artemisinin drug forms are still playing an active role in parasite killing.
Further DHA IC50 increases above 20-fold had no discernable effect on treatment outcome and failure rates remained lower than those seen when the IC50's were correlated thus implying that while not as potent as AR and DHA it still plays an active role in parasite killing. For both ACTs, increases in failure rate as a result of increasing artemisinin resistance were much larger if the IC50s of the artemisinin drug forms were simultaneously increased. Rapid loss of protection was most noticeable for AS-MQ with small IC50 increases (20 and 40-fold), well within the range of natural variation [22], increasing failure rates by 65–70% (Figure 2, panel E). Loss of protection was more gradual following AR-LF treatments (Figure 2, panel F) but both ACTs showed failure approaching those of the of the monotherapies as artemisinin IC50s increased to 100-fold greater than the mean.
The PCT appears to be determined predominantly by the level of resistance to the artemisinin component with the initial level of partner drug resistance being relatively unimportant (Figure 3). This was particularly evident following AR-LF treatment where increasing the IC50 of LF had no discernable effect on PCT (Figure 3, panels B and D) while increasing MQ resistance only caused the PCT to vary by up to one day (Figure 3, panels A and C). When the IC50s of the two artemisinin species were increased simultaneously, the addition of artemisinin to the monotherapy reduced PCTs by approximately 2 to 3 days for both ACTs. As seen with the treatment failures (Figure 2), increasing the IC50 of AS/AR or DHA independently had little/no effect on PCT (Figure 3, panels A to D) and PCT did not approach that of the monotherapy because the other artemisinin species retained its effectiveness. When the IC50s were increased simultaneously both artemisinin species lost their effectiveness (Figure 3, panels E and F) while the PCT increased almost linearly with increasing artemisinin resistance and approached the PCTs seen with monotherapies (Figure 3, panels E and F).
The extended PK/PD mechanism based modelling was applied to ACTs and produced results and predictions consistent with field data on failure rates [25] and increasing PCT associated with resistance. The main operational concern surrounding the evolution of artemisinin resistance is that it will lead to clinical failure in patients treated with ACTs [26]. Obviously, if the partner drug is effective as a monotherapy, then the presence or absence of artemisinin resistance has no clinical effect. Problems arise as resistance spreads to the partner drugs, a process slowed by the addition of an artemisinin [27]. The results clearly show that adding AS to a failing drug (MQ) reduced the treatment failure rates by up to 90% (relative reduction) but did not result in a fully effective ACT (Figure 2, panel E). This observations is in line with the findings of the International Artemisinin Study Group who performed a meta-analysis of individual patients from 16 randomised trials (n = 5948) studying the effect of adding AS to either CQ, AQ, SP or MQ [25]. While the total population failure rates were reduced by 42–65% when averaged across all drug regimens, the addition of AS to MQ monotherapy reduced failure rates by approximately 90–95% [25]. The results for AR-LF show that the addition of AR with default IC50 values was sufficient to save a failing LF monotherapy by reducing failure rates to <1% for all levels of partner drug resistance regardless of whether the IC50s of the AR and DHA are increased simultaneously or independently (Figure 2, panels B, D and F). However, this observation was much more difficult to validate than those of AS-MQ as there is almost no published data on the in vivo efficacy of LF monotherapy and so it is impossible to quantify the proportion of failures averted specifically by the addition of AR. We also note that for both ACTs, only when the IC50s were correlated did increasing the IC50 eventually lead to failure rates approximately equal to those of the monotherapy therefore removing any benefit afforded to the partner drug by the artemisinin. These occurred after 50–100 fold increases in artemisinin IC50 which is large, but around the same magnitude as the natural variation observed in field isolates [22]. The key question is whether the IC50s are correlated; field data suggest they are (Text S1).
Increasing PCTs are currently being observed in the field [7], [26], [28], [29], [30]; Dondorp et al. [31] for example, show that parasites resistant or tolerant to artemisinins take 3 or 4 days to parasites as compared with less than 2 days for artemisinin sensitive parasites; this pattern was also apparent in the results presented here (Figure 3). The simulated results showed the initial level of resistance to the partner drug had very little effect on the PCT and whilst this may seem strange it can be explained relatively easily. While the partner drug is undeniably important when determining the treatment outcome (i.e. success or failure), the PCT is determined almost solely by the short-lived but fast-acting artemisinin component, which causes a rapid decline in parasite numbers but is not present long enough to completely clear the parasite load [13]. As with dug failure rates, PCT only approached those of the monotherapies when the IC50s were increased simultaneously again consistent with field data that the IC50s are correlated (Text S1). For both ACTs, PCT began to increase after relatively small increases in artemisinin IC50 of 20- to 40-fold (within the range of natural variation [22]).
The results shown on Figure 2 illustrate an important factor not generally recognised when considering how resistance may arise to artemisinins and other drugs whose converted and unconverted forms are both active: if resistance arises to only one form, then the other form may retain sufficient activity to compensate. This is well illustrated by AR-LF in Figure 2 where increasing resistance to either AR alone (Figure 2B) or DHA alone (Figure 2D) has virtually no impact on failure rates which only start to escalate if resistance occurs simultaneously to both forms (Figure 2F). It is therefore essential to consider whether mutations that encode resistance to one form are likely to simultaneously encode cross-resistance to the other form (so that IC50s are correlated), or whether the mutations are specific to individual drug forms (in which case IC50s are uncorrelated). When considering the likelihood of cross-resistance, it is important to realise that cross-resistance and mode of drug action are related, but distinct entities. Drugs with identical modes of action may show complete cross-resistance if mutations occur at their site of action which prevents both/all forms of the drug from binding therefore blocking their activity. Alternately, resistance may emerge through mutations that alter the drugs' ability to reach or accumulate at their site of action. Malaria is often characterised by the latter where mutations in membrane transporters, notably mdr and crt, are implicated in resistance to a range of antimalarial drugs [32]. These transporters depend more on the chemical scaffold (charge and structure) of the drug than its active site so it is not a priori certain that cross-resistance will inevitable occur between a parent drug and its active metabolite. A lack of cross resistance would be hugely beneficial as it means parasites would have to evolve resistance to both forms of the drug but, unfortunately, our simulations suggest a model of complete cross resistance provides the best fit to the malaria observations that IC50s are likely to be correlated (discussed further in SI).
Drug IC50 values are estimated either from parasites taken from a patient's primary infection or from laboratory isolates. The IC50 values of the artemisinins and their active metabolite DHA vary widely in the literature and their reported values appear to be highly dependent on the source from which they were estimated. For example, Brockman et al. [33] show the mean IC50 of AR was approximately 4-fold higher than DHA (4.83 and 1.22 respectively) when measured in patients from Thailand but were approximately equal (3.4 and 3.6 in 1996 and 3.1 and 4.0 in 1998 respectively) when measured in K1 laboratory isolates. The 4-fold lower DHA IC50 measured in patients may result in a higher level of effectiveness of DHA in their patient population. What is not generally realised is that both artemisinin components are potentially important in determining treatment outcome; for example Saralamba et al. [34] simply stated that in their patients “the total drug exposure of AS was <10% that of DHA” and so choose to ignore the parasiticidal effect of AS. This may be true on average, but there is huge variation in how patients metabolise different forms of the drug so it entirely plausible that some patients will slowly convert artesunate but rapidly clear DHA, in which case the former would have the larger killing effect. In particular, changing IC50 simply translates into how long the drug is killing at near-maximal rates in the few hours following treatment (Figure S2). Importantly, this means that artemisinin therapy given as artesunate or artemether has an inherent therapeutic safety margin: If one component of the artemisinin is metabolised quickly or has a particularly high volume of distribution, there is still a second active component present within the patient that is likely to retain therapeutic effectiveness.
Increasing tolerance/resistance to artemisinins was modelled using the standard assumption that it will arise through increased IC50 values. Artemisinin resistance may be atypical in this respect as it appears to manifest through increased clearance times of parasites following treatment with unchanged IC50, possible due to the drug(s) having activity against a more restricted range of stages in the malaria cell cycle (see below). The mechanistic approach assumes instantaneous killing of parasites irrespective of their stage, so deceased activity against some stages would be manifested as decreased drug maximal killing rate (Vmax in Equation 1) in the methodology; interesting this parameter was found to be a far more potent determinant of resistance than the IC50 [13]. It would be possible to re-run the above simulations altering Vmax rather than IC50 but we chose to use the more conventional approach in the first instance as we consider this primarily a computational paper; we shall explore this approach in future studies applying the methodology more specifically to malaria.
Malaria differs from many other pathogens in having a distinct 48 hour intracellular cycle that essentially consists of invasion of red blood cells (RBC), digestion of host haemoglobin, parasite multiplication within the RBC, cell rupture and re-invasion of new RBCs. Drugs consequently have different stage specificity profiles depending on what metabolic processes are occurring in each stage (for example, many drugs target haemoglobin digestion so are primarily active against parasites in this stage of their cycle). Our analyses ignored these drug stage-specificities. It would however be easy to re-compute the dynamics using one hour time steps and using a 48 hour array to move parasites through the 48 hour development cycle as done previously [35], [36], [37]. We chose not to do so for two main reasons. Firstly, stage specificity requires that PD parameters be specified for each stage and that the initial distribution of parasite stages in the infection be specified. Secondly, and more importantly in our opinion, is that the PK/PD computations assume instantaneous killing of parasites depending on current drug concentration whereas, in reality, there is a delay in killing. The delayed killing can be incorporated into the methodology by postulating a hypothetical ‘metabolite’ whose production or elimination is disrupted by the drug, and that parasite death occurs as a function of metabolite level; the time taken for metabolite levels to reach ‘lethal’ levels introduces a time-lag into the killing [38], [39]. This is an elegant way of incorporating a delay but it requires further parameterisation of the metabolite's production and elimination, specification of a killing rate as a function of metabolite level, and calibration against field data. Patel and colleagues [38] estimated the delay in artemisinin killing as around 5 hours. A recent study attempted to simulate ACT dynamics using a stage structured approach and concluded that it did not match well field data [36]; we are unsurprised because the short-term dynamics will be critically dependent on stage-specific PD parameterisation and no time lag was built into the model. Hence, our approach was to ignore short-term dynamics and run the enhanced PK/PD methodology, ignoring stage specify and delayed drug action [40]; the objective was to simulate the fate of the infection over the longer term rather than the dynamics immediately post-treatment. Consistency of our results with field and clinical observations suggest this is a robust approach but it is important to recognise the alternative modelling approaches can be designed, and that our enhanced PK/PD methodology can easily form the basis for an improved stage-specific model run in 1-hour time steps.
The rationale behind this paper is that combining good quality field and clinical data into a sophisticated PK/PD model should allow a thorough investigation of ACT effectiveness in the context of increasing artemisinin tolerance/resistance. It therefore provides a methodological framework for clinical pharmacologists to interpret their results. However the predictive power of mathematical modelling is governed by the crucial step of model calibration and the availability of comprehensive, good quality PK/PD data in the literature is surprisingly scarce (Supporting Information, part 2). This has the potential to limit the usefulness of models as predictive tools. Given the amount of effort and resources required to conduct PK/PD studies and that their explicit aim is usually to improve human therapy, it seems appropriate to consider how best to report such studies for maximum impact. We therefore make three specific suggestions that authors may consider to maximise their studies' chance of influencing policy choice. Firstly, all available population PK/PD data, including those required purely for intermediate calculations should be reported. For example, terminal elimination rates are invariably reported but parameters required in their calculation, for example volumes of distributions (often confounded with bioavailability) are often omitted [41]. We are uncomfortable with the rationale underlying the common assertion that DHA is the main active species during artemisinin treatment (see above and Figure S2); we would therefore recommend that PK parameters for parent species such as artesunate and artemether also be measured and reported. Secondly, the nature and extent of natural variation in the parameters are vitally important and can result in some patients developing low drug concentrations possibly leading to therapeutic failures or high concentrations potentially leading to toxicity. The distributions (normal, log-normal, etc) with their associated coefficients of variations (CV) are therefore almost equally important as their mean values. For example, many authors cite CV estimates larger than the mean, which obviously indicates a non-normal distribution: such data are much more useful if accompanied by their distributions (herein we were forced to assume they were log-normal). Finally, there are wide variations in reported mean values between studies; these are generally ascribed to sampling different populations or age groups but a more critical appraisal in terms of any impact of different methods of analysis would also be helpful. An excellent example is that of Tan et al. [42] who, after describing the population PK of AS and DHA in healthy patients, compare their results with those of other AS and DHA PK studies and provide a detailed discussion explaining how and why the results may differ.
We would emphasise that our choice of specific studies to parameterise the simulation should not be regarded as prescriptive or judgemental; as described above, the choice was often problematic. Few, if any, PK/PD studies produce all the parameters required to evaluate their impact on therapeutic outcome. PK studies often focus on a single drug in a combination and lack local estimates of parasite drug sensitivity, while PD studies generally lack accompanying PK estimates. Consequently, we have focused on developing a methodology that individual researchers can calibrate as they wish; we provide a mechanism by which their results can be integrated with the results of other studies to gauge their implications for drug effectiveness.
Despite the caveats mentioned above, our results and implications are clear. The kill rate of both artemisinin forms appears to be important in determining treatment outcome and their IC50's are likely to be correlated. AS-MQ is more sensitive to increases in artemisinin drug resistance than AR-LF with the number of failures increasing quickly with relatively small increases in AS and DHA IC50s. Both ACTs show increasing PCT associated with increasing artemisinin IC50, an observation already seen in the field [5], [6], [7], [8], [9]. Our results suggest this is indicative of a rapid loss of protection provided by the artemisinins against the partner drug(s). If, or when, resistance against the partner drug starts to increase, most plausibly driven by mismatched half-lives [43], [44], [45], then a rapid reduction in ACT clinical effectiveness is likely to occur. We conclude that policies designed to isolate and minimise the spread of artemisinin resistance are to be greatly encouraged [26].
|
10.1371/journal.ppat.1003571 | Schistosoma mansoni Mucin Gene (SmPoMuc) Expression: Epigenetic Control to Shape Adaptation to a New Host | The digenetic trematode Schistosoma mansoni is a human parasite that uses the mollusc Biomphalaria glabrata as intermediate host. Specific S. mansoni strains can infect efficiently only certain B. glabrata strains (compatible strain) while others are incompatible. Strain-specific differences in transcription of a conserved family of polymorphic mucins (SmPoMucs) in S. mansoni are the principle determinants for this compatibility. In the present study, we investigated the bases of the control of SmPoMuc expression that evolved to evade B. glabrata diversified antigen recognition molecules. We compared the DNA sequences and chromatin structure of SmPoMuc promoters of two S. mansoni strains that are either compatible (C) or incompatible (IC) with a reference snail host. We reveal that although sequence differences are observed between active promoter regions of SmPoMuc genes, the sequences of the promoters are not diverse and are conserved between IC and C strains, suggesting that genetics alone cannot explain the evolution of compatibility polymorphism. In contrast, promoters carry epigenetic marks that are significantly different between the C and IC strains. Moreover, we show that modifications of the structure of the chromatin of the parasite modify transcription of SmPoMuc in the IC strain compared to the C strain and correlate with the presence of additional combinations of SmPoMuc transcripts only observed in the IC phenotype. Our results indicate that transcription polymorphism of a gene family that is responsible for an important adaptive trait of the parasite is epigenetically encoded. These strain-specific epigenetic marks are heritable, but can change while the underlying genetic information remains stable. This suggests that epigenetic changes may be important for the early steps in the adaptation of pathogens to new hosts, and might be an initial step in adaptive evolution in general.
| Schistosoma mansoni is a parasitic worm and agent of a disease that causes a considerable economic burden in African and South American countries. The propagation of the parasite requires passage through a freshwater snail of Biomphalaria genus. In the field, actually very few snails are infected. This is due to the fact that specific strains of the parasite can infect only specific strains of the snail. Comparative studies have shown that this so-called compatibility is based on the expression of a family of genes that are called SmPoMucs. We have shown previously that all parasites strains possess the repertoire of all SmPoMuc genes but every strain and even every individual parasite expresses only a subset. These differences could be due to DNA sequence differences in the regions that control gene expression, but here we show that these regions are nearly identical. Instead, the chromatin structure shows strain-specific characteristics. This means that the parasite can adapt to different snail strains simply by changing its chromatin structure and not necessarily the DNA sequence. If this holds true for other parasites, then we have to rethink the way parasite evolution is currently imagined but this also provides a new potential entry point to control the spread of diseases.
| The interaction of hosts and parasites is one of the best-studied examples of evolution in a changing environment [1]. Their reciprocal antagonistic co-evolution can be illustrated by an arms race in which host and parasite develop mechanisms to circumvent counter-measures developed by their opponents [2], [3]. Under certain conditions, parasite virulence and host defence can be in equilibrium leading to a phenomenon called compatibility. Compatibility occurs in a host-parasite system when the parasite species is capable of infection and transmission through the host species [4]. The phenomenon that some parasite strains are compatible with certain host strains but not with others (and vice versa) is called compatibility polymorphism. This phenomenon was described in the platyhelminth Schistosoma mansoni and its intermediate host, the mollusc Biomphalaria glabrata [5]. S. mansoni is a human parasite whose life cycle is characterised by the passage through two obligatory sequential hosts: the fresh-water snail B. glabrata (or dependent on the geographical location other Biomphalaria species) for asexual replication, and humans or rodents as hosts for sexual reproduction [6]. The molecular mechanisms underlying compatibility polymorphism between S. mansoni and B. glabrata were recently investigated by comparing the proteomes of two S. mansoni laboratory strains: one strain that is compatible (the C strain) and one that is incompatible (the IC strain) with the same reference B. glabrata strain from Brazil [7]. The study identified S. mansoni Polymorphic Mucins (SmPoMucs) as key markers for compatibility (see [4] for a recent review). SmPoMuc glycoproteins have a mucin-like structure with an N-terminal domain containing a variable number of tandem repeats (VNTR) [8]. SmPoMuc proteins are highly polymorphic [8] and interact with the Fibrinogen RElated Proteins (FREPs) of the mollusc [9]. FREPS are diversified antigen recognition molecules playing a central role in the secondary immune response to digenetic trematodes [10], [11], [12]. The extraordinary level of SmPoMuc polymorphism is generated by a complex cascade of mechanisms, a “controlled chaos”, acting at the transcriptional, translational and post-translational level [8]. SmPoMucs are encoded by a multigene family with at least 10 members that are organised in 4 clusters on the genome. They recombine frequently and generate new alleles [8]. Each individual miracidium (the larva that infects the mollusc) expresses only a specific subset of SmPoMuc genes. The mechanisms controlling this expression polymorphism of SmPoMucs remained unclear. Our recent finding that Trichostatin A, a modifier of chromatin structure, influences SmPoMuc transcription patterns [13] suggests that epigenetic mechanisms participate in transcription control.
Epigenetic information is information on the status of gene activity that is heritable, for which changes are reversible and that is not based on the DNA sequence [14], [15], [16]. The scientific debate about the reason of the evolution of an epigenetic inheritance system (EIS) in most organisms is intense. Others and we have suggested that EIS provides a basis for modifications in the reaction norms that do not require changes of genotypes [17], [13], resulting in increased phenotypic plasticity at the individual level or increased phenotypic variability at the population level. If EIS influences the capacity to generate different phenotypes, both the better adapted phenotype and the capacity to generate this phenotype will be selected for and carried into the next generation. This hypothesis has been largely validated in the malaria parasite Plasmodium falciparum which displays “Clonally Variant Gene Expression” (CVGE) [18]. Genes that show CVGE are present in multicopy, such that individual parasites within an isogenic population express these genes at very different levels, often fully active or completely silenced. Their transcriptional patterns are clonally transmitted to the next generations through asexual multiplication, and stochastic changes of the transcription level occur at low frequency. This bet hedging strategy allows for stochastic generation of phenotypic diversity and can be controlled by epigenetic based events, similar to those described for the var gene family. The var genes encode the red blood cell surface antigen P. falciparum erythrocyte membrane protein 1 (PfEMP-1) and their “CVGE” regulation strategy is responsible for surface antigen variation that ultimately results in immune evasion. In this context, the EIS that leads to “CVGE” allows for rapid adaptation to the ever-changing vertebrate immune environment. In S. mansoni miracidia, we have shown that epigenetic-based events influence the phenotypic plasticity in populations [13] and particularly regulate SmPoMuc gene expression. To gain further insight into the precise mechanism of regulation of these genes, in the present study we investigated the genetic and epigenetic changes that occurred during the evolution of the phenotypic compatibility polymorphism in two S. mansoni strains. We focused on the sequences of the promoters of active SmPoMuc genes and investigated whether there exist differences in the promoter sequences between S. mansoni compatible and incompatible strains. Our study reveals that IC and C strains display very little within strain genetic variability, and limited nucleotide differences between promoter sequences of the two strains, but show strong chromatin structure differences. These chromatin structures are heritable throughout the life cycle and transmitted to the next generation, therefore demonstrating that EIS can control a heritable adaptive trait, such as compatibility polymorphism.
SmPoMuc genes are classified into 4 groups (Roger et al. 2008) according to their 3′region: group 1 to 4. Group 3 is itself divided into subgroups (3.1, 3.2, 3.3 and 3.4). SmPoMucs genes have a 5′ region containing a variable number of tandem repeats (exon2), which have been previously called r1 and r2 [8]. r2 exclusively occurs in the group1 and 2 and the intermingled r1–r2 exclusively occurs in the subgroup 3.1, which is present in several copies with either the r1–r2 intermingled repeats or with r1. Due to the very high degree of sequence similarity between the SmPoMuc groups, specific transcriptional analyses of the different SmPoMuc groups were only possible for groups 1, 2 and 3.1(r1–r2). The transcription levels of these groups were compared between miracidia of the IC and C strains. SmPoMuc gene groups 1, 2 and 3.1(r1–r2) are 2.2 to 4.9, 2.5 to 6.7 and 18.6 to 59.7 fold more transcribed in the IC than in the C strain, respectively (fig. 1). The 3.1 subgroup containing intermingled r1–r2 repeats is highly transcribed in the IC strain but was practically undetectable in the C strain. This result is consistent with a previous study on individuals of the IC and C strains, which showed that variants containing the r1–r2 combinations are only expressed in the IC strain [8].
To investigate the mechanisms underlying differences of transcription between SmPoMuc groups and subgroups, we characterized the minimal promoter region of the SmPoMuc genes. We sequenced a region spanning 1.04 to 2.00 kb upstream of the transcriptional start site (TSS) for 4 groups of SmPoMuc (Groups 1, 2, 3.1 and 3.1(r1–r2). We produced a PCR product of a 996 bp of the region of the promoter of the group 3.1(r1–r2) and a PCR product of 1002 bp of the group 3.1 just upstream of the transcriptional start site. Plasmids containing these sequences upstream of a reporter gene (EGFP) were transfected into HeLa cells and fluorescence was observed under a microscope (fig. 2). These experiments showed that these sequences are sufficient to drive the heterologous expression of the reporter gene and contain the minimal promoter sufficient for transcription.
As a first approach to investigate a putative genetic basis for the difference in transcription levels between strains, we investigated the paralogous and orthologous relationships between the four groups of SmPoMuc gene promoters and between the two S. mansoni IC and C strains using phylogenetic analysis, reciprocal BLAST dot-plots and comparison of repetitive elements, duplication, recombination events and gene conversions (fig. 3). We annotated the sequences and visualised them by colour-coding of blocks with less than 95% identity (fig. 3). A recombination event was detected using BootScan [19], [20], Maximum Chi Square [21], [22] and Sister Scanning [23] methods in RDP3 and the recombination break points were putatively identified (fig. 3). In both strains we observed one duplication in group 3.1(r1–r2) promoters resulting in an insertion, several insertions/deletions (indels) including one large deletion in group 3.1 promoters and probably a recombination event from the group 2 to group 1 promoter. High similarity to a repeated DNA element was detected in the group 2 promoter; however, it constituted only a small fragment of the complete repeat – 61 bp out of 385 bp of the DIVER2 LTR (Drosophila).
The estimated divergence time between the IC and C S. mansoni strains is about 400 years [6] and the promoter sequences between the two strains are highly conserved (0.000–0.004 net substitutions per site, Table 1). The number of fixed differences between the two strains varied between 0 in the promoter region of SmPoMuc group 2 genes, to 3 in group 3.1, 4 in group 1 and 8 in group 3.1(r1–r2) (Table 1). No substitution was observed in the TATA signal, nor in the TSS regions or in putative regulator binding sites of the promoters between the two strains. SmPoMuc promoter sequences were divided into four paralogous sequence groups and sequence differences between strains (orthologous relationships) within groups were much less than the differences observed between groups of the SmPoMuc gene family - net substitutions per site varied from 0.000–0.004 within groups of promoter sequences between strains compared to 0.024–0.041 between promoter groups (Table 1). The number of SmPoMuc promoter sequence differences between strains was equal to or slightly higher than the number of sequence differences for the promoter of the single copy gene SmFTZ-F1 [24] which shows no difference between strains (Table 1). Six of 14 microsatellite loci also showed no sequence differences between the two strains (one unique allele). The two strains share the molecular evolution and phylogeny of the promoter region of the four groups of the SmPoMuc gene family (fig. 3) – indels, recombination and duplication events. These findings indicate that the divergence between groups of the SmPoMuc gene family from a common gene ancestor is ancient and largely predates the time of separation between the IC and C strains.
At this stage of the study we hypothesized that SmPoMuc expression differences in C and IC strains could be due to nucleotide differences in the promoter regions of the genes. The sequencing of 1.4 kb of SmPoMuc group 1 promoter region for 20 and 18 individuals of the IC and C strains respectively, revealed a very low number of alleles and genotypes (Table 2) – one genotype in the IC strain and 3 genotypes in the C strain. In the C strain, sequence variation was minimal, with the three alleles differing by only one base pair from each other, resulting in insignificant nucleotide diversity (Table 2). All individuals were homozygotes. The IC strain allele of the SmPoMuc promoter group 1 differed from the three C strain alleles by four to five base pairs, a sequence divergence of 0.29 to 0.36%. In summary, nucleotide sequence differences between the two strains are surprisingly small.
Promoter diversity within strain and divergence between strains of SmPoMuc group 1 genes were similar to those of 14 microsatellite loci that can be used to reflect genome-wide diversity and divergence [25]. The promoter diversity of SmPoMuc group 1 was 0.00 (one allele) in the IC strain compared to 0.22 (3 alleles) in the C strain (Table 2), while expected heterozygosity was 0.000 (one allele) for both strains for 14 microsatellite loci (Data not shown). All individuals were homozygotes. Six out of 14 microsatellite loci showed no divergence between the two strains. At eight microsatellite loci, the IC strain alleles differed from the C strain alleles by one to eleven microsatellite repeats. The promoter region of the single copy SmFTZ-F1 gene displayed a unique sequence common to the two strains. We estimated extremely high and significant genetic differentiation between the two strains for both SmPoMuc group 1 promoter sequences and microsatellite loci using θ, ΦST and RST estimators (Table 3). However, we detected almost no heterozygotes and highly significant inbreeding coefficients f in both strains and for both SmPoMuc group 1 promoter sequences and the microsatellite loci (Table 3). Therefore the high values of divergence are likely the result of the bottleneck induced during the care of the life cycle in the laboratory in the two strains as discussed previously [25]. Nonetheless, the distribution of alleles matched the pattern of differentiation as we detected fixed alleles that were different in the two strains. We reasoned that the small genetic differences in the promoter region are simply a by-product of clonality and not the reason for expression differences. We therefore explored an alternative hypothesis, i.e. that the expression differences are due to dissimilarity in the epigenetic information.
As the difference in SmPoMuc transcription phenotype cannot easily be explained by genetic differences in the promoter region, we investigated the putative implication of epigenetic mechanisms. As a previous study had shown that histone modifications are clearly involved in S. mansoni epigenetic mechanisms [13], [26], we tried to influence the epigenotype and phenotype (SmPoMuc expression pattern) of S. mansoni using trichostatin-A (TSA) that is a specific and reversible inhibitor of class I and II histone deacetylases (HDAC). Treatment with this drug prevents histone deacetylation and is expected to increase the overall acetylation of histones and therefore gene expression [26][27]. The influence of TSA treatment on the transcription of SmPoMuc genes (group 1, 2 and 3.1(r1–r2) of both C and IC strains was tested in miracidia larvae exposed during 4 h to the drug. A Friedman non-parametric test was performed to test the significance of the TSA effect (Figure S1). We observed a statistically significant increase in transcription of groups 1 and 2 after TSA treatment in the IC strain only (p-value = 0.05). This indicates that changes in histone acetylation correlate with increased expression for SmPoMuc group 1 and 2 in the IC strain and has no effect in the C strain. Control genes were also tested for their response to TSA in order to determine that its effect was not pleiotropic. No effect of TSA was observed for these genes (GAPDH, Smp_011030, Smp_152710.1, Smp_054160, Smp_158110.1, GST.B, Glyaxalase, data not shown).
Since the TSA treatment influences overall histone acetylation, it could not be excluded that the observed effect is an indirect one and that SmPoMuc expression control is posttranscriptional and/or posttranslational such as selective RNA or protein degradation. We reasoned that in the offspring of crosses between the IC and C strains transcriptional control would produce an additive pattern of SmPoMuc proteins, while control by selective degradation of gene products would produce a subtractive pattern. Western blots show that in miracidia that are produced from crosses between the strains an additive pattern of the C and IC specific bands can be observed (fig. 4). This indicates that regulation operates at the transcriptional and not the post-transcriptional level and further supports the view that chromatin structure plays a role in the generation of specific SmPoMuc profiles for each strain.
Since all our experiments had delivered results in favour of a difference in chromatin structure of the SmPoMuc locus between strains, we decided to investigate the chromatin status in these loci. The occurrence of DNA methylation in S.mansoni is currently debated [28][29]. To test for DNA methylation in the promoter region of SmPoMucs we performed bisulfite genomic sequencing of DNA from miracidia using in-vitro methylated DNA as a positive control. We did not detect any methylated cytosine in the target region while 98% of the CpGs of in-vitro methylated DNA scored methylation positive. Our results are in line with earlier results showing that DNA methylation is rare from genes in S.mansoni [29][28]. We then performed Chromatin ImmunoPrecipitation (ChIP) experiments to check for histone modifications in the promoter regions. Due to the high similarity between the different groups of SmPoMuc promoters, ChIP-qPCR (quantitative Polymerase Chain Reaction) analysis was possible only in degenerate regions. Therefore, the chromatin structure analysis was performed on the promoter regions of SmPoMuc groups 1, 3.1 and 3.1(r1–r2). ChIP experiments were performed using an antibody that recognised Histone 3 acetylated on lysine 9 (H3K9Ac) and Histone 3 tri-methylated on lysine 4 (H3K4Met3) which are euchromatic marks and an antibody that recognised H3 tri-methylated on lysine 9 (H3K9Met3), which is a heterochromatic mark. Immunoprecipitation with the antibody that targets H3K4Met3 did not show any enrichment in the SmPoMuc region tested for either the IC or C strains whereas controls, αTub (Smp_090120.2) and 28S (Z46503.1) were positive (data not shown). The H3K4Met3 mark is usually very sharp and difficult to localise by target approach.. Both SmPoMuc group 1 and 3.1(r1–r2) from the IC strain displayed a higher level of H3K9Ac compared to the C strain (fig. 5). Consistent with this result, the C strain displayed a higher level of the heterochromatic mark (H3K9Met3) for group 1 and 3.1(r1–r2). These results have been obtained with several generations of the parasite, demonstrating that the phenotype is transmitted to the next generation.
In the IC strain, epigenetic marks showed differences among SmPoMuc groups 1, 3.1 and 3.1(r1–r2) (Figure S2). The promoter of group 3.1(r1–r2) is the most acetylated and the least heterochromatic. This result is consistent with expression analysis after TSA treatment where no effect of TSA was observed for the expression of group 3.1(r1–r2). This absence of an effect of TSA may be explained by the fact that acetylation on this promoter is already saturated and cannot be further increased as previously observed for H4 acetylation in the promoter region of HDAC1 in S. Mansoni [26].
The chromatin status in the promoter sequence of SmPoMuc groups 1, 3.1 and 3.1(r1–r2) was also investigated in the IC strain in cercaria and adults where SmPoMuc genes are not expressed. The level of the heterochromatic and euchromatic marks was the same as in miracidia and this level was maintained through several generations (Figure S3).
The host-parasite arms race determines that variability-generating processes are crucial for survival on both sides of the interaction (red queen hypothesis, [2]). The mechanisms that are responsible for these (heritable) phenotypic variations are a current and fundamental question in evolutionary biology. Traditionally, random genetic changes are seen as the sole source of phenotypic variation. But the picture is probably more complex: heritable adaptive phenotypic shifts could be partly controlled by epigenetic factors that were underrated until recently [30], [31]. A high rate of heritable epigenetic changes would generate phenotypic variation, which in turn could allow a rapid response to selection pressures [13]; [32]. This could allow for a transient and efficient response to changes in the environment, and could subsequently be followed by stabilization through genetic changes [33], [34]. Epigenetic modifications affect the transcription status of a gene in a heritable way without changes in the DNA sequence [14], [15], [16] and epigenetic information can be based on a chromatin marking system. Chromatin exists either as a relaxed structure that is permissive to gene expression and is called euchromatin, or as a condensed structure that is typically silent and is called heterochromatin [35]. Therefore, these different chromatin states alter gene expression and, ultimately, influence phenotypic outcomes without changes to the DNA sequence. The evolutionary implications of epigenetic inheritance systems and their potential link to stress-induced phenotypic variation have been discussed in several models [36], [37], [38], [31], [39], [40], [41] as well as in the specific context of host-pathogen interaction [42].
While it is clear now that induced epigenetic modifications are heritable [43], there are very few reports that show that epigenetic events lead to modification of gene expression profiles, production of new phenotypes and adaptation to the environment [44].In the present work, we addressed the question of the relative importance of genetic and epigenetic differences between two strains of S. mansoni that show clear differences in an ecological important adaptive trait: the capacity to infect their intermediate host. We had previously identified the SmPoMuc genes as surface molecular markers important for host compatibility. These markers encode mucins that display an extraordinary level of polymorphism, although they are produced from a relatively small number of very similar genes.As we had shown that nucleotide differences in the coding region could not explain differences in transcription, we focused therefore on the promoter regions in the present work. Our comparative survey of sequence variation in the different groups of SmPoMuc gene family from IC and C strains revealed a high level of conservation of the promoter sequences of SmPoMuc genes between the two strains. The molecular evolution of SmPoMuc promoters was uniform between all strains analysed, IC, C and NMRI. The sequence differences between the IC, C and NMRI strains within each group of SmPoMuc promoter were small, and the number of substitutions between the IC and C strains was equal or slightly higher than in the monomorphic single-copy gene SmFTZ-F1 and consistent with sequence differences at 14 microsatellite loci. To assess whether substitutions between the two strains could have an effect on transcription, we searched for functional regions of the active promoters. None of the substitutions between the IC and C strains occurred in the TATA signal, putative transcription factor binding sites or TSS regions. The nucleotide differences between the two strains consisted of zero in group 2 to eight substitutions in group 3.1(r1–r2), resulting in net nucleotide substitutions per site similar or lower than the ones observed in presumably neutral SmPoMuc introns (Table 2). At the population level, our analysis of SmPoMuc group 1 promoters in the IC and C strains revealed very low allelic and nucleotide variability within strain and high allele frequency differences between the IC and C strains due to fixed substitutions. All individuals were homozygotes at SmPoMuc group 1 promoter, similarly to the genotypes at 14 microsatellite loci, suggesting that S. mansoni strains present genome-wide homozygosity. Both strains are characterised by a high significant inbreeding coefficient, resulting from high clonality in the two strains [25], which may have arisen because of the bottleneck due to the strain maintenance in laboratory conditions. Despite the lack of diversity within strains, alleles fixed in each strain for the SmPoMuc group 1 promoter and nine microsatellites were different, resulting in high genetic differentiation between the two strains as estimated by FST. This contrasted with the promoter of the single-copy gene SmFTZ-F1 and six microsatellite loci, which displayed a unique sequence common to the two strains.
In summary, our analysis of the genetic information shows that (i) both strains are genetically monomorphic, including the SmPoMuc promoter regions, (ii) both strains are different in terms of alleles, i.e. they do not share the same alleles, but (iii) these alleles are similar or display low number of base substitutions (outside functional regions). It could be argued that the small nucleotide differences observed between the two strains are sufficient to provoke modulation of histone modification. Such a leverage effect of SNPs cannot be excluded but has so far not been observed in heavily studied models such as human, Drosophila melanogaster and Arabidopsis thaliana. It could also be the case that strain-specific loci exist that regulate the chromatin structure of the SmPoMuc genes in trans or in cis (upstream of the minimal functional promoter). However, previous work has compared the proteomes of both C and IC strains [7] and did not pinpoint any major regulators that may be responsible for such a phenotype. In view of these results, we argue that genetic differences between sequences within each group of SmPoMuc promoters were unlikely to solely dictate the high level of variation in SmPoMuc transcription and compatibility polymorphism phenotypes.
We therefore further investigated the epigenetic basis for such phenotypes. TSA treatment was used to study the impact of overall acetylation status of histones on miracidia larvae where SmPoMuc is expressed. This drug is known to be a specific histone deacetylase (HDAC) inhibitor and has been previously shown to influence phenotypic traits in S. mansoni [13]. A dose dependant effect of TSA was observed for SmPoMuc expression (all groups taken together) in the IC strain whereas no effect was observed in the C strain. This result suggests that the acetylation status of histones in the promoter sequences is differentially regulated between the IC and C strains. HDACs seem to play a more prominent role in regulating the acetylation level in the IC strain that allowed us to pinpoint a TSA effect in this strain. More specifically, we report a TSA effect on groups 1 and 2 of the IC strain whereas no effect is observed for group 3.1(r1–r2) for which acetylation is the strongest. This also suggests that a differential regulation by HDAC exists between the SmPoMuc groups in the same strain. Further support for regulation on transcriptional level comes from a crossing experiment in which strain hybrids were produced. Western blots show that in the hybrids, both the C-specific and the IC-specific SmPoMucs are expressed. One could hypothesize that production of SmPoMuc variants is due to post-transcriptional strain-specific regulation. In this scenario all genes would be expressed, but the gene products would be processed in a strain-specific form. In the hybrids, in which the hypothetical post-transcriptional regulation pathway for both strains is present, we should have seen a diminution of non-IC and the non-C SmPoMuc forms. This was not the case. In summary, all lines of evidence point towards a chromatin-based regulation of SmPoMuc expression.
The chromatin configuration was further investigated by ChIP analysis using antibody that recognises heterochromatic and euchromatic marks. ChIP results clearly demonstrate that different epigenetic marks occur on the SmPoMuc promoter of group 1 and group 3.1(r1–r2) between the IC and C strains likely resulting in a different chromatin configuration. On these loci, chromatin is indeed more enriched in H3 acetylated on lysine 9 in the IC compared to the C strain and less enriched in the opposite mark, H3 trimethylated on lysine 9. Therefore, the local chromatin structures differ between the two strains for groups 1 and 3.1(r1–r2) and are consistent with expression data as stronger acetylation correlates with enhanced expression. Importantly, H3K9Met3 and H3K9Ac marks are maintained through the cercarial and adult stages at which the genes are not expressed. This persistence of the chromatin mark throughout other stages of the S. mansoni life cycle is a crucial result as this is a necessary condition for the epigenetic mechanism to act as a heritable trait. Similarly, several CVGE genes of P. falciparum that display a bistable chromatin state to regulate their expression in the intraerythrocytic stages have been shown to maintain their epigenetic marks during trophozoite and schizont stages, the other asexual stages at which these genes are not expressed [45].
It is now established that the phenotype is not onlya product of genetic processes, but expression of an ensemble that is composed of genetic and epigenetic components. Others and we have proposed that this additional system allows for rapid adaptive evolution without necessarily changing the genotype initially. A theoretical framework for this model was provided by Pal and Miklos (1999) [17], and more recently by Klironomos, Berg and Collins (personal communication). Essentially, these authors propose that a higher rate of random changes in epigenetic marks compared to genetic mutations transmitted from one generation to the next in a population generates increased phenotypic variations that can be selected for if the environment changes. In this sense, epigenetic modifications provide a source of rapid and reversible phenotypic variation and are therefore expected to be major players in the context of host-pathogen interaction where selection pressures are strong and evolution is fast [42], [18]. In this context, epigenetic based events to generate variability of surface antigens of parasites perfectly matched to this theory. For exemple, VSP diversification of Giardia sp. likely occurs by epigenetic mechanisms involving the histone acetylation status [46] and/or RNAi [47]. Chromatin remodeling proteins and histone modifications have been shown to play a role in VSG expression site silencing [48] and Plasmodium Var diversification is orchestrated by multiple epigenetic factors including monoallelic transcription at separate spatial domains at the nuclear periphery, differential histone marks on otherwise identical var genes, and var silencing mediated by telomeric heterochromatin [49]. On the host side, genetic and epigenetic crosstalks have been previously demonstrated in the generation of a high level of polymorphism of the receptors of the adaptative immune system [50], [51]. Therefore, all these variability generating mechanisms are examples of local adaptation to an ever-changing environment where epigenetic based events are used to rapidly produce new phenotypes and potentially induce rapid evolutionary change of genes that are under pressure. In our work, we show that two population of S. mansoni with distinct phenotypic traits, in particular their compatibility with a reference host, show low nucleotide differences in both coding sequence and promoters of SmPoMuc but high epigenetic differences in the promoter regions. Both parasite populations are in a situation where the fitness value of genetically encoded phenotypes has not changed significantly, but epigenetic variations have produced phenotypic variants that are adapted to different environments (compatible hosts).
While we have compared only South American strains, our observations suggest a scenario for the adaptation of S. mansoni to the new world host: in the 15th–16th century the ancestral strain of contemporary strains IC and C migrated via the slave trade from Africa to the West Indian Islands and the South American continent, respectively [6]. There, they had to adapt to a new intermediate host. The initial bottleneck resulting from the migration of only a limited number of parasites and the expected strong selective pressure acting on both genetic and epigenetic variants of the key-molecules for compatibility with the new snail hosts, SmPoMucs, may have significantly reduced genetic and epigenetic variation in the newly formed laboratory IC and C strains compared to the ancestral strain. Now, it is likely that epigenetic variation retained from the ancestral strain and the higher rate of occurrence of epigenetic changes in subsequent generations, rather than the strain genetic variation, enabled the parasite to adapt rapidly to their host and new environment. A conundrum with the “epigenetic mutation system first” hypothesis is that epigenetic information concerns the transcriptional activity of a gene but not its coding potential, in other words, a gene can be switched on and off by the surrounding chromatin but the resulting protein cannot be changed. Loss of function of genes can easily be imagined through an epigenetic mechanism, but for gain of function a complex inhibitor-based mechanism would be necessary. The classical Ohno hypothesis of gene duplications as way to provide material for evolution [52] could deliver a solution. Rodin and Riggs have shown that duplicated genes have a tendency to be heterochromatic [53]. It is interesting to note that the SmPoMuc proteins, essential for host compatibility, are encoded by duplicated genes. Our analysis shows that the duplication events predate the IC/C separation and occurred in the strain's common ancestor, i.e. gene duplication was not a result of divergence of the two strains. We postulate that SmPoMuc duplicated genes provide an additional system for phenotypic variation. Duplicated genes are randomly modulated in their relative transcriptional activity through chromatin structure changes as evidenced by our current and previous results [13], resulting in new combinations of expressed SmPoMuc genes and subsequent increased phenotypic variation. If the parasite encounters new intermediate hosts, the probability for the phenotypes to match is increased, thus allowing for adaptive evolution.
Therefore, our work shows that in a gene family that codes for an adaptive phenotypic trait, epigenetic changes are more important than genetic changes. This finding provides support for theoretical models of adaptive evolution in which epimutations occur more rapidly than mutations.
The French Ministère de l'Agriculture et de la Pêche and French Ministère de l'Education Nationale de la Recherche et de la Technologie provided permit A 66040 to our laboratory for experiments on animals and certificate for animal experimentation (authorization 007083, decree 87–848) for the experimenters. Housing, breeding and animal care followed the national ethical requirements.
A compatible strain (C) (Brazilian strain), an incompatible S. mansoni strain (IC) (Guadeloupean strain), the reference NMRI S. mansoni strain (Puerto Rican strain) and a reference mollusc strain (B. glabrata BRE isolated from Brazil) were used in this study. For initial breeding, each strain was maintained in its sympatric (compatible) B. glabrata strain, and in hamsters (Mesocricetus auratus) as described previously [54]. Adult worms and miracidia were obtained as described previously [8].
Individual B. glabrata snails were infested with a single miracidium to obtain cercarial clonal populations. Subsequently the sex of the cercariae was determined as described previously [55]. Strain hybrids of S. mansoni were produced by infection of mice or hamster with 300 cercariae: 200 males from a clonal cercarial population combined with 100 females from another clonal cercarial population. Different combinations of parental cercariae of the IC and C strains were used, thus generating worm couples in which the male is C and the female is IC or vice versa. Eggs were recovered from infected (3 to 6) mice (Mus musculus) 12 weeks post-infection. Livers were collected and homogenized, and eggs were filtered and washed. Miracidia were allowed to hatch in spring water and were concentrated by sedimentation on ice for 15 minutes.
1000 Miracidia were incubated in 350 µl UTCD buffer (ultrapure urea 8 M, Tris 40 mM, DTT 65 mM, CHAPS 4%), two hours at room temperature. The extract was cleared by centrifugation for 30 minutes at 1500 g, and the supernatant was collected. Total proteins (5 µg per sample) were separated by 10% SDS-PAGE gel electrophoresis before being blotted on a nitrocellulose membrane (Trans-Blot turbo, Bio-Rad). The membrane was blocked with 5% non-fat dry milk in TBST (TBS buffer containing 0.05% tween 20) one hour at room temperature, and incubated with the primary antibody “anti-SmPoMuc” diluted 1/500 in TBST for 90 minutes at room temperature. This rabbit polyclonal antibody was produced according to standard procedures and was shown to recognise all the SmPoMuc groups [9]. Then, the membrane was incubated with secondary antibody (peroxidase conjugated, purified anti-rabbit IgG) diluted 1/5000 in TBST for 1 hour. After washing 3 times for 10 minutes in TBST, the detection was carried out using the ECL reagents and the ChemiDoc MP Imaging system – BioRad).
We searched for sequences of promoter regions of SmPoMuc genes in the genomic database of the S. mansoni NMRI strain (assembly version 3.1) using BLAST searches. Contigs matching to SmPoMuc genes were assembled with the Sequencher software (Gene Codes Corporation) to recover the sequences of the promoter regions of the genes. From the BLAST search and manual assemblage of relevant contigs, scaffolds of promoter regions were constructed for the different SmPoMuc genes in groups 1–4. Primers were designed on these contigs to amplify the promoter regions of the different SmPoMuc genes in the C and IC strains of S. mansoni. The DNA templates to generate PCR products were either genomic DNA (C and IC strains), a BAC library (NMRI strain) or a phage library (IC strain). Genomic DNA was extracted from adult worms as described previously [8]. The production of the phage library is described below. Promoter regions were amplified using the Advantage 2 PCR Enzyme System (Clontech) (Table S1 for primer sequences, amplified fragment lengths and sources of DNA). PCR products were either cloned into pCR-XL-TOPO (TOPO TA Cloning kit for sequencing, Invitrogen) and plasmid DNA was purified using the Wizard Plus SV Miniprep DNA purification system (Promega), or sequenced directly. We sent PCR amplificons or plasmids containing the promoter regions to GATC (GATC Biotech, Germany) for cycle sequencing in both directions and performed primer walking up to 2.0 kb upstream of the transcription start sites (TSS) of SmPoMuc genes (for primer sequences see Table S2). We checked trace data and aligned nucleotide sequences manually using the BioEdit software. We scanned the promoter sequences for putative regulator binding sites using the web based interface Program NSITE (Softberry Inc.) (http://linux1.softberry.com/berry.phtml?topic=nsite&group=programs&subgroup=promoter).
The presence of multiple copies of some SmPoMuc genes sometimes prevented the amplification of a single copy and assembly of a gene with its corresponding promoter. To address this problem, we constructed a phage library of the IC strain using the Lambda Fix II vector system from Stratagene. The expected size of inserts was 15 to 23 kb corresponding to the size range of SmPoMuc genes (10–30 kb). Details of the construction of the phage library and screening are available at http://methdb.univ-perp.fr/epievo/. Genome coverage of the library was four fold. The library was screened for SmPoMuc genes using as a probe UR1, a highly conserved intronic sequence spanning the region between two repeat units of the SmPoMuc genes [8]. The probe was labeled with the DIG High Prime DNA Labeling and Detection Starter Kit II using Random primed DNA labeling with digoxigenin-dUTP, alkali-labile and chemiluminescence with CSPD (Roche). Screening was performed according to the manufacturer's instructions. Secondary and tertiary screening rounds were performed with the same probe to isolate individual phage clones. Phages that scored positive for SmPoMuc repeat units were screened by PCR using a combination of diagnostic primers for each group of SmPoMuc genes (Table S2) with the Advantage 2 PCR Enzyme System (Clontech). Selected phages were subsequently purified and used as templates to PCR amplify SmPoMuc group 3.1(r1–r2) as described in the section “PCR screening for promoters of SmPoMuc genes, cloning and sequencing”.
We used DnaSP to characterise promoter sequence variation within and between groups of SmPoMuc promoter sequences as the number of polymorphic sites, number of mutations between strains, net number of substitutions per site between strains and between groups of SmPoMuc promoter sequences.
We amplified and sequenced the promoter region of the SmFTZ-F1 gene. This gene encodes the nuclear receptor fushi tarazu-factor 1alpha and its promoter has been fully characterised [24] in 1 and 2 individuals of S. mansoni strains IC and C, respectively, from genomic DNA with primers Smftzf1-F (5′-ATGAGATGTTTCTGAGCAATGGC-3′) and Smftzf1-R (5′-TCTTCTCGTAGCTGAATCTGACC-3′) using the Advantage 2 PCR Enzyme System (Clontech). PCR amplicons were then sequenced and analysed for sequence variation and gene diversity as described above.
We amplified 996 kb of the SmPoMuc group 3.1(r1–r2) promoter and 1002 kb of the SmPoMuc group 3.1 promoters. These sequences are located just upstream of the transcriptional start site and have been amplified from the IC strain. These sequences were amplified using primers containing SacI and BamHI restriction sites (Table S2). The PCR product was gel-purified (Wizard SV gel and Clean-Up system,Qiagen), digested with both restriction enzymes and cloned into a SacI and BamHI digested pEGFP-1 reporter vector with T4 DNA ligase (New England Biolabs). The construct was verified by sequencing both DNA strands. Plasmids pEGFP-1 and pCMV-EGFP driving EGFP expression, under the control of the CMV-promoter, were used as negative and positive controls in the transfection assay.
A 3.3 kb region of the SmPoMuc group 1 gene promoter region was amplified using primers SmpomucpromGP3.1.f2 and BR2 (Table S1) in individuals of each of S. mansoni IC and C strain. The PCR products span from 1.8 kb upstream of the TSS to the first repeat unit of the SmPoMuc gene and cover the promoter region. 1.4 kb of the promoter region was sequenced for 20 and 18 individuals of the IC and C strains, respectively, by primer walking (Table S2). We used Arlequin 3.1 to characterise SmPoMuc group 1 promoter diversity within the two strains as the expected unbiased gene diversity, the nucleotide diversity, corrected for sample size and incorporating nucleotide information [62]. We tested for sequence variation between the two strains using population comparisons and differentiation in Arlequin 3.1. Estimations incorporated Tamura-Nei distances between sequences and allele frequencies (Nei's Φ-estimator of FST). The significance of genetic differentiation was tested by permuting the alleles among all samples 2,000 times. We also estimated the inbreeding coefficient in each strain using f and genetic differentiation between the two strains using FST estimator θ ([63], incorporating allele frequencies only). Inbreeding coefficients and genetic differentiation for departure from the null hypothesis (f = 0, θ = 0) were tested using 2,000 permutations in GENETIX 4.05 [64].
Nineteen individuals of each of the IC and C strains were genotyped using 14 microsatellite loci [25]. We estimated genetic diversity of microsatellite loci as the mean number of alleles per locus (A) and observed and expected unbiased heterozygosities (HO and Ĥ? respectively) under the assumption of Hardy–Weinberg equilibrium [62]. We estimated the inbreeding coefficient f in each strain, genetic differentiation between the two strains RST estimator [65], [66] and the FST estimator θ as above.
Trichostatin-A (TSA) (invivoGen met-tsa-5) was dissolved in ethanol to 20 mM and added to the 1000 IC or C miracidia pool at 20 µM and 200 µM during 4 h. We had shown previously the effect of TSA at these concentrations on development, morphology, mobility and gene expression without any cytotoxicity for the larvae [13], [27]. To the untreated control, an equal volume of ethanol was added (mock treatment). After 4 h, metamorphosis arrest was observed for larvae treated with TSA at 200 µM as expected for a positive effect with this drug [27]. Miracidia were then spun down at 12,000 g during 5 min and suspended in 100 µl of lysis buffer (Dynabeads mRNA DIRECT Micro kit, Dynal Biotech) in RNase-free tubes and stored at −80°C. Messenger RNAs were extracted using the Dynabeads mRNA isolation Kit according to the manufacturer's instructions. mRNA poly-A residues were eluted from the surface of the paramagnetic beads by a final denaturation step of 10 min at 75°C in 20 µl of Tris-HCl 10 mM. cDNA synthesis was carried out using 10 µl of mRNA in a final volume of 20 µl according to manufacturer's instructions (0.5 mM dNTPs, 0.01 mM DTT, 1× first strand buffer, 2 U RNase out, 10 U SuperScript II RT (Invitrogen) during 50 min at 42°C). After reverse transcription, the cDNAs were purified with the PCR clean-up system (Promega) and eluted into 100 µl 10 mM Tris/HCl (ph 7.5).
Specific primers for qPCR from groups 1, 2 and 3.1(r1–r2) were designed based on sequence alignment performed on cDNA variant representative of each group (Table S2). Their specificity was tested using as template a plasmid in which a cDNA variant of group 1, 2 or 3.1(r1–r2) was cloned. Group 4 genes contain a STOP codon in exon 8 of the gDNA sequence and their cDNA has never been detected. Therefore, transcripts of the group 4 genes were not targeted in this study. Other subgroups were not studied as it was not possible to design specific primers to amplify them. qPCR amplifications were performed as described below. Results were normalised with the αTub gene. The 2ΔCt value was calculated. Statistical tests were performed on at least 3 different biological samples.
Native chromatin immunoprecipitation was performed as described before [67]. Briefly, antibodies against histone isoforms were used to precipitate chromatin in miracidia from IC and C strains (Table S3). DNA was extracted from the precipitated complex and analysed by qPCR using specific primers of SmPoMuc groups 1, 3.1 and 3.1(r1–r2). Primers specifically targeting these genes were designed based on sequence alignment of SmPoMuc promoter sequences (Table S2). We tested their specificity using as templates plasmids with promoters of group 1, 3.1 or 3.1(r1–r2). It was not possible to design primer sets that would hybridize specifically to the promoter sequences of the other groups or subgroups because conservation in the sequences resulted in cross-amplification between these groups. The amount of target DNA recovered in the immunoprecipitated fraction was quantified by calculating the percent input recovery (% IR) normalised with the percent input recovery obtained with a reference locus (αTub) as previously described [67].
Bisulfite genomic sequencing was carried out as described in [68]) on gDNA extracted from miracidia from the NMRI strain. Amplification was performed using primers BS.IC-1-Group1/1111-1715.48f GATATGTTTTAAGAAGTAGAAAAGAATATT, BS.IC-1-Group1/1111-1715.508r ATAAAAATTTTACAACCACCTACTC and BS.IC-1-Group3.1/421-952.29f ATTGTTTTTTTTAATTTTAGATATGTTTTA and two rounds of PCR. 1 µl of each PCR products were cloned into the TOPO TA vector (Invitrogen) and sequenced. In-vitro methylation with M.SssI (NEB) was done as recommended by the supplier. A total of 20 sequences (7 M.SssI treated positive controls and 13 target miracidial gDNA) were aligned with the genomic sequence from GenBank (Bioedit) to visualise the sites of methylated cytosine.
qPCR amplifications were performed with 2.5 µl of immunoprecipitated DNA or cDNA in a final volume of 10 µl on a LightCycler® 480 II Real Time instrument (1.5 µl H20, 0.5 µM of each primer, 5 µl of master mix). The following protocol was used: denaturation, 95°C for 10 minutes; amplification and quantification (40 times): 95°C for 10 seconds, 60°C for 10 seconds, 72°C for 20 seconds; melting curve, 65–97°C with a heating rate of 0.11°C/s and continuous fluorescence measurement, and a cooling step to 40°C. For each reaction, the cycle threshold (Ct) was determined using the “2nd derivative” method of the LightCycler® 480 Software release 1.5. PCR reactions were performed in duplicate and the mean value of Ct was calculated. Correct melting curves were checked using the Tm calling method of the LightCycler® 480 Software release 1.5. The amplification of a unique band was verified by electrophoresis separation through a 2% agarose gel for each qPCR product.
JQ615951–JQ615966.
|
10.1371/journal.pcbi.1001072 | Ribavirin-Induced Anemia in Hepatitis C Virus Patients Undergoing Combination Therapy | The current standard of care for hepatitis C virus (HCV) infection – combination therapy with pegylated interferon and ribavirin – elicits sustained responses in only ∼50% of the patients treated. No alternatives exist for patients who do not respond to combination therapy. Addition of ribavirin substantially improves response rates to interferon and lowers relapse rates following the cessation of therapy, suggesting that increasing ribavirin exposure may further improve treatment response. A key limitation, however, is the toxic side-effect of ribavirin, hemolytic anemia, which often necessitates a reduction of ribavirin dosage and compromises treatment response. Maximizing treatment response thus requires striking a balance between the antiviral and hemolytic activities of ribavirin. Current models of viral kinetics describe the enhancement of treatment response due to ribavirin. Ribavirin-induced anemia, however, remains poorly understood and precludes rational optimization of combination therapy. Here, we develop a new mathematical model of the population dynamics of erythrocytes that quantitatively describes ribavirin-induced anemia in HCV patients. Based on the assumption that ribavirin accumulation decreases erythrocyte lifespan in a dose-dependent manner, model predictions capture several independent experimental observations of the accumulation of ribavirin in erythrocytes and the resulting decline of hemoglobin in HCV patients undergoing combination therapy, estimate the reduced erythrocyte lifespan during therapy, and describe inter-patient variations in the severity of ribavirin-induced anemia. Further, model predictions estimate the threshold ribavirin exposure beyond which anemia becomes intolerable and suggest guidelines for the usage of growth hormones, such as erythropoietin, that stimulate erythrocyte production and avert the reduction of ribavirin dosage, thereby improving treatment response. Our model thus facilitates, in conjunction with models of viral kinetics, the rational identification of treatment protocols that maximize treatment response while curtailing side effects.
| The treatment of HCV infection poses a major global health-care challenge today. The current standard of care, combination therapy with interferon and ribavirin, works in only about half of the patients treated. Because no alternatives are available yet for patients in whom combination therapy fails, identifying ways to improve response to combination therapy is critical. Increasing exposure to ribavirin does improve response but is associated with the severe side-effect, anemia. One way to maximize treatment response therefore is to increase ribavirin exposure to levels just below where anemia becomes intolerable. A second way is to supplement combination therapy with growth hormones, such as erythropoietin, that increase the production of red blood cells (erythrocytes) and compensate for ribavirin-induced anemia. Rational optimization of combination therapy thus relies on a quantitative description of ribavirin-induced anemia, which is currently lacking. Here, we develop a model of the population dynamics of erythrocytes in individuals exposed to ribavirin that quantitatively describes ribavirin-induced anemia. Model predictions capture several independent observations of ribavirin-induced anemia in HCV patients undergoing combination therapy, estimate the threshold ribavirin exposure beyond which anemia becomes intolerable, suggest guidelines for the usage of growth hormones, and facilitate rational optimization of therapy.
| 130–170 million people worldwide are currently infected with hepatitis C virus (HCV) [1]. Over 70% of HCV infections become chronic and if untreated may lead to cirrhosis and hepatocellular carcinoma, necessitating liver transplantation [1]. The standard of care for HCV infection involves combination therapy with pegylated interferon and ribavirin [2]. Ribavirin alone does not elicit a lasting antiviral response [3]–[6], yet it substantially improves treatment response in combination with interferon [7]–[11]. For instance, whereas ∼29% of the patients treated with interferon exhibited a sustained virological response (SVR), the response rate increased to ∼56% upon addition of ribavirin [8]. Ribavirin, however, is associated with the side-effect, hemolytic anemia, which often renders therapy intolerable [4], [12]–[15]. With the standard ribavirin dosage of 1000–1200 mg/day, 54% of the patients treated experienced a decline in the hemoglobin (Hb) level of over 3 g/dL, and 10% of the men and 7% of the women treated experienced an Hb decline of over 5 g/dL (normal Hb range: 14–16 g/dL) [15]. This drop in Hb often necessitates a reduction of ribavirin dosage, which significantly compromises treatment response [7], [13], [14], [16]. The probability of achieving SVR is estimated to decrease from ∼65% to ∼45% when ribavirin dosage is reduced from ∼15 mg/kg to ∼7 mg/kg of body weight, in combination with pegylated interferon at 1.5 µg/kg of body weight [7]. Patients receiving fewer than 60% of the planned ribavirin doses had lower response rates [14], indicating that lower cumulative ribavirin exposure results in poorer treatment response [13], [16]. The rates of relapse of infection following the end of treatment also increased upon lowering ribavirin dosage [14], [16]. In a recent clinical trial where interferon was employed with telaprevir, a promising new inhibitor of HCV protease, response rates were lowest in patients who were not administered ribavirin [17], underscoring the importance of ribavirin in achieving SVR.
Alternatives for patients who do not respond to combination therapy do not exist yet [2], [18]. Significant efforts are underway therefore to identify treatment protocols that maximize response rates to combination therapy while curtailing side-effects [16], [19]–[25]. A particularly promising strategy is to supplement combination therapy with growth hormones, such as erythropoietin, that stimulate erythropoiesis and thus avert the reduction of ribavirin dosage, potentially improving treatment response [26]–[31]. The predominant mechanism(s) of the anti-HCV activity of ribavirin remain to be established [32]–[34]. Mathematical models of viral kinetics have been developed that describe the antiviral activity of interferon and the enhancement of treatment response rates due to ribavirin, and are being extended to predict the impact of new antiviral drugs [34]–[42]. Ribavirin-induced anemia, on the other hand, remains poorly understood [13], [24], [25], [43]–[47] and precludes rational optimization of combination therapy.
Here, we construct a mathematical model of the population dynamics of erythrocytes that quantitatively describes ribavirin-induced anemia and informs future strategies for improving outcomes of combination therapy. Model predictions capture experimental observations of the accumulation of ribavirin in erythrocytes and the ensuing Hb decline in HCV patients following the onset of combination therapy, estimate the enhanced turnover rate of erythrocytes during therapy and the threshold ribavirin exposure beyond which anemia is intolerable, present guidelines for the optimal usage of growth hormone supplements, and provide a framework, in conjunction with models of viral kinetics, for rational optimization of combination therapy.
Prior to the onset of treatment with ribavirin, the population of erythrocytes (RBCs) in an HCV infected individual is constant; a balance exists between RBC production and death (Fig. 1). Following the onset of treatment, ribavirin administered orally gets rapidly transported from the plasma to RBCs, where it is phosphorylated to its mono-, di- and tri-phosphate analogs (RMP, RDP, and RTP) [48]. Phosphorylated analogs are neither easily metabolized nor transported out of RBCs [48]. Consequently, ribavirin accumulates inside RBCs in the form of its phosphorylated analogs; the total intracellular concentration of ribavirin can be >100-fold its extracellular concentration [47]. This dramatic accumulation of ribavirin may induce oxidative damage and result in enhanced extra vascular death of RBCs [12]. Indeed, RBC lifespan decreased from 107±22 d in HCV patients not exposed to ribavirin to 39±13 d in HCV patients undergoing treatment with ribavirin [49], [50]. The shortened RBC lifespan creates an imbalance between RBC production and death and results in a decline in the RBC population. Accordingly, Hb levels drop and patients become anemic. We construct a mathematical model to describe this dynamics of ribavirin-induced anemia (Methods).
We consider a recent study of the time-evolution of and in 19 Japanese patients following the onset of combination therapy [47]. In this latter study, no reduction of ribavirin dosage is reported. The patients were divided into two groups based on whether <1000 µM (7 patients) or >1000 µM (12 patients); the data are reported as the average within each group. We fit model predictions of and to the data of the former 7 patients using , , and as adjustable parameters. (Interferon may also induce anemia, but does so to a much smaller extent than ribavirin [13]. We therefore assume that the Hb decline in patients undergoing combination therapy is primarily due to ribavirin.) We fix the remaining parameters based on previous studies or from analysis of independent experiments (Methods). Model predictions provide good fits to the data and yield estimates of , , and (Fig. 5A). The fits suggest that our model is able to describe the underlying dynamics of ribavirin-induced anemia in HCV patients.
Interestingly, with the same parameter values, our model captures changes in Hb and from the other 12 Japanese patients, as well as an independent data set of Hb decline in another group of HCV patients undergoing combination therapy [29] (Fig. 5B), validating our best-fit parameter estimates. Further, with the same parameter values, we estimate that the RBC lifespan is 38 days (95% CI: 19–55 days) in Japanese patients with <1000 µM and 33 days (95% CI: 14–53 days) in Japanese patients with >1000 µM. These estimates of the RBC lifespan are in close agreement with independent estimates, 39±13 days, from measurements of alveolar carbon monoxide [49], [50], presenting another successful test of our model. Finally, we find that our predictions of the dependence of on and using the same parameters above are also in agreement with observations in the Japanese patients [47] (Fig. 5C,D). Our model thus presents a robust description of ribavirin-induced anemia in HCV patients undergoing combination therapy.
Our model has several clinical implications. First, it enables estimation of the threshold ribavirin exposure beyond which anemia is intolerable. Current treatment guidelines recommend a reduction of ribavirin dosage when Hb decreases below 10 g/dL. We apply our model to predict as a function of . We find that on average (when = 14.4 g/dL) <10 g/dL when >13 µM (Fig. 6A). Thus, steady state plasma concentrations above 13 µM would render ribavirin therapy intolerable. While the dependence of the peak plasma concentration on dosage following a single ribavirin dose has been determined [48], the dependence of on dosage remains to be established. A description of the multiple dose pharmacokinetics of ribavirin, which also remains elusive [6], [34], [48], [52], [53], would establish the dosage corresponding to of 13 µM that would render ribavirin intolerable.
Second, when is above the threshold, our model allows estimation of the increase in RBC production, which may be achieved by administration of exogenous growth hormones such as recombinant erythropoietin, necessary to avert the currently recommended reduction of dosage. Because growth hormones also have side-effects [29], [54], one strategy is to use them at levels just enough to increase to 10–12 g/dL (rather than the pretreatment level), which renders ribavirin tolerable [16]. We apply our model to predict the level of RBC production necessary for achieving of 10–12 g/dL for different values of (Fig. 6B). Thus, when = 15 µM, RBC production rates of 8.44 and 10.2 million cells s−1 are necessary for ensuring of 10 and 12 g/dL, respectively. Increase in endogenous erythropoietin levels during therapy, also observed experimentally [45], [55], results in an enhanced production rate of 8.1 million cells s−1, which is 3.5-fold higher than the basal production rate (here 2.3 million cells s−1 in the absence of ribavirin) but inadequate to achieve the desired . Hormone supplements may be employed to provide the balance of 0.34 or 2.1 million cells s−1 increase in the RBC production rate to ensure of 10 or 12 g/dL, respectively. This deficiency in RBC production that hormone supplements must compensate increases with ribavirin exposure (Fig. 6B).
The ability to enhance treatment response rates renders ribavirin central to the treatment of HCV infection. Maximizing the benefit of ribavirin to patients requires striking the right balance between its antiviral activity and its treatment-limiting side-effect, hemolytic anemia. Rational approaches to therapy optimization thus rely on quantitative descriptions of both the antiviral and the hemolytic activities of ribavirin. Extant mathematical models predict the enhancement in treatment response due to ribavirin [34]–[42]. Ribavirin-induced anemia, however, remains poorly described and limits our ability to maximize treatment response. Here, we fill this gap by constructing a model of the population dynamics of RBCs that quantitatively describes ribavirin-induced anemia. By assuming that intracellular accumulation of ribavirin enhances RBC death rate in a dose-dependent manner, our model captures several independent observations of ribavirin-induced anemia in HCV patients undergoing combination therapy. In particular, our model predicts the dynamics of the accumulation of ribavirin in RBCs and the resulting decline of Hb in patients following the onset of therapy, estimates the reduced lifespan of RBCs during therapy, and describes inter-patient variations in the severity of anemia, thus presenting a robust description of ribavirin-induced anemia, which, in conjunction with models of viral kinetics, may facilitate identification of treatment protocols that maximize the impact of ribavirin in the treatment of HCV infection.
Our model has clinical implications. First, it allows estimation of the threshold ribavirin exposure beyond which ribavirin-induced anemia becomes intolerable. For instance, with model parameters that describe ribavirin-induced anemia in the patients we considered (Fig. 5), we estimate that steady state plasma ribavirin concentrations above 13 µM would render ribavirin therapy intolerable. Determining dosage levels corresponding to this steady state plasma concentration requires knowledge of the pharmacokinetics of ribavirin, which is currently lacking [6], [34], [48], [52], [53]. Ribavirin pharmacokinetics is peculiar because of an unusually long elimination phase that follows rapid absorption and distribution phases upon oral dosing [48]. Standard absorption-elimination models of drug pharmacokinetics are unable to describe this long elimination phase. Models that include additional compartments have been proposed to capture the three-phase pharmacokinetics of ribavirin [52], but the biological origin of these compartments remains unclear. An additional complication is that the half-life of the elimination phase increases from 79 h following a single dose to 274–298 h following multiple dosing [48], suggesting that parameters that describe single dose pharmacokinetics may not apply to multiple dose pharmacokinetics. In the absence of rigorous models of ribavirin pharmacokinetics, one may have to rely on empirical relationships between the dosage and the resulting steady state plasma concentration following multiple dosing (e.g., [56]) to establish the dosage that would ensure tolerability of ribavirin while maximizing treatment response.
Second, our model suggests guidelines for the usage of hormone supplements, such as erythropoietin, which enhance RBC production and improve the tolerability of ribavirin. For instance, we predict that when ribavirin accumulates to a plasma concentration of 15 µM, the associated enhanced RBC death rate elicits a natural response that increases RBC production 3.5-fold, from 2.3 to 8.1 million cells s−1. This response, however, is inadequate to suppress ribavirin-induced anemia adequately and renders ribavirin intolerable. We estimate then that growth hormone supplements must increase RBC production rate by an additional 0.34–2.1 million cells s−1 to render ribavirin tolerable. This compensation that hormone supplements must provide increases with ribavirin accumulation. Identifying the dosage of the growth hormones that induces the necessary RBC production requires knowledge of the dose-response relationships and of the pharmacokinetics of the growth hormones, which are yet to be fully elucidated [26]–[31].
Third, genetic variations that resulted in a deficiency in the enzyme inosine triphosphatase (ITPA) were recently found to protect HCV patients against ribavirin-induced anemia [51]. Deficiency in ITPA causes an increase in inosine triphosphate levels in RBCs, which is thought to interfere with RTP activity and thereby suppress the hemolytic potential of ribavirin. Because deficiency in ITPA is a clinically benign condition, therapeutic intervention to suppress ITPA presents a promising new strategy to curtail ribavirin-induced anemia without compromising the antiviral activity of ribavirin [51]. Our model may be adapted to inform the development of such an intervention strategy. In our model, the dependence of the death rate of RBCs on ribavirin accumulation, determined by Eq. (2) (Methods), would now be a function of the ITPA level. Thus, experiments that determine how variations in the ITPA level both in the absence and in the presence of ribavirin influence RBC lifespan would provide the necessary inputs for our model to account explicitly for the role of ITPA in ribavirin-induced anemia. The resulting model would enable determination of the minimal inhibition of ITPA necessary to maintain ribavirin-induced anemia within tolerable limits. Conversely, using information of the ITPA level intrinsic to a patient, the model can be applied to predict the maximum ribavirin dosage that the patient can tolerate, thus presenting an avenue for personalizing the treatment of HCV infection.
We consider the RBC population in an individual at time t following the onset of treatment with ribavirin (t = 0) (Fig. 1). RBCs produced at different times in the interval from 0 to t will have been exposed to ribavirin for different durations and accordingly have different intracellular levels of ribavirin. We define as the population of RBCs that contain ribavirin phosphorylated analogs, RXP, which comprises RMP, RDP, and RTP, at concentrations between and at time . is thus the number density of RBCs containing RXP at concentration C at time t. The time evolution of is governed by the following equation (Text S1)(1)
The first term on the right-hand-side in Eq. (1) represents the change in due to intracellular phosphorylation of ribavirin. is the net rate of increase of C due to phosphorylation, is the intracellular concentration of (unphosphorylated) ribavirin, is the phosphorylation rate and is the rate of loss, including by possible slow dephosphorylation, of RXP. In vitro studies of ribavirin uptake by RBCs observe rapid (<10 min) equilibration of intracellular and extracellular ribavirin [53], [57]. We assume therefore that , the concentration of ribavirin in plasma. With twice daily oral administration of ribavirin, rises from zero at and reaches an asymptotic maximum, , so that , where is the characteristic timescale of the accumulation of ribavirin in plasma [6], [38].
The second term on the right-hand side of Eq. (1) accounts for the loss of RBCs due to their death. We assume that the death rate, D, of RBCs increases with as follows(2)where is the death rate of RBCs in the absence of ribavirin, is that value of at which the death rate doubles (or the lifespan halves) compared to that in the absence of ribavirin, and , analogous to the Hill coefficient, determines the sensitivity of to changes in . (A saturable form for D(C) appears inconsistent with available data; see Text S2, Fig. S1.)
Equation (1) is constrained by the initial condition that in all cells at the start of therapy, so that , where N0 is the population of RBCs at t = 0, and is the Dirac delta function, which satisfies and . In other words, the Dirac delta function ensures that no cells have RXP at non-zero concentrations at t = 0. A second constraint on Eq. (1) is imposed by the boundary condition that when >0, newborn cells contain no RXP so that (Text S1) where(3)is the rate of production of RBCs at time t.
The production of RBCs by the bone marrow is regulated by a negative feedback mechanism involving the hormone erythropoietin [58]. Recent studies on modeling erythropoiesis elucidate the complexities involved in a quantitative description of this feedback mechanism [59]–[65]. Here, we employ Eq. (3) to capture the essential features of this negative feedback: As the population of RBCs, , decreases, P increases. is the maximum production rate of RBCs, which occurs when N is vanishingly small, is that value of the RBC population per unit volume of blood () at which , is the volume of blood, and , analogous to the Hill coefficient, determines the sensitivity of to changes in . Eq. (3) provides good fits to independent measurements of the dynamics of the recovery of RBCs following phlebotomy (Text S3, Fig. S2).
Equations (1)–(3) present a model of the population dynamics of RBCs in individuals undergoing treatment with ribavirin. We solve the equations (see below) and obtain the population density, , and the corresponding cumulative population, , using which we predict the time-evolution of the hemoglobin level in blood, (where is the volume of a single erythrocyte); the average concentration of ribavirin in RBCs, ; and the average RBC lifespan, , where is the average death rate of RBCs.
Equation (1) along with the initial and boundary conditions is equivalent to the following set of differential equations obtained using the method of characteristics (Text S4)(4)where Si(t) is the subpopulation of cells born within an interval of that survive at time t. Ci(t) is the concentration of RXP in the latter cells at time t. We solve Eq. (4) along with Eqs. (2) and (3) with d using a program written in MATLAB (Text S5). We validate our solution methodology against an analytical solution that can be obtained in the limiting case when the RBC death rate is independent of RXP accumulation (Text S6, Fig. S3). We also ensure that d allows accurate integration of Eq. (4) without compromising computational efficiency (Fig. S4). From the solution, we calculate the quantities of interest, viz., , , , and .
We employ the following values of the model parameters unless stated otherwise. The average RBC lifespan in normal man is ∼120 days [49], [66], which corresponds to d−1. We let b = 7 following earlier studies [59] and obtain cells d−1 from an independent analysis of blood loss experiments (Text S3). We fix and [67]. Using [47], we get . We obtain from the initial steady state . Further, we let [47] and because ribavirin accumulates in plasma to its maximum concentration in ∼4 weeks, we set [6], [38]. The remaining parameter values , , , and are obtained from best-fits of our model predictions to experimental data (Fig. 5A). We summarize model parameters and their values in Table 1.
We fit model predictions to experimental data (Fig. 5A) using the nonlinear regression tool NLINFIT in MATLAB.
|
10.1371/journal.ppat.1003914 | Genome-Wide RNAi Screen Identifies Broadly-Acting Host Factors That Inhibit Arbovirus Infection | Vector-borne viruses are an important class of emerging and re-emerging pathogens; thus, an improved understanding of the cellular factors that modulate infection in their respective vertebrate and insect hosts may aid control efforts. In particular, cell-intrinsic antiviral pathways restrict vector-borne viruses including the type I interferon response in vertebrates and the RNA interference (RNAi) pathway in insects. However, it is likely that additional cell-intrinsic mechanisms exist to limit these viruses. Since insects rely on innate immune mechanisms to inhibit virus infections, we used Drosophila as a model insect to identify cellular factors that restrict West Nile virus (WNV), a flavivirus with a broad and expanding geographical host range. Our genome-wide RNAi screen identified 50 genes that inhibited WNV infection. Further screening revealed that 17 of these genes were antiviral against additional flaviviruses, and seven of these were antiviral against other vector-borne viruses, expanding our knowledge of invertebrate cell-intrinsic immunity. Investigation of two newly identified factors that restrict diverse viruses, dXPO1 and dRUVBL1, in the Tip60 complex, demonstrated they contributed to antiviral defense at the organismal level in adult flies, in mosquito cells, and in mammalian cells. These data suggest the existence of broadly acting and functionally conserved antiviral genes and pathways that restrict virus infections in evolutionarily divergent hosts.
| West Nile virus (WNV) is an insect-borne virus that has re-emerged globally and for which there are no specific therapeutics or vaccines. We set out to identify cellular factors that impact infection using Drosophila as a model insect. Using a genome-wide RNAi screen we identified a large number of genes that altered WNV infection. We focused on genes that restricted infection and validated 50 genes that were conserved from insects to humans that inhibited infection. Since WNV is a flavivirus, we tested whether additional flaviviruses were restricted by these genes and found that 17 also had antiviral activity against Dengue virus. There are additional families of insect-transmitted viruses that infect humans. Accordingly, we tested whether these genes also were antiviral against the bunyavirus Rift Valley Fever virus, the alphavirus Sindbis virus and the rhabdovirus Vesicular Stomatitis virus. From this analysis, we identified seven genes that are antiviral against all of these divergent arthropod-borne pathogens expanding our knowledge of cell-intrinsic immunity in insects. Lastly, we found that XPO1 and the Tip60 complex had antiviral activity in mammalian cells. These data demonstrate the existence of previously unknown antiviral genes that restrict infection of multiple viruses across divergent hosts.
| Historically, West Nile virus (WNV) epidemics were observed in Africa, the Middle East, Europe, India, Australia, and parts of Asia, In 1999, WNV entered into the North America as part of an outbreak of neuroinvasive disease in New York City [1], and since then has become endemic in the United States with large numbers of cases occurring annually in different regions of the country. Indeed, the occurrence, size, and severity of outbreaks in humans overall have increased worldwide since the mid 1990s [2], with a large outbreak in Texas in 2012 leading to many fatalities [3], [4]. Different strains of WNV, with variable worldwide distributions, exhibit significant differences in pathogenesis. In humans infected with North American WNV strains, approximately 80% of infections are asymptomatic, with 20% developing WNV fever and other relatively mild symptoms, and 1% progressing to encephalitis, meningitis, or flaccid paralysis [2]. In contrast, WNV-Kunjin, endemic in Australia, has not been associated with any human fatalities or severe disease [5]. The natural transmission cycle of WNV is between mosquitoes and birds, with humans, horses, and other vertebrates being incidental dead-end hosts [2]. WNV is a member of the Flavivirus genus, which includes many globally important vector-borne pathogens, such as Dengue (DENV), yellow fever (YFV), tick-borne encephalitis (TBEV), and Japanese encephalitis viruses (JEV) [6]. DENV is endemic in more than 110 countries with 3.6 billion people at risk, and 390 million people infected yearly [7], [8]. At present, there are no specific antiviral therapies against any flavivirus, and only three insect-borne flaviviruses have approved vaccines for humans (YFV, TBEV, and JEV) [9].
Flaviviruses are small (∼50 nm diameter) enveloped viruses that contain a single-stranded, positive-sense RNA genome of ∼11-kb with a 5′ cap, but unlike mRNA, lack a 3′ polyadenylated tail [10]. WNV enters both vertebrate and invertebrate cells through clathrin-mediated endocytosis [11], and then traffics to an acidic compartment that facilitates viral fusion with endosomal membranes and release of the nucleocapsid into the cytoplasm [12]. The viral genome encodes one open reading frame and is translated as a single polyprotein at the rough endoplasmic reticulum (ER), which is subsequently processed by both viral and cellular proteases into 3 structural and 7 non-structural viral proteins [13]. Viral RNA replication occurs within cytoplasmic complexes associated with perinuclear membranes requiring lipid rearrangements [14], [15], [16], [17], and progeny viruses bud into the ER and traffic through the Golgi network where virions are processed into mature particles prior to exocytosis [18].
While there has been extensive study into the cellular pathways that are hijacked to facilitate WNV infection in mammalian cells, less is known about the cell-intrinsic pathways that restrict WNV in insects and whether these pathways have conserved roles in vertebrates. Furthermore, Flaviviruses belong to a larger group of vector-borne RNA viruses (including Togaviruses and Bunyaviruses), raising the possibility that these viruses as a group may be restricted using shared host defense pathways. Indeed, RNA interference (RNAi) is recognized as a major antiviral mechanism in insects and is active against all human arthropod-borne viruses tested including the flaviviruses WNV and DENV [19], [20]. The Jak-STAT and Toll pathways also are active in diverse insect hosts and restrict flavivirus infection in mosquitoes [19], [20]. Indeed, many antiviral pathways active in vector insects were first shown to restrict viral infection in the fruit fly (Drosophila melanogaster) model. This is in part due to the depth of Drosophila genome annotation, powerful genetic tools, potent gene silencing by RNAi, limited genetic redundancy, a high percentage of identifiable functional orthologs in both mosquitoes and vertebrates, lack of an acquired immune system, and that Drosophila can be experimentally infected by a large number of human arthropod-transmitted viruses. Furthermore, RNAi screening is robust in Drosophila cells and has been used effectively to analyze host-pathogen interactions and identify genes involved in antiviral defense including components of the RNAi silencing machinery [21], [22], [23], [24]. Additionally, findings in Drosophila have been extended to mosquitoes and mammals further validating this approach [23], [24], [25], [26], [27], [28], [29], [30], [31].
In this study, we used Drosophila to identify cell-intrinsic antiviral genes that restrict WNV and hypothesized that a number of these would restrict other insect-borne viruses, and some might have conserved roles in vector insects such as mosquitoes. Since many antiviral pathways (e.g., autophagy, Jak/Stat and Toll pathways) are active both in mammals and insects, we speculated that some of these newly identified factors also would confer antiviral activity in mammalian cells. To identify such genes using an unbiased approach we performed a genome-wide high-content RNAi screen in Drosophila cells to identify cellular factors that limited WNV infection. To date, RNAi screens have mainly focused on cellular factors usurped by pathogens to promote infection. While 22 restriction factors have been identified as anti-flaviviral in genome-wide RNAi screens [24], [32], only 2 of these are conserved between humans and insects. We optimized the assay for the discovery of restriction factors and identified 50 genes that when silenced resulted in enhanced WNV infection in Drosophila cells. All 50 are conserved in mosquitoes and 86% have clearly defined human orthologs. Furthermore, 17 of these genes had antiviral activity against multiple flaviviruses, and 7 genes were antiviral against a diverse panel of additional vector-borne RNA viruses. We focused on two broadly acting conserved genes, dRUVBL1 (pontin) and dXPO1 (embargoed), and found both restricted viral infection in adult flies, were antiviral in mosquito Aedes aegypti cell culture as well as in human cells. Furthermore, since WNV is neurotropic we tested whether RUVBL1 contributes to control of WNV in neurons and found it to be antiviral in these cells. Mechanistically, our studies establish that dRUVBL1 along with other members of the Tip60 histone acetylase complex are antiviral suggesting a role for this complex in virus restriction. Furthermore, we found that dXPO1 controls the nuclear export of specific host mRNAs, including the mRNA encoding Aldolase, which we identified as antiviral. Collectively, we identified additional novel, broadly acting cell-intrinsic antiviral genes in Drosophila at least some of which function in mosquito and vertebrate cells.
To identify cellular factors that restrict WNV infection, we first characterized the infection of a pathogenic North American WNV isolate (New York 2000) (referred to as WNV) [33], in Drosophila DL1 cells. WNV successfully infected and produced infectious virions from DL1 cells, although infection levels were substantially lower than that observed in human cells (Figure S1A and B in Text S1). Kinetic experiments revealed that peak immunofluorescence signal of virally produced NS1 protein was 48 hours post infection (hpi), a time point prior to substantial virus spread (Figure S1B and C in Text S1). We next tested whether WNV infection of Drosophila cells was dependent on similar entry and replication pathways as in mammalian and mosquito cells. Chlorpromazine, an inhibitor of clathrin-mediated endocytosis, blocks entry of WNV in both mammalian and mosquito cells [34], [35], and also effectively inhibited WNV infection of Drosophila DL1 cells (Figure S1D in Text S1). Ribavirin, a nucleoside analog and a inhibitor of Flavivirus replication in many mammalian cell types [36], also inhibited WNV infection of Drosophila cells (Figure S1E in Text S1).
Next, we optimized RNAi in a 384-well format using dsRNAs against β-galactosidase (βgal) as a negative control, and dsRNA against the WNV genome as a positive control (Figure 1A and B). We also included dsRNA targeting Ars2, a gene that we previously established as antiviral in Drosophila against many unrelated RNA viruses [21]. By selecting the infection level at ∼7%, this maximized the fold-change in infection upon loss of Ars2, allowing us to focus the assay on genes which restrict infection. This approach contrasts with previous screens that used a higher infection level and focused on genes that promote infection [24], [32]. Briefly, DL1 cells were seeded onto 384 well plates pre-arrayed with dsRNAs, incubated for 3 days for effective knockdown of target genes, and infected with WNV (Multiplicity of infection (MOI) of 10) for 48 hours. Cells were fixed, permeabilized and stained for the viral protein NS1 [37] and counterstained for nuclei. Automated microscopy and image analysis calculated the cell number per well (nuclei) and number of infected cells (WNV NS1) to measure the percent infection. As expected, we observed a decrease in infection after treatment with dsRNA against WNV. Importantly, we also observed a robust increase in WNV infection upon loss of Ars2 (Figure 1A and B); thus these optimized conditions were used for RNAi screening.
A genome-wide RNAi screen was performed in duplicate and statistical analysis identified 537 genes (3.6% of the Drosophila genome) that when silenced had a significant effect on the percentage of WNV infected cells, with a robust Z score of ≥2 or ≤−2 in both replicates (p<0.001; ∼40% change; Figure 1C). None of the non-targeting controls spotted on each plate were identified whereas 100% of the positive control dsRNAs spotted on each plate against WNV genome and Ars2 were identified. Silencing of 376 of these 537 genes resulted in decreased viral infection, indicating WNV was dependent on these genes for replication (viral sensitivity factors (VSF)). Silencing of 161 genes resulted in increased WNV infection suggesting they normally restrict replication (viral resistance factors (VRF)). As WNV infects mosquitoes, birds, and vertebrates we were interested in those genes having orthologs in hosts that normally encounter the virus, rather than genes annotated as Drosophila specific, as flies are not natural hosts. Analysis of this candidate gene list revealed that ∼59% of the genes have orthologs in both humans and mosquitoes (p<0.0001), with Drosophila-specific genes being greatly under-represented (∼23% of the total; p<0.0001) (Figure 1D). Of the 537 genes identified in the primary screen, 147 were cytotoxic (robust Z score<−2 in duplicate; ∼15% decrease in cell number) and were excluded from further analysis. Only one gene had a robust Z score>2 in duplicate but did not validate subsequently. Additionally, 131 genes were not clearly conserved in mosquitoes or humans (as determined by Homologene) and also were excluded from further analysis. Of the 280 remaining genes, we set out to validate all of the genes except for a handful that were members of complexes in which we identified >2 components. In those cases, we chose to validate representative genes from these complexes (Table S1). To do this, we generated independent dsRNA reagents that targeted 217 genes and screened this secondary gene set under two conditions: we infected cells at a low level of infection (4%) to maximize identification of genes that restricted infection, and at a higher level of infection (18%) to maximize validation of genes that promote infection. Of the 217 genes, 121 validated (56%): 82 genes (68%) facilitated WNV infection (VSFs) and 39 genes (32%) restricted infection (VRFs). We also validated a total of 23 genes from larger complexes (Table S1, Table S2, and Figure S1F in Text S1). If we include the remaining 17 genes in the complex, the screen identified 96 VSFs and 50 VRFs in total (Figure 1E; Table S1 and S2).
Bioinformatics analysis was used to identify processes or pathways that impact WNV infection. First, we performed Gene Ontology enrichment analysis on the VSF and VRF gene sets independently (Figure 1F and G) and found biological pathways including vesicle-mediated transport and membrane modifications were enriched within the VSF data set, consistent with the important role of vesicular trafficking and membrane modifications in WNV entry and replication [6]. Second, we used several functional annotation metrics to place these validated genes into cellular pathways and sub-cellular compartments most likely relevant to WNV infection (red genes, VRF; green genes, VSF; black genes not tested but in validated complexes; Figure S1G in Text S1). We identified 29 genes involved in endocytosis and endosomal acidification, a known entry pathway for WNV. Furthermore, although we tested and validated only 4 of the components in the signal recognition particle complex, we identified 6 subunits of this complex in our primary screen, supporting the importance of targeting the WNV polyprotein to the ER for proper translation and processing. These findings suggest that this screen was robust and identified important host factors that promote infection.
The VRFs were highly conserved (86% have human orthologs) and fell into distinct enriched groups. Two of the three categories involved RNA metabolism, including RNA transcription, which may be involved in an antiviral transcriptional program [31]. In fact, 28% of the WNV VRFs (p<0.00012) have a function within the nucleus suggesting a complex host response to infection since WNV replicates exclusively in the cytoplasm (Figure S1F in Text S1).
While few antiviral pathways have been described in Drosophila, the well characterized ones (e.g., RNA silencing machinery) appear to inhibit infection of diverse viruses [38], [39]. Given this, we explored whether the anti-WNV factors identified also restricted other viral pathogens. We tested two additional flaviviruses: the WNV strain Kunjin (CH 16532; WNV-KUN) and Dengue virus (Drosophila adapted Dengue-2 (DENV)). In addition we tested three additional human vector-borne viruses: Sindbis virus (HRsp; SINV), Rift Valley Fever virus (MP12; RVFV), and vesicular stomatitis virus (Indiana; VSV) (Figure 2A). All of these are enveloped RNA viruses transmitted to vertebrates by an insect vector. While mosquitoes are the natural vector for WNV, WNV-KUN, DENV, SINV and RVFV, sandflies are the primary vector for VSV. The flaviviruses and SINV are positive sense RNA viruses, whereas RVFV and VSV are negative sense. RVFV has a tri-segmented genome, while the other viruses encode a non-segmented genome. Thus, these viruses represent divergent families and genomic architectures.
We and others have previously infected Drosophila with WNV, DENV, SINV, RVFV and VSV [24], [25], [40], [41], [42]. However, WNV-KUN infection of Drosophila has not been characterized. WNV-KUN is a less pathogenic strain of WNV endemic to Oceania [5]. We found that WNV-KUN, analogous to WNV New York, productively infected Drosophila cells (Figure S2A and B in Text S1). Next, we optimized conditions for RNAi screening in 384 well plates using both negative and positive control dsRNAs based upon our previous studies selecting conditions to identify restriction factors for WNV-KUN, DENV, SINV, RVFV and VSV (Figure S2C–G in Text S1) [25], [40], [41]. We screened the validated WNV gene set in duplicate against each virus, and Z-scores were calculated (Table S3). We used hierarchical clustering to compare the VRF gene dependencies of all six viruses (Figure 2B). The four positive sense viruses clustered together (flaviviruses WNV, WNV-KUN, and DENV, and alphavirus SINV), while RVFV and VSV, the two negative sense viruses clustered together. This suggests the gene signature of restriction is related to a fundamental aspect of viral structure.
The WNV VRFs had a high propensity to impact infection by multiple different viruses. There was a high concordance of gene dependencies across the three flaviviruses; 31 genes (86%) restricted WNV-KUN and 22 genes (61%) restricted DENV (Figure 2C and Table S3). There also was a large overlap between WNV and SINV VRFs (64%), while less so with RVFV (38%) and VSV (25%). Thus, many anti-WNV factors appear broadly antiviral against other flaviviruses and an unrelated positive strand RNA virus in insect cells. The degree of VRF overlap diminished as the viruses became more disparate (Figure 2C). Nonetheless, we identified 7 host factors that significantly restricted infection by all six vector-borne viruses tested (p<0.05): dXPO1 (emb), dRUVBL1 (pont), dYARS (Aats-tyr), dEIF1B (CG17737), dPPM1L (CG7115), dCTNS (CG17119) and dICT1 (CG6094). All seven of these VRF genes have human and mosquito orthologs (Figure 2D).
Among the validated WNV VRFs, genes with putative nuclear roles were enriched (p<0.00012) and included dRUVBL1 (pontin, also known as Tip49), which was antiviral against all six viruses. RUVBL1 is an ATP-binding protein belonging to the AAA+ (ATPase associated with diverse cellular activities) family of ATPases implicated in diverse cellular pathways in the nucleus and cytoplasm [43], [44], [45], [46], [47], [48], [49]. First, we validated the antiviral activity of dRUVBL1 using independent dsRNA targeting dRUVBL1 outside of the screening format and observed a significant increase (p<0.05) in infection by WNV, WNV-KUN, DENV, SINV, RVFV and VSV compared to control (Figure 3A and B). There was no impact on cell number upon depletion of dRUVBL1 (Figure S3A in Text S1). Next, using quantitative RT-PCR (RT-qPCR) as an independent assay, we found that both WNV and VSV RNA levels were increased upon dRUVBL1-depletion compared to the control (Figure 3C and D).
One advantage of the Drosophila system is the powerful genetic tools including the availability of genome-wide in vivo RNAi transgenic flies. Furthermore, Drosophila are not hematophagous, so they can be challenged easily and safely with highly pathogenic human viruses. We took advantage of WNV-KUN as it is a BSL2 agent in comparison to the more virulent North American WNV strains, which require a BSL3 facility [50]. Wild-type flies were permissive to WNV-KUN infection as measured by plaque assay and exhibited no increase in mortality compared to control flies (Figure S3B and C in Text S1). This is consistent with the natural infection of mosquitoes where limited pathogenesis is observed, and similar to our observations with other vector-borne viruses (VSV, SINV, RVFV) that display limited pathology upon viral infection [25], [40], [41]. However, loss of innate immune defenses in Drosophila or mosquitoes can render insects more susceptible to infection and result in increased viral replication and mortality [19], [20], [21], [40], [51], [52], [53]. Because null mutants in dRUVBL1 are lethal, we took advantage of inducible RNAi transgenic flies [54]. We used the GAL4/UAS system to promote expression of a UAS- inverted repeat (IR) transgene that bears long hairpin dsRNA against dRUVBL1 to target the endogenous transcript in vivo. We induced expression of the transgene using a heat shock (hs) promoter in adult flies allowing us to bypass any developmental requirements. Indeed, expression of the hairpin during development was lethal (data not shown). Importantly, heat shock driven dRUVBL1 RNAi flies had decreased dRUVBL1 mRNA (Figure S3D in Text S1). Next, dRUVBL1-depleted (hs-GAL4<dRUVBL1 IR) and control flies (hs-GAL4<+) were challenged with WNV-KUN and survival was monitored. Unchallenged dRUVBL1-depleted flies exhibited no increase in mortality nor did control flies challenged with WNV-KUN. Notably, the majority of WNV-KUN infected dRUVBL1-depleted flies succumbed to infection (p<0.01, Figure 3E). We next tested if there was an impact on viral load. Groups of 15 flies were challenged, and whole animals were crushed, and assayed for WNV-KUN by plaque assay in four independent experiments (shown as individual dots). We observed modest, but increased viral loads in dRUVBL1-depleted animals compared to controls (set to 1) at day 6 post infection; similar results were observed at day 9 post infection (Figure 3F, not shown).
We subsequently explored the requirement of dRUVBL1 during VSV infection, the best-studied human arbovirus in flies, and most divergent from WNV of the vector-borne viruses tested (Figure 2C). Again, while uninfected flies or wild type control flies challenged with VSV exhibited little mortality, flies depleted for dRUVBL1 and challenged with VSV showed an increase in mortality after infection (p<0.01, Figure 3G). Groups of 15 flies were challenged, and whole animals were crushed, and assayed for VSV by plaque assay in seven independent experiments (shown as individual dots). We observed modest, but increased viral loads in dRUVBL1-depleted animals compared to controls (set to 1) at day 6 post infection (Figure 3H). Together, these results demonstrate the important and broad-spectrum antiviral requirement for dRUVBL1 both in vitro and in vivo in Drosophila.
dRUVBL1 has been shown to function in many complexes, most often in conjunction with another AAA+ ATPase, dRUVBL2 (reptin, also known as Tip48) (depicted in Figure 4A) [43], [54]. Indeed, structural and functional analysis of human and yeast RUVBL1 and RUVBL2 suggest these proteins work as a scaffold in addition to functioning as ATPases, potentially explaining their association with a diverse set of cellular complexes [44]. dRUVBL1, along with dRUVBL2, is involved in chromatin remodeling, most notably in the Ino80 and Tip60/Swr1 complexes [55], [56]. Furthermore, roles in transcriptional regulation facilitating the activity of c-Myc and β-catenin also have been reported in Drosophila and human cells [57], [58]. Additional roles for dRUVBL1 and dRUVBL2 have been described in snoRNA maturation, nonsense mediated mRNA decay, and telomere maintenance. Lastly, dRUVBL1 also has been implicated in chromatin remodeling with the Drosophila Trithorax complex, although this is thought to be independent of dRUVBL2 [46]. Based on these possible functions, we tested components of these complexes for their impact on viral infection to identify which of the putative dRUVBL1 containing complex(es) mediated the antiviral activity. We designed dsRNAs against dRUVBL2 along with the indicated genes in each of the complexes in Figure 4A. Cells were treated with these dsRNAs, along with β-gal (negative) and dRUVBL1 (positive) controls. Importantly, no impact on cell viability was observed (Figure S4A in Text S1). Next, the dsRNA treated cells were infected with either WNV or VSV. Depletion of c-Myc, arm (Drosophila β–catenin), Smg1, Fib and Nop60B did not impact WNV or VSV infection levels (Figure 4B and C). In contrast, dRUVBL2 was antiviral against both VSV and WNV (Figure 4B and C). Depletion of both dRUVBL1 and dRUVBL2 together did not increase infection beyond that observed with silencing of either gene, indicating their effect was not additive (data not shown). Increased WNV and VSV infection also was observed when dTIP60 (Tip60), dEP400 (domino (dom)) and dSMARCA4 (Brahma (brm)) were depleted. Since dRUVBL1, dRUVBL2, dEP400 and dTIP60 are all antiviral, and members of the Tip60 complex, these data suggest that a major antiviral role of dRUVBL1 is through its function in the Tip60 complex.
Next, we tested whether Tip60 also restricted infection of adult flies. Indeed, we depletion of Tip60 using in vivo RNAi led to decreased survival of flies challenged with WNV-KUN and VSV but did not impact survival of unchallenged animals (Figure S4E–G in Text S1). Thus, the Tip60 complex also has antiviral roles in vivo.
Mosquitoes are the natural vectors for WNV, WNV-KUN, DENV, RVFV and SINV although the particular mosquito species that transmit each of these viruses varies [59], [60]. In contrast, the primary vector for VSV is the sandfly, although VSV has been isolated from mosquitoes [61]. Aedes aegypti is the primary vector species for DENV transmission, and can be infected by RVFV, WNV, WNV-KUN, and SINV [62]. Furthermore, the Aedes aegypti genome has been sequenced [63] and the Aedes aegypti cell line Aag2 is amenable to RNAi and routinely used as a model for mosquito cell studies [64].
We designed dsRNAs against Aedes aegypti RUVBL1 (AAEL004686), RUVBL2 (AAEL010341), and TIP60 (AAEL014072) orthologs. Prior to infection, Aag2 cells were treated with these dsRNAs or with dsRNAs against Bgal or the viral genome as negative and positive controls, respectively. Loss of RUVBL1, RUVBL2, or TIP60 mosquito orthologs did not affect cell number (Figure S4B in Text S1) but led to a significant increase in WNV-KUN infection (p<0.05, Figure S4C in Text S1 and Figure 4D). Similarly, each of these genes had antiviral effects against VSV, as silencing resulted in increased infection (p<0.05, Figure S4D in Text S1 and Figure 4E). These data indicate that members of the Tip60 complex also have antiviral activity in cells from a mosquito vector.
dXPO1 (embargoed (emb), also known as CRM1), another broadly antiviral gene identified in our screen, is a nuclear export receptor conserved from yeast to humans. XPO1 shuttles proteins and RNAs from the nucleus to the cytoplasm [65], [66]. To validate the role of dXPO1 in viral infection we tested whether an independent dsRNA against dXPO1 modulated infection. Silencing of XPO1 with an independent dsRNA did not impact cell number (Figure S5A in Text S1) but resulted in to 2 to 4-fold increases in the percentage of cells infected with WNV, WNV-KUN, DENV, SINV, RVFV or VSV as measured by microscopy (p<0.05, Figure 5A and B). Consistent with this, loss of dXPO1 led to a ≥6-fold increase in both WNV and VSV RNA, as measured by RT-qPCR (p<0.05, Figure 5C and D). Thus, a loss of dXPO1 expression leads to increased viral replication in Drosophila cells.
Next, we assessed whether dXPO1 was antiviral in vivo in adult flies. Null mutants of dXPO1 are lethal [67] so we again used an inducible RNAi and observed in vivo silencing of the mRNA (Figure S5B in Text S1). We then challenged control (hs-GAL4>+) or dXPO1-depleted (hs-GAL4>dXPO1 IR) flies with vehicle, WNV-KUN or VSV. While unchallenged flies or control challenged flies did not exhibit increased mortality, dXPO1-depleted flies challenged with either WNV-KUN or VSV had increased mortality (p<0.01, Figure 5E and F). Furthermore, dXPO1-depleted flies had modestly increased WNV-KUN viral loads, as measured by plaque assay of whole flies in four independent experiments (individual dots) relative to control (set to 1) (Figure 5G). And increased VSV loads, as measured by plaque assay of whole flies in three independent experiments (individual dots) relative to control (set to 1) (Figure 5H). These results establish that dXPO1 is required for antiviral defense both in cells and at the organismal level in adult flies.
To assess whether XPO1 also had antiviral activity in the vector mosquito cells, we treated Aag2 cells with dsRNAs against the Aedes aegypti XPO1 ortholog (AAEL001484) or against Bgal or the viral genome as negative and positive controls, respectively. These cells were subsequently challenged with WNV-KUN or VSV. While depletion of XPO1 did not affect cell number (Figure S5B in Text S1), we observed a significant increase in the percentage of Aag2 cells infected with WNV-KUN or VSV (p<0.05, Figure S5C and D in Text S1 and Figure 5I and J).
Since dXPO1 is as a nuclear export receptor, we speculated that dXPO1-dependent regulation of either host genes required for infection or virus-induced antiviral genes may account for the antiviral activity. Indeed, antiviral transcriptional programs have been shown to restrict viral infections in Drosophila [19], [31], [52], [68], [69]. Leptomycin B (LMB) is a potent and specific inhibitor of dXPO1 mediated nuclear export [70]. Previous work demonstrated that LMB treatment of Drosophila cells altered the nuclear export of only 85 mRNAs (<2% of the transcripts surveyed) [71]. One gene, bsg, was XPO1-dependent and required for WNV infection. However, this cannot explain the phenotype of XPO1 because bsg was required for WNV infection and not the other viruses that are sensitive to XPO1 restriction (Table S3). Moreover, 2 XPO1-dependent genes also were transcriptionally induced by VSV infection (CG4294, CG30389) [31]. We generated dsRNAs targeting CG4294 and CG30389 but observed no impact on WNV-KUN or VSV infection (Figure 6C and D). None of the 50 VRFs from our screen were within this set; however, data mining of an RNAi screen with VSV in DL1 cells (S. Cherry unpublished data) identified one additional gene (Aldolase, dALDOA) from this LMB-dependent gene set that showed antiviral activity against VSV (Figure 6A). We generated an independent dsRNA targeting dALDOA and observed that depletion did not affect cell number (Figure S6A in Text S1) but resulted in a 1.5 to 2.5-fold increase (p<0.05) in the percentage of cells infected with WNV-KUN and VSV, respectively (Figure 6C and D). This suggests that dXPO1-dependent mRNA export of dALDOA contributes to the defense against multiple virus families.
Aldolase is a critical enzyme in glycolysis, catalyzing the conversion of fructose 1,6-biphosphate to glyceraldehyde-3-phosphate (G3P) and dihydroxyacetone phosphate (DHAP) (Figure 6B). However, Aldolase may have functions apart from glycolysis, as its expression but not all core glycolytic enzymes are increased in response to LPS treatment [72]. To define whether the antiviral activity of Aldolase was related to glycolysis we performed two complementary experiments. First, we depleted Drosophila cells of additional enzymes essential for glycolysis (Phosphoglucose isomerase (Pgi), Phosphofructokinase (Pfk), Phosphoglycerate kinase (Pgk), and Phosphoglycerate mutase (Pglym87)) (Figure 6B). Depletion of these canonical glycolysis enzymes had no impact on cell number (Figure S6A in Text S1) or WNV-KUN and VSV infection (Figure 6C and D). Second, to overcome the fact that RNAi is incomplete, and that these are enzymes which may be fully active at low levels, we took advantage of two specific and potent glycolysis pathway inhibitors, Dichloracetic Acid (DCA), which inhibits the enzyme pyruvate dehydrogenase kinase, and a hexokinase inhibitor (3Br) [73]. Neither of these treatments impacted cell number (Figure S6B–E in Text S1) or WNV-KUN and VSV infection of Drosophila cells (Figure 6B–D). Together, these data suggest the antiviral effect of Aldolase is not mediated through the glycolysis pathway.
As dRUBVL1 and dXPO1 are conserved from insects to mammals, we tested whether silencing of these genes in human cells impacted infection. For these studies, we transfected human osteosarcoma U2OS cells with siRNAs against a non-targeting control, hRUVBL1 or hXPO1. Three days later, we confirmed silencing of these genes by RT-qPCR (Figure S7A and B in Text S1) with no impact on cell number (Figure S7C in Text S1). Next, the cells were infected with WNV-KUN (MOI of 0.5), and infection levels were monitored using immunofluorescence 20 hpi. Loss of either hRUVBL1 or hXPO1 resulted in a 2 to 3-fold increase in the percentage of WNV-KUN-infected cells, as measured by microscopy (p<0.05, Figure 7A). Consistent with this, we observed an increase in viral RNA levels in cells depleted of hRUVBL1 or hXPO1 as measured by Northern blot and quantified (p<0.05, Figure 7B). Similarly, depletion of hRUVBL1 or hXPO1 enhanced VSV infection, as measured by the percentage of infected cells (p<0.05, Figure 7C) or levels of viral RNA (p<0.05, Figure 7D). Furthermore, we tested whether RUVBL1 likely acted through the same Tip60 complex as we found in Drosophila. To this end, we obtained independent siRNAs against hTIP60 (KAT5) and confirmed they reduced TIP60 expression in human 293T cells as measured by RT-qPCR (Figure S7D in Text S1). Furthermore, we observed significantly increased WNV infection in the depleted cells (Figure 7E). Together, these data suggest that the Tip60 complex is antiviral against multiple viruses and in disparate hosts ranging from insects to vertebrates.
Our initial studies in Drosophila were performed in a single round of infection suggesting that the requirements for the genes in the viral lifecycle included: entry, uncoating, translation, polyprotein processing, and RNA replication. To study the step in the viral lifecycle impacted by the Tip60 complex in mammalian cells we took advantage of a human cell line (293T) that stably maintains a subgenomic WNV replicon expressing GFP [74], [75]. If these genes restricted infection downstream of entry, but upstream of assembly, they should restrict the replication of this WNV replicon. Indeed, siRNA depletion of hRUVBL1 or hTIP60 led to increased levels of WNV replicon replication as measured by immunoblot (Figure S7E in Text S1). Therefore, the action of these genes is at the step of translation, polyprotein processing, or RNA replication.
Since WNV is a neurotropic virus we tested whether RUVBL1 restricted infection in primary neuronal cultures. We prepared cerebellar granule cell neurons from wild-type C57BL/6 mice and transduced them with lentiviruses expressing either a control shRNA, or 4 independent shRNAs against RUVBL1. Three days later, we challenged the cells with WNV (MOI = 0.1), and harvested virus in the supernatant 24 hours later. Notably, all four independent shRNA depleted RUVBL1 to varying extents (Figure S7F in Text S1), and the level of depletion correlated with a significant increase in viral titers (p<0.05, Figure 7F). These data demonstrate that RUVBL1 restricts WNV infection in primary neurons.
To confirm a role for hXPO1 in antiviral defense in human cells using a small molecule inhibitor to complement our RNAi studies, we treated U2OS cells with the XPO1 export inhibitor LMB and monitored WNV-KUN or VSV infection. Treatment with LMB significantly enhanced (2–3 fold) viral replication by both viruses (p<0.05, Figure 7G and H), as measured by an increase in the percentage of infected cells. LMB treatment did not impact cell number (Figure S7G in Text S1). Furthermore, siRNA-mediated depletion of hXPO1 or LMB treatment of 293T cells carrying a WNV replicon revealed that the dependence was again downstream of entry and upstream of assembly since both perturbations led to increased levels of replication (Figure S7E and S7H in Text S1). These data suggest that the hXPO1 has antiviral activity through the regulation of XPO1-dependent cargo export downstream of entry in evolutionarily diverse cell types from insects to mammals.
Genome-wide RNAi screens have been employed to identify cellular factors required by viruses to successfully infect cells as well as factors that, if left unmodulated by the virus, serve to suppress infection. In addition, this screening approach can identify pathways that regulate the expression and activity of direct antiviral factors, orchestrating a robust antiviral response. Since our goal was to identify conserved inhibitory pathways that span insects and mammals with a particular interest in those having broad antiviral activity against disparate viruses, we performed a genome-wide RNAi screen in Drosophila in which we deliberately set a low infection rate, thereby sensitizing our assay to detect factors that when suppressed result in higher levels of infection. This is in contrast to previous genome-wide flavivirus RNAi screens, which targeted a higher level of infection and so led to the identification of a larger number of genes that promote infection [24], [32]. Nonetheless, our screen was sufficiently sensitive and robust to enable us to identify 96 genes that promoted WNV infection. Enriched gene ontology categories included pathways such as clathrin-mediated endocytosis and endosomal acidification that are required for flavivirus entry and were identified by earlier RNAi screens.
We identified 50 restriction factors, greatly expanding the number of cell intrinsic anti-WNV factors known [32], [76], [77], [78], [79], [80]. We compared our restriction factors with previous studies (Table S4). A genome-wide siRNA screen against WNV in human cells identified 22 genes that were antiviral of which 6 had Drosophila homologs; none of which were within our validated antiviral genes [32]. A genome wide screen against hepatitis C virus, a distantly related Flaviviridae family member, in human cells identified 25 antiviral genes of which 12 had Drosophila orthologs; again, none of which were within our gene set [81]. Two screens querying the antiviral role of interferon stimulated genes (ISGs) against flaviviruses were recently published [80], [82]; however, none of our antiviral genes are known ISGs. The Schoggins screen identified 47 ISGs that when ectopically expressed restricted a flavivirus amongst which there were 12 homologs in Drosophila; none of which we identified as antiviral in our screen. The Li screen identified 47 ISGs that when depleted by RNAi restricted infection amongst which 13 had homologs in Drosophila; none of which were identified in our screen. None of the Drosophila homologs from any of these screens were within any other screen making conclusions difficult. Additional screens performed at low levels of infection may reveal additional intrinsic restriction factors.
Unexpectedly, our antiviral gene set was enriched for nuclear functions such as RNA metabolism and transcription even though WNV replicates exclusively in the cytoplasm. This observation suggested that we had uncovered pathways and processes that orchestrate an antiviral response rather than factors that interact directly with the virus. If this were the case, as is seen with antiviral interferon (IFN) responses in mammals, we reasoned that many of the VRFs might have antiviral activities against additional viruses. This in fact proved to be the case - not only did we discover a high degree of concordance between the WNV, WNV-KUN and DEN VRFs (WNV-KUN 86%, 31 genes; DENV 61%, 22 genes), we identified seven genes that restricted infection of all six different arboviruses tested, which included both positive and negative-sense RNA genomes: dYARS (Aats-tyr), dEIF1(CG17737), dPPM1L (CG7115), dCTNS (CG17119), dICT1 (CG6094), dXPO1 (emb), and dRUVBL1 (pont). Since none of these genes have been suggested previously to have an antiviral role in insects, we chose two genes for more detailed analysis: dRUVBL1 and dXPO1.
RUVBL1 had antiviral activity in Drosophila and mosquito cells. Depletion of dRUVBL1 in adult flies converted a non-pathogenic infection by WNV-KUN or VSV into a pathogenic infection with increased mortality and viral replication. These data suggest that RUVBL1 has a highly conserved role in antiviral defense in insects, including mosquito vectors. RUVBL1 is an AAA+ ATPase implicated in many cellular pathways [43] and that interacts with a number of other molecules that impact its function. By methodically suppressing each of its known interacting partners, we found that components of the Tip60 chromatin-remodeling complex (TIP60, EP400 and RUVBL2) that regulates transcription [43] were antiviral against multiple viruses in Drosophila and mosquito cells. dTip60 also was antiviral in adult flies. Furthermore, silencing of RUVBL1 led to increased viral infection in human cultured cells and primary mouse neurons. Silencing of TIP60 in human cells also rendered them more susceptible to WNV infection. Together, these results suggest a conserved role for the Tip60 complex in antiviral defense across phylogeny. WNV subgenomic replicons were used to show that the requirement for these genes in restriction is downstream of entry and upstream of viral assembly, suggesting a restriction of translation, polyprotein processing and/or RNA replication. While further investigation is required to determine the Tip60 targets that are responsible for the antiviral activity and the precise step of the lifecycle impacted, the identification of a chromatin remodeling complex as broadly antiviral is intriguing. Innate immunity is controlled, in large part, through the tight regulation of sequential gene expression programs that have effector function to restrict pathogen replication. We recently characterized a complex and rapid transcriptional antiviral host program active in insects that includes both primary responses which are translation-independent and secondary responses that are translation-dependent [31]. Half of this response was controlled at the level of transcriptional pausing, which also plays a role in innate immune responses in mammals [31], [83], [84], [85]. This antiviral transcriptional program was active against a broad panel of viruses, as we found with the Tip60 complex here. Thus, we hypothesize that the Tip60 chromatin remodeling complex may contribute to the orchestration of this sophisticated antiviral transcriptional response.
A recent study found that RUVBL2 was antiviral against influenza virus by interfering with nucleoprotein (NP) oligomerization that drives viral RNA polymerase activity in the nucleus; however, in contrast to our findings, this effect was independent of RUVBL1 function [86]. This may be a distinct and direct role for RUVBL2 in influenza replication independent of the role for the Tip60 complex in antiviral defense.
Many viruses, including those used in our studies inhibit host transcriptional responses to prevent the induction of antiviral mRNAs including IFN genes. Whether viruses target this Tip60 complex to block an antiviral transcriptional program is unknown. Tip60 is degraded by a number of nuclear viruses including HIV, adenovirus, papilloma virus and cytomegalovirus to promote viral replication [87], [88], [89], [90]. This is thought to alleviate its repression of early gene transcription. Whether these virus interactions alter the activity of the Tip60 complex on antiviral gene expression remains unknown.
The second broadly-acting VRF we investigated was XPO1. At the organismal level depletion of dXPO1 enhanced viral replication and mortality by both WNV-KUN and VSV. Data from yeast and humans suggest that XPO1 is a nuclear export receptor responsible for the translocation of RNAs and proteins from the nucleus to the cytoplasm [91], [92]. However, more recent studies have suggested that the mRNA cargo dependent on XPO1 is limited [71], [93]. Inhibition of XPO1 either by RNAi or using the specific inhibitor LMB, which blocks the nuclear export function of XPO1, resulted in increased viral infection, which suggests the antiviral role of XPO1 is at the step of nuclear export. As the vector-borne RNA viruses studied here replicate exclusively in the cytoplasm, we hypothesize that XPO1 transports cellular mRNAs critical for an antiviral response. Indeed, viruses including VSV (used in our study), HIV, VEEV, ebolavirus and picornaviruses inhibit nuclear export of antiviral genes including ISG mRNAs required for defense in mammalian cells [94], [95], [96], [97], [98], [99], [100], [101]. Consistent with this, LMB inhibition of XPO1 mediated export in human cells suppressed the export of IFNα1 mRNA [99].
While a role for nuclear export in antiviral immunity has been described in mammalian cells, its function in insect immunity was unknown. To identify the particular mRNAs responsible for the antiviral effects of XPO1 in Drosophila we mined a microarray study of Drosophila cells that found less than 2% of the mRNAs tested (85 mRNAs) exhibited nuclear export dysregulation upon LMB treatment [71]. Depletion of one XPO1-dependent mRNA, dALDOA (Aldolase A), resulted in enhanced virus infection in Drosophila cells. This suggests that the transport of dALDOA mRNA plays a role in the innate immune response to vector-borne viral infections. As part of the glycolysis pathway, ALDOA enzymatically cleaves fructose 1,6-bisphosphate (F-1,6-BP) into glyceraldehyde 3-phosphate (G3P) and dihydroxyacetone phosphate (DHAP). Our RNAi and pharmacological experiments suggested the mechanism of viral suppression by ALDOA was independent of its effects on glycolysis. Future studies will be required to define mechanistically how ALDOA acts to inhibit viral infections in insect cells.
Collectively, we have begun to describe a series of conserved pathways, including transcriptional pausing, chromatin remodeling and RNA export that likely regulate the expression of gene sets whose products are antiviral, perhaps in a direct way. For the most part, virus-host interaction studies have often concentrated on proteins that interact directly with the virus. Our work has revealed pathways that orchestrate larger responses, and this confers potent antiviral activity against a broad range of divergent viruses. Clearly, there is still much to be learned about the cellular factors critical for an innate immune response to vector-borne viruses in both vertebrate and invertebrate hosts. Our identification of these cell-intrinsic antiviral genes restricting WNV, and in many cases additional viruses, provides new opportunities for understanding the control mechanisms and larger antiviral programs active against globally relevant classes of emerging viral pathogens.
This animal studies were carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee at the Washington University School of Medicine (Assurance Number: A3381-01).
DL1 and Aag-2 cells were grown as previously described [31]. BHK, U2OS, and 293T cells were maintained as previously described [102]. 293T cells harboring WNV subgenomic replicon were maintained as previously described [74], [75]. Cerebellar granule cell neurons from neonatal (E6) wild-type C57BL/6 mice were generated from cerebella dissected in HBSS and dissociated in 1 mg/ml trypsin with 125 U/ml DNAse (Sigma, St. Louis, MO) for 20 min. Enzymatic digestion was quenched with DMEM/10% FCS and the tissue was pelleted, washed in HBSS, dissociated by trituration through a P-200 pipette tip and layered on a Percoll gradient. Cells were plated in neurobasal media (Gibco) supplemented with B-27 serum-free supplement (Gibco) on poly-D-lysine (PDL)-treated dishes for 1 hour to remove adherent glial cells. Nonadherent cells were then washed in HBSS, counted plated on PDL-coated wells in serum-free DMEM (supplemented with N2 growth medium (Gibco, Grand Island, NY) and 20 mM KCl. Cultures were >95% pure and were used 3 to 4 days later for lentivirus infections [103]. Antibodies were obtained from the following sources, anti-WNV NS1 (9-NS1; [37]), anti-RVFV N (1D8 – gift from C. Schmaljohn), anti-hTIP60 (abcam, ab23886) and Alexa-488 donkey anti-mouse secondary (Jackson Immunochemicals). The following inhibitors were used: Leptomycin B (SIGMA) 50 ng/ml; Dichloracetic Acid (SIGMA) 60 µM; Hexokinase II inhibitor II (Calbiochem) 0.1 mM.
West Nile virus (WNV lineage I strain 3000.0259 New York 2000) was generated in BHK cells, concentrated using Centricon Plus-70 (Millipore), and ultracentrifuged through a sucrose cushion as described previously [104]. The WNV-KUNV isolate (CH16532) was a generous gift of R. Tesh (World Reference Center of Emerging Viruses and Arboviruses, Galveston, TX) was propagated using the same protocol as WNV. DENV (gift from M. Garcia-Blanco) was grown as previously described [24]. SINV was propagated as previously described [25]. RVFV strain MP12 was propagated as described [41]. VSV was grown as described [40]. All MOIs were determined on BHK cells.
For the primary WNV screen dsRNAs targeting 13,071 genes were pre-arrayed in thirty-two 384-well plates at 250 ng per well (Ambion). 16,000 DL1 cells were seeded in serum-free Schneider's media (10 uL/well). One hour later complete media was added (20 uL/well). Three days post plating, cells were infected with WNV at an MOI of 10 (10 uL/well). 48 hours post infection cells were fixed (4% formaldehyde), and a mAb against WNV NS1 (9-NS1) was used to identify infected cells (anti-mouse Alexa-fluor488 (Jackson Immunochemicals)) and counterstained with Hoechst 33342 to monitor nuclei. 3 images per well were captured at 20× using an automated microscope (ImageXpressMicro) and analyzed using MetaXpress software. Average infection and nuclei number were calculated for each site and averaged for each well. The percent infection was log-transformed, and the median and interquartile range were used to calculate a z-score: (log10(%infection)-log10(median))/(IQR*0.74) for each plate. The entire screen was performed in duplicate and those wells with Robust Z-scores≥to 2.0 or ≤to −2.0 in both replicates were considered ‘hits’.
Similar to the primary screen, secondary screen plates were arrayed with dsRNA (250 ng) targeting a different region of the genes identified in the primary screen (DRSC). WNV infections were performed in duplicate at a higher (20) and lower (5) MOI (18% and 4% respectively). Infections with other viruses used the same protocol as WNV and were fixed at the following hours post-infection: WNV-KUN - 48 hrs (MOI = 10), DENV - 72 hrs (MOI = 10), SINV – 40 hrs (MOI = 5), VSV – 24 hrs (MOI = 1), RVFV MP12 – 30 hrs (MOI = 1). Robust Z-scores in each duplicate viral infection set of ≥1.5 or ≤1.5 in duplicate (∼40% change) were considered ‘hits’ (p<0.009); none of the negative controls (non-targeting) were identified as positive, and all of the positive controls (dsRNA against virus genome) were identified.
The functional annotation and clustering of WNV ‘hits’ was performed using the DAVID Bioinfomatics resource. Homologene (NCBI) was used to identify orthologs. Gene Cluster and TreeView were used to generate heat maps.
All flies were maintained on standard medium at room temperature. Flies carrying UAS-dXPO1 IR (VDRC v3347) or UAS-dRUVBL1 IR (VDRC v105408) were crossed to heat shock (HS)-GAL4 flies (Bloomington) at room temperature. On the day of injection, the progeny were heat-shocked at 37°C for 1 hour and shocked every 2 days throughout the experiment. Adults of the stated genotypes were challenged with WNV-KUN or VSV as previously described [22]. Groups of at least 20 flies were challenged for mortality studies. For viral titers, groups of 15 flies per experimental treatment were crushed and processed at 6 days post-infection for plaque assays on BHK cells [31].
Total RNA was isolated from infected cells using Trizol (Invitrogen). For northern blots, RNA species were analyzed as previously described [105]. For RT-qPCR, cDNA was generated using random hexamers to prime reverse transcription reactions using MMLV reverse transcriptase. cDNA samples were treated with 100 U DNase I (Qiagen) according to manufacturer's protocol. Quantitative PCR (qPCR) was performed with the cDNA using Power SYBR Green PCR Master Mix (Applied Biosystems) and primers targeting VSV and WNV (VSV For-CGGAGGATTGACGACTAATGC, Rev-ACCATCCGAGCCATTCGA: WNV For-ACATCAAACGTGGTTGTTCCGCTG, Rev-TTGAGGCTAGAGCCAAGCATAGCA) in accordance with manufacturer's protocol. qPCR conditions were as follows; initial 94°C for 5 min, then 30 cycles of 94°C for 30 sec, 55°C for 30 sec, and 72°C for 30 sec. Relative viral copy numbers were generated by normalizing to cells treated with control dsRNA.
For siRNA treatments, mammalian cells were reverse transfected with 20 nM siRNA (Ambion: Negative Control #1 (AM4611), GFP (AM4626), XPO1 (s14937), RUVBL1 (s16370)), KAT5 (s20630, s20631) using HiPerfect according to the manufacturers protocol (Qiagen). 60 hours post-transfection, for immunofluorescence, cells were replated in a 96-well format and infected with either VSV or WNV-KUN virus 12 hours later. For FACS the cells were infected with WNV for 24 hours and stained for NS1 or TIP60 [82] or infected and processed for northern blot or immunoblot at 12 hr p.i. for VSV and 20 hr p.i. for WNV-KUN.
For shRNA treatments, lentiviruses (pLK0.1) encoding shRNA targeting RUVBL1 (clone 1: GCTGGAGATGTGATTTACATT; clone 2: GCTGGCAAAGATCAATGGCAA; clone 3: GCCACAGAGTTTGACCTTGAA; clone 4: GCAAGATATTCTGTCTATGAT) or a control (luciferase) were obtained from RNAi Core facility at Washington University School of Medicine. Lentivirus particles were generated after co-transfection of HEK-293T cells with packaging plasmids. Supernatants were collected at 48 hours later and added to neuron cultures. Three days after transduction, neurons were infected with WNV (New York 1999 strain) at an MOI of 0.1. One day later, supernatants were harvested and titered for virus infection by focus-forming assay [106].
One day prior to infection U2OS or DL1 cells were seeded in 96-well plates at 20,000 or 70,000 cells per well respectively. Leptomycin B (SIGMA) was added at 50 ng/ml to mammalian cells cells 2 hr prior to infection with WNV-KUN or VSV. Dichloroacetic Acid (SIGMA) or Hexokinase II inhibitor II (Calbiochem) were added to U2OS or DL1 cells respectively, 30 minutes prior to infection. Cells were fixed 20 (WNV-KUN) or 12 (VSV) hr post infection, stained and imaged as previously described.
|
10.1371/journal.pntd.0005633 | First evidence of lymphatic filariasis transmission interruption in Cameroon: Progress towards elimination | Lymphatic filariasis (LF) is among the 10 neglected tropical diseases targeted for control or elimination by 2020. For LF elimination, the World Health Organization (WHO) has proposed a comprehensive strategy including (i) interruption of LF transmission through large-scale annual treatment (or mass drug administration (MDA)) of all eligible individuals in endemic areas, and (ii) alleviation of LF-associated suffering through morbidity management and disability prevention. In Cameroon, once-yearly mass administration of ivermectin and albendazole has been implemented since 2008. The aim of this study was to assess progress towards the elimination goal, looking specifically at the impact of six rounds of MDA on LF transmission in northern Cameroon.
The study was conducted in the North and Far North Regions of Cameroon. Five health districts that successfully completed six rounds of MDA (defined as achieving a treatment coverage ≥ 65% each year) and reported no positive results for Wuchereria bancrofti microfilariaemia during routine surveys following the fifth MDA were grouped into three evaluation units (EU) according to WHO criteria. LF transmission was assessed through a community-based transmission assessment survey (TAS) using an immunochromatographic test (ICT) for the detection of circulating filarial antigen (CFA) in children aged 5–8 years old.
A total of 5292 children (male/female ratio 1.04) aged 5–8 years old were examined in 97 communities. Positive CFA results were observed in 2, 8 and 11 cases, with a CFA prevalence of 0.13% (95% CI: 0.04–0.46) in EU#1, 0.57% (95% CI: 0.32–1.02) in EU#2, and 0.45% (95% CI: 0.23–0.89) in EU#3.
The positive CFA cases were below WHO defined critical cut-off thresholds for stopping treatment and suggest that transmission can no longer be sustained. Post-MDA surveillance activities should be organized to evaluate whether recrudescence can occur.
| Lymphatic filariasis (LF) affects more than 120 million people worldwide, and is considered the second leading cause of permanent and long-term disability. In response to the important burden of this disease, the World Health Organization (WHO) elaborated a strategic plan to eliminate LF as a public health problem through annual preventive chemotherapy (PC), repeated for at least six years, and reaching at least 65% of the population at risk. To date, about 5.63 billion cumulative treatments have been delivered since 2000, and more than 300 million people no longer require PC thanks to successful implementation of the WHO strategy. In Cameroon, PC for LF has been implemented since 2008. The aim of this study was to assess whether the transmission of LF has been interrupted. Cross-sectional surveys were conducted in three evaluation units (EU) in northern Cameroon. The LF prevalence observed in each of these EU was lower than the threshold of infection below which transmission is likely no longer sustainable, suggesting that the transmission of LF has been interrupted in the study area.
| Lymphatic filariasis (LF) is among the most widespread neglected tropical diseases. In the mid-1990s, it was reported that about 1.4 billion people were exposed to the disease worldwide, of whom 120 million were infected and more than 40 million disfigured by the disease [1].
One of the core resolutions of the 50th World Health Assembly held in 1997 was to eliminate LF as a public health problem (resolution WHA50.29). To address this global concern, the World Health Organization (WHO) proposed a comprehensive elimination strategy including (i) transmission interruption in endemic communities (so-called mass drug administration or MDA strategy), and (ii) implementation of interventions to prevent and manage LF-associated disabilities (so-called morbidity management and disability prevention or MMDP strategy) [2]. The Global Programme to Eliminate Lymphatic Filariasis (GPELF) was launched in 2000, by the WHO, to elaborate specific plans and coordinate control efforts to reach this ambitious goal. In the MDA strategy, LF must be mapped and preventive chemotherapy (PC) implemented to treat the entire eligible population (areas where prevalence of antigenaemia is ≥ 1%). In areas where onchocerciasis is endemic and where Wuchereria bancrofti prevails, the recommended PC is a single dose of a bi-therapy (150 μg/kg of body weight ivermectin in combination with 400 mg albendazole), administered once yearly [3,4]. Since this treatment is not macrofilaricidal, adult worms can remain viable for about six years and the delivery of several rounds of MDA appeared crucial. It is now accepted that annual MDA should be repeated for at least 5 years at adequate levels of coverage, estimated to be at least 65% of the total population in endemic areas (“effective” MDA), to ascertain that the level of infection in the community will be reduced to levels below which transmission cannot be sustained, even after MDA has been stopped [5]. Recent estimates of the impact of MDA during the past 13 years revealed that more than 96 million LF cases were prevented or cured, although as many as 36 million cases of hydrocele and lymphedema remain [1]. However, data reporting interruption of LF transmission are scanty, especially in Sub Saharan Africa where the disease represents one-third of the global burden [6].
Cameroon is known to be endemic to onchocerciasis [7,8] and bancroftian filariasis [9,10], and MDA against LF have been implemented since 2008. Indeed, ivermectin and albendazole have been distributed by community drug distributors (CDDs) following the community directed treatment with ivermectin (CDTI) approach. This strategy has already been implemented 15–20 years earlier to fight against onchocerciasis. As such, the strategy was already well mastered by CDDs and was ongoing smoothly at the time MDAs against LF were implemented, following a door-to-door approach. This study aimed at assessing whether the transmission of LF has been successfully halted in areas where six MDA rounds have already been delivered.
This study was carried out in 2014 in the North and Far North Regions of Cameroon, situated between latitudes 7° and 12°N, and longitudes 12° and 16°E. Five health districts or implementation units (Ngong, Poli, Tcholliré, Rey-Bouba in the North Region and Mokolo in the Far-North Region), with rural to semi-urban settings, were included in this study. These implementation units (IU) were organized into three evaluation units (EU) (Fig 1) according to the criteria described in the WHO monitoring and evaluation manual [2]. In 2014, the population of each of these two Regions was estimated to two millions, children aged 6–7 years old representing about 10% of the general population [11].
A cross sectional study was carried out following the recommendations described in the WHO manual for national elimination programs [2].
The flow chart below (Fig 2) describes the different steps taken in the LF elimination process, thus conferring the eligibility of the targeted implementation units to the transmission assessment survey step.
All relevant data for this study were recorded into a purpose-built Microsoft Access database and subsequently exported into PASW Statistics version 18 (SPSS Inc., Chicago, IL, USA) for statistical analyses. The prevalences of infection were expressed as the percentage of infected children (harboring CFA) among the total number of children examined; the 95% confidence interval (CI) was calculated using the Wilson method not corrected for continuity [18]. Chi-square tests were used to compare LF prevalence between sexes and age groups, as well as the computed threshold of infection prevalence below which transmission is likely no longer sustainable, so-called critical cut-off threshold, against the observed proportion of ICT positive cases.
This study was conducted as part of the action plan of the national program to eliminate lymphatic filariasis in Cameroon. Ethical clearance was granted by the Cameroon National Ethics Committee for Human Health Research (N°2014/09/491/CE/CNERSH/SP). Before enrolment, the objectives and schedule of the study were explained to the eligible population and individuals willing to participate signed two inform consent forms, and kept a copy. The second copy was stored at the Centre for Research on Filariasis and other Tropical Diseases. Even after minors assenting, the approval of their parents or legal guardians was necessary before any procedure. Each enrollee was assigned a unique identifier and his data analyzed anonymously. Positive cases were referred to CDDs and health officers for a close follow-up during next treatments, and their parents or legal guardians warned about the situation to further insure a better compliance. Although no guidelines are given in the TAS manual [2], the number of positive cases- that can be up to 18 as was the case in the present study -, appears as a real concern in a context where MDA has to be halted if the EU passes TAS. In this context, we have recommended to treat these rare positive cases with ivermectin during the MDA campaign plan just after the survey, then by a long course of doxycycline (4–6 weeks) when they get above 8 years and MDA no longer available.
A total of 97 communities (EAs) were surveyed in the three EUs, and 5292 children (48.9% females) examined. These children were aged 5–8 (median age: 6) years old. Among the 5292 enrollees, 4171 (78.8%) were aged 6–7 years old (initial target), a small proportion being aged 5 (11.8%) or 8 (9.4%) years old. A total of 1595 children were examined in EU#1, 1919 in EU#2 and 1778 in EU#3, the expected sample size being reached in all the three EUs (Table 1).
Prevalence of W. bancrofti circulating antigens, assessed using ICT card test, was equal to 0.13% (95% CI: 0.04–0.46) in EU#1, 0.57% (95% CI: 0.32–1.02) in EU#2, and 0.45% (95% CI: 0.23–0.89) in EU#3 (Table 1). The overall prevalence was 0.40% (95% CI: 0.26–0.61), with 80.95% positive cases aged 6–7 years old. The prevalence of LF was similar, both between age groups and sexes (p > 0.7408). The spatial distribution of positive cases was in general over-dispersed (both among health districts and EAs), except in the EU#2 where 8 children (1.07%; 95% CI: 0.54–2.10) with W. bancrofti circulating antigens were found in the Ngong health district, 6 of them belonging to two EAs.
The total number of LF positive cases was 2 in EU#1, 11 in EU#2 and 8 in EU#3, all below the critical cut-off threshold (18 in each EU) generated by the Survey Sample Builder. As compared to the threshold of infection prevalence below which transmission is likely no longer sustainable, the proportion of positive cases was significantly lower in the EU#1 (Chi-square = 12.68; p = 0.0004) and EU#3 (Chi-square = 5.27; p = 0.02), but not significantly lower in EU#2 (Chi-square = 3.48; p = 0.06).
In Cameroon, MDA against LF, using the combination of ivermectin and albendazole, started in 2008 in the North and Far North Regions. In 2014, five health districts (Mokolo, Ngong, Poli, Rey-Bouba and Tcholliré) completed six MDA rounds, and successfully passed the assessment of impact of MDA on LF infection after the fifth round of MDA (post 5th MDA survey). The objective of the present study was thus to check whether the transmission of the disease has been successfully halted.
Based on historical data [10], sentinel sites’ survey data and/or kriging data [9,12], the North and Far North Regions were previously highly endemic for LF, and were reported among the most prevalent over the country. In 2014, LF prevalence observed in each of the three EUs investigated—in average equal to 0.40% (95% CI: 0.26–0.61)—was significantly lower than the threshold below which the transmission of the disease can no longer be sustained. Indeed, it was accepted that in areas where W. bancrofti is endemic and Anopheles or Culex is the principal vector, this target threshold must be < 2% antigenaemia prevalence [2]. In Cameroon, LF entomological data are very scanty but malaria data can be informative. Although Anopheles gambiae and Anopheles funestus have been found naturally infected with W. bancrofti [10], malaria entomological data have shown that the most abundant vectors in the Northern Cameroon are from genera Anopheles and Culex (Nwane, personal communication).
The number of positive antigenaemia cases observed in each of the three EUs surveyed was below the critical cut-off generated by the SSB (18 CFA positive cases), indicating that the area successfully “passed” TAS, and a cessation of MDA in the constituting communities should be envisioned. Indeed, the sample sizes and critical cut-off values were chosen so that an EU has (i) at least a 75% chance of passing TAS if the true prevalence of antigenaemia is 1.0% (half the target level if the vector is Anopheles or Culex), and (ii) no more than about a 5% chance of passing (incorrectly) TAS if the true prevalence of antigenaemia is ≥2.0% [19,20]. The importance of transmission assessment surveys as an evaluation tool for stopping MDA have been previously demonstrated in a multicenter evaluation using different approaches or study designs [21]. Moreover, the validity of TAS was also proven in long term post-MDA surveillance, although complementary test (antibody and xenomonitoring) appear of interest to ascertain the interruption of transmission during post-MDA surveillance [22–24].
It is important to notice that the interruption of transmission was achieved despite the fact that in some health districts, the effective treatment coverage was not reached for one or two rounds, although globally higher than 65% (S1 Table). TAS was considered for these health districts for three main reasons: (i) long lasting insecticidal nets (LLINs) have been distributed in the study area in the framework of malaria control program activities. Indeed, between 2003 and 2010, more than two millions LLINs have been distributed to pregnant women and children under 5 years old. In the framework of LLINs universal coverage for the control of malaria, a total of 21,028,770 LLINs have been distributed in the entire country in 2011 and 2016 (with 73% coverage in 2011 and 88% coverage in 2016), on the basis of one LLIN for every 2.2 households [25]. The impact of LLINs on prevalence and intensity of LF infection is now widely accepted [26,27], and it was shown that a sustained reduction in LF prevalence can be reached in spite of missed rounds of MDA [28]. In addition to these efforts related to the known usefulness of LLINs, the relatively poor compliance observed at the beginning of this large scale control strategy against malaria, and to some extent against LF, was improved over time thanks to communication and sensitization of populations [29]. It seems worth to mention that although insecticide resistance has been reported in several foci in Cameroon, it was demonstrated that LLINs might still offer some protection against the resistant Anopheles gambiae s.l. populations in northern Cameroon [30]. (ii) Ivermectin has been widely distributed in the study area since 1987 (Fig 2). Indeed, in 1987–1989, limited MDA campaigns were organized in the framework of a phase IV trial of ivermectin conducted in the Vina Valley located in the North Region of Cameroon [31]. In 1992, the Ministry of Health (MoH) and the River Blindness Foundation (RBF) began to broaden distribution of ivermectin, with the assistance of non-governmental developmental organizations (NGDOs), through mobile teams/outreach approach [32]. Since 1997–1998, the African Program for Onchocerciasis Control (APOC) joined the coalition to support annual delivery of ivermectin through community-directed treatment with ivermectin (CDTI) [33]. Although ivermectin is not macrofilaricidal, it is highly microfilaricidal and repeated treatments might have significantly contributed to the interruption of LF transmission [34,35]. Moreover, it was demonstrated that the transmission can be interrupted earlier than expected in areas previously treated for onchocerciasis [36]. (iii) Last but not the least, the prevalences in the study areas were relatively low when MDAs against LF began, suggesting that in such context, LF endemicity can be quickly lowered to level under which transmission cannot be sustained.
After six years of MDA (ivermectin in combination with albendazole), the transmission of LF was interrupted in five IUs (Mokolo, Ngong, Poli, Tchollire and Rey-Bouba heath districts) of the North and Far North Regions. These results support the cessation of MDA in these IUs, but this decision needs further thinking. It was demonstrated that MDA can be safely stopped in some but not all local government areas of Plateau and Nasarawa States in Nigeria [37], suggesting that the cessation of MDA can be feasible in the IUs investigated in northern Cameroon, even if the transmission of LF might be ongoing in the neighboring IUs. This is likely in accordance with the focal LF transmission that might be occurring in Cameroon. Also, the LF prevalences were relatively low at the beginning of MDAs, and the neighboring EUs has already completed at least four effective rounds of MDA when mass treatments can be halted as a consequence of transmission interruption.
However, epidemiological surveys conducted in northern Cameroon in 2008–2010 showed that mass ivermectin distributions had significantly lowered prevalence and intensity of onchocerciasis, but the transmission of the disease was yet to be interrupted [33]. In such circumstances where onchocerciasis transmission is still ongoing in these (and the neighboring) IUs, the interruption of treatments (IVM + ALB) might need further thinking. It is accepted that in areas where onchocerciasis is endemic, ivermectin can be used solely after interruption of LF transmission but this might be challenging while conducting surveillance activities to investigate potential recrudescence of LF. Another important challenge to take into account is the endemicity of soil transmitted helminthiasis (STH) since both IVM and ALB are effective against the parasites responsible of these diseases, especially in areas where STH control is not performing well. In such circumstances, it appears useful to investigate the situation of onchocerciasis and STH, especially now rapid diagnostic tests are being releasing for these diseases. This will help taking the decision about stopping MDA not only according to the evidence of LF transmission interruption, but also to the situation of STH and onchocerciasis in the selected EU. These additional data, collected in an integrated manner during TAS surveys, will be really cost effective and provide more insights in decision making.
|
10.1371/journal.pntd.0001326 | Schistosomiais and Soil-Transmitted Helminth Control in Niger: Cost Effectiveness of School Based and Community Distributed Mass Drug Administration | In 2004 Niger established a large scale schistosomiasis and soil-transmitted helminths control programme targeting children aged 5–14 years and adults. In two years 4.3 million treatments were delivered in 40 districts using school based and community distribution.
Four districts were surveyed in 2006 to estimate the economic cost per district, per treatment and per schistosomiasis infection averted. The study compares the costs of treatment at start up and in a subsequent year, identifies the allocation of costs by activity, input and organisation, and assesses the cost of treatment. The cost of delivery provided by teachers is compared to cost of delivery by community distributers (CDD).
The total economic cost of the programme including programmatic, national and local government costs and international support in four study districts, over two years, was US$ 456,718; an economic cost/treatment of $0.58. The full economic delivery cost of school based treatment in 2005/06 was $0.76, and for community distribution was $0.46. Including only the programme costs the figures are $0.47 and $0.41 respectively. Differences at sub-district are more marked. This is partly explained by the fact that a CDD treats 5.8 people for every one treated in school.
The range in cost effectiveness for both direct and direct and indirect treatments is quantified and the need to develop and refine such estimates is emphasised.
The relative cost effectiveness of school and community delivery differs by country according to the composition of the population treated, the numbers targeted and treated at school and in the community, the cost and frequency of training teachers and CDDs. Options analysis of technical and implementation alternatives including a financial analysis should form part of the programme design process.
| Schistosomiasis and soil-transmitted helminth control programmes are important, relatively low cost means to improve the health of those affected, in particular rural school age children. It can also reduce schistosomiasis related morbidity in their later lives. The paper presents information on the implementation and costs of a large scale national programme in Niger. The total economic cost per treatment was $0.58. This includes programme, government and international costs. Two systems, school based and community delivery were used to treat children and targeted adults. Contrary to findings in some countries we find that school based delivery is less cost effective than community delivery. This is due to the low proportion of the population targeted and treated by the school based system. Treating adults as well as children increased the numbers treated and reduced the overall cost per treatment. Prevalence and infection is higher in children than adults and overall effectiveness in terms of infection averted is affected. The cost per infection averted is assessed for direct treatment and direct and indirect treatment effects. The study expands the evidence available for decision makers involved in programme planning and design, funding and implementation.
| Schistosomiasis is one of the most prevalent chronic infectious diseases found world-wide and is associated with anaemia, chronic pain, diarrhoea, and under nutrition. It is recognised as a major public health problem in many rural areas, particularly in school-age children. With affordable and sustained control measures morbidity and transmission can be decreased.
Robust studies on the implementation of large scale control of schistosomiasis and soil transmitted helminths (STH) are required to strengthen the evidence base on the cost-effectiveness and affordability of such investment [1]–[3]. In particular, evidence on effectiveness is needed to support the strategic planning for expanded treatment and global coverage as well as for national vertical and integrated Neglected Tropical Disease (NTD) programmes. The objectives of this study are to identify: the cost of the Mass Drug Administration (MDA) programme; the cost per person treated; and the costs of treatment as delivered by school based staff and community distributers.
A number of studies have identified the costs of targeted and MDA for schistosomiasis control [4]–[9]. These have, with the exception of [4], [9], provided empirical evidence of school based approaches. This paper provides the cost of MDA treatment and compares the costs of the school and community based distribution systems used. It assesses the evidence from other MDA programmes taking account of factors such as the level of school enrolment, coverage levels to consider the general guidelines that can be taken from these works.
In 2004 Niger established a national programme to control schistosomiasis and soil-transmitted helminths (PNLBG) supported by the Schistosomiasis Control Initiative (SCI), funded by the Bill and Melinda Gates foundation [10]. Its objective in line with Resolution WHA54.19 was to treat 75% of school age children at risk of infection and in communities where prevalence is over 50% to also treat at risk adults. The purpose being to reduce the morbidity related to schistosomiasis infection to a level at which it would not constitute a public health problem [11].
The primary school net enrolment rate (NER) in 2004 in Niger was 41% (UNESCO UIS global education database Table 5, Enrolment ratios by International Standard Classification of Education (ISCED) level), lower in rural areas; and considerably less than the rate of 68% for Sub Saharan Africa (SSA). To achieve high treatment coverage in targeted school age children and at risk adults two treatment strategies, school-based and community-based distribution, were established. Treatment for S. haematobium was provided every two years in most endemic areas, and annually in high prevalence areas to reduce initial levels. School-based distribution was provided by trained teachers who distributed the drugs to students in the schools. Children not attending school could receive treatment either in the schools or from the Community Drug Distributor (CDD) at home or at another fixed treatment location.
The MDA programme was established and rolled out over 2 years from April 2004 across 40 districts in all 7 regions of the country, including the capital city Niamey. The programme activities were implemented progressively commencing in the Tilaberi and Dosso regions. In 2004/05, 1,627,828 treatments were delivered in 22 districts and in 2005/06 2,683,121 treatments were delivered in 40 districts. Figure 1 outlines the main MDA programme activities.
Initial meetings and agreement of the programme with the regional and district administrations were followed by a prevalence survey and mapping to prioritise areas for MDA. A national workshop and practical field sessions to develop capacity in the diagnosis of schistosomiasis was organised. Further capacity-building workshops and training for staff in organisation, management and implementation of MDA was provided to key district health and educational inspectorate staff in an initial national workshop. These staff then trained clinic staff, health workers and head teachers though district meetings. Training was provided on the calculation of drug requirements, drug distribution organisation, the management of side effects and the reporting of results. Community and school-based drug distributors were trained in the use of dose poles (to determine drug dosage), and the completion of treatment forms.
The national programme developed, piloted, printed and delivered information, education and communication (IEC) materials for distribution to the districts. These materials included posters and a booklet for use in schools and communities as well as technical sheets for those administering the drugs. Drugs were procured centrally by SCI on behalf of countries which SCI supported in West and East Africa [12]. The drugs were sent directly from the National store to the districts. The districts and inspectorates repacked the drugs and IEC material for distribution to or collection by clinics and schools. Social mobilisation activities were undertaken at various administrative levels. National radio and television broadcasts were undertaken in three local languages and in French, organised by the PNLBG; local radio broadcasts were organised by the districts; village criers were organised by clinics to inform communities about the logistics of the MDA.
A rally to launch the campaign was undertaken and organised by an host region; it involved a day of speeches, dance and hospitality supported by national, regional and programme dignitaries and was broadcast on national radio and television.
Treatments were delivered at schools by teachers; CDDs provided treatment by going from door to door and at other fixed points supervised by clinic, district and regional staff. Technical and management support and supervision were provided to the districts by national staff.
At the end of the MDA unused drugs and monitoring reports were collected by national staff. A one day post MDA evaluation meeting was held in each participating region attended by national, regional and district staff. A summary of the partner roles and responsibilities is identified in Table 1.
A protocol for the cost and resource use study was agreed with Niger Ministries of Health and Education at national and local government level. Written informed consent to participate in the programme longitudinal surveillance and monitoring research was obtained from the children's parents or guardians, or head teachers according to the study protocol approved by St Mary Research Ethics Committee of Imperial College, UK, 2003, (EC No 03.36, R&D No: 03/SB/003E) and amended 2005 St Mary's (REC Ref: AM2003).
This was a retrospective study which covered a two year period from April 2004 to May 2006, including the first and second years of MDA and related programme activities in four health districts.
All data on first year costs at national, regional, district, and sub district levels were taken from the PNLBG accounts and receipts and records of staff missions or activities. Second year cost data for national and regional level activities were taken from receipts. District and sub district, school and community MDA resource use data for 2005 were collected in June 2006 through a retrospective survey. The four health districts: Kollo, Tera, Tilaberi, (Tilaberi region) and Gaya (Dosso region) are all located in the Niger River Valley, and were in the first phase of implementation. The control programme had previously established sites for longitudinal monitoring of prevalence and morbidity in these districts.
The cost survey was designed to collect data on the time taken and resources used by district and sub district health staff, and by the education inspectorate and school staff in the 2005 MDA delivery. Questionnaires were designed and tested, and a 2 day training workshop was undertaken in 2006 to familiarise and train the schistosomiasis MDA district co-ordinators, responsible for the data collection.
The questionnaires covered the usage of: vehicles, fuel and other equipment and materials used in the MDA and in training, time spent in different activities by staff at regional, district, and clinic level, payments made to CDDs, and for local services. Survey data on 2005 MDA delivery costs were verified with PNLBG receipts. Drug usage was obtained from district and from PNLBG records. Coverage figures were obtained from district treatment registers and the national annual treatment summary.
Longitudinal surveillance and monitoring data were obtained from records of the Centre de Recherche Medicalé et Sanitaire (CERMES) and from the PNLBG register of activities and receipts.
The study examines the economic costs of the MDA programme in its first and second years. The economic costs include the full value of the resources used. Where this is not adequately represented by the financial or market cost, an opportunity cost is used (see Table S4). The main cost elements include: the programme specific expenditure; the opportunity cost or value of government contributions related to in-kind costs of using local government staff and vehicles and the value of CDD's time (taken as the daily agricultural labour rate); and the international costs of programme co-ordination, reporting and technical support.
Programme costs include directly incurred capital costs; recurrent costs; and variable costs. Capital costs incurred by the programme included central level purchase of Information Technology (IT), medical and laboratory equipment and other electrical and mechanical goods and furniture used to equip the PNLBG office including the purchase of four 4×4 vehicles for PNLBG (Table S1). Capital costs were annualised over their useful lives (Table S2) using a discount rate of 3%. This represents the annual cost of owning and operating an asset over its lifespan.
Programme recurrent costs including staff costs, office and vehicle running costs, and programme variable costs were collected from the programme records, accounts and receipts. These costs were apportioned in relation to the time spent by programme staff on MDA activities and the proportion of that time allocated to study areas.
Variable costs related to perdiems, materials and services incurred in relation to the programme activities. Centralised activities (e.g. organisation and provision of national training for all district technicians, planning and organisation) and regional activities (e.g. MDA launch) were equally apportioned in relation to the number of districts in the MDA and share of the four study districts. Location specific activity costs such as supervision, mapping, central delivery of drugs to districts were allocated on the basis of costs incurred in the study districts.
Sentinel monitoring informs the national treatment strategy. The costs of sentinel site monitoring were apportioned to the study districts on the basis of study area treatments relative to national treatments.
Government staff costs were based on salary costs collected through questionnaire and the Government salary grid. District and sub district vehicle usage was calculated from questionnaire returns and costed using hire rates. These values are estimated to reflect the opportunity cost of using the resources for the MDA rather than for an alternative activity.
Costs were collated and classified by three levels of organisation (national, regional & district, or community), type of activity (training, support & supervision, baseline & monitoring, reporting, evaluation, advocacy, mobilisation & IEC) and cost type (fuel, transport, materials, services, drugs, per diems, temporary contracts and office related recurrent costs).
Prices are in constant 2005 terms (Table S4). Foreign exchange was converted at the fixed rate of CFA 655/Euro and $1.244/Euro (http://www.federalreserve.gov/release/January 2 2009). Discounted economic analysis was undertaken using discount rate of 3% in line with World Bank rates [13]. The cost of a treatment includes both albendazole and praziquantel.
Community and school based delivery was and is practised nationally. The costs incurred by the two systems were equally attributed at national, regional and district level. It is at sub district level that the systems differ in the organisation and implementation of the delivery activities. The school and sub district delivery services used a partial analysis which took account of these cost components only.
The cost of delivery using a CDD and of using a teacher was calculated. These costs included per diems and travel allowances for CDD and head teacher training; allowances for delivery (applicable only for CDD), health clinic staff costs for CDD selection (per diems and fuel) and supervision (fuel only). The training of one or more teachers and their supervision in schools was undertaken by the school head, no financial cost was incurred. Joint activity costs of the district health and education inspectorate (training, drug repacking, drug delivery to sub districts and schools and supervision) would be incurred despite the system. These have not been included in the partial analysis but an allowance has been estimated to allow comparability with other MDA programmes.
The effectiveness of treatment was calculated as the difference between the population with schistosomiasis infection at baseline and follow-up survey. The prevalence rates used are from a longitudinal health impact study (Nadine Seward (2007) Niger Three Years Data Analysis, SCI internal report (unpublished)).
To assess the effectiveness of the programme's direct and direct and indirect treatment effects an assessment of the impact in the treated population and in the targeted population was made. Treatment costs were calculated as the number of treatments in each year multiplied by the full economic cost in 2004/5 and in 2005/6.
Eight schools and four communities located in areas highly endemic for schistosomiasis took part in a longitudinal health impact study. The study used baseline and longitudinal follow-up surveys one year post treatment to monitor: parasitological indicators (prevalence and intensity of helminth disease examining stool and urine samples following standard procedures using kato katz and filtration methods [14]); morbidity indicators (anaemia and associated pathology of schistosomiasis, assessed by ultrasound examination following standard protocols developed by WHO) and general indicators of height and weight. The baseline survey enrolled 1659 children from 8 different schools in 3 regions prior to the first MDA campaigns of 2004 and 2005. The number of children enrolled from each school ranged from 179 to 299; with almost equal numbers of children in age groups of 7, 8 & 11 years old. Of those recruited 1193 (72%) were followed-up successfully at year 1 and year 2 surveys. Adults and adolescents were monitored in 4 sites in a single region. A total of 484 adolescents and adults were recruited at baseline. Of these, 143 (30%) were followed-up successfully at both year 1 and year 2 surveys. The sample sizes was estimated using the same criteria as described in [15].
The surveyed sites mirror the MDA treatment and represent MDA performance in targeted populations taking into account the treated and untreated participants in proportion to the MDA coverage. Any indirect effect of reduced infection in untreated pupils resulting from changes in the force of infection is reflected in the intensity of infection [16] which is related to prevalence ([17] provides more detail). To assess the wider impacts on the community, untreated first year students were monitored in the schools. Adults and adolescents were monitored at four sites.
The total economic cost of the programme including programme specific expenditure, national and local government costs and international technical support and programme co-ordination in four study districts, over two years, was US$ 456,718 (Table 2); an economic cost per treatment of $0.58. Excluding international costs, the programme and government expenditure was $0.54 per treatment. The programme expenditure per treatment was $0.44. The average drugs cost was $0.28 per treatment. The numbers treated in these two years totalled 818,562 (781,883, discounted at 3%).
The distribution of costs between the programme, the government and international support are shown in Table 2. Drugs accounted for 49% of the total economic cost (65% of programme expenditure), variable costs accounted for 19% of the economic cost (26% of programme expenditure). Overall there was little difference in the total economic cost of the programme in the four districts between the first and second years. However the total economic cost per treatment in the first year was $0.68 and in the second year was $0.51. Cost differences are shown in Table 3 and discussed below.
Excluding the MDA drug costs, the economic cost of the programme in the four districts in the second year was 29% less costly than the first year and treated 25% more people. Higher costs in the first year of the programme are seen in programme costs and international support. Three factors contribute to this. The cost of the initial start up activities incurred in the first year only. The activities involved advocacy, development of IEC materials, prevalence surveys and data collection for planning and the establishment of monitoring and evaluation (M&E) activities, in particular the longitudinal monitoring sites (illustrated in figures 2 and 3), and repair and maintenance of the national office. In the second year the programme was scaled up. This reduced the apportioned share of recurrent and capital programme costs and international costs allocated to the study area. In 2004/05 22 districts were treated and in 2005/06 40 districts were treated. Within the study area the population treated in the second year which was 25% more than those treated in the first year.
The distribution of variable expenditure (excluding drugs) by activity in the study area is presented in Figures 2 and 3. These show the relatively large proportion of expenditure on training and on MDA delivery. It also highlights activities mainly undertaken at establishment. Total programme variable costs in 2004/05 were $ 51,970 and in 2005/06 were $ 40,318, 22% less than those in the first year.
M&E costs include costs of process monitoring in 2004/5, annual district and regional evaluations and programme health impact monitoring undertaken through the National sentinel sites. These costs amounted to an average of 13% of variable costs over the 2 years. Table 4 presents the average allocation of cost by category (capital, recurrent and variable) and type of input. Labour related costs (salary plus per diems) and vehicle and fuel costs account for 64% and 19% of all costs excluding drugs.
Sensitivity analysis was undertaken on major cost items. A 10% increase in the cost of drugs would result in a 4.9% increase in the total economic cost of treatment ($456,718), and a 6.5% increase in the current programme cost ($342,226). A 10% increase in perdiems and allowances would result in a 1.1% increase in the total economic cost of treatment, or a 1.5% increase in the programme cost, a 4.2% increase in the programme cost excluding drug costs. A 10% increase in wages and salaries would result in a 1.5% increase in the economic cost of treatment. It would impact most on the government sector and distributer opportunity costs increasing costs by 7.7%. The increase on the programme cost would be 0.2% The sensitivity of total economic cost to a saving in teacher training costs was explored. This assumed community distributers would undertake the school treatments for the same fixed allowance. Savings in teacher training allowances are assumed, but not the economic cost of their time which would still be required to support distributers in school based treatment. Any savings in teacher time would be offset by the increased opportunity cost of time for community distributers. The impact of the net saving on the total economic cost would be 2.9%. This is equivalent to a saving of 4.1% in the programme cost. This provides an approximate scale of magnitude within which to assess comparative costs of sub district delivery systems below.
Sub district costs (i.e. clinic, school and community costs) account for the largest portion of the economic cost by administrative level. This is 23% of the total economic cost (based on Table 2); so, it is important to understand the allocation and usage.
Sub district variable programme costs include head teacher and CDD per diems for training and CDD payments for distribution. Sub district government costs include the opportunity cost for the use of motorbikes (11%) and labour (89%). The opportunity cost of labour is principally accounted for by the time of the teacher and head teachers (61%), of the clinic staff in supervision (20%) and CDD time for training and distribution (19%).
Table 5 presents the characteristics and costs of sub district delivery. The economic cost per school based treatment and per CDD treatment delivered was $0.36 (range $0.26–$0.55) and $0.06 (range $0.04–$0.07) respectively. The programme cost per school based treatment and per CDD treatment delivered was $0.09 (range $0.07–$0.15) and $0.03 (range from $0.03–0.04) respectively.
The full economic delivery cost of school based treatment in 2005/06 was $0.76, and community treatment was $0.46. If only programme costs are included this figures are $0.47 and $0.41 respectively.
The difference in costs is in part explained by the fact that a CDD delivers 5.8 treatments for every one delivered in school. On average each CDD delivered 407 treatments while each school delivered 70.
Over the 2 treatment cycles 530,300 treatments were provided to an estimated 317,549 adults and 288,262 treatments were provided to 241,218 children in the study areas in the regions of Dosso and Tilaberi. Coverage in the target population in Gaya, Dosso was 78% in both years, and was 69% and 71% in the three districts monitored in Tilaberi.
Two estimates of the cost of treatment per case of infection averted (Table 6) are presented. One includes only the direct impacts of treatment and the other includes the direct and indirect impacts of treatment. They provide minimum and maximum limits of the true value. This is discussed further in the next section.
Including only the direct impacts on the treated population the average cost per infection averted in treated children in the 4 districts over the two years was $1.10. Of the 317,549 adults treated in a single round the cost per infection averted was $4.4 and over two rounds it is estimated to be $6.5. The overall cost per infection averted in the treated population of children and adults is calculated as $2.5.
If indirect treatment effects are included the average cost per infection averted in targeted (treated and untreated) children in the 4 districts over the two years was $0.78. Of the 446,180 adults targeted in a single round the cost per infection averted was $3.08 and over two rounds it is estimated to be $4.6. The overall cost per infection averted in the targeted population of children and adults is calculated as $1.78.
The higher cost of infection averted in adults reflects the lower base prevalence rate. The longitudinal adult cohort followed up over the 2 year period suffered high drop-out rates and its composition was significant different at the 0.05 significance level. Males in particular those who were infected with S. haematobium infection were more difficult to retain at follow-up. The resulting cohort of 116 adults had a lower proportion of males, and had a lower base rate of infection (24.1%, (95% CI:16.35–31.93)) as compared with the original baseline sample (39.62% (95%CI:35.23–44.01)). To avoid this issue the results for the sample monitored at the first year follow up are used and it is conservatively assumed that the prevalence in the second follow up did not change.
The cost per treatment and prevalence figures relate to the study sample of four districts located in the Niger River Valley. This was and is an area of high disease prevalence and high population density relative to other parts of the country. The costs per person treated may be higher in lower density and more remote areas. Likewise the cost per infection averted will be greater in sub populations with lower changes in prevalence.
The cost study relied on the survey work for details of district and sub district resource use. As the survey was undertaken almost a year after the MDA, it may have been affected by recall bias. A further limitation of the study is the delay in final analysis. This limited follow-up work; the recent MDA's confuse recall and some key people have changed positions and locations. However, the issues raised by the study analysis are still relevant and worthy of further investigation.
One of the strengths of the study is the availability and use of MDA health impact monitoring results to assess programme effectiveness. Cost effectiveness studies are often obliged to use trials data concerned with the efficacy of treatment [3] rather than programme effectiveness as monitored in Niger. This potentially allows us to capture direct and indirect treatment effects as identified by Miguel and Kremer [18] and French et al [16]. Both identify direct and indirect effects of treatment in terms of the reduced transmission of Schistosoma mansoni in children in the community including children not treated. The direct and indirect impact quantified here represents a best case or maximum impact. Further work based on intensity and force of infection based on the work of French et al [16] is required to refine and triangulate the estimate. Estimating only the direct effects of treatment in the treated population provides a conservative estimate of infection averted. Due to the definition of prevalence used in the study and the data available, the prevalence estimate in the treated population is under estimated (and consequently cost per treatment is overestimated). The true value of the infections averted is believed to be between these two estimates. There is a 40–42% difference in the cost per infection averted between the “best” and worst case scenarios. However the magnitude of difference underlines the importance of refining the method and developing more robust estimates.
The most effective means of delivering helminth treatment to school age children has been debated in various papers. The Partnership for Child Development (PCD) [6] provided evidence from Ghana and Tanzania on the cost of large scale treatment in schools and the potential savings in using the existing school infrastructure for treatment. For many recipients, access to the more numerous schools is more convenient than attending more distant health facilities [19]. However, where school enrolment is low and particular groups (for example girls or the poorest children) are under-represented, there is a need for additional methods of reaching target populations [7]. Studies in Tanzania [20] and Uganda [21] have examined the effectiveness of Community Directed Treatment (ComDT) and school based treatment in terms of coverage for enrolled and non enrolled children. In Tanzania coverage in both systems was similar, whilst in Uganda coverage rates under ComDT was higher; the associated costs are not discussed.
The evidence on the cost effectiveness of three recent large scale helminth MDA studies in Sub-Saharan Africa is summarised in Table 7. This presents the characteristics: treatment strategies, distribution methods, coverage levels, activities costed, study duration and the number of treatments rounds provided. Each of these affects the cost and technical effectiveness of the programme.
Prevalence and mapping data facilitate treatment prioritisation of endemic areas. This reduces the numbers treated, easing pressure on constrained budgets, but allows the option, to treat targeted at risk adults. Niger and Uganda used a targeted approach. Burkino Faso undertook a blanket approach and treated all school children in the first treatment.
Burkina Faso has the lowest financial cost per treatment but excludes start up, mapping and M&E costs included in the studies (see Table7); it also has the highest coverage. Coverage is a key factor in determining costs per treatment; in particular capital and recurrent costs. Children accounted for all treatments in Burkina Faso and 53% of treatments in the Niger study. Niger's targeted strategy reduced the numbers of children requiring treatment. This eased the budget constraint allowing the targeted treatment of adults, and increased the scale economies of the programme.
Niger and Uganda provide a measure of the effectiveness of treatment. Infections averted are used based on anaemia in Uganda and schistosomiasis in Niger. The use of a technical measure of effectiveness provides the opportunity to assess, both ex-ante and ex-post the potential cost effectiveness of alternative strategies.
The sub-district school based delivery cost per treatment is similar between Niger and Burkina Faso (a difference of 10%). Uganda district school based delivery costs (allowing for central programme costs) are almost 45% greater than Niger's. The reason for this is not clear. It may be that central costs are included in activities other than “programme costs” Sub district community delivery cost per treatment in Niger are significantly lower than Burkina Faso due to the high numbers targeted and treated. The low levels of enrolment, low school numbers targeted and lower coverage rates all add to the relative cost per person of the Niger school based system.
The cost per person treated can be reduced either by increasing national coverage or by an improvement in resource efficiency. Alternative means of school based implementation are available and school based treatment can be delivered by teachers, health workers or CDDs, training may or not be required annually. However the differences in delivery will impact on treatment acceptability [19] and coverage, collaboration and motivation as well as the variable costs.
The distribution of programme costs in Niger, suggest cost savings would have greatest impact in drugs and training. Drugs are a major component of the treatment cost and account for almost half of the economic cost in the current study, between 27%–46% per district of the economic cost in Uganda and 69% of the financial cost in Burkina Faso. The central procurement (undertaken in 2004/06 by SCI) improved the buyer's market power, described by Fenwick and Thompson [12]. On average 3.1 praziquantel tablets and 1.4 albendazole tablets were consumed per treatment in the current study. Using an estimate of actual against planned praziquantel usage (3.5 tablets/adult and 2.5 tablets/child), gave an average difference of 6% with a range of −4% to 23% across the four districts. The average rate is considerably more than the wastage rate of 1% assumed in [5]. This range shows a considerable discrepancy between districts and emphasises the importance of robust drug monitoring and reporting system.
Targeted treatment of at risk adults in high endemic areas is used in Niger and Uganda in line with WHO guidelines. As the cost per infection averted in adults was 3.5 greater than for a child, it is important to understand the economic value of adult treatment. One approach is to assess the direct benefits in terms of impact on adult productivity and the value of this productivity. USAID Famine Early Warning System Network (FEWS-NET): Niger Livelihoods Profiles, provide a useful description of livelihood profiles in these areas. Two studies, with agricultural production comparable to that in the Niger study, report the impact of schistosomiasis on agricultural labour productivity [22], [23]. The impact of schistosomiasis treatment on family labour in paddy rice growing systems in Mali [22] found that health is improved due to schistosomiasis treatment. As a result the time available for farm work by family workers increased by 69 days/ha. Much of this time was invested in the cultivation of additional non irrigated land, (0.47 ha) (if it is assumed a family has 7-10
members, the average improvement / person would be 10-7 days.). The days of family farm labour lost due to schistosomiasis and other parasitic and non parasitic infections was assessed in rain fed farm systems in Benue State, Nigeria [23]. In these systems 46% of time lost due to illness was related to schistosomiasis, an average loss of 18.7 working days per adult. The cost per adult of infection averted in Niger in the first round is estimated here as $4.3 (Table 6) equivalent to 3 days of labour (based on the agricultural day rate ($1.4) in 2005, a year of famine) or 2.3 days (based on rate of $1.9 in a normal year). This indicates the potential economic net gain from adult treatment. The gain could be greater depending on adult rates of re-infection and consequent treatment need.
The cost of treatment per person is driven by the scale of treatment. The strategy, in Niger, to include targeted adults as well as school age children has increased the treatment numbers and reduced the cost per person treated and increased effectiveness. A conservative estimate of cost effectiveness over 2 years for the treated population is estimated to be $1.1 per infection averted for children and $6.5 for adults.
This study used a targeted treatment strategy; 53% of treatments were to children, but only 16% of the population treated received school based treatment. Under these conditions community based treatment was more cost effective than school based treatment. In Burkina Faso, only school age children were treated; 40% of these received school based treatment); the school based system was more cost effective per treatment. However, the school and community based distribution systems serve overlapping groups in the population; and was designed to facilitate access to treatment for different groups and support a coverage rate of 75% or more in target populations. Any improvement in either system must be the result of improved resource use or increased coverage at the district and programme level if the change is not to impact on the effectiveness of the other system.
In designing cost effective and sustainable programmes factors relating to: the treatment strategy, the demographic mix of the population served, system acceptability to stakeholders and the coverage rate need to be taken into account along with logistic issues such as health staff availability. Economic and financial assessment of alternative implementation plans should be undertaken for the project or programme design. This would support decision makers and programme managers, provide financial evidence in planning discussions and negotiations and potentially reduce the need for programme changes to improve cost effectiveness.
|
10.1371/journal.ppat.0030032 | Transport of Streptococcus pneumoniae Capsular Polysaccharide in MHC Class II Tubules | Bacterial capsular polysaccharides are virulence factors and are considered T cell–independent antigens. However, the capsular polysaccharide Sp1 from Streptococcus pneumoniae serotype 1 has been shown to activate CD4+ T cells in a major histocompatibility complex (MHC) class II–dependent manner. The mechanism of carbohydrate presentation to CD4+ T cells is unknown. We show in live murine dendritic cells (DCs) that Sp1 translocates from lysosomal compartments to the plasma membrane in MHCII-positive tubules. Sp1 cell surface presentation results in reduction of self-peptide presentation without alteration of the MHCII self peptide repertoire. In DM-deficient mice, retrograde transport of Sp1/MHCII complexes resulting in T cell–dependent immune responses to the polysaccharide in vitro and in vivo is significantly reduced. The results demonstrate the capacity of a bacterial capsular polysaccharide antigen to use DC tubules as a vehicle for its transport as an MHCII/saccharide complex to the cell surface for the induction of T cell activation. Furthermore, retrograde transport requires the functional role of DM in self peptide–carbohydrate exchange. These observations open new opportunities for the design of vaccines against microbial encapsulated pathogens.
| Microorganisms are comprised of proteins, carbohydrates, lipids, and nucleic acids. Current immunologic paradigms state that activation of T lymphocytes required for humoral and cellular immune responses resulting in immunologic memory to the pathogens is solely brought about by proteinaceous antigens, processed and degraded to small peptides, loaded onto major histocompatibility complex (MHC) molecules, and transported as MHC/peptide complexes to the cell surface, where the MHC/peptide complex is recognized by the T cell antigen receptor. The findings of the present study elucidate the mechanism of MHC class II (MHCII)–dependent presentation of the bacterial capsular polysaccharide of Streptococcus pneumoniae serotype 1 (Sp1) that results in effective T cell activation. Sp1 is transported in MHCII-positive tubules from lysosomal compartments to the plasma membrane for presentation. In the absence of the DM molecule, known as an editor and catalyst of self and foreign peptide exchange, retrograde transport of carbohydrate/MHCII complexes resulting in dendritic cell engagement with T cells in vitro and T cell–dependent immune responses to the polysaccharide in vivo fail. The results suggest a fundamental shift in the immunologic paradigm, offering previously unrecognized opportunities for the design of new classes of vaccines against infectious diseases.
| The immune response to polysaccharide antigens is considered T cell–independent [1]. However, emerging evidence suggests that bacterial polysaccharides from Streptococcus pneumoniae, Bacteroides fragilis, and Staphylococcus aureus activate CD4+ T cells in vivo and in vitro due to their zwitterionic charge motif within each repeating unit [2,3]. Nuclear magnetic resonance (NMR) structural studies of zwitterionic polysaccharides (ZPSs) such as the capsular polysaccharides PS A2 from B. fragilis and Sp1 from S. pneumoniae serotype 1 reveal the formation of extended right-handed helices with repeated 20 Å negatively charged grooves and positive charges located on the outer surfaces of the lateral boundaries [4,5]. A minimum molecular weight of ZPS >5 kDa and ≤17 kDa is required for the elucidation of antigenicity [6]. ZPSs induce CD4+ T cell activation in the presence of B cells, monocytes, and dendritic cells (DCs) [7] and have been demonstrated to correct systemic T cell deficiencies [8]. Animals lacking αβCD4+ T cells fail to develop abscesses in response to ZPS [9]. We and others have shown that T cell activation by the ZPS requires the costimulatory molecules B7–2, CD40, and the major histocompatibility complex (MHC) class II protein HLA-DR [7,10,11]. ZPSs locate in endosomal compartments and co-immunoprecipitate with HLA-DR [7,12]. These studies indicate similarities between ZPS and peptide antigen presentation to CD4+ T cells by antigen-presenting cells (APCs).
Antigen processing and presentation to CD4+ T cells by the MHCII endocytic pathway has been considered strictly limited to protein antigens [13]. A complex set of interlinked factors, including the local pH, is likely to influence the activity of the processing enzymes. The endosomal pH in APCs is regulated by proinflammatory and anti-inflammatory cytokines [14] and microbial products such as bacterial lipopolysaccharide (LPS) [15]. LPS triggers enhanced vacuolar proton ATPase function in immature DCs (iDCs), lower endosomal or lysosomal pH, and more efficient antigen processing and a rapid and transient boost of MHCII synthesis [15,16]. Another important event in antigen processing and presentation is the removal of class II invariant chain (CLIP) occupying the peptide binding groove by the MHCII homolog DM (HLA-DM in humans and H2-M in mice). DM further stabilizes the empty MHCII molecule and assists in peptide selection [17–19]. In the absence of DM, peptide editing fails, leading to the appearance of weakly bound peptides, including CLIP [20]. CLIP also qualifies as an endogenous regulator in DCs in priming of T helper cells by antagonizing the polarization towards the TH1 phenotype [21]. Recent studies show that LPS challenge induces tubules from lysosomes, which transport MHCII to the cell surface [22–24]. In the case of protein-loaded lysosomes, protein is transported in the MHCII tubules to the cell surface for presentation of the peptide, formation of biological interaction with T cells, and induction of T cell–dependent immune responses [24,25]. Although lysosomal compartments contain abundant glycosidases that act on sugar linkages, presumably with high specificity, they are considered the end stations for carbohydrates. MHCII tubules formed after LPS challenge in iDCs do not transport carbohydrates such as dextran from lysosomes to the cell membrane [24]. However, Cobb et al. recently showed that a nitric oxide–dependent processing mechanism of the ZPS PS A1 in early endosomes, resulting in the generation of low molecular weight, antigenic fragments [12]. This finding indicates that antigen processing is not limited to proteins only. Our recent observation that blockade of endo/lysosomal acidification inhibits carbohydrate-induced T cell activation [11] suggests that intravesicular acidic pH and DM activity [26] are required for polysaccharide binding to MHCII in lysosomal compartments and/or retrograde transport of carbohydrate/MHCII complexes.
Here, we report the mechanism of retrograde transport of the bacterial capsular polysaccharide Sp1 from S. pneumoniae in live cells and the essential role of DM in this process. Sp1 traffics through endosomal compartments to acidic lysosomes and is transported in MHCII-positive tubules for presentation to the cell surface and engagement of T cells. We show that the DM molecule is required for the retrograde transport and the cellular immune responses in vitro and in vivo. The data close a gap in our understanding of the new paradigm of MHCII presentation of bacterial carbohydrate antigens.
Sp1 induces T cell activation in the presence of B cells, monocytes, and DCs [7]. In an experimental model of abscess formation, besides macrophages [27], CD11c-positive DCs play an important role. They migrate into the peritoneal cavity upon Sp1 challenge and are retrieved in the abscess capsule (Figure 1). Live cell imaging showed that in iDCs, part of Sp1 is internalized into early endosomes as indicated by partial co-localization with Rab5 and BCECF-dextran (dextran-2′,7′-bis-(2-carboxyethyl)-5-(and-6)-carboxyfluorescein) (Figure 2A and 2B). Co-localization of Sp1 with Rab7 and dextran, markers for late endosomes and lysosomes, and with LysoTracker, an acidotropic marker for lysosomes, demonstrated that Sp1-containing compartments fuse with late endosomes and lysosomes (Figure 2C–2E). Sp1 co-localized with ovalbumin, a conventional protein antigen processed and presented by the MHCII pathway (Figure 2F). In order to test whether Sp1 is internalized and represented by recycling receptors, we performed live cell imaging with Rab11b-EGFP fusion protein, a marker for recycling endosomal compartments. Co-localization of Sp1 with Rab11b was not observed during an observation interval of 5 min to 24 h in the absence and presence of LPS treatment (Figure 2G). These results demonstrate that after internalization, Sp1 gains access to endocytic compartments where antigenic epitopes are loaded to MHCII molecules. As presentation of Sp1 by recycling receptors is largely excluded, the question arises whether and how Sp1 is transported to the cell surface.
Lysosomes constitute the terminal compartment of the endocytic pathway where exogenous components are generally degraded. Recent studies with green fluorescent protein (GFP)–tagged MHCII have shown that after LPS stimulation of iDCs, MHCII molecules are transported via tubules that originate from lysosomes to the plasma membrane [22–24]. We transfected iDCs of C57BL/6 wild-type (WT) mice with MHCII-GFP (I-Eα-EGFP) to investigate the presentation mechanism of Sp1. Flow cytometry analysis revealed that surface expression of assembled I-A and I-E molecules in transfected DCs was similar to I-A surface expression in non-transfected cells (Figure S1). Thus, MHCII-GFP was fully functional and appeared to exhibit the same general pattern of intracellular transport as endogenous MHCII in iDCs and mature DCs (mDCs). In iDCs, MHCII-GFP co-localized extensively with Sp1-Alexa 594 in lysosomes (Figure 3A). No MHCII-GFP was found on the cell surface. Within 4 h after LPS stimulation, numerous extensive tubules extended from the perinuclear area, which were intensely labeled for both MHCII and Sp1 (Figure 3B; see also Video S1). All tubules were yellow, indicating that Sp1 is exclusively transported in MHCII-positive tubules.
The observation that lysosomes can form dynamic and motile MHCII/Sp1-containing tubules does not prove that these structures mediate the transfer of MHCII/Sp1 complexes to the cell membrane. To determine directly whether these tubules not only move to the periphery but also actually reach the plasma membrane, we imaged Sp1 transport in DCs using combined epifluorescence (EPI) and total internal reflectance fluorescence microscopy (TIR-FM) [28]. We observed Sp1-Alexa 594–containing tubules exiting lysosomes and associating with the plasma membrane (Figure 3C; see also Video S2). With time, as the tubule approached the membrane, the red-colored EPI signal decreased while the bright yellow TIR-FM signal increased. After 4 h of LPS stimulation of Sp1-biotin–treated DCs, we detected significant amounts of Sp1 on the APC surface by fluorescence-activated cell sorting (FACS) analysis (Figure 3D), demonstrating that MHCII/Sp1 fuses with the cell membrane for cell surface presentation of Sp1. Concomitantly, Sp1-treated DCs induced engagement with CD4+ T cells, while LPS stimulation of DCs in the absence of Sp1 did not induce conjugate formation with T cells (Figure 4A and 4B). Analysis of CD69, the early activation marker, showed that although the majority of naïve CD4+ T cells form transient interactions with Sp1-pulsed DCs, a maximum 8% of naïve T cells are activated by Sp1-treated DCs at 10 h of incubation (Figure 4C), indicating proper cell surface T cell stimulatory function of MHCII/Sp1 complexes.
The proposed functions of CLIP are that of a precursor peptide to be exchanged for foreign antigenic peptides, and of a regulator in priming TH cells by antagonizing the polarization towards the TH1 phenotype [21]. It was shown that regardless of the presence and type of protein antigen provided to mDCs and loaded onto MHCII, the number of surface CLIP/MHCII complexes remained unchanged [21]. Here, in contrast to ovalbumin as a control antigen, the incubation of maturing DCs with Sp1 resulted in a 57% decrease in Δ-CLIP surface expression (Figure 5A), whereas HLA-DR expression remained unaltered. The ratio of the mean fluorescence intensities (MFIs) of CLIP/MHCII for iDCs, mDCs, and ovalbumin-treated DCs was 0.5, and 0.3 for Sp1-treated DCs. This observation suggests that CLIP is displaced when Sp1 is present and that reduced CLIP surface presentation modulates Sp1-mediated T cell immune responses. To investigate whether Sp1 treatment also affects presentation of self peptides other than CLIP, we performed matrix-assisted laser desorption and ionization mass spectrometry (MALDI-MS) of MHCII precipitates from T2.DR4.DM transfectants. The composition of the self-peptide repertoire with CLIP as the major representative remained unaltered when we compared MALDI-MS spectra obtained in the absence and presence of Sp1 (Figure 5B). In summary, Sp1 provided to DCs and loaded onto MHCII leads to a reduction in the number of surface self-peptide/MHCII complexes with CLIP/MHCII as the principal subset.
Sp1 is a highly charged molecule and might be exchanged for peptides in an antigen site due to its stronger electrostatic forces. However, it is possible that DM as a catalyzer of peptide exchange and editor of peptide/MHCII binding might also be required for carbohydrate/peptide exchange. To assess the catalytic activity of DM, we first investigated whether DM is required for T cell–dependent immune responses to Sp1 in vivo. In an experimental model for abscess formation, unlike WT mice, animals lacking DM are not able to form abscesses in response to Sp1 (Figure 6A). Twenty-four hours before challenge, CD4+ T cells from WT mice were adoptively transferred to DM−/− mice per intravenous route to compensate for the 3- to 4-fold reduction of mature CD4+ T lymphocytes and for the diminished T cell repertoire selection of DM−/− mice [29]. Analysis of the cells migrating into the peritoneal cavity 24 h after polysaccharide challenge showed that the total number of cells did not differ in WT and DM−/− mice (Figure 6B, left panel). In both groups, about 40% of influx cells were macrophages (not shown). In contrast to WT mice, the peritoneal influx of CD4+ T cells was significantly reduced in DM−/− mice (Figure 6B, right panel). Adoptive intraperitoneal transfer of Sp1-pulsed APCs from WT mice fully reconstituted the CD4+ influx in DM−/− mice. Analysis of the peritoneal lavage did not reveal a CD4+ T cell influx to the APC transfer alone. These findings indicate that the CD4+ T cell influx in response to Sp1 depends on DM expression in peritoneal APCs.
We investigated the role of DM in the retrograde transport of Sp1 in DCs and in the initiation of T cell–dependent immune responses in vitro. We transfected iDCs of DM−/− mice of the H-2b haplotype with I-Eα-EGFP. The dependency of I-A molecules on DM to function in antigen presentation has been characterized extensively in the DM−/− mouse strain used [29–31]. It also has been shown that I-E is dependent on DM for peptide loading, as evidenced by the abundance of CLIP occupying the MHCII groove in DM−/− mice [32]. Flow cytometry analysis revealed that after LPS stimulation of DM−/− iDCs, assembled I-E and I-A molecules appeared at the cell surface with similar quantities and kinetics as in WT DCs (Figure S2). In DM−/− iDCs, MHCII co-localized extensively with Sp1 in lysosomes (Figure 7A). Within 4 h of stimulation with LPS, tubules extended from the perinuclear area, which were intensely labeled for MHCII-GFP and were devoid of Sp1-Alexa 594 (Figure 7B). At this time point and within the next 6 h, all tubules were green, indicating that Sp1-Alexa 594 is not transported in tubules with MHCII-GFP in DCs lacking DM. To provide functional evidence for the requirement of DM for Sp1 presentation in DCs, we examined the effect of the absence of DM on APC/T-cell engagement in vitro. iDCs from DM−/− and WT mice were pulsed with Sp1-Alexa 594 in the presence or absence of LPS for different time intervals. Pulsed DCs were incubated with carboxyfluorescein succinimidyl ester (CFSE)–labeled CD4+ T cells from WT mice and examined for DC/T-cell conjugate formation by fluorescent microscopy. After the addition of LPS, WT DCs showed a significant increase of APC/T-cell conjugates, which peaked at 4 h to 10 h (Figure 7C). In contrast to WT DCs, DM−/− DCs pulsed with Sp1-Alexa 594 did not induce significant conjugate formation with CD4+ T cells.
The new paradigm of MHCII-restricted presentation of carbohydrates leaves open obvious questions regarding the precise mechanism of bacterial capsular carbohydrate interactions with MHCII molecules. Here we provide evidence that internalization of polysaccharides is followed by intracellular transport and presentation on the cell surface by newly synthesized MHCII molecules. We show that in DCs, Sp1 migrates in tubules as carbohydrate/MHCII complexes to the cell surface to induce T cell–dependent immune responses in vitro and in vivo. Sp1/MHCII retrograde transport requires the editor protein DM.
Intracellular tracking of Sp1 reveals partial co-localization with BCECF-dextran and Rab5, markers for early endosomes, that might reflect different pathways for Sp1-containing pinocytic and endocytic vesicles [33] or an intermediate status during fast maturation of Sp1-containing vesicles into late endosomes and lysosomes. In early endosomes, Sp1 could be subjected to oxidation by free radicals as was shown for PS A1 [12]. There is an increasing acidification of Sp1-containing intracellular compartments. We previously demonstrated that Sp1-induced T cell activation depends on the acidic lysosomal pH and that Sp1 induces maturation of human monocyte-derived DCs [11]. Besides proteases, lysosomes also contain abundant glycosidases, such as fucosidases and galactosidases [34]. At a later stage of the endocytic pathway at an optimal acidic pH of maturing DCs [15], glycosidases may trim Sp1 to smaller molecular sizes, forming conformations that facilitate Sp1 anchoring and binding to MHCII and promote optimal generation of T cell epitopes.
In contrast to non-charged dextrans, Sp1 is transported from lysosomes to the cell surface in MHCII-positive tubules like conventional protein antigens HEL and Ova [24,25]. Confocal, EPI/TIR fluorescence microscopy, and FACS analyses demonstrate time-dependent retrograde transport and cell surface presentation of Sp1 on maturing DCs, indicating fusion of Sp1-carrying tubules with the plasma membrane for formation of an immunological synapse required for proper activation of T cells. Presentation of Sp1 on the DC surface results in conjugate formation with a considerable number of T cells from non-Sp1–primed naïve mice. Although mDCs are known to attract and cluster with naïve T cells [35], it is possible that T cells from non–germ-free animals that are colonized with the ubiquitous gut organism B. fragilis are primed by the ZPS from B. fragilis and cross-react with Sp1 [36]. Indeed, about 8% of the naïve CD4+ T cells become activated by Sp1-treated DCs. The drastically different immunogenic properties of Sp1 are brought about by specific biochemical characteristics by which Sp1 distinguishes itself from other carbohydrates such as dextrans. At an acidic lysosomal pH similar to the Sp1 isoelectric point of 3.5 (unpublished observation), an optimal equilibrium of positively charged free amino groups and negatively charged carboxyl groups is available and provides a large number of binding sites to associate with MHCII molecules. The high density of alternating opposite charges is exposed on the outmost surface of the molecule. Maximum binding would be achieved via abundant electrostatic interactions supplemented by the potential for numerous hydrogen bonds to hydrophilic hydroxyls and van der Waals interactions.
The proper balance of TH1 and TH2 immunologic responses is critical to maintain balance in the immune system's task to fight microbial antigens. It has been shown that increased representation of CLIP/MHCII complexes antagonize polarization of T cells towards the TH1 phenotype [21]. Here, we show reduction of CLIP cell surface presentation in mDCs possibly caused by antigenic exchange of CLIP with Sp1 and/or modulation of MHCII/self-peptide retrograde transport by Sp1. Inhibition of presentation of CLIP/MHCII in Sp1-treated mDCs might contribute to the establishment of a TH1/TH2 balance towards the TH1 phenotype as has been described for the ZPS of the symbiotic intestinal bacteria B. fragilis [8]. It might also be responsible for low Sp1-specific antibody production (unpublished data) and might modulate the immune response to ZPS during abscess formation and adhesion.
Beyond the functional role of DM in peptide exchange, our data suggest an extension to antigenic exchange with carbohydrates. Three functions have been described for DM: 1) to catalyze the removal of CLIP or non-CLIP peptides and their exchange by heterogeneous peptides [17–19]; 2) to serve as a molecular chaperone, preventing non-specific aggregation of the temporarily empty αβ dimers following CLIP release [37]; and 3) to function as a peptide editor, positively selecting peptides that can stably bind to a particular class II allele [18,38]. Besides facilitating Sp1 binding through catalytic release of CLIP and other peptides, DM might select those Sp1 length variants for binding that form the most stable complexes due to their optimal structural and electrostatic features. Although in DM-deficient DCs accumulation of Sp1 in endocytic compartments, retrograde transport, and surface expression of MHCII complexed with either CLIP or self peptides is normal, they are inefficient in transporting Sp1 from endocytic compartments to the cell surface and initiating conjugate formation with naïve CD4+ T cells.
So far, it is not possible to rule out internalization and presentation of ZPS by recycling MHCII or presentation similar to superantigens [7,12]. Recycling MHCII binds to peptides in early endosomes and traffics between early endosomes and cell membrane. Indeed, Sp1 partially co-internalizes with transferrin (not shown), a marker for recycling receptor-mediated endocytosis. However, Sp1 is directed from early endosomal compartments to late endosomes and lysosomes, where it co-localizes with newly synthesized MHCII and does not locate in recycling endocytic Rab11b-positive compartments. Furthermore, the dependency of Sp1 biological activity on DM and on the retrograde tubular transport of MHCII clearly argues against a presentation mechanism similar to superantigens and by recycling MHCII molecules.
Taking the results together, we show that bacterial polysaccharide–induced APC/T-cell conjugate formation and T cell–dependent immune responses depend on retrograde transport via MHCII tubules and the functional role of DM.
S. pneumoniae type 1 capsular polysaccharide complex was obtained from the American Type Culture Collection (http://www.atcc.org) and further purified to obtain homogeneity as described previously [11]. High-resolution (500 MHz) proton NMR spectroscopy [5] revealed that Sp1 was free of contaminating protein and nucleic acids. Endotoxin was not detectable in Sp1 by the limulus test with sensitivity of <8 pg LPS/mg Sp1. As control antigens, ovalbumin–fluorescein isothiocyanate (FITC), ovalbumin–Alexa Fluor 594, dextran–Alexa Fluor 488, and dextran–Texas Red Molecular Probes were used (http://probes.invitrogen.com).
For intracellular tracking, BCECF-dextran, LysoTracker Red DND-99, dextran–Alexa Fluor 488, and ovalbumin-FITC were obtained from Molecular Probes.
Sp1 is a linear polymer of an average molecular size of 90 kDa corresponding to 167 trisaccharide repeating units with a respective molecular size of 537 Da. Each repeating unit of Sp1 contains one positively and two negatively charged groups with galacturonic acid (GalA, residues a and c) and 2-acetamido-4-amino-2,4,6-trideoxygalactose (Aat, residue b) with a sequence of →3)-α-D-GalA (a)-(1→3)-α-D-Aat (b)-(1→4)-α-D-GalA (c)-(1→ [5,39]. The adjacent hydroxyl groups on residue c (molecular weight 175) were oxidized by sodium m-periodate (Sigma, http://www.sigmaaldrich.com) treatment in molar ratios ranging from 1:0.1 to 1:0.5 for 90 min at room temperature in the dark to create highly reactive aldehyde functional groups [40]. The reaction was stopped by addition of ethylene glycol (Sigma). After gel filtration chromatography with a PD-10 column (Amersham, http://www.amersham.com), Sp1 was labeled by formation of covalent hydrazone linkages between aldehydes and EZ-Link Biotin-Hydrazide (Pierce, http://www.piercenet.com), Alexa Fluor 488 hydrazide, and Alexa Fluor 594 hydrazide (Molecular Probes) following the instructions of the manufacturer. After reduction of residual aldehydes of biotinylated Sp1 (Sp1-biotin), Alexa Fluor 488–labeled Sp1 (Sp1-Alexa 488), and Alexa Fluor 594–labeled Sp1 (Sp1-Alexa 594) by base treatment at pH 9.0 for 60 min, the glycoconjugate was separated from unbound labeling agents by three consecutive runs on PD-10 columns. The degree of biotinylation was determined with the ImmunoPure 2-(4′-hydroxyazobenzene benzoic acid) (HABA) and ImmunoPure Avidin (Pierce) reagents, following the instructions of the manufacturer. This method allows the calculation of mol biotin per mol Sp1 and number of biotin molecules per repeating units. Labeled Sp1 carried a biotin molecule on every 20th repeating unit (Sp1-biotin), which corresponds to one label per 11-kDa fragment by 1H NMR spectroscopy and showed the same chemical shifts as native Sp1 (Figure S3A and S3B). The additional signals obtained for Sp1-biotin originated from EZ-Link Biotin-Hydrazide (Figure S3B, upper spectrum). All mice challenged with Sp1-biotin developed intraabdominal abscesses to the same degree as native Sp1 (Figure S3C). Sp1 labeled with Alexa Fluor hydrazide 488 (Sp1-Alexa 488) or Alexa Fluor hydrazide 594 (Sp1-Alexa 594) preserved its intact structure and in vivo immune responses (not shown). These controls demonstrated that the biological activity of labeled Sp1 used in our studies is indistinguishable from that of unlabeled Sp1.
NMR spectra were obtained from a sample of 2 mg of purified Sp1, Sp1-biotin, or Sp1-Alexa 488, which was exchanged with 2H2O once and redissolved in 0.7 ml of 2H2O as described previously [5]. NMR experiments were performed on a Bruker DRX 500 instrument (Bruker, http://www.bruker.de) with a proton resonance frequency of 500.13 MHz. The 1H spectra were recorded at 80 °C in 2H2O using presaturation to suppress the water signal. Chemical shifts were referenced in relation to 1H2HO resonance at 4.36 ppm.
Animal experiments were performed in accordance with the guidelines of German animal protection legislation (license number 50.203.2-K 16,3/02). In abscess induction studies, B6129SF2/J (WT) and H2-Dmatm1Luc (DM−/−) [29] obtained from Charles River Laboratories (http://www.criver.com) were injected intraperitoneally with Sp1 (100 μg of Sp1 in PBS mixed with sterile cecal content adjuvant [SCCA]; 1:1 v/v, 0.2 ml total volume) [9]. Then, 24 h before challenge, 2 × 107 CD4+ T cells (>95% purity) from WT mice were adoptively transferred to DM−/− mice per intravenous route. Six days after challenge, mice were macroscopically examined for the presence of abscesses within the peritoneal cavity by two double-blinded examiners. Abscesses were isolated and their diameter was measured.
The cellular influx into the peritoneal cavity was assessed at 24 h following challenge with Sp1. As in abscess induction studies, 2 × 107 CD4+ T cells from C57BL/6 (WT) mice obtained from Charles River Laboratories were adoptively transferred to DM−/− mice per intravenous route 24 h before challenge. WT mice were challenged intraperitoneally with Sp1. DM−/− mice were either challenged intraperitoneally with Sp1, Sp1 plus 2 × 107 WT APCs, or 2 × 107 WT APCs alone. APCs were purified from the peritoneal lavage followed by CD4+ T cell depletion (<0.05% CD4+ T cells) of WT mice challenged 24 h before adoptive transfer. Mice underwent peritoneal lavage with 4 ml of ice-cold PBS. A total cell count was performed by trypan blue staining with a hemocytometer. Each sample was then analyzed by flow cytometry for different cell types. The absolute number of each respective cell type present was calculated by taking its respective frequency and multiplying it by the total number of cells per ml lavage obtained from each mouse. In each experiment, four to six mice per group were tested. The experiment was performed three times in an independent manner.
Frozen sections of abscesses were fixed in cold acetone for 10 min followed by blocking of endogenous peroxidase with peroxidase blocking solution (DAKO, http://www.dako.com) for 10 min at room temperature. The CD11c antibody (N418, supernatant; 1:100 diluted) was then overlayed and the slides incubated in a humid chamber for 45 min. With TRIS washes between every step, a biotinylated link antibody (Becton Dickinson, http://www.bdbiosciences.com) was applied for 45 min followed by a streptavidin-alkaline phosphatase (DAKO) for 10 min. After another wash, the substrate (Vector NovaRed; Vector Laboratories, http://www.vectorlabs.com) was added and the slides were incubated in the dark for 20 min. After a TRIS wash, the slides were counter stained, mounted, and viewed using a Zeiss Axiophot microscope with photographic capabilities (http://www.zeiss.com).
DCs were generated from mouse bone marrow by adapting a previously described method [41]. In brief, bone marrow cells from H2-Dmatm1Luc [29] (DM−/−) and C57BL/6 (WT) mice that have a mutation that abolishes production of the MHCII I-Eα chain [31,42] were cultured in RPMI supplemented with 5% FBS, 500–1,000 U recombinant mouse granulocyte/macrophage-colony stimulating factor (GM-CSF), 20 μg per ml gentamicin, and 50 μM 2-mercaptoethanol (DC medium). DC medium was exchanged in two-day intervals. DCs were isolated by magnetic cell sorting with a CD11c-specific monoclonal antibody (mAb) (Miltenyi Biotec, http://www.miltenyibiotec.com). CD11c-positive iDCs were imaged on days 4 and 5 of culture. mDCs were generated by adding LPS O26:B6 (100 ng/ml) (Sigma) to disaggregated and replated cultures.
Infection of proliferating precursors with retrovirus containing I-Eα-EGFP was performed on day 2 by adapting a method described previously by Chow and coworkers [24]. I-Eα-EGFP, kindly provided by I. Mellman, was cloned into LZRS-pBMN using EcoRI sites. This viral vector was transfected into ΦNX-ecotropic cells using calcium chloride. Virus was collected in DC medium for 24 h, supplemented with polybrene and HEPES, and added to the DC culture for infection. Cells were spun at 32 °C, 2,500 rpm for 2 h. Virus was removed and fresh medium added. Expression was assayed 48 h after infection.
Localization of Sp1 in recycling, early, and late endosomes was performed by adenoviral infection of live dendritic cells with Rab11b-EGFP, Rab5-GFP, and Rab7-GFP–containing mammalian expression plasmids [43]. DNA fragments containing Rab fusion protein constructs and CMV promoter were amplified in the double digested promoter-less transfer vector pEntry148AU-MCS. For packaging into adenovirus particles, constructs were recombined into pAd/PL-DEST vector (Invitrogen, http://www.invitrogen.com). Adenoviral stocks were produced in 293A cells after transfection with plasmid DNA and Lipofectamine. Infection of DCs with Rab5-EGFP, Rab7-EGFP, and Rab11b-EGFP was performed on day 4. iDCs were spun at 2,500 rpm at 37 °C for 120 min in virus-containing medium supplemented with 10 mM HEPES. After replacement with DC medium, expression was checked by fluorescence microscopy after 24 h. Transfection efficiency was 50% to 70%.
Human DCs were differentiated from peripheral blood mononuclear cells as described [44]. Monocytes were isolated from peripheral blood mononuclear cells by positive selection by anti-CD14 magnetic beads (Miltenyi Biotec) and cultured in complete RPMI medium containing 50 ng/ml GM-CSF and 3 ng/ml IL-4. Maturation was induced at day 4 by the addition of LPS from Salmonella abortus equi (1 μg/ml) (Sigma).
Staining of surface molecules was performed using PE- or FITC-conjugated anti-CD4 (clone L3T4), anti-CD11c (clone Hl3), anti-CD69 (clone HI.2F3), anti-I-Eαβ (clone 14.4.4S), anti-I-Aβ (clone AF6–120.1), anti-HLA-DR/CLIP complex (clone Cer.CLIP), and anti-HLA-DR (clone L243) (BD Pharmingen, http://www.bdbiosciences.com). For Sp1 surface presentation studies, iDCs were incubated for 30 min at 37 °C without or with Sp1-biotin (200 μg/ml), washed, and incubated for 30 min to 8 h at 37 °C in LPS-containing medium (100 ng/ml) before staining with streptavidin-FITC (Sigma) for 30 min and washing. Cells prepared for flow cytometry were analyzed—after gating for viable cells by forward and side scatter and by propidium iodide staining—by FACScan (Becton Dickinson) using CELLQuest software (Becton Dickinson). The results were expressed as MFI, or as percentage (%) of fluorescence-labeled APCs of the whole APC population.
To investigate intracellular trafficking of Sp1, live cell imaging was performed. Cells were plated for 30 min on poly-d-lysine–pre-coated number 1.5 coverslips attached to 35-mm dishes (MatTek, http://www.mattek.com), and fresh medium was added. To study mechanisms of internalization of Sp1, cells were incubated with competitors or chemical inhibitors for 30 min at 4 °C or 37 °C before Sp1 treatment. To monitor Sp1, APCs were loaded with markers for cellular compartments before or at the same time point of Sp1 addition. Cells were washed before and after Sp1 treatment three times in ice-cold medium. Inverted fluorescent microscopy was performed on an Olympus IX81 microscope (http://www.olympus-europe.com/microscopes/index.htm). Temperature control at 37 °C was achieved with a heating dish. Acquisition was performed using AnalySIS Imaging System software (Olympus, http://www.olympus.de). Confocal microscopy was done on a PerkinElmer UltraView LCI spinning disc system (http://las.perkinelmer.com) equipped with a suitable multi-band beamsplitter and a MellesGriot Omnichrome 643-RYB-A02 ArKr gas laser (http://www.mellesgriot.com) providing 488-nm and 568-nm lines for excitation. A Nikon Plan Fluor ×100 1.3NA oil immersion objective (http://www.nikon.com) and 525/50 and 607/45 emitter filters were used for GFP, FITC, Alexa 488, and Alexa 594, and Texas Red stains, respectively.
Multi-color TIR-FM and EPI was performed on an Olympus Biosystems Cell-R system equipped with a stabilized Xenon arc lamp and dual coupling for Coherent Sapphire 488–20 and Compass 250M-50 diode lasers (http://www.coherent.com) providing 488 nm and 532 nm excitation light, respectively. On confocal microscopy, EPI, and TIR-FM systems, environmental condition was controlled by a custom incubator (EMBL GP 168) that provides a 37 °C and 5% CO2 atmosphere. Images were exported to TIFF images, processed using Adobe Photoshop version 6.0 (http://www.adobe.com), and converted into QuickTime movies using Graphic Converter version 3.8 (Softguide, http://www.softguide.de).
Investigation of DC/T-cell conjugate formation was performed as previously described [45]. In brief, DCs from C57BL/6 WT and DM−/− mice were loaded with Sp1-Alexa 594 for 30 min or left untreated and washed. DCs (0.5 × 105/ml) were then treated with LPS (100 ng/ml) for different time intervals, washed, and mixed with CFSE-labeled CD4+ T cells (1.5 × 105) from C57BL/6 WT mice. Cells were centrifuged for 5 min at 50g to increase cell interactions, and incubated at 37 °C for 20 min. The cells were gently transferred to poly-d-lysine–pre-coated number dishes. After incubation at 37 °C for 30 min, T-cell–DC conjugates were subjected to imaging by fluorescent microscopy. CFSE-labeled T cells were distinguished from GFP-labeled DCs by morphology. Three independent experiments were performed and the number of CFSE-positive CD4+ T cells interacting with 100 Sp1-Alexa 594–positive DC was counted in a blinded manner as previously described [22].
For the investigation of WT CD4+ T cell activation induced by Sp1-treated DCs, the same protocol as for DC/T-cell conjugate formation was applied with some modifications. DCs were treated with 100 μg/ml Sp1 for 45 min or left non-treated. Analysis of the expression of CD69, the early activation marker on CD4+ T cells, was performed by flow cytometry at different time points after addition of T cells to DCs.
T2.DR4.DM transfectants, expressing the MHCII molecules HLA-DR4 and DM, respectively, were maintained in RPMI 1640 supplemented with 10% FCS. T2 is a BxT cell hybrid with a large deletion in the MHCII locus and does not express endogenous MHCII proteins. Cells at a density of 6 × 105 cells/ml were treated with Sp1 (200 ug/ml) for 20 h, washed with PBS, and lysed (6 × 106/ml) at 4 °C in lysis buffer of 20 mM and 5 mM MgCl containing 1% Triton X-100 and protease inhibitors. The cells were precipitated with mAb L243 (recognizing antigen/HLA-DR complexes) conjugated to sepharose beads. Peptides were eluted with 0.1% trifluor-acetic acid. MALDI-MS analysis was done as described [21] on a Reflex III mass spectrometer (Bruker).
Comparison of groups with regard to abscess formation was made by chi-square analysis. Results of the various groups in peritoneal cellular influx and APC/T-cell engagement assays were compared by Student's t test. |
10.1371/journal.pcbi.1002031 | Identification of Hammerhead Ribozymes in All Domains of Life Reveals Novel Structural Variations | Hammerhead ribozymes are small self-cleaving RNAs that promote strand scission by internal phosphoester transfer. Comparative sequence analysis was used to identify numerous additional representatives of this ribozyme class than were previously known, including the first representatives in fungi and archaea. Moreover, we have uncovered the first natural examples of “type II” hammerheads, and our findings reveal that this permuted form occurs in bacteria as frequently as type I and III architectures. We also identified a commonly occurring pseudoknot that forms a tertiary interaction critical for high-speed ribozyme activity. Genomic contexts of many hammerhead ribozymes indicate that they perform biological functions different from their known role in generating unit-length RNA transcripts of multimeric viroid and satellite virus genomes. In rare instances, nucleotide variation occurs at positions within the catalytic core that are otherwise strictly conserved, suggesting that core mutations are occasionally tolerated or preferred.
| The expanding diversity of noncoding RNA discoveries is revealing a broader spectrum of roles RNA plays in cellular signaling and in biochemical functions. These discoveries in part are being facilitated by the expanding collection of genomic sequence data and by computational methods used to search for novel RNAs. In addition to searching for new classes of structured RNAs, these methods can be used to reevaluate the distributions of long-known RNAs. We have used a bioinformatics search strategy to identify many novel variants of hammerhead self-cleaving ribozymes, including examples from species in all three domains of life. New architectural features and novel catalytic core variants were identified, and the genomic locations of some hammerhead ribozymes suggest important biological functions. This ribozyme class promotes RNA cleavage by an internal phosphoester transfer reaction by using a small catalytic core. The simple sequence and structural architecture coupled with the general utility of RNA strand scission may explain its great abundance in many organisms.
| Hammerhead ribozymes [1] represent one of five distinct structural classes of natural self-cleaving RNAs identified to date [2]. The first hammerheads were discovered in viroids and plant satellite RNA viruses where they process RNA transcripts containing multimeric genomes to yield individual genomic RNAs [1], [3], [4]. Representatives of this ribozyme class have been studied extensively for the past 25 years because their small size and fundamental catalytic activity make them excellent models for RNA structure-function research [5].
Although a minimal three-stem junction constitutes the catalytic core of the ribozyme (Figure 1A), additional sequence and structural elements form an extended hammerhead motif [6], [7] that yields robust RNA cleavage activity under physiological concentrations of Mg2+. Specifically, tertiary interactions form between the loop of stem II and either an internal or terminal loop in stem I that increase activity of the core by several orders of magnitude under low magnesium conditions. Identification of this tertiary substructure in high-speed hammerhead ribozymes [5], [8] resolved a long-standing paradox between biochemical data and atomic-resolution structures of minimal hammerhead ribozymes [5].
Several searches for new examples of hammerhead ribozymes have been performed previously [9]–[11] by taking advantage of the wealth of knowledge derived from mutational and biochemical analyses of various hammerhead ribozymes. By carefully establishing descriptors of the minimum functional consensus motif, dozens of new hammerhead representatives have been found in the parasitic worm Schistosoma mansoni [12], Arabidopsis thaliana [13], in mouse [14] and very recently in bacteria and human [15], [16]. A similar bioinformatics search for RNA structures homologous to hepatitis delta virus (HDV) ribozymes [17] revealed that representatives of this self-cleaving ribozyme class are far more widely distributed in many organisms. Moreover, among numerous noncoding RNA candidates revealed by our recent bioinformatics efforts was a distinct architectural variant of hammerhead ribozymes (see below). Given these observations, we speculated that far more hammerhead ribozymes may exist in the rapidly growing collection of genomic sequence data.
Using a combination of homology searches we found thousands of new hammerhead ribozyme sequences in all domains of life. These ribozymes are observed in the eubacterial and archaeal domains, as well as in fungi and humans. Moreover, many of the newfound hammerhead ribozymes exploit a pseudoknot interaction to form the tertiary structure necessary to stabilize the positioning of stems I and II. We also identified a number of active sequence variants that suggest the hammerhead consensus is more variable than previously thought.
Although the biological functions of these hammerhead ribozymes remain unproven, some could be involved in gene regulation based on their genomic contexts, similarly to what has been proposed for the mouse hammerhead and human HDV ribozymes [14], [17], [18]. Although glmS ribozymes [19] are known to control gene expression by using a metabolite as an active site cofactor to promote mRNA cleavage, gene regulation by other ribozymes such as the hammerhead might rely on protein- or small-molecule-mediated allosteric control of self-cleavage activity.
We used a comparative genomics pipeline [20] integrating homology searches [21] and the algorithms RNAMotif [22] and CMFinder [23] to identify structured RNAs in available sequences [20]. In addition to many novel motifs, we identified numerous examples of RNAs that conform to the well-established consensus sequence for hammerhead self-cleaving ribozymes (Figure 1A). We eventually conducted a comprehensive search of all available genomic DNA, which allowed us to expand the collection of hammerhead ribozymes from ∼360 previously known examples to more than 10,000 (Figure 1B; see sequence alignments in Dataset S1).
A large number of additional hammerhead ribozymes were identified in metazoans, including mosquitoes and sea anemones. While many hammerhead ribozymes associated with repeated elements were previously found in various species of Dolichopoda (cave crickets) [24], [25] and Schistosoma mansoni (parasitic worm) [12], they represent only a small fraction of all occurrences. Aedes aegypti (mosquito), Nematostella vectensis (sea anemone), Xenopus tropicalis (frog) and Yarrowia lipolytica (fungus) also appear to have hammerhead ribozymes associated with interspersed repeat elements, which are found in multiple copies in their genomes. Interestingly, we uncovered instances of this motif in humans, and the genetic contexts of two of these are conserved among many mammals.
Our search efforts also extended the range of known hammerhead ribozymes beyond the eukaryotic domain of life. At least three representatives are present in archaea and hundreds are present in bacteria (Figure 1B), where many are in proximity to integrase genes frequently grouped with prophages. Although the precise biological functions of these hammerheads remain unclear, the fact that nearly all carry conserved sequence and structural features (Figure 1A) previously proven to promote RNA cleavage by internal phosphoester transfer suggests that they also promote high-speed RNA cleavage. Almost without exception, the catalytic core of each representative matches the consensus hammerhead sequence. Also, the three base-paired stems enclosing the catalytic core typically show variability in sequence and length, with stem II commonly formed by as few as two base pairs.
However, several novel features for this ribozyme class were observed among the expanded list of representatives. Among the notable variants are the first natural examples of “type II” hammerhead architectures (Figure 2A), wherein stems I and III are closed by hairpin loops while stem II lacks a loop. Although type II hammerheads are functional [12], they were paradoxically thought to be absent in nature. Our findings reveal that all three circularly permuted architectures indeed are common in nature. Specifically, type II hammerhead ribozymes are very common in eubacteria and are also present in some archaeal species.
The type II hammerhead consensus identified in our bioinformatics search included a putative conserved pseudoknot linking the loop of stem I with the 3′ tail extending from the right shoulder of stem II (Figure 2A). Indeed, a majority of type II hammerhead motifs have potential pseudoknots of four or more base pairs between loop I and nucleotides immediately downstream of stem II. On further examination we found that pseudoknots can be formed by numerous representatives of all three hammerhead types (Figure 2B), suggesting that the tertiary structure required to stabilize the parallel assembly of stems I and II is commonly achieved by this base-paired substructure (see below).
On many occasions, multiple hammerhead ribozymes are arranged in close proximity to flank individual genes or short blocks of genes in prophage genomes, although the identities of these genes are not constant (Figure 3). Such arrangements imply that long bacteriophage RNA transcripts may be processed into operon- or single-gene-length mRNAs, although other possibilities exist. For example, some hammerheads may not be functional, or successive ribozyme-mediated cleavage and ligation reactions could yield spliced or circular RNA products, although we were unable to detect either type of product in this study (data not shown).
Three tandem hammerhead arrangements from Clostridium scindens, Azorhizobium caulinodans (Figure S1) and Agrobacterium tumefaciens (Figure S2) were tested for cleavage activity during in vitro transcriptions of constructs corresponding to ∼2 kb fragments of the native polycistronic RNAs. In each case, cleavage products were observed that correspond to the sizes expected if all ribozymes were active and efficiently promoted self-cleavage reactions.
Ribozymes from the triple hammerhead arrangement of A. tumefaciens flank ardA, a gene involved in protecting phages from bacterial restriction enzymes, and another gene of unknown function (Figures 3 and S2). These ribozymes exhibit self-cleavage activity in vivo following cloning and transcription of the appropriate A. tumefaciens DNA fragment in E. coli (Figure S3). Although the biological purpose of this triple arrangement is unknown, the ardA gene is located immediately downstream of a hammerhead ribozyme in three strains of Legionella, suggesting that ribozyme action may be important for this gene.
Previous studies demonstrated that non-Watson/Crick contacts between the terminal or internal loops in stems I and II play a critical role in forming the tertiary structure necessary for high-speed hammerhead function [6], [7]. However, many newfound hammerhead representatives instead are predicted to use a pseudoknot interaction to stabilize the parallel alignment of stems I and II (Figure 2). This prediction was assessed by conducting a series of RNA cleavage assays using various mutants of the type II hammerhead ribozyme from a metagenome dataset (Figure 4) and from several other sources (see Figure 2).
A bimolecular construct based on the wild-type (WT) ribozyme sequence exhibits an observed rate constant (kobs) for RNA cleavage of greater than 1.2 min−1 under single-turnover conditions and simulated physiological conditions (23°C, 0.5 mM MgCl2, 100 mM NaCl, 50 mM Tris-HCl [pH 7.5 at 23°C]). All deletions or other mutations that are predicted to disrupt the pseudoknot substructure drastically reduce cleavage activity (Figure 4). For example, deleting two nucleotides from the 3′ terminus to reduce the pseudoknot from six to four base pairs caused the kobs to decrease by a factor of ∼20, and deleting another two nucleotides from this terminus reduced activity by a factor of more than 100 compared to WT. Mutating the pseudoknot has a similar effect, while the compensatory mutation restores high activity (Figure 4).
Deletions or insertions of nucleotides surrounding the pseudoknot also reduced kobs values by orders of magnitude. Moreover, stabilizing stem I by adding two base-pairs, or stabilizing stem II by adding one additional base-pair also decreased ribozyme activity substantially. All of these mutations are located outside of the highly conserved ribozyme core and are designed to promote local structure formation. However, these mutations change the relative positions of nucleotides that form the pseudoknot, which likely disrupts the proper orientation of this tertiary structure critical for high-speed activity.
We also assessed pseudoknot formation by subjecting the longer of the two strands that form the bimolecular construct to in-line probing [26], which is an assay that can be used to map structured versus unstructured portions of RNA molecules. The pattern of spontaneous RNA fragmentation is consistent with formation of the pseudoknot in the absence of the second strand (data not shown). Likewise, in-line probing of this portion of bimolecular constructs from two other hammerhead ribozymes indicates that pseudoknot formation occurs even in the absence of the remaining portion of the ribozyme (data not shown).
All hammerhead ribozymes representatives were examined for the presence of a pseudoknot contact between stems I and II, revealing that approximately 40% likely use this structural constraint (Figure S4). Stem lengths appear constrained by this base pairing (Figure 4), but the constraints do not follow a simple rule and seem to vary for different types of hammerheads. The identification of pseudoknot interactions between these two substructures expands the known tertiary interactions described previously [6], [7] that are essential for high activity. However, there are many hammerhead ribozyme examples that do not appear to use these contacts, suggesting other types of interactions exist or that none are used in some cases (Figure S5).
The importance of conserved catalytic core nucleotides has been well established by numerous previous studies [5], including the use of systematic mutational analyses [27] and in vitro evolution [28], [29]. Some RNAs with core mutations do retain modest levels of cleavage activity, but the decreases are generally assumed to render the ribozyme biologically non-functional. Despite the fact that the core is exceptionally well conserved, three hammerhead ribozymes previously identified from viroids have core nucleotides that deviate from the consensus [30], [31], suggesting some changes do preserve biological function (shown in blue, Figure 5A and 5B).
Our expanded collection of hammerhead representatives revealed additional examples of core variation (Figure 5A). Most of the known interactions and important chemical groups within the core are minimally affected in these variants. However, some interactions predicted to be important based on atomic-resolution structural models are disrupted in some cases. Several ribozymes with variant core sequences were assayed to determine how these changes affect RNA cleavage activity.
Some of these variant cores carry compensatory changes that prevent severe alteration to the active structure (Figure 5). For example, core nucleotides C3 and G8 form a base pair, and these nucleotides covary to U3 and A8 in several hammerhead ribozyme examples. Ribozymes containing covarying nucleotides at these positions had already been proven to be active in vitro [32], but covariation at these positions had not previously been observed in nature.
A hammerhead sequence found in an intergenic region of bacteriophage Bcep176 (Figure S6) carries an A6C variation that is expected to disrupt at least one hydrogen bond and potentially two. Correspondingly, we observe a kobs of less than 0.1 min−1, which is in agreement with the low activity that a previous mutational analysis of the core revealed for changes at this position [27]. Similarly, low activity of an insertion observed after A6 (called A6a in Figure 5B) is consistent with the fact it should disrupt a hydrogen bond observed in the crystal structure because the phosphate connecting A6 to N7 interacts with U4. Changing the backbone conformation at this position would be expected to be detrimental to an active core.
An insertion is likely to be easier to accommodate if the phosphate backbone is protruding out of the otherwise compact structure. Thus U13a, (Figure 5A and 5B) which is inserted in the “GAAA” region of the core, could point outside of the core, resulting in minimal structural change. A sequence with U15.1–A16.1 instead of A-U, usually considered essential, self-cleaves, albeit less efficiently than a typical hammerhead ribozyme. This is likely caused by the loss of an interaction observed between A15.1 and G5. WT ribozymes have been shown to exhibit at least 10-fold greater activity compared to mutants at nucleotides 15.1 and 16.1 examined in previous in vitro studies [27], [33].
The activities of these core variants are consistent with the findings of previous biochemical studies that assessed the importance of individual chemical groups for activity. For example, the U15.1–A16.1 and C6 mutations are expected to disrupt the core, and did result in low, but detectable, activity. Additionally, for some predicted ribozymes that have mutations expected to be highly disruptive, no activity was detected (Figure S7).
In addition to exhibiting variation of the core, some hammerhead ribozymes have very weak stems. In particular, stem II often consists of only two base-pairs and even a single base-pair in one case (Figure 4). It is even more surprising that stem II can start with a U10.1–U11.1 mismatch (Figure 5) since this is the most conserved base-pair of the hammerhead consensus, aside from A15.1–U16.1 (Figure 1A). However, this U-U mismatch had already been shown to support higher levels of cleavage activity than any other mispaired combination [27]. Weak stems III were also very common (Figure S8).
Several hammerhead ribozyme representatives were identified among sequences derived from viral fractions of solar salterns (see sequence alignments in Dataset S1). Solar salterns consist of a series of interconnected pools of increasing salinities, and culminate in crystallizer ponds from which various salts are precipitated and harvested. These saturating brines are inhabited predominantly by extreme halophiles of the archaeal domain, and these organisms contend with the acute hypersaline environment primarily by maintaining high intracellular concentrations of K+ ions [34]. Therefore, we speculated that hammerhead variants from this source might become active in high salt.
Three of the hammerhead examples from this environment carry short insertions in the catalytic core near the C3 nucleotide and P1 stem (Figure 6A). Such changes in this local region of the catalytic core are unprecedented among reported examples of hammerhead ribozymes. Furthermore, based on the atomic resolution structure of the hammerhead active site [8], insertions of this type are expected to destabilize the catalytic core. It is important to note that one of the sequences derived from saltern metagenomes had a typical consensus, so it appears that alteration of the catalytic core is not a requisite feature of hammerhead ribozymes from extremely halophilic environments. However, these three unusual variants were found only in the genomes and metagenomes of solar salterns.
To examine whether these alterations of the catalytic core reflect adaptations to hypersaline conditions, we prepared wild-type and mutant versions of HHmeta (Figure 6A) derived from saltern metagenomic data. Only very low levels of self-cleavage activity were detected for HHmeta during transcription in vitro (data not shown), despite the presence of 15 mM MgCl2. In contrast, an engineered mutant in which the two-nucleotide insertion (G2a and C2b) was removed to create a consensus catalytic core undergoes nearly quantitative self-cleavage during transcription (data not shown). Thus, the unusual insertion sequence in this saltern-derived hammerhead ribozyme impairs cleavage activity under standard assay conditions.
To test whether elevated salt concentrations can rescue this deficiency, we first determined the kobs for self-cleavage of HHmeta in a high concentration of monovalent ions alone. HHmeta undergoes self-cleavage with a kobs of 4×10−4 min−1 in 4 M LiCl (data not shown). For comparison, a consensus hammerhead ribozyme catalyzes strand scission with a kobs of 0.17 min−1 under similar conditions [35], a kobs that is 425-fold faster than that of the saltern-derived variant.
To assess whether more appreciable activity of HHmeta requires elevated levels of divalent metal ions, we measured kobs values over a range of Mg2+ concentrations. The activity of the variant increases with increasing Mg2+ levels (Figure 6B), mirroring the behavior of consensus hammerhead ribozymes [36]. However, HHmeta requires substantially higher Mg2+ concentrations to achieve comparable kobs, such that a Mg2+ concentration of 300 mM is necessary to attain a kobs of ∼0.13 min−1. Values for kobs are slightly improved at higher Mg2+ concentrations when reactions are supplemented with 3 M KCl (Figure 6B), with the monovalent ions likely providing additional structure stabilization. Conversely, the added KCl results in slightly decreased kobs values in the lower range of Mg2+ concentrations, due presumably to competition with Mg2+-binding sites [37]. Nonetheless, it is clear that the concentration of Mg2+, and not that of monovalent cations, has the most pronounced effect on the self-cleavage activity of HHmeta. Mg2+ ions are smaller and more densely charged than monovalent ions, and thus might more effectively stabilize the active structure of HHmeta through low-affinity, diffuse interactions [38]. Elevated Mg2+ concentrations might be important for global folding of HHmeta, or could be necessary to compensate for the putative destabilized active site of the variant. It is also possible that Mg2+ ions provide a larger direct contribution to catalysis in HHmeta than in consensus hammerhead ribozymes.
Our homology searches reveal the presence of nine regions in human genomic DNA that conform to the consensus for hammerhead ribozymes (see sequence alignments in Dataset S1). Two candidates (Figure 7A and 7B) appear to be conserved among some other vertebrates, and therefore were chosen for experimental validation. These two candidates are the same that have been reported very recently [15]. Robust self-cleaving activity of one representative, termed “C10 hammerhead”, was observed during in vitro transcription for both human and pig sequences (Figure 7C). As do many new-found hammerhead ribozymes noted above, this RNA appears to use pseudoknot formation to stabilize the active structure. As expected, a truncated form of the ribozyme that lacks the five base-pair pseudoknot is inactive when assayed at 0.5 mM MgCl2 (data not shown).
The C10 hammerhead is found within an intron in the 5′ untranslated region (UTR) of C10orf118 (Figure 7D), which is a gene of unknown function that is conserved throughout mammals. The C10 hammerhead is present in all examined sequenced mammalian species with the exception of mouse and rat, which do not carry an intron in the 5′ UTR of this gene. The biological significance of C10 hammerhead self-cleavage is not clear. Genbank and GeneCards EST data indicate that the RNA is expressed in at least 18 tissues [39], [40] (Figure S9), and RT-PCR on the first exon of C10orf118 yields product that demonstrate expression of the gene in four human cell lines (Figure S9). One possibility is that cells control 5′ UTR splicing by controlling hammerhead action.
The second human hammerhead we subjected to further analysis, termed “RECK hammerhead”, resides in an intron of the gene for RECK (reversion-inducing cysteine-rich protein with Kazal motifs), a negative regulator of certain metalloproteinases involved in tumor suppression [41]. This arrangement is conserved in all mammals and birds examined (Figure 7E). The ribozyme appears to lack a pseudoknot, but perhaps interactions between loop II and stem I substitute for this tertiary contact as is observed for many hammerhead representatives. The RECK hammerhead also tested positively for cleavage in vitro (Figure 7C). According to EST data (I.M.A.G.E. consortium) [42], the exons flanking the hammerhead-containing intron appear to be alternatively spliced, and are usually absent from RECK transcripts expressed in nervous system tissue, although they are present in the corresponding RNAs from most other tested tissues. Interestingly, two ESTs from Bos taurus have sequences corresponding exactly to the hammerhead's 3′ cleavage product fused with those matching RNA components of U snRNPs (U5 and U6, EST accession numbers are DV870859.1 and DV835419.1), suggesting that this ribozyme may be active in vivo.
The application of increasingly powerful bioinformatics algorithms to the expanding collection of DNA sequence data is facilitating the discovery of novel noncoding RNAs and revealing new locations for previously known examples. A recent report [17] revealed additional representatives of the HDV self-cleaving ribozyme class, which are widely distributed among many organisms. Previously, this ribozyme had been considered one of the least commonly occurring of the self-cleaving RNA classes. In the current study, we expand the number of reported hammerhead ribozymes by more than an order of magnitude compared to what was known previously, and we have identified members of this ribozyme class in all domains of life. Our findings strongly suggest that hammerhead ribozymes comprise the most abundant self-cleaving ribozyme class in nature. Almost simultaneously, three groups have recently used computational methods to discover additional hammerhead ribozymes. These efforts revealed hammerhead ribozymes in bacteria and various eukaryotes, although their methods differed from ours and were not used to identify variants from the consensus [15], [16], [43], [44].
Previous in vitro selection studies demonstrated that hammerhead ribozymes are among the first self-cleaving motifs to emerge from random-sequence populations [45], [46]. These findings suggest that this is one of the simplest ribozyme architectures that can cleave RNA efficiently and that this simplicity ensures multiple evolutionary origins. This latter conclusion also is supported by our observation that type I, II and III hammerhead motifs are very common, which would be unlikely if all hammerhead ribozymes descended from a single founding example of a given type.
Although the hammerhead consensus is highly conserved, there are rare instances in which the catalytic core is altered. Previous studies have established that mutations at most positions in the core resulted in drastic loss of activity [5], [27], and consequently such variants are not expected to be found in nature. Nevertheless, three divergent cores were previously shown to exhibit self-cleavage activity [30], [31], and we add eight additional variants to this collection (Figure 5A, 5B and 6). It is likely that any adverse effects resulting from the variant cores are offset by stabilizing influences from tertiary contacts outside the active site, which would permit physiologically relevant activities of these natural variants. Consistent with this hypothesis is the observation that the U4C variant that considerably decreases activity in vitro maintains sufficient activity in vivo to permit viroid infectivity [30].
The diversity of structural alternatives observed in our hammerhead collection hints at the inherent difficulty in any effort to comprehensively identify ribozyme representatives. Including more core variations or distal structure variations in search outputs will result in larger numbers of false positives. Given the simplicity of the motif, sequences that conform to the consensus are expected to occur by chance in large sequence databases, even if some of them might be incapable of folding into an active hammerhead.
Although numerous hammerhead examples can be discovered by comparative sequence analyses, the identification of those that are biologically relevant will ultimately require experimentation in vivo. For example, viroid hammerhead sequences can experience mutations at high frequency, and most of these mutations result in non-infectious phenotypes [47], but some are still infectious in spite of a less active ribozyme [30]. In this study, we tested 18 hammerhead ribozymes conforming to the consensus, with 14 exhibiting activity in vitro. No cleavage was detected for the remaining four examples under our assay conditions, although two of these inactive RNAs are derived from Aedes and Nematostella, organisms in which active hammerheads might require dimeric conformations (Figure S8). It is thus possible that these ribozymes follow a more complex folding pathway that is more difficult to reproduce experimentally. However, some other inactive candidates are more likely to be false positives, such as a putative type II hammerhead in humans, which lacks conservation of the hammerhead structure in closely related species (see sequence alignments in Dataset S1).
The previous absence of known natural examples of type II hammerheads suggested that this architecture might not be biologically useful. However, our findings demonstrate that all types of hammerhead ribozymes are exploited naturally. Nevertheless, the vast majority of hammerhead ribozymes associated with repeated genetic regions in eukaryotes are of type I. This is most likely due to the evolutionary origin of the repeats, wherein the initial sequence carried a type I hammerhead that was widely propagated. Alternatively, it is possible that repeat propagation may require a type I hammerhead architecture. For example, if the ribozyme was involved in cis cleavage and trans ligation reactions to DNA, then type I ribozymes are the only architecture that would provide a 2′,3′-cyclic phosphate terminus and the bulk of the catalytic architecture to ligate to a separate nucleic acid strand carrying a 5′ hydroxyl group. This ligation reaction between RNA and DNA with a type I hammerhead architecture has been previously demonstrated [48]. This is only an example of how type I hammerhead could have been favored.
Based upon the abundance of hammerhead motifs we find associated with DNA repeats, self-cleaving ribozymes appear to be especially common in selfish elements (Figure 1B). This trend is also evident for group I and group II self-splicing introns [49], which commonly are associated with selfish elements. Moreover, other self-cleaving ribozyme classes may have similar distributions, as is evident from the recent report of HDV ribozyme representatives associated with R2 retrotransposons [50]. A possible outcome of these arrangements is that some selfish element harboring a ribozyme will occasionally integrate at a site where the ribozyme provides a selective advantage to the host. Strongly suggestive of this scenario is the striking similarity between the two most conserved vertebrate hammerhead ribozymes (pink regions Figure 7A and 7B) and the repeat-associated hammerhead sequences found in Xenopus (see AAMC01XXXXXX accession numbers in sequence alignments of Dataset S1). Hence, a hammerhead-containing element in an ancestral amphibian, apparently still active in some contemporary frogs, might have been retained in C10 and RECK introns because of advantages provided by self-cleavage at these sites, but would have been lost at most other positions.
The hammerheads in viroids process multimeric genomic RNAs, and in such cases constitutive RNA cleavage may be desirable. However, it is possible that some of the hammerheads of retroelements or bacteriophages will have more diverse functions, such as regulated RNA cleavage. This seems likely for the two validated hammerheads found in human introns, wherein the utility of constitutive cleavage activity would be difficult to rationalize. It is notable that the hammerhead ribozyme recently reported in mouse [14], [51] has a very large loop structure that could be naturally exploited for ribozyme control [18]. Similarly, the slower ribozyme variants in bacteriophages might become more active under the appropriate physiological conditions or upon interaction with molecular signals.
For some hammerhead variants such as HHmeta, activity may be facilitated by extreme salt concentrations. In vitro assays reveal that HHmeta requires at least 75 mM MgCl2 to attain biologically relevant kobs values (greater than 0.1 min−1). For most organisms, this divalent magnesium concentration is not attained. However, for microbes inhabiting certain environments, such as the Dead Sea or high salinity zones of solar salterns, growth has been reported in extracellular MgCl2 concentrations ranging from 0.6 to >2 M [52], [53]. Importantly, for certain extreme halophiles grown in medium containing 0.75 M Mg2+, estimates of the intracellular Mg2+ concentrations range as high as 0.42 M [54]. Such a high-salt environment for HHmeta might relax the need for strict conservation of the catalytic core. The variant hammerhead may thus function constitutively in an extremely halophilic host, perhaps fulfilling an RNA processing role.
Alternatively, it is possible that HHmeta and related variants have been selected to function as gene control elements that modulate the expression of associated genes in response to fluctuating intracellular salt concentrations. HHmeta was identified in a metagenome survey as part of a short sequence fragment, and therefore its genomic context is unknown. However, the structurally analogous hammerhead ribozyme variant HHphage (Figure 6A), which resides within the completed genome sequences of haloviruses HF1 and HF2, is in each case positioned only 13 nucleotides upstream of the start codon corresponding to an ATP-dependent DNA helicase. HF1 and HF2 are highly related lytic bacteriophages targeting extreme halophiles of the archaeal domain, and possess linear double-stranded DNA genomes [55]. The HHphage-associated helicase gene is located in the section of the genome containing early genes, which are presumably involved in initiating virus replication, and corresponds to the first of several ORFs within a polycistronic transcript [56]. Intriguingly, the 5′ end of this major transcript was mapped using primer extension [56] to within three nucleotides of the HHphage cleavage site, suggesting that this hammerhead ribozyme variant is active in vivo.
Dilution of the environment is highly toxic for obligate extreme halophiles. Accordingly, for certain bacteriophages that infect these organisms, virulence is tightly controlled in response to salt concentrations [57]. This allows bacteriophage to proliferate more aggressively when dilution threatens the viability of their hosts. Conversely, when salt levels are saturating, a carrier state is established in which phage DNA is propagated with a minimal burden on the host organism [58]. It is conceivable, then, that an appropriately tuned hammerhead ribozyme variant could be utilized by a halovirus to modulate the stability of a key transcript in a salt-dependent manner, thereby acting as a component of this regulatory response.
The discovery of thousands of new hammerheads in all three domains of life provides many opportunities to examine the functions and biological utilities of these ribozymes in their natural contexts. The activities of some representatives may be regulated by RNA folding changes induced by changes in protein, metabolite, or metal ion concentrations, similar to the structure modulation observed with riboswitches. Previous engineering efforts produced numerous examples of allosteric hammerhead ribozymes or other RNAs, establishing a precedent for ligand-mediated regulation of ribozyme function [59], [60]. In this context, the pseudoknot interactions identified in our study could be more easily manipulated to create regulated allosteric ribozymes via rational design.
Type II hammerheads were uncovered by a comparative genomics method described previously [21], [67]. Briefly, clusters of homologous non-coding sequences were analyzed with CMfinder to predict secondary structures and iterative homology searches conducted with RaveNnA [21]. A series of descriptors for RNAMotif were also used to find new hammerheads (descriptors in Text S1). All new hammerheads were combined with previously known examples and used as three updated alignments, type I-II-III, to perform homology searches on all RefSeq version 37 and available environmental sequences [21] using Infernal [68].
For final alignments, possible false positives were eliminated based on three criteria. First, any mutation in the core disqualified the hit. For this purpose, the consensus core was considered to be: C3, U4, G5, A6, N7, G8, A9, G12, A13, A14, A15, U16 and H17, where “N” means any nucleotide and “H” means A, C or U. Second, any mispairing directly adjacent to the core in stems I, II or III (i.e., N10.1–N11.1, N1.1–N2.1, and A15.1–U16.1) also led us to reject the hit. Finally, multiple mispairs or bulges in short stems resulted in candidate disqualification. The list of rejected hits consisted mainly of cryptic mutant hammerheads that are part of repeated elements, but those occurring in typical gene contexts (e.g., prophage) were often tested, as they were considered likely functional variants. Initially rejected hits were included in hammerhead alignments if activity could be measured.
To produce in vitro transcription templates, PCR was performed using genomic DNA isolated from Agrobacterium tumefaciens, Azorhizobium caulinodans (ATCC), Clostridium scindens (ATCC), PaP3 bacteriophage (kind gift of Professor Fuquan Hu) [69], pork chops (Shaw's Supermarket) and human whole blood (Promega). In cases where genomic DNA was unavailable, templates were constructed from chemically synthesized oligodeoxynucleotides (see Table S1). Transcriptions were generally conducted in 80 mM HEPES–KOH (pH 7.5 at 23°C), 24 mM MgCl2, 2 mM spermidine, 40 mM DTT, 2 mM of each ribonucleotide and 40 U µl−1 of purified T7 RNA polymerase. For ribozyme assays in trans, RNA was purified using denaturing PAGE, visualized by UV shadowing, and eluted in 200 mM NaCl, 10 mM Tris-HCl (pH 7.5 at 23°C), and 1 mM EDTA. RNA was then precipitated in ethanol, and the resulting pellet was rinsed in 70% ethanol and resuspended in water. Concentration was measured by UV spectrophotometry with a Nanodrop ND8000 (ThermoScientific).
For 5′ labeling, RNA was dephosphorylated with calf intestinal phosphatase (NEB) according to the manufacturer's instructions. Following phosphatase inactivation at 94°C for 3 minutes, 1 pmole of dephosphorylated RNA was typically used for 5′-end-labeling with T4 polynucleotide kinase (NEB) and [γ-32P]ATP according to the manufacturer's instructions. Labeled RNA was gel-purified as described, but visualized by autoradiogram.
To design bimolecular constructs, loop III was opened and base pairs were added to stabilize stem II by extending it to at least seven base pairs. Both RNA molecules were then transcribed from different synthetic DNA templates. RNA designated as the “ribozyme” (the strand not containing the cleavage site) was used in 200-fold excess for single-turnover kinetics. Typically, ∼5 nM radiolabeled substrate and 1 µM ribozyme were heated together at 65°C for two minutes in a 10 µl volume containing 100 mM Tris-HCl (pH 7.5 at 23°C) and 200 mM NaCl. After cooling to room temperature and removing time zero aliquots, 10 µl MgCl2 was added to a final concentration of 500 µM, unless otherwise stated. Reactions were stopped at various times with 5 volumes of stop buffer (80% formamide, 100 mM EDTA, 0.02% bromophenol blue and 0.02% xylene cyanol).
All time points for a given experiment were analyzed on the same denaturing gel, ranging from 6% to 20% polyacrylamide, depending on substrate and product sizes. After drying the gel, radiolabeled species were imaged using a Storm 820 PhosphorImager and analyzed with ImageQuant software (Molecular Dynamics). Values for kobs were derived from the slope of the line obtained by plotting the natural logarithm of the fraction of precursor RNA remaining versus time. Calculations were performed assuming first order reaction kinetics using data points corresponding to the first 5% to 30% of the reaction. Many ribozymes exhibit biphasic reaction kinetics. For these, we used SigmaPlot (SYSTAT) to fit the curves to the equation F = a(1−e−bt)+c(1−e−dt) by non-linear regression, where “F” is the fraction cleaved, “t” is time, “a” is the fraction cleaved where RNA molecules are cleaved at kobs “b” (the larger kobs) and “c” is the fraction cleaved at a kobs “d” (the smaller kobs) [70]. kobs values reported for rapidly cleaving ribozymes should be considered lower bounds due to the limitations of manual pipetting.
To estimate kobs values for reactions in cis, ribozyme cleavage time courses were performed during transcriptions in vitro. Transcriptions were assembled in either 80 mM HEPES–KOH (pH 7.5 at 23°C), 24 mM MgCl2, 2 mM spermidine, 40 mM DTT or in 10 µl volumes containing 50 mM Tris-HCl (pH 7.5 at 23°C), 100 mM NaCl, 10 mM MgCl2, 2 mM each rNTP, and 40 units µl−1 T7 RNA polymerase. Polymerization was allowed to proceed for 5 minutes at 37°C, at which point 5 µl of an equivalent mixture was added that also contained trace amounts of [α-32P]UTP and [α-32P]GTP. Incubations were continued at 37°C, and 1 µl aliquots were removed at various time points and added to 14 µl of stop buffer. Due to the initially low levels of incorporation of radiolabeled nucleotides, the earliest time point that can practically be assessed is 20 seconds. Note that, due to the requirements of T7 RNA polymerase, the Mg2+ concentrations used in these assays are considerably higher than those used for assays in trans. Note also that HHmeta, because of its requirement for particularly high Mg2+ concentrations, was able to be isolated in precursor form from standard in vitro transcriptions, and was subsequently assayed in cleavage assays in cis.
|
10.1371/journal.ppat.1004940 | Geometric Constraints Dominate the Antigenic Evolution of Influenza H3N2 Hemagglutinin | We have carried out a comprehensive analysis of the determinants of human influenza A H3 hemagglutinin evolution. We consider three distinct predictors of evolutionary variation at individual sites: solvent accessibility (as a proxy for protein fold stability and/or conservation), Immune Epitope Database (IEDB) epitope sites (as a proxy for host immune bias), and proximity to the receptor-binding region (as a proxy for one of the functions of hemagglutinin-to bind sialic acid). Individually, these quantities explain approximately 15% of the variation in site-wise dN/dS. In combination, solvent accessibility and proximity explain 32% of the variation in dN/dS; incorporating IEDB epitope sites into the model adds only an additional 2 percentage points. Thus, while solvent accessibility and proximity perform largely as independent predictors of evolutionary variation, they each overlap with the epitope-sites predictor. Furthermore, we find that the historical H3 epitope sites, which date back to the 1980s and 1990s, only partially overlap with the experimental sites from the IEDB, and display similar overlap in predictive power when combined with solvent accessibility and proximity. We also find that sites with dN/dS > 1, i.e., the sites most likely driving seasonal immune escape, are not correctly predicted by either historical or IEDB epitope sites, but only by proximity to the receptor-binding region. In summary, a simple geometric model of HA evolution outperforms a model based on epitope sites. These results suggest that either the available epitope sites do not accurately represent the true influenza antigenic sites or that host immune bias may be less important for influenza evolution than commonly thought.
| The influenza virus is one of the most rapidly evolving human viruses. Every year, it accumulates mutations that allow it to evade the host immune response of previously infected individuals. Which sites in the virus’ genome allow this immune escape and the manner of escape is not entirely understood, but conventional wisdom states that specific “immune epitope sites” in the protein hemagglutinin are preferentially attacked by host antibodies and that these sites mutate to directly avoid host recognition; as a result, these sites are commonly targeted by vaccine development efforts. Here, we combine influenza hemagglutinin sequence data, protein structural information, IEDB immune epitope data, and historical epitopes to demonstrate that neither the historical epitope groups nor epitopes based on IEDB data are crucial for predicting the rate of influenza evolution. Instead, we find that a simple geometrical model works best: sites that are closest to the location where the virus binds the human receptor and are exposed to solvent are the primary drivers of hemagglutinin evolution. There are two possible explanations for this result. First, the existing historical and IEDB epitope sites may not be the real antigenic sites in hemagglutinin. Second, alternatively, hemagglutinin antigenicity may not be the primary driver of influenza evolution.
| The influenza virus causes one of the most common infections in the human population. The success of influenza is largely driven by the virus’s ability to rapidly adapt to its host and escape host immunity. The antibody response to the influenza virus is determined by the surface proteins hemagglutinin (HA) and neuraminidase (NA). Among these two proteins, hemagglutinin, the viral protein responsible for receptor binding and uptake, is a major driver of host immune escape by the virus. Previous work on hemagglutinin evolution has shown that the protein evolves episodically [1–3]. During most seasons, hemagglutinin experiences mostly neutral drift around the center of an antigenic sequence cluster; in those seasons, it can be neutralized by similar though not identical antibodies, and all of the strains lie near each other in antigenic space [4–7]. After several seasons, the virus escapes its local sequence cluster to establish a new center in antigenic space [7–9].
There is a long tradition of research aimed at identifying important regions of the hemagglutinin protein, and by proxy, the sites that determine sequence-cluster transitions [4, 6, 10–21]. Initial attempts to identify and categorize important sites of H3 hemagglutinin were primarily sequence-based and focused on substitutions that took place between 1968, the emergence of the Hong Kong H3N2 strain, and 1977 [10, 11]. Those early studies used the contemporaneously solved protein crystal structure, a very small set of mouse monoclonal antibodies, and largely depended on chemical intuition to identify antigenically relevant amino-acid changes in the mature protein. Many of the sites identified in those studies reappeared nearly two decades later, in 1999, as putative epitope sites with no additional citations linking them to actual immune data [4]. Those sites and their groupings are still considered the canonical immune epitope set today [3, 16, 22]. While the limitations of experimental techniques and of available sequence data in the early 1980’s made it necessary to form hypotheses based on chemical intuition, these limitations are starting to be overcome through recent advances in experimental immunological techniques and wide-spread sequencing of viral genomes. Therefore, it is time to revisit the question of whether or not our current understanding of the host immune response is reflected in the observed patterns of influenza hemagglutinin evolution. For example, at least one recent model has suggested that the hemagglutinin protein may evolve to modulate receptor-binding avidity rather than to modulate antibody-binding [23]. Moreoever, since the original epitope set was identified via sequence analysis, we do not even know whether bona-fide immune-epitope sites actually exist, i.e., sites which represent a measurable bias in the host immune response. Most importantly, even if immune-epitope sites do exist and can be experimentally identified, it is possible that they do not experience more positive selection than other important sites in the protein.
Some recent studies have begun to address these questions indirectly, via evolutionary analysis. For example, over the last two decades, virtually every major study on positive selection in hemagglutinin has found some but never all of the historical epitope sites to be under positive selection [3, 16, 18, 19, 23]. Furthermore, each of these studies has found a set of sites that are under positive selection but do not belong to any historical epitope. Finally, because every study identifies slightly different sites, there seems to be no broad agreement on which sites are under positive selection [12, 16, 18, 19]. The sites found by disparate techniques are similar but they are never identical.
To dissect the determinants of hemagglutinin evolution, we here linked several predictors, including relative solvent accessibility, the inverse distance from the receptor-binding region, and IEDB immune epitope data, to site-wise evolutionary rates calculated from all of the human H3N2 sequence data for the last 22 seasons (1991–2014). We found that, individually, all these predictors explained approximately 15% of evolutionary rate variation. After controlling for biophysical constraints with relative solvent accessibility and function with distance to the receptor-binding region, the remaining predictive power of either IEDB or historical categories was relatively low. In addition, we found that current IEDB data does not reflect the historical epitope sites or their groups. Finally, by explicitly accounting for RSA, proximity, and host immune data, we found that we could predict nearly 35% of the evolutionary rate variation in hemagglutinin, nearly twice as much variation as could be explained by earlier models.
Our overarching goal in this study was to identify specific biophysical or biochemical properties of the mature protein that determine whether a given site will evolve rapidly or not. As a measure of evolutionary variation and selective pressure, we used the metric dN/dS. dN/dS can measure both the amount of purifying selection acting on a site (when dN/dS ≪ 1 at that site) and the amount of positive diversifying selection acting on a site (when dN/dS ≳ 1). For simplicity, we will refer to dN/dS as an evolutionary rate, even though technically it is a relative evolutionary rate or evolutionary-rate ratio. We built an alignment of 3854 full-length H3 sequences spanning 22 seasons, from 1991/92 to 2013/14. We subsequently calculated dN/dS at each site, using a one-rate fixed-effects likelihood (FEL) model as implemented in the software HyPhy [24].
Several recent works have shown that site-specific evolutionary variation is partially predicted by a site’s solvent exposure and/or number of residue-residue contacts in the 3D structure [19, 20, 25–30] (see Ref. [31] for a recent review). This relationship between protein structure and evolutionary conservation likely reflects the requirement for proper and stable protein folding: Mutations at buried sites or sites with many contacts are more likely to disrupt the protein’s conformation [30] or thermodynamic stability [32]. In addition, there may be functional constraints on site evolution. For example, regions in proteins involved in protein–protein interactions or enzymatic reactions are frequently more conserved than other regions [27, 33, 34]. However, these structural and functional constraints generally predict the amount of purifying selection expected at sites, and therefore they cannot identify sites under positive diversifying selection. Moreover, the short divergence time of viruses causes the systematic biophysical pressures that predict much of eukaryotic protein evolution to be much less dominant in viral evolution [28]. Thus, we set out to find a constraint on hemagglutinin evolution that was related to the protein’s role in viral binding and fusion.
A few earlier studies had shown that sites near the sialic acid-binding region of hemagglutinin tend to evolve more rapidly than the average for the protein [4, 20, 21]. Furthermore, when mapping evolutionary rates onto the hemagglutinin structure, we noticed that the density of rapidly evolving sites seemed to increase somewhat towards the receptor-binding region (Fig 1A). Therefore, as the primary function of hemagglutinin is to bind to sialic acid and induce influenza uptake, we reasoned that distance from the receptor-binding region of HA might serve as a predictor of functionally driven HA evolution. We calculated distances from the sialic acid-binding region (defined as the distance from site 224 in HA), and correlated these distances with the evolutionary rates at all sites. We found that distance from the receptor-binding region was a strong predictor of evolutionary rate variation in hemagglutinin (Pearson correlation r = 0.41, P < 10−15).
Next, we wanted to verify that this correlation was representative of hemagglutinin evolution and not just an artifact of the specific site chosen as the reference point in the distance calculations. It would be possible, for example, that distances to several spatially separated reference sites all resulted in similarly strong correlations. We addressed this question systematically by making, in turn, each individual site in HA the reference site, calculating distances from that site to all other sites, and correlating these distances with evolutionary rate. We then mapped these correlations onto the structure of hemagglutinin, coloring each site according to the strength of the correlation we obtained when we used that site as reference in the distance calculation (Fig 1B). We obtained a clean, gradient-like pattern: The correlations were highest when we calculated distances relative to sites near the receptor-binding site (with the maximum correlation obtained for distances relative to site 224), and they continuously declined and then turned negative the further we moved the reference site away from the apical region of hemagglutinin (Fig 1B). This result was in stark contrast to the pattern we had previously observed when mapping evolutionary rate directly (Fig 1A). In that earlier case, while there was a perceptible preference of faster evolving sites to fall near the receptor-binding site, the overall distribution of evolutionary rates along the structure looked mostly random to the naked eye. We thus found a geometrical, distance-based constraint on hemagglutinin evolution: Sites evolve faster the closer they lie toward the receptor-binding region.
We also evaluated how proximity to the receptor-binding region performed as a predictor of dN/dS in comparison to the previously proposed structural predictors relative solvent accessibility (RSA) and weighted contact number (WCN). We found that among these three quantities, proximity to the sialic acid-binding region was the strongest predictor, explaining 16% of the variation in dN/dS (Pearson r = 0.41, P < 10−15, see also Fig 2 and S1 Fig). RSA and WCN explained 14% and 6% of the variation in dN/dS, respectively (r = 0.37, P < 10−15 and r = 0.25, P = 7 × 10−9). Proximity to the sialic acid-binding region and RSA were virtually uncorrelated (r = 0.08, P = 0.09) while RSA and WCN correlated strongly (r = −0.64, P < 10−15). These results suggested that proximity to the sialic acid-binding region and RSA should be used jointly in a predictive model.
Because hemagglutinin has, in addition to its function as a receptor-binding protein, a host of other intermediate functional states during the viral fusion process, we also tested the ability of structural metrics from the post-fusion state to predict hemagglutinin evolutionary rate [35]. We found no significant metric, either RSA or proximity, derived from the post-fusion state. (Complete data and analysis scripts are available in the accompanying Github repository, see Methods for details.)
Another potential functional constraint on hemagglutinin evolution is a bias in the human immune system. This bias, generally referred to as antigenicity, describes the extent to which the human immune system does a better job attacking one region of a protein compared to another. Conventional wisdom states that functionally important sites in the protein that are targeted by antibodies will evolve more rapidly to facilitate immune escape. And indeed, our results from the previous subsection have shown that proximity to the receptor-binding region is a good predictor of evolutionary variation. However, if substitutions to avoid direct antibody binding are the primary cause of positive selection, then we would expect antigenic sites on hemaggalutinin to serve as a substantially better predictors of adaptation than proximity to the receptor-binding site alone.
For influenza hemagglutinin H3, there exists a list of canonical, historical epitope sites that are commonly considered to represent this bias [4]. However, these sites were not primarily defined based on actual immunological data, and they have not been re-validated since the late 1990s even though more experimental data is now available. (See Discussion for details on the history of the historical epitope sites.) Before we could generate a combined evolutionary model, we therefore considered it essential to validate the antigenic groups with available immunological data. As it turns out, the majority of antigenic data available did not agree with the historical epitope sites (S1 Text). Therefore, we used both the historical epitope sites and a set of IEDB re-defined epitopes for further modeling.
A detailed explanation of our re-grouping based on IEDB data is available in S1 Text. It is important to note that these groups are not intended to represent a new canonical set of hemagglutinin epitopes. Indeed, the data from which they were derived is limited and relatively poorly annotated. However, considering the magnitude of the difference between the historical epitopes and the available IEDB data we considered it imperative to include IEDB derived epitopes in our analysis.
Thus, we considered both the historical epitope groups (Bush 1999) and the IEDB derived epitopes 1–4, defined in S1 Text. Because a site’s epitope status is a categorical variable, we calculated variance explained as the coefficient of determination (R2) in a linear model with dN/dS as the response variable and epitope status as the predictor variable. We found that IEDB epitopes explained 15% of the variation in dN/dS, comparable to RSA and proximity. In comparison, the historical epitopes alone explained nearly 18% of the variation in dN/dS, outperforming all other individual predictor variables considered here (Fig 2 and Table 1). However, as discussed in S1 Text, the available IEDB data suggest that not all of the historical sites may be actual immune epitope sites. Therefore, we suspected that some of the predictive power of historical sites was due to these sites simply being solvent-exposed sites near the receptor-binding region. We similarly wondered to what extent the predictive power of the IEDB epitope sites was attributable to the same cause, since, in fact, both historical and IEDB epitope sites showed comparable enrichment in sites near the sialic acid-binding region and in solvent-exposed sites (S2 Fig). Therefore, we analyzed how the variance explained increased as we combined epitope sites (IEDB or historical) with either RSA or proximity or both.
We found that epitope status, under either definition (IEDB/historical), led to increased predictive power of the model when combined with either RSA or proximity (Fig 2). However, a model consisting of just the two predictors RSA and proximity, not including any information about epitope status of any sites, performed even better than any of the other one- or two-predictor models, explaining 32% of the variation in dN/dS (Fig 2). Adding epitope status to this best-performing two-predictor model resulted in only minor improvement, from 32% to 34% variance explained in the case of IEDB epitopes and from 32% to 37% variance explained in the case of historical epitope sites (Fig 2 and Table 1).
The geometrical constraints RSA and proximity explained more variance in dN/dS than did epitope sites, but were they also better at predicting sites of interest? Because dN/dS can measure purifying as well as positive diversifying selection, the percent variance in dN/dS that a model explains may not necessarily accurately reflect how useful that model is in predicting specific sites, e.g. sites under positive selection. For example, one could imagine a scenario in which a model does exceptionally well on sites under purifying selection (dN/dS ≪ 1) but fails entirely on sites under positive selection (dN/dS > 1). Such a model might explain a large proportion of variance but be considered less useful than a model that overall predicts less variation in dN/dS but accurately pinpoints site under positive selection. Therefore, we wondered whether epitope sites might do a poor job predicting background purifying selection but might still be useful in predicting sites with dN/dS > 1. We found, to the contrary, that neither the historical nor the IEDB epitope sites could reliably predict sites with dN/dS > 1, alone or in combination with RSA (Fig 3A–3D). Proximity to the receptor-binding site, on the other hand, correctly predicted four sites with dN/dS > 1, even in the absence of any other predictors. Notably, all models we considered here were robust to cross-validation. The cross-validated residual standard error was virtually unchanged from its non-cross-validated value in all cases (Table 1). Because proximity clearly identified four points with high dN/dS, we also verified that the proximity–dN/dS correlation was not caused just by these four points. We removed from our data set the four points that had both predicted and observed dN/dS > 1, and found that a significant proximity–dN/dS correlation remained nonetheless (r = 0.17, p = 0.00001).
Finally, we compared the predictions from the geometrical model of hemagglutinin evolution to results from a recent study of antigenic cluster transitions; that study found seven sites near the receptor-binding region which were critical for cluster transitions according to hemagglutinin inhibition (HI) assays with ferret antisera [21]. The sites identified in Ref. [21] were 145, 155, 156, 158, 159, 189, and 193. For comparison, our geometric model (with predictors RSA and 1/Distance) predicted none of these sites to be under positive selection. Sites predicted to have dN/dS > 1 were instead 96, 137, 138, 143, 222, 223, 225, and 226. Moreover, out of the seven sites from Ref. [21], only one (site 145) had an observed dN/dS significantly above 1. By contrast, four of the eight sites predicted under the geometric model to have dN/dS > 1 did indeed have dN/dS significantly above 1. Thus, the sites that determine the major antigenic changes in the virus did not at all overlap with the sites expected and observed to be under the greatest evolutionary pressure. When investigating the location of these sites in detail, we found that all of the sites we predicted to have dN/dS > 1 were located just basal to the receptor-binding site, whereas nearly all of the sites from [21] (with the exception of 145, the site with dN/dS > 1) were located on the apical side of the receptor-binding site (Fig 4).
In summary, we have found that two simple geometric measures of a site’s location in the 3D protein structure, solvent exposure and proximity to the receptor-binding region, jointly outperformed, by a wide margin, any previously considered predictor of evolutionary variation in hemagglutinin, including immune epitope groups. In fact, the vast majority of the variation in evolutionary rate that was explained by the historical epitope sites was likely due to these sites simply being located near the receptor-binding region on the surface of the protein. However, historical epitope sites, in combination with solvent exposure and proximity, had some residual explanatory power beyond even a three-predictor model that combined the two geometric measures with IEDB immune-epitope data. We suspect that this residual explanatory power reflects the sequence-based origin of the historical epitope sites. To our knowledge, the historical epitope sites were at least partially identified by observed sequence variation, so that, to some extent, these sites are simply the sites that have been observed to evolve rapidly in hemagglutinin.
We have conducted a thorough analysis of the determinants of site-specific hemagglutinin evolution. Most importantly, we have found that immune epitopes, defined either by IEDB data or historically by sequence analysis, account for a relatively small portion of influenza evolution. In addition, we have found that neither epitope definition could be used to predict hemagglutinin sites under positive selection. By contrast, a simple geometric measure, receptor-binding proximity, is both a combined strong predictor of evolutionary rate and is the only quantity that can predict sites with dN/dS > 1. In addition, we have shown that a simple linear model containing three predictors, solvent accessibility, proximity to the receptor-binding region, and IEDB epitopes, explains nearly 35% of the evolutionary rate variation in hemagglutinin H3. Taken together our analysis suggests that one of two possible explanations must be true. First, it is possible that hemagglutinin antigenicity is not a strong direct driver of influenza adaptive evolution. Second, alternatively, the current IEDB data and historical epitopes may simply be insufficient and/or incorrect. Such a situation would explain why neither epitope definition can explain much evolutionary rate variation beyond the geometric constraints, and why neither epitope definition can predict sites under positive selection.
Efforts to define immune epitope sites in H3 hemagglutinin go back to the early 1980’s [10]. Initially, epitope sites were identified primarily by speculating about the chemical neutrality of amino acid substitutions between 1968 (the year H3N2 emerged) and 1977, though some limited experimental data on neutralizing antibodies was also considered [10, 11]. In 1981, the initial four epitope groups were defined by non-neutrality (amino-acid substitutions that the authors believed changed the chemical nature of the side chain) and relative location, and given the names A through D [10]. Since that original study in 1981, the names and general locations of H3 epitopes have remained largely unchanged [4, 16]. The sites were slightly revised in 1987 by the same authors and an additional epitope named E was defined [11]. From that point forward until 1999 there were essentially no revisions to the codified epitope sites. In addition, while epitopes have since been redefined by adding or removing sites, no other epitope groups have been added [3, 16, 18]; epitopes are still named A–E. In 1999, the epitopes were redefined by more than doubling the total number of sites and expanding all of the epitope groups [4]. At that time, the redefinition consisted almost entirely of adding sites; very few sites were eliminated from the epitope groups. Although this set of sites and their groupings remain by far the most cited epitope sites, it is not particularly clear what data justified this definition. Moreover, when the immune epitope database (IEDB) summarized the publicly available data for influenza in 2007, it only included one IEDB B cell epitope in humans (Table 2 in [36]). Although there were a substantial number of putative T cell epitopes in the database, a priori there is no reason to expect a T cell epitope to show preference to hemagglutinin as opposed to any other influenza protein; yet it is known that several other influenza proteins show almost no sites under positive selection. Moreover, it is known that the B cell response plays the biggest role is maintaining immunological memory to influenza, and thus it is the most important arm of the adaptive immune system for influenza to avoid.
The historical H3 epitope sites have played a crucial role in molecular evolution research. Since 1987, an enormous number of methods have been developed to analyze the molecular evolution of proteins, and specifically, to identify positive selection. The vast majority of these methods have either used hemagglutinin for testing, have used the epitopes for validation, or have at some point been applied to hemagglutinin. Most importantly, in all this work, the epitope definitions have been considered fixed. Most investigators simply conclude that their methods work as expected because they recover some portion of the epitope sites. Yet virtually all of these studies identify many sites that appear to be positively selected but are not part of the epitopes. Likewise, there is no single study that has ever found all of the epitope sites to be important. Even if the identified sites from all available studies were aggregated, we would likely not find every site among the historical epitopes in that aggregated set of sites.
Given all of this research activity, it seems that the meaning of an immune epitope has been muddled. Strictly speaking, an immune epitope is a site to which the immune system reacts. There is no a priori reason why an immune epitope needs to be under positive selection, needs to be a site that has some number or chemical type of amino acid substitutions, or needs to be predictive of influenza whole-genome or hemagglutinin-specific sequence cluster transitions. Yet, from the beginning of the effort to define hemagglutinin immune epitopes, such features have been used to identify epitope sites, resulting in a set of sites that may not accurately reflect the sites against which the human immune system produces antibodies.
Ironically, this methodological confusion has actually been largely beneficial to the field of hemagglutinin evolution. As our data indicate, if the field had been strict in its pursuit of immune epitopes sites, it would have been much harder to produce predictive models with those sites, in particular given that IEDB data on non-linear epitopes have been sparse until very recently. By contrast, the historical epitope sites have been used quite successfully in several predictive models of the episodic nature of influenza sequence evolution. In fact, in our analysis, historical epitopes displayed the highest amount of variance explained among all individual predictors (Fig 2). We argue here that the success of historical epitope sites likely stems from the fact that they were produced by disparate analyses each of which accounted for a different portion of the evolutionary pressures on hemagglutinin. Of course, it is important to realize that some of this success is likely the result of circular reasoning, since the sites themselves were identified at least partially from sequence analysis that included the clustered, episodic nature of influenza hemagglutinin sequence evolution.
Despite the success of historical epitope groups, they only predict about 18% of the evolutionary rate-variation of hemagglutinin for the entire phylogenetic tree. Since many of these sites likely are not true immune epitopes (and therefore not host dependent), one might ask which features of the historical epitope sites make them good predictors. We suspect that they perform well primarily because they are a collection of solvent-exposed sites near the sialic acid-binding region (see S2 Fig). We had shown previously that sites within 8 Šof the sialic acid-binding site are enriched in sites under positive selection, compared to the rest of the protein [20]. A similar result was found in the original paper by Bush et al. [4]. However, the related metric of distance from the sialic acid-binding site has not previously been considered as a predictor of evolution in hemagglutinin. Furthermore, before 1999, most researchers thought the opposite should be true; that receptor-binding sites should have depressed evolutionary rates [4]. Even today the field seems split on the matter [21]. As we have shown here, the inverse of the distance from sialic acid is a relatively strong quantitative predictor of hemagglutinin evolution; by itself this distance metric can account for 16% of evolutionary rate-variation. Moreover, by combining this one metric with another to control for solvent exposure, we can account for more than a third of the evolutionary rate variation in hemagglutinin. For reference, this number is larger than the variation one could predict by collecting and analyzing all of the hemagglutinin sequences that infect birds (another group of animals with large numbers of natural influenza infections), and using those rates to predict human influenza hemagglutinin evolutionary rates [20].
In terms of re-grouping IEDB immune data, it is important to note that the IEDB has major limitations; not all existing (not to mention all possible) immunological data have been added. Further, the extent to which certain epitopes (e.g., stalk epitopes) have been mapped may be more reflective of a bias in research interests among influenza researchers than a bias in the human immune system. Also, until recently, the ability to generate unbiased high-affinity antibodies to influenza has been limited [37, 38]. Therefore, in our re-derivation of epitope groupings, we are certainly missing sites or may be incorrectly grouping the ones that we have. Our analysis of epitope sites will likely have to be redone as more data become available. However, we expect that as more non-linear data become available, they will broadly follow the trend observed in the linear epitope data, that is, the more antibodies are mapped, the more sites in the hemagglutinin protein appear in at least one mapping, until virtually every site in the entire hemagglutinin protein is represented. Under this scenario, the ability to predict evolution from immunological data would become worse, not better, as more data are accumulated.
One additional caveat comes from any potential effect of glycosylation on influenza immune escape. Glycosylations on hemagglutinin can have a major effect on receptor and antibody binding [13]. In addition, the number of glycosylations in H3 hemagglutinin has increased since initial introduction of pandemic H3N2 in 1968 [13]. However, a priori there is no reason to believe that glycosylation will either increase or decrease dN/dS at individual sites or groups of sites; it could affect dN/dS in either direction, in particular if direct antibody escape is not the primary driver of hemagglutinin evolution. Moreover, there is no clear way to incorporate glycosylation into our regression model. In the future, investigating changing glycosylation patterns throughout the evolution of H3 hemagglutinin may yield important insights into influenza adaptation and immune escape.
Why do geometric constraints (solvent exposure and proximity to receptor-binding site) do a good job predicting hemagglutinin evolutionary rates? Hemagglutinin falls into a class of proteins known collectively as viral spike glycoproteins (GP). In general, the function of these proteins is to bind a host receptor to initiate and carry out uptake or fusion with the host cell. Therefore, a priori one might expect that the receptor-binding region would be the most conserved part of the protein, since binding is required for viral entry. Yet, in hemagglutinin sites near the binding region are the most variable in the entire protein. There are at least two possible models that might explain this observation. First, conventional wisdom says that in terms of host immune evasion, antibodies that bind near the receptor-binding region may be the most inhibitory, and hence mutations in this region the most effective in allowing immune escape. Viral spike GPs have a surface that is both critical for viral survival and is sufficiently long lived that a host immune response is easily generated against it. There are likely many other viral protein surfaces that are comparatively less important or sufficiently short lived during a conformational change that antibody neutralization is impractical. Thus, the virions that survive to the next generation are those with substantial variation at the surface or surfaces with high fitness consequences and a long half-life in vivo. Evolutionary variation at surfaces with low or no fitness consequences, or at short-lived surfaces, should behave mostly like neutral variation and hence appear as random noise, not producing a consistent signal of positive selection. Second, according to the avidity modulation model of Hensley et al. [23], it is possible that antibody inhibition is not overcome by escaping the antibody directly. Considering the fact that neither historical nor IEDB immune epitopes vastly out-performed our simple distance metric, we think that our results support a model which does not expect an evolutionary bias based on antibody binding sites. However, it remains a possibility that the historical epitopes and current IEDB data are simply wrong about which sites and groups of sites the human immune system attacks. Either way, our work highlights the need for a paradigm shift in the field.
We also need to consider that actual epitope sites, i.e., sites toward which the immune system has a bias, may not be that important for the evolution of viruses. An epitope is simply a part of a viral protein to which the immune system reacts. Therefore, it represents a host-centered biological bias. The virus may experience stronger selection at regions with high fitness consequences but that generate a relatively moderate host response compared to other sites with low fitness consequences that generate a relatively strong host response. Moreover, there is little reason to believe that influenza must escape an antibody by directly reducing the binding of that antibody. There are other possible scenarios for immune evasion, e.g. avidity modulation as stated above. Thus, we expect that the geometric constraints we have identified here will be more useful in future modeling work than the IEDB epitope groups we have defined. Moreover, we expect that similar geometrical constraints will exist in other viral spike glycoproteins, and in particular in other hemagglutinin variants.
By contrast to the clear geometric constraints we observed for the pre-fusion structure, we found no comparable result for the post-fusion structure. There are perhaps several good reasons to expect this result. First, the transition state is likely very short-lived, such that the human immune system is not able to generate antibodies against it. Second, due to the short-lived functional nature of the transition state, there is likely relatively little selection for folding stability. Therefore, for the post-fusion structure we do not expect to observe the RSA–rate correlation that exists in the pre-fusion structure and in most other proteins. Third, models describing the transition from the pre-fusion to the post-fusion state show that the HA1 chain dissociates from the HA2 chain [39]. Subsequently, the HA2 chain carries out virtually all of the fusogenic functions. Thus, the HA1 chain is likely the functional unit in the first step of entry and the HA2 chain is likely the functional unit in the second. However, there is almost no rapid evolution happening in the HA2 chain, i.e., the HA2 chain does not seem to experience any positive diversifying selection.
Remarkably, the sites we found that experienced the most positive selection showed minimal overlap with the sites found to be minimally sufficient for explaining the major antigenic transitions in H3N2, as determined by HI assays with ferret antisera [21]. While both groups of sites lie near the sialic-acid binding region, the vast majority of positively selected sites are located basally to sialic acid whereas sites identified by HI assays lie predominantly on the apical side (Fig 4). This finding suggests that HI assays and positive selection analyses reflect distinct biological mechanisms. For example, HI assays might not accurately reflect selection pressures in vivo. Such a result would suggest that influenza is not under pressure to directly escape antibody binding. Alternatively, it is possible that the standard manner for obtaining ferret antisera simply may not represent a good proxy for the cyclical nature of human influenza infections [40]. Indeed, recent evidence suggests that, at least for the pandemic H1N1 strain, cyclical infections can shift the antibody response toward the receptor-binding region [41]. In future work, disentangling the different mechanisms reflected by HI assays and by positive-selection analyses will likely be crucial for improved prediction of HA evolution and of optimal vaccine strains.
All of the data we analyzed were taken from the Influenza Research Database (IRD) [42]. The IRD provides IEDB immune data curated from the data available in the Immune Epitope Database [43].
We used sequences that had been collected since the 1991–1992 influenza season. Any season before the 1991–1992 season had an insufficient number of sequences to contribute much to the selection analysis. The sequences were filtered to remove redundant sequences and laboratory strains. The sequences were then aligned with MAFFT [44]. Since it is known that there have been no insertions or deletions since the introduction of the H3N2 strain, we imposed a strict opening penalty and removed any sequences that had intragenic gaps. In addition, we manually curated the entire set to remove any sequence that obviously did not align to the vast majority of the set; in total the final step only removed about 10 sequences from the final set of 3854 sequences. For the subsequent evolutionary rate calculations, we built a tree with FastTree 2.0 [45].
To compute evolutionary rates, we used a fixed effects likelihood (FEL) approach with the MG94 substitution model [24, 46, 47]. We used the FEL provided with the HyPhy package [24]. For the full setup see the linked GitHub repository (https://github.com/wilkelab/influenza_HA_evolution). As is the case for all FEL models, an independent evolutionary rate is fit to each site using only the data from that column of the alignment. Because our data set consisted of nearly 4000 sequences, almost every site in our alignment had a statistically significant posterior probability of being either positively or negatively selected after adjusting via the false discovery rate (FDR) method. As shown in Fig 3, all evolutionary rates fall into a range between dN/dS = 0 and dN/dS = 4.
We computed RSA values as described previously [28]. Briefly, we used DSSP [48] to compute the solvent accessibility of each amino acid in the hemagglutinin protein. Then, we used the maximum solvent accessibilities [49] for each amino acid to normalized the solvent accessibilities to relative values between 0 and 1. We found that RSA calculated in the trimeric state produced better predictions than RSA calculated in the monomeric state. Thus, we used multimeric RSA in all models in this study. Both multimeric and monomeric RSA are included in the supplementary data.
To create the structural heat map of correlations shown in Fig 1B, we first needed to calculate the correlations between evolutionary rates and pairwise distances, calculated in turn for each location in the protein structure as the reference point for the distance calculations. Conceptually, we can think of this analysis as overlaying a grid on the entire protein structure, where we first calculate the distance to various grid points from every Cα in the entire protein, and then compute the correlation between the set of distances to the sites on the grid and the evolutionary rate at those sites. In practice, we calculated the distance from each Cα to every other Cα. We then colored each residue by the correlation obtained between evolutionary rates and all distances to its Cα.
All statistical analyses were performed using R [50]. We built the linear models with both the lm() and glm() functions. For cross validation, we used the cv.glm() function within the boot package. Residual standard error values were computed by taking the square root of the delta value from cv.glm(). With the exception of graph visualizations, all figures in this manuscript were created using ggplot2 [51].
A complete data set including evolutionary rates, epitope assignments, RSA, and proximity to the receptor-binding site is available as S1 Dataset. Raw data and analysis scripts are available at https://github.com/wilkelab/influenza_HA_evolution. In the repository, we have included all human H3 sequences from the 1991–1992 season to present combined into a single alignment. We have cleaned the combined data to only include sequences with canonical bases, non-repetitive sequences, and we have hand filtered the data to ensure all included sequences align appropriately to the 566 known amino acid sites. In addition, we have built a tree and visually verified that there were no outlying sequences on the tree for the combined set.
The site-wise numbering for the H3 hemagglutinin protein reflects the numbering of the mature protein; this numbering scheme requires the removal of the first 16 amino acids in the full-length gene. Thus, for protein numbering purposes, site number 1 is actually the 17th codon in full-length gene numbering. The complete length of the H3 hemagglutinin gene is 566 sites while the total length of the protein is 550 sites. It is important to point out that the mature H3 protein has two chains (HA1 and HA2) that are produced by cutting the presursor (HA0) protein between sites 329 and 330 in protein numbering. In addition, as a result of cloning and experimental diffraction limitations, most (or likely all) hemagglutinin structures do not include some portion of the first or last few amino acids of either chain of the mature protein, and crystallographers always remove the C-terminal transmembrane span from HA2. For example, the structure we used (PDBID: 4FNK) in this study does not include the first 8 amino acids of HA1, the last 3 amino acids of HA1, or the last 48 amino acids of HA2. As a result, HA1 includes sites 9–326 and HA2 includes sites 330–502. The complete data table in the project repository lists the gene sequence from one of the three original H3N2 (Hong Kong flu) hemagglutinin (A/Aichi/2/1968), the gene numbering, the protein numbering, the numbering of one H3N2 crystal structure, historical immune epitope sites from 1981, 1987 and 1999, and every calculated parameter used (and many others than were not used) in this study. In general, the most common epitope definitions in use today are those employed by Bush et. al 1999 [4]. Throughout this work, we refer to the Bush et. al 1999 epitopes as the“historical epitope sites”.
|
10.1371/journal.pbio.1002073 | Cortical Hierarchies Perform Bayesian Causal Inference in Multisensory Perception | To form a veridical percept of the environment, the brain needs to integrate sensory signals from a common source but segregate those from independent sources. Thus, perception inherently relies on solving the “causal inference problem.” Behaviorally, humans solve this problem optimally as predicted by Bayesian Causal Inference; yet, the underlying neural mechanisms are unexplored. Combining psychophysics, Bayesian modeling, functional magnetic resonance imaging (fMRI), and multivariate decoding in an audiovisual spatial localization task, we demonstrate that Bayesian Causal Inference is performed by a hierarchy of multisensory processes in the human brain. At the bottom of the hierarchy, in auditory and visual areas, location is represented on the basis that the two signals are generated by independent sources (= segregation). At the next stage, in posterior intraparietal sulcus, location is estimated under the assumption that the two signals are from a common source (= forced fusion). Only at the top of the hierarchy, in anterior intraparietal sulcus, the uncertainty about the causal structure of the world is taken into account and sensory signals are combined as predicted by Bayesian Causal Inference. Characterizing the computational operations of signal interactions reveals the hierarchical nature of multisensory perception in human neocortex. It unravels how the brain accomplishes Bayesian Causal Inference, a statistical computation fundamental for perception and cognition. Our results demonstrate how the brain combines information in the face of uncertainty about the underlying causal structure of the world.
| How can the brain integrate signals into a veridical percept of the environment without knowing whether they pertain to same or different events? For example, I can hear a bird and I can see a bird, but is it one bird singing on the branch, or is it two birds (one sitting on the branch and the other singing in the bush)? Recent studies demonstrate that human observers solve this problem optimally as predicted by Bayesian Causal Inference; yet, the neural mechanisms remain unclear. By combining psychophysics, Bayesian modelling, functional magnetic resonance imaging (fMRI), and multivariate decoding in an audiovisual localization task, we show that Bayesian Causal Inference is performed by a neural hierarchy of multisensory processes. At the bottom of the hierarchy, in auditory and visual areas, location is represented on the basis that the two signals are generated by independent sources (= segregation). At the next stage, in posterior intraparietal sulcus, location is estimated under the assumption that the two signals are from a common source (= forced fusion). Only at the top of the hierarchy, in anterior intraparietal sulcus, the uncertainty about the world’s causal structure is taken into account and sensory signals are combined as predicted by Bayesian Causal Inference.
| Our senses are constantly bombarded with many different signals. Imagine you are crossing a street and suddenly hear a loud motor noise. Is that motor noise coming from the car on the opposite side of the street or from a rapidly approaching car that you have not yet spotted? To locate the source of the motor noise more precisely, you should integrate the auditory signal with the sight of the car only if the two inputs pertain to the same object. Thus, estimating an environmental property (e.g., spatial location) in multisensory perception inherently relies on inferring whether sensory signals are caused by common or independent sources [1,2].
Past research in perception and cue combination has mostly ignored the causal inference problem and focused on the special case in which sensory signals arise from a common source. A large body of research has demonstrated that observers integrate signals near-optimally weighted by their reliability in these “forced fusion” settings [3–9]. Yet, in our complex natural environment forced fusion would be detrimental and the brain needs to balance integration and segregation according to the underlying causal structure (i.e., common versus independent sources) [10].
Hierarchical Bayesian Causal Inference provides a rational strategy to arbitrate between information integration and segregation in perception and cognition. In case of a common source, signals should be integrated weighted by their relative sensory reliabilities [3,4]. In case of independent sources, they should be processed independently. Critically, the observer does not know the underlying causal structure and needs to infer it from spatiotemporal or higher order (e.g., semantic) congruency cues [2]. To account for the uncertainty about the causal structure, an observer should compute a final estimate by averaging the estimates (e.g., spatial location) under the two potential causal structures weighted by the posterior probabilities of these structures (i.e., model averaging).
Indeed, recent psychophysics and modeling efforts have demonstrated that human observers locate audiovisual signal sources in line with Bayesian Causal Inference by combining the spatial estimates under the assumptions of common and independent sources weighted by their posterior probabilities [2]. For small spatial disparities, audiovisual spatial signals are integrated weighted by their relative sensory reliabilities leading to strong crossmodal spatial biases [3]; for large spatial disparities, these crossmodal biases are greatly attenuated [11,12], because the final spatial estimate relies predominantly on the segregated option.
However, the neural mechanisms that enable Bayesian Causal Inference are unknown. In particular, it is unclear whether the brain encodes the spatial estimates under the assumptions of common and independent sources in order to perform Bayesian Causal Inference. Does the brain explicitly represent several spatial estimates that enter into Bayesian Causal Inference?
We combined psychophysics, Bayesian statistical modeling, and a multivariate functional magnetic resonance imaging (fMRI) decoding approach to characterize how the human brain performs Bayesian Causal Inference along the auditory [13] and visual [14] spatial cortical hierarchies. During fMRI scanning, we presented five participants with synchronous auditory (white noise) and visual (Gaussian cloud of dots) spatial signals that were independently sampled from four possible locations along the azimuth (i.e., −10°, −3.3°, 3.3°, or 10°) (Fig 1A). Further, we manipulated the reliability of the visual signal by varying the standard deviation of the visual cloud (2° or 14° standard deviation). Participants were asked selectively to report either the visual or the auditory signal location (without feed-back). Thus, the 4 (auditory locations) × 4 (visual locations) × 2 (visual reliability) × 2 (visual versus auditory report) factorial design included 64 conditions (Fig 1B). Importantly, as auditory and visual spatial locations were sampled independently on each trial, our design implicitly manipulated audiovisual spatial disparity, a critical cue informing the brain whether signals emanate from common or independent sources (cf. supporting S5 Table).
At the behavioral level, we first investigated how participants integrate and segregate sensory signals for auditory and visual spatial localization. Fig 2 shows the histograms of response deviations as a function of task-relevance (i.e., auditory versus visual report), audiovisual spatial disparity, and visual reliability. If participants were able to determine the location of the task-relevant auditory or visual signal precisely, the histogram over response deviations would reduce to a delta function centered on zero. Thus, the difference in widths of the histograms for auditory and visual report indicates that participants were less precise when locating auditory (green) as compared to the visual signals (red). Likewise, as expected visual localization was less precise for low (red dashed) relative to high visual (red solid) reliability. Importantly, for auditory localization, the response distribution was shifted towards a concurrent spatially discrepant visual signal. This visual spatial bias on the perceived auditory location was increased when the visual signal was reliable, thus replicating the classical profile of the spatial ventriloquist effect [3]. Moreover, it was more pronounced for 13.3° than for 20° disparity. In other words, as expected under Bayesian Causal Inference, the influence of a concurrent visual signal on the perceived auditory location was attenuated for large spatial discrepancies, when it was less likely that auditory and visual signals came from a common source.
Next, we analyzed visual and auditory localization reports more formally by comparing three models. (i) The full-segregation model assumes that auditory and visual signals are processed independently. (ii) The forced-fusion model assumes that auditory and visual signals are integrated weighted by their reliabilities in a mandatory fashion irrespective of the environmental causal structure. (iii) The Bayesian Causal Inference model computes a final auditory (or visual) spatial estimate by averaging the spatial estimates under forced-fusion and full-segregation assumptions weighted by the posterior probabilities of each causal structure (i.e., model averaging, see S3 and S4 Tables for other decision functions). Using a maximum likelihood procedure, we fitted the parameters (e.g., visual variances σV12 − σV22 for the two reliability levels) of the three models individually to each participant’s behavioral localization responses. Bayesian model comparison corroborated previous results [2] and demonstrated that the Bayesian Causal Inference model outperformed the full-segregation and forced-fusion models (82.4% variance explained, exceedance probability of 0.95) (Table 1). In other words, human observers integrate audiovisual spatial signals predominantly when they are close in space and hence likely to come from a common source.
Next, we asked how Bayesian Causal Inference emerged along the auditory and visual cortical hierarchies (Fig 3). In particular, Bayesian Causal Inference entails four spatial estimates: the full-segregation unisensory (i) auditory (ŜA,C = 2) and (ii) visual estimates (ŜV,C = 2), (iii) the “audiovisual forced-fusion estimate” (ŜAV,C = 1), and (iv) the final Bayesian Causal Inference estimate (ŜA & ŜV, pooled over conditions of auditory and visual report) that is obtained by averaging the forced-fusion and the task-relevant unisensory estimates weighted by the posterior probability of each causal structure. We obtained these four spatial estimates for each of the 64 conditions and each participant from the Causal Inference model fitted individually to participant’s behavioral data (Fig 3B, bottom). Using cross-validation, we trained a support vector regression model to decode each of these four spatial estimates from fMRI voxel response patterns in regions along the cortical hierarchies defined by visual retinotopic and auditory localizers (Fig 3C). We quantified the decoding accuracies for each of these four spatial estimates in terms of their correlation between (i) the spatial estimates obtained from the Causal Inference model fitted individually to participants’ localization responses (i.e., training labels for fMRI decoding) and (ii) the spatial estimates decoded from fMRI voxel response patterns. To determine which of the four spatial estimates is primarily encoded in a particular region, we computed the exceedance probability that a correlation coefficient of one spatial estimate was greater than that of any other spatial estimate by bootstrapping the decoding accuracies (Fig 3D).
The profile of exceedance probabilities demonstrates that Bayesian Causal Inference is performed by a hierarchy of multisensory processes in the human brain: At the bottom of the hierarchy, in auditory and visual areas, location is represented on the basis that the two signals are generated by independent sources. Thus, primary sensory areas predominantly encoded the spatial estimate of their preferred sensory modality under information segregation, even though they also showed limited multisensory influences as previously reported [15–21]. At the next stage, in posterior intraparietal sulcus (IPS1–2), location is estimated under the assumption that the two signals are from a common source. In other words, IPS1–2 represented primarily the reliability-weighted integration estimate under forced-fusion assumptions. It is only at the top of the hierarchy, in anterior intraparietal sulcus (IPS3–4), that the uncertainty about whether signals are generated by common or independent sources is taken into account. As predicted by Bayesian Causal Inference, location is estimated in IPS3–4 by combining the full-segregation and the forced-fusion estimates weighted by the posterior probabilities of common and independent sources. Thus, according to Bayesian Causal Inference the spatial estimates in IPS3–4 should be influenced by task-irrelevant sensory signals primarily for small spatial disparities, when signals were likely to be generated by a common event. Critically, while no region could uniquely be assigned one type of spatial estimate, the profile of exceedance probabilities reveals a hierarchical organization of the computational operations in human neocortex.
Recent elegant neurophysiological research in non-human primates has shown how single neurons and neuronal populations implement reliability-weighted integration under forced-fusion assumptions [22–24]. In other words, they presented visual and vestibular signals only with a very small discrepancy, so that signals could be assumed to arise from a common source. Yet, to our knowledge this is the first neuroimaging study that moves beyond traditional forced-fusion models and demonstrates how the brain performs hierarchical Bayesian Causal Inference [2]. Thus, future neurophysiological and modelling research will need to define how single neurons and neuronal populations implement computational operations of Bayesian Causal Inference, potentially via probabilistic population codes [25].
Accumulating evidence has suggested that multisensory interactions are pervasive in human neocortex [18,26–31] starting already at the primary cortical level [15–21]. Indeed, our multivariate decoding analysis also revealed multisensory influences ubiquitously along the auditory and visual processing streams with limited multisensory influences emerging already in primary sensory areas.
To link our study more closely with previous fMRI results of spatial ventriloquism, we have interrogated our data also with a conventional univariate analysis of regional blood-oxygen-level dependent (BOLD) responses. Converging with our model-based findings, this conventional analysis also suggested that low-level sensory areas are predominantly driven by signals of their preferred sensory modality (e.g., visual cortex by visual signals). Yet, in line with previous reports [27,31], visual signals influenced the BOLD response already in the “higher auditory area” (hA) encompassing the planum temporale. Moreover, while activations in parietal areas were still influenced by visual location, they were progressively susceptible to effects of task-context mediated either directly or in interaction with visual reliability (see supporting results and discussion in S1 Fig, S2 Table, and S1 Text). Thus, both the regional BOLD response and the spatial representations encoded in parietal areas and to some extent in auditory areas were influenced by whether the location of the visual or the auditory signal needed to be attended to and reported in line with the principles of Bayesian Causal Inference.
As the current paradigm manipulated the factor of task-relevance over sessions, participants knew the sensory modality that needed to be reported prior to stimulus presentation. Thus, the regional BOLD-response in higher auditory cortices is likely to be modulated by attentional top-down effects [32–36]. Future studies may investigate Bayesian Causal Inference when auditory and visual report trials are presented in a randomized fashion to minimize attention- and expectation-related effects. Alternatively, studies could factorially manipulate (i) the attended and (ii) the reported sensory modality. For instance, participants may be cued to attend to the auditory modality prior to stimulus presentation and yet be instructed to report the visual modality after stimulus presentation.
Yet, despite these attempts Bayesian Causal Inference may inherently entail processes associated with “attentional modulation” in a wider sense, as it computationally requires combining the multisensory forced-fusion estimate with the “task-relevant” unisensory estimate. Critically, however, the effects of attentional modulation or task-relevance invoked by Bayesian Causal Inference should interact with the spatial discrepancy between the sensory signals. Effects of task-relevance should be most pronounced for large spatial discrepancies.
In conclusion, the multivariate analysis based on Bayesian Causal Inference moves significantly beyond identifying multisensory interactions, towards characterizing their computational operations that prove to differ across cortical levels. This methodological approach provides a novel hierarchical perspective on multisensory integration in human neocortex. We demonstrate that the brain simultaneously encodes multiple spatial estimates based on segregation, forced fusion, and model averaging along the cortical hierarchy. Only at the top of the hierarchy, higher-order anterior IPS3–4 takes into account the uncertainty about the causal structure of the world and combines sensory signals as predicted by Bayesian Causal Inference. To our knowledge, this study is the first compelling demonstration of how the brain performs Bayesian Causal Inference, a statistical operation fundamental for perception and cognition.
The study was approved by the human research review committee of the University of Tuebingen (approval number 432 2007 BO1). After giving written informed consent, six healthy volunteers without a history of neurological or psychiatric disorders (all university students or graduates; 2 female; mean age 28.8 years, range 22–36 years) participated in the fMRI study. All participants had normal or corrected-to-normal vision and reported normal hearing. One participant was excluded because of excessive head motion (4.21/3.52 standard deviations above the mean of the translational/rotational volume-wise head motion based on the included five participants).
The visual stimulus was a cloud of 20 white dots (diameter: 0.43° visual angle) sampled from a bivariate Gaussian with a vertical standard deviation of 2.5° and a horizontal standard deviation of 2° or 14° presented on a black background (i.e., 100% contrast). Participants were told that the 20 dots were generated by one underlying source in the center of the cloud.
The auditory stimulus was a burst of white noise with a 5 ms on/off ramp. To create a virtual auditory spatial signal, the noise was convolved with spatially specific head-related transfer functions (HRTFs) thereby providing binaural (interaural time and amplitude differences) and monoaural spatial filtering signals. The HRTFs were pseudo-individualized by matching participants’ head width, height, depth, and circumference to the anthropometry of participants in the CIPIC database [37]. HRTFs from the available locations in the database were interpolated to the desired location of the auditory signal. The behavioral responses from the auditory localizer session (see below) indicated that participants were able to localize the virtual auditory spatial signals in the magnetic resonance (MR) scanner. They were significantly better than chance at discriminating whether two subsequent auditory signals were presented from the same or different locations (mean accuracy = 0.88; mean d’ = 3.14, p = 0.001 in a one sample t-test against zero).
In a spatial ventriloquist paradigm, participants were presented with synchronous, yet spatially congruent or disparate visual and auditory signals (Fig 1A). On each trial, visual and auditory locations were independently sampled from four possible locations along the azimuth (i.e., −10°, −3.3°, 3.3°, or 10°) leading to four levels of spatial discrepancy (i.e., 0°, 6.6°, 13.3°, or 20°). In addition, we manipulated the reliability of the visual signal by setting the horizontal standard deviation of the Gaussian cloud to 2° (high reliability) or 14° (low reliability) visual angle. In an inter-sensory selective-attention paradigm, participants reported their auditory or visual perceived signal location and ignored signals in the other modality. For the visual modality, they were asked to determine the location of the center of the visual cloud of dots. Hence, the 4 × 4 × 2 × 2 factorial design manipulated (i) the location of the visual stimulus ({−10°, −3.3°, 3.3°, 10°}, i.e., the mean of the Gaussian); (ii) the location of the auditory stimulus ({−10°, −3.3°, 3.3°, 10°}); (iii) the reliability of the visual signal ({2°,14°}, standard deviation of the Gaussian); and (iv) task-relevance (auditory-/visual-selective report) resulting in 64 conditions (Fig 1B). Please note that in contrast to our inter-sensory attention paradigm, Koerding and colleagues [2] employed a dual task paradigm where participants reported auditory and visual locations on each trial. Thus, the two paradigms differ in terms of attentional and task-induced processes.
On each trial, synchronous audiovisual spatial signals were presented for 50 ms followed by a variable inter-stimulus fixation interval from 1.75–2.75 s. Participants localized the signal in the task-relevant sensory modality as accurately as possible by pushing one of four spatially corresponding buttons. Throughout the experiment, they fixated a central cross (1.6° diameter).
To maximize design efficiency, stimuli and conditions were presented in a pseudorandomized fashion. Only the factor task-relevance was held constant within a session and counterbalanced across sessions. In each session, each of the 32 audiovisual spatial stimuli was presented exactly 11 times either under auditory- or visual-selective report. On average, 5.9% of the trials were interspersed as null-events in the sequence of 352 stimuli per session. Each participant completed 20 sessions (ten auditory and ten visual localization reports; apart from one participant who performed nine auditory and 11 visual localization sessions). Before the fMRI study, participants completed one practice session outside the scanner.
Audiovisual stimuli were presented using Psychtoolbox 3.09 (www.psychtoolbox.org) [38] running under MATLAB R2010a (MathWorks). Auditory stimuli were presented at ~75 dB SPL using MR-compatible headphones (MR Confon). Visual stimuli were back-projected onto a Plexiglas screen using an LCoS projector (JVC DLA-SX21). Participants viewed the screen through an extra-wide mirror mounted on the MR head-coil resulting in a horizontal visual field of approximately 76° at a viewing distance of 26 cm. Participants performed the localization task using an MR-compatible custom-built button device. Participants’ eye movements and fixation were monitored by recording participants’ pupil location using an MR-compatible custom-build infrared camera (sampling rate 50 Hz) mounted in front of the participants’ right eye and iView software 2.2.4 (SensoMotoric Instruments).
To address potential concerns that our results may be confounded by eye movements, we evaluated participants’ eye movements based on eye tracking data recorded concurrently during fMRI acquisition. Eye recordings were calibrated with standard eccentricities between ±3° and ±10° to determine the deviation from the fixation cross. Fixation position was post-hoc offset corrected. Eye position data were automatically corrected for blinks and converted to radial velocity. For each condition, the number of saccades (defined by a radial eye-velocity threshold of 15° s−1 for a minimum of 60 ms duration and radial amplitude larger than 1°) were quantified (0–875 ms after stimulus onset). Fixation was well maintained throughout the experiment with post-stimulus saccades detected in only 2.293% ± 1.043% (mean ± SEM) of the trials. Moreover, 4 (visual location) × 4 (auditory location) × 2 (visual reliability) × 2 (visual versus auditory report) repeated measure ANOVAs performed separately for (i) % saccades or (ii) % eye blinks revealed no significant main effects or interactions.
To characterize how participants integrate auditory and visual signals into spatial representations, we computed the deviation between the responded location and the mean responded location in the corresponding congruent condition for each trial and in each subject. For instance, for trial i (e.g., auditory location = 3.3°, visual location = −3.3°, visual reliability = low, visual report) we computed the response deviation by comparing the responded visual location in trial i to the mean responded visual location for the corresponding congruent condition (e.g., auditory location = −3.3°, visual location = −3.3°, visual reliability = low, visual report). We then averaged the individual histograms of response deviations across subjects (Fig 2, for an additional analysis of response accuracy see supporting results in S1 Table and S1 Text). Fig 2 shows the histograms of the response deviations as a function of task-relevance, visual reliability and audiovisual disparity (i.e., disparity = visual location − auditory location). Please note that we flipped the histograms for negative spatial disparities and auditory report and the histograms for positive spatial disparities and visual report, so that for both types of reports increasing disparity corresponded to a rightward shift of the task-irrelevant signal in Fig 2. We then combined the histograms for positive and negative spatial disparities to reduce the number of conditions and the complexity of Fig 2.
Details of the Bayesian Causal Inference model of audiovisual perception can be found in Koerding and colleagues [2]. The generative model (Fig 3B) assumes that common (C = 1) or independent (C = 2) sources are determined by sampling from a binomial distribution with the common-source prior P(C = 1) = pcommon. For a common source, the “true” location SAV is drawn from the spatial prior distribution N(μP, σP). For two independent causes, the “true” auditory (SA) and visual (SV) locations are drawn independently from this spatial prior distribution. For the spatial prior distribution, we assumed a central bias (i.e., μP = 0). We introduced sensory noise by drawing xA and xV independently from normal distributions centered on the true auditory (respectively, visual) locations with parameters σA (respectively, σV). Thus, the generative model included the following free parameters: the common-source prior pcommon, the spatial prior variance σP2, the auditory variance σA2, and the two visual variances σV2 corresponding to the two visual reliability levels.
Under the assumption of a squared loss function, the posterior probability of the underlying causal structure can be inferred by combining the common-source prior with the sensory evidence according to Bayes rule (cf. S5 Table):
p(C=
1∣xA,xV)=
p
(
xA,xV∣C=1)pcommon
p(xA,xV)
(1)
In the case of a common source (C = 1) (Fig 3B left), the optimal estimate of the audiovisual location is a reliability-weighted average of the auditory and visual percepts and the spatial prior.
In the case of independent sources (C = 2) (Fig 3B right), the optimal estimates of the auditory and visual signal locations (for the auditory and visual location report, respectively) are independent from each other.
To provide a final estimate of the auditory and visual locations, the brain can combine the estimates under the two causal structures using various decision functions such as “model averaging,” “model selection,” and “probability matching” [39]. In the main paper, we present results using “model averaging” as the decision function that was associated with the highest model evidence and exceedance probability at the group level (see S4 Table; please note that at the within-subject level, model averaging was the most likely decision strategy in only three subjects, see S3 Table, and Wozny and colleagues [39]). According to the “model averaging” strategy, the brain combines the integrated forced-fusion spatial estimate with the segregated, task-relevant unisensory (i.e., either auditory or visual) spatial estimates weighted in proportion to the posterior probability of the underlying causal structures.
Thus, Bayesian Causal Inference formally requires three spatial estimates (ŜAV,C = 1, ŜA,C = 2, ŜV,C = 2) which are combined weighted by the posterior probability of each causal structure into a final estimate (ŜA / ŜV, depending on which sensory modality is task-relevant).
We evaluated whether and how participants integrate auditory and visual signals based on their behavioral localization responses by comparing three models: (i) The observers may process and report auditory and visual signals independently (i.e., the full-segregation model, Equation 3). (ii) They may integrate auditory and visual signals in a mandatory fashion irrespective of spatial disparity (i.e., the forced-fusion model, Equation 2). (iii) The observer may perform Bayesian Causal Inference, i.e., combine estimates from the forced-fusion and the task-relevant estimate from the full-segregation model weighted by the probability of the underlying causal structures (Equations 4 and 5, i.e., model averaging, for other decision functions see S3 Table and S4 Table).
To arbitrate between full segregation, forced fusion, and Bayesian Causal Inference, we fitted each model to participants’ localization responses (Table 1) based on the predicted distributions of the auditory spatial estimates (i.e., p(ŜA|SA,SV)) and the visual spatial estimates (i.e., p(ŜV|SA,SV)). These distributions were obtained by marginalizing over the internal variables xA and xV that are not accessible to the experimenter (for further details of the fitting procedure see Koerding and colleagues [2]). These distributions were generated by simulating xA and xV 5,000 times for each of the 64 conditions and inferring ŜA and ŜV from Equations 1–5. To link p(ŜA|SA,SV) and p(ŜV|SA,SV) to participants’ auditory or visual discrete localization responses, we assumed that participants selected the button that is closest to ŜA or ŜV and binned the ŜA and ŜV accordingly into a histogram (with four bins corresponding to the four buttons). Thus, we obtained a histogram of predicted auditory or visual localization responses for each condition and participant. Based on these histograms we computed the probability of a participant’s counts of localization responses using the multinomial distribution (see Koerding and colleagues [2]). This gives the likelihood of the model given participants’ response data. Assuming independence of experimental conditions, we summed the log likelihoods across conditions.
To obtain maximum likelihood estimates for the parameters of the models (pcommon, σP, σA, σV1 − σV2 for the two levels of visual reliability; formally, the forced-fusion and full-segregation models assume pcommon = 1 or = 0, respectively), we used a non-linear simplex optimization algorithm as implemented in MATLAB’s fmin search function (MATLAB R2010b). This optimization algorithm was initialized with 200 different parameter settings that were defined based on a prior grid search. We report the results (across-subjects' mean and standard error) from the parameter setting with the highest log likelihood across the 200 initializations (Table 1). This fitting procedure was applied individually to each participant’s data set for the Bayesian Causal Inference, the forced-fusion, and the full-segregation models.
The model fit was assessed by the coefficient of determination R2 [40] defined as
R2=1−exp(−2n(l(ß^)−l(0)))
where l(ß^) and l(0) denote the log likelihoods of the fitted and the null model, respectively, and n is the number of data points. For the null model, we assumed that an observer randomly chooses one of the four response options, i.e., we assumed a discrete uniform distribution with a probability of 0.25. As in our case the Bayesian Causal Inference model’s responses were discretized to relate them to the four discrete response options, the coefficient of determination was scaled (i.e., divided) by the maximum coefficient (cf. [40]) defined as
max(R2)=1−exp(2nl(0))
To identify the optimal model for explaining participants’ data, we compared the candidate models using the Bayesian information criterion (BIC) as an approximation to the model evidence [41]. The BIC depends on both model complexity and model fit. We performed Bayesian model selection [42] at the group level as implemented in SPM8 [43] to obtain the exceedance probability for the candidate models (i.e., the probability that a given model is more likely than any other model given the data).
A 3T Siemens Magnetom Trio MR scanner was used to acquire both T1-weighted anatomical images and T2*-weighted axial echoplanar images with BOLD contrast (gradient echo, parallel imaging using GRAPPA with an acceleration factor of 2, TR = 2,480 ms, TE = 40 ms, flip angle = 90°, FOV = 192 × 192 mm2, image matrix 78 × 78, 42 transversal slices acquired interleaved in ascending direction, voxel size = 2.5 × 2.5 × 2.5 mm3 + 0.25 mm interslice gap).
In total, 353 volumes times 20 sessions were acquired for the ventriloquist paradigm, 161 volumes times 2–4 sessions for the auditory localizer and 159 volumes times 10–16 sessions for the visual retinotopic localizer resulting in approximately 18 hours of scanning in total per participant assigned over 7–11 days. The first three volumes of each session were discarded to allow for T1 equilibration effects.
Ventriloquist paradigm. The fMRI data were analyzed with SPM8 (http://www.fil.ion.ucl.ac.uk/spm) [43]. Scans from each participant were corrected for slice timing, were realigned and unwarped to correct for head motion and spatially smoothed with a Gaussian kernel of 3 mm FWHM. The time series in each voxel was high-pass filtered to 1/128 Hz. All data were analyzed in native participant space. The fMRI experiment was modelled in an event-related fashion with regressors entering into the design matrix after convolving each event-related unit impulse with a canonical hemodynamic response function and its first temporal derivative. In addition to modelling the 32 conditions in our 4 (auditory locations) × 4 (visual locations) × 2 (visual reliability) factorial design, the general linear model included the realignment parameters as nuisance covariates to account for residual motion artefacts. The factor task-relevance (visual versus auditory report) was modelled across sessions.
The parameter estimates pertaining to the canonical hemodynamic response function defined the magnitude of the BOLD response to the audiovisual stimuli in each voxel. For the multivariate decoding analysis, we extracted the parameter estimates of the canonical hemodynamic response function for each condition and session from voxels of the regions of interest (= fMRI voxel response patterns) defined in separate auditory and retinotopic localizer experiments (see below). Each fMRI voxel response pattern for the 64 conditions in our 4 × 4 × 2 × 2 factorial design was based on 11 trials within a particular session. To avoid the effects of image-wide activity changes, each fMRI voxel response pattern was normalized to have mean zero and standard deviation one.
Decoding of spatial estimates. To investigate whether and how regions along the auditory and visual spatial processing hierarchy (defined below; cf. Fig 3C) represent spatial estimates of the Causal Inference model, we used a multivariate decoding approach where we decoded each of the four spatial estimates from the regions of interest: (i) the full-segregation visual estimate: ŜV,C = 2, (ii) the full-segregation auditory estimate: ŜA,C = 2, (iii) the forced-fusion audiovisual estimate: ŜAV,C = 1, and (iv) the Bayesian Causal Inference (i.e., model averaging) estimate: ŜA & ŜV, pooled over auditory and visual report (i.e., for each condition we selected the model averaging estimate that needs to be reported in a particular task context). Thus, our decoding approach implicitly assumed that the forced-fusion as well as the auditory and visual estimates under full segregation are computed automatically irrespective of task-context. By contrast, the final auditory or visual Bayesian Causal Inference estimates are flexibly computed depending on the particular task-context according to a decision function such as model averaging. After fitting the Causal Inference model individually to behavioral localization responses (see above), the fitted model predicted these four spatial estimates’ values in 10,000 simulated trials for each of the 64 conditions. The spatial estimates’ values as an index of participants’ perceived location are of a continuous nature. Finally, we summarized the posterior distribution of spatial estimates (i.e., participant’s perceived location) by averaging the values across those 10,000 simulated trials for each of the four spatial estimates separately for each condition and participant. Please note that using the maximum a posteriori estimate as a summary index for the posterior distribution provided nearly equivalent results.
For decoding, we trained a linear support vector regression model (SVR, as implemented in LIBSVM 3.14 [44]) to accommodate the continuous nature of these mean spatial estimates that reflect the perceived signal location for a particular condition and subject. More specifically, we employed a leave “one session” out cross-validation scheme: First, we extracted the voxel response patterns in a particular region of interest (e.g., V1) from the parameter estimate images pertaining to the magnitude of the BOLD response for each condition and session (i.e., 32 conditions × 10 sessions for auditory report + 32 conditions × 10 sessions for visual report = 640 voxel response patterns). For each of the four spatial estimates (e.g., ŜV,C = 2), we trained one SVR model to learn the mapping from the condition-specific fMRI voxel response patterns (i.e., examples) to the condition-specific spatial estimate’s values (i.e., labels) from all but one session (i.e., 640 − 32 = 608 voxel responses patterns). The model then used this learnt mapping to decode the spatial estimates from the 32 voxel response patterns from the single remaining session. In a leave-one-session-out cross-validation scheme, the training-test procedure was repeated for all sessions. The SVRs’ parameters (C and ν) were optimized using a grid search within each cross-validation fold (i.e., nested cross-validation).
We quantified the decoding accuracies for each of these four spatial estimates in terms of the correlation coefficient between (i) the spatial estimates obtained from the Causal Inference model fitted individually to a participant’s localization responses (i.e., these spatial estimates were used as training labels for fMRI decoding, e.g., ŜV,C = 2) and (ii) the spatial estimates decoded from fMRI voxel response patterns using SVR. To determine whether the spatial estimates (i.e., labels) can be decoded from the voxel response patterns, we entered the Fisher z-transformed correlation coefficients for each participant into a between-subject one-sample t-test and tested whether the across-participants mean correlation coefficient was significantly different from zero separately for the (i) segregated auditory or (ii) visual, (iii) forced-fusion audiovisual, or (iv) auditory and visual Bayesian Causal Inference estimate. As these four spatial estimates were inherently correlated, most regions showed significant positive correlation coefficients for several or even all spatial estimates (see S6 Table). Thus, to determine which of the four spatial estimates was predominantly represented in a region, we computed the exceedance probabilities (i.e., the probability that the correlation coefficient of one spatial estimate is greater than the correlation coefficient of any other spatial estimate) using non-parametric bootstrapping across participants (N = 1,000 times). For each bootstrap, we resampled the 5 (= number of participants) individual Fisher z-transformed correlation coefficients with replacement from the set of participants for each of the four spatial estimates and formed the across participants’ mean correlation coefficient for each of the four spatial estimates [45]. In each bootstrap, we then determined which of the four spatial estimates obtained the largest mean correlation coefficient. We repeated this procedure for 1,000 bootstraps. The fraction of bootstraps in which a decoded spatial estimate (e.g., the segregated auditory estimate) had the largest mean correlation coefficient (indexing decoding accuracy) was defined as a spatial estimate’s exceedance probability (Fig 3D). Please note that under the null hypothesis, we would expect that none of the four spatial estimates is related to the voxel response pattern resulting in a uniform distribution of exceedance probabilities for all four spatial estimates (i.e., exceedance probability of 0.25).
Auditory and visual retinotopic localizer. Auditory and visual retinotopic localizers were used to define regions of interest along the auditory and visual processing hierarchies in a participant-specific fashion. In the auditory localizer, participants were presented with brief bursts of white noise at −10° or 10° visual angle (duration 500 ms, stimulus onset asynchrony 1 s). In a one-back task, participants indicated via a key press when the spatial location of the current trial was different from the previous trial. 20 s blocks of auditory conditions (i.e., 20 trials) alternated with 13 s fixation periods. The auditory locations were presented in a pseudorandomized fashion to optimize design efficiency. Similar to the main experiment, the auditory localizer sessions were modelled in an event-related fashion with the onset vectors of left and right auditory stimuli being entered into the design matrix after convolution with the hemodynamic response function and its first temporal derivative. Auditory responsive regions were defined as voxels in superior temporal and Heschl’s gyrus showing significant activations for auditory stimulation relative to fixation (p < 0.05, family-wise error corrected). Within these regions, we defined primary auditory cortex (A1) based on cytoarchitectonic probability maps [46] and referred to the remainder (i.e., planum temporale and posterior superior temporal gyrus) as higher-order auditory area (hA, see Fig 3C).
Standard phase-encoded retinotopic mapping [47] was used to define visual regions of interest (http://sampendu.wordpress.com/retinotopy-tutorial/). Participants viewed a checkerboard background flickering at 7.5 Hz through a rotating wedge aperture of 70° width (polar angle mapping) or an expanding/contracting ring (eccentricity mapping). The periodicity of the apertures was 42 s. Visual responses were modelled by entering a sine and cosine convolved with the hemodynamic response function as regressors in a general linear model. The preferred polar angle was determined as the phase lag for each voxel, which is the angle between the parameter estimates for the sine and the cosine. The preferred phase lags for each voxel were projected on the reconstructed, inflated cortical surface using Freesurfer 5.1.0 [48]. Visual regions V1–V3, V3AB, and IPS0-IPS4 were defined as phase reversal in angular retinotopic maps. IPS0–4 were defined as contiguous, approximately rectangular regions based on phase reversals along the anatomical IPS [49]. For the decoding analyses, the auditory and visual regions were combined from the left and right hemispheres.
|
10.1371/journal.pbio.1001524 | A Conserved Role for Human Nup98 in Altering Chromatin Structure and Promoting Epigenetic Transcriptional Memory | The interaction of nuclear pore proteins (Nups) with active genes can promote their transcription. In yeast, some inducible genes interact with the nuclear pore complex both when active and for several generations after being repressed, a phenomenon called epigenetic transcriptional memory. This interaction promotes future reactivation and requires Nup100, a homologue of human Nup98. A similar phenomenon occurs in human cells; for at least four generations after treatment with interferon gamma (IFN-γ), many IFN-γ-inducible genes are induced more rapidly and more strongly than in cells that have not previously been exposed to IFN-γ. In both yeast and human cells, the recently expressed promoters of genes with memory exhibit persistent dimethylation of histone H3 lysine 4 (H3K4me2) and physically interact with Nups and a poised form of RNA polymerase II. However, in human cells, unlike yeast, these interactions occur in the nucleoplasm. In human cells transiently depleted of Nup98 or yeast cells lacking Nup100, transcriptional memory is lost; RNA polymerase II does not remain associated with promoters, H3K4me2 is lost, and the rate of transcriptional reactivation is reduced. These results suggest that Nup100/Nup98 binding to recently expressed promoters plays a conserved role in promoting epigenetic transcriptional memory.
| Cells respond to changes in nutrients or signaling molecules by altering the expression of genes. The rate at which genes are turned on is not uniform; some genes are induced rapidly and others are induced slowly. In brewer's yeast, previous experience can enhance the rate at which genes are turned on again, a phenomenon called “transcriptional memory.” After repression, such genes physically interact with the nuclear pore complex, leading to altered chromatin structure and binding of a poised RNA polymerase II. Human genes that are induced by interferon gamma show a similar behavior. In both cases, the phenomenon persists through several cell divisions, suggesting that it is epigenetically inherited. Here, we find that yeast and human cells utilize a similar molecular mechanism to prime genes for reactivation. In both species, the nuclear pore protein Nup100/Nup98 binds to the promoters of genes that exhibit transcriptional memory. This leads to an altered chromatin state in the promoter and binding of RNA polymerase II, poising genes for future expression. We conclude that both unicellular and multicellular organisms use nuclear pore proteins in a novel way to alter transcription based on previous experiences.
| The nuclear pore complex (NPC) is a conserved macromolecular structure that mediates the essential transport of molecules between the nucleus and the cytoplasm [1]. The NPC is an 8-fold symmetric channel derived from ∼30 proteins associated with cytoplasmic filaments and a nucleoplasmic “basket” [2],[3]. Natively unstructured NPC proteins rich in phenylalanine-glycine repeats line the channel of the NPC and interactions of these proteins with transport factors facilitates selective transport of proteins and mRNPs [2],[4]–[6]. Proteins that make up the basket-like structure on the nucleoplasmic face of the NPC and the fibrils on the cytoplasmic face of the NPC play key roles in regulating nuclear transport and mRNP remodeling [4],[7].
Nuclear pore proteins also physically interact with chromatin to regulate transcription of certain genes. In Saccharomyces cerevisiae and Drosophila melanogaster, many active genes physically interact with nuclear pore proteins [8]–[11]. Interaction with the NPC has been proposed to promote stronger transcription [9],[12]–[16], to mediate epigenetic regulation [17]–[19], to promote chromatin boundary activity [20],[21], and to provide negative feedback in signaling pathways [22]. However, the exact biochemical nature of these roles, their generality, and their conservation is unclear.
In yeast, some of the inducible genes that relocate from the nucleoplasm to the NPC upon activation [such as GAL1 (GenBank Accession CAA84962.1) and INO1 (GenBank Accession CAA89448.1)] remain at the nuclear periphery for multiple generations after repression, a phenomenon called epigenetic transcriptional memory [17]. The persistent association of genes with the NPC is not associated with transcription, but promotes faster reactivation [17],[18],[23]. In the case of the GAL genes, this leads to significantly faster reactivation compared with activation [17],[24]. This is not always true; in the case of the INO1 gene, perhaps because of the rate at which cells sense the activating signal (inositol starvation) during reactivation, the rate of reactivation is slower than the rate of activation [17],[18]. However, interaction with the NPC after repression specifically promotes INO1 reactivation because when it is lost, the rate of reactivation is slowed [18].
Active INO1 and recently repressed INO1 interact with the NPC by distinct mechanisms. Interaction of active INO1 with the NPC involves cis-acting “DNA zip codes” called gene recruitment sequences (GRSs) in the promoter [15]. Interaction of recently repressed INO1 with the NPC is independent of the GRSs and requires a different zip code called a memory recruitment sequence (MRS), as well as the histone variant H2A.Z (GenBank Accession CAA99011.1) and the nuclear pore protein Nup100 (GenBank Accession CAA81905.1) [18], which is homologous to Nup98 in metazoa. Whereas GRS-mediated interaction of active INO1 with the NPC promotes stronger transcription [15], MRS-mediated interaction of recently repressed INO1 with the NPC promotes incorporation of H2A.Z into the promoter and allows RNA polymerase II (RNAPII) to bind, poising the gene for future reactivation [18]. Mutations in the MRS, loss of H2A.Z, or loss of Nup100 specifically block interaction of recently repressed INO1 with the NPC, leading to loss of RNAPII from the recently repressed promoter and slower reactivation [18]. However, these mutations have no effect on the rate of initial activation or the ultimate steady-state levels of INO1 mRNA [17],[18],[23]. Thus, the interaction of genes with the NPC can both promote stronger expression and, by a distinct mechanism, poise recently repressed genes for future reactivation.
Stress-inducible genes utilize a related type of transcriptional memory. Previous exposure of yeast cells to high salt leads to faster activation of many genes induced by oxidative stress [19]. Similar to INO1 transcriptional memory, this effect persists for four to five generations, suggesting that salt stress establishes an epigenetic change that promotes the rate of activation of these genes. The faster rate of activation of these genes is dependent on the NPC protein Nup42 (GenBank Accession EEU07798.1) [19]. MEME analysis of the promoters of 77 genes exhibiting stress-induced transcriptional memory identified a DNA element very similar to the INO1 MRS element [19]. GAL gene transcriptional memory has been suggested to depend on the NPC-associated protein Mlp1 (GenBank Accession CAA82174.1) [23],[25]. Therefore, although there are some gene-specific features, aspects of the molecular mechanism of INO1 transcriptional memory are shared by diverse yeast genes.
Nuclear pore proteins also interact with metazoan genes to promote their transcription. Inhibiting histone deacetylase activity using trichostatin A in human cells leads to derepression of hundreds of genes, many of which physically interact with the NPC [8]. In Drosophila, nuclear pore proteins interact with the hsp70 locus, the X chromosome in male flies [26]–[28], and genome-wide, thousands of genes [9]. However, in metazoans, some nuclear pore proteins localize both at the NPC and in the nucleoplasm and genes that interact with nuclear pore proteins can localize either at the nuclear periphery or in the nucleoplasm [16],[29]. In Drosophila, of the 18,878 genes that interact with the nuclear pore protein Nup98, 3,810 interacted exclusively with NPC-associated Nup98 and 11,307 interacted exclusively with nucleoplasmic Nup98 (GenBank Accession NP_651187.2) [9]. As in yeast, the interaction of genes with nuclear pore proteins also promotes transcription in flies [9],[16].
Here we sought to explore the role of nuclear pore interactions with genes in promoting transcriptional memory in humans. HeLa cells treated with IFN-γ show much faster and stronger expression of certain target genes if they have previously encountered IFN-γ [30]. This effect persists for up to four cell divisions (96 h), suggesting that it is epigenetically inherited through mitosis [31]. This phenomenon is also associated with changes in chromatin structure; dimethylated histone H3 lysine 4 (H3K4me2) remains associated with the promoter of the interferon-γ (IFN-γ)-inducible gene HLA-DRA for up to 96 h after treatment with IFN-γ [31]. Here, we have determined the scope of IFN-γ transcriptional memory in human cells and compared it with the molecular mechanisms of INO1 transcriptional memory in yeast. Hundreds of the genes that are induced by IFN-γ exhibit transcriptional memory. Following expression, yeast and human genes that exhibit transcriptional memory are marked by H3K4me2 and associate with both a poised RNAPII and Nup100/Nup98 (GenBank Accession AAH12906.2) for up to four generations. Loss of Nup100 in yeast, or transient knockdown of Nup98 in HeLa cells, leads to loss of RNAPII and H3K4me2 from recently expressed promoters and a slower rate of reactivation of genes that exhibit memory. Thus, Nup100/Nup98 is required for epigenetic transcriptional memory, a mechanism conserved from yeast to humans.
After yeast cells are shifted from medium lacking inositol into medium containing inositol, INO1 transcription is rapidly repressed and the mRNA returns to baseline within ∼30 min [17],[18],[23]. RNAPII dissociates from the body of the gene after addition of inositol, but it remains associated with the INO1 promoter for up to four generations after repression [17],[18],[23]. This form of RNAPII is unphosphorylated on the carboxy terminal domain (CTD) on serine 5 (associated with transcription initiation) and serine 2 (associated with transcription elongation), suggesting that it represents a preinitiation form [18]. To explore the nature of RNAPII that is associated with the recently repressed INO1 promoter, we monitored the association of preinitiation complex (PIC) components before, during, and after expression of INO1. We performed ChIP using strains expressing Tandem Affinity Purification (TAP)-tagged components of TFIID, TFIIA, TFIIB, TFIIF, TFIIE, TFIIH, TFIIK, TFIIS, and Mediator from cells grown in long-term repressing (+inositol), activating (−inositol), or recently repressed (−ino→+ino, 3 h) conditions. None of these PIC components bound to the long-term repressed INO1 promoter (Figure 1A). However, like RNAPII, the PIC components TFIID, TFIIA, TFIIB, TFIIF, TFIIE, and TFIIH bound to the promoter both when the gene was active and after repression (Figure 1A). In contrast, three PIC components interacted with the active promoter, but not the recently repressed promoter: the TFIIK component Kin28 (GenBank Accession CAA64904.1, the kinase that phosphorylates serine 5 on the CTD) [32],[33], the TFIIS component Ctk1 (GenBank Accession CAA81980.1, the kinase that phosphorylates serine 2 on the CTD) [34]–[36], and the Mediator component Gal11 (GenBank Accession CAA99056.1) [35],[36]. This is consistent with the conclusion that the RNAPII that binds to the recently repressed INO1 promoter is not phosphorylated on Ser2 or Ser5 of the CTD [18]. We confirmed that Ser5 phosphorylated RNAPII did not remain associated with the INO1 promoter after repression using a monoclonal antiphospho Ser5 CTD antibody (mAb 4h8; Figure S1). Together, these results suggest that a novel, partially assembled PIC associates with the recently repressed INO1 promoter. Furthermore, binding of these components is not sufficient to induce transcription, suggesting that INO1 reactivation is regulated by controlling the association of Mediator, TFIIK, and/or TFIIS.
INO1 transcriptional memory requires an 11 base pair cis-acting element called the MRS in the promoter [18]. Mutation of the MRS blocks interaction of recently repressed INO1 with the NPC, incorporation of the histone variant H2A.Z, and binding of RNAPII to the recently repressed INO1 promoter, resulting in a slower rate of reactivation of INO1 [18]. When inserted at an ectopic locus, the MRS is sufficient to induce both H2A.Z incorporation and interaction with the NPC [18]. To test if the MRS was also sufficient to induce the association of a poised RNAPII at an ectopic locus, we inserted the MRS adjacent to the URA3 locus (GenBank Accession AAB64498.1) [18] and performed ChIP for RNAPII. We fixed and harvested cells that had been shifted from activating to repressing conditions for 3 h so that we could simultaneously monitor the recovery of the endogenous INO1 locus as an internal positive control. Although RNAPII associated with the recently repressed INO1 promoter under these conditions, it did not associate with URA3 or URA3:MRS or a negative control locus (the GAL1 promoter; Figure 1B). Therefore, the MRS is not sufficient to recapitulate all facets of INO1 transcriptional memory and assembly of the PIC requires other features of the promoter. This suggests that the interaction with the NPC and the incorporation of H2A.Z occur upstream of, and presumably promote, assembly of the PIC.
The HLA-DRA gene in HeLa cells (GenBank Accession CAG33294.1, encoding the HLA class II histocompatibility antigen DRα chain) exhibits a form of transcriptional memory in response to IFN-γ. Cells previously treated with IFN-γ induce HLA-DRA more rapidly and more robustly in response to subsequent exposure to IFN-γ (Figure 2A) [31]. Not all IFN-γ-inducible genes behave this way; another IFN-γ-inducible gene, CIITA (GenBank Accession NP_000237.2), does not display transcriptional memory [31]. Similar to INO1 transcriptional memory, this type of transcriptional memory is epigenetically inherited, persisting through at least four cell divisions in HeLa cells (96 h) [31]. These similarities led us to ask if these two systems utilize related molecular mechanisms.
We used ChIP to examine the association of RNAPII with the HLA-DRA promoter before (uninduced), during (induced), or at various times after treatment with IFN-γ [48 h (∼2 cell divisions) and 96 h (∼4 cell divisions) postinduction]. Prior to IFN-γ treatment, RNAPII was not associated with the promoter or the coding sequence of HLA-DRA and CIITA (Figure 2B). This is consistent with the undetectable levels of HLA-DRA and CIITA mRNA before IFN-γ treatment (Figure 2A). During IFN-γ treatment, RNAPII associated strongly with the promoter and the coding sequence of both HLA-DRA and CIITA (Figure 2B). After removing IFN-γ, RNAPII remained associated with the HLA-DRA promoter, but not the coding sequence, for up to 96 h (Figure 2B). RNAPII did not associate with CIITA after removing IFN-γ and the levels of RNAPII associated with GAPDH (GenBank Accession AAH83511.1) promoter and coding sequence were consistent under all three conditions (Figure 2B). Therefore, HLA-DRA memory correlates with persistent RNAPII binding to the previously induced promoter through at least four cell divisions.
Following treatment of cells with IFN-γ, the signaling and transcriptional response can persist even after washing, presumably because of persistent association of IFN-γ with the IFN-γ receptor and signaling through the JAK/STAT pathway (Figure S2A and B) [37]. For this reason, we trypsinized and split the cells after removing IFN-γ in all of our experiments, which leads to the levels of HLA-DRA mRNA returning to baseline levels within 6 h (Figure S2A). Likewise, treatment of HeLa cells with IFN-γ immediately after trypsinizing did not result in expression of HLA-DRA (Figure S2B), suggesting that trypsin digestion blocks IFN-γ signaling. Thus, the association of RNAPII with the HLA-DRA promoter is not due to persistent expression. This is consistent with loss of RNAPII from the HLA-DRA coding sequence after removing IFN-γ and splitting (Figure 2B).
We also tested if previous expression of HLA-DRA is necessary for transcriptional memory. Exposure of cells to IFN-γ for 2 h, followed by splitting the cells, does not result in significant HLA-DRA expression (Figure S2B). However, this brief exposure to IFN-γ is sufficient to induce a faster rate of reactivation 48 h later (∼2 cell divisions; Figure S1C). Thus, previous expression of HLA-DRA is not necessary to induce future transcriptional memory.
In metazoans, transcription is regulated both by blocking RNAPII recruitment and by blocking RNAPII elongation [38]–[42]. In the latter case, RNAPII binds to the promoter, initiates transcription, and then pauses at the 5′ end of the gene due to regulation by negative elongation factor (NELF; GenBank Accession AAI10499.1) and DRB sensitivity-inducing factor (DSIF; GenBank Accession BAA24075.1) [43]. This paused RNAPII is phosphorylated on Ser5 of the CTD, but unphosphorylated on Ser2 [38],[44],[45]. Transcription of such genes is stimulated by recruitment of the kinase P-TEFb, which phosphorylates Ser2 and allows elongation [46]. In yeast, ChIP using a monoclonal anti-phospho serine 5 antibody (mAb 4h8) recovers active, but not recently repressed INO1 promoter (Figure S2). We performed ChIP using mAb 4h8 to ask if the RNAPII associated with the previously expressed HLA-DRA promoter is postinitiation or preinitiation. Ser5 phosphorylated RNAPII bound to the active HLA-DRA promoter in cells exposed to IFN-γ, but not with the previously expressed HLA-DRA promoter after removal of IFN-γ (Figure 2C). Therefore, similar to yeast INO1, the HLA-DRA promoter associates with a preinitiation form of RNAPII for several cell divisions after removing IFN-γ.
In yeast, distinct Nups interact with active [11] and recently repressed INO1 [18]. To test if HLA-DRA interacts with Nups, we performed ChIP using the mAb 414 monoclonal antibody, which recognizes Phe-x-Phe-Gly repeats present in several nuclear pore proteins [47],[48]. We observed strong interaction of Phe-x-Phe-Gly repeat proteins with the active HLA-DRA promoter and a weaker interaction after removing IFN-γ (Figure 3A). This pattern was very similar to the interaction of the Phe-x-Phe-Gly repeat protein Nup2 (GenBank Accession AAB67259.1) with the INO1 promoter in yeast [18]. The interaction was also specific; mAb 414 did not recover the HLA-DRA coding sequence (not shown) nor the promoters of CIITA, GAPDH, and β-ACTIN (Figure 3A). This suggests that Phe-x-Phe-Gly Nups interact with the active and recently expressed HLA-DRA promoter.
The yeast nuclear pore protein Nup100 interacts with the INO1 promoter specifically after repression, and not during activation [18]. We performed ChIP using an antibody against Nup98, a human homologue of Nup100 [1], and analyzed the interaction with the HLA-DRA promoter. Nup98 did not interact with the HLA-DRA promoter before or during IFN-γ treatment (Figure 3B). However, for up to 96 h (∼4 cell divisions) after removal of IFN-γ, we observed a clear and specific association of Nup98 with the HLA-DRA promoter. Therefore, similar to the specific interaction of Nup100 with the recently repressed INO1 promoter, Nup98 interacts specifically with the recently expressed HLA-DRA promoter.
In Drosophila, genes interact with Nups both at the NPC and in the nucleoplasm [9],[16]. In particular, Nup98 has been shown to localize both at the NPC and the nuclear periphery [49]. To test if the HLA-DRA interaction with Nup98 occurs at the NPC, we localized the HLA-DRA gene with respect to the nuclear periphery using DNA fluorescence in situ hybridization (FISH). Cells were fixed before (uninduced), during (induced), or after (48 h postinduction) treatment with IFN-γ and processed for DNA-FISH using fluorescent probes generated by nick translation of bacterial artificial chromosomes (BACs). Using confocal microscopy, we measured the distance from the individual HLA-DRA foci to the edge of the Hoescht fluorescence within individual z slices (Figure 3C). The distribution of these distances was plotted for ∼300 foci. Under all three conditions, the HLA-DRA gene localized in the nucleoplasm (Figure 3D). Active HLA-DRA was somewhat more nucleoplasmic than the preinduced HLA-DRA (p = 0.004, two-tailed t test) or postinduced HLA-DRA (p = 0.001). The position of CIITA with respect to the nuclear periphery did not change under these conditions (Figure S4). This suggests that HLA-DRA interacts with Nups away from the NPC, in the nucleoplasm.
To probe the generality of IFN-γ-induced transcriptional memory throughout the human genome, we performed expression microarrays on cDNA from cells treated with IFN-γ. We compared samples from time points either during initial activation or during reactivation after 48 h without IFN-γ (∼2 cell divisions, as in Figure 2A). The log2 ratios relative to the initial time point (0 h) were calculated by averaging between replicates and, for genes with multiple probes, between probes. Based on the initial activation after addition of IFN-γ, we identified a subset of 664 genes that were induced ≥2 fold between 6 h and 24 h (Table S1). Gene ontology (GO) analysis revealed that this subset of genes was highly enriched for terms related to innate immunity: “regulation of immune system process” (p = 7.76×10−23), “response to interferon gamma” (p = 3.99×10−18), and “response to cytokine stimulus” (p = 1.16×10−17; Table S2) [50]. We then used k means clustering to organize this subset into clusters on the basis of their behaviors during activation and reactivation (Figure 4A and Table S1).
Cluster 1 includes 218 genes that were modestly induced during activation and more strongly induced during reactivation (see average behavior in Figure 4B). This cluster was strongly enriched for genes involved in inflammation and genes regulated by infection (Table S3). Cluster 2 includes 403 genes that were induced equivalently during activation and reactivation (Figure 4C). This cluster was enriched for GO terms associated with innate immunity (Table S4). Cluster 3 includes 42 of the most strongly induced genes that were, nonetheless, induced more rapidly during reactivation (Figure 4D). This cluster includes HLA-DRA (Figure 4A, arrow) and was highly enriched for GO terms associated with cytokine signaling generally and IFN-γ signaling in particular (Table S5).
Many of the genes in Cluster 1 and most of the genes in Cluster 3 displayed mRNA profiles consistent with transcriptional memory. We analyzed genes from each of these clusters by RT qPCR to confirm this behavior. The Cluster 1 gene HLA-DQB1 (GenBank Accession AAA59770.1) and the Cluster 3 genes HLA-DPB1 (GenBank Accession AAA59837.1) and OAS2 (GenBank Accession AAH10625.1) displayed significantly faster and/or stronger activation in cells previously exposed to IFN-γ (Figure S5B–D). HLA-DQB1 and HLA-DPB1 encode the HLA class II histocompatibility antigen DQα and DPβ chains, respectively, and OAS2 encodes a 2′-5′-oligoadenylate synthetase [51],[52]. However, the clustering algorithm did not result in perfect segregation of genes with memory from genes without memory; the gene CIITA, which does not exhibit obvious transcriptional memory (Figure S5A) [31], was within Cluster 1. Regardless, these results suggest that a large subset of the genes induced by IFN-γ exhibit stronger or more rapid induction in response to IFN-γ if the cells have been previously exposed to IFN-γ.
To test if other genes that exhibit memory are regulated by the same mechanism as HLA-DRA, we used ChIP against RNAPII and Nup98 before, during, and after IFN-γ treatment. Both RNAPII (Figure S5E) and Nup98 (Figure 4E) bound to the promoters of HLA-DQB1, HLA-DPB1, and OAS2 for up to 96 h (∼4 cell divisions) after removing IFN-γ. Neither RNAPII (Figure S5E) nor Nup98 (Figure 4C) bound to the coding sequences of these genes after removal of IFN-γ. Therefore, the interaction of RNAPII and Nup98 with recently expressed promoters is a general feature of IFN-γ memory.
Transcriptional memory in yeast and humans is associated with changes in chromatin. In yeast, the histone variant H2A.Z is incorporated into a single nucleosome in the INO1 promoter after repression and loss of H2A.Z blocks INO1 transcriptional memory [17],[18]. It is unclear if H2A.Z is involved in IFN-γ transcriptional memory; HLA-DRA memory is associated with a very slight increase in H2A.Z incorporation into the promoter after removal of IFN-γ (Figure S5F). However, HLA-DRA transcriptional memory is associated with persistent dimethylation of histone H3 lysine 4 (H3K4me2) [31]. Histone H3 in nucleosomes at the 5′ end of actively transcribed genes are trimethylated on lysine 4 (H3K4me3) [53],[54]. Whereas the H3K4me3 mark is lost from the promoter of HLA-DRA after removal of IFN-γ, the H3K4me2 mark remains (Figure S6A and B) [31]. In contrast, both marks are lost from the CIITA promoter after removing IFN-γ (Figure S6A and B) [31].
To test if H3K4me2 was associated with INO1 transcriptional memory in yeast, we examined the association of H3K4me3 and H3K4me2 with the INO1 promoter under long-term repressing, activating, or recently repressed conditions. As a control, we used a strain in which the MRS had been mutated and INO1 transcriptional memory is blocked [18]. H3K4me3 was only associated with the active INO1 promoter (Figure 5A). However, H3K4me2 was associated with both the active and recently repressed INO1 promoters (Figure 5B). The persistence of H3K4me2 after repression required the MRS (Figure 5B). Therefore, INO1 transcriptional memory is also associated with dimethylation of H3K4.
We next asked if the machinery responsible for methylation of H3K4 was required for other aspects of INO1 transcriptional memory; namely, poised RNAPII association and localization at the nuclear periphery after repression. In yeast strains lacking either the histone methyltransferase Set1 (GenBank Accession AAB68867.1) or E2 ubiquitin-conjugating enzyme Rad6 (GenBank Accession CAA96761.1), all di- and tri-methylation of H3K4 is lost [55]. Loss of these enzymes did not affect the localization of active INO1 to the nuclear periphery (Figure 5C) or interaction of RNAPII with active INO1 (Figure 5D). However, loss of either Set1 or Rad6 specifically disrupted both localization of recently repressed INO1 at the nuclear periphery (Figure 5C) and RNAPII binding to the recently repressed INO1 promoter (Figure 5D). This suggests that H3K4me2 at the INO1 promoter is required for INO1 transcriptional memory.
The MRS is necessary for both incorporation of H2A.Z [18] and the persistent dimethylation of H3K4 (Figure 5B) at the recently repressed INO1 promoter. Integration of the MRS at ectopic sites is sufficient to induce H2A.Z deposition [18]. Therefore, we asked if the MRS was also sufficient to induce dimethylation of H3K4 at an ectopic locus. We performed ChIP against H3K4me2 in a strain in which either the MRS or the nonfunctional mrs mutant was integrated beside URA3 [18]. We observed a robust signal for H3K4me2 associated with URA3:MRS but not with URA3:mrsmut (Figure 5E) or a control locus (the coding sequence of the repressed gene PRM1; not shown). We also observed a small but reproducible increase in H3K4me3 at URA3:MRS compared with URA3:mrsmut, although this level was significantly lower than the level associated with the active INO1 promoter (Figure S6C). Therefore, the MRS is sufficient to induce both H2A.Z incorporation and dimethylation of H3K4, recapitulating the chromatin changes associated with INO1 transcriptional memory.
Loss of H2A.Z or H3K4 dimethylation leads to loss of INO1 transcriptional memory and both of these modifications require the MRS (Figure 5) [17],[18]. We next asked if the methylation of H3K4 at the recently repressed INO1 promoter required H2A.Z. In wild-type and htz1Δ strains, we observed similar levels of H3K4me3 (Figure S6D) and H3K4me2 (Figure 5F) at the active and recently repressed INO1 promoter. Therefore, dimethylation of H3K4 at the recently repressed INO1 promoter requires the MRS, but not H2A.Z. Furthermore, although strains lacking H2A.Z show slower INO1 reactivation kinetics, loss of peripheral localization of recently repressed INO1, and loss of RNAPII from the INO1 promoter after repression [18], the INO1 promoter is still marked by H3K4me2. This suggests that dimethylation of H3K4 occurs upstream of, or independent of, H2A.Z deposition to promote INO1 transcriptional memory.
INO1 transcriptional memory requires Nup100 for both rapid reactivation and RNAPII association after repression [18]. Because we observed a specific physical interaction of Nup98 with the previously induced HLA-DRA promoter, we asked if Nup98 was required for IFN-γ transcriptional memory. We used transient siRNA knockdown to reduce the levels of Nup98 prior to expression and ChIP analysis (schematized in Figure 6A). Transient knockdown reduced Nup98 during the time course of the experiment; 5 d after transfection, Nup98 protein levels were still reduced, while at earlier times Nup98 was not detected (Figure 6B). Both nuclear pore-associated Nup98 and nucleoplasmic Nup98 were depleted by this treatment (Figure S7). We did not observe a significant change in the growth rate or morphology of the cells subjected to this treatment.
Knockdown of Nup98 had no apparent effect on RNAPII binding to active HLA-DRA (Figure 6C), on the rate of initial activation of HLA-DRA (Figure 6D), or on the association of RNAPII with the CIITA gene (Figure 6C). However, knockdown of Nup98 blocked binding of RNAPII to the HLA-DRA promoter following removal of IFN-γ (Figure 6C) and dramatically reduced the rate of reactivation of HLA-DRA (Figure 6D). This suggests that Nup98 is required for HLA-DRA transcriptional memory.
Because Nup98 bound to the promoters of the Cluster 1 gene HLA-DQB1 and the Cluster 3 genes HLA-DPB1 and OAS2 after removal of IFN-γ (Figure 4E), we tested the effect of Nup98 knockdown on the transcriptional memory of these genes. Transient knockdown of Nup98 reduced the rate of reactivation of all three genes (Figure 6E, F, and G). In the cases of HLA-DQB1 and OAS2, this effect was specific for reactivation. However, in the case of HLA-DPB1, we also observed a slower rate of activation (Figure 6F). Therefore, knockdown of Nup98 affects the transcriptional memory of several human genes, although this effect may not be memory-specific in all cases.
To explore the role of Nup98 in regulating chromatin structure, we asked if the dimethylation of H3K4 associated with transcriptional memory required Nup98. We performed ChIP against H3K4me2 in cells knocked down for Nup98 and quantified the enrichment of this mark at the promoters of HLA-DRA, HLA-DPB1, HLA-DQB1, OAS2, and CIITA (Figure 7A). We observed H3K4me2 at the recently expressed promoters of HLA-DRA, HLA-DPB1, HLA-DQB1, and OAS2, and this mark was lost when Nup98 was knocked down (Figure 7A). Consistent with the impaired activation of HLA-DPB1 in the absence of Nup98 (Figure 6G), we also observed a slight decrease in the level of H3K4me2 associated with the active HLA-DPB1 promoter when Nup98 was knocked down (Figure 7A). Therefore, Nup98 is required for dimethylation of H3K4 at the promoters of genes with transcriptional memory.
To confirm that the effects of NUP98 knockdown are specific, we tested the effect of knockdown of NUP107, a component of the core channel of the NPC [56], on H3K4 dimethylation of IFN-γ-inducible promoters. Using the same strategy to knockdown Nup107, the protein was effectively depleted (Figure 7B). H3K4me2 levels associated with the promoters of HLA-DRA, HLA-DQB1, HLA-DPB1, and OAS2 after removal of IFN-γ remained high in the absence of Nup107 (Figure 7C). Therefore, the effects of Nup98 knockdown on transcriptional memory were specific.
We next asked if dimethylation of H3K4 at the recently repressed INO1 promoter requires yeast Nup100. We grew wild-type and nup100Δ yeast strains in repressing, activating, and recently repressed conditions and performed ChIP against H3K4me2 and H3K4me3 (Figure 7D and Figure S8). Cells lacking Nup100 exhibited normal trimethylation (Figure S8) and dimethylation of H3K4 at the active INO1 promoter but did not maintain H3K4me2 at the recently repressed INO1 promoter (Figure 7D). Therefore, in yeast and human cells, Nup98/Nup100 is required for maintenance of histone H3 dimethylation during transcriptional memory.
Dimethylation of H3K4 is generally associated with the 5′ coding sequences of actively transcribed genes [57]. However, H3K3me2 of promoter regions, often associated with ncRNAs, leads to recruitment of the Set3 histone deacetylase complex (Set3C) and transcriptional repression [58]. We wondered if Set3C had any role in INO1 transcriptional memory. This is somewhat complicated by the poor expression of INO1 in mutants lacking Set3 (GenBank Accession EEU07596.1), suggesting that Set3 might have multiple effects on INO1 expression [59]. However, we tested if Set3 plays a role in binding of RNAPII to the recently repressed INO1 promoter. In mutant strains lacking Set3, RNAPII failed to remain associated with the INO1 promoter after repression (Figure 7E). This suggests that recognition of H3K4me2 by Set3 is required for INO1 transcriptional memory.
Here we further define the molecular mechanism of INO1 transcriptional memory in yeast and demonstrate that aspects of this mechanism are conserved in human cells. Despite over a billion years of evolutionary time [60], in both systems, transcriptional memory requires the interaction of the nuclear pore protein Nup98/Nup100 with recently expressed promoters. This interaction leads to (1) dimethylation of histone H3 lysine 4 in promoter nucleosomes, (2) binding of a preinitiation form of RNAPII, and (3) faster reactivation. Loss of Nup98 in human cells and Nup100 in yeast leads to loss of H3K4me2 and RNAPII from the recently expressed promoters and a slower rate of reactivation. Therefore, Nup98/Nup100-dependent epigenetic transcriptional memory is a conserved mechanism of transcriptional regulation in both unicellular and multicellular organisms.
Our work suggests that transcription can be regulated at three distinct stages: RNAPII recruitment, transcription initiation, and transcription elongation. In yeast, the primary mechanism by which transcription is regulated is through recruitment of RNAPII/PIC to the promoter and inducible genes tend to be devoid of RNAPII when uninduced or repressed [61]–[64]. However, under certain circumstances, a preinitiation form of RNAPII can associate with inactive yeast promoters. For example, in stationary phase cells, unphosphorylated RNAPII is associated with hundreds of inactive promoters, and this has been suggested to poise these genes for future activation [65]. Likewise, in metazoans, transcription can be regulated both at the level of PIC assembly and after initiation, at the level of transcription elongation [38],[44],[66]–[68]. Promoter-proximal pausing requires NELF and DSIF and is relieved by recruitment of pTEF-b [46],[69],[70]. Brewer's yeast lacks a homologue of NELF, and there is no conclusive evidence for RNAPII pausing [71]. Therefore, promoter-proximal pausing may be a metazoan-specific form of regulation [64].
Our results suggest that, for certain genes, the mechanism of regulation depends on the history of the cells. Transcription of such genes is regulated by either preventing RNAPII/PIC recruitment (under long-term repressing conditions) or allowing RNAPII/PIC recruitment but preventing transcription initiation (under recently repressed conditions). In the case of the INO1 gene, whereas none of the PIC components bound to the long-term repressed INO1 promoter, most of them bound to the recently repressed INO1 promoter (Figure 1). This form of PIC is distinct from the PIC that associates with active INO1: TFIIK, Mediator, and TFIIS are absent and RNAPII remains unphosphorylated and fails to initiate. Thus, while the regulation of long-term repressed INO1 prevents binding of RNAPII and the rest of the PIC, the regulation of recently repressed INO1 occurs at a subsequent step. This suggests that the rate-limiting step in derepression is different for long-term repressed INO1 and recently repressed INO1.
In both yeast and humans, transcriptional memory is inherited through cell division. In the case of yeast, the INO1 gene remains at the nuclear periphery, associated with the NPC for ∼3–4 generations (≥6 h) in both the mother and daughter cells [17]. Likewise, RNAPII remains associated with the INO1 promoter over the same number of generations [18]. HeLa cells exposed to IFN-γ exhibit faster and more robust activation of IFN-γ-inducible genes for up to ∼4 generations (96 h) after removing IFN-γ [31]. Binding of RNAPII and Nup98 to, and dimethylation of histone H3 lysine 4 over, the promoters of genes exhibiting transcriptional memory persists over the same number of generations. Therefore, the poised, preinitiation state is heritable through several generations, suggesting that it represents an epigenetic state.
Changes in chromatin composition and modification are necessary for transcriptional memory and presumably allow binding of RNAPII/PIC to recently expressed promoters in yeast and humans. For genes that display transcriptional memory in both yeast and humans, histone H3 within promoter nucleosomes was unmethylated on lysine 4 prior to induction (Figures 5 and 7). After expression, histone H3 within promoter nucleosomes was dimethylated on lysine 4 (H3K4me2; Figures 5 and 7) [31]. In human cells, this correlates with lower nucleosome occupancy of the recently expressed HLA-DRA promoter compared with the uninduced promoter [31]. In yeast, a cis-acting element necessary for INO1 transcriptional memory (the MRS) was necessary for dimethylation of H3K4 after repression and insertion of the MRS at an ectopic locus is sufficient to induce both H2A.Z incorporation [18] and dimethylation of H3K4 (Figure 5). Finally, loss of the enzymes responsible for H3K4 methylation (Set1 or Rad6; Figure 5) or recognition of H3K4me2 (Set3; Figure 7) led to loss of INO1 transcriptional memory. Although it is still formally possible that Set1 methylates another protein required for transcriptional memory, the connection to the cis-acting MRS element and the requirement for Rad6 and Set3 (Figure 7) suggest that Set1 methylation of H3K4 is required for transcriptional memory.
Binding of nuclear pore proteins to the promoters of genes impacts both their transcription and their epigenetic regulation. In yeast and Drosophila, nuclear pore proteins interact with the promoters of active genes and this interaction is required for their full expression [9],[12]–[16]. We have found another role for these interactions. The yeast nuclear pore protein Nup100 is required for INO1 transcriptional memory and the homologous human protein Nup98 is required for IFN-γ-mediated memory.
One complication of any experiment manipulating Nup98 levels is that both Nup98 and Nup96 are generated from a single transcript [72]. After nuclear import, this protein undergoes autocatalytic cleavage, producing two proteins [73]. Indeed, knockdown of Nup98 also leads to depletion of Nup96 (Figure S9). Therefore, our results raise the possibility that both Nup98 and Nup96 impact transcriptional memory, especially since Nup96 has been implicated in promoting expression of interferon-responsive genes in mouse [74] and the Nup96 homologue Nup145C (GenBank Accession P49687) is required for localization of recently repressed INO1 at the nuclear periphery in yeast [18]. Both Nup96 and Nup98 localize in the nucleoplasm and at the NPC [49],[75]. Our data suggest that Nup98 plays a direct and specific role: Nup98 binds to the promoters of genes that exhibit transcriptional memory specifically after removal of IFN-γ and knockdown affects reactivation rate without affecting activation rate. Also, although loss of Nup96 is lethal [74],[76], we did not observe a strong defect in the growth of cells transiently knocked down for Nup98, suggesting that Nup96 function was not completely depleted. And finally, knockdown of Nup107, another component of the same subcomplex of the NPC as Nup96, had no effect on H3K4 dimethylation of promoters of primed genes. We conclude that our data support a role for Nup98, and potentially Nup96, in transcriptional memory. Future work will be required to separate these two roles.
Phe-x-Phe-Gly repeat proteins that are recognized by mAb 414 interact with the promoter of both active HLA-DRA and recently expressed HLA-DRA. Nup98, which possesses related repeated motif (Gly-Leu-Phe-Gly), interacts specifically with the promoter of recently expressed HLA-DRA. This suggests that active HLA-DRA and recently expressed HLA-DRA interact with two distinct sets of nuclear pore proteins. This conclusion is very similar to what we have observed for INO1; active INO1 and recently repressed INO1 interact with different Nups, and localization of active and recently repressed INO1 at the nuclear periphery requires different Nups and different DNA elements [18]. The interaction of active and recently expressed HLA-DRA with Nups occurs in the nucleoplasm. Consistent with previous work [49], this suggests that in both Drosophila and human cells there are two pools of nuclear pore proteins: a pool at the NPC and a pool in the nucleoplasm. Although the interaction between genes and Nup100 occurs at the NPC in yeast and the interaction with Nup98 occurs in the nucleoplasm in human cells, the biochemical outputs of these interactions are conserved.
After removing IFN-γ, the HLA-DRA gene shows increased colocalization with PML bodies, nuclear “dots” that are enriched for the promyelocytic leukemia factor (PML) [31]. PML bodies increase in number after IFN-γ treatment and depletion of PML led to a decrease in both the rate of HLA-DRA reactivation and loss of H3K4me2 after removing IFN-γ [31]. This suggests that relocalization of HLA-DRA to these structures is required for IFN-γ memory. Foci of Nup98 in the nucleoplasm do not colocalize with PML bodies [49], suggesting that HLA-DRA may not colocalize with Nup98 foci. It will be important to understand how PML bodies and nuclear pore proteins impact each other in this process.
The role of H2A.Z in transcriptional memory is controversial. Whereas H2A.Z is required for INO1 transcriptional memory and is intimately connected to MRS function [18], loss of H2A.Z affects both the rate of activation and reactivation of GAL genes and its role in GAL gene transcriptional memory has been challenged [77]–[79]. The MRS from the INO1 promoter is sufficient to induce both incorporation of H2A.Z and dimethylation of H3K4. However, loss of H2A.Z did not block dimethylation of H3K4, suggesting that H3K4me2 occurs upstream of, or in parallel to, H2A.Z deposition. In HeLa cells, we observed only a slight increase in H2A.Z levels associated with the HLA-DRA promoter after removing IFN-γ (Figure S5F). Although a large increase in H2A.Z association is not necessary for H2A.Z to play a role in memory, it is possible that H2A.Z functions as a gene-specific regulator of a more general system to promote reactivation.
Our results suggest that, under certain circumstances, Nup98 regulates H3K4 methylation of a gene that colocalizes with PML bodies. Translocations that result in fusion of PML with the retinoic acid receptor α lead to loss of PML bodies, altered transcription, and acute promyelocytic leukemia [80],[81]. Translocations that result in fusion of Nup98 to transcription factors lead to acute myeloid leukemias [82]–[85]. Finally, translocations that result in fusions of >60 different genes with the H3K4 methyltransferase MLL result in acute lymphoblastic leukemia, in part through altered Hox gene expression [86]–[88]. These striking similarities raise the possibility that these translocations impact the expression of an overlapping set of genes through similar mechanisms, perhaps involving transcriptional priming.
Transcriptional memory plays a broad role in gene priming of IFN-γ-responsive genes (Figure 4). Hundreds of genes displayed faster or stronger induction kinetics in cells that have previously been exposed to IFN-γ and that this effect persists for days in rapidly doubling HeLa cells. This suggests that transcriptional memory can qualitatively alter the response of a system to a particular stimulus. If so, it is possible that the response of cells to other stimuli can be modulated by transcriptional memory. For example, similar to the phenomenon of stress cross-protection in yeast [19], the rate of induction in response to one cytokine could also be qualitatively or quantitatively altered by previous exposure to a different cytokine. Because transcriptional memory is epigenetically inherited through several cell divisions, it could alter the response of cells, tissues, or whole organisms to persistent or episodic stimuli over long timescales. If so, then it might play an important role in pathological inflammation [89],[90].
Unless otherwise noted, chemicals used were obtained from Sigma Aldrich and enzymes were from New England Biolabs. BACs were from Invitrogen. 8WG16 antibody was obtained from Covance, anti-Ser5P CTD (cat no. ab5408), anti-Phe-x-Phe-Gly m414 (cat no. ab50008), anti-Nup98 (cat no. ab45584), anti-Nup96 (cat no. ab124980), anti-Nup107 (cat no. ab85916), anti-H2A.Z (cat no. ab4174), anti-H3K4me2 (cat no. ab32356), and anti-H3K4me3 (cat no. ab1012) were from AbCam. IFN-γ was from PBL Biomedical.
Yeast strains used in this study are listed in Table S6. Strains with the MRS or the mrs mutant elements have been described [18] and were created as described [15].
For yeast experiments, ChIP was performed as described [18]. For TAP-tagged ChIP experiments, Pan Mouse IgG Dynabeads from Invitrogen were used. For HeLa experiments, cells were trypsinized and fixed using 1% formaldehyde for 15 min at 25°C. Cross-linking was quenched using 150 mM glycine, and cells were harvested by centrifugation and washed twice with ice-cold PBS. Cells were lysed in 10 ml MC lysis buffer (10 mM NaCl, 10 mM Tris-HCl, 3 mM MgCl2, 0.5% NP-40) and nuclei were recovered by centrifugation at 1,350 rpm twice and snap frozen in liquid nitrogen. Fixed nuclei were resuspended in 1 ml lysis buffer (10 mM Tris-HCl, 100 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.5% N-lauroylsarcosine) with protease inhibitors (Roche) and sonicated using a Branson 450 microtip 16 times for 15 s at setting 5 to generate ∼500 bp average fragment size. 1% Triton X-100 and 0.1% sodium deoxycholate were added back and then chromatin was spun at ∼16,100× g at 4°C for 15 min. The supernatant fraction was added to antibody and Dynabeads overnight at 4°C. The beads were recovered and washed four times with lysis buffer+Triton X-100 and sodium deoxycholate. Immunoprecipitated chromatin was eluted in 100 µl TE+1% SDS at 65°C for 15 min. Input and IP fractions were treated with 50 µg RNase A and 100 µg Proteinase K for 1 h at 42°C, before reversing crosslinks overnight at 65°C. DNA was extracted with phenol∶chloroform∶isoamyl alcohol, and chloroform and 2 µg linear acrylamide was added prior to ethanol precipitation. Samples were washed with 70% ethanol and resuspended in 30 µl TE. qPCR was performed as described [12] using oligonucleotides listed in Table S7.
For reactivation experiments, HeLa cells were grown to ∼50% confluence, treated with 50 ng/mL of IFN-γ in DMEM supplemented with calf serum and antibiotics for 24 h, washed extensively with PBS, trypsinized, and seeded to plates at appropriate densities that would lead to the same confluence when the cells were harvested. Transfections were performed using Lipofectamine 2000 (Invitrogen) and siRNA smart pools for Nup98, Nup107, or scrambled (Thermo Fisher) according to the manufacturer's recommendations. RNA was harvested using Trizol Reagent (Invitrogen) according to the manufacturer's recommendations.
HeLa cells were treated with IFN-γ as indicated. Cells were then trypsinized and adhered to polylysine-treated slides. Cells were fixed with formaldehyde for 15 min at 25°C and then washed with PBS+0.5% Triton X-100 several times. Slides were then treated with 0.1 M HCl on ice for 15 min, and then in 50% formamide/2× SSC for 30 min at 80°C. Fish probes were generated from BACs using FISH Tag DNA kit 488 (Invitrogen). Probes were added to coverslip and then cells were covered, sealed with rubber cement, and heated at 80°C for 4 min, followed by incubation overnight at 37°C in the dark. Slides were then washed 3 times with 2× SSC at 37°C, 3 times with 0.1× SSC, and stained with Hoechst in 0.1× SSC at 25°C for 10 min, followed by 4× SSC/0.2% Tween-20, mounted in Vectashield, sealed with nail polish, and z stacks of images were obtained using a Leica SP5 confocal microscope with a 100× objective. Measurements of the distance from FISH probe signal to nuclear periphery were made using ImageJ for individual z slices.
HeLa cells were treated with IFN-γ for 0, 6, or 24 h for initial activation or after a previous 24 h treatment followed by a 48 h rest period (reactivation). RNA samples were isolated using Trizol (Invitrogen). RNA was DNase I treated and then reverse transcribed using Superscript III (Invitrogen). For qPCR experiments, primer locations are shown in Figure S3. For microarrays, the second strand was synthesized using second strand synthesis kit (New England Biolabs). cDNA from two biological replicates was then labeled and hybridized to Agilent 128×135K arrays using human genome build hg18. Log2 ratios were generated using DNAstar Arraystar software (Roche). Averaged, normalized array data for the subset of genes that were induced ≥2 fold on average between 6 h and 24 h were organized using k means clustering by Cluster and visualized using Treeview.
Media was removed from cells, and cells were scrapped off of plates in PBS using a rubber scraper. Cells were pelleted at 1,500 rpm at 4°C. Pellets were resuspended in whole cell extract buffer (50 mM Tris, 280 mM NaCl, 0.5% NP-40, 0.2 mM EDTA, 2 mM EGTA, 10% glycerol) with DTT, sodium vanadate, and protease inhibitors. Lysates were incubated on ice for 20 min and then spun at 13,200 rpm at 4°C. Supernatant was harvested, and protein concentration was quantified using BCA assay (Pierce) and frozen in liquid nitrogen. 75 µg of each sample was separated on a 10% SDS Tris-MOPS gel, transferred to nitrocellulose, and incubated overnight with antibodies against Nup98 GAPDH in TBST+5% skim milk at 4°C. Blots were then washed twice with TBST, incubated with secondary antibody conjugated to HRP, and exposed to Enhanced Chemiluminescence reagents (Pierce) and imaged using a UVP BiospectrumAC Imaging System.
Yeast strains were harvested, fixed with methanol, and processed for microscopy as described [91].
Cells were harvested by centrifugation, washed 3 times with PBS, and adhered to polylysine-treated slides for 3 min at RT. Cells were fixed with 4% paraformaldehyde in PBS at RT for 15 min, washed twice with PBS, and then permeabilized with PBS+0.5% Triton X-100 at RT for 30 min. Cells were then blocked with PBS+3% BSA+0.1% Triton X-100 for 1 h at room temperature, and then incubated with primary antibodies (either m414 with Nup96 or m414 with Nup98) overnight at 4°C. The following day, cells were washed twice with PBS and then incubated with secondary antibody (Goat anti Rabbit Alexafluor488 or Goat anti Mouse Alexafluor 594) in blocking buffer for 2 h at room temperature, washed twice with PBS, and mounted with Vectashield. Cells were imaged on a Leica SP5 confocal microscope.
|
10.1371/journal.pgen.1006207 | Mutational Biases Drive Elevated Rates of Substitution at Regulatory Sites across Cancer Types | Disruption of gene regulation is known to play major roles in carcinogenesis and tumour progression. Here, we comprehensively characterize the mutational profiles of diverse transcription factor binding sites (TFBSs) across 1,574 completely sequenced cancer genomes encompassing 11 tumour types. We assess the relative rates and impact of the mutational burden at the binding sites of 81 transcription factors (TFs), by comparing the abundance and patterns of single base substitutions within putatively functional binding sites to control sites with matched sequence composition. There is a strong (1.43-fold) and significant excess of mutations at functional binding sites across TFs, and the mutations that accumulate in cancers are typically more disruptive than variants tolerated in extant human populations at the same sites. CTCF binding sites suffer an exceptionally high mutational load in cancer (3.31-fold excess) relative to control sites, and we demonstrate for the first time that this effect is seen in essentially all cancer types with sufficient data. The sub-set of CTCF sites involved in higher order chromatin structures has the highest mutational burden, suggesting a widespread breakdown of chromatin organization. However, we find no evidence for selection driving these distinctive patterns of mutation. The mutational load at CTCF-binding sites is substantially determined by replication timing and the mutational signature of the tumor in question, suggesting that selectively neutral processes underlie the unusual mutation patterns. Pervasive hyper-mutation within transcription factor binding sites rewires the regulatory landscape of the cancer genome, but it is dominated by mutational processes rather than selection.
| Regulatory regions of the genome are important players in cancer initiation and progression. Here, we study the patterns of mutations accumulating at short DNA segments bound by regulatory proteins (transcription factor binding sites) across many cancer types and in the human population. We find strikingly high rates of mutation at active regulatory sites across different cancers, relative to matched control sequences. This excess of mutations disrupts the binding sites of particular factors, such as CTCF, and is likely to be driven by selectively neutral processes, such as the replication timing of the genomic regions concerned. However, binding sites involved in regulatory chromatin structures suffer particularly high levels of mutation, suggesting the frequent disruption of such structures in cancers.
| Most large-scale surveys of somatic mutation in cancer have focussed on protein-coding sequences, and catalogues of genes that carry recurrent mutations already number in the hundreds [1–3], but it has long been speculated that driver mutations are likely to exist in the 98% of the genome sequence outside protein-coding exons [4]. The landscape of somatic mutation in cancer is complex, whole genome sequencing (WGS) data have revealed variable mutational spectra across cancers, some associated with particular mutagens, some with defects in DNA repair or replication fidelity, and others with unknown etiology [5]. In spite of this, cancers can be classified based upon the constellations of genomic, epigenomic and transcriptomic features they possess, indicating broad changes in regulation during tumour evolution [6].
Over the past decade, our view of transcriptional regulation in the human genome has changed radically as large consortia have profiled chromatin features across multiple cell types [7], including extensive catalogues of active regulatory elements [8]. At the same time, new technologies have allowed the exploration of chromatin conformation within nuclei, revealing maps of three-dimensional nuclear architecture, e.g. Rao et al. [9]. The most recent studies of WGS data derived from tumours have made use of these new perspectives, studying patterns of recurrent mutations in putatively functional regulatory sites [10–12]. However, accurately detecting elevated rates of mutation at relatively small numbers of regulatory sites presents major challenges for analysis. Firstly, there are wide variations in the mutational spectra experienced by different cancer types and individual tumours [2]. Secondly, the success of searches for recurrently mutated genomic regions is heavily dependent upon the number of samples available, and even large studies have proved under-powered to detect known hotspots at regulatory loci [11]. Thirdly, the reliable detection of elevated mutation at particular sites requires careful comparisons with control sites, accounting for the features associated with the sites under scrutiny, such as nucleotide composition, fine scale chromatin accessibility and replication timing [11,13]. Some studies of mutation at regulatory sites have suffered from low sample sizes per cancer type but were still able to identify a number of recurrently mutated promoters [14], for example the telomerase reverse transcriptase (TERT) gene in melanomas [15].
Predicting the functional impact of mutations occurring within noncoding regions also remains challenging. Studies of coding sequence variation in cancers have often sought evidence for variants subject to positive selection as a proxy for functional significance [3]. However, this is complicated by a widespread increase in functional (nonsynonymous) mutations, reflecting the relaxation of purifying selection in cancers relative to the germline [16]. Current strategies include the use of regions annotated as functional based upon ChIP-seq data that is restricted to a small fraction of DNA binding proteins [10], and the use of regulatory compendia scores [11]. Robust measures of selection traditionally use comparisons of putatively functional and non-functional sites (e.g. nonsynonymous and synonymous sites), but this has been lacking in studies of selection at regulatory sites in cancer.
Here, we exploit the unprecedented volumes of data produced recently by cancer WGS projects [5,17] and examine the likely functional consequences of mutations at regulatory sites. We develop novel approaches to explore the strength and directionality of selection exercised at these sites, controlling for the mutational spectra seen across cancer types and the variation in mutation rates across the human genome. Significant enrichments of somatic mutations are evident at the binding sites of several transcription factors, particularly CTCF, pointing to elevated mutation rates or suppressed surveillance and repair. These enrichments disproportionately involve mutations predicted to weaken or abolish binding at functional regulatory sites, and we find little evidence for selection preserving binding sites in cancer. However, we discover mutational foci across cancers that are predicted to alter chromatin organisation, and intriguing differences emerge in the patterns and extent of regulatory disruption seen between cancer types.
We compiled a total of 9,958,580 somatic single base substitutions across 1,574 tumour samples from 11 different tumour types; consistent with previous studies [2,5], there was a high degree of variation in substitution rates amongst tumour types (Table 1). DNase hypersensitive sites containing sequence-specific transcription factor (TF) binding motifs have previously been shown to closely match signals obtained from Chip-Seq data and can hence be used as a proxy for TF occupancy [18–20]. We established the genomic locations for constitutive DNase hypersensitive sites, active in most cell types, spanning a total of 3.92MB in the human reference genome (see Methods section). Next, we scanned the genome for matches to 118 known binding motifs of 81 transcription factors, and those motif matches inside constitutive DNase regions were labeled as “putatively functional” TFBSs. We found a total of 197,374 functional TFBSs (S1 Dataset), spanning 1.39MB of the genome and containing a total of 4,782 somatic mutations across the 11 cancer types (Table 1). For each motif matrix, we also compiled a list of control TFBSs, i.e. sequences that match a given TF binding motif, but are located outside any regions of open chromatin or genic regions, and are therefore unlikely to be bound, functionally active TFBSs (see Methods section). For each matrix, we compiled the same number of functional and control TFBSs (listed in S1 Dataset). The median distance between functional and control motifs was 10.6KB, with 90% of functional-control sites being less than 55KB apart. Functional motifs showed significantly higher conservation scores across 35 mammals than control motifs, consistent with their differing importance in biological fitness (see Methods).
Considering each TFBS matrix separately, the total number of mutations increased linearly with the length of sequence encompassed by the TFBSs as expected (S1 Fig). This was also true for control TFBSs in cancer and for high frequency germline variants, i.e. 1000 Genomes Project (1KG) polymorphisms at both functional and control TFBSs (S1 Fig). However, in the combined dataset across cancer types, we found a marked genome-wide excess of somatic mutations at functional TFBSs. This excess was seen relative to control motifs and compared with 1KG polymorphism rates (Fig 1A and 1B and S1 Table; χ2-test with Yate’s correction: χ2 = 298.2; p < 10−4). Stratifying the data by the type of binding motif, the vast majority of TFBSs (78%, 92/118 matrices) showed an excess of substitutions at functional binding sites compared to control sites (Fig 2), with 27 TFBSs showing significant enrichment for mutations (Fisher’s exact test p < 0.05), and none with significant depletion. Accordingly, putatively active TFBSs are common targets for mutations in cancer and, on average, these sites mutate at higher rates than inactive control sites.
We also observed an increase of somatic mutations at functional TFBSs compared to the regions of open chromatin that they occur within: functional TFBSs mutated at significantly higher rates than constitutively open DNase regions (S2 Fig; 0.00348 versus 0.00336 mutations bp-1; χ2 = 4.35, p < 0.05). This increase is seen in spite of the fact that constitutively DNase accessible regions suffered higher mutation rates than both the mappable portion of the genome as a whole (0.00321 mutations bp-1; S2 Fig; χ2 = 25.26, p < 10−6), and the ENCODE DNase master sites, which are regions that are accessible in any of the 125 ENCODE cell lines (0.00301 mutations bp-1; S2 Fig; χ2 = 152.89, p < 10−15). Thus, TFBSs within DNase regions suffer unusually high mutation rates, even relative to the generally elevated mutation rates seen at regions of accessible chromatin, consistent with a mutational cost of factor binding.
To quantify the deleteriousness of somatic mutations in TFBSs, we calculated the reduction in the position weight matrix (PWM) score caused by a substitution [21]. Specifically, we calculated the PWM-score for each mutated binding site and compared this to the PWM-score for the reference sequence from the human genome build (hg19), i.e. we calculated the statistic PWM-score(ALT/REF). On average, 1KG polymorphisms reduced the PWM-score to a greater extent at control sites than at functional TFBSs (Fig 1C), as expected if purifying selection in extant human populations often acts to remove deleterious variants at functional sites. In stark contrast, the PWM-score(ALT/REF) values generated by somatic mutation in cancer are statistically indistinguishable between functional and control TFBSs (Fig 1C), suggesting a widespread loss of selective constraint at these sites in cancer. Next, we calculated the ratio of the PWM-score(ALT/REF) in functional, relative to control binding sites for all 118 motifs in both cancer and 1KG; for 68 motifs, the reduction in the PWM-score was greater in cancer than in 1KG (Fig 2), with 4 motifs attaining statistical significance. Hence, in cancer, functional binding sites do not only acquire an excess of mutations, but the changes introduced by these mutations often lead to PWM-scores that are predicted to be more deleterious than substitutions tolerated as polymorphisms. Intriguingly, two TFBS motifs (ZNF263 and NRF1) had significantly increased relative PWM-scores in cancer compared to 1KG (S1 Dataset), suggesting binding is enhanced in cancers, and raising the possibility of adaptive evolution at these particular classes of binding sites in cancer.
CTCF binding sites are among the most common TFBSs in the genome (S1 Dataset), and we found the CTCF-motif to be recurrently mutated at position 9 across cancer types (Fig 3A), a pattern that was previously seen in CTCF-TFBSs identified via Chip-exo of CTCF in a colorectal cell line [10]. Note that the majority of our constitutive CTCF-TFBSs (8,795 out of 10,763) overlap with those identified by Katainen et al. [10]. The distribution of mutations within functional CTCF TFBSs in our dataset was significantly different from that of 1KG polymorphisms (Fisher’s exact test, p < 10−5; S2 Dataset), with the central nucleotide known to be constrained at the population level but highly mutated in cancer (Fig 3A) [10]. Most substitutions at position 9 of the CTCF-motif are T>G, T>C and T >A in cancer (Fig 3A), and mutations away from T at this information-rich central motif position are expected to lead to reduced binding of CTCF [22]. Overall, we observe an exceptionally high mutational burden at functional CTCF binding sites in cancer (3.31-fold excess) relative to control sites, and we demonstrate that this effect is seen across cancer types (Fig 2). This unusual accumulation of substitutions could conceivably be the result of selective processes or mutational bias during cancer evolution. In either case, the mechanisms that lead to a specific site of the motif being subject to high rates of substitution, remain elusive.
We stratified our samples into five mutational spectra (S3 Fig and S2 Table), based upon the genome-wide occurrence of substitutions in their trinucleotide context, consistent with previous studies (see Methods). Since we subdivide the data into only five signatures, a one-to-one comparison with the 21 mutational signatures of Alexandrov et al. [5] is not possible. However, we observe a similar grouping of lung adenocarcinoma samples (in mutational group 1, characterized by C>A mutations; Alexandrov et al.’s signatures 4 and 5), and observe an overrepresentation of C>T changes across most cancer samples. Interestingly, the excess of T>G/C/A mutations at position 9 of the CTCF-motif was only seen in mutational spectra 3 and 5 (S4 Fig), and it was strongest in spectrum 3 which also shows the strongest T>C signature. In contrast, tumours in spectrum 1 do not show the elevated substitution rate at position 9. Similarly, the total number of mutations in functional motifs, relative to control motifs, is not elevated in spectrum group 1, as it is for samples in spectra 2, 3 and 5 (S3 Table). Thus, the increase in mutation at CTCF binding sites is driven by mutations at position 9, which is heavily mutated in particular subsets of samples with a common mutational signature and indicative of the dominant underlying mutational process.
It has recently been shown that liver cancer is particular prone to asymmetries of A>G/T>C mutations in relation to the transcribed and untranscribed DNA strands [23], and we observe a similar genome-wide trend for the liver cancer samples (S4 Table) here. A:T nucleotides were more prone to mutate to G:C when the ‘A’ nucleotide occurred on the non-transcribed strand and the ‘T’ was on the transcribed strand. Interestingly, the same trend was also seen for the subset of functional CTCF sites that fall into transcribed genomic regions, and these sites mutated at much higher rates than the genome wide average (S4 Table); this further supports the notion that mutations at CTCF-TFBSs follow genome-wide trends in mutational bias.
CTCF has long been known to have important architectural roles in chromatin structure [24,25]. Rao et al. [9] found that CTCF binding sites delineate a hierarchy of chromatin loops (indicating peaks of Hi-C contact frequencies), and regulatory domains (median size 185KB) that compartmentalize the genome into self-interacting units. The majority of points in the genome marking the beginnings and ends of chromatin loops (loop anchor points) are bound by CTCF, and are thought to link regulatory sites to target promoters. The majority (55–75%) of loop anchor points are conserved across human cell types, and across mammals; many of these loops also demarcate the boundaries of self-interacting regulatory domains [9]. Using a sliding window approach, we found the number of functional CTCF motif instances to increase sharply at chromatin loop anchor points and domain boundaries (Fig 4A and 4C). Functional CTCF motifs were strikingly prone to mutation if they were located within chromatin loop anchor points (Fig 4B and S5 Table), with a similar (though non-significant) trend evident at domain boundaries (Fig 4D), whereas there was no significant enrichment of mutated control motifs (S5 Table). Further, position 9 of the CTCF-motif was more highly mutated when the binding site was located inside a loop anchor point. Inside loop anchor points, 204 out of 792 observed substitutions (26%) were at position 9 of the motif, compared with 15% (83/539) in functional motifs outside loop anchors, despite the motifs having very similar sequence composition inside and outside loop anchor points (S5 Fig). The mutation rate was approximately three-fold higher within CTCF sites within loop anchor points, compared to the rate observed within anchor points in general (S2 Fig; χ2 = 1242.00, p < 10−15), supporting the idea that the CTCF-motif is a hotspot of mutation within this specific chromatin context.
Given the limited numbers of mutations recorded in some tumour samples, we could not rigorously determine if CTCF motifs were highly mutated inside chromatin anchor loops across all tumour types. However, an excess of mutations inside loop anchor motifs was observed in all cancer types with a sufficiently high number of CTCF mutations, i.e. whenever the power to detect this difference in mutation rates at alpha = 0.05 was 80% or greater (S6 Fig). Thus, CTCF sites involved in higher order chromatin structures appear to suffer the highest mutational burden, and chromatin organization may be affected by this increased mutational input across several cancer types. S6 Table lists the number of CTCF-mutations for each cancer type and shows that the highest mutation rates at functional CTCF sites per individual are suffered by liver and lung tumours, with substantial mutational loads also seen for breast, pancreas and lymphoma samples. In contrast, the relatively numerous (Table 1) medulloblastoma and astrocytoma samples show orders of magnitude lower rates per individual, suggesting that different cancer types experience very different degrees of CTCF binding site disruption (S6 Table).
Using the GREAT tool [26] with default parameters, we tested for enrichments of functional annotations at genomic regions associated with mutated functional CTCF-sites. We found modest over-representation of certain functional categories, including biological processes associated with the regulation of cellular secretion, and several cancer-associated MSigDB entries, such as down-regulated genes predicting poor survival of patients with thyroid carcinoma (S7 Table).
We further explored the chromatin context of mutated TFBS instances, examining whether particular functional chromatin states were associated with the propensity of a particular TFBS to undergo mutation (see Methods). Among the 118 TFBSs tested, the mutational load of only five TFBSs (E2F1_MA0024.2; CTCF_MA0139.1, CTCFL_MA0531.1; E2F4_MA0541.1 and YY1_MA0095.2) showed an uneven distribution among chromatin states (Chi-Squared Test, p < 10−3). In each case, there was an excess of mutations in insulator regions (S3 Dataset). In particular, 16–17% of the CTCF functional binding sites allocated to the “insulator” chromatin state carried a mutation in at least one sample, whereas CTCF TFBSs in “promoter”, “enhancer” and “transcription” regions were mutated less often (5–10% of functional sites). This suggests that CTCF binding sites are particularly prone to mutation when they are involved in specific chromatin contexts. This appears to reflect variation in the rates of somatic mutation in DNAse hypersensitive sites in general, which was 0.0039 per base pair in accessible regions classified as “insulator”, but only 0.0032 in regions classified as “promoter”, “enhancer” and “transcription” (χ2 = 128.61, p < 10−15).
We used logistic regression to assess which genomic parameters were prominently associated with a high rate of substitution across the 118 TFBSs. Factors, which significantly affected the propensity of a binding site to undergo mutation in cancer, included replication timing, the identity of the TFBS matrix, the functionality (i.e. DNase status) of sites and whether sites were present at loop anchor positions (S4 Dataset). Logistic regression analysis confirmed that functional binding sites consistently mutate more often than control sites, that the positioning within loop anchor points increases a binding site’s chance of mutation, and that different binding motifs mutate at distinct rates. In addition, late replication was significantly associated with higher rates of mutation in the regression model, consistent with a general role for replication timing in the nucleotide substitution rate [27,28]. In fact, when we correct for replication timing, the difference in mutation rates between CTCF motifs inside and outside chromatin loop anchor points diminishes (S4 Dataset and S7 Fig). These CTCF binding sites might otherwise have been regarded as candidates for the apparent action of selection in cancer, given their specialized roles as well as the elevated frequencies and specific patterns of mutation observed. It is therefore striking that even for these sites mutational bias emerges as a convincing explanation for the patterns observed.
Motivated by the patterns of site-specific mutation accumulation in CTCF, we investigated the pattern of substitutions on a per-site basis for all 118 TFBSs, but found few examples of selection acting to preserve motif integrity. For example, ZBTB33, a regulator of the Wnt signaling pathway, binds to methylated 5'-CGCG-3', and showed evidence for preservation of its target TFBSs in 1KG data (Fig 3C). By contrast, in cancer, ZBTB33 binding sites were highly mutated at positions 5 and 8, reflecting the high mutational input evident at ZBTB33 control motifs (Fig 3C). The significantly elevated numbers of mutations at these motifs were accompanied by a reduced PWM-score for the ZBTB33 motif in cancer (S1 Dataset). Examination of most TFBSs suggests a similar situation, but the USF1 binding motif (MA0093.2) was a rare exception. Functional USF1 TFBSs showed a depletion of substitutions compared to flanking regions—in the 1KG polymorphism as well as the cancer dataset—but this depletion was absent at control sites (Fig 3B). In addition, mutations at functional USF1 binding sites reduced the PWM-score to a much lesser degree than control sites in cancer (Fig 2 and S1 Dataset). Due to the relatively modest number of mutations present at USF1 sites in the current data, the comparison with 1KG PWM-scores was not statistically significant, but these observations are consistent with motif preservation at USF1 binding sites in cancer. The complete dataset for each of the 118 matrixes, their controls sites, flanking regions and 1KG comparison, are provided in S7 Fig.
We found no evidence that significantly mutated binding motifs are more likely to be bound by transcription factors which have been reported to suffer recurrent protein coding sequence mutations, i.e. genes that are found in the Cancer5000 gene set of Lawrence et al. [2] (S8 Table; Fisher’s test N.S.). This suggests that mutations at TFBSs and those within coding regions have largely independent impacts on regulatory dysfunction in cancer. Further, we found little recurrence of mutations at individual functional binding sites: the most highly mutated positions inside motif instances were mutated in only five out of the 1,574 tumor samples each, at chr6:73122103, chr2:49173806 and chr2:49173798, affecting the binding motifs of CTCF/YY1 and CTCF/CTCFL, respectively. The chr6:73122103 site was also previously found to be mutated in 3.5% of colorectal cancer samples [10]. In contrast, the two most highly mutated sites across cancer genomes in protein coding sequence are a known mutational hotspot in codon 12 of the KRAS gene; these sites carried substitutions in 257 and 67 tumors, respectively. Thus, in contrast to coding sequences, where specific loci suffer detectably higher mutation rates, the mutational burden at regulatory sites requires a genome-wide perspective, encompassing many individual sites that belong to a given class of TFBS.
In spite of the broad loss of constraint seen across TFBSs in cancer, it was possible to discern differences among cancer types, even with the limitations and caveats of the current data. We found that the particular binding motifs mutated in functional, relative to control sites and 1KG polymorphisms differed markedly over different cancer types (Fig 5; complete dataset in S5 Dataset). Stratifying the data by cancer type reduces the mutation counts in each category, but suggests that lung adenoma tumours (which also possess a distinctive mutational profile; S2 Table) may accumulate more mutations at functional TFBSs compared to other cancer types, with the notable exception of CTCF binding sites. Within cancer types, we observe large variation in the numbers of mutations on a per-patient basis (Fig 5). The high rate of TFBS mutations in liver cancer is in part driven by a small number of outlier patients with exceptional biases to mutation in functional rather than control motifs (Fig 5). With larger cancer sequencing datasets it is likely that such variation among cancer types will become clearer, promising a new perspective on cancer genomics.
We have shown that functional regulatory elements suffer elevated rates of somatic mutations in cancer that based upon the accumulation of substitutions relative to matched control sites appear deleterious to regulatory protein binding. These striking patterns of mutation differ across TFBSs and cancers, and yet a high attrition of CTCF sites is a notably general feature. The unusual patterns of mutation seen at CTCF sites suggest widespread alterations to regulatory chromatin architectures across the genome, underpinned by strong mutational biases rather than selective processes. This raises the possibility that regulatory ‘driver’ mutations in cancers may arise as a byproduct of such biases superimposed upon a genome-wide relaxation of selective constraint at regulatory sites.
The strongest impact of mutation on functional CTCF sites in the current data was observed in liver cancer samples, which showed the most dramatic increase in numbers of mutations observed (S2 Table). We have shown that, by examining aggregated sites across the genome, it is possible to detect these patterns rigorously, while controlling for the influences of sequence composition and regional variation in mutation rates. However, it is important to note that these patterns will remain undiscovered by conventional approaches, most of which are based upon identifying individual genomic regions subject to recurrent mutations, and make it difficult to correct for compositional bias. This is exemplified by a recent publication describing the liver samples studied here, which assessed mutation rates within 500bp genomic windows, did not correct for compositional bias, and was therefore unable to detect the genome-wide increase in mutation rates at CTCF sites [29].
Regions of open chromatin have previously been shown to mutate at a decreased rate [13,28,30,31], presumably as such regions are more accessible to the DNA repair machinery. However, these analyses were based on sections of large, often multi-megabase regions, rather than the short binding motifs, about 10-20bp in size, examined here. Michaelson et al. [32] found DNAse I sites often to be de novo mutated in the germline, especially when the applied window sizes were small, i.e. 10 or 100bp. Recent studies [10,33–35] have suggested possible mechanisms for increased mutation rates at TFBSs, including the perturbation of lagging-strand replication at strong binding sites, and differential accessibility of binding sites to the nucleotide excision repair machinery. An emerging theme here is that there may be a general mutational burden to regulatory function, where the action of sequence specific binding to DNA interferes with normal replication, damage, surveillance and repair processes. The breadth of effects we observe genome wide, across many transcription factors and tissues of origin, suggests that these are pervasive influences on the mutagenicity of the genome. As the net effect is one of increased mutation rate specifically at functional regulatory sites, it will be important in future studies to explore the mechanistic nature of these interactions and the relative importance of replication, repair and exogenous mutagenesis to the locally elevated mutation rates.
We have shown that the mutation mediated decay of TFBSs can be observed across cancer types and binding motifs, and there appears to be no widespread purifying selection to counteract this. Among 118 motifs tested, not a single motif was significantly depleted for mutations at functional sites, relative to comparisons with control sites or population variation (1KG), suggesting that most binding sites for most known transcription factors are dispensable for tumor survival. Further, considering the per-site mutation rates within motifs, we often observe the same patterns of substitutions at control and functional sites, e.g. CpG mutations, suggesting that the accumulation of substitutions at TF binding sites is mostly driven by mutational rather than selection processes. Finally, the recurrence of mutations in functional TFBSs was two orders of magnitude lower than at sites of recurrent mutation in protein coding regions, consistent with the notion that no individual TF binding site in our dataset is likely to be a major driver of tumorigenesis. However, this does not mean that the aggregated, genome-wide impact of mutations across many TF binding sites is negligible. For example, the widespread disruption of CTCF-binding sites may have drastic consequences for the chromatin organisation and hence regulation of tumour gene expression [36], and possibly for the stable transmission of DNA in subsequent cell divisions [37]. Cancers with a strong A:T>G:C mutational signature were particularly affected by CTCF binding site mutations, and such cancers may show higher degrees of regulatory instability. Consistent with our results a recent study showed that the disruption of chromatin boundary sites may activate proto-oncogenes in T-cell acute lymphoblastic leukemia, and observed a similar excess of mutations at CTCF sites [38].
Many previous studies (e.g. [29]) have used comparisons between binding sites and their flanking regions to assess the relative somatic mutation rates at such sites. Given the inevitable differences in sequence composition between binding sites and flanks, and the large literature supporting the role of compositional bias in mutation rates [2,11], this is a challenging strategy. In addition, since TFBSs are highly clustered in the genome, the neighbouring regions of any given motif may also act as binding sites for other factors, potentially affecting flanking rates of mutation. Third, it has also recently been shown that immediately flanking regions per se may undergo increased rates of mutation [33], which is consistent with the mutational input observed at CTCF TFBSs (Fig 3). In this study, we use a metric comparing the rates of mutation in functional versus control motifs of matched length and composition, circumventing biases introduced by differences in nucleotide sequence composition of the binding site or its flanks. Nevertheless, for comparison with prior studies in S1 Dataset, we compare the number of mutations in functional and control sites seen for each binding motif, relative to their 100bp flanking regions.
One should note that our global analysis, in common with others to date, was limited by the heterogeneity of substitution rates across tumour types and by the numbers of mutations found within TF binding sites, which bounded the statistical power of our analyses; further, all p-values shown are uncorrected for multiple testing of 118 binding motifs. Thus, it was not always possible to meaningfully stratify results by mutational signature group or tissue of origin. Considering each tumour type separately, it appears that some cancer samples have a reduced proportion of mutations in functional motifs compared to control sites (Fig 5). However, the number of samples and/or the overall rate of mutation within these cancers are relatively low, which increases sampling bias. In our genome-wide pan-cancer analyses, the weaker patterns seen in these tumours is overridden by cancers such as lung adenoma and liver cancer, which show an excess of mutations at functional sites (Table 1; Fig 5). Thus, with additional cancer WGS data to explore, many new insights into the regulatory genomics of cancers should be possible.
To detect functional regulatory binding sites in the genome, we used a combination of computational prediction and experimental data: Position weight matrices for 118 transcription factor binding motifs (85 from ensemble Biomart at http://grch37.ensembl.org/biomart/martview/9620562a1888b791f43eb69ee9adcaf0 and 33 additional motifs from Jaspar [39] at http://jaspar.genereg.net/) were used as input to FIMO (of the MEME suite [40]), to find predicted motif matches in the genome. The maximum p-value for a motif match was set as the default (p < 4.4e-05); if more than 300,000 motif instances were found, the motifs with the largest p-values were iteratively dropped. We intersected these motif matches with experimentally defined open chromatin regions: UCSC DNase master sites were downloaded from the UCSC genome browser (http://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg19&g=wgEncodeAwgDnaseMasterSites), and DNAse footprints came from Thurman et al. [8], with footprints calculated as in Neph et al. [19]. In order to avoid the erroneous classification of binding sites as active in tumour tissue, we only considered putative binding sites in constitutively open chromatin, i.e. in UCSC chromatin regions that were DNAseI accessible in at least 113 out of 125 ENCODE cell types, or within DNAse footprints that were found in at least 39 out of 41 tissues. We conservatively limited our analysis to these putatively functional binding sites in constitutively DNAseI hypersensitive sites, and accordingly, expect a relative underrepresentation of tissue-specific binding sites in our dataset. The aim was to enrich our ‘functional’ sites for active TF binding relative to control sites. Note that, due to partial positional overlap of motifs, 44% (2,123 out of 4,782) of the somatic substitutions found within functional sites affected more than one TFBS, supporting the functional significance of these sites. As control motifs, we chose FIMO motif matches that were located outside open chromatin regions/DNAseI sites in any tissue of the ENCODE and Thurman datasets; in addition, control motifs had to be in the mappable regions of the genome (i.e. outside DUKE and Dac excluded regions [41]) and more than 2kb upstream of known genes. To minimize the difference in the mutation rate among functional and control TFBSs, we position matched each functional motif instance with a nearest control motif, choosing, for each functional TFBS, the closest motif from the pool of possible control sites. Functional and controls TFBSs both had high and comparable uniqueness scores (S9 Table), suggesting that mutations can be detected in both regions. We note that the GERP conservation score [42] across whole genome alignments of 35 mammals (http://genome.ucsc.edu/) is, on average, higher for functional TFBSs than for control motifs (S9 Fig); this is expected if functional motifs are under purifying selection, and has no impact on our analysis. Functional motifs match the input position weight matrices slightly better than control motifs, with median PWM-scores of 8.73 and 8.33, respectively (S9 Fig). However, since we measure the reduction in score relative to the reference allele, this should have negligible consequences for our analysis, and, consistent with this, the reduction in score is lower for functional TFBSs in the 1KG data, even though functional motifs start off with slightly higher scores (see Results section).
We downloaded whole genome mutation annotation format (maf) files for 11 tumour types from public data resources: 507 samples came from Alexandrov et al. [5], and a further 1,067 non-embargoed samples (free of all publication moratoria) came from Release_17 of the ICGC [17], including the projects LINC-JP, BRCA-UK, LIRI-JP, CLLE-ES, MALY-DE, PBCA-DE, EOPC-DE, PRAD-CA, PRAD-UK, PACA-AU, LICA-FR and PACA-CA. The maf files had previously been filtered for germline variants, i.e. they only included somatic mutations. 1KG polymorphism data (vcf files) were from EBI (ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/). Somatic point mutations and 1KG common SNPs with a frequency of >5% were intersected with our set of functional binding sites and control motif sites.
PWM-scores [21] were calculated for each motif site that carried somatic substitutions or polymorphisms, and this score was compared to the reference allele, i.e. the motif instance in the human reference assembly (hg19). The relative reduction or increase in PWM-score for each binding site was calculated as PWM-score(ALT)/ PWM-score(REF), thereby controlling for variation in information content between motifs.
To assess the impact of mutations in cancer with regards to the number of mutations per motif site and the predicted change in PWM-score, we divided the data into four separate categories: 1) somatic mutations at functional sites; 2) 1KG polymorphisms at functional sites; 3) somatic mutations at control sites; 4) 1KG polymorphisms at control sites. Variants with a frequency > 5% in the 1KG dataset may be neutral, advantageous or mildly deleterious, but are unlikely overall to be under strong purifying selection. Accordingly, the level of 1KG polymorphism at functional sites, relative to control sites for the same motif, gives an indication of the level of constraint for a given class of binding sites [43] and can be compared to the patterns of mutation seen in cancers.
The significance of enrichment or depletion of mutations inside functional TFBSs in cancer was assessed using Fisher’s exact test for mutation counts in the four classes of sites: functional and control sites in cancer and 1KG, respectively. To assign a p-value to the reduction in the PWM-score, we used the methods of Price and Bonett [44] and calculated, for each binding motif, the confidence intervals for the ratios of median relative PWM-scores in cancer (functional/control) and 1KG (functional/control) separately, and assessed the extent to which they overlapped.
Aggregate mutation/polymorphism counts were produced for each binding motif and sample; the shape of the distribution between cancer and 1KG samples (visualized as barplots in S8 Fig) was compared using Fisher’s exact test.
Mutational spectra were calculated by counting the number of each of the 96 possible substitution types for each cancer sample, and dividing this vector by the expected number of substitutions, which was based on the trinucleotide count in the human reference sequence and assuming that a substitution from any nucleotide to any other is equally likely [5]. The Manhattan distance between each sample-specific mutational spectrum (scaled to a total sum of one) was calculated, with a dendrogram based on hierarchical clustering to relate samples. To avoid errors due to sampling of low mutation counts, the dendrogram shown in S3 Fig only included samples with at least 7000 mutations. Samples were allocated to five different spectra based on their clustering in the dendrogram.
We divided CTCF-binding regions of the genome, which also overlap transcribed regions, into two groups, based on whether DNA is transcribed from the reference strand or its complement according to the ENSEMBL annotation of hg19. A total of 44,072 and 40,507 basepairs overlap functional CTCF motifs and are transcribed from the reference and complement strands respectively, excluding sites that are transcribed bi-directionally. Next, we counted the number of A>G and T>C changes at CTCF sites in liver cancers; we assessed whether the reference “A” nucleotide was on the transcribed or the non-transcribed strand of DNA (with its complement, “T”, being on the other strand), and calculated the strand bias of these mutation classes as in Haradhvala et al. [23]. We repeated the same procedure for all liver somatic mutations that fell into unidirectionally transcribed regions of the genome (612MB and 587MB of DNA for reference and complement strands respectively).
Chromatin loop anchor positions and chromatin domain boundaries based on the Hi-C data of GM12878 (the cell line with the highest resolution of 950bp from Rao et al. [9]) were obtained from NCBI GEO (Accession GSE63525). Across domain boundaries and loop anchor points reported by Rao et al. [9], we counted the number of somatic mutations and the number of CTCF motif instances. We do not have Hi-C data for the tumour samples in this study; however, to assess if an increase in mutations at CTCF-TFBSs inside loop anchor points is seen across different cell lines, we repeated the analysis with loop anchor points called in IMR90, HMEC, NHEK, K562, HUVEC, HeLa, and KBM7 cell lines [9].
ChromHMM tracks [45] were downloaded from the UCSC Genome Bioinformatics site (http://genome.ucsc.edu/) for GM12878, H1-hESC and K562 cell lines. These datasets were intersected with the genomic location of all functional motifs, classifying each motif into falling into one of six chromatin “colors”, i.e. “promoter” (red), “enhancer” (yellow), “insulator” (blue), “transcription” (green), “repressed” (grey) and “low signal” (white). For each Matrix, we counted the number of mutated and intact functional binding sites, using a Chi-Squared test to assess if different chromatin states showed different propensities for mutation.
A logistic regression model was constructed, modeling the binary outcome variable “mutated/not mutated” in the combined cancer dataset; this variable describes if a given binding site at a particular genomic location is mutated in any of the cancer samples. As predictor variables, we used the replication timing data of Chen et al. [46], “Matrix” as a factor with 118 different levels representing the different TFBS motifs included, a binary “Functionality” (i.e. functional vs. control) variable and the binary classifier of whether the binding motif was inside or outside a chromatin loop anchor point [9]. The Wald test was used to test for the significance of individual predictor variables within the model. The fraction of predicted mutated motif positions was calculated for each functional matrix inside or outside loop anchors respectively, keeping replication time constant.
|
10.1371/journal.ppat.1005889 | Microbiome Composition and Function Drives Wound-Healing Impairment in the Female Genital Tract | The mechanism(s) by which bacterial communities impact susceptibility to infectious diseases, such as HIV, and maintain female genital tract (FGT) health are poorly understood. Evaluation of FGT bacteria has predominantly been limited to studies of species abundance, but not bacterial function. We therefore sought to examine the relationship of bacterial community composition and function with mucosal epithelial barrier health in the context of bacterial vaginosis (BV) using metaproteomic, metagenomic, and in vitro approaches. We found highly diverse bacterial communities dominated by Gardnerella vaginalis associated with host epithelial barrier disruption and enhanced immune activation, and low diversity communities dominated by Lactobacillus species that associated with lower Nugent scores, reduced pH, and expression of host mucosal proteins important for maintaining epithelial integrity. Importantly, proteomic signatures of disrupted epithelial integrity associated with G. vaginalis-dominated communities in the absence of clinical BV diagnosis. Because traditional clinical assessments did not capture this, it likely represents a larger underrepresented phenomenon in populations with high prevalence of G. vaginalis. We finally demonstrated that soluble products derived from G. vaginalis inhibited wound healing, while those derived from L. iners did not, providing insight into functional mechanisms by which FGT bacterial communities affect epithelial barrier integrity.
| The female genital tract (FGT) is a key mucosal surface in the context of HIV transmission. Lactobacillus species are beneficial to the FGT, while Garderella vaginalis and other anaerobic bacteria are detrimental. Bacterial vaginosis (BV) is an inflammatory condition characterized by an outgrowth of G. vaginalis and other anaerobes, which is linked to increased HIV acquisition rates. However, the mechanism behind this remains unknown. Here, we used a novel proteomic approach to simultaneously evaluate host and bacterial functions in the FGT. We found that women with G. vaginalis-dominated FGT bacterial communities always displayed markers of decreased epithelial barrier integrity, and decreased wound healing capacity. We also demonstrated that the abundance of proteins from G. vaginalis associated with these signatures of disrupted epithelial integrity. Finally, we showed that products derived from G. vaginalis prevented healing of wounded cell monolayers while products derived from L. iners maintained the ability of the cell monolayers to close wounds. This study provides novel mechanistic insight into the link between BV and increased HIV acquisition rates.
| Mucosal surfaces exposed to the external environment contain distinct bacterial communities that exist in relationship with the host and can contribute to health and functioning. These bacterial communities have been linked to several human diseases and overall health [1], and can vary between individuals, but also over time within the same person [2]. In the female genital tract (FGT), colonization by Lactobacillus species and other lactate-producing bacteria helps to inhibit colonization by pathogenic bacteria [3]. However, colonization by more diverse communities of anaerobic bacteria, notably Gardnerella vaginalis, is common [4], and often associated with the development of bacterial vaginosis (BV) [5]. BV is highly prevalent, affecting 4–58% of women globally; some areas, such as sub-Saharan Africa have rates as high as 55% [6]. BV is associated with significant health consequences, including pre-term birth, post-partum endometriosis, pelvic inflammatory disease, upper reproductive tract infections, and increased susceptibility to sexually transmitted infections (STI’s) [7, 8], with HIV being highly significant [9, 10]. Indeed, a recent meta-analysis linked BV to a 60% increase in HIV acquisition rates [11]. However, while these relationships between microbial composition and vaginal health have been described epidemiologically, there is limited understanding about the mechanisms underlying the impact of bacterial dysbiosis on the vaginal mucosa.
Maintenance of the mucosal barrier is critical for preventing invading microorganisms, including HIV, from penetrating into tissues and entering circulation [12]. Bacterial diversity in the FGT has been strongly associated with negative consequences for FGT mucosa. Highly diverse communities dominated by G. vaginalis and Prevotella are associated with upregulated expression of Toll-like Receptor (TLR) and NFkB pathways, leading to increased pro-inflammatory cytokine concentrations and activation of immune cells [13]. While it is widely appreciated that BV is associated with inflammation, the mechanism that elicits this inflammation or the bacterial proteins associated with inflammation remain unresolved [14], which may partly explain the limited effectiveness of antimicrobial treatment for BV [15–17]. Bacterial metabolites including hydrogen peroxide, antimicrobial peptides, and acids that reduce the FGT pH have been proposed to have an important impact in sustaining mucosal health [3]. Furthermore, the integrity of mucosal epithelial surfaces has been shown to depend on bacterial community composition in other diseases [18], and has been proposed to be important in the FGT during bacterial dysbiosis [19], but this has not been extensively studied. Each of these factors likely impact disease susceptibility independently, and a Lactobacillus-dominant microbiota likely contributes to many of these factors to maintain the function of the healthy FGT and inhibit infections. Taken together these studies suggest that host-microbe interactions are key to understanding negative consequences on vaginal health, yet this interaction remains poorly defined in human cohorts [20].
We sought to better understand the relationship between mucosal health and bacterial diversity using a combination of metaproteomics and metagenomics, which to our knowledge represents the first attempt at integrating these approaches to study the FGT. Indeed the functional diversity of the bacterial proteome, and how this relates to FGT health and inflammation has not been assessed comprehensively, and has largely been limited to 16S rRNA gene sequencing. Thus, we hypothesized that bacterial protein factors can influence FGT mucosal health and affect disease susceptibility. Here we characterized FGT bacterial communities in two distinct human cohorts, longitudinally and cross-sectionally, in asymptomatic and symptomatic women with BV, uncovering bacterial-host interactions leading to wound healing impairment.
Cervicovaginal secretion samples from two cohorts of women were evaluated to understand the mucosal environment associated with bacterial dysbiosis. We first assessed mucosal changes in women at BV+ or BV- time points (Cohort 1, n = 10), through a combination of mass spectrometry (MS) and 16S rRNA gene sequencing. MS analysis identified 1123 unique proteins, including 434 human and 689 bacterial proteins from 64 species. To assess the diversity of the bacterial proteome, we quantified the relative proteome load of each bacterial genus in each sample by summing the total number of protein spectral counts assigned to each genus, an approach previously shown to directly correlate with colony-forming units [21]. We clustered the bacterial proteomes from the twenty samples using unsupervised hierarchical clustering. Two major bacterial proteomes were identified, dominated by either Lactobacillus iners (Group 1, or G1) or Gardnerella vaginalis (Group 2, or G2) (Fig 1A, species-level taxonomy shown in S1A Fig), which were used for downstream comparisons. In G1, L. iners proteins accounted for 87–100% of the total protein load while in G2, G. vaginalis proteins accounted for 48–96%. Compared to those in G1, the bacterial proteomes in G2 displayed significantly higher species diversity (S1C Fig). G2 profiles also had higher overall bacterial protein load when normalized to the total protein content (0.34 log10, +2.2 fold higher; S1E Fig). L. iners dominated the FGT bacterial proteome of eight of the 10 patients from Cohort 1 at the time point without clinically diagnosed BV, but not the remaining two. Patient “10“, at the time point without BV, displayed a high abundance of G. vaginalis with a lower abundance of L. iners, and Patient “6”at the time point without BV had high levels of Lactococcus lactis and Streptococcus mitis. In contrast, all samples taken during episodes of BV had high abundances of proteins from G. vaginalis, Prevotella spp., Streptococcus mitis, Escherichia coli, and Atopobium vaginae, which have been previously identified as part of BV-associated bacterial communities [7].
Bacterial community composition for G1 and G2 was confirmed by 16S rRNA gene sequencing (Fig 1B). According to 16S rRNA gene sequencing, G1 communities were dominated by Lactobacillus spp. (26–99% of the total community), and G2 communities were dominated by Gardnerella, but at lower proportions than were detected by MS (17–66% of the total community). As with MS, bacterial genera detected in BV-positive individuals by 16S rRNA gene sequencing included Sneathia, Prevotella, Atopobium, Megasphera, and others. 16S rRNA gene sequencing also detected greater bacterial diversity in the G2 samples compared to G1 (S1C Fig). Several species detected by 16S and not by MS included Leptotrichia, Fastidiosipila, Shuttleworthia, and Aerococcus. Overall, this demonstrates significant heterogeneity in the structure of FGT bacterial communities between clinically defined BV and asymptomatic time points, that G. vaginalis and other anaerobes associate with BV, and that specific species dominate the bacterial proteome landscape in mucosal secretions.
Mucosal samples from a separate group of 31 women from North America (Cohort 2) were analyzed to further evaluate associations between FGT bacterial proteome diversity and BV. MS analysis of Cohort 2 samples showed similar trends to that of Cohort 1 (Fig 1C, species-level taxonomy shown in S1B Fig), including Lactobacillus spp.-dominant (G1) and G. vaginalis-dominant (G2) communities. A wider distribution of lactobacilli including L. iners, L. crispatus, and L. jensenii was observed in G1 in Cohort 2 than Cohort 1. Varying abundances of other BV-associated bacteria including Prevotella spp., Atopobium vaginae, Mobiliuncus mulieris, and Sneathia sp. were also observed. In Cohort 2, there was no difference in the species diversity of G1 compared to G2 (S1D Fig). While all 7 women with BV clustered into G2, 46% of participants demonstrated a G. vaginalis-dominated proteome despite a lack of clinical BV diagnosis, consistent with the observation that not all women have Lactobacillus-dominant FGT microflora despite low Nugent scores. Also similar to Cohort 1, G2 in Cohort 2 had higher overall microbial proteome burden than G1 (1.5-fold higher; S1F Fig), indicating further changes in bacterial community function. This agrees with other studies showing Lactobacillus dominance varies between 37–90% of women, with greater diversity and variation in African women [5, 13, 22].
As bacterial diversity has been associated with other biological variables, such as concurrent STI’s [23] and hormonal contraceptive usage [10], we compared clinical characteristics between Lactobacillus and Gardnerella-dominant groups (Cohort 1-Table 1; Cohort 2-Table 2). With the exception of BV status, we found no differences between G1 and G2 with respect to age, contraceptive use, antimicrobial usage, last menstrual period, detectible STI’s, or sexual practices in either cohort. There were differences in Amsel’s criteria collected from Cohort 2 (S1 Table) between G1 and G2, where vaginal pH, clue and white blood cell presence was higher in women with G2 bacterial proteome profiles, in agreement with clinical BV status. Overall, there was no evidence to support that vaginal bacterial profiles were related to exogenous hormonal contraceptive use, the menstrual cycle, sexual behaviors, or concurrent STI’s in these cohorts.
As the functional diversity of FGT-resident bacteria remained undefined we characterized the major bacterial pathways present in G1 and G2 profiles (Fig 2). Major functional categories represented in either group included transport and catabolism (G1 average 19.5%; G2 average 12.6%), carbohydrate metabolism (1 average 14.5%; G2 average 15.7%), as well as nucleotide and amino acid metabolism (G1 average 2.6%/1.8%; G2 average 2.0%/1.4%).
However, unique functional signatures were observed between G1 and G2 FGT bacterial communities. Across both cohorts, the G1 group showed significant enrichment of proteins involved in transport and catabolism (6.9% higher), energy metabolism (5.6% higher), and folding, sorting, and degradation (4.9% higher), while G2 was highly significantly enriched in membrane transport functions (22% higher). Twelve bacterial proteins were significantly differentially abundant after multiple comparison correction between G1 and G2 in Cohort 1 (S2 Table). Proteins enriched in G1 mostly belonged to L. iners proteins and were involved in homolactic fermentation of carbohydrates including glyceraldehyde-3-phosphate dehydrogenase (GAP-DH), pyruvate kinase (PK), and lactate dehydrogenase (LDH). Proteins enriched in G2 were all G. vaginalis proteins and included a MalE-type ABC sugar transport system periplasmic component (MAL-E ABC) and an alpha-1,4 glucan phosphorylase, an enzyme that degrades starch and glycogen, suggesting that G. vaginalis directs its metabolism towards liberation and uptake of extracellular saccharides. Although many of these proteins were also differentially abundant between G1 and G2 in Cohort 2 they did not pass multiple comparison correction. Overall this shows that ‘core’ functional pathways necessary to host-associated bacterial life within the FGT include carbohydrate, amino acid, and translational machinery, while perturbations to membrane transport and carbohydrate catabolism are likely important for pathogenic states.
Bacterial dysbiosis impacts HIV acquisition risk [10, 11, 24], reproductive health [7], and mucosal cellular activation [13], but the effect on the FGT is not well defined. Our analysis revealed that 69/434 (15.8%, 15 passing 5% FDR) and 64/434 (14.7%, 19 passing 5% FDR) host proteins were significantly differentially abundant between G1 and G2 profiles in Cohort 1 and 2, respectively. For Cohort 2, comparison based on bacterial groups rather than Nugent score criteria yielded greater host proteome differences, statistically (9.2% vs. 15.8%, P<0.05), and in magnitude (5 vs. 6 Log2 Fold Change; S1G/S1H Fig), suggesting that bacterial community composition, rather than clinical BV criteria, more accurately classifies mucosal inflammation. This comparison was not possible for Cohort 1, as all G2 profiles had clinically defined BV. Hierarchical cluster analysis revealed that longitudinal changes from G1 to G2 profiles in Cohort 1 were clearly distinguishable by two major branches of host proteins (S2A/S2B Fig). Proteins more abundant in G1 (Branch 1) associated with epidermis development and the cornified envelope, whereas G2 (Branch 2) showed increased factors involved in cytoskeletal-binding, threonine proteases involved in proteasome activity, as well as vesicular components and the melanosome. Many of these included S100 proteins and innate immune factors, important for antimicrobial defense based on gene ontology (DMBT1, CADH1, S10A7, EFHD2, S10AB, S10A6, TGM3, K2C1 S10A2). Similarly, in Cohort 2, hierarchical cluster analysis showed that proteins more abundant in G1 (Branch 1) associated with epidermis development, structural molecular activity, and the cornified envelope, while proteins elevated in G2 (Branch 2) also included ectoderm development and differentiation, although were related to cytoskeletal activity (S2C/S2D Fig). Many of these are important for leukocyte-mediated immunity and wounding responses based on their gene ontology (A1AT, IC1, GELS, CO3, PEBP1, PRDX1, PRDX2, CO4A, ANXA8).
Seventeen proteins were differentially abundant across both cohorts (Fig 3A). Host proteins more abundant in G2 profiles included apoptotic regulators (PRDX, NDKB, CADH1) and leukocyte migration factors (PLST), while G1 profiles showed increased keratinization, epidermis development, and cornified envelope (INVO, SPR1A) factors (Fig 3B). Of particular interest, the abundances of INVO and SPR1A were 14.7 and 7.2-fold lower in women G2 microbial profiles Cohorts 1 and 2, respectively. In Cohort 2, INVO and SPR1A were lower for women with G2 microbial profiles even if they had not been clinically diagnosed for BV (Fig 3C). These proteins are known to act as scaffolding for epidermal layers and are important for proper barrier function [25], and immunohistochemical analysis confirmed the presence of INVO and SPR1A in cervical and vaginal tissues, where they strongly associated with the squamous epithelium and stratum corneum in healthy FGT tissue (S3 Fig). Collectively these data show an association of heightened immune activation, apoptosis, and decreased epithelial barrier function in women with G. vaginalis-dominated bacterial profiles and that these effects are evident in G. vaginalis-dominated communities in the absence of clinical diagnosis.
Due to the strong association of epithelial development pathways with different bacterial groups, we compared cornified envelope factors INVO and SPR1A to bacterial proteins. Nineteen bacterial proteins had strong associations in at least one comparison against either INVO or SPR1A after correcting for multiple comparisons. Proteins from L. iners that positively correlated with INVO and SPR1A were involved in Catabolism and Energy Metabolism pathways, including glycolysis and homolactic fermentation of sugars (Embden-Meyerhoff-Parnas (EMP) pathway) (Fig 4A/4B). These included a putative fructose 1,6-bisphosphate aldolase (PFBA), PK, GAP-DH, and LDH, as well as a ferritin-like protein (FLP), which is important for sequestering excess iron and preventing oxidative damage [26]. Bacterial proteins that negatively correlated with INVO and SPR1A belonged to alternate sugar metabolism pathways (phosphoketolase pathway), transport functions, and amino acid catabolism. The majority of these belonged to G. vaginalis (Fig 4C/4D), including D-xylulose 5-phosphate/D-fructose 6-phosphate phosphoketolase (XFBP), a putative sugar-binding secreted protein (P-SBSP), MAL-E ABC, an extracellular solute binding protein (ESBP), and glycine oxidase (GOx). A membrane protein from Prevotella sp., sharing sequence homology with SusD-like (starch-binding) protein, was also negatively associated (S4 Fig). Many associations with vaginal pH were also observed, including negative associations with enzymes from L. iners (PK, LDH, elongation factor tu, and FLP) (S5A Fig), and positive associations with enzymes from G. vaginalis, including GOx, which catalyzes the conversion of glycine into glyoxalate, ammonia, and hydrogen peroxide (S5B Fig). Therefore a clear relationship between metabolic function, epithelial barrier protein levels, and vaginal pH was observed, demonstrating that these microbial pathways may be an important component of mucosal barrier disruption and vaginal health.
The association of bacterial communities with barrier integrity proteins led us to hypothesize that wound-healing capacity may be supported or inhibited by specific bacterial species and/or their products. We thus performed a classical wound-healing assay wherein we cultivated relevant cervical cell line (HeLa CCL-2) in the presence of supernatants derived from cultures of L. iners or G. vaginalis. Prior to adding culture supernatants, a wound was induced by scratching HeLa cell monolayers. Incubation of scratched monolayers with L. iners culture supernatant did not alter wound healing compared to the control incubations. However, incubation with G. vaginalis culture supernatants significantly reduced wound healing after 24 hours compared to both the control and L. iners conditions (Fig 5). These results confirm a relationship between soluble compounds produced by the major bacterial species of the G1 and G2 profiles and wound healing capacity. This implicates these species as important components or drivers of epithelial barrier repair, maintenance, and disruption in the FGT.
In this study, we demonstrated a novel metaproteomic approach to simultaneously assess bacterial diversity, abundance, and function, along with host barrier and inflammation processes, providing mechanistic insight relevant to women’s health. We described distinct vaginal bacterial proteome profiles that were dominated by Lactobacillus spp. (G1) or G. vaginalis (G2), where the latter associated with BV, increased community diversity, and significant divergence from normal metabolic function. We next demonstrated that bacterial functional profiles were significantly associated with cornified envelope factors in the FGT, and this was affected even in the absence of clinical diagnosis. Finally, we found that predominant species identified in this study, specifically G. vaginalis and L. iners, generate soluble products that disrupt or maintain the ability of cervical epithelial cells to repair and close wounds. Therefore, impaired wound healing is a potential mechanism by which key bacterial species may impact mucosal barrier function and therefore disease and/or HIV/STI infection risk. The association of vaginal inflammation and inflammatory vaginal bacteria with HIV susceptibility indicates that targeting this mechanism may lead to novel prevention strategies for HIV.
While increased diversity of bacterial communities has been linked to better mucosal functioning in the gut [27, 28], low-diversity bacterial communities are beneficial for the FGT [5], where increased diversity is strongly associated with BV [7]. Consistent with previous observations [5], many of the women in Cohort 2 with G. vaginalis-dominated communities were asymptomatic for BV (61%), further supporting the fact that Nugent score is underestimating the extent of non-Lactobacillus dominant communities. However, the effects on host epithelial pathways, including decreased integrity and increased inflammatory pathways were still evident in the absence of clinical diagnosis. Lactobacillus spp. and G. vaginalis proteins comprised more of the soluble proteome load than might be inferred from 16S rRNA gene sequencing, suggesting that these bacteria dominate the metabolic landscape of the FGT.
Metagenomic studies of the human microbiome have shown that core metabolic function is less variable than the community composition [2]. In agreement with this, we observed that the majority of assigned protein functions did not vary significantly, which likely represent core metabolic functions. However, some functions varied between G1 and G2, including increased carbohydrate metabolism, energy production, and folding/sorting functions in G1 to enhanced membrane transport and secretion of extracellular products in G2, with L. iners and G. vaginalis dominating these key functions. The increased abundances of enzymes important for sugar transport and starch and glycogen catabolism in G2 suggest that G. vaginalis may outcompete Lactobacillus spp. for the uptake of carbohydrate substrates. This agrees with a recent study showing that women with BV have significant metabolite alterations in cervicovaginal mucous, including lower levels of carbohydrates, amino acids, and lactate, accompanied by increased levels of amino acid catabolites and polyamines [29]. Overall, this demonstrates that increased bacterial diversity is associated with changes in key metabolic pathways, which allows for better understanding of dysbiosis in the FGT.
We found that G1 profiles from both cohorts strongly associated with cornified envelope factors, especially INVO and SPR1A, which are expressed in the upper layers of the vaginal and cervical epithelia, and aid in maintaining epithelial integrity. Our group has previously reported that increased levels of cervocovaginal CD4+ T cells associated with lower levels of cornified envelope factors [30], demonstrating the important link between vaginal epithelial integrity and HIV acquisition risk. G1 profiles were also associated with higher levels of antimicrobial peptides, such as dermcidin, which is important for host defense against microorganisms [31]. In comparison, the G2 bacterial profiles correlated with lower cornified envelope and epithelial barrier factors, increased cytoskeletal elements important for cell migration, and increased proteasome factors. This agrees with other studies showing that BV associates with activation of innate immune and inflammation pathways in the FGT, including increased complement [32], proteasome levels [33], and pro-inflammatory cytokines and activated CD4+ T-cells [13]. Importantly, G2 bacterial proteome profiles associated with decreased abundances of INVO and SPR1A regardless of clinical BV status. This finding demonstrates that current methods used to diagnose BV likely underestimate the true extent of bacterial dysbiosis on mucosal barrier function the FGT, as the Nugent Scores were poor predictors of BV, especially for Cohort 2. Thus, new methods to detect and treat G. vaginalis in the FGT could aid in reducing HIV acquisition risk by promoting mucosal and epithelial barrier integrity, and reduced inflammation.
Catabolic enzymes involved in homolactic fermentation of glucose from Lactobacillus, such as L-lactate dehydrogenase, correlated with higher epithelial barrier proteins, while membrane transporters, extracellular proteins, and alternate routes of carbohydrate metabolism (heterolactic fermentative or phosphoketolase pathways) from G. vaginalis were negatively correlated. In addition, GOx, was strongly correlated with increased vaginal pH, implicating a role of this enzyme in altered vaginal pH during dysbiosis. To our knowledge this is the first time these bacterial enzymes have been associated with epithelial disruption signatures and vaginal pH. Collectively, this shows a relationship between bacterial community structure, metabolic function, disruption of epithelial proteins important for barrier integrity, and overall vaginal health.
We also demonstrated that G. vaginalis culture supernatants inhibited healing of scratched HeLa cell monolayers while, L. iners culture supernatants maintained effective wound healing. Based on these data, G. vaginalis is likely an important component or a potential driver of subverting the wound healing process. While acknowledging that HeLa cell monolayers do not completely recapitulate the squamous epithelium or immune environment of the FGT, this nevertheless supports would healing as an underlying mechanism. Taken collectively, and considering the metaproteomic, metagenomic, and in vitro models, these data suggest that G. vaginalis releases a variety of extracellular products in the vaginal compartment that aid in uptake for nutrients, alter the vaginal microenvironment, contribute to innate immune activation, and prevent healing of the epithelial barrier. Future studies to identify exact protein pathways involved, how they may be altered, and more advance animal and engineered tissue models would help better decipher these host-microbiome interactions.
It is important to compare discuss the benefits and limitations of metaproteomics compared to 16S rRNA-based techniques to characterize microbial communities in the vaginal compartment. Both techniques are quantitative and spectral counts by MS have been shown to correlate directly to colony-forming units [21]. An advantage of 16S over MS is greater resolution of the overall community structure, and while we showed high sensitivity to identify species that were at 0.1% of the population by MS, 16S captured more overall bacterial species. It is likely that the larger dynamic range of the proteome over the genome is a large contributing factor to this observation. Both 16S and MS rely on curated databases to identify species and are subjected to this same limitation in availability and extensiveness of libraries. While databases for 16S rRNA genes are likely more comprehensive, proteomic libraries are growing and becoming more available. MS is advantageous in that it can provide direct species-level identification, which is not achievable through high-throughput 16S rRNA gene sequencing methods. Furthermore, metaproteomic analysis reveals bacterial functional and metabolic activity, which is not provided by 16S-based approaches. Prior studies have attempted to alleviate this using MS to correlate metabolite abundances with species abundances [34], through metagenomic studies [35], or by employing computational methods to estimate bacterial community functional capacity based on 16S rRNA gene signatures [36], but nevertheless represent indirect methods to evaluate bacterial community functionality. While 16S rRNA gene sequencing is a popular and well-validated method for studying microbial communities, the use of metaproteomic approaches provides complimentary and invaluable data on community structure, function, and host inflammation to better study host-bacterial relationships.
Our data provide novel mechanistic insight of how dysbiosis of vaginal bacterial communities may directly increase host susceptibility to infection through the disruption of epithelial barriers, inhibition of wound repair, and induction of inflammation. In the context of HIV transmission, inhibition of wound repair is under studied and may represent underlying mechanisms in other risk factors for HIV, including hormonal contraceptive usage, intravaginal practices, and other STI’s. These pathways may also impact the effectiveness or responsiveness to mucosa-targeted prevention technologies for other infections, such as microbicides or vaccines for HIV. In summary, this study delineated functional configurations of microbial communities that impact vaginal health during BV, providing new information on host-bacterial interactions, enabling future experiments to probe host-microbe relationships in the FGT that could have important implications for women’s health.
All women who participated in this study provided written informed consent. The studies were approved by the University of Washington Human Subjects Review Committee, the Kenya Medical Research Institute (KEMRI), Human Subjects Committee of the University of Illinois at Chicago, and the Research Ethics Board of the University of Manitoba.
Vaginal swabs were eluted with 2 x 250ul washes in PBS (pH 7.0). Swab eluates (Cohort 1) or CVL samples (Cohort 2) were then centrifuged in SpinX tubes with a bonded fritted bottom (Corning, Corning, NY), and protein content determined by BCA assay (Novagen, Bilerica, MA). Proteins were then denatured, reduced, alkylated, digested into peptides, and prepared for mass spectrometry as described previously [38]. Detailed methods for this process are available in S1 Methods.
Briefly, peptide samples were injected into a nano-flow LC system (Easy nLC, Thermo Fisher) connected inline to a LTQ Orbitrap Velos (Thermo Fisher) mass spectrometer, and analyzed in a label-free manner as described previously [38]. Peptide identity searching was performed with Mascot v2.4.0 (Matrix Science) against a manually curated database comprised of the SwissProt Human & Bacteria (June 2015) and UniProtKB/Trembl All Bacteria databases (August 2015). A decoy database was included to determine the rate of false discovery. Protein identifications were confirmed using Scaffold (v 4.4.1, Proteome Software) with confidence thresholds set at 95% protein identification confidence, requiring at least 2 unique peptides and 80% peptide identification confidence. A combination of label-free methods was used for protein quantitation: spectral counting (for microbial proteins and bacterial diversity clustering, see below) and area-under-the-curve quantitation (Progenesis LC-MS software (v4.0, Nonlinear Dynamics)). Criteria for assigning presence of microbial proteins included those that had at least 1 peptide in one sample, and at least 2 peptides per protein across all samples. These parameters resulted in a false discovery rate below 3.1% based on the search results run against Mascot’s generated decoy database. For the latter, only proteins that had an average co-variance of <25% (575 proteins), as determined through measurements of standard reference sample run at 10 sample intervals (total 7 times), were utilized in downstream analysis to exclude proteins with higher technical measurement variability. Complete details of liquid chromatography and mass spectrometry instrument settings are as described previously [38].
Biological/molecular functions and cellular components were annotated based on gene ontologies using the DAVID Bioinformatics Resource (v6.7) [41], which calculates a modified Fisher’s Exact P value to determine the probability that the association between each protein in the dataset and functional pathway is random. Functional categories were considered to be those with P-values < 0.05 (Benjamini Hochberg adjusted) and at least 3 proteins selected to be positive associations.
HeLa (ATCC CCL-2) cells were obtained as a gift from the laboratory of Dr. Shiu-Lok Hu (University of Washington), and were maintained in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 4.5 g/L glucose, L-glutamine, 10% (v/v) fetal bovine serum (Corning), and 1% (v/v) penicillin/streptomycin/amphotericin B solution (Gibco). HeLa cells were incubated at 37°C with air/5% CO2 atmosphere. Gardnerella vaginalis ATCC 14018 and Lactobacillus iners ATCC 55195 were obtained from the American Type Culture Collection, and were maintained using HBT-Bilayer medium (BD) and NYCIII liquid medium with incubation at 37°C with air/5% CO2 atmosphere. Frozen stocks were stored in 20% (v/v) glycerol at -80°C.
To assess the impact of different bacteria on the ability of cervical epithelial cells to repair wounds, we utilized the well-established in vitro scratch assay [42]. To prepare live bacteria and culture supernatants for the wound-healing assay, overnight cultures of L. iners and G. vaginalis in NYCIII medium were grown as described above. Wells of a 24-well tissue culture plate (Corning) were initially seeded with 50,000 HeLa cells in a volume of 500 μL DMEM and incubated at 37°C under 5% CO2 until a confluent cell monolayer had formed. Monolayers in each well were then scratched using a sterile P200 pipette tip. Live bacteria, bacterial culture supernatants, or control solutions were then added to the wells. Images at five reference points per well were captured using a Nikon Eclipse TS100 microscope equipped with a Nikon DS-Ri1 camera and the size of the scratch at each reference point was manually analyzed using the ImageJ software. The size of the wound was determined immediately after beginning (t = 0) the experiment and then again after 24 hours (t = 24) of incubation at 37°C with air/5% CO2 atmosphere. Additional information on wound-healing assays is available in S1 Methods.
|
10.1371/journal.pntd.0000393 | Two Distinct Triatoma dimidiata (Latreille, 1811) Taxa Are Found in Sympatry in Guatemala and Mexico | Approximately 10 million people are infected with Trypanosoma cruzi, the causative agent of Chagas disease, which remains the most serious parasitic disease in the Americas. Most people are infected via triatomine vectors. Transmission has been largely halted in South America in areas with predominantly domestic vectors. However, one of the main Chagas vectors in Mesoamerica, Triatoma dimidiata, poses special challenges to control due to its diversity across its large geographic range (from Mexico into northern South America), and peridomestic and sylvatic populations that repopulate houses following pesticide treatment. Recent evidence suggests T. dimidiata may be a complex of species, perhaps including cryptic species; taxonomic ambiguity which confounds control. The nuclear sequence of the internal transcribed spacer 2 (ITS2) of the ribosomal DNA and the mitochondrial cytochrome b (mt cyt b) gene were used to analyze the taxonomy of T. dimidiata from southern Mexico throughout Central America. ITS2 sequence divides T. dimidiata into four taxa. The first three are found mostly localized to specific geographic regions with some overlap: (1) southern Mexico and Guatemala (Group 2); (2) Guatemala, Honduras, El Salvador, Nicaragua, and Costa Rica (Group 1A); (3) and Panama (Group 1B). We extend ITS2 Group 1A south into Costa Rica, Group 2 into southern Guatemala and show the first information on isolates in Belize, identifying Groups 2 and 3 in that country. The fourth group (Group 3), a potential cryptic species, is dispersed across parts of Mexico, Guatemala, and Belize. We show it exists in sympatry with other groups in Peten, Guatemala, and Yucatan, Mexico. Mitochondrial cyt b data supports this putative cryptic species in sympatry with others. However, unlike the clear distinction of the remaining groups by ITS2, the remaining groups are not separated by mt cyt b. This work contributes to an understanding of the taxonomy and population subdivision of T. dimidiata, essential for designing effective control strategies.
| The Chagas disease parasite, transmitted to humans by triatomine bugs, remains a leading cause of heart and digestive disease in Latin America. Pesticide spraying has effectively halted transmission in most of southern South America, especially where the bugs live exclusively inside houses. In Mesoamerica, bugs living in the forest readily reinfest treated houses. In addition, one of the main species of insect that transmits Chagas in Mesoamerica, Triatoma dimidiata, although it looks similar in different localities, may consist of genetically distinct populations, even different species, which differ in how efficiently they transmit the parasite: characteristics which confound control efforts. Nuclear and mitochondrial DNA were analyzed to characterize different populations of T. dimidiata from Mexico and Central America. Both the nuclear and mitochondrial DNA show that there is a very distinct population of T. dimidiata, perhaps even a different species, that lives in very close proximity with other T. dimidiata in Mexico and Guatemala. The nuclear DNA divides the remaining T. dimidiata into three additional genetically distinct groups. However, the mitochondrial DNA does not distinguish these additional groups. This study helps inform control efforts by showing where genetically distinct populations of T. dimidiata occur.
| Chagas disease is considered the largest parasitic disease burden in Latin America with a cost of the loss of 667,000 Disability Adjusted Life Years (DALYs) in 2002 [1]. Trypanosoma cruzi, the parasite that causes Chagas disease, infects approximately 9.8 million people in the Americas [2] with 200,000 new Chagas cases annually [3]. Thus, Chagas disease remains a serious health problem in the Americas. Most transmission occurs by contamination with the parasite-containing feces of triatomine insect vectors (Hemiptera:Reduviidae). There is no vaccine available and treatment shows limited effectiveness, comes with troublesome side effects, and is out of reach of most people in endemic countries. Therefore, as with most parasitic infections, control of transmission by the vectors is the control strategy of choice.
A greater than 94% reduction in Chagas transmission has been realized in South America through the Southern Cone Initiative, a concerted effort of the Ministries of Health and the World Health Organization [4]. This initiative is focused on reduction in vector populations by residual pesticide application in houses and mandatory blood screening. The fact that the targeted S. American vectors are almost exclusively domestic (remain in houses) has greatly facilitated the control efforts. In fact, in regions with significant peridomestic and sylvan vector populations (e.g. the Gran Chaco region), control is still a challenge [5]. In Central America, Mexico, and regions of S. America, many vectors occupy sylvan and peridomestic as well as domestic habitats which poses serious challenges to control efforts as these extra-domiciliary sites serve as reservoirs to repopulate treated houses.
The Triatominae comprises 141 species grouped into 18 genera forming six tribes ([6],[7] and references therein) and more than half of these are naturally infected with T. cruzi [8]. The most important genera involved in Chagas transmission include: Triatoma, Rhodnius and Panstrongylus and epidemiologically significant species within these genera have been described for most endemic regions [8]. The risk particular species pose for transmission to humans is affected by several aspects of triatomine biology and behavior such as: food preference and frequency of feeding, infestation and crowding indices, likelihood and conditions for dispersal, fecundity, and especially the degree of adaptation to the domestic environment which puts them in close contact with human hosts. As humans move into infested sylvan areas, vectors appear to be able to adapt to human dwellings and some species have evolved towards domesticity [9]. An unambiguous identification of vector species and an understanding of the divisions among taxa in endemic areas are critical to an understanding of the epidemiology of the disease and to its control.
Considerable morphological variation of Triatoma dimidiata (Latreille, 1811), the most important Chagas vector in Central America [10], historically has led to splitting, merging and name changes of the species (reviewed in [11]). T. dimidiata is found from Mexico, throughout Central America, and in northern S. America including: Colombia, Ecuador and into to northern Peru. First two separate species were described: Conorhinus dimidiatus from Ecuador, Costa Rica, and Panama; and C. maculipennis from Mexico [12] then quickly synonymized [13], as has occurred again more recently under its current name, T. dimidiata, considering that the variation was “roughly clinal in nature” [14]. However, its taxonomy remains problematic as results of recent morphological evaluation have led taxonomists to assert that T. dimidiata is more aptly considered a species complex [15]. Analysis of antenna sensilla [16], head morphometry [17], cuticular hydrocarbon patterns [18] suggested T. dimidiata was divided into two, three or four taxa, respectively, dividing roughly between Southern Mexico and Guatemala, often with population outliers, such as the Lanquin cave population in Guatemala and the Yucatan, Mexico population (sometimes, but not always, grouped with nearby Peten, Guatemala). Cytogenetic analysis shows three distinct “cytotypes”, however, the divisions differ from those described using phenotypic markers. For example, cytogenetics distinguishes isolates from Peten, Guatemala and Yucatan, Mexico from each other and all other isolates [19], however, other markers, such as cuticular hydrocarbon patterns, show them clustering [18].
Molecular tools are increasingly being used to clarify relationships at all taxonomic levels including tribes, genera, species, subspecies and even populations. The internal transcribed spacer 2 (ITS2) has been particularly useful for analyzing populations of arthropod vectors. ITS2 is part of the rDNA cistron found between the 5.8S and 28S rDNA and is present in hundreds of tandemly repeated copies in the eukaryotic genome. Since the role of ITS2 is to assist with the processing of the 45S precursor RNA to rRNA subunits, only sequences required for its secondary structure need to be conserved [20]. Its high rate of mutation has made it useful for distinguishing species [21],[22], uncovering cryptic species [23], and importantly, identifying species responsible for human infection [24]. Assays based on ITS2 can then be developed to identify species [23]. It is often necessary to include several molecular markers to unambiguously resolve taxonomies [25]. The mitochondrial gene, cytochrome b (mt cyt b), codes for a protein involved in the electron transport chain. Since it is mitochondrial (as opposed to the nuclear ITS2), and protein coding, it provides a distinct marker for taxa subdivision.
Within the Triatominae, ITS2 has become increasingly important [26] and been used to identify two major clades in the Triatomini; one consisting of North and Central American species and the other South American [27], to demonstrate that certain populations were introduced from elsewhere [27],[28], and challenge previous taxonomic arrangements [29].
Among the Triatomini, mt cyt b has been used to understand phylogenetic relationships and population genetic structure, challenge taxonomic status, infer ancestral populations and source of reinfesting populations [30]–[36]. Mitochondrial cyt b has been used to understand divisions among triatomine complexes, but has not yet been used to determine T. dimidiata taxa.
The clinal variation among T. dimidiata populations, suggested by Lent and Wygodzinsky [14] was supported by a study of the male external genitalia on a limited number of samples [37]. Clinal variation was also initially supported by ITS2 studies showing Southern Mexican populations nearly indistinguishable (Yucatan excluded) but increasing differences when compared to Central American populations (Nicaragua and Honduras) [27]. Preliminary ITS2 data showed three distinct taxa, rather than clinal variation; the divisions were: (1) southern Mexico, (2) Central America and (3) Yucatan, Mexico grouped with Peten, Guatemala, this latter group as separate as a different species [11]. In fact, cryptic species may exist in the Yucatan [27],[38]. An “outlier species” is also suggested by cuticular hydrocarbon patterns [18] and cytogenetics [19]. However, these “outliers” are sometimes found in different geographic regions from the putative cryptic species and since different analyses were done on different specimens, it is impossible to tell if they are identifying the same putative cryptic species. Recently, additional ITS2 analyses have shown that two distinct taxa exist in another state in the Yucatan peninsula, Campeche, Mexico, (one taxon includes samples from Central America); the taxa occupying different geographic regions and habitats [39]. And very recently, in isolates from: Mexico, Guatemala, Honduras, Nicaragua, Panama, Colombia and Ecuador, 31 T. dimidiata ITS2 haplotypes were identified falling in four distinct groups, referred to as groups 1A and B, 2 and 3, including one that is proposed to be a separate species (Group 3, T. sp. aff. dimidiata) [38]. This proposed cryptic species (Group 3) was found in Chiapas and Yucatan, Mexico; Peten, Guatemala; and Yoro, Honduras. So it is clear that the most diversity of T. dimidiata is found in the region encompassing southern Mexico through northern Guatemala (perhaps as far south as Honduras) and extending east through the Yucatan peninsula and this is the region where both T. dimidiata (or subspecies of T. dimidiata) and T. sp. aff. dimidiata (Group 3, proposed cryptic species) occur. Nothing is yet known about populations in Belize. We analyzed 53 T. dimidiata samples across this most diverse region and Belize by ITS2 and partial mt cyt b sequences from a subset of these samples to further understand the taxonomic subdivisions of T. dimidiata among Mesoamerican populations.
Knowing the clear identity of the vector species, the dividing lines between different populations and the mechanisms maintaining these divisions is critical to effective control of transmission of Chagas disease [40]. Identification of genetically similar populations could suggest shared characteristics such as: food and habitat preference, tendencies towards domestication, feeding and mating behavior, time and conditions of dispersal, fecundity, etc. Many of these characteristics are directly related to vector competence. The degree of subdivision will indicate the risk of repopulation from nearby populations following control and the degree of genetic variation within a population can suggest the risk of acquisition of insecticide resistance. Genetic markers can also identify the source of re-infesting insects [41]. An understanding of the mechanism of population subdivision may lead to novel control strategies.
The sample information for the T. dimidiata specimens studied with ITS2 are shown in Table 1. Triatoma samples used for the study of mt cyt b sequence are given in Table 2. All T. dimidiata were identified using the key of Lent and Wygodzinsky [14]. Bugs were collected during 2000–2007 by trained personnel using the person-hour collection method or in the case of some Yucatan, Mexico and Belize samples, by householder collection. All samples were collected inside houses (domestic), except those indicated as peridomestic (collected in outbuildings, woodpiles, etc. nearby the house), or sylvan (forest, Table 1). The bug's legs were removed and stored at −4°C in 95% alcohol with 5% glycerol until DNA isolation.
DNA was isolated from two (adults) or three (nymphs) bug legs exactly as described in Dorn et al. [42] using the method originally from [43] with modifications as described in [44] and below. Briefly, bug legs from individual bugs were separately ground using a Kontes pestle in 100 µl grind buffer (0.1 M NaCl, 0.2 M sucrose, 50 mM EDTA, 100 mM Tris-HCl [pH 8.0–9.0], 0.05%SDS). The debris was removed by centrifuging the lysate briefly at 14,000×g. The homogenate was incubated at 65°C for 15–30 min. 8 M potassium acetate was added for a final concentration of 1 M potassium acetate and the solution incubated for 15 minutes on ice to precipitate the SDS. The sample was centrifuged at 14,000×g at 4°C for 10 minutes and the supernatant transferred to a cold 1.5 ml microfuge tube. 2.5× volumes of 100% ethanol were added to precipitate the DNA. The sample was then incubated on ice for at least 10 minutes and centrifuged for 20 min at 14,000×g at 4°C. The resultant pellet was washed with 70% ethanol, allowed to dry and then resuspended in 50 µl sterile TE buffer (10 mM Tris-Cl, pH 7.5, 1 mM EDTA, pH 8.0) containing 1 U RNAase A (Sigma-Aldrich Co., St. Louis, MO). The samples were then stored at −80°C until amplification. Several DNA samples were left at room temperature for ∼1 month due to power failure following Hurricane Katrina. Only those that still amplified following this treatment were used in the analysis.
The ITS2 region was amplified from 3% of the isolated DNA in a 50 µl reaction (3.5 mM MgCl2, 2 U Taq DNA polymerase), exactly according to manufacturer's instructions (Applied Biosystems, Foster City, CA) using primers that anneal to the conserved 5.8S and 28S rDNA that flank the ITS2 region [27]: 5′-CTAAGCGGTGGATCACTCGG-3′ (5,8T) and 5′-GCACTATCAAGCAACACGACTC-3′ (28T). Amplification conditions were as follows: one cycle at 94°C 2 min; followed by 30 cycles of: 94°C - 30 sec, 58°C - 30 sec, 72°C - 30 sec and a final polishing step of 72°C for 7 min. The mitochondrial cytochrome b gene (mt cyt b) was amplified using primers CYTB7432F, 5′-GGACG(AT)GG(AT)ATTTATTATGGATC, and CYTB7433R, 5′GC(AT)CCAATTCA(AG)GTTA(AG)TAA [45] using amplification conditions: one cycle at 94°C 3 min; followed by 30 cycles of: 94°C – 1 min, 45°C – 1 min, 72°C – 1 min and a final polishing step of 72°C for 10 min. Ten percent of the amplified product was visualized by Agarose Gel Electrophoresis and UV transillumination and successfully amplified products purified using QIAquick PCR purification kit or the QIAquick Gel Extraction Kit (QIAGEN, Valencia, CA). Both strands were completely sequenced and sequences edited using MacVector software (version 9.5, Accelrys, San Diego, CA) or Bioedit version 7.0.9 [46] and ClustalW version 2.0 [47] was used to align the data from within Bioedit or MacVector. The Staden Package (version 1.6.0) [48],[49] was used to obtain the haplotype sequences and group identical sequences. DnaSP v 4.50.3 was used to find the polymorphic and parsimony informative sites in the haplotypes [50].
The average evolutionary distance within and between the groups was determined by a Kimura 2-parameter distance calculation (MacVector 10.0), based on the assumption that all sites evolve at the same rate and counting only substitutions, indels are excluded [51]. The estimated time of divergence (below diagonal, Table 3) was calculated using the Kimura 2-parameter distance and the base substitution rate, r = 44.1–99.4 X −10 per site per year calculated by Bargues et al. [52] for ITS2 in Triatomini.
Hierarchical partitioning of molecular variance was tested using AMOVA [53] in Arlequin ver. 3.0 [54].
Maximum Parsimony (MP) analysis was conducted on all characters equally weighted both including and excluding gap characters. Heuristic searches were conducted in PAUP*, version 4.0b10 [55] using 1,000 random taxon addition replicates, holding one single tree in each step and using TBR (Tree Bisection and Reconnection), a branch swapping algorithm for tree search. To estimate clade support 1,000 bootstrap replicates were subject to heuristic searches using 1,000 random taxon additions and TBR branch swapping, giving 463 constant characters, 26 parsimony-informative variable characters.
ML analysis was conducted only on DNA characters, heuristic searches using 1000 random addition replicates and TBR branch swapping were completed under the best fit model (K81uf+I+G), selected by Akaike Information Criterion (AIC) implemented in Modeltest 3.7. The parameters used were: A(0.3191), C(0.1063), G(0.1371), T(0.4375), Nst = 6, Rmat = (1.0000 9.7678 3.4276 3.4276 9.7678) Rates = gamma Shape = 0.9635 Pinvar = 0.7352 [55].
Bayesian phylogenetic analysis was conducted in MrBayes version 3.1 [56]. The sequences were analyzed under the K81uf+I+G model. Clade support was estimated using a Markov Chain Monte Carlo (MCMC) algorithm [56], set to analyze 8 linked chains (sequential heat = 0.1) with four independent runs for 2,000,000 generations sampling every 100 generations. Stability of the process was assessed by plotting the likelihood scores against generation time and 25% of the trees were discarded as part of the burn in.
The full-length ITS2 [38] and mt cyt b sequences [57] (Harris, KD and Beard, CB) available on GenBank are included for comparison. The sequence of T. pallidipennis was used as an outgroup [34],[38] (Accession no. for ITS2, AJ286882).
Median-joining network analysis was performed using Network (version 4.5.0.0, Fluxus Technology, Suffolk, England; fluxus-engineering.com) [58].
The portion of the DNA containing the ribosomal ITS2 sequence was amplified from 53 specimens of T. dimidiata using primers to the 5.8S and 28S rDNA resulting in approximately 900 bp fragments. The ITS2 region was identified according to Bargues et al. [38] and ranged from 489–499 bp in T. dimidiata (Table 1). The sequences were strongly A+T biased at 75–76%. DNA for amplification of the mt cyt b gene was available from a subset of six samples (three each from Belize and Yucatan, Mexico) and trimmed to 665 nt for comparison with sequences available on GenBank (Table 2).
We found 15 ITS2 haplotypes in the 53 sequences and compared these to the 31 available full-length ITS2 haplotypes (from 137 specimens) in GenBank (Fig 1). (The remaining 24 sequences in GenBank are truncated therefore haplotypes cannot be assigned and they are excluded from the analysis). Seven were among the 31 T. dimidiata haplotypes previously identified [38] and eight are unique (found in Belize, Guatemala and Mexico) to give a total of 39 haplotypes. Our phylogenetic and AMOVA analyses support the four distinct groups previously identified [38]: 1A (a Central American cluster, which we now extend into Costa Rica), 1B (Panama and Colombia [also includes one southern Mexico sample]), 2 (a southern Mexico cluster, which we now extend into southern Guatemala), and 3 (a quite distinct taxon found in Yucatan, Mexico Peten, Guatemala, and Cayo, BZ) (Figs 2 & 3). The complex microsatellite repeat around nucleotides ∼44–72 (5′-TT(AT)5TTT(AT)7-3′), shows SNPs and indels among individuals, however changes do not correlate with particular groups. There is a Group 3 signature sequence present around nucleotides ∼307–320 of 5′-CTGTATAAAACAAT-3′. The following four SNPs distinguish Group 2 individuals: a T to C transition at position 213, a G to T transversion at position 400, an A to G transition at position 404, and an A to G transition at position 485.
Separate MP, ML and Bayesian analysis of the ITS2 datasets did not generate discordant topologies among them. ML phylogenetic analysis shows the four groups with strong bootstrap support for each node and Group 2 derived from Group 1B (Fig 3). Results of the median-joining network analysis using ITS2 data show the same four groupings as the ML phylogram (data not shown) with Groups 1A and 1B the closest and central, Group 2 a bit more distant from those two and Group 3 the furthest. Interestingly, one of the newly reported haplotypes, H34, from Veracruz, MX, is in an intermediate position between Groups 1 and 2, which may represent a transitional state between the two or hybridization and further concerted evolution. Although it clusters with Group 1B on the ML tree (Fig 3), the 5′ end of the haplotype appears most closely related to Group 2 (Fig 1). Haplotype H10, representing the Lanquin, GT cave population, clusters with, but is the most distinct from the Central American group, 1A (Fig 3).
The average evolutionary distance within and between groups, calculated by the Kimura 2-parameter model, shows less than 0.5% divergence within each group (0.1–0.4%, diagonal, Table 3) with Group 1A (Central America) showing the greatest intragroup divergence (0.43%). Among groups, 1A and 1B are the most closely related (0.8% substitution), with Group 2 showing 2–2.5 times greater distance with groups 1A and 1B than between the latter two. Group 3 shows the greatest divergence from the three other groups, ∼3–4% which, by the molecular clock, translates into a time of 1.6–5.2 mya since the last common ancestor.
A hierarchical partitioning of variance shows that by far, most of the variance (90.4%, Table 4) can be accounted for by differences among the groups, which is highly significant (p<0.001) and supports the four-group classification [38]. The variance component among countries within groups and among individuals within countries is also significant, although with smaller values (Table 4).
The largest of the fixation indices was nearly one (FST = 0.944), indicating nearly completely distinct haplotypes among countries within groups. FCT is also nearly equal to one (FCT = 0.902) showing the significant variance among groups. These two indices show the importance of the groups, which explains most of the observed variance between haplotypes. FCS indicates the importance of the variance between countries compared to the variances among and within countries (FCS = 0.428). This FCS shows that the variance among countries is slightly less than the variance among haplotypes within countries. The high amount of heterogeneity within countries could mean some populations within a country are isolated from others.
Our ITS2 results support a cluster of “southern Mexican isolates” (Group 2, which we show extends into Guatemala) distinct from a Central American (Group 1A, overlapping in Guatemala) (Fig 2). Our data extends the range of Group 2 considerably southwards to include Quiche and Baja Verapaz, Guatemala, and also Belize, identifying overlap of Groups 1A and 2 in Guatemala. The remainder of Group 1A haplotypes are found in Central America (and our data extends this group southwards into Costa Rica), and 1B found in Panama and Colombia, with the exception of one Veracruz, Mexico haplotype (H34) that clusters with this latter group by ML phylogenetic analysis; however, the long length of the branch indicates it is diverged from Group 1B.
Unlike the geographically localized Groups 1A, 1B and 2, the more divergent taxon, Group 3, shows a scattered distribution, occurring along with Group 2 in Peten, Guatemala and Yucatan, Mexico (Fig 2). We identified four new haplotypes in this group and extend its distribution to Cayo, Belize. In contrast to reports associating distinct groups with specific localities and habitats [39], our data show several distinct groups are in sympatry, sometimes within the same city (Merida, MX) or even microhabitat (palm trees) in the same archeological site (Yaxhá, GT).
To check if Group 3 truly represents a distinct taxon in sympatry with other taxa we compared the ITS2 phylogenies with mt cyt b phylogenies from the six samples for which we had DNA available, representing ITS2 Groups 2 and 3 from Yucatan, Mexico and Belize. Those from ITS2 group 3 (MxYuMe02, MxYuMe03 and BzCCCa34 clearly fall into a distinct taxon also with mt cyt b sequence data, clustering with other isolates from Yucatan, Mexico, where other Group 3 individuals have been identified. As seen with ITS2, the distance is nearly as large as the distance from a different species, T. pallidipennis, used as an outgroup. In addition, network analysis shows a tight clustering, even closer than ITS2, quite separate from all others (Fig 4).
Interestingly, what are clearly distinct taxa as ITS2 Groups 1A and 2 (1B may not be represented in our mt cyt b data as we do not have samples from Panama and Colombia) are unresolved by mt cyt b data (<75% bootstrap value, Fig 5). Both the ML and network analysis show more of a clinal variation than distinct groups (Figs 4 and 5). So it appears that the distinct taxon (Group 3 by ITS2) is well supported, both by ITS2 and mt cyt b. However, the groupings of 1 and 2 are less clear with mt cyt b than ITS2. Southern Mexico and Central America isolates are more similar with mt cyt b than ITS2. More samples and markers will be needed to resolve this issue.
The sizes of the ITS2 sequences ranged from a small of 489 bp (in Group 3) to 499 bp (in Group 2), the latter only slightly larger than previously published T. dimidiata sequences (up to 497 bp) [38]. In addition, the AT bias is evident (75–76%) and is similar to that found in Panstrongylus species (75–79%) [29] and other Triatomini (77%) and Rhodniini (76%) [27].
The previously recognized complex microsatellite ((AT)5TTT(AT)7) was also present in the majority of all the haplotypes (69%), and in all groups, but interestingly is longer than that identified in other members of the phyllosoma which are (AT)4TTT(AT)5–6) [35], thus providing further evidence that T. dimidiata may be diverging from the phyllosoma complex. Our data also shows SNPs present within the microsatellite sequence which may limit the use of this sequence for assigning individuals to groups as has been recently proposed [35]. SNPs in ITS2 that are outside the microsatellite sequence were identified that are diagnostic for Groups 2 and 3. The Group 3 signature can be used to directly identify this group by PCR amplification without the need for sequencing (Dumonteil, et al., unpublished data). So ITS2 can be used to distinguish groups although the region containing the microsatellite sequence may not be the most useful portion.
Since its discovery, the species T. dimidiata, has been split and merged many times (reviewed in Dorn, et al. [11]). At the population level, genetic characters show that among domestic populations, geographically close T. dimidiata are similar [42] whereas generally the geographically more distant populations are diverged [59]; results expected if ancestral populations became separated. Until recently, the divergence was considered to be “clinal in nature” and “not segregated into clearly separable allopatric populations” [14]. However, recently taxonomists have suggested that T. dimidiata is more aptly considered a species complex [15] and rather than “clinal” differentiation, recent phenotypic and genetic data suggest T. dimidiata may be divided into distinct taxa.
Results using particular phenotypic or genetic markers divide T. dimidiata into two (antenna sensilla, [16] and male genitalia (Monroy, et al., unpublished data), three (head morphometry [17] and cytogenetics [19]) or four (cuticular hydrocarbon patterns [18]) distinct taxa. However, the divisions among these taxa do not always agree among different markers (e.g. different “cytotypes” are found in Peten, Guatemala and Yucatan, Mexico however, isolates from these two regions are grouped together [and different from all other isolates] by cuticular hydrocarbon patterns). Often there is an outlier population, but this is sometimes identified in Yucatan, Mexico, and/or Peten, Guatemala (by cuticular hydrocarbons, cytogenetics, male genitalia) or the Lanquin caves in Alta Verapaz, Guatemala or Boavita, Colombia (by head morphometry). To see if different markers are identifying the same or different taxa, it will be necessary to analyze a suite of markers on the same individuals.
Recently 31 haplotypes of ITS2 sequence from T. dimidiata across its geographic range grouped into four taxa by phylogenetic analysis [38]. Our ITS2 data from an additional eight haplotypes supports this classification and we show that the “southern Mexico” Group 2 extends well into southern Guatemala where it overlaps with the Central American Group 1A. This division between a southern Mexican (and Guatemalan) and Central American T. dimidiata (ITS2 estimated time of divergence of 1.02–2.47 my) is also supported by antenna structure, head morphometry, and cuticular hydrocarbon analysis. (Cytogenetics and morphometry of male genitalia do not resolve these two groups). However, these two groups are more closely clustered with mt cyt b data than is seen with ITS2 and more of a clinal variation is seen (Figs 4 & 5). Before we support the proposal to reassign the subspecies designations, T. dimidiata maculipennis, to the ITS2 southern Mexican group 2, and T. dimidiata dimidiata, to the ITS2 Central American taxon, Group 1A, it will be important to look at additional gene sequences and to use multiple markers, phenotypic and genotypic, all on the same samples.
Our data extends the reach of ITS2 Group 1A as far south as Costa Rica. With very few isolates of Group 1B examined so far, and most of these from Colombia, the division between Subgroups 1A and 1B by ITS2 is unclear and needs examination of additional samples in Panama and further north in Central America. In addition, subgroups 1A and 1B show the least intergroup divergence and additional samples, studied with multiple markers, will be needed to see if this is a true division or the “clinal” variation noted by Lent and Wygodzinsky [14]. Indeed, the topology of our ML tree using ITS2 differs somewhat from that published by Bargues, et al. [38] as Group 2 is derived from 1B in our tree and 1A in theirs.
Group 3, the putative cryptic species (T. spp. aff. dimidiata, [38]) is confirmed by our ITS2 data and mt cyt b sequence data, 100% bootstrap support, clearly separate from all other T. dimidiata. It appears to be widespread as we find it in Peten, Guatemala; Yucatan, Mexico, and Belize and it was previously shown to be as far north as Chiapas, MX and as far south as Yoro, Honduras [38]. Interestingly, we have clearly shown that Group 3 exists in sympatry with Group 2 in Peten, Guatemala and Yucatan, Mexico; more isolates are needed to see if this holds true for Belize as well. Finding the distinct groups in the same city or microhabitat in the same archeological site suggests that geographic separation is not essential for reproductive isolation. In addition, we see no association with habitat and group as sylvan and domestic samples are found in Groups 1A, 2 and 3 (we have incomplete habitat information for samples published by Bargues, et al. to assess Group 1B). Cross-breeding experiments are ongoing (Monroy, et al.) to begin to understand the mechanism of this reproductive isolation.
A large body of literature shows that one of the major Chagas vectors in Mesoamerica, T. dimidiata, varies enormously in genetic, phenotypic traits and behaviorally across its geographic range (reviewed in [11]). The studies described here using ITS2 as well as mt cyt b here show a clear separation of the putative cryptic species. However groupings of the remaining populations seem to differ between these two markers. Clearly, information from more genes is needed to clearly understand the division among T. dimidiata taxa. Distinct taxa have significance for the epidemiology of the disease, e.g. in different localities where T. dimidiata is the only Chagas vector, the seropositivity rate in humans differs dramatically, e.g. from 0–18.5% in regions in Guatemala [60]. Distinct taxa may also affect control outcomes. Since 1997, the Central America Initiative for the Control of Chagas disease has shown dramatically different results following insecticide spraying in houses, e.g. in Nicaragua, the bugs did not return [61]; in stark contrast to rapid reinfestation in Jutiapa, Guatemala [62]. It is important to understand how much of the differences in epidemiology and control outcomes are due to distinct taxa of T. dimidiata. The area of Peten, Guatemala has not been included in the control program since most populations are sylvan. Deforestation and increasing encroachment of human populations in the area means that T. dimidiata could become domesticated in this region. It is critical to realize that there are at least two distinct T. dimidiata populations in this area (and in Mexico and Belize) as control measures are designed. This work has begun to clarify the taxonomic status of T. dimidiata from different geographic regions. For effective control it will be imperative to understand the mechanisms maintaining this reproductive isolation and the epidemiological importance of distinct taxa.
|
10.1371/journal.pgen.1002403 | The RNA–Methyltransferase Misu (NSun2) Poises Epidermal Stem Cells to Differentiate | Homeostasis of most adult tissues is maintained by balancing stem cell self-renewal and differentiation, but whether post-transcriptional mechanisms can regulate this process is unknown. Here, we identify that an RNA methyltransferase (Misu/Nsun2) is required to balance stem cell self-renewal and differentiation in skin. In the epidermis, this methyltransferase is found in a defined sub-population of hair follicle stem cells poised to undergo lineage commitment, and its depletion results in enhanced quiescence and aberrant stem cell differentiation. Our results reveal that post-transcriptional RNA methylation can play a previously unappreciated role in controlling stem cell fate.
| We demonstrate that the RNA methyltransferase activity of Misu/NSun2 is required for the proper maintenance of the epidermal differentiation program, and thus post-transcriptional mechanisms are involved in controlling the balance between stem cell self-renewal and differentiation.
| Stem cells are defined by their ability to continuously maintain their population (self-renewal) while generating progeny (differentiation). During self-renewal, stem cells have to avoid cell cycle exit and differentiation; however, when differentiating they have to evade uncontrolled proliferation. Thus, the question of how the balance between self-renewal and commitment is regulated is highly relevant to a fundamental understanding of cell differentiation and cancer.
The hair follicle offers an excellent model to study stem cell fate, as it undergoes cyclic bouts of growth (anagen), apoptosis-mediated regression (catagen) and rest (telogen) [1]. Multipotent hair follicle stem cells, located in a special microenvironment called the bulge, are slow cycling but exhibit long-term contribution to all hair compartments [2], [3]. At the early onset of hair growth, single bulge cells migrate out of their niche in telogen and undergo proliferation as progenitors before they differentiate into hair [4], [5].
Once a stem cell has left its niche, intrinsic and environmental cues converge to balance proliferation of progenitors with lineage-specific differentiation. For example, c-Myc is known to control the balance between stem cell expansion and differentiation [6], [7], [8]. When activated in epidermal stem cells, Myc triggers their exit from the stem cell compartment, induces proliferation of progenitors, and subsequently leads to lineage-specific differentiation [9], [10], [11].
Because the nucleolar protein Misu/NSun2 (Myc-induced SUN-domain-containing protein) is direct target gene of c-Myc [12], we considered the possibility that its RNA methyltransferase activity may represent a novel mechanism to regulate stem cell fate. Misu catalyzes the formation of 5-methylcytidine (m5C) in tRNA and possibly other RNA species [12], [13], [14]. Whereas the function of (cytosine-5) methylated DNA has been extensively analyzed [15], the functional roles of methylated RNA are largely unknown [16]. To date, Misu (NSun2) is one of only two identified m5C methylases with substrate specificity towards tRNA [13], [17], [18], [19].
Here, we demonstrate that Misu is required for normal tissue homeostasis in vivo. Expression of Misu is up-regulated in the hair follicle bulge at the entry of anagen. Deletion of Misu prolongs stem cell quiescence leading to a delay in initiation of anagen. Thus, our data reveals a post-transcriptional modification as a novel mechanism used by stem cells to precise and temporally accurate balance self-renewal and differentiation.
To functionally analyse the role of Misu in vivo, we generated a reporter-tagged loss-of-function mouse model using an ES cell line carrying a Gene Trap in intron 8 of the NSUN2 gene (GGTC-clone ID: D014D11). The methyltransferase activity of Misu is mediated by the conserved SUN domain encoded by exon 2 to 12 (Figure S1A; red box) [19]. Insertion of the Gene Trap and fusion to the reporter β-galactosidase leads to disruption of the catalytic domain of Misu (Figure S1A; blue box) [20]. We confirmed deletion of Misu and presence of β-galactosidase by gene-specific PCR (Figure S1B). Quantitative real-time PCR (QPCR) using a probe for NSun2 showed substantial reduction of full-length Misu in skin of heterozygous (+/−) and loss of expression in homozygous mice (−/−) (Figure S1C). Western Blot analysis demonstrated that Misu −/− mice lacked Misu protein (Figure S1D).
Heterozygous mice for Misu deletion were viable and did not display an obvious phenotype (Figure 1A). Homozygous mice were also viable, but were appreciably smaller than their wild type (wt) and heterozygous littermates (Figure 1A). At three months of age, Misu −/− mice weighed around 30% less than controls (Figure 1B). The smaller size of Misu −/− mice was confirmed by a second knock-out mouse model Nsun2tm1a(EUCOMM)Wtsi generated by the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk) (data not shown). Misu −/− males were sterile, and both genders had pelage that was comparable to controls after birth, but showed cyclic alopecia at around 10 months of age (Figure 1C, 1D).
Both, the yeast and human orthologue of Misu have previously been shown to catalyse the methylation of cytosine at position 34 (C34) in intron-containing pre-tRNALeu (CAA) [13], [19]. To investigate whether mouse Misu also methylated tRNALeu (CAA), we used RNA bisulfite sequencing, which allows analysis of RNA methylation patterns in their native sequence context [21]. We find that intron-containing pre-tRNALeu (CAA) was methylated at C34 in wild-type mice but lacked the modification when Misu was deleted (Figure 1E).
We concluded that Misu was indispensable for methylation of tRNALeu (CAA) in skin. Based on our observation that older mice show signs of alopecia in the absence of Misu we speculated that Misu might have a functional role in maintaining skin homeostasis in the long-term.
To determine the role of Misu in skin, we began by examining endogenous expression of Misu during embryogenesis and skin development by staining for LacZ (Figure 2A–2Q). Misu-expression was detected from E3.5 in the inner cell mass of the blastocyst (Figure 2A). After implantation and gastrulation, at E6.5, Misu was observed throughout the extra-embryonic ectoderm (Figure 2B, 2B′), which gives rise to the nervous system and epidermis. Starting from E9.5, expression of Misu became more restricted and at E13.5 and E14.5 Misu was enriched in developing whiskers (arrow) and eyes (arrowhead) (Figure 2C–2F). From E15.5, when the interfollicular epidermis (IFE) begins to stratify and follicular morphogenesis starts by forming hair placodes (Figure 2G), highest expression of Misu was found in the suprabasal layer of IFE (Figure 2G–2I; arrows).
After birth, from postnatal day 1 (P1), expression of Misu in the IFE waned (Figure 2J, 2K, 2N, arrows) but increased in the matrix of growing (anagen) hair follicles (Figure 2L, 2M, arrowhead). At the end of follicular morphogenesis, from around P14 to P19, when hair follicles progress through the destructive (catagen) and resting (telogen) phase of the hair cycle, expression of Misu was absent (data not shown and Figure 2N). At early anagen (P23), Misu was up-regulated in the hair germ (HG; arrowhead) and weakly expressed in the bulge (Bu; arrow) (Figure 2O; insert). Throughout anagen, Misu was highly expressed in matrix but was also found in differentiated lineages of the hair follicles (Figure 2P, 2Q, Figure S2). We confirmed the dynamic expression pattern of Misu in skin by detecting endogenous RNA levels of Misu during morphogenesis (M), catagen (C), telogen (T), and anagen (A) of the first postnatal hair cycle by QPCR (Figure 2R).
In conclusion, Misu was dynamically expressed during morphogenesis and in adult skin. In adult skin, expression of Misu was up-regulated in the bulge and hair germ as soon as the hair follicle entered its growing phase (anagen). During anagen, cells of the hair germ give rise to the hair matrix, which showed highest expression of Misu. Matrix cells, which are often referred as to transit amplifying (TA) cells because they only survive through anagen [22], are stem cell progenitors that divide a finite number of times until they become differentiated.
Misu was originally identified as a transcriptional target of c-Myc in skin [12]. However, activation of c-Myc triggers epidermal stem cells to differentiate mainly into lineages of IFE and sebaceous glands [7]. Since a role for c-Myc in regulating bulge stem cells has yet not been identified and c-Myc expression levels remained unchanged during the hair cycle (Figure 3A), we asked whether expression of Misu might also be regulated by hair-specific transcription factors. We examined the mouse Misu proximal promoter and detected a putative Lef1-binding motif (Figure 3H). As described for Misu, also expression of Lef1 increased when hair follicles entered anagen (Figure 2R; Figure 3B, 3C) and expression of Misu and Lef1 overlapped in hair follicles at both early and late stages of anagen (Figure 3F, 3G).
To validate Misu as a target gene of Lef1, we performed chromatin immunoprecipitation (ChIP) in mouse epidermis in anagen using an antibody for Lef1 (Figure 3I). We detected Lef1-binding to the promoters of Misu and Msx2, a known target gene of Lef1 [23]. We further confirmed Misu as a transcriptional target of the Lef1/β-catenin complex by luciferase assays using the Misu promoter (pMisu) and full-length Lef1 or ΔN63Lef1 (ΔLef1), a mutant construct lacking the β-catenin binding motif. The assays were performed in presence or as an additional negative control absence of β-Catenin (Figure 3J). The Lef7 synthetic promoter (pLef7) served as positive control (Figure 3J). Luciferase activity using the Misu promoter increased around two-fold when the Lef1/β-catenin complex was present compared to the controls (Figure 3J). Finally, we confirmed in vivo that Misu RNA levels decreased when ΔLef1 was over-expressed in the mouse epidermis (K14ΔLef1) (Figure 3K). We concluded that expression of Misu is up-regulated by Lef1 when hair follicles enter anagen.
The complete lack of expression of Misu in adult skin in both the IFE and the bulge in either the catagen or telogen phase of the hair cycle excludes its expression in quiescent stem cells. However, Misu-expression, detected by LacZ staining, was up-regulated in the bulge region (arrows) and the hair germ (HG) as early as telogen to anagen transition (Figure 4A, 4B; Figure S3A, S3B). We confirmed expression of Misu protein in the bulge (arrows) and the hair germ (arrowheads) in anagen by immunoflourescence using an antibody against mouse Misu (Figure 4C–4E; Figure S3C).
We next asked whether Misu-positive cells in the bulge were indeed stem cells. Hair follicle stem cells can be isolated by fluorescence activated cell sorting (FACS), based on high expression of CD34 and α6 integrin (Itgα6) (Figure 4F–4H) [24]. Progenitor cells of the hair germ are characterized by high expression of P-cadherin and low expression of Itgα6 (Figure 4I–4K; Figure S4A–S4C) [5]. Both, stem and progenitor populations were sorted from Misu +/− mice at the onset of anagen (P21) and tested for expression of Misu based on β-galactosidase activity (LacZ) using fluorescein di-galactoside (FDG) (Figure 4F–4K). Around 12% of bulge stem cells (Itgα6high/CD34+ve) and 17% of progenitor cells in the hair germ (Itgα6low/P-cadherinhigh) expressed Misu (Figure 4F–4K). No signal for FDG was detected in keratinocytes from wild-type mice (Figure 4G, 4J).
To further confirm co-expression of Misu with stem cells markers, we isolated FDG+ve and FDG−ve keratinocytes from Misu +/− mice at P21 (Figure S5A, S5B). QPCR analysis demonstrated that the stem cell markers CD34, NFATc1, and Lgr5 were enriched in Misu-expressing cells (FDG+ve) (Figure 4L–4O). Both populations expressed Itgα6 at similar levels (Figure S5C) but FDG+ve cells were also enriched for the hair germ markers Lef1, Wnt5a, and Sox6 (Figure 4P; Figure S5D, S5E). Expression of differentiation markers was comparable or decreased compared to FDG−ve cells (Figure S5F, S5G). We concluded that Misu was expressed in both bulge stem cells and cells of the hair germ at initiation of anagen.
To test whether Misu might induce bulge stem cells to enter cell cycle at the transition of telogen to anagen, we FACS-sorted bulge stem cells, based on high expression of CD34 and Itgα6 (Figure S6A) and early progenitor cells of the hair germ, based on high expression of P-cadherin and low expression of Itgα6 (Figure S6B) at the onset of anagen at P21 in wild-type and Misu −/− mice. We then determined the cell cycle profile for all sorted populations (Figure 5A–5I). We did not observe any difference in cell cycle profile when the whole epidermal population was analyzed, and as expected most of the cells were in G1 (Figure 5B, 5C). However, the percentage of Misu −/− bulge stem cells (Itgα6high/CD34+ve) in S- and G2/M-phase of the cell cycle was significantly reduced compared to wild-type cells at anagen initiation (Figure 5D, 5E). In contrast, the cell cycle profile of progenitor cells (Itgα6low/P-cadherinhigh) in Misu −/− and wild-type epidermis was comparable (Figure 5F, 5G). At later stages, in anagen at P24, the percentage of Misu −/− bulge stem cell population dividing was comparable to that of wild-type controls (Figure 5H, 5I; Figure S6C). These data indicated that Misu was important to stimulate cell cycle entry of bulge stem cells at initiation of anagen and depletion of Misu might increase the quiescent phase of bulge stem cells.
To test our hypothesis that lack of Misu resulted in increased number of quiescent bulge stem cells, we labelled Misu −/− and wild-type mice with BrdU, and detected label-retaining cells (LRC) after a chase period of four months (Figure 6A–6D). A long chase period allows detecting differences not only in the number of LRC but also in the intensity of the BrdU-label, which correlates with number of divisions. The number of LRC in Misu-depleted tail hair follicles was significantly increased in the outer follicles of triplets (Figure 6A–6C), which are known to go through the hair cycle concurrently, whereas the central follicle cycles asynchronously and usually has fewer LRC (data not shown) [25], [26], [27]. Additionally, the intensity of the BrdU-label was higher in Misu −/− skin suggesting that those cells divided slower than in their wt controls (Figure 6D).
Flow cytometry for CD34 and Itgα6 in anagen (P30), catagen (P40) and telogen (P49) confirmed that loss of Misu resulted in a two-fold increase of bulge stem cells only in telogen of the hair cycle (Figure 6E, 6F; Figure S7A). The number of hair germ cells in telogen (P49) was unchanged (Figure S7B). Strikingly, we found that the increase of bulge stem cells in Misu −/− epidermis was due to an enrichment of a distinct cell population with lower Itgα6 expression compared to wild-type skin (Figure 6E; red line), which might represent those bulge cells that failed to enter the cell cycle at the initiation of anagen (Figure 5D). If our hypothesis was correct CD34+ve but Itgα6low cells from Misu −/− mice should be more quiescent and therefore express higher levels of bulge stem cell markers than the comparable wild-type population.
To test our hypothesis, we sorted bulge stem cells into two populations: L (CD34+ve/Itgα6low) and H (CD34+ve/Itgα6high) (Figure 6G). We then analysed both populations for expression of bulge markers in wild-type and Misu −/− epidermis (Figure 6H). We found that population L obtained from Misu −/− mice showed consistently higher expression of the stem cell markers FGF-18, Itgα6, Sox9, CD34 and NFATc1 relative to wild-type mice (Figure 6H; Figure S8E). Importantly, population L did not show increased expression of the hair germ markers Wnt5a and Sox6, or Lgr5, a marker for cycling stem cells (Figure 6H). In contrast, population H from Misu −/− mice expressed similar levels of stem cell markers than wild-type mice, but showed an increase in the expression of hair germ markers. Thus, our data indicated that depletion of Misu resulted in an accumulation of quiescent CD34+ve/Itgα6low stem cells in the bulge.
To exclude the possibility that the increase in the quiescent stem cell population in Misu −/− mice was due to the general lack of Misu rather than being tissue-autonomous to skin, we generated a conditional knockout mouse model for Misu in the epidermis (Methods) (Figure S8A–S8C). Like in Misu −/− animals, also in mice with conditionally deleted Misu in the basal, undifferentiated layers of the epidermis (K14MisuΔ/Δ), a distinct cell population of stem cells with lower Itgα6 expression accumulated in the bulge (Figure S8D; red line in right hand panel) and expressed higher levels of stem cell markers than the Cre-negative Misuf/f controls (Figure S8E).
In vivo lineage tracing experiments suggest that a sub-population of bulge cells migrate into the hair germ in telogen to then undergo cell division at the onset of anagen [4], [28]. We speculated that stem cells in the bulge of Misu −/− mice might fail to commit in the bulge and do not migrate into the hair germ. To determine whether lack of Misu affected stem cell migration into the hair germ, we labelled Misu −/− new born mice and control littermates with BrdU at late morphogenesis and chased them for one hair cycle until the second postnatal telogen (P47) We then detected quiescent bulge stem cells using an antibody to BrdU and compared the number and location of label-retaining cells (LRC) in the upper (high) and lower bulge region and the hair germ (Figure 6I–6M; Figure S9A–S9F) [4].
Wild-type and Misu −/− epidermal cells incorporated BrdU at the same rates (Figure S9A, S9B). After the chase period at P47, we found a significant higher number of fully labelled LRC in the high bulge area of Misu −/− epidermis (Figure 6J–6L). Moreover, the number of hair follicles without any LRC in the lower bulge or hair germ was three-fold higher in Misu −/− skin than wild-type littermates (Figure 6M). Accordingly, we also found that the intensity of the BrdU-label increased in the higher bulge and decreased in the lower bulge and hair germ when Misu was depleted (Figure S9C–S9F). Importantly, at this stage of the hair cycle the majority of bulge cells are yet not dividing, as shown by labelling for Ki67 (Figure 6J, 6K), excluding the possibility that Misu −/− cells diluted the label by cell divisions.
Our data showed that the increased quiescence of bulge stem cells in Misu −/− skin correlated with an accumulation of LRC in the upper part of the bulge, indicating that Misu −/− bulge stem cells are delayed in generating committed progenitor cells of the hair germ.
To test whether Misu-deletion led to increased self-renewal capacity of stem cells in vitro, we measured the colony forming efficiency (CFE) of sorted Itgα6high/CD34+ve bulge stem cells (Figure 7A). Out of all cell populations derived from skin, epidermal stem cells exhibit the highest CFE [29]. When seeded in clonal density, keratinocytes derived from Misu −/− mice formed more colonies than wild-type littermates (Figure 7A), indicating that Misu −/− cells have a higher self-renewal capacity than wild-type cells. Similarly, unsorted keratinocytes obtained from Misu −/− epidermis showed higher CFE than control keratinocytes (Figure 7B, 7C). Although expression of Misu was undetectable in adult IFE (Figure 2N), cultured keratinocytes obtained from back skin expressed high levels of Misu RNA and protein (Figure 7D; data not shown).
Our data indicated so far, that loss of Misu increased accumulation of bulge stem cells in their niche at late telogen leading to an increased self-renewing but quiescent stem cell population. Therfore, we speculated that Misu −/− hair follicles should be delayed in entering anagen. Indeed, depletion of Misu led to a delay of entry into the first and second synchronized hair cycle in males (Figure 7E–7L; Figure S10B, S10C). Compared to males, females exhibit a delayed hair cycle progression of around 2 days, yet even in Misu −/− females the percentage of hair follicles in the first anagen at P25 was lower than in their wild-type controls (Figure S10A; Table S1). Similarly, male mice with conditionally deleted Misu in the epidermis (K14MisuΔ/Δ), displayed a delay in entering the first adult hair cycle compared to controls (Misuf/f) (Figure 7M; Figure S10F–S10K).
Later entry into anagen in Misu-depleted skin resulted in delayed differentiation of matrix cells (Figure 7N–7U). The number of Lef1-postive cells, marking lineage committed hair follicle cells, was lower in matrix cells of Misu −/− mice (Figure 7N, 7O; arrows) [30]. Accordingly, expression of Dlx3, Gata3 and BMP signalling, as determined by staining for phosphorylated Smad1/5/8, were absent in Misu −/− anagen hair follicles (Figure 7P–7U; arrows). Once Misu −/− mice entered anagen, the hair follicles were morphologically indistinguishable from those of wild-type mice (Figure S10F, S10E, S10L–S10S). Our data suggested that Misu plays a role in accurately timing lineage commitment of hair follicle progenitor cells.
Here we show through generating general and skin-specific loss-of-function mouse models that the RNA methyltransferase Misu (NSun2) defines expanding, committed progenitor populations in mammalian skin. Expression of Misu is absent in adult mouse interfollicular epidermis and the quiescent phases of the hair cycle (catagen and telogen). As the hair follicle enters its growing phase (anagen), Misu is expressed in the bulge and hair germ, both of which contain multipotent stem cells [3], [31], [32]. Cells in the hair germ give rise to transit amplifying (TA) cells in the hair matrix [4], [5]. Matrix cells, which collectively show the highest expression of Misu of all cell types, subsequently differentiate into all hair lineages [3].
Uniquely at the transition of telogen to anagen, Misu is co-expressed with markers for both quiescent (CD34+ve/Itgα6high) and cycling (Lgr5+ve) stem cells from the bulge and the hair germ respectively. However, in contrast to those stem cell populations, Misu- expressing cells exhibit reduced self-renewal capacity. Thus, although Misu could reversibly commit stem cells to differentiate [33], we clearly show that stem and committed progenitor cells co-exist within the bulge, the hair follicle stem cell niche. Our findings now raise the question of how only a few selected stem cells are activated during each hair cycle.
Misu is required for cellular division of bulge cells only at the telogen-to-anagen transition, indicating that its function is temporarily and spatially controlled in a strict manner. One key pathway that drives epidermal stem cells from telogen into anagen and specifies hair follicle lineages is the canonical Wnt pathway. Activation of Wnt signaling by transient expression of N-terminally truncated β-catenin in the epidermis is sufficient to induce ectopic hair follicle formation [34], [35], [36], [37], [38]. Conversely, when the pathway is inhibited by β-catenin ablation or expressing of a ΔNLef1 mutant, hair follicle formation is impaired [30], [39], [40], [41]. The identification of Misu as a direct downstream target of Lef1 further supports Misu as a key component during lineage commitment of bulge stem cells at the initiation of anagen. Thus, our data support a model in which stem and committed progenitors are distinct populations within the hair follicle that can be distinguished by their expression of Misu.
Misu belongs to a large family of highly conserved methyltransferases, modifying cytosine-5 in RNA (RNA:m5C-MTase) [16]. Misu/NSun2 is the human orthologue of S. cerevisiae Trm4, both of which have substrate specificity towards tRNA [12], [13], [19]. Like for human and yeast NSun2, we confirm pre-tRNALeu (CAA) as a direct target substrate for mouse Misu in vivo. Although post-transcriptional methylation of tRNA at cytosine-5 is one of the most frequently encountered modifications, Dnmt2 and Misu are as yet the only identified tRNA:m5C MTases [13], [17], [18], [42]. Although it has been recently shown that m5C methylation protects tRNA from cleavage and degradation, the biological function this may mediate remains unclear [17], [43].
Similar to the depletion of Misu in mice, loss of Dnmt2 in zebrafish also results in reduced body size and impaired differentiation of specific tissues [44]. It remains, however, unclear why the RNA-methyltransferase activity of both proteins is critical in maintaining tissue homeostasis [44]. One possible mechanism of Misu's function is that methylated tRNA species could be directly involved in regulating stem cell differentiation. An alternative and intriguing possibility could be that methylation regulates the cleavage of tRNAs or intron-splincing of pre-tRNAs to generate products with microRNA-like features, which may offer an additional mechanism of post-transcriptional control [17], [45]. Recent RNA deep-sequencing data identified a new set of small RNAs derived from tRNAs, including intron sequences, which are associated with Dicer and Argonaute proteins, strongly suggesting a role of these fragments in RNA silencing [46], [47].
On the other hand, modified nucleotides in tRNA are well known to affect their structural and metabolic stability, and thus are likely to influence directly the rate or efficiency of protein translation [42], [48]. Interestingly, protein translation is hierarchically controlled during stem cell self-renewal and differentiation; while parsimonious during self-renewal, it enhances during differentiation [49]. This increased protein synthesis capacity in stem cells could allow rapid elevation of translational rate in response to differentiation signals [49].
In summary, we demonstrate that the RNA methyltransferase Misu (Nsun2) stimulates a sub-population of stem cells to leave the hair bulge and become committed progenitor cells in the hair germ. Thus we identify post-transcriptional RNA modifications as a novel mechanism by which stem cells control the balance between stem cell self-renewal and differentiation.
All mouse husbandry and experiments were carried out according to the local ethics committee under the terms of a UK Home Office license.
Misu −/− mice were derived using 129S2 ES cell line carrying a Gene Trap in intron 8 of the NSUN2 gene (GGTC-clone ID: D014D11) generated by the German Gene Trap Consortium, and then mated with C57Bl6/J CBA F1 mice. F1 progeny was subsequently inbred. A PCR-based strategy was developed to distinguish the wild-type and Misu gene trap alleles. The primers are as follows: SR2, (5′GCC AAA CCT ACA GGT GGG GTC TTT) and B34 (5′- TGT AAA ACG ACG GGA TCC GCC) amplify a fragment 650 bp of the ß-geo cassette. The primers Misu-Int8-5′ (5′AGG TGG ACC TGA TCA TGG AG) and Misu-Int8-3′ (5′-AGGG AGG GTC TGG AAA GATG) amplify a fragment of 500 bp of the wild-type allele.
Mice containing a floxed allele of the NSUN2 gene were obtained by first crossing Nsun2tm1a(EUCOMM)Wtsi mice, generated by the Wellcome Trust Sanger Institute, with transgenic mice expressing Flp recombinase to delete the LacZ-neo cassette [50]. The offspring, containing two LoxP sites flanking exon 6, were then crossed with KRT14-cre mice (The Jackson Laboratory). In KRT14-cre mice, Cre-recombinase is expressed under the control of the keratin14 (KRT14) -promoter leading to deletion of Misu in all basal, undifferentiated cells of the epidermis (K14MisuΔ/Δ).
All mouse lines were bred to a mixed genetic background of CBA×C57BL/6J. Primers to identify Misuwt, Misuflox and MisuΔ alleles are F1 (5′CCC CCA CTG CTG CTC AAC G) and R1 (5′CAA TGC CAC CAC AAC CTC CTT).
Total of 97 Misu −/− mice with their wild-type controls and 29 K14MisuΔ/Δ with their Misuf/f control mice were genotyped and grouped by gender. Samples were taken from same regions from dorsal skin and process for H&E staining. Each mouse was classified into specific stages of the hair cycle based on established morphological guidelines [51]. The total numbers were represented as percentages.
Bisulfite conversion of tRNA was carried out as previously described [21]. cDNA was PCR amplified using primers specific for the deaminated sequences of tRNA-998Leu (CAA) and tRNA-1911Leu (CAA). tRNALeu (CAA) sequences were obtained from the Genomic tRNA Database (http://gtrnadb.ucsc.edu/). Primer sequences are as follows: Fw_Leu_De: 5′GAT GGT TGA GTG GTT TAA GGT GTT, Rv_Leu998_De: 5′CAC CTC CAA AAA AAA CCA AAA C and Rv_Leu1911_De: 5′CAC CTC CAT TCA AAA ACC AAA AC.
DNA label-retaining cells (LRC) were generated by repeated BrdU (Sigma) injections of neonatal mice at P10 [25]. For LRC assays animals were chased for the times indicated. For tracing migration of bulge LRC into the hair germ, animals were sacrificed at P47 [4]. LRC were detected by BrdU immunostaining in tail skin whole mounts. Z-stack volumes of random areas of the slide were collected using a confocal microscope (Leica SP5). Maximum projected images were quantified using Volocity software (PerkinElmer). Images were segmented to identify and measure intensity of BrdU-positive cells constrained to specific Regions Of Interest (ROI) defining the whole hair follicle (bulge and hair germ), high and low bulge region and hair germ. Frequency distributions of BrdU intensity were calculated with Microsoft Office Excel 2007 software (Microsoft).
Immunostainings were performed on 10 µm paraffin sections. After citrate epitope retrieval, sections were permeabilized for 5 minutes with 0.2% Triton×100 at room temperature, blocked for 1 hour with 5% FCS and incubated overnight with the appropriate antibody dilution. Immunostaining on cryosections or cultured cells was performed as for paraffin after fixation for 10 minutes in 4% paraformaldehyde at room temperature. Tail epidermal whole mounts were prepared and immunolabelled as described previously [25], [36].
For LacZ staining, whole mounts of embryos or freshly obtained skin samples were fixed for 30 minutes at room temperature in buffer containing 0.1 M phosphate buffer, 5 mM EGTA, 2 mM MgCl2 and 0.2% glutaraldehyde. Samples were then washed three times for 15 minutes each in wash buffer (2 mM MgCl2 and 0.1% Nonidet P40 in 0.1 M phosphate buffer) and stained for 12 hours in a solution consisting of 1 mg/ml X-gal (Melford), 5 mM K3Fe(CN)6 and 5 mM K4Fe(CN)6 in wash buffer. The skin samples were then embedded in paraffin, sectioned at 10 µm and stained with eosin or used for stainings.
Primary antibodies were used at the following dilutions: rabbit monoclonal antibody to Ki67 (1∶100; SP6, Vector Labs), rabbit polyclonal anti mouse keratin 14 (1∶2000; Covance), rabbit polyclonal anti mouse keratin 10 (1∶500; Covance), mouse monoclonal anti Gata3 (1∶50; HG3-31, Santa Cruz Biotechnology), rabbit polyclonal anti Lef1 (1∶50; Cell Signaling Technology), mouse polyclonal anti Dlx3 (1∶200; Abnova), guinea pig polyclonal anti hair keratins 31, 71 and 72 (1∶200; Progen), rabbit polyclonal anti keratin 6 (1∶5.000; Babco), rabbit polyclonal to phosphor-Smad1/5/8 (1∶50; Cell Signaling Technology), rabbit polyclonal CUK-1079-A antibody to mouse Misu (1∶1000; produced by Covalab), mouse monoclonal to keratin 15 (1∶1000) [25]), rat monoclonal anti BrdU (1∶100; Abcam), goat polyclonal anti P-cadherin (1∶100; R & D Systems), rat monoclonal anti α6 integrin (1∶500; GoH3, AbD Serotec), Secondary antibodies (Alexa Fluor 594- and 488-conjugated anti-rabbit, mouse, rat and guinea pig, Invitrogen) were added at a dilution of 1∶500 for 1 hour at room temperature together with DAPI to label nuclei.
White field images were acquired using an Olympus IX80 microscope and a DP50 camera. Confocal images were acquired on a Leica TCS SP5 confocal microscope. Z-stacks were acquired at 100 Hz with an optimal stack distance and 1024×1024 dpi resolutions. Z-stack projections were generated using the LAS AF software package (Leica Microsystems). All the images were processed with Photoshop CS4 (Adobe) software.
Proteins were extracted from cultured keratinocytes or total skin. 1 cm2 pieces of total back skin (dermis and epidermis) were snap-frozen in liquid N2, transferred to lysis buffer (1% NP-40, 200 mM NaCl, 25 mM Tris-HCl, pH 8, 1 mM DTT) including protease inhibitor cocktail (Roche) and homogenised for 30 seconds. Samples were incubated on ice for 20 minutes. Protein lysates were cleared by centrifugation at 13,000 rpm. Total protein concentration was quantified using Dc Protein Assay (Bio-Rad). Equal amounts of protein were run in 7.5% polyacrylamide gels and blotted onto Hybond-P PVDF membranes (GE Healthcare), which were incubated in TBST-blocking solution (Tris-buffered saline, pH 8.8, with 5% skimmed milk powder). Blots were incubated overnight at 4°C with primary antibodies, washed and incubated with the appropriate HRP-conjugated secondary antibodies (GE Healthcare). α-Tubulin (Sigma) was used as a loading control. The chemiluminescent signal was detected using the ECL Plus Detection System (GE Healthcare).
Total RNA from mouse skin or cultured keratinocytes was prepared using Trizol reagent (Invitrogen) according to the manufacturer's instructions. Total RNA from flow-sorted cells was purified using Pure-Link RNA Micro Isolation Kit (Invitrogen). Double-stranded cDNA was generated from 1 µg total RNA using Superscript III First-Strand Synthesis kit (Invitrogen) and random hexamer primers (Promega). A minimum of two independent biological and three technical replicates was analysed. In case of Itgá6/CD34+ve-sorted cells one sample was pooled from four mice. For FDG-sorted cells one sample was pooled from one mouse.
Real-time PCR amplification and analysis was conducted using the 7900HT Real-Time PCR System (Applied Biosystems). The standard amplification protocol was used with pre-designed probe sets and TaqMan Fast Universal PCR Master Mix (2×) (Applied Biosystems). Probe set Mm00520224_m1 and Mm00487803_m1 were used to amplify mouse NSun2 (Misu) and c-Myc from total skin. The following probes were used to amplify selected genes from flow-sorted cells: α6 integrin (Mm01333831_m1), FGF-18 (Mm00433286_m1), CD34 (Mm00519283_m1), NFATc1 (Mm00479445_m1), Sox9 (Mm00448840_m1), Lgr5 (Mm00438890_m1), Sox6 (Mm00488393_m1), Wnt5a (Mm00437347_m1), Lef1 (Mm00550265_m1), Gata3 (Mm01337569_m1) and Keratin 72 (Mm00495207_m1). GAPDH expression (4352932E) was used to normalize samples using the ΔCt method.
The −2068 to −48 bp DNA fragment of the mouse Misu promoter was cloned into the pGL3-Basic vector (Promega). Plasmids for the reporter assays included pCDNA 3.1-hLef1-V5 (Lef1), pCDNA 3.1-ΔN63-hLef1-V5 (ΔLef1), and pCDNA 3.1-S33Y mCTNNB1 (β-catenin). pLef7-fos-luc (pLef7) was kindly provided by R. Grosschedl [52]. Hela cells were grown in DMEM (Invitrogen) supplemented with 10% fetal calf serum (FCS) in a humidified atmosphere at 37°C and 5% CO2. Cells were transiently co-transfected with the promoter construct, pRL-TK renilla as an internal control and the indicated plasmids using Lipofectamine LTX transfection reagent (Invitrogen). After recovery, cells were grown in media containing 0.2%FCS. Luciferase activity was measured 36–48 hours after the transfection using the Dual-Luciferase Reporter Assay System (Promega) on Glomax (Promega). Each transfection was carried out in triplicate and the experiment was repeated twice.
To isolate mouse keratinocytes from dorsal back skin we rinsed mouse back skin in 10% Betadine and 70% ethanol and washed it in PBS. The dermal side was thoroughly scraped to remove excess fat. The tissue was then floated on 0.25% Trypsin without EDTA (Invitrogen) for 2 hours at 37°C or overnight at 4°C. The epidermis was subsequently scraped from the dermis, minced using scalpels, disaggregated by gentle pipetting and filtered through a 70 ìm cell strainer. Trypsin was inactivated by addition of low-calcium medium with 10% FCS. The cells were pelleted and resuspended in the following antibodies: PE-conjugated Itgá6 (clone GoH3, eBiosciences), Alexa Fluor 647-conjugated CD34 (RAM34, eBiosciences) and goat polyclonal anti-P-cadherin (R & D Systems). After incubation for 45 minutes at 4°C, cells were washed twice in PBS. For detection of P-cadherin, cells were incubated for 10 minutes at 4°C with anti-goat Alexa Fluor 647-congugated secondary antibody (Invitrogen).
Cells were gated using forward versus side scatter to eliminate debris. Doublet discrimination was carried out using pulse width. The viable cells were then gated by their exclusion of DAPI using a 450/65 nm filter. Itgá6 PE stained cells were detected using a 580/30 nm filter and Alexa 647 cells were detected using a 670/30 nm filter. Cells were sorted with a MoFlo high-speed sorter (Beckman Coulter).
For cell cycle analysis of Itgá6high/CD34+ve or Itgá6low/P-cadherinhigh cell populations, cells were fixed with 1% paraformaldehyde for 5 minutes after immunolabelling, transferred to cold 70% ethanol and incubated for at least 1 hour before stained with propidium iodide (PI) or DAPI (Sigma). After incubating the cells for 1 hour in RNase, analysis was carried out on a CyAN ADP analyzer (Beckman Coulter).
For detection of intracellular β-galactosidase activity, mouse keratinocytes from mice in early anagen (P21) were loaded with fluorescein-di-β-D-galactopyranoside (FDG) (Sigma) using hypotonic shock. Briefly, equal volume of cells was mixed with warm 2× hypothonic shock solution (2 mM FDG in water) and incubated for 30 seconds at 37°C, then cold media was added. Cells were washed and subsequently stained with the appropriate antibodies. Cells were then sorted based on FDG (fluorescence detected using a 530/40 nm filter), after gating out dead cells (based on DAPI staining).
Mouse keratinocytes were isolated as described above and cultured on mitomycin-treated J2-3T3 feeder cells on collagen type I (BD Biosciences) coated plates (BD Falcon). Mouse keratinocytes were grown in low-calcium FAD media (one part Ham's F12, three parts Dulbecco's modified Eagle's medium, 18 mM adenine and 0.05 mM calcium) supplemented with 10% FCS and a cocktail of 0.5 µg/ml of hydrocortisone, 5 µg/ml insulin, 10−10 M cholera enterotoxin, and 10 ng/ml epidermal growth factor (HICE cocktail) and maintained in a humidified atmosphere at 32°C and 8% CO2.
Clonal growth was assayed by culturing 500 to 2500 Itgá6high/CD34+ve sorted cells or 5000 to 10000 viable epidermal cells per well in 6-well plates for 3 weeks. Cells were fixed and stained with 1% Rhodamine B. Three independent experiments were conducted.
Mouse keratinocytes were isolated from mouse skin in anagen and processed according Chromatin immunoprecipitation Assay Kit (Upstate). Chromatin was incubated with control anti-rabbit IgG and anti-Lef1 (Cell Signaling) antibody overnight at 4°C. The samples were eluted after washing. PCR reactions were performed by sets of specific primers: Misu TCT GTG CGG TCC TTT CTA CC (forward) and CGC GTC CTG CTA GCT ATG TT (reverse); Msx2 AAG GGA GAA AGG GTA GAG (forward) and CCC GCC TGA GAA TGT TGG (reverse) and GAPDH TAC TAG CGG TTT TAC GGG CG (forward) and TCG AAC AGG AGG AGC AGA GAG CGA (reverse).
The significance of quantitative data was tested using the unpaired, two-tailed Student's T test.
|
10.1371/journal.pbio.1002026 | Specific Phosphorylation of Histone Demethylase KDM3A Determines Target Gene Expression in Response to Heat Shock | Histone lysine (K) residues, which are modified by methyl- and acetyl-transferases, diversely regulate RNA synthesis. Unlike the ubiquitously activating effect of histone K acetylation, the effects of histone K methylation vary with the number of methyl groups added and with the position of these groups in the histone tails. Histone K demethylases (KDMs) counteract the activity of methyl-transferases and remove methyl group(s) from specific K residues in histones. KDM3A (also known as JHDM2A or JMJD1A) is an H3K9me2/1 demethylase. KDM3A performs diverse functions via the regulation of its associated genes, which are involved in spermatogenesis, metabolism, and cell differentiation. However, the mechanism by which the activity of KDM3A is regulated is largely unknown. Here, we demonstrated that mitogen- and stress-activated protein kinase 1 (MSK1) specifically phosphorylates KDM3A at Ser264 (p-KDM3A), which is enriched in the regulatory regions of gene loci in the human genome. p-KDM3A directly interacts with and is recruited by the transcription factor Stat1 to activate p-KDM3A target genes under heat shock conditions. The demethylation of H3K9me2 at the Stat1 binding site specifically depends on the co-expression of p-KDM3A in the heat-shocked cells. In contrast to heat shock, IFN-γ treatment does not phosphorylate KDM3A via MSK1, thereby abrogating its downstream effects. To our knowledge, this is the first evidence that a KDM can be modified via phosphorylation to determine its specific binding to target genes in response to thermal stress.
| Histone methylation regulates gene expression and can have drastic consequences for health if the process is defective. Histone lysine demethylases (KDMs) counteract the activity of methyl-transferases and remove methyl group(s) from histones. KDM3A is a H3K9me2/1 demethylase that performs diverse functions via the regulation of its target genes, which are involved in spermatogenesis, metabolism, and cell differentiation. However, the mechanisms underlying KDM3A regulation of specific genes at specific times are largely unknown. Here we found that a physiological stress—elevated temperature—induces KDM3A phosphorylation in human cells via the MSK1 kinase. This phosphorylated form of KDM3A directly interacts with the transcription factor Stat1, which enables Stat1 to recruit KDM3A to Stat1-binding sequences at the promoters of specific target genes. KDM3A then acts to demethylate H3K9me2/1 at these targets, thereby causing specific gene expression in response to the thermal stress. We conclude that heat shock can affect the expression of many genes in human cells via a novel activation mechanism that is centered around the phosphorylation of KDM3A.
| Histone modifications, such as methylation and acetylation, regulate RNA synthesis [1],[2]. Unlike the activating impact of acetylation, the methylation of lysine residues in histones can exert either an activating or a repressive effect on genes, depending on the number of methyl groups that are added and the position of the lysine residue in the histone tail [3]. For example, the di- or tri-methylation of lysine (K) 9 on histone (H) 3 (H3K9me2/3), H3K27me2/3, and H4K20me3 is repressive, whereas that of H3K4me3 and H3K36me3 enhances the transcription of their target genes [4]–[6]. A major breakthrough in this field was the discovery that the methylation of histone tails is a reversible process. This discovery was based on the identification of two classes of histone lysine demethylases (KDMs), namely the FAD-dependent amine oxidase LSD1 [7] and the Jumonji C (JmjC) domain demethylases, a family of Fe2+- and 2-oxoglutarate-dependent KDMs [8].
Among the JmjC domain demethylases, KDM3A (also known as JHDM2A or JMJD1A) was first identified as a testis-enriched zinc finger protein that is highly expressed in male germ cells and is involved in germ cell development [9]. KDM3A was later identified as an H3K9me2/1 demethylase that activates the expression of the androgen receptor (AR) gene via an androgen-dependent pathway [10]. Furthermore, KDM3A has been demonstrated to regulate genes that are involved in spermatogenesis [11],[12], metabolism [13], and cell differentiation [14]. With such a broad functional diversity, the mechanism by which KDM3A regulates the appropriate gene(s) in vivo at the appropriate time and targets the appropriate element is of great interest.
Post-translational protein modification is very important for determining the function of proteins, including JmjC domain-containing proteins such as PHF8, which is phosphorylated by cyclin-dependent kinases (CDK), inducing the dissociation of PHF8 from chromatin [15]. PHF2 is enzymatically inactive in isolation, but PKA-phosphorylated PHF2 in complex with ARID5B displays H3K9Me2 demethylase activity [16]. PKCα–phosphorylated LSD1 forms a complex with CLOCK:BMAL1 to facilitate E-box-mediated transcriptional activation [17]. However, it is unknown whether KDM3A is phosphorylated, and the consequences of such a modification are also unknown.
In this study, we demonstrate that MSK1 is activated and specifically phosphorylates KDM3A at Ser264 under heat shock. The phosphorylated KDM3A (p-KDM3A) is enriched at the regulatory regions of gene loci and co-localizes with Stat1 in the human genome. Extensive experiments indicate that p-KDM3A directly interacts with and is recruited by Stat1 to mediate chromatin remodeling and the expression of its target genes in response to heat shock.
Histone modifications are recognized by specific proteins, including transcription factors (TFs), thereby mediating functional signaling to affect chromatin condensation or remodeling near target genes [2],[18],[19]. Methylated H3K9, a repressive histone mark, must be recognized and demethylated during the initiation of gene activation. Among the identified KDMs, KDM3A was the only KDM that targeted an IFNγ-activated sequence (GAS) in heat-shocked Jurkat cells (S1 Figure). Using an antibody against pan-phosphorylated serine (p-Ser) to detect the proteins immunoprecipitated for phosphorylated KDM3A, we found that KDM3A was phosphorylated after 30 or 60 min of heat shock at 42°C (the treatment of cells at 42°C for 60 min is generally defined as “heat shock” or abbreviated as “HS” in this study; it should be otherwise indicated when a shorter incubation time is applied) (Fig. 1A). This phosphorylation occurred within the first 661 aa of the N-terminus of KDM3A (Fig. 1B). Analysis of mutants in which serine was substituted with alanine at 264, 265, 445, and 463 aa of KDM3A revealed that only the S264A mutant abrogated the HS-induced phosphorylation of KDM3A (Fig. 1C). Next, we generated an antibody against a serine-phosphorylated peptide (cVKRK(p)SSENNG) and verified its efficacy via western blot (S2 Figure). Phosphorylated Ser264-KDM3A (p-KDM3A) was confirmed to be specifically induced under HS (Fig. 1D).
To explore the upstream kinase responsible for KDM3A phosphorylation under heat shock, mitogen- and stress-activated protein kinase 1 (MSK1) was considered as the most likely candidate because Jil1, the Drosophila ortholog of human MSK1, is activated in response to heat shock [20]. Because the activation of MSK1 can be identified based on its phosphorylation at S376 (p-MSK) [21], an antibody against p-MSK was used. An increased level of p-MSK was detected following extended incubation of the cells under HS (Fig. 1E). In co-IP assays with antibody targeting either MSK1 or KDM3A, co-IP of KDM3A and MSK1 in their phosphorylated forms was found only under HS. In contrast, the non-phosphorylated forms of MSK1 and KDM3A were unable to interact with one another under physiological condition (Fig. 1F). Furthermore, this interaction in heat-shocked cells was not affected by introducing either a dominant negative mutant of MSK1 or the S264A mutant of KDM3A (S3 Figure).
Next, we analyzed the specificity of activated MSK1 for KDM3A via an in vitro kinase assay using γ-32P-ATP to label the phosphorylated substrate. We demonstrated that only the GST-fused wild-type N-terminal KDM3A (1-394 aa), but not the S264A mutant (S/A), was phosphorylated by MSK1 based on 32P labeling (central panel of Fig. 1G). Then, MSK1 was incubated in the two GST-fused KDM3A protein fragments as described above, resulting in the specific phosphorylation of wild-type but not mutant KDM3A in vitro (Fig. 1H). Furthermore, we performed an in vitro kinase assay followed by mass spectrometric analysis to determine the specific target serine of MSK1 between the two successive serine residues at 264 and 265 aa in the synthesized KDM3A peptide (Fig. 1I). These in vitro data demonstrated that MSK1 specifically phosphorylates S264 of KDM3A.
To determine the effect of S264 phosphorylation on KDM3A, the demethylase activity of this enzyme was examined in vitro. However, no clear changes in the activity of KDM3A with or without S264 phosphorylation were detected (S4 Figure). Then, chromatin immunoprecipitation sequence (ChIP-seq) was performed to determine the global occupancy of p-KDM3A. Chromatin fragments were immunoprecipitated using an antibody against p-KDM3A from Jurkat cells subjected to HS (+) or not (-) or using a native KDM3A antibody from Jurkat cells not subjected to HS. A heat map containing more than 25,000 elements (gene promoters) was generated using seqMINER [22], and the results presented in four rows based on the antibody used and the heat-shock status. These elements were separated into three clusters, consisting of 12,719 elements in cluster 1 (top), 5,304 elements in cluster 2 (middle), and 7,120 elements in cluster 3 (bottom) (right panel, Fig. 2A). The MetaGene profiles indicated that the reads were enriched at the transcription start site (TSS) in cluster 1 genes, whereas both the TSS and the body of the genes were enriched in those of cluster 2 (top and middle, left panel, Fig. 2A). We analyzed all of the significant peaks in each sequencing sample using SICER V1.1 [23]. The percentages of the peaks of p-KDM3A that occupied the 2,700-MB mappable genome were 0.49% (HS-) and 0.42% (HS+), and their distributions across the genome are shown in a pie chart (Fig. 2B and S1 Table). The peaks were significantly enriched in the upstream regulatory region (approximate 10-fold, all p<1×10−100). By screening the differential SICER intervals near gene promoters (from −5 kb to approximately +2 kb) (FDR threshold 10−20), KDM3A and the non-treated or heat-shocked p-KDM3A target genes were identified, as shown in the Venn diagrams (Fig. 2C and listed in S2 Table). Gene Ontology (GO) and MSigDB Pathway analyses were performed on the target genes using GREAT 2.0.2 [24] (Fig. 2D and S5 Figure).
Next, we performed a TF motif analysis of the p-KDM3A-binding regions under HS using MEME [25],[26] and found that two of the three most common motifs (RGRAA and CSDGGA) correspond to Stat1-binding sites, indicating the genomic co-localization of p-KDM3A with Stat1 (Fig. 2E, S6 Figure, and S3 Table). Then, we determined the nearest gene locus in the top 68 sites of p-KDM3A binding that displayed the most significant difference between the HS and control conditions (S4 Table) to determine the binding peaks of p-KDM3A at four gene loci, DNAJB1, SERPINH1, SMIM20, and RNASEK, each of which is on a distinct chromosome in Jurkat cells (Fig. 2F, bottom panel). In addition, profiles of the Stat1-binding peaks in HeLa S3 cells treated with or without IFN-γ [27] were used as a reference (top panel).
To further illustrate the relationships between p-KDM3A occupancy and the expression of selected genes, ChIP-quantitative PCR (ChIP-qPCR) and reverse transcription quantitative PCR (RT-qPCR) were performed. The data demonstrated that the occupancy of p-KDM3A at all four gene loci examined (top panel, Fig. 2G) and the mRNA expression of all of these genes were enhanced under HS (bottom panel, Fig. 2G), suggesting a correlation between these two events in heat-shocked cells.
To determine the interaction between p-KDM3A and Stat1, we used antibodies targeting each protein to immunoprecipitate (IP) cell extracts for co-IP assays. We demonstrated that KDM3A and Stat1 interacted with one another only under HS (Fig. 3A). Based on a GST pull-down assay, MSK1 initially bound and phosphorylated KDM3A in vitro, but only p-KDM3A interacted with GST-Stat1 (Fig. 3B). By introducing S/A point mutations into KDM3A, we demonstrated that KDM3A-S264A, but not KDM3A-S265A, lacked this binding between KDM3A and Stat1 under HS (Fig. 3C), indicating that phosphorylation of KDM3A at S264 is critical for Stat1 binding. Next, we mutated S264 of KDM3A to aspartate (S/D) to mimic the phosphorylation of KDM3A at S264 in these cells. KDM3A-S/D co-immunoprecipitated with Stat1 even without HS (Fig. 3D), suggesting that although HS induces phosphorylation of both the Y701 and S727 residues of Stat1 [28], this phosphorylation was not required for Stat1 to interact with either p-KDM3A or KDM3A-S264D.
Then, we determined which region of Stat1 is required for its interaction with KDM3A-S264D in these cells. Among the Stat1 fragments S1, S2, S4, and S5 that interacted with KDM3A-S/D (Fig. 3E, top of right panel), the fragment S5 (residues 129-317, left panel) were the least required for this interaction. Based on GST pull-down assays, only the recombinant 1-394 fragment of KDM3A in its S264D form pulled down S5-Stat1 (Fig. 3F). Based on co-IP assays, HA-tagged Stat1 (129-317) interacted with full length S/D-KDM3A (Fig. 3G) and the shorter fragment S/D-KDM3A (214-306) (Fig. 3H), indicating that this 93-aa fragment of KDM3A interacts with Stat1. By performing another co-IP using an antibody against FLAG to detect FLAG-tagged KDM3A (214-306), we identified the 231-317 aa fragment of Stat1 was co-precipitated (Fig. 3I); this interaction between S264D-KDM3A (214-306) and Stat1 (231-317) was further confirmed in Fig. 3J.
Data from Fig. 1 and Fig. 3 revealed that p-MSK1 only interacted with p-KDM3A under HS, and p-KDM3A interacted with Stat1 even in its non-phosphorylated form. To address the detail correlations of MSK1, KDM3A, and Stat1 in heat-shocked cells, we further showed that p-MSK1 can be co-precipitated by a 214/306aa fragment of KDM3A under HS, suggesting a likely kinase versus substrate interaction for the phosphorylation of KDM3A at S264 (S7A Figure). Furthermore, the interaction of Stat1 and p-KDM3A was enhanced by extended incubation under HS, but not the interaction with p-MSK1 in the same cells and was not in the least enhanced (S7B Figure). However, the fact that the 93aa fragment of p-KDM3A could be co-precipitated by a 213/317aa fragment of Stat1 under HS indicates that the phosphorylated Y701 and S727 of Stat1 were not required for its interaction with p-KDM3A (Fig. 3J). Taken together, these results suggest these three factors do not exist in a complex, but sequentially take parts in the two functional stages: (1) activated MSK1 interacts and phosphorylates KDM3A-S264 under HS and (2) the recruitment of p-KDM3A via Stat1 to the promoter of target gene for HS inducing activation.
Next, we analyzed the MetaGene profile of p-KDM3A at the gene locus encoding hsp90α (hsp90aa1) under HS, which indicated the reads were enriched around the TSS of a cluster 1 gene. p-KDM3A under HS was markedly enriched at the TSS that is dominant over either non-heat shock p-KDM3A or non-phosphorylated KDM3A without HS (Fig. 4A). Interestingly, the p-KDM3A-enriched TSS region coincidently displays IFNγ-induced Stat1 binding at the hsp90α gene locus in HeLa S3 cells (Fig. 4A, top panel) according to Robertson et al [27]. Therefore, hsp90α is appropriately selected as a representative gene to further evaluate the mechanism underlying the targeting and functions of p-KDM3A in the human genome.
ChIP assays were then performed to examine the occupancy of p-KDM3A in the upstream sequences, its impact on the H3K9me2 level and in chromatin remodeling of hsp90α. We demonstrated that p-KDM3A was gradually enriched near the GAS element of hsp90α over time under HS (Fig. 4B), while the level of endogenous H3K9me2 decreased (Fig. 4C). This result suggests that p-KDM3A is directly involved in the demethylation of H3K9me2. Interestingly, once Stat1 was knocked down using a specific shRNA, the heat-shock-induced occupancy of p-KDM3A was abrogated in these cells (Fig. 4D), moreover, KDM3A-S/D mimic was no longer occupied even without HS (S8 Figure). In contrast, Stat1 binding remained following KDM3A knockdown (S9C Figure). ChIP/reChIP assays also demonstrated that p-KDM3A occupancy at the GAS element is Stat1-dependent (Fig. 4E). For DNase I hypersensitivity analysis, we set the sensitivity level without DNase I to 1.00 on the y-axis, representing a 100% “resistance” to this enzyme. As the amount of DNase I increased, the resistance to DNase I digestion significantly decreased in the upstream region of hsp90α in mock shRNA-transfected cells under HS (Fig. 4F, filled bars in left panel). In contrast, the HS-mediated changes in DNase I sensitivity at the GAS element were absent from KDM3A shRNA-transfected cells (Fig. 4F, right panel). Furthermore, in non-functional KDM3A H1120Y mutant (DN-KDM3A)-transfected cells [10], a similar profile lacking any clear changes in HS-dependent DNase I sensitivity was found (Fig. 4G). These data indicate that HS-mediated DNase I sensitivity at the GAS element is dependent on KDM3A demethylase activity. The HS-induced activation of hsp90α, as revealed by RT-qPCR analysis of its mRNA expression, was markedly reduced in KDM3A-knockdown cells (Fig. 4H) and in DN-KDM3A-transfected cells (Fig. 4I).
Jil1, the Drosophila ortholog of human MSK1, is activated in response to heat shock [20] and phosphorylates H3 to elicit chromatin relaxation, facilitating the binding of additional regulatory proteins [21]. In this study, we demonstrated that MSK1 is also activated in heat-shocked cells, as shown in Fig. 1E. To further address the detailed functions of MSK1 in KDM3A, we transfected the cells with either shRNA (i-MSK1) or a dominant negative (DN) mutant of MSK1; the phosphorylation of KDM3A at S264 under HS was blocked in these cells compared to the wild-type control cells (Fig. 5A and 5B and S10A–C Figure). However, similar to KDM3A knockdown, MSK1 knockdown did not affect the occupancy of Stat1 upstream of hsp90α (S10D Figure). i-MSK1 and DN-MSK1 also significantly impaired the mRNA expression of hap90α under HS (Fig. 5C), similar to the results using i-KDM3A and DN-KDM3A (Fig. 4H and 4I). These results indicate that MSK1 is the critical kinase that is responsible for the phosphorylation of KDM3A at S264 under HS. Then, we demonstrated that these reduced expression profiles in the presence of i-MSK1 and DN-MSK1 were based on a change in the occupancy of KDM3A at the GAS of hsp90α (Fig. 5D); a high expression level of H3K9me2 was detected (Fig. 5E). Furthermore, using the S264A mutant of KDM3A, the MSK1-mediated occupancy of KDM3A at the GAS was abolished (Fig. 5F), the expression levels of H3K9me2 remained elevated (Fig. 5G), and HS-induced mRNA gene expression was markedly reduced (Fig. 5H). In contrast, using the S265A mutant of KDM3A, identical results were obtained compared to wild-type KDM3A, as shown in the respective figures. Additionally, the importance of residue S264 of KDM3A was further demonstrated in KDM3A-S264A-transfected cells, which exhibited strongly reduced HS-induced DNase I hypersensitivity at the GAS region of hsp90α (Fig. 5I). It is, therefore, notable that the occupancy of p-KDM3A at GAS is required for KDM3A to display its demethylase activity on H3K9me2 and elicit chromatin remodeling at the GAS to activate the hsp90α gene.
MSK1 is a major kinase responsible for the phosphorylation of histone H3, including at S10 and S28 [29], and the phosphorylation of H3S10 facilitates the accessibility and transcriptional competence of a specific chromatin region in the genome [18],[30],[31]. Next, we demonstrated via western blot that the expression of phosphorylated H3S10 (p-H3S10) increased in heat-shocked Jurkat cells and was inhibited by transfection with specific MSK1 shRNA (Fig. 5J and 5K). A ChIP assay also verified the inhibitory effect of this shRNA on the occupancy of p-H3S10 at the GAS region under HS (Fig. 5L). In addition, the ChIP assay revealed that HP1α, the only HP1 isoform in the GAS region of hsp90α, is expressed at high levels preceding HS and reduced rapidly to minimal level within the first 30 min of HS treatment in Jurkat cells (Fig. 5M and 5N). Because the expression of p-H3S10 at the GAS was accompanied by an increase in acetylation of H3K9 but not H3K14 upon HS treatment [28], the phosphorylation of H3S10 by MSK1 may provide an open chromatin structure to recruit p-KDM3A via Stat1, thus facilitating the binding of additional regulatory proteins. This explained why the HS-induced DNase I hypersensitivity was severely impaired by the knockdown of MSK1 (Fig. 5O). Although the outcome elicited by MSK1 was similar with that of the KDM3A-S264A transfected (Fig. 5I), it may indicate that a novel aspect of MSK1 functioned on human chromatin remodeling under heat shock.
We previously reported that in contrast to HS treatment, IFN-γ treatment does not induce the expression of hsp90α or other related genes, such as CIITA-pIV, in Jurkat cells [28]. In this study, we demonstrated that p-KDM3A occupied at the GAS region of hsp90α (Fig. 4B), and its expression is efficiently induced under HS (Fig. 4H and 4I). IFN-γ did not induce the mRNA expression of this gene, independent of the presence of KDM3A in these cells (Fig. 6A). Unlike HS treatment, as shown in Fig. 1D and 1E, IFN-γ treatment did not induce the expression of MSK1 or activate the kinase activity of MSK1 (Fig. 6B), thus preventing the specific phosphorylation of KDM3A at S264 in IFN-γ-treated cells (Fig. 6C). These data indicate that only HS treatment activates MSK1 to phosphorylate KDM3A at S264, but this pathway is not activated in IFN-γ–treated cells. Therefore, we conclude that the expression level of p-KDM3A is the critical difference between the impact of HS and IFN-γ on the activation of their target genes in Jurkat cells.
To determine the mechanism by which p-KDM3A differentially functions in cells under different treatments, we transfected the cells with mutant KDM3A-S264D to mimic the phosphorylation of the critical S264 of KDM3A. We demonstrated that KDM3A-S264D occupied the GAS element of hsp90α either with or without HS treatment (Fig. 6D) and strongly reduced the H3K9me2 expression to the basal level (Fig. 6E). In contrast, hsp90αmRNA expression and DNase I hypersensitivity for the KDM3A-S264D mutant were similar to those for the wild-type enzyme under HS but not the control conditions (Fig. 6F and 6G).
Then, the aforementioned transfected cells were treated with IFNγ. The ectopically expressed KDM3A-S264D was efficiently recruited to the GAS region of hsp90α and the expression level of H3K9me2 was markedly reduced in the presence or absence of IFN-γ. However, wild-type and S264A mutant KDM3A did not bind to the GAS in IFNγ-treated cells and did not display any demethylase activity on H3K9me2 (Fig. 6H and 6I). Notably, KDM3A-S264D, but not the wild-type or S/A mutant counterparts, rendered hsp90α to be susceptible to IFN-γ treatment, as that shown under HS (Fig. 6J, slanted line-filled bars compared to the open bars).
The above results indicate that in untreated Jurkat cells, the ectopic KDM3A S/D mutant occupied the GAS and decreased the H3K9me2 level, but for an unknown reason, hsp90αmRNA expression was not induced. Therefore, we transfected wild-type and S/D mutant KDM3A into Jurkat cells to examine the occupancy of the Brg1 chromatin remodeling complex at the GAS before and after HS treatment or after IFNγ treatment. The ChIP data indicated that only when KDM3A-S/D was transfected did Brg1 efficiently occupy the GAS following both HS (Fig. 6K) and IFNγ treatment (Fig. 6L), but this binding was never constitutive at the GAS. However, transfected KDM3A and its S/A, S/D mutants did not affect Stat1 binding at the GAS (S11 Figure). This result agrees with our previous report that Brg1 is only recruited by p-Stat1 that is induced in response to HS treatment [28]. In IFNγ-treated cells, p-Stat1 also occupied the GAS [32], possibly providing a docking site for KDM3A-S/D and activating hsp90α. Therefore, it is conceivable that Stat1-mediated p-KDM3A recruitment is necessary but not sufficient for gene activation (Fig. 7). Our data indicate that the level of gene activation under HS or IFN-γ treatment is determined by the potential for an external stimulus to activate MSK1, which phosphorylates KDM3A. The two-step model in Fig. 7 shows that, first, MSK1-phosphorylated KDM3A is recruited by Stat1 to remove the repressive mark H3K9me2, and second, p-Stat1 mediates Brg1 complex recruitment to fully activate the target gene.
KDM3A is the second identified JmjC domain lysine demethylase (JHDM2A) that is specific for the demethylation of H3K9me2/me1. This demethylase contains a JmjC domain at 1058-1281 aa and a zinc finger domain at 662-687 aa [10]. Although certain TFs can induce KDM3A expression [13],[33]–[35] or interact with KDM3A [11],[14],[36], our understanding of the relationship between its modification and function has not been fully elucidated since its discovery.
In this study, we demonstrate that KDM3A is phosphorylated at S264 by MSK1 under heat shock. Specifically, S264 of KDM3A is approximately 400 residues from the N-terminus of the zinc finger domain, which performs no known function [10]. We then perform ChIP-Seq analysis to determine the genome-wide distribution of HS-induced p-KDM3A in Jurkat cells. To our surprise, ChIP-Seq data have shown that either with or without HS, the peaks of p-KDM3A could occupy the mappable genome at a comparable percentage. We then analyze the MetaGene profiles of p-KDM3A under HS, which shows the reads are enriched around the TSS at all of the five gene loci encoding the hsp90α (Fig. 4A) and the other genes (Fig. 2F); while those of the constitutive p-KDM3A only show much lower or minimal occupancy at these loci (the fourth versus the third rows in the bottom panels of Fig. 4 and 2). This finding suggests the p-KDM3As, either induced under stress (HS) or expressed in the normal life cycle of the cells, are functionally diverse through distribution to each distinct gene locus in the genome. In addition, the occupancy of p-KDM3A on Myo7B-Lims2 site is reduced under HS. The p-KDM3A in non-HS cells is likely phosphorylated by other kinase(s) or even the constitutively expressed MSK1 (Fig. 1E). These kinases can be activated via specific signaling pathway(s), such as IFNα [21], and exhibit their own function(s) on the specific constitutively expressed genes in the cells.
The TF motifs from ChIP-Seq data indicate that the p-KDM3A-bound sites are similar to those of some TFs, including Stat1. The phosphorylation of S264-KDM3A is a prerequisite for its efficient interaction with the TF Stat1, and residues 231-317 in the coiled-coil domain of Stat1 interact with the p-KDM3A in vitro. We suggest that this Stat1/p-KDM3A interaction represents a TF that directs KDM3A to an appropriate upstream element of its target gene to demethylate H3K9me2.
Because MSK1 is activated in response to a vast array of environmental stress stimuli via the p38 or ERK pathway to phosphorylate histone and HMG proteins [37],[38], MSK1 is involved in chromatin remodeling [21],[39]. We demonstrate that MSK1 is activated by HS but not IFNγ treatment and that p-KDM3A efficiently reduces the level of H3K9me2 at the GAS of hsp90α and renders this region susceptible to DNase I treatment.
Our data suggest that the p-KDM3A-mediated reduction in H3K9me2 expression is a major step of gene activation in Jurkat cells. Because no gene expresses efficiently in the presence of high level of H3K9me2 in Jurkat and Raji cells in response to either HS or IFNγ treatment (S12 Figure and ref. [28]). Hence the outcome of gene activation under HS or IFN-γ treatment is determined by the potential for the stimulus to activate MSK1 to phosphorylate KDM3A.
KDM3A-S264D was used in this study to mimic the function of phosphorylated KDM3A-S264 in vivo. We demonstrate that this S264D mutant directly interacts with Stat1 to occupy the GAS element regardless of heat shock. Although the KDM3A-S264D mutant constitutively binds to the GAS element, H3K9me2 remains at a basal level under IFN-γ treatment, similar to the results under HS treatment; in contrast, non-phosphorylated KDM3A does not interact with Stat1, is not recruited to the GAS element, and does not reduce the level of H3K9me2 when exposed to IFN-γ.
H1120 in the JmjC domain is indispensable for the demethylase activity of KDM3A [10]. However, the phosphorylation of KDM3A-S264 exerts the same effects, including H3K9me2 reduction and DNase I hypersensitivity at Stat1 target genes. Therefore, it is logical to propose that the Stat1-mediated recruitment of the p-KDM3A represents a specific pathway by which the demethylase activity of KDM3A is regulated under heat shock.
In summary, heat shock is a physical stimulus that broadly affects the expression of a variety of genes in human cells, likely in a general manner. In addition to the activation of the well-accepted heat shock factor and heat shock element (HSF/HSE) pathways to induce expression of heat-shock-related genes, we present a novel, generalized heat-shock-induced activation mechanism that is centered on the phosphorylation of KDM3A. (1) p-KDM3A-S264 is enriched genome-wide at the promoter region of several genes, including heat-shock-related genes, under heat shock; (2) p-KDM3A is guided by a TF to the binding element of TF in the genome; (3) the genomic occupancy of p-KDM3A at its target genes is a prerequisite for the demethylase activity of KDM3A in situ; and (4) the phosphorylation of KDM3A is specifically dependent on the upstream stimulus-dependent kinase activity of MSK1 in HS- but not IFN-γ-treated Jurkat cells.
Antibodies against KDM3A, p-MSK1, GAPDH, H3K9me2, and H3K9me3 and recombinant activated MSK1 were purchased from Millipore Biotech (Billerica, MA, United States). The FLAG and M2 antibodies were purchased from Sigma. The GST, MSK1, MSK2, HA, and Stat1 antibodies were purchased from Santa Cruz Biotechnologies (Santa Cruz, CA, US). The anti-phosphorylated serine (p-Ser) (antibody catalog number AB1603) was purchased from Merck (Darmstadt, Germany). A specific antibody against p-S264-KDM3A was produced by Beijing B&M Biotech (Beijing, China) using the synthesized peptide VKRKSSENNG, corresponding to residues 260–269 of KDM3A, as an antigen.
The FLAG-tagged MSK1 eukaryotic expression plasmid was constructed by cloning MSK1 into the pcDNA6-FLAG vector using a PCR product from a Jurkat cell cDNA library. We inserted point mutations at amino acids 165 (D to A) and 565 (D to A) in full-length FLAG-MSK1 to produce DN-MSK1 [40]. The FLAG-tagged KDM3A eukaryotic expression plasmid was a gift from Dr. Zhong-Zhou Chen of China Agricultural University. We inserted a point mutation at amino acid 1120 (H to Y) to produce DN-KDM3A [10], and we generated five individual point mutants of KDM3A: S264A, S265A, S445A, S463A, and S264D. The KDM3A fragment from 214-306 was subcloned using the PCR product of full-length FLAG-KDM3A. The MSK1 and KDM3A shRNA oligonucleotide sequences were designed by OriGene Technologies, Inc. (Rockville, MD, USA) and inserted into the HindIII/BamHI site of the pRS vector. shRNA-Stat1 was purchased from OriGene Technologies, Inc. The truncation mutants of Stat1 (S2 and S4-S6) were described previously [28]. A new construct of S3 (317–750 aa) was subcloned using the PCR product of full-length HA-Stat1 (S1). We constructed Stat1 (129–235) and Stat1 (231–317). The primers that were used to generate the MSK1, KDM3A, and Stat1 mutant plasmids are listed in S5 Table.
RT-qPCR was performed as described previously [41],[42]. The relative expression levels of DNAJB1, SERPINH1, SMIM20, RNASEK, and HSP90AA1 (hsp90α) were normalized to those of GAPDH using the comparative CT method according to the manufacturer's instructions (Rotor-Gene RG-3000A Real-Time PCR System, Corbett Research, Australia). The specific primers corresponding to the above genes are listed in S6 Table. The experiments were repeated at least three times, and statistical analysis was performed on the individual experimental sets. All of the values in the experiments are expressed as the means ± SD.
The ChIP assays were performed as described previously [41],[42]. The primers used for DNAJB1, SERPINH1, SMIM20, RNASEK, and HSP90AA1 (hsp90α) are listed in S7 Table. The percentage of ChIP DNA relative to the input was calculated and expressed as the mean ± SD of three independent experiments [43].
For ChIP-reChIP analysis [28], first, Jurkat cells were transiently transfected with FLAG-tagged Stat1 expression plasmids prior to further treatment. The chromatin fragments from the sonicated cells with or without HS treatment were used as the input, which was then immunoprecipitated using an anti-Flag M2 affinity gel (F1). Aliquots of the F1 chromatin fragments were reverse cross-linked to obtain DNA for qPCR assays or were saved for re-IP using an antibody against KDM3A or p-KDM3A for reChIP assays (F2). The DNA that was extracted from the chromatin fragments subjected to reChIP was re-amplified using the primer sets used for qPCR. The amount of KDM3A or p-KDM3A that was recruited by the antibody against Stat1 at 42°C was quantified relative to that recruited at 37°C, which was normalized to 1.
For ChIP-Seq, the chromatin fragments of 1×107 Jurkat cells with or without HS treatment were immunoprecipitated using IgG or an antibody against KDM3A or p-KDM3A. The DNA fragments were end-repaired, adenylated, ligated to adaptors, and PCR-amplified for 18 cycles. The PCR products corresponding to bp 250-450 were gel-purified, quantified and stored at −80°C until use for sequencing. For high-throughput sequencing, the libraries were prepared according to the manufacturer's instructions, and to the samples were analyzed using an Illumina GAIIx system for 80-nt single-end sequencing (ABLife, Wuhan, China).
The data were analyzed using Active Motif; the flow chart of analysis is shown in S13 Figure. After removing the adaptors and low-quality bases, the reads (36 bp in length) were mapped to the human genome (hg19) using the BWA algorithm with the default settings. The clean reads that passed through the Illumina purity filter and aligned with less than two mismatches and without duplicates were saved as BED files for use in subsequent analyses. The mapped reads were inserted into seqMINER to obtain the Meta Gene distribution profile, and the genes were distributed into three clusters based on their distribution profiles. The reads files were converted to Wig files, which were inserted into the IGV 2.3 Genome Browser with the peak height set at 4–24 to determine the peak binding profiles.
For peak calling, the mapped BED files were inserted into SICER V1.1 [23] (estimated false discovery rate [FDR] threshold = 1×10−10; window size: 200 bp; fragment size: 200 bp; gap size: 200 bp; hg19 genome database) and MACS 1.4.2 (p-value cutoff = 1×10−7; tag size: 36 bp; band width: 150 bp; model fold = 8, 24) [44] using the pooled input (control/heat shock) and IgG experiment reads files as backgrounds. The NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) accession number for the ChIP-seq data is GSE62309.
The GO and MSigDB Pathway analyses were conducted using GREAT 2.02 on the SICER intervals data limited to the regulator regions (from −5 kb to approximately +2 kb of the TSS). The pathway analysis database in GREAT is the MSigDB from the Gene Set Enrichment Analysis. The binomial p-value reflects the significance of the targeted genes enriched in a GO term.
To identify the genome sites with more p-KDM3A after heat shock, we used the p-KDM3A HS (+) MACS interval peaks in Active Regions (in locations where only one sample had an interval, which defines the Active Region) to perform a sample comparison with peak metrics against the p-KDM3A HS (−). The unique intervals were annotated into genes (between 10 kb upstream and 10 kb downstream). The GO analysis of these genes was described above.
Transcription factor motifs were identified around p-KDM3A SICER islands (FA files) after heat shock using MEME (version 4.9.1) [45]. The database JASPAR_CORE_2014_vertebrates was used.
Jurkat cells were transiently transfected with shRNA-MSK1 or shRNA-KDM3A. A total of 1×107 cells were washed twice in PBS, and the nuclei were extracted as described above and digested with DNase I (ranging from 0 to 80 units/ml) on ice for 10 min. The DNase I digestion was terminated by incubating in stop buffer (Promega, M6101) at 65°C for 10 min. Then, the nuclei were digested with 50 µg/ml RNase A at 37°C for 60 min and 50 µg/ml proteinase K at 50°C overnight. The genomic DNA was purified via phenol/chloroform extraction and ethanol precipitation [46],[47]. Aliquots of 10 µg DNA were purified for qPCR using the primers described for the ChIP-qPCR assays.
The GST-Stat1 fusion protein was expressed in Escherichia coli (BL21 DE3) and purified using glutathione-sepharose. GST and GST-Stat1 were bound to glutathione-sepharose, and 10 µl packed beads containing 5 µg the GST or GST-Stat1 fusion protein were incubated in the product of the kinase assay for MSK1 and KDM3A. After overnight incubation at 4°C, the beads were washed three times, and the bound proteins were analyzed via western blot.
The Co-IP analyses were performed using approximately 500 µg protein samples that were incubated in a specific antibody for 2 hr at 4°C. In total, 20 µl Protein A (or G)-agarose were added, and the samples were incubated at 4°C overnight. Then, the pellets were washed with RIPA buffer, followed by the addition of 40 µl 1× Laemmli buffer. Then, the samples were resuspended and boiled. The samples were separated via SDS-PAGE and analyzed via sequential western blot using individual antibodies [48].
Recombinant MSK1 (Millipore Biotech) was incubated in 1 µg purified wild-type or mutant KDM3A (1-394) in the presence of 50 µM ATP or 5 µCi [γ-32P]ATP in kinase buffer (10 mM Tris, pH 7.4; 10 mM MgCl2, 150 mM NaCl) for 30 min at 30°C. The reaction products were resolved via SDS–PAGE for western blot using specific antibodies; alternatively, the 32P-labeled proteins were visualized via autoradiography. Recombinant MSK1 was incubated in 1 µg of the synthesized peptide cVKRKSSENNG, corresponding to residues 260-269 of KDM3A, in the presence of 50 µM ATP in kinase buffer for 30 min at 30°C. The reaction products were purified for mass spectrometric analysis (Institute of Microbiology, CAS, China). Recombinant MSK1 was incubated in full-length GST-KDM3A for the kinase assay; then, 2 µg histone from HeLa cells was added to demethylation buffer (50 mM Tris, pH 8.0, 50 mM NaCl, 2 mM L-ascorbic acid, 1 mM α-ketoglutarate, 50 µM Fe(NH4)2(SO4)2) at 37°C for 2 hr, and the reaction was terminated by adding SDS-PAGE loading buffer. The results were analyzed via western blot using specific antibodies.
The numerical data in all figures are included in S1 Data.
|
10.1371/journal.pntd.0000801 | Impact of Increased Economic Burden Due to Human Echinococcosis in an Underdeveloped Rural Community of the People's Republic of China | Ningxia is located in western People's Republic of China, which is hyperendemic for human cystic echinococcosis (CE) throughout the entire area with alveolar echinococcosis (AE) hyperendemic in the south. This is in part due to its underdeveloped economy. Despite the recent rapid growth in P.R. China's economy, medical expenditure for hospitalization of echinococcosis cases has become one of the major poverty generators in rural Ningxia, resulting in a significant social problem.
We reviewed the 2000 inpatient records with liver CE in surgical departments of hospitals from north, central and south Ningxia for the period 1996–2002. We carried out an analysis of health care expenditure of inpatient treatment in public hospitals, and examined the financial inequalities relating to human echinococcosis and the variation in per capita income between various socioeconomic groups with different levels of gross domestic product for different years. Hospital charges for Yinchuan, NHAR's capital city in the north, increased approximately 35-fold more than the annual income of rural farmers with the result that they preferred to seek health care in local county hospitals, despite higher quality and more efficient treatment and diagnosis available in the city. Household income levels thus strongly influenced the choice of health care provider and the additional expense impeded access of poor people to better quality treatment.
Information on socioeconomic problems arising from echinococcosis, which adds considerably to the burden on patient families and communities, needs to be collected as a prerequisite for developing policies to tackle the disease in rural Ningxia.
| This paper compares medical expenditure for hospital treatment of echinococcosis in NHAR, western People's Republic of China, for different years, different regions and different socioeconomic groups. The results show that the level of household income strongly influences health care decisions. This study represents an effort to determine the effect of hospital charges for inpatient treatment of echinococcosis on the choice of provider in NHAR, and quantitatively examines this topic for the rural poor. The findings show that low income individuals from rural areas opted to visit a local county hospital rather than an urban hospital for hydatid surgery despite the inferior infrastructure, personnel and general health care facilities available. There are a number of policy implications. For example, enhancing the quality and service of county hospitals in rural areas will benefit those with lower incomes, thus improving access of rural residents to health facilities for higher quality diagnosis and efficient treatment. Thus, we advocate that government policy should be to increase investment in health care in poor rural areas, and to launch relevant medical aid projects to help those in poverty.
| Since the inception of market reforms in the early 1980s, the annual health expenditure of People's Republic of China (P.R. China) has increased consistently [1], [2]. But, contrary to this increase, two national healthcare surveys [3], [4] showed that health insurance decreased from 30.2% of coverage in 1993 to 23.6% in 1998. Moreover, the rural insurance system has almost collapsed [5]–[8]. Thus, the majority of rural residents have to pay all their medical expenses personally and are therefore ‘out-of-pocket’ for any health services they require. The rapid escalation of medical costs, largely due to the frequent use of advanced medical technologies and over-prescription of drugs by health care providers [1], [9], accompanied by a lack of insurance coverage, has inevitably caused severe financial hardship for many households in P.R. China. This is particularly so for low income rural families. In this way, poor health is an important source of transient poverty.
Ningxia Hui Autonomous Region (NHAR) is an underdeveloped provincial autonomous region in western P.R. China [10]. Human cystic echinococcosis (CE) is hyperendemic throughout NHAR whereas CE and alveolar echinococcosis (AE) are co-hyper-endemic in the southern areas [11], due, in part, to poverty and poor economic development. Sheep farming is the main source of income due to the prevailing socio-religious Islamic culture and the favourable conditions for growing crops and breeding livestock [11], [12]. This article focuses on the analysis of health care expenditure from public hospital records in NHAR, and examines the health and financial inequalities relating to human echinococcosis. It also examines variation in per capita income between various socioeconomic groups and different regions with different levels of gross domestic product (GDP) for NHAR. The purpose of this study was to gain some insight into the inequality in health services and their utilization, and health care expenditure through detailing hospital charges and average personal income in NHAR, so as to inform policy development, a pre-requisite for better understanding the interactions of poor households with health systems in different contexts so as to promote equitable and universal access to basic health care.
NHAR is one of five provincial level autonomous regions in P.R. China. With a total area of 66,400 Km2, the Hui, one of the officially recognized nationalities of P.R. China, make up 34% of the total population of 5.9 million. NHAR is located in the middle reaches of the Yellow River in northwest P.R. China, and it has five regional prefecture municipalities or cities: Yinchuan and Shizuishan in the north, central Wuzhong and Zhongwei, and Guyuan in the south (Fig. 1). NHAR is mostly deserted and is sparsely settled, but the vast plain of the Yellow River in the north has been irrigated for centuries; over the years an extensive system of canals has been built. Desert and grazing land make up most of the area. Extensive land reclamation and irrigation projects have increased cultivation. The fertile plain is ideal for the cultivation of fish and favorable agricultural conditions for growing crops and breeding livestock in north and central NHAR. This greatly benefits the northern Yellow River regions, notably Wuzhong municipality, including Qingtongxia, Wuzhong City and Zhongwei county, as well as the capital city Yinchuan, which is the economic and cultural centre, and is the most developed region in NHAR. This is in direct contrast to south NHAR (Fig. 1) which is mountainous with low production levels and poor economic development. This variability in natural environment and resources in the different geographical areas has led to economic disparity between south, central and north NHAR (see Tables 1 and 2). There are two geographical and administrative centres in NHAR: Guyuan in the south and Yinchuan in the north. The hospitals in Guyuan (The Second Provincial Hospital of NHAR) and in Yinchuan are responsible for patients from the whole of NHAR and receive financial support from both the local and central NHAR governments.
The data used in this study mainly involved two components. The first component comprised regional/local economic data for different years and socioeconomic groups and population demographic searches by locality; these data were compared with relevant hospital data. The second component included hospital inpatient records, particularly hospitalization charges, which were collected for all echinococcosis patients who had previously registered in a public hospital in NHAR. The full details of the data sources are described in the following sections.
The searched data included local population demographic status (using census data); GDP per year; per capita GDP and per capita income and/or per capita deposited income by locality; and basic livelihood proportions of per capita expenditure of households in various socioeconomic population groups. Available statistical data for NHAR [13] were used to calculate the average per capita income level for different socioeconomic groups/localities by comparison with the average entire rural GDP level (Table 3), which obviously reflects the living standard levels of different socioeconomic groups. All other data used in this study were obtained from searches of locally available information cited in the references [13]–[16]. The search methods for obtaining the sourced data included electronic publications of the local NHAR literature, annual local government reports/documents and social science abstracts for NHAR.
We reviewed the records of 2000 patients from 24 public hospitals in 20 administrative districts across the whole of NHAR [17]. The hospitals included public hospitals in 18 of these districts, two in Guyuan and four in Yinchuan. All patients, between 1996 and 2002, underwent surgery for removal of liver hydatid cysts in the surgical departments of these public hospitals in north, central and south NHAR (Fig. 1). Data were extracted on sex, age, occupation, nationality and domicile, duration of hospitalization, cost of each procedure for diagnosis and treatment, and cost of hospital stay.
The extent of urban-rural income inequality can be manifested by the real per capita urban-to-rural consumption ratio as shown in Table 3. Consumption is considered to be a better measure of living standards than income [18].
The costs of hospital charges for hydatid surgery mainly included general expenses (hospital stay, staff, and medications), surgical costs (surgery fee, costs for anesthetization and monitoring, and other expenditure, such as oxygen, blood transfusion costs, emergency equipment and essential drugs). Other components of the hospital charges were radiology/imaging examinations, laboratory testing, and exploratory surgery/post-surgical nursing and prescription charges.
The economic consequences of illness and costs for health care and treatment in public hospitals in various regions of NHAR were compared across different socioeconomic groups and also by their geographic location.
Financial estimates were based on the actual costs of the treatment of echinococcosis cases in public hospitals in various regions in NHAR. A comparison of increased charges for hospital treatment between different years (comparing 2002 with1998) and between different regions (different locations with differing economic levels) was also carried out.
The GDP is a common measure of welfare for socioeconomic development. Generally, per capita income estimates are based on the present value of income/earnings. The deposited per capita income was chosen in this study to be a financial measurement for three reasons: (i) it is an after-tax income for the individual; (ii) it already includes both non-labour and labour income; (iii) it incorporates in-kind incomes as well.
The ‘per capita’ income and household expenditure for education, health care, food, food for livestock and purchase of chemical fertilizer for crops, were compared to estimate livelihood levels. The cash income per person increased as household income increased, as also did expenditure on education, health care and food, suggesting that using per capita income/deposited income can reflect real differences in the income levels of households (see Table 3, data of livelihood expenditure weights from the 2004 Statistics Year Book for NHAR) [13]. Therefore, the direct cost burden is a measure of the proportion of per capita income that can represent the household burden in most rural families in NHAR. In this study, the direct cost of hospitalization for echinococcosis cases was measured as the proportion of household expenditure that was spent on health care, indicating the financial burden placed on the household by the cost of seeking treatment.
The mean, standard deviation (SD), and the percentiles were calculated for general descriptions of inpatient records, including demographic data for age, occupation, clinical information for liver CE lesion size/numbers, and data for length of hospital stay, which included length of pre- and post-surgery stay. An analysis of hospital charges comparing per capita income/per capita deposited income was carried out to demonstrate what proportion of average individual income was used for hospitalization costs. This can reflect the average economic burden in various socioeconomic groups of people. Pearson's chi-square test for independence was used in order to test differences between years, locations and people-groups. The level of statistical significance was set at P = 0.05 unless otherwise stated. All statistical analyses were performed using SPSS 13.0 software (SPSS Inc. Chicago, III, USA).
Ethical clearance for the study was given by the Ningxia Medical University Ethics Committee and The University of Queensland Ethics Committee, and sanctioned by local hospital representatives.
Farmers (76%) were the main group of echinococcosis patients, followed by students (12.4%), workers including those self employed (5.2%), cadres (4.8%) and others (2%). Females outnumbered male patients with a ratio of 1.38. For 2000, the population ratio (0.95 females to males) [14] showed a significantly higher morbidity in females for echinococcosis (P<0.01). In south NHAR, Hui Chinese accounted for 55% of patients which matched the population composition. For north and central NHAR, 29% patients were Hui, again similar to the population composition ratio in these areas. Age at diagnosis ranged from 4–78 years (mean ± SD, 37.0±16.9) in the south, and from 6–78 years (mean ± SD, 40.0±17.0) in the north and central regions, showing a slight younger diagnosis/age in the south.
In Yinchuan, 57.7% patients came from other counties compared with the small number (28.7%) of patients from the Yinchuan area; of those who domiciled in areas other than Yinchuan, the majority (63.0%) came from central NHAR with a minority from the north (20.0%) and south (17.0%). Patients from the Guyuan hospital records were domiciled almost equally from Guayuan and from areas out of Guyuan. Among the latter, the majority (96.3%) came from various counties belonging to Guyuan prefecture in south NHAR, and were administered by the prefecture government, with a minority from north (1.0%) and central (2.7%) NHAR (Table 4).
Small hydatid cysts (≤5 cm diameter) accounted for 10–12%, middle-sized cysts (5–10 cm) accounted for 50–52%, large cysts (≥10 cm) accounted for 33–37% and very large cysts (≥20 cm) accounted for 0.3–1.8% of echinococcosis patients. The majority of patients (81%) had single cysts, whereas15% had two cysts, and 4% had three or more cysts.
The duration of hospitalization pre-surgery ranged from 0–65 days (mean ± SD; 5±5 days), the number of post-surgery hospitalisation days ranged from 1–63 (mean ± SD; 10±6 days), and the average number of hospitalization days was 14±8 (mean ± SD) for all pooled records. The major reason for prolonged hospitalization prior to surgery was due to supportive treatment to improve the general health condition of the patient or for reducing the surgical risk due to other disorders, such as co-infection. Prolonged hydatid cyst drainage was the major cause of prolonged stay post-surgery. On occasions, some patients stayed for less than one day pre-surgery due to an emergency such as anaphylactic shock caused by cystic rupture which required immediate surgery.
The individual components of hospital charges were compared for echinococcosis case treatment in the rural (those county hospitals in the south) and urban (four public hospitals in Yinchuan city) hospitals (Table 5). The general expenses (hospital stay/bed fee) which accounted for 5.5% of the total for rural hospitals and 11.4% in Yinchuan, and other charges that mainly indicated costs for occupation of a hospital room by accompanying family members, or renting toilet containers for the patient, accounting for 6.4% costs in the rural areas and 3.7% in Yinchuan. The comparisons between rural and Yinchuan hospitals showed the differences were significant (P<0.05). None of the other costs showed significant differences though different percentage values were apparent when the comparisons were carried out between the rural and Yinchuan hospitals. These included the costs of surgery (36.6% in rural versus 42.1% in Yinchuan); of routine clinical laboratory tests (1.1% versus 2.2%); of medical/physical checks and pathology examinations (1.5% versus 1.8%); of provision of drugs (31.4% versus 22.0%); and of post-surgery nursing (4.1% versus 3.7%). There were also similar costs between the rural and the urban Yinchuan hospitals when a comparison was made of imaging charges for X-ray, ultrasound, and computerized tomography (CT)/magnetic resonance imaging (MRI) scanning (1.7%, 2.2% and 9.2–9.5%), respectively.
The different disease burden, caused by charges for hospitalization, was measured by the comparison of the different proportions of hospital-charges per capita GDP averaged between different years, between different locations and/or between different socioeconomic groups of people. A comparison (Table 1) of inpatient hospital charges for echinococcosis cases in different areas between 1998 and 2002 showed that these had increased significantly. However, a comparison of the increased ratio of local GDP with the increased ratio for the average local hospital charge for 1998 and 2002, showed they were the same for rural counties (1.3 versus 1.3) and Guyuan in the south (2.0 versus 2.0). In contrast, hospital charges increased more than the per capita GDP in Wuzhong, central NHAR in 2002 compared with 1998 (1.9 versus 1.5); for Yinchuan, the ratio of hospital charges decreased compared with the increase in GDP over this period (1.6 versus 2.3). Due to their geographic location and administrative seniority, Guyuan and Yinchuan Hospitals generally act as the centres for the health care of all people in NHAR. However, Guyuan city hospital charges were 1.3 times higher than local counties in 2002 compared with 0.9 times in 1998; in comparison, Yinchuan city hospital charge were 1.9 times higher than local county hospitals in 2002 and 1.7 times higher in 1998 (Table 1).
A comparison of costs for echinococcosis inpatients in different hospitals in 2002 with the local income levels at the same time (Table 2) showed the local health care costs (surgery or hospitalization charges) accounted for 32% of the per capita income and 54% of the deposited income in Yinchuan in the north, but accounted for 50%, 61% and 65% per capita income and 103%, 133% and 147% deposited income in Wuzhong, Guyuan and local counties in south NHAR with low GDP, respectively. Based on the charges at various level hospitals, we calculated the costs for a poor farmer from a rural area seeking inpatient healthcare at different levels of the public hospital system. The gap between income and hospital payment accounted for 163% of per capita income and 1800% of deposited income if a poor person sought treatment at a local county hospital; it accounted for 208% of per capita income and 2285% of deposited income if sought in Guyuan (defined as a rural area), and 318% of per capita income and 3495% of deposited income if sought in Yinchuan (defined as an urban area) (Table 2).
Although average incomes have risen recently in south NHAR, health care costs have also increased considerably. The rise in the average cost of an admission to a county hospital in south NHAR was similar to the growth in average income, while the average cost of admission in Yinchuan, the capital city of NHAR, in the north, has become relatively more expensive. Evidence for increased disease burden, particularly for rural patients, was demonstrated by the fact that the average hospital charge in Yinchuan was 1.7 times higher than that for a rural county hospital in 1998 and this rate increased to 1.9 times in 2002; the rate of charge differences for Guyuan versus rural county hospitals also increased from 0.85 to 1.26 times (1998–2002) (Table 1). Therefore, the first choice for clinical consultation and hospitalization for most of the rural population in NHAR is the local county hospital because of its closer proximity and the lower health care costs charged.
This study has provided clear evidence that the cost of hospitalization imposes a severe burden on many rural households, particularly poorer families, in NHAR. Households had to spend a large proportion of their annual income on medical costs. Substantial numbers of rural households had to borrow money to pay hospital fees, because the payments were greater than their annual income. This situation is similar to other settings [15] where the most common reason for not seeking hospital inpatient treatment is financial hardship. Such financial problems restrict poor people access to quality health care. The most severe effects are on those who did not seek hospital treatment because of their inability to afford treatment and consequently their sickness goes untreated. Such people are at risk of further suffering and deterioration in health [16], [19]. In addition, some patients may have limited access to professional health care services, poor access to drugs, are only able to purchase drugs from cheaper sources, or seek treatment at cheaper private clinics where the health care facilities generally are limited [20]. This situation not only generates an unhealthy, irrational use of drugs, but also wastes scarce financial resources. Due to the financial constraints, poor people delay seeking care until an emergency situation arises, but this delay often forces them to seek care at a more expensive level, typically in hospitals. This may also be a major reason for our previous report of hospitalized patients with echinococcosis having more severe disease [16]. There are two main reasons why only a low percentage (17.0%) of patients from poor rural areas in south NHAR attends hospitals in Yinchuan. They may either be unwilling to pay transportation charges or the more expensive hospital fees compared with those of the local county hospitals are prohibitive. The facilities in the county hospitals are generally of poorer quality and diagnostic procedures and treatment are less efficient than in urban hospitals. Household income levels strongly influenced the choice of hospital and the additional expense impeded access of poor people to better quality health care.
Hospitals in rural P.R. China rarely provide non-medical care (e.g. food, drinking water, toilet facilities) for inpatients who are cared for by their relatives, generally farm labourers. In addition, a sick household member is unable to work or do heavy tasks for long periods of time; prolonged illness and recuperation can therefore influence production and income for one or more production cycle (eg. annual/seasonal harvest), which increases the long-term financial burden on the household.
Rural families are faced with a range of disease problems, but the need for medical help clearly exceeds their ability to pay for treatment. This inevitably results in poor livelihood and financial difficulty. People use a variety of ways to finance unforeseen health expenditure; they use savings or sell consumables, borrow money or purchase health care on credit, often leaving them with substantial debt [19]. The families of many patients cut down on food to offset the cost of borrowing, even sacrificing investment in future productivity, such as withdrawal of their children from school to support the family through labour, and to save school fees [20], [21]. This inevitably triggers a vicious circle of impoverishment and causes more debt [22] because these poorly educated children will obviously be part of low-income socio-status groups in the future. The high cost that poor rural people have to pay to health providers for treatment is an unfortunate path that leads this group from illness to poverty.
Economically vulnerable people have little capita assets or savings to withstand any short-term reduction in income without falling quickly into poverty. So, any disaster, such as illness or an accident, requiring time off work, may bring about a financial crisis leading to a rapid deterioration in livelihood. Therefore, poverty resulting from illness, due especially to human echinococcosis, has become a significant social problem in rural NHAR [10].
The costs for a hospital bed and the period spent in hospital for hydatid surgery and treatment were higher in urban than rural hospitals in NHAR. Reduction in these costs by hospital management would allow the limited financial resources to be used for diagnosis and treatment of cases. As well, the socioeconomic problems arising from echinococcosis which add to the burden on patients' families and the community were not taken into account in the current calculations. This additional information needs to be collected as a pre-requisite for developing more comprehensive government policies to tackle the burden of echinococcosis in rural NHAR. We advocate that the local and central Chinese governments should increase investment in health care generally in poor rural areas, and launch relevant medical aid projects to help those in poverty, and improving the equity of the health care system. In order to improve access of rural residents to health care services, it is important to locate health facilities and personnel rationally, thereby reducing the financial commitment and distance as obstacles to better health care.
Hospital costs for each hydatid disease surgical case were the best assessment of the average expenditure per patient for NHAR hospitals although the total illness burden costs will be underestimated. This study has some limitations in that it is hospital-based, in which other health care costs, such as consultation fees, transport, food costs, and the convalescence, rehabilitation, and the social consequences of disability have not been included. In addition, since this set of data was obtained from retrospective hospital records, the economic status of each patient could not be obtained. Therefore, the financial burden analysis was based on comparisons between the average of the aggregate patient group charges and the average local level of per capita income. Such aggregated correlations only reflect information at the group level, and do not target an individual patient. However, taking a population level perspective into account, this study provides important information for future policy development to tackle the public health issues for this rural population, and elsewhere, where the situation may be comparable to that in rural P.R. China.
|
10.1371/journal.pcbi.1004486 | A Dynamic Gene Regulatory Network Model That Recovers the Cyclic Behavior of Arabidopsis thaliana Cell Cycle | Cell cycle control is fundamental in eukaryotic development. Several modeling efforts have been used to integrate the complex network of interacting molecular components involved in cell cycle dynamics. In this paper, we aimed at recovering the regulatory logic upstream of previously known components of cell cycle control, with the aim of understanding the mechanisms underlying the emergence of the cyclic behavior of such components. We focus on Arabidopsis thaliana, but given that many components of cell cycle regulation are conserved among eukaryotes, when experimental data for this system was not available, we considered experimental results from yeast and animal systems. We are proposing a Boolean gene regulatory network (GRN) that converges into only one robust limit cycle attractor that closely resembles the cyclic behavior of the key cell-cycle molecular components and other regulators considered here. We validate the model by comparing our in silico configurations with data from loss- and gain-of-function mutants, where the endocyclic behavior also was recovered. Additionally, we approximate a continuous model and recovered the temporal periodic expression profiles of the cell-cycle molecular components involved, thus suggesting that the single limit cycle attractor recovered with the Boolean model is not an artifact of its discrete and synchronous nature, but rather an emergent consequence of the inherent characteristics of the regulatory logic proposed here. This dynamical model, hence provides a novel theoretical framework to address cell cycle regulation in plants, and it can also be used to propose novel predictions regarding cell cycle regulation in other eukaryotes.
| In multicellular organisms, cells undergo a cyclic behavior of DNA duplication and delivery of a copy to daughter cells during cell division. In each of the main cell-cycle (CC) stages different sets of proteins are active and genes are expressed. Understanding how such cycling cellular behavior emerges and is robustly maintained in the face of changing developmental and environmental conditions, remains a fundamental challenge of biology. The molecular components that cycle through DNA duplication and citokinesis are interconnected in a complex regulatory network. Several models of such network have been proposed, although the regulatory network that robustly recovers a limit-cycle steady state that resembles the behavior of CC molecular components has been recovered only in a few cases, and no comprehensive model exists for plants. In this paper we used the plant Arabidopsis thaliana, as a study system to propose a core regulatory network to recover a cyclic attractor that mimics the oscillatory behavior of the key CC components. Our analyses show that the proposed GRN model is robust to transient alterations, and is validated with the loss- and gain-of-function mutants of the CC components. The interactions proposed for Arabidopsis thaliana CC can inspire predictions for further uncovering regulatory motifs in the CC of other organisms including human.
| The eukaryotic cell cycle (CC) in multicellular organisms is regulated spatio-temporally to yield normal morphogenetic patterns. In plants, organogenesis occurs over the entire lifespan, thus CC arrest, reactivation, and cell differentiation, as well as endoreduplication should be dynamically controlled at different points in time and space [1]. Endoreduplication is a variation of the CC, in which cells increase their ploidy but do not divide. Normal morphogenesis thus depends on a tight molecular coordination among cell proliferation, cell differentiation, cell death and quiescence. These biological processes share common regulators which are influenced by environmental and developmental stimuli [1–3]. It would not be parsimonious to depend on different regulatory circuits to control such interlinked cellular processes, CC behaviors and responses. Thus we postulate that a common network is deployed in all of them. Such overall conserved CC network may then connect to different regulatory networks underlying cell differentiation in contrasting tissue types or to signal transduction pathways elicited under different conditions, and thus yield the emergence of contrasting cellular behaviors in terms of cycling rate, entrance to endocycle, differentiation, etc.
Furthermore, the overall CC behaviors are widely conserved and robust among plants and animals. Hence, we aim at further investigating the collective behavior of the key upstream regulators and studied CC components to understand the mechanisms involved in the robustness of CC regulation under changing developmental stages and environmental conditions faced by plants along their life-cycles. Previous studies, that have shown the oscillatory behavior of several transcription factors, that had not been associated as direct regulators of the CC, support our proposed hypothesis [4]. We thus propose to uncovering the set of necessary and sufficient regulatory interactions underlying the core regulatory network of plant CC, including some key upstream transcriptional regulators.
Computational tools are essential to understanding the collective and dynamical behavior of these components within the regulatory networks involved. As a means of uncovering the main topological and architectural traits of such networks, we propose to use Boolean formalisms that are simple and have proven to be useful and powerful to follow changes in the activity of regulators of complex networks in different organisms and biological processes [5, 6].
Although the key CC components have been described in different organisms, the complexity and dynamic nature of the molecular interactions that are involved in CC regulation and the emergence of the cyclic behavior of the CC molecular components are not well understood yet. The use of systemic, dynamic and mathematical or computational approaches has been useful towards this already. Previous models have focused mainly on yeast and animal systems and have been useful to analyze many traits of CC behavior such as robustness, hysteresis, irreversibility and bistability [7–11]. The latter two properties have been validated with experimental data [12–14].
We herein summarize the main traits and components of the eukaryotic CC. The molecular CC regulators have been described and they are well conserved across distantly related organisms [15, 16]. CC progression is regulated by Cyclin-Dependent Kinases (CDKs) [17] that associate with different cyclins to confer substrate specificity [18]. CDK-cyclin complexes trigger the transition from G1 (Gap 1) to synthesis phase (S phase) in where the genome is duplicated, and from G2 (Gap 2) to mitotic phase (M phase) for the delivery of the newly duplicated DNA to the two daughter cells [19] (see for a review [17, 20]). The CDK-cyclin activity also regulates the cell transit between G and S phases during the endoreduplication process [21, 22].
Two CDKs (CDKA and CDKB) are involved in CC regulation. CDKA;1-CYCDs and CDKA;1-CYCA3 complexes regulate G1/S and S phase progression [23–25]; while CDKB-CYCA2 and CDKB-CYCBs regulate G2/M phase and M progression [26–28]. Thus CDK-cyclin activity is finely-tuned by phosphorylation, interactions with CDK inhibitors such as Kip-related proteins (KRPs), and degradation of cyclins and KRPs by Skp1/Cullin/F-box (SCF), as well as by the anaphase-promoting complex/Cyclosome (APC/C) [29–31]. Besides these components, plant CC machinery has a greater number of CC regulators than other eukaryotes and some of those components such as the CDKB are plant-specific.
Several key transcriptional regulators participate in the G1/S and G2/M transitions [32]. The E2F/RBR pathway regulates G1/S transition by transcriptional modulation of many genes required for CC progression and DNA replication [33, 34]. While E2Fa and E2Fb with their dimerization partner (DP) activate transcription of a subset of S phase genes, E2Fc-DP represses transcription [35]. The function of E2Fa and E2Fb is inhibited by their interaction with RBR [36]; in G1/S transition CDKA;1-CYCD-mediated RBR hyperphosphorylation, releases E2Fa/b-DP heterodimers allowing transcriptional activation of E2Fa and E2Fb targets. Simultaneously the E2Fc-DP transcriptional inhibitor is degraded [37].
Little is known about the regulation of G2/M transition in plants, however a class of conserved transcription factors belonging to the MYB family has been described, that seem to have key roles in CC regulation. MYB transcription factors have a prominent role during G2/M transition, by regulating, for example, CYCB1;1 which is determinant in triggering mitosis [38–43]. For the mitosis exit, APC/C mediates degradation of the mitotic cyclins as CYCB1;1 and CYCA2;3, inactivating CDK-cyclin complexes. CCS52A2, an activator subunit of APC/C, is transcriptionally inhibited by E2Fe [44].
Some previous models have recovered the limit cycle attractor as well for CC components [45–48]. A pioneer model of the CC focused on mitotic CDK-cyclin heterodimer and a cyclin protease oscillatory behavior [49]. On the other hand, Novak and Tyson incorporated additional nodes and interactions to model the G1/S and G2/M transitions of the S. pombe CC [50, 51]. They also analyzed evolutionary roles of CC regulators [52], mutant phenotypes [53], stable steady states [7] and the role of cues such as cell size or pheromones in CC progression [54, 55]. Additionally, comprehensive CC continuous models [45] and generic modules for eukaryotic CC regulation [56, 57] have been proposed.
In addition to continuous formalisms, CC models have used discrete approaches as Boolean models for yeast and mammalian systems [46–48, 58–61], and more recently, hybrid models for mammalian cells have been published [62]. Subsequently, time-delayed variables [63] and variables defining CC events [47, 48] were incorporated. Time robustness was improved with specifications of the temporal order with which each component is activated [60]. Recent published reports on CC dynamics use steady state probability distributions and potential landscapes, and highlight the enormous potential of CC models to characterize normal and altered regulation of mammalian CC [64, 65].
Yeast CC Boolean models with summatory thresholds [58, 59], incorporated self-degradation for proteins, but did not incorporate several negative regulators explicitly. In a later work [61], nodes were kept active when the summatory effect of their regulators was greater than the activation threshold, which implies self-degradation of the protein, when such summatory is equal to or below the threshold. Fauré and Thieffry have transformed CC Boolean models, that use threshold functions, to models with a combinatorial scheme, and they have also presented a broader discussion about these two approaches to logical frameworks [66].
Two Boolean models of budding yeast CC and another one of mammalian CC recover cyclic attractors [46–48]. The mammalian CC model [46] also recovers a fixed-point attractor corresponding to G0. In another study, Fauré and collaborators integrated three modules to yield a comprehensive model for the budding yeast CC GRN [47]. The components included variables to represent cellular growth, citokinesis, bud formation, DNA replication and the formation of the spindle. The yeast CC model by Irons also included variables of CC events (e.g. bud formation or DNA replication) as well as time delays [48]. In contrast to other eukaryotes, in Arabidopsis thaliana (A. thaliana herein) very few attempts have been made to integrate available experimental data on CC regulators using mechanistic models. Only a study that considers the G1/S transition has been proposed and contributed to show some additional conserved features of this CC control point among eukaryotes [67].
We integrated available experimental data on 29 A. thaliana regulatory interactions involved in CC progression into a Boolean discrete model, that recovers key properties of the observed plant CC. The regulatory network, that we put forward, also incorporates three uncovered interactions, based on animal systems (E2Fb → SCF, CDKB1;1-CYCA2;3 ⊣ E2Fa, APC/C ⊣ SCF), as well as 16 interactions based on bioinformatic analyses. Therefore, the latter proposed interactions constitute new predictions that should be tested experimentally. The use of yeast or animal data is supported by the fact that main CC components or regulatory motifs are conserved among eukaryotes [16]. In our model, we include solely molecular components and avoid artificial self-degradation loops, which have been used for recovering the limit cycle attractor. We validated the model simulating loss- and gain-of-function lines, and hence demonstrate that the Boolean network robustly implements true dynamical features of the biological CC regulatory network under wild type and genetic alterations. Possible artifacts due to the discrete dynamical nature of the model used, and of its synchronous updating scheme, were discarded by comparing the Boolean model results to those of a continuous approximation model. The continuous model indeed recovers the robust limit cycle that mimics the dynamical behavior of CC components under a wide range of parameters tested. Finally, we provide novel predictions that can be tested against biological experimental measurements in future studies. The model put forward constitutes a first mechanistic and integrative explanation to A. thaliana CC.
We proposed a Boolean approach to integrate and study the qualitative complex logic of regulation of the molecular components underlying the CC dynamics. We formalized available experimental data on logical functions and tables of truth that rule how the state of a particular component is altered as a function of the states of all the components that regulate it. In a Boolean model each node state can be 0, when the expression of a gene or other type of molecular component or complex of such components is unexpressed or “OFF”, or 1 when it is expressed, or “ON”. Nodes states are updated according to the function: Xi(t+1) = Fi(Xi1(t), Xi2(t), …, Xik(t)), where Xi(t+1) is the state of Xi gene at time t+1 and Xi1(t), Xi2(t), …, Xik(t) is the set of its regulators at time t. The set of logical rules for all the network components defines the dynamics of the system. By applying the logical rules to all nodes for several iterations, the dynamics of the whole network can be followed until it reaches a steady state; a configuration or set of configurations that does not change any more or are visited in a cyclical manner, respectively. Such state is called an “attractor”. Single-point attractors only have one GRN configuration, or cyclic attractors with period n, which have n configurations that are visited indefinitely in the same order. In this paper we propose a GRN model that converges to a single limit cycle attractor that recovers the CC molecular components’ states of presence (network configuration) in a cyclic pattern that mimics the pattern observed for the molecular components included in the model along the different CC phase.
A. thaliana CC Boolean model has the following assumptions:
Nodes represent mRNA, proteins or protein complexes involved in CC phase transitions. Node state “ON” is for the presence of regulator, and “OFF” is for absence; in the latter case, it may also indicate instances in which a component may be present but non-functional due to a post-translational modification.
The state of the RBR (RETINOBLASTOMA-RELATED) node corresponds to a 1 or “ON” when this protein is in its hypo-phosphorylated form and therefore is ready to inhibit E2F transcription factors.
When a particular CDK is not specified, a cyclin can form a complex with CDKA;1, a kinase that is always present because it is expressed in proliferative tissues [68] during the complete CC.
E2Fa, E2Fb and E2Fc need dimerization partner proteins (DPa or DPb) for its DNA-binding. Given that DP expression does not change drastically in CC [69], we assumed that the state of these heterodimers is given only by the presence of E2F factors.
The Boolean logical functions integrate and formalize experimental data available mainly for the A. thaliana root apical meristem, however some data from leaves were considered, and we assumed that these are also valid for CC regulation in the root meristem. Also, data from other systems and data obtained by sequence promoter analysis were considered as indicated in each case [27, 39, 40, 67, 70–85] (summarized in Table 1).
The dynamics of complex formation (such as CDK-cyclin and KRP1, or RBR and E2F factors) are specified directly in the Boolean function of their target genes. For instance, the logic rule for E2Fb is E2Fa & !RBR, indicating that E2Fb state is “ON” when it is transcriptionally activated by E2Fa free of RBR. All E2Fa targets also included in their logical rules RBR, as is shown in S1 Text. Then, the presence of KRP1 or RBR in a logical rule does not imply that they are regulators acting directly on the corresponding target.
The updating scheme for the node states was synchronous.
Most regulatory interactions and logical rules were obtained from the A. thaliana data [20, 21, 25–27, 29, 30, 35, 37, 38, 40, 43, 44, 78–80, 85–103] (detailed in Table 2). A. thaliana CC-dependent expression data for validation was obtained from: [72–74]. The consensus site used for MYB77 was CNGTTR, according to: [75, 76], while that for MYB3R4 was AACGG according to: [43]. The motifs were searched in the regulatory sequences of all network nodes using Pathmatch tool (http://arabidopsis.org/cgi-bin/patmatch/nph-patmatch.pl) of TAIR. Regulatory sequences in TAIR10 Loci Upstream Sequences-1000bp and TAIR10 5’ UTRs datasets were used.
We used BoolNet [104] (a library of R language [105]) and Atalia(Á. Chaos; http://web.ecologia.unam.mx/achaos/Atalia/atalia.htm) to simulate the CC GRN dynamics and perform robustness, and mutant analyses. Systematic alterations in Boolean functions for robustness analyses were done with Atalia, while stochastic perturbations in random networks to compare attractor’s robustness were done with BoolNet. For random perturbations made in transitions between network configurations or in Boolean functions, the “bitflip” method was applied. To validate the GRN model proposed here, we used BoolNet and simulated loss- and gain-of-function mutations for each node, by skipping the node’s logical rule and setting the respective gene to “0” and “1”, respectively.
For the continuous model, we followed [106, 107]. In the continuous version of the model the rate of change for each xi node is represented by a differential equation that comprises production as well as decay rates:
d x i d t = - e 0 . 5 h + e - h * ( ω i ) ( 1 - e 0 . 5 h ) * ( 1 + e - h * ( ω i - 0 . 5 ) ) - γ i x i (1)
The parameter h determines the form of the curve; when h is very close to 0, the curve becomes a straight line, while with values close to 100, the curve approximates a step function. The parameter ωi is the continuous form of Fi(Xi1(t), Xi2(t), …, Xik(t)) used in the Boolean model, and γi is its degradation rate. Detailed information about the continuous model can be found in S2 Text.
The CC model proposed here integrates and synthesizes published data for A. thaliana CC components interactions, as well as some molecular data from other organisms (mammal and yeast), that we propose as predictions for A. thaliana CC regulation, and assume to be conserved among all eukaryotes. The whole set of interactions and nodes included in the model and detailed in Tables 1 and 2 are shown in Fig 1. Four types of molecular interactions can be distinguished: (i) transcriptional regulation, (ii) ubiquitination, (iii) phosphorylation and (iv) physical protein-protein interactions. Additionally, an in silico analysis of transcription factors and promoters was carried out, in order to further substantiate 16 predicted interactions in the GRN (these are: E2Fb → MYB77; MYB77 → E2Fe, MYB3R1/4, KRP1, CYCB1;1, CYCA2;3, CDKB1;1 and CCS52A2; MYB3R1/4 → SCF, RBR, CDKB1;1, CYCA2;3, APC/C, KRP1, E2Fc and MYB3R1/4). The logical rules are available in S1 Text.
Our results show that the nodes and interactions considered are sufficient to recover a single robust cyclic steady state, and thus the cyclic behavior of the components considered. Such behavior closely resembles the periodic patterns observed during actual CC progression, Fig 2. The first two columns or network configurations match a G1 state, given that during the early G1 phase, the CDKA;1-CYCD3;1 complex is absent or inactive by the presence of KRP1 [92, 93, 108]. The CDKA;1-CYCD3;1 state is given only by the presence of CYCD3;1 since CDKA;1 is always expressed in proliferative cells [68]. To facilitate understanding, in Fig 2 the complex CDKA;1-CYCD3;1 is shown instead of only CYCD3;1. The absence of mitotic cyclins (CYCA2;3 and CYCB1;1) at this stage [28, 38], as well as the APC/C presence until the early G1 phase, which is needed for the mitosis exit, also coincides with experimental observations [44, 109, 110]. The presence of the RBR protein in G1-phase implies an inactive state of the E2F, as expected [33, 111, 112]. Then, the third column resembles G1/S transition, where the presence of CDKA;1-CYCD3;1 complex would be inducing RBR phosphorylation and its inactivation [32]. In the fourth configuration, the S-phase is represented by RBR inactivation and E2Fa/b transcriptional activation [113]. In the fifth and sixth configuration, E2Fc state returns to “ON” but RBR state is kept in “OFF”, which indicates that transcription driven by E2Fa and E2Fb can still happen. Indeed, the E2Fb factor appears from the fifth configuration and it is consistent with their function regulating the expression of genes needed to achieve the G2/M transition. In the sixth configuration, MYB77 is turned on, although in synchronization experiments it has been observed to be on until the beginning of mitosis [73]. During G2-phase the MYB transcription factors and KRP1 are expressed [31, 73, 93], the former would maintain dimers of CDKA;1 and mitotic cyclins inactive; and together, this data is consistent with what is observed in the seventh configuration of the CC attractor. In the eighth column, KRP1 is lost because it was phosphorylated by CDKB1;1-CYCA2;3, which is active in the G2/M transition and the onset of mitosis [27]. The phosphorylation of KRP1 drives its degradation and posterior activation of mitotic complexes such as CDKA;1-CYCB1;1 to trigger mitosis [21, 78] (configuration 9 and 10 in Fig 2). The lack of APC/C at the onset of mitosis is determinant for the accumulation of the mitotic cyclins, but APC/C presence is necessary for the mitosis exit [110], which occurs in the eleventh configuration of the attractor (Fig 2). Thus, our CC GRN model recovers a unique attractor of eleven network configurations (Fig 2), which shows a congruent cyclic behavior of its components with that observed experimentally. This result validates that the proposed set of restrictions converge to a single cyclic behavior, which is independent of the initial conditions. A further validation of the proposed CC model, would imply that the recovered cyclic attractor is robust to permanent alterations, as is the case for real CC behavior that is highly robust to external and internal perturbations [14, 58, 114, 115].
To provide further validation for the proposed CC regulatory network, we performed robustness analyses of the attractor to four types of alterations in the logical functions of the model. First, we altered the output of each logical rule by systematically flipping one by one, each one of their bits. We found that 87.47% of the perturbed networks recovered the original attractor, while 1.77% of the altered networks maintained the original attractor and produced new ones (see supplementary material S3 Text for details). In contrast, the remaining 10.76% of alterations reduced the number of network configurations of the original attractor. In the second robustness analysis, after calculating the transitions between one network configuration to the next one, one bit (i.e. the state of a node) of this next configuration is randomly chosen and its value changed. Then, the network is reconstructed and its attractors recovered again. This procedure was repeated 100 times, thus we found that in 88.2 ± 3.2 out of the 100 perturbations (mean ± SD) the original attractor was reached. These results suggest that the proposed GRN for A. thaliana CC is robust to alterations as expected and in coincidence with previous GRN models proposed for other developmental processes [116, 117].
To confirm that the robustness recovered in these two types of analyses is a specific property of the network under study, we performed robustness analyses of randomly generated networks with similar structures (same number of input interactors for the logical functions) to the one proposed here for the A. thaliana CC regulatory network, and compared the above robustness analyses results to those recovered for equivalent analyses for the random networks. We generated 1000 random networks. Then, 100 copies of the random and of our network were done. In each copy we randomly flipped the value of one bit in one logical function (to confirm the first robustness analysis), or in one next configuration (for the second robustness analysis). When perturbations are made in logical functions, the A. thaliana CC GRN recovers its attractor in 68% of perturbations, while the median of percentage of cases in which such attractor was recovered in the random networks was only 18.55% (mean 19.12%±13.86 SD, Fig 3A). The difference between the 68% of this latter analysis and the 87.47% of the first robustness analysis could be due to sampling error. If transitions between network configurations are perturbed, the median of original attractors recovered in random networks is 24.2% (mean 24.6% ± 18.2 SD). In contrast, the original attractor of A. thaliana CC GRN was found in 88% of perturbed networks starting with that grounded on experimental data (Fig 3B). These results confirm that the CC GRN proposed here is much more robust than randomly generated networks with similar topologies and suggests that its robustness is not due to overall structural properties of the network.
Boolean models can produce cyclic dynamics as an artifact due to their discrete nature and the time delays implied. To address this issue we approximated the Boolean model to a continuous system of differential equations following [106, 107, 118, 119]. To recover steady states of such continuous system, the continuous versions of the GRN were evaluated for 1000 different randomly picked initial conditions (See S2 Text). In all cases and independently of the methodology (i.e. [106, 107] or [118, 119]), we recovered the same limit cycle steady state. In the continuous model, key cyclins for the main phase transitions, CYCD3;1 and CYCB1;1, have an oscillatory behavior that is not attenuated with time (Fig 4). Importantly, this result is robust to changes in the decay rates or alterations of the h parameter that affects the shape of activation function (see details in S2 Text); the limit cycle was recovered in 92.86% of the cases. The results of the continuous model corroborate that the limit cycle attractor recovered by the Boolean version, is not due to an artifact associated to the discrete and synchronous nature of the Boolean model, but is rather an emergent property of the underlying network architecture and topology. In addition, the recovery of the cyclic behavior of the continuous model constitutes a further robustness test for the Boolean model.
Previous studies have also tested asynchronous updating schemes [46]. In this study we have used a continuous form of the model to discard that the recovered cyclic attractor is due to an artifact owing to the discrete and synchronous nature of the model used. Future studies could approach analyses of asynchronous behavior of the model by devising some priority classes distinguishing fast and slow processes, and thus refining the asynchronous attractor, under a plausible updating scheme. On the other hand, biological time delays may be involved in CC progression, but they are not enough for irreversibility. The CC unidirectionality has been proposed to be a consequence of system-level regulation [120], here we hypothesize that the ordered transitions of A. thaliana CC are an emergent property of network architecture and dynamics.
An additional validation analysis for the proposed A. thaliana CC model implies simulating loss- and gain-of-function mutations and comparing the recovered attractors with the expression profiles documented experimentally for each mutant tested. We simulated mutants by fixing the corresponding node to 0 or 1 in loss- and gain-of-functions mutations, respectively. The recovered altered configurations are summarized in S4 Text, and in Table 3 as well as in Table 4 for gain- and loss-of-function mutants, respectively. The simulated mutant attractors are coherent with experimental data in most cases [2, 21, 23, 30, 35, 37, 43, 44, 76, 79, 80, 88, 90–93, 103, 108, 109, 111, 113, 114, 121–129]. In Fig 5 we show a representative example of attractors recovered by simulations of CDKB1;1 and KRP1 loss-of-function and APC/C and E2Fa gain-of-function mutants. It is noteworthy that several simulated mutants, such as mitotic cyclins or B-type CDK loss-of-function, converge to a cyclic attractor that corresponds to the configuration observed under an endoreduplicative cycle (e.g. Fig 5A). In such attractors, endoreduplication inductors, such as APC/C, KRP1 and E2Fc [37, 78, 130] are present, at least in some network configurations (Fig 5A, 5C and 5D-right). Another outstanding feature of these mutant attractors is that, although mitotic CDK-cyclin complex may be present, it is inhibited by KRP1, therefore there is no CDK-cyclin activity to trigger the onset of mitosis. These data are coincident with the reported regulation during the onset of endoreduplication [21]. In the attractors where E2Fa coincides with alternating states of RBR, it suggests that DNA replication may occur (Fig 5). Likely due to plant redundancy, some mutations do not produce an obvious impaired phenotype. Such is the case of KRP1 loss-of-function, in which loss-of-function simulation, a cyclic attractor identical to the original one is recovered, as is expected (see Table 4), because such mutants do not show an evident altered CC behavior (Fig 5B) [93].
Interestingly, the simulation of a constitutively active APC/C also converges to a single cyclic attractor, which corresponds to an endoreduplication cycle, since it has Gap and S phases, but lacks an M-phase configuration. This coincides with the experimental observation that the overexpression of one of the APC/C subunits (CCS52A) promotes entry to an endocycle [44] (see Table 3). Another interesting example is the gain-of-function mutation of E2Fa that yields two cyclic attractors, one corresponding to the normal CC cycle and the other one to an endocycle (Table 3). It has been shown that this gene is required for both processes [111] that are apparently exclusive, although in both processes the DNA replication occurs and among E2Fa targets there are genes required for S-phase. Thus our model suggests that the regulation of E2Fa at the end of G2 phase is decisive for CC exit and transition to endoreduplication. In this E2Fa gain-of-function simulation, we found an inconsistency with APC/C because this E3 ubiquitin ligase is decisive for endoreduplication, while in the simulated attractor is only present in one network configuration (Fig 5D-right). Such behavior observed in the endoreduplication attractor for E2Fa gain-of-function leads to unstable activity in the CDK-cyclin complex (Fig 5D), thus suggesting that the increase in APC/C is required for endoreduplication entry as well as its progression. In the attractor of the simulated APC/C gain-of-function, the states of the CYCD3;1, SCF, E2Fb, E2Fc and MYB nodes are more stable than in endoreduplication attractors of CDKB1;1 loss-of-function or E2Fa gain-of-function, where E2Fb, E2Fc and MYB factors expression states alternate between “ON” and “OFF” (Fig 5).
We highlight APC/C gain-of-function simulations, as it provides a possible mechanism for plant hormones action over the CC machinery and, thus how such key morphogens regulate cell proliferation patterns. Recently, Takahashi and collaborators reported a direct connection between cytokinins and CC machinery in A. thaliana root [131]. The authors showed that ARR2, a transcriptional factor of cytokinins signaling, induces expression of APC/C activator protein CCS52A1. Our simulated APC/C gain-of-function is congruent with that observation, since it reproduces the configuration attained by a cell entering an endocycle when APC/C activity is enhanced (Fig 5C), as it happens at the elongation zone of A. thaliana root. Therefore, our model is able to recover the attractors of loss- and gain-of-function mutant phenotypes reported experimentally, and it thus provides a mechanistic explanation for observed patterns of expression in both normal CC and during endoreduplication cycles or endocycle.
We test if the CC GRN recovers the periodic patterns observed in synchronization experiments of A. thaliana CC molecular components. Interestingly, the E2Fc repressor and KRP1 are regulators that have two short lapses of expression in the attractor recovered in the continuous model (Fig 6), and experimentally they also show two peaks of expression when synchronized with aphidicolin [74]. In such synchronization experiments, the expression of E2Fc increases from late S to middle G2, but then it decreases dramatically in late G2. In the model, E2Fc appears from S to G2 phase, and then a second increment of E2Fc expression in G2/M is observed. The latter correspondence is a further validation of the CC GRN model proposed here. Furthermore, synchronization experiments using sucrose have shown that KRP1 is expressed previous to G1/S transition and before mitosis [132], in a similar way that occurs in the model. More recently it has been proposed that KRP1 has a role during G1/S and G2/M transitions [93]; the latter should be important for endoreduplication control [78]. Once again, such roles and expression profiles are consistent with the recovered active state of KRP1 in our model.
In contrast with the consistent behaviors of E2Fc and KRP1 components to recovered results with our model, E2Fe results do not coincide with previous observations. In our model this E2F factor presents only one peak from S to early M phase, but according to synchronization experiments [69], E2Fe has two peaks of expression. One of its peaks is due to regulation by other E2F family factors during S phase, while the G2/M peak could be due to MSA elements. Indeed, when the regulatory motifs for E2F binding are deleted from E2Fe, it can still be expressed although at lower levels [96], suggesting that additional transcription factors regulate its expression. Such factors could belong to the MYB family as suggested for the A. thaliana CC GRN proposed here.
The canonical cyclic behavior of eukaryotic cells as they go from DNA duplication to cytokinesis suggests that a conserved underlying mechanism with shared molecular components and/or regulatory logic should exist. While yeast and animal CC have been thoroughly studied and modelled, plant CC is less studied and no comprehensive model for it has been proposed.
In this study we put forward a Boolean model of the A. thaliana CC GRN. We show that this model robustly recovers a single cyclic attractor or steady state with 11 network configurations. Such configurations correspond to those observed experimentally for the CC components included here at each one of the CC stages. In addition, the canonical order of sequential transitions that is recovered also mimics the observed temporal pattern of transition from one configuration to another one along the CC (Fig 2). The fact that the 16,384 initial conditions of the proposed system converge to this single cyclic attractor already suggests that the GRN comprises a robust module that integrates the necessary and sufficient set of components and interactions to recover molecular oscillations experimentally observed. The proposed GRN is also robust to alterations, being similarly robust to previously published models for other cell differentiation or developmental modules [116, 117, 133]. The model is validated because it recovers A. thaliana wild type and altered (in gain- and loss-of-function) configurations and cycling behaviors. The comparison between experimentally observed and recovered gene configurations is summarized in Tables 3 and 4.
Some cyclins such as CYCD3;1 and CYCB1;1, important components during G1/S and G2/M transitions, show a mutually exclusive regulation, as occurs in a predator-prey Lotka-Volterra dynamical system [134], even though they do not interact directly. Their mutual exclusion is achieved thanks to the coordinated expression of genes with specific proteolytic degradation capacity. Our cyclic attractor shows two transcriptional periods, one of them in S-phase regulated by E2F-RBR pathway, and the second one operating at a time previous to M-phase and regulated by MYB transcription factors. The SCF and APC/C ubiquitin ligases work during G2-to-M phases, and during mitosis exit, respectively. Therefore, the fourteen nodes and their interactions proposed in the CC GRN constitute a necessary and sufficient set of restrictions to recover the oscillations of node states characteristic of CC phases.
Two alternative possibilities could drive CC progression in actual organisms. The first would imply that transitions from one CC state to the next would require external cues, like the cell size. The alternative possibility is that CC progression and the temporal pattern of transitions among stages are both emergent consequences of an underlying complex regulatory network, and do not require external cues, or these only reinforce such temporal progression emergent from complex underlying regulatory interactions. Our CC GRN model supports the latter. This does not imply that several internal or external signals or molecules, such as hormones or other types of cues could alter the CC. Therefore, the two alternative possibilities are not exclusive but they likely complement or enhance each other. Indeed, A. thaliana CC is regulated by plant hormones, light, sucrose, osmotic stress [135] or oxidative stress [136]. These could now be modelled as CC modulators.
In the model proposed here we avoided redundancy. For instance, the KRP1 node represents the KRP family members that share several functions. Also the metaphase-anaphase transition could be added to the model when more data about APC/C regulation (i.e. negative feedback loop comprising CDK-cyclin complexes, or the regulation of Cdc20 homologues) becomes available in plants. Apparently, these simplifications did not disrupt the main features of the A. thaliana CC, since the cyclic behavior distinctive of the CC components was correctly recovered.
Our proposed GRN model suggests some predictions regarding the regulation of certain CC components in A. thaliana. Such predictions can be classified into two types. The first type pertains to those recovered by in silico promoter analysis. The predictions of the second type were inferred from data of other eukaryotes, because they seem to imply conserved components and some evidence from A. thaliana suggested that these interactions are part of the CC GRN in A. thaliana. Three interactions belong to the second type, E2Fb → SCF, CDKB1;1-CYCA2;3 ⊣ E2Fa and APC/C ⊣ SCF (see Table 1 for a synthesis of hypothetical interactions). Although some evidence supports the idea that these interactions could exist in A. thaliana, they should be corroborated with additional experimental examination.
Our model provides a dynamic explanation to the cyclic behavior of certain transcription factors and predicts a novel interaction for E2F and MYB regulators; they connect waves of periodic expression that seem to be key for the robust limit cycle attractor that characterizes CC behavior. Interestingly, previous studies have shown that such periodic transcription can be maintained even in the absence of S-phase and mitotic cyclins [4], which underpin the role of a transcription factor network oscillator for the correct CC progression [137]. A regulatory interaction between E2F and MYB factors (or among the equivalent regulators) may be conserved among other eukaryotes (e.g. mammals and yeast), but there is no experimental support yet for it in A. thaliana. After looking for the same direct evidence in A. thaliana and not finding it, we thought about an alternative regulatory mechanism that consists in transcription factors acting between E2F and MYB. Hence, we decided to analyze the important transcription factor families known so far, to find out if one of their members could be mediating the regulation between E2F and MYB. The TCP (for Teosinte branched 1, Cycloidea, PCF) and the MYB family were chosen because they have been reported to be involved in CC regulation [42]. Based on their gene expression patterns and promoter sequence analysis, MYB77 was our best candidate: it is expressed at the beginning of M phase, and could be regulated by E2F and regulator of MYB (see Table 1). A second possibility might be that several tissue-specific transcription factors are involved in E2F-MYB genetic regulation (e.g. GL3, MYB88, SHR/SCR [17], MYB59 [138] or even members of the MADS box gene family could be implied). Indeed, we have recently documented that a MADS-box gene, XAL1, encodes a transcription factor that regulates several CC components (García-Cruz et al., in preparation).
Differences among eukaryotic CCs allow us to recognize or characterize alternative mechanisms for the regulation of CC. The first difference between GRN of A. thaliana CC and that of other eukaryotes, concerns the number of duplicates of some key regulators. A. thaliana has up to ten copies of some of the genes that encode for CC regulators (e.g. families of cyclins or CDK), while yeast, mammals or the algae Ostreococcus tauri, have much fewer duplicates [20, 139–141]. The only exception concerns the homologues of Retinoblastoma protein, of which there are three members in humans and mouse, and only one copy in A. thaliana [127]. Future models should address the explicit role of CC duplicated components in the plastic response of plant development to environmental conditions. Being sessile, such developmental adjustments, as plants grow under varying environments, are expected to be more important, complex and dynamic than in motile yeast and animals. One possibility is that different members of the same gene family are linked to different transduction pathways of signals that modulate CC dynamics.
The second difference among A. thaliana and other CC was regarding the transcriptional regulation throughout the GRN underlying it. For instance, S. cerevisiae does not have RBR or E2F homologues, but instead has Whi5, Swi4,6 and Mbp1 proteins which perform equivalent regulatory functions to the former CC components [142, 143]. S. cerevisiae does not have any MYB transcription factors but it presents other transcriptional regulators, such as Fkh1/2, Ndd1 and Mcm1 [142, 144, 145], which regulate the G2/M transition in a similar way to MYBs in mammals.
Contrary to the conservation in G1/S transition [15, 67], molecular components controlling G2/M transition seem to vary among different eukaryotes. It seems that molecules such as WEE1 kinase and CDC25 phosphatase are not conserved. In A. thaliana, CDC25-like has phosphatase and arsenate-reductase functions [146], while A. thaliana WEE1 phosphorylates monomeric CDKA;1 in vitro [147], and Nicotiana tabacum WEE1 inhibits CDK activity in vitro [148]. However the lack of any obvious mutant phenotype of CDC25 or WEE1 loss-of-function mutants predicts that these genes are not involved in the regulation of a normal CC. Additionally, although WEE1 has a role during DNA damage [146, 149], does not seem to have a CDKA;1 recognition domain [150]. CDC25-like does not have the required sites for CDKA;1 recognition [150]. In summary, the positive regulatory feedback between CDKA;1 and CDC25-like, as well as the mutual-inhibitory feedback loop between CDKA;1 and WEE1, seem not to be conserved in A. thaliana.
Given all that evidence for G2/M regulation, we integrated the regulatory interactions between stoichiometric CDK inhibitor (KRP1), B-type plant specific CDK and MYB transcriptional factors. It is not surprising that there are clear differences between plant G2 phase regulation and that of other organisms, because variations in this control point could define cell fate. Although differences among the A. thaliana CC GRN uncovered here and that of yeasts and animals have now become clear, we think that the basic regulatory CC module reported here, will be a useful framework to incorporate and discover new components of the CC GRNs in plants and also in other eukaryotes.
Despite the fact that our CC GRN model recovers observed CC stage configurations and their canonical pattern of temporal transitions, it did not recover an alternative attractor that corresponds to the endocycle. We hypothesize that the same multi-stable GRN underlies both states, and additional components yet to be connected to the CC GRN will ensure a cyclic attractor corresponding to the complete CC, and another one with shorter period corresponding to the endocycle. In its present form, our model suggests that CYCD3;1 function, which has been associated with the proliferative state [108] and with a delay in the endocycle onset [23], is important to enter the endocycle. Besides, it also has been reported that CYCD3;1 plays a role in G1/S transition [121] and regulates RBR protein during DNA replication [89]. Furthermore, the endoreduplication attractor obtained in some of our mutant simulations (e.g. Fig 5A, 5C and 5D-right) also supports the role of CYCD3;1 in entering an endocycle.
The GRN model of A. thaliana CC could help to identify physiological or developmental interactions involved in the tight relationship between proliferation and differentiation observed during different stages of development [1, 88, 108, 109, 126]. Previous to cell division, the cell senses its intracellular and environmental conditions to arrest or promote CC progress. Such cues directly affect the CC machinery, which does not depend on a master or central regulator.
CC control is the result of a network formed by feedback and feedforward loops between complexes of CDK-cyclin and its regulators. It is not evident how complex dynamical processes such as CC progression emerge from simple interactions among components acting simultaneously. The proposed CC GRN will be very helpful to study how cell proliferation/differentiation decisions and balance keeps a suitable spatio-temporal control of CC during plant growth and development.
|
10.1371/journal.pcbi.1003250 | The Lazy Visual Word Form Area: Computational Insights into Location-Sensitivity | In a recent study, Rauschecker et al. convincingly demonstrate that visual words evoke neural activation signals in the Visual Word Form Area that can be classified based on where they were presented in the visual fields. This result goes against the prevailing consensus, and begs an explanation. We show that one of the simplest possible models for word recognition, a multilayer feedforward network, will exhibit precisely the same behavior when trained to recognize words at different locations. The model suggests that the VWFA initially starts with information about location, which is not being suppressed during reading acquisition more than is needed to meet the requirements of location-invariant word recognition. Some new interpretations of Rauschecker et al.'s results are proposed, and three specific predictions are derived to be tested in further studies.
| There is a mild form of modern “mind-reading” that involves, with heavy fMRI apparatus and software assistance, to guess from brain signals alone the locations of words that have been seen by a (consenting) subject. The recent surprise brought to us by Rauschecker et al. is not that we can currently do that, but that we can do it in a brain region that had until now been largely taken to discard information pertaining to location — the so-called Visual Word Form Area (VWFA). The contribution of our article is to explain this phenomenon in a principled manner, using computational modeling. The gist of our account is that the VWFA starts out with location information, which is indeed progressively discarded as the region maturates but only in as much as actually required to recognize words presented at different retinal locations (a necessary feat when one learns how to read). This “lazy VWFA” account captures many of the findings reported by Rauschecker et al. in a simple model with very few parameters, and it makes specific predictions that would falsify the model immediately were they to be found incorrect.
| Until recently the undisputed agreement amongst essentially all researchers in the field of visual word recognition, the current authors included, was that the Visual Word Form Area (VWFA hereafter, [1]) is a location-invariant area: that it is the seat of a computing device for “word form” representations whose mechanism —while still unknown— abstracts away from irrelevant properties such as location (see for instance [2]–[4]). This view was more than just a default prior consistent with the locus of the VWFA in the left fusiform gyrus, far down in the ventral visual processing stream. It was also suggested by analogy with the most successful hierarchical network models of invariant object recognition [5], which systematically claim location-invariance in their top layers [6], [7]. It thus came as some surprise when Rauschecker et al. demonstrated that a “blind” classifier was indeed able to categorize, with high accuracy, BOLD activation patterns evoked in the VWFA into the locations at which they had been seen by the subject [8]. Although the notion of a location-sensitive VWFA had been previously evoked by one early fMRI study [9], which explicitly manipulated word location and found support for a posterior-to-anterior gradient of sensitivity in the VWFA, the study of Rauschecker et al. is inconsistent with this account because at least in some subjects, both the posterior and the anterior portions of the VWFA were found to be sensitive to (opposite) locations in the visual field [8].
Why then should the VWFA be sensitive to the location of a word? Computational models ought to help shed light on this question, by showing how certain representations develop through learning. Embarrassingly enough however, there is currently no computational model that makes even so much as an attempt to capture how the VWFA is inserted within the network of brain areas described by Rauschecker et al., let alone attempting to describe the internal organization of the VWFA. But in trying to answer this question, we can do the next best thing and gain some insights from a class of computational models of location invariant word recognition [10]–[12]. These models all consist in a simple feedforward network that learns to recognize words independently of where they have been presented on the input space (in this article, a two-dimensional input space). Thanks to their simplicity, these models have been analyzed [13] and studied in a number of variants with an english or a french vocabulary, words of different lengths, and different visibility assumptions.
The network presented in Figure 1 is the latest instantiation of this class of models. It is not designed to investigate such questions as the role of feedback or of hemispheric integration in reading, and focuses exclusively on how location invariant word recognition might be achieved. The input layer is a location specific bank of letter detectors that codes for the presence of letters at specific horizontal and vertical locations, which would be consistent with any retinotopically organized region or group of regions between V2 and VO in the network described by Rauschecker et al. The hidden layer is where the code for any visual word stimulus —or any visual nonword stimulus— is computed, and thus it is functionally equivalent to the VWFA. The output of the network consists in a bank of location invariant word units, one for each word in the vocabulary, that may be usefully construed as word meaning representations in the pars triangularis of Broca's area (i.e. a location invariant area that receives connections from the VWFA). Every unit in the hidden layer/VWFA is assumed to receive connections from every input unit and to send connections to every output unit, initially with randomly weighted connections. Under the influence of the backpropagation learning algorithm [14], the network learns to change these connections in order to associate location specific letter inputs (e.g. ) to a location invariant output (e.g. LIFE).
Our simulation procedure is described in the “Models” section, and followed in its principal lines the study of Rauschecker et al., whose material consisted of 4 letter English words and who used a linear classifier to sort activation patterns from V1/V2 and the VWFA, evoked by word stimuli presented at 6 possible locations along the horizontal and vertical axes, into 2 or 6 target locations. After training the network to recognize a vocabulary of 100 english 4-letter words presented at 7×7 possible locations, we likewise presented word inputs at 6 possible locations along the horizontal and vertical axes and collected either the input patterns or the hidden patterns (which in the rest of the article are respectively being compared to the human data for V1/V2 and for the VWFA). These patterns were then fed to a linear classifier who learned to classify them either in 2 location categories or in 6, and was tested on its ability to generalize to new patterns.
Table 1 reports the average classification performance on input patterns and hidden patterns from all trained networks, when words were presented horizontally or vertically and when patterns were classified either into 2 or 6 classes. Let us first consider the performance of the classifier for 2 location classes.
For two target classes, input representations were classified with perfect accuracy and in a way that mimics performance on human V1/V2 BOLD patterns with horizontally presented words (model 100%, human V1/V2 93%, chance 50%) and vertically presented words (model 100%, human V1/V2 92%, chance 50%). Performance for hidden representations, while overall inferior, followed the same pattern. Classification accuracy remained well above chance and at almost identical levels for horizontal and for vertical representations, which again compares well to the human data (horizontal model 79.8%, horizontal VWFA 76%; vertical model 77.1%, vertical VWFA 74%), and establishes that just like expert human readers, the trained model is location sensitive both at the input and hidden layer along the horizontal and vertical directions.
Classification performance for six location classes is potentially the most interesting, as it allows for a more fine-grained assessment of the location information present in word representations. For word representations in the input layer, performance was almost at ceiling along the horizontal and vertical axes (horizontal input layer 93.3%, vertical input layer 100%, chance 16.7%). This should be compared to classification performance on V1/V2 fMRI signals in humans, which while above chance was clearly less important than in the model (horizontal V1/V2 66.7%, vertical V1/V2 76.0%). However the observed pattern of results was very similar between humans and model in three respects: first in that they both showed a superior location sensitivity along the vertical axis, second because more classifications were made on locations adjacent to the target than to non-adjacent locations, and third because this adjacency effect was stronger along the horizontal axis than the vertical one (horizontal model 100%, horizontal human 56.3%; vertical model 57.1%, vertical human 40.6%).
Classification scores were weaker for hidden representations but still largely above chance, and location sensitivity was superior along the vertical axis than along the horizontal one (horizontal hidden layer 34.0%, vertical hidden layer 57.3%, chance 16.7%). This again mirrored qualitatively what is observed in humans (horizontal VWFA 26.2%, vertical VWFA 31.2%), including the fact that classifiers made more misclassifications on adjacent locations (horizontal model 49.5%, horizontal human 52.9%; vertical model 44.5%, vertical human 49.1%). We note that the superior strength of classification signals along the vertical axis in the model could explain why detecting adjacency effects in the VWFA for horizontally presented words is hard to achieve: the horizontal signal would be lost much faster as a function of white noise in the hidden patterns than the vertically presented signal.
Figures 2 and 3 provide a visual comparison between our simulations and the human data obtained by Rauschecker et al., for horizontally and vertically presented words, respectively. The agreement between model and human data is generally good, with a few visible discrepancies. Both Figures 2 and 3 (upper pannels) show that input classification is too good in the model as compared to V1/V2. Although some white noise was introduced in classified patterns in order to acknowledge the imprecision of fMRI measurements, this parameter was not fitted to the data, and increasing noise could bring input classification to the same level of performance. The model also brings support to the idea that increased classification accuracy on more central locations could be a byproduct of cortical magnification [8]. Indeed Figures 2 and 3 (lower pannels) show that this secondary phenomenon could not be reproduced in a simple model without cortical magnification, and if anything the opposite is observed as the more peripheral locations are the best classified. Finally, fMRI patterns are slightly more “hemifield specific” when elicited horizontally than vertically, with visibly less misclassifications being made across visual fields in Figure 2 than in Figure 3. This aspect of the data is beyond the scope of our single-hemispheric model.
Although the proportion of adjacent misclassifications helps to convey how similar word patterns are as a function of stimulus location, we also attempted to quantify this similarity more precisely by introducing an index that returns 0 for chance classification (no adjacency effect, a uniform confusion matrix) and increases as the distance between the actual and guessed locations decreases, to reach 1 for perfect classification at all locations (i.e. perfect classification and an identity confusion matrix, see Model section). Roughly speaking the adjacency index (AI) behaves like the geometric mean of the general classification accuracy and the adjacency of the misclassifications. Confusion matrices for the input patterns had a AI of 0.98 along the horizontal axis and 1 along the vertical axis, against 0.66 and 0.47 respectively for humans in V1/V2. Confusion matrices for hidden patterns had smaller AIs, reaching 0.19 for horizontal patterns and 0.36 for vertical patterns, to be respectively compared with AIs of 0.36 and 0.27 for humans in the VWFA.
According to the inference that high adjacency effects constitute evidence for an underlying retinotopy in the classified patterns, these results should imply a retinotopic organization of both the input and the hidden layer in the model along the vertical and horizontal axes. This might come as some surprise to the reader because although the input layer in the model is, by construction, retinotopically organized, units in the hidden layer have no contiguity: unlike the input layer, the hidden layer is a simple bag of units, and because each unit has the same total receptive field over the input layer, there can be no induced topology. Therefore this layer cannot be retinotopic in the accepted sense that contiguous units should code for contiguous inputs, and we must conclude that retinotopic organization is not the only way to account for the adjacency effects reported by Rauschecker et al. We will return to the significance and interpretation of these results in the discussion section.
Figure 4 shows the average evolution of recognition accuracy and location sensitivity in 10 networks, as assessed at 20 epochs of training (to be called steps hereafter, referring to steps rather than directly to epochs is necessary given that networks needed different numbers of epochs to reach criterion: if network X took longer to train than network Y, one training step for X contains more epochs than one training step for Y). At each step location sensitivity was measured just as before, by the generalization performance of a linear classifier on sorting 40 randomly chosen hidden network patterns into 2 or 6 location classes.
Two aspects of the data are immediately obvious. First, before training (step 0) networks have as expected no recognition ability, but nonetheless they already exhibit location sensitive hidden representations, a sensitivity which peaks after the first weight modifications. From the first step onwards, the situation then reverts as networks slowly give away location sensitivity in exchange for word recognition performance. However by the time networks have met the task's requirement on word recognition, not all location sensitivity has been lost and classification accuracy is still well above chance in all conditions. Second, it can be seen that the same interaction previously found at the end of training, between the mode of presentation (horizontal or vertical) and the mode of classification (into 2 or 6 classes), extends throughout training. Specifically, performance is equally good for horizontal and for vertical patterns only when classifying into 2 classes, but much better for vertical patterns when classifying into 6 classes.
The fact that location sensitivity in the model varies inversely with recognition accuracy strongly suggests that it could be linked to the extent of the acquired vocabulary. In other words, hidden representations would end up being less and less location specific for training sets of increasing sizes. To verify this claim, we generated 50 new networks that were identical in everything but for their random initial connectivity and the training regimes they received. 10 networks were assigned to each of 5 different training regimes that used increasingly large training sets (50, 100, 150, 200 and 250 words in our simulation). As in our previous simulations, the hidden patterns for 40 randomly selected words were then collected for each of the 50 networks along the horizontal and vertical axis, and subjected to 2 dedicated classifiers that categorized them either into 2 or 6 classes.
The results are presented in Figure 5. The same general pattern of location sensitivity is found as in previous simulations: location sensitivity being everywhere above chance and showing an interaction between mode of presentation and classification type. Critically however, the results confirm that location sensitivity in the model decreases with vocabulary size as assessed by classification accuracy for 6 location classes. Although classification accuracy for 2 classes did not significantly vary across vocabulary size, the more sensitive classification accuracy for 6 classes exhibits a clear linear decrease in generalization performance from a vocabulary of 50 words (mean horizontal accuracy = 45.7%, mean vertical accuracy = 62.8%) to a vocabulary of 250 words (mean horizontal accuracy = 28.2%, mean vertical accuracy = 45.43%). This establishes that networks give up more location specificity as the vocabulary load increases, which translates into the prediction that readers with a larger estimated vocabulary should have statistically less location-specific representations in the VWFA.
The predictions we have established so far on the time-course of learning and the mature vocabulary size were derived from general learning properties of the network, by averaging classifier performance across random samples of words. But because due to their different confusability, not all the words in a training set are equally easy to learn, one would expect location specificity to be an item-level property. In a last simulation we therefore resort to an other way of testing the model at the item level, by varying the proportion of highly confusable words in the training set.
In a task of location invariant word recognition, anagrams are expected to be the most difficult items to classify: whereas non-anagrams can be recognized just by the set of letters they are made of and without recourse to positional information, anagrams cannot. More generally, distinguishing between two anagrams at several locations would in average require to overcome more interference from shared letters at the same location than when distinguishing between two non-anagrams (even if they share some letters). One would then expect the network to assign more location specific representations to anagrams. This idea is put to the test in a third simulation, where 10 new networks were trained on a lexicon specially designed with 50 normal words and 50 anagrams. As previously, networks had random initial weights and were trained until convergence to criterion. Classification was then performed on horizontally or vertically presented patterns, into 2 or 6 location classes. Unlike the previous simulations however, two different linear classifiers were used to operate respectively on anagram and non-anagram patterns.
Figure 6 illustrates the results. Classification accuracy on normal words revealed exactly the same location sensitivity pattern as previously found: accuracy was well above chance, and there was an interaction between mode of presentation and mode of classification, location sensitivity being marginally higher for horizontal patterns when sorted into 2 location classes (mean horizontal accuracy = 70.0%, mean vertical accuracy = 66.7%), but much higher for vertical patterns when classified into 6 classes (mean horizontal accuracy = 39.4, AI = 0.19; mean vertical accuracy = 57.7, AI = 0.40). Location sensitivity for anagrams, while following generally the same pattern, was higher than for normal words for 2 classes (mean horizontal accuracy = 76.7, mean vertical accuracy = 80.0) as for 6 classes (mean horizontal accuracy = 47.2, AI = 0.63; mean vertical accuracy = 62.2, AI = 0.58). Note that the adjacency index penalizes nonadjacent classification errors and therefore does not always follow the direction of mean accuracies: adjacency is marginally higher along the horizontal axis than the vertical one, despite a strong difference in mean accuracies that goes in the opposite direction. This simulation brings support to the idea that location sensitivity is an item-level property in the model, and makes the prediction that at least in adults, classifying VWFA activation patterns for anagrams should be easier than for normal words.
The model's agreement with experimental data would be less impressive if it had been obtained at the expense of fitting a long list of parameters, or making implausible hypotheses. It would also not be informative of the phenomenon of location-sensitivity in the VWFA if it couldn't explain the reason for it, or if it couldn't make new predictions. In what follows we examine how the model fares in all these respects.
The starting point of this class of models is that the VWFA is engaged in recognizing words in the face of stochastically located inputs, and that using a minimal feedforward network, one can ask how the system solves this task at the exclusion of all others and still get meaningful insights. The approach has proved successful in the past and here we have described an instance of the model that succeeds quantitatively in reproducing several surprising results of Rauschecker et al. To do this we assume only that a supervised learning algorithm is mapping location specific letter patterns produced by normally distributed word stimuli into abstract word representations. Let us look at these practical hypotheses one by one.
There is good evidence for location specific letter representations coming from behavioural and fMRI studies [9], [15], although fMRI data points to the posterior VWFA itself for the locus of these detectors. We would simply note that the complexity of the input code does not appear to play any role in the evolution of location sensitivity in the model, and one would expect the same results with less integrated inputs such as location specific letter features, or with more complex inputs such as location specific letter combinations. Indeed according to the model the critical characteristic of the input is its location specificity: it is this property which, together with random initial connection weights, ensures that the VWFA will start in a location sensitive condition. That location specific inputs should be present from the very early stages of reading is not necessarily in tension with the previously acquired location invariant object recognition skills, or with the mastering of the alphabet by children before they start to read. The general question of how children operate the transition from recognizing single letters to recognizing words is, to this day, an open-problem, but it is conceivable that children achieve this feat by harnessing intermediate stages in the letter recognition system where representations are still location specific (for instance the above mentioned location specific letter features).
The hypothesis of stochastically distributed training inputs is demanded by the well-documented stochastic component of ocular saccades during reading (see [16] for a review) as by the existence of a preferred viewing position [17] —if not simply by fixational eye movements such as microsaccades, which are correlated with bursts of neural activity in the early visual system [18]. Although the lopsided 2D-normal distribution we have used for inputs is numerically arbitrary, it is qualitatively conservative given the horizontal direction of reading in English, from which one would expect much less vertical than horizontal variance in eye fixations. As for the modeling choice that only the central location should be trained to 100% accuracy, it is supported by the fact that the probability of refixation to a given word increases with the distance of the actual landing site to the preferred viewing position [19], showing that expert readers are not trained to achieve perfect recognition for all positions. Although no data has been gathered on variability in fixations along the vertical axis, it is unrealistic to assume that there would be less refixations along the vertical axis than there are along the horizontal one. For these reasons the training protocol that uses normally distributed training patterns with a distribution lopsided on the x-axis and where only recognition at the central location was required to reach perfect scores, appears to us to be the most conservative.
Another modeling choice that should be discussed here is the absence of hemifields and hemispheres in the model. The main rationale behind this choice is that, as intriguing and significant as are the findings of Rauschecker et al. pertaining to the right homologue of the VWFA, these were not obviously relevant to the phenomenon of location sensitivity, which is the main focus of the present work. Our modeling choice therefore should be seen as a simplifying assumption rather than as reflecting a strong theoretical statement. As we have indicated, our interpretation is that the model is operating exclusively in the left hemisphere, with early integration of information coming from the right hemisphere resulting in the activation of all adequate location specific letter detectors, even if the corresponding letters were initially perceived in the left hemifield.
Finally the use of the backpropagation algorithm in a single hidden layer network might be seen as an efficient shortcut for more plausible learning algorithms operating over deeper, hierarchically organized networks. Critically, the property exhibited by our model to learn identical weights for the same letters at different locations appears not to be limited to backpropagation or to visual words, as it has recently been mirrored on a non-linguistic training base and with the Trace rule, a hebbian learning rule with a temporal window [20].
Having defended the model's assumptions, let us consider the account it gives of the phenomena described by Rauschecker et al. According to the model, the VWFA starts with representations that are location-specific and display adjacency effects, but are not at all selective for visual words. In our simulations this initial blindness to identity but sensitivity to location is reflected by word recognition being initially absent, while the classifier is still able to sort hidden word patterns by location (see Figure 4), and classified patterns have a large adjacency index along both axes.
In the model this is a consequence of the initially random connection weights afferent to the VWFA, which will conserve the location-specificity of its inputs. Adjacency effects in the hidden layer are the product of the retinotopy of the hidden layer, as well as of the differential training exposure. The fact that adjacency effects can be observed in a hidden layer that doesn't have any topology demonstrates that adjacency effects cannot be taken as a marker of retinotopy: although a retinotopic organization must imply adjacency effects, the converse does not necessarily hold.
Hidden patterns in the model therefore already start out location sensitive, being a product of location specific inputs propagated by random connection weights. But our simulations also show that the weight modifications at the very first epoch of training produce a burst of location sensitivity. This instantly brings classification performance on hidden patterns to peak values, from which they will then decrease slowly over the course of learning. It is well-known that in backpropagation networks the early weight modifications tend to be the strongest, since they are proportional to the error, thereby explaining the observed burst [21]. But it is perhaps not straightforward why this should go in the direction of more location sensitivity rather than less, or why location sensitivity should slowly decrease from the first epoch to the last. Analyses of previous instances of the model show that during training, backpropagation solves the task of location invariant word recognition by trying to assign the same weights to the same letter inputs seen at different locations [13]: in other words the model slowly turns into a symmetry network [22], and as it does so it naturally looses location sensitivity. However this does not explicitly require to destroy all information about where a word was presented, and therefore the mature representations still exhibit location sensitivity. This is a fortiori true for items that cannot be sorted out simply by considering letter identities, such as anagrams, for which letter representations will need to remain more location specific. In this view, the results of Rauschecker et al. obtain because, at least when it comes to location invariance, the VWFA chooses the path of least effort.
Apart from providing a principled explanation of location sensitivity and adjacency effects, we see that this “lazy VWFA” account makes three testable predictions. A first prediction is that fMRI activation patterns in the VWFA should be less location specific for adults than for first-grade children, who are in the process of learning to read (see Figure 4, decreasing location sensitivity over training steps). A second prediction is that word patterns should be harder to classify in subjects with a higher estimated vocabulary (see Figure 5). Finally, a third prediction is that word patterns should be better classified when they are evoked by anagrams (see Figure 6). These predictions appear to be unavoidable in the sense that they fall out directly of the account itself, and that we expect that none of the few parameters of the model —learning rate, number of hidden units, variance of the input distribution— could be manipulated to change them. More predictions may be derived, especially concerning the impact of lexical frequency and neighborhood on location sensitivity. Although new simulations with training sets varying along these two factors would be required to draw firm predictions, from the observed impact of exposure and of letter overlap we expect that word frequency and neighborhood should be respectively negatively and positively correlated with location sensitivity.
We have presented a simple learning account of location sensitivity in the VWFA, whereby maturation in this brain area is seen as a process of finding the minimal departure from an initially location sensitive connectivity, that could eventually achieve invariant word recognition. The model reproduces experimental data under parsimonious assumptions, helps to clarify some of the original data interpretations, and allows us to make testable predictions. It is also notable that none of the hypotheses we have made in this model –namely the existence of stochastically distributed retinotopic inputs, random initial connectivity, and an incremental error-correction learning algorithm– are a priori specific to visual words, and therefore a similar learning account may apply to other types of visual expertise.
Several new analyses should be carried out to elucidate the experimental data reported by Rauschecker et al. For one thing and if the VWFA is to serve any purpose whatsoever, the activation patterns of its word exemplars ought to be better classified by identity than by location. Figure 7 suggests that this is indeed the case in the model, since by the end of training the model achieves very good recognition accuracy on all sufficiently exposed locations, which is exactly equivalent to a linear classifier like a perceptron network producing high scores on classifying hidden word representations by identities. Although the percentage of correct classification by identity is not to be found in Rauschecker et al.'s article, it would serve as a simple but critical validation of the approach. The sparseness of BOLD signals evoked by visual words, as defined for instance in [6], could also be usefully contrasted with the representations that are used in widespread and neurally inspired computational models that deal with location invariant object recognition [6], [7]. Finally if the model we have presented turns out to be warranted by subsequent studies, an instructive future step would be to address the laterality questions raised by Rauschecker et al., by running the same computational analyses on a model that explicitly distinguishes between left and right hemispheres — as for instance in [10] (see also [23]). Using two distinct hidden layers that would stand for the VWFA and its right homologue, one could hope to gain insights as to whether and how the “complementary” character of location information that Rauschecker et al. reported in these regions could indeed develop, and how it would interact with cross-hemispheric connectivity.
The model is a standard three-layer feedforward network trained under the backpropagation algorithm [14]. It has 70 location banks (10 horizontal times 7 vertical locations) of 26 letter units as an input layer, each sending connections to 50 hidden units, which in turn are fully connected to N word units in the output layer (N varied throughout the simulations, from 50 to 200 words). Initial connection weights were randomly drawn from a uniform distribution with support [−0.5,0.5]. Using random initial weights is standard practice in connectionist modeling, Input and output layers use a localist coding scheme, whereby one and only one unit stood for a given word (or a given letter/location). While input units were clamped to binary values during stimulus presentation, all other units computed their activation as a function of the net input they received, using a standard sigmoid function . The model was trained using a vocabulary of N words on the task of associating sequences of letters seen at different locations (e.g. ) to invariant word units (e.g. life,file,life).
One epoch of training consisted in presenting every word in the training base exactly once. Although all input words were therefore seen with equal frequency, their locations were not uniformly distributed, but were randomly chosen anew every 5 epochs following a gaussian distribution centered on location x = 4,y = 4 (the central location), with a larger spread along the horizontal axis ( = 2.5) than the vertical one ( = 1.5), as shown in Figure 7 (top). Networks were trained for as many epochs as necessary to achieve perfect recognition within a radius of the central location. Unlike in previous instances of the model, for plausibility and also for convenience and speed of simulations the radius was chosen to be one in all simulations (perfect recognition was only demanded at the central location). Even with this relaxed criterion however, by the end of training a large measure of location invariance has been obtained for every word in the training set and in a way that was proportional to exposure (see Figure 7 (bottom)). For every simulation, data analysis was carried out on 10 networks instantiated with different initial weight conditions.
Once a network had been trained successfully, we randomly selected 40 words from its vocabulary and fed either their corresponding input patterns or their hidden activation patterns, obtained at the locations of interest, to a linear classifier. The locations were 6 vertical locations centered horizontally, and 6 horizontal locations centered vertically, emulating the 12 presentation conditions of Rauschecker et al. A constant amount of white noise (mean = 0.0, variance = 0.025) was also added to the patterns before classification, in order to acknowledge noise in fmri recordings (if only qualitatively). The classifier was a simple linear perceptron network, a single input layer fully connected to a single output layer with L units, and initial connection weights randomly and uniformly chosen between −0.1 and 0.1. It was trained using the delta rule (learning rate r = 0.0001) for 500 epochs on all but 6*L of the selected items, either to classify patterns in one of two location categories for C = 2 (which depending on the condition would correspond to “left” or “right”, or to “up” and “down”) or when L = 6 to classify patterns precisely into the 6 locations. In a generalization phase, we used the remaining 6*L items to test the classifier's ability to categorize new patterns. The random word selection, training and testing of the classifier were repeated for 10 runs.
To quantify the adjacency effect revealed by a classification with L location classes, we built an adjacency index that returns one for perfect classification (an identity confusion matrix), and zero in the case of random classification (a uniformly distributed confusion matrix). This is achieved by extracting the mean and standard deviation of the error distribution between guessed locations and target locations, and letting the index vary like the product of the opposite of these moments. To reflect the fact that classification errors confined to adjacent target locations reveal more adjacency than when they are randomly distributed, the index should also incorporate a contrast to the maximum standard deviation obtained in the case of a uniform distribution. An index that meets all of these criteria is:Which can be interpreted as the geometric mean of the classification accuracy and the error distribution's departure from uniformity.
|
10.1371/journal.pgen.1004570 | Pooled Segregant Sequencing Reveals Genetic Determinants of Yeast Pseudohyphal Growth | The pseudohyphal growth response is a dramatic morphological transition and presumed foraging mechanism wherein yeast cells form invasive and surface-spread multicellular filaments. Pseudohyphal growth has been studied extensively as a model of conserved signaling pathways controlling stress responses, cell morphogenesis, and fungal virulence in pathogenic fungi. The genetic contribution to pseudohyphal growth is extensive, with at least 500 genes required for filamentation; as such, pseudohyphal growth is a complex trait, and linkage analysis is a classical means to dissect the genetic basis of a complex phenotype. Here, we implemented linkage analysis by crossing each of two filamentous strains of Saccharomyces cerevisiae (Σ1278b and SK1) with an S288C-derived non-filamentous strain. We then assayed meiotic progeny for filamentation and mapped allelic linkage in pooled segregants by whole-genome sequencing. This analysis identified linkage in a cohort of genes, including the negative regulator SFL1, which we find contains a premature stop codon in the invasive SK1 background. The S288C allele of the polarity gene PEA2, encoding Leu409 rather than Met, is linked with non-invasion. In Σ1278b, the pea2-M409L mutation results in decreased invasive filamentation and elongation, diminished activity of a Kss1p MAPK pathway reporter, decreased unipolar budding, and diminished binding of the polarisome protein Spa2p. Variation between SK1 and S288C in the mitochondrial inner membrane protein Mdm32p at residues 182 and 262 impacts invasive growth and mitochondrial network structure. Collectively, this work identifies new determinants of pseudohyphal growth, while highlighting the coevolution of protein complexes and organelle structures within a given genome in specifying complex phenotypes.
| Cellular processes in eukaryotes are brought about through the contributions of large gene sets, and a continuing obstacle in studying these processes lies in the identification of critical constituent genes. The yeast pseudohyphal growth transition is an important example of a complex cellular growth transition. During pseudohyphal growth, yeast cells form connected chains or filaments, constituting a means of foraging for nutrients under conditions of nitrogen and/or glucose limitation. Yeast pseudohyphal growth has been studied for over two decades as a model of signaling systems controlling stress responses, cell shape, and fungal virulence. Hundreds of genes are required for pseudohyphal growth, however, and the critical genes that determine the filamentous phenotype have not been elucidated. Towards this goal, we implemented a genetic approach to identify alleles linked with the pseudohyphal growth phenotype. These studies identified previously unstudied variation in proteins functioning in a complex that controls cell polarity and in a protein of the mitochondrial inner membrane. This work indicates that proteins in complexes and organelles have coevolved within a given genome to yield distinct outputs and phenotype, while highlighting the application of an approach that is useful for the analysis of complex phenotypes in many eukaryotes.
| The budding yeast Saccharomyces cerevisiae undergoes a pronounced growth transition in response to nitrogen limitation or glucose limitation, forming multicellular pseudohyphal filaments that can spread outward from a colony and/or invade the surface of a solid growth substrate [1], [2]. Yeast pseudohyphal filament formation is a presumed foraging mechanism, accomplished through underlying changes in cell adhesion, cell cycle progression, and budding [1], [3], [4]. During pseudohyphal growth, yeast cells remain physically connected after cytokinesis via mechanisms encompassing the regulated expression and shedding of the flocculin Flo11p [5]–[7]. Cells undergoing pseudohyphal growth exhibit increased apical growth through reorganization of the actin cytoskeleton, regulation of polarity proteins, and delayed G2/M progression [8]–[12].
The molecular basis of yeast pseudohyphal growth has been studied extensively as a model of conserved signaling pathways controlling cell morphogenesis and polarity. Furthermore, related processes of filamentous development in the principal opportunistic human fungal pathogen Candida albicans are required for virulence, and signaling pathways between the related yeasts are conserved [13]. Classic studies of pseudohyphal growth in S. cerevisiae have resulted most prominently in the identification of core pseudohyphal growth signaling modules encompassing the Kss1p mitogen-activated protein kinase (MAPK) cascade, the cAMP-dependent protein kinase A (PKA) pathway, and the AMP-activated protein kinase ortholog Snf1p [14]–[20]. The pseudohyphal growth MAPK cascade encompasses Ste11p, Ste7p, and the MAPK Kss1p [10], [14]. Kss1p phosphorylates the Ste12p transcription factor, resulting in dissociation of the negative regulatory Dig1p and Dig2p interactors and binding of a Ste12p-Tec1p heterodimer to target promoters, such as the FLO11 promoter [21]–[23]. Tpk2p, a catalytic subunit of PKA, phosphorylates the Flo8p transcription factor, promoting Flo8p binding and transcriptional activation at the FLO11 promoter and other regulatory sites [17], [24]–[26]. In response to glucose limitation, FLO11 transcription is regulated by Snf1p; the Snf1p-Gal83p isoform promotes cell adhesion during invasive filamentation by antagonizing Nrg1p- and Nrg2p-mediated repression of FLO11 [19], [27].
While the central components of these signaling pathways have been identified, the scope of the yeast pseudohyphal stress response is broad [28]–[33], and the mechanisms enabling these genes and gene products to drive pseudohyphal filamentation are incompletely defined, as are the genetic determinants within this gene set that underlie filamentation. To further dissect pseudohyphal growth pathways, we implemented a linkage study, coupling whole genome sequencing with pooled segregant analysis. The results present previously unidentified genetic determinants of yeast invasive growth and indicate the coevolution of proteins within complexes in driving phenotype.
For linkage analysis, we selected as parents the non-filamentous S288C-derived strain BY4741 and the filamentation-competent strains Σ1278b and SK1 [34], [35]. Filamentous-form growth in haploid strains is classically assessed using the plate-washing assay of Gimeno et al. [1] to identify pseudohyphal cells that have invaded the agar substrate. The invasive phenotype of each parent strain in this assay is indicated in Fig. 1A. The experimental design of the linkage study is presented in Fig. 1B. The non-invasive S288C-derived strain was mated with each of the filamentous Σ1278b and SK1 strains, and the resulting diploid strain from each cross was sporulated. Meiotic progeny from dissected tetrads were assayed for agar invasion by plate-washing, and spores indicating strongly non-invasive or invasive phenotypes were pooled for subsequent linkage analysis. Only spores resulting from complete meiosis were included in these phenotypic pools, and intermediate filamentation phenotypes were excluded from subsequent analysis to provide the greatest likelihood of identifying allelic variation with a strong effect on filament formation. Genomic DNA was extracted from each segregant pool and subjected to high-throughput sequencing that yielded greater than 100-fold coverage per pool.
From the BY4741-by-Σ1278b cross, 31 complete tetrads (124 spores) were screened for agar invasion, identifying 37 strongly invasive spores and 63 non-invasive spores (Fig. 2A). The segregant pools were sequenced, and candidate determinants of the invasive phenotype were identified using a linkage LOD score of greater than 3 as an arbitrarily defined cut-off. Table S1 provides a listing of these alleles, encompassing only variants that are in protein-coding sequence and that are non-synonymous with respect to the encoded amino acid sequence. This allele set affects 50 genes in eleven linkage blocks physically located on seven yeast chromosomes. Figure S1 summarizes the available functional information for this gene set.
Representative plots of non-synonymous allelic variation with respective LOD scores are graphed in Fig. 2B for chromosomes V and IX, highlighting the pseudohyphal growth transcription factor gene FLO8 and the flocculin effector gene FLO11. FLO8 is a pseudogene in S288C-derived strains [36], and in this analysis, the BY4741 allelic variant containing a premature translational stop at codon 142 of the FLO8 sequence yielded a LOD score greater than 17 (Fig. 2B and C). The FLO11 locus exhibits fifteen allelic changes linked with invasive growth phenotypes (Fig. 2B and C). Previous studies identified allelic variation in FLO11 sequence encoding amino- and carboxy-terminal regions linked with the ability to form biofilms on the surface of wine [37]. We recovered these as well as additional sites of DNA sequence variation in FLO11, with the Σ1278b-encoded alleles indicating linkage with strong invasive growth. The FLO11 sequence contains an internal repeat region that is a source of allelic variation between some strains and colonies [38], [39]; however, we did not observe a change in the number of these repeats between BY4741 and Σ1278b. Collectively, the identification of these known pseudohyphal growth genes demonstrates the relevance of results obtained from our pooled segregant analysis.
To further identify important determinants of invasion, we screened candidates from Table S1 as follows: 1) we generated gene deletions and assayed for invasive growth phenotypes (Table S2), and 2) for genes yielding deletion phenotypes, we generated mutants with swapped alleles to identify genetic variants required for invasive growth in Σ1278b. In particular, we focused on alleles of genes that contributed to cell polarity, cell cycle progression, cell morphology, and cell responses to nitrogen/carbon limitation, as these are hallmark characteristics of filamentation.
By this approach, we identified variation in PEA2 as an important part of the genetics underpinning invasive growth. Pea2p localizes to sites of polarized growth as a component of a protein complex, termed the polarisome [40],[41]. PEA2 is required for wild-type invasive growth, mating projection formation, and bipolar bud site selection in diploids [42], [43]. In the filamentous Σ1278b strain, PEA2 codon 409 specifies methionine rather than the leucine residue encoded in the S288C-derived reference genome. The pea2-M409 allele was linked with invasive growth, and generation of an integrated site-specific mutation (pea2-M409L) reconstituting the S288C-encoded PEA2 allele in Σ1278b resulted in decreased invasive growth (Fig. 3A). Relative to wild type Σ1278b, the cell morphology of the pea2-M409L mutant is altered, exhibiting decreased elongation (Fig. 3B); over a population of 200 cells, the percentage of pea2-M409L cells with a length:width ratio of less than 1.5 was nearly four-fold the corresponding percentage in a wild type strain. In addition, the pea2-M409L mutant is impaired in Kss1p MAPK signaling activity. The Kss1p kinase activates the Ste12p/Tec1p transcription factor complex, which recognizes a regulatory element (FRE) in the FLO11 promoter. The plasmid-based Pflo11-9/10-lacZ construct contains the Ste12p/Tec1p-responsive region of the FLO11 promoter fused to lacZ [6], and, by this reporter, the pea2-M409L mutant yields significantly decreased Ste12p/Tec1p-dependent transcriptional activation of FLO11 relative to wild-type Σ1278b (Fig. 3C). In contrast, the pea2-M409L mutation results in wild-type levels of a similarly designed FLO11 promoter fusion responsive to the PKA pathway effector Flo8p (Fig. 3C) [6].
Under conditions of vegetative growth haploid yeast cells bud in an axial pattern, with new buds emerging adjacent to the preceding bud site [44]. Haploid cells undergoing pseudohyphal growth, however, adopt a predominantly unipolar budding pattern wherein the first bud forms distal to the original cell division site, and subsequent buds cluster in the distal pole [1], [10]. Here, we find that in the Σ1278b background the pea2-M409L mutant, corresponding to the S288C-encoded PEA2 allele, exhibited a decrease in unipolar budding and an increase in axial budding relative to wild type (p<0.001), with levels intermediate between an otherwise isogenic wild-type strain and a pea2Δ mutant (Fig. 3D). For this analysis, we examined a population of invasive cells exhibiting three or more bud scars, such that patterns of axial, unipolar, bipolar, and random budding could be reliably distinguished [44], [45]. This budding phenotype was evident in invasive cells, but not in an equally sized population of cells scraped from the surface of an agar plate. Previous studies have indicated that the majority of bud sites are distal in a pea2Δ mutant [12]; results here also indicate that the majority of bud sites are distal in pea2 mutants, but budding pattern analysis does indicate that Pea2p residue 409 impacts unipolar budding in invasive haploid cells.
In the polarisome complex, Pea2p binds the scaffolding protein Spa2p, a large coiled-coil domain-containing protein required for polarisome function [41], [46]. Here, we assessed the possibility that allelic variation at the PEA2 locus impacts Spa2p binding, using Protein A (ProA)-tagged Pea2p variants to recover by co-immunoprecipitation Spa2p tagged at its amino terminus with the hemagglutinin (HA) epitope. In the Σ1278b strain, the Pea2p-M409-ProA variant recovered more HA-Spa2p than the Pea2p-L409-ProA variant, representing the S288C-encoded PEA2 allele (Fig. 3E).
The BY4741-by-SK1 cross was implemented as described in Experimental Procedures, and phenotypic analysis of meiotic progeny identified 51 and 24 strongly invasive and non-invasive spores, respectively (Fig. 4A). Subsequent deep-sequencing of the phenotypic pools identified allelic variation linked with invasive growth phenotypes within eleven separated locus blocks encompassing 88 genes exhibiting non-synonymous changes and a LOD score of greater than 4 (Table S3). A functional breakdown of these genes is indicated in Figure S2. In this analysis, we used a higher LOD score relative to the Σ1278b-by-BY4741 cross in order to limit the number of selected allelic variants to a manageable size for further study, as the SK1 and BY4741 genomes are more divergent (99.5% sequence identity) than the Σ1278b and BY4741 genomes (99.7% identity) [47]. Very few allelic variants linked with invasive growth in Σ1278b were also identified in SK1, aside from a few sequences near FLO8 that are unlikely to be causative. The set of identified alleles was primarily distinct between the two linkage studies, and deletion phenotypes for tested genes in SK1 are indicated in Table S4.
In haploid spores from this cross, genetic variation is most strongly linked with the invasive growth phenotype over a region of roughly 80,000 bp on chromosome XV, encompassing SFL1 (Fig. 4B and C). SFL1 encodes a transcriptional repressor of pseudohyphal growth that functions by binding to the FLO11 promoter, thereby blocking transcriptional activation [24], [25], [48]–[51]. Consistent with its function in repressing FLO gene expression, deletion of SFL1 results in exaggerated invasive growth [28], [32], [48]. Interestingly, the SK1 strain contains an allelic variant of SFL1 with respect to S288C-derived strains, resulting in the conversion of codon 477 (CAA encoding glutamine) to a TAA stop codon (Fig. 4D). This premature stop codon truncates SFL1 prior to the sequence encoding a domain (AA 571–658) that is strongly similar to a conserved region in Myc oncoproteins [48]. Previous studies have identified hyperactive filamentation in a mutant of the CEN.PK 113-7D background upon introduction of a premature translational stop at SFL1 codon 320 (Q320-stop) [51]. Here, we found that allelic variation in SFL1, encompassing a premature stop codon (C1430T, Q477-stop) in the SK1 background, is linked to the aggressively invasive phenotype of SK1 relative to BY4741.
In the BY4741-by-SK1 cross, allelic variation in MDM32 was linked to invasive growth more strongly than any other identified locus, with a LOD score of 9. As indicated in Fig. 4B–D, MDM32 is found on chromosome XV, and relative to BY4741, the SK1 allele of MDM32 encodes Ser182 and Phe262 rather than Cys and Leu, respectively. MDM32 encodes a protein complex subunit of the mitochondrial inner membrane required for membrane organization, the maintenance of elongated mitochondrial morphology, and mitochondrial DNA nucleoid stabilization [52]. Mitochondrial function is required for pseudohyphal growth, as filamentation-competent strains of S. cerevisiae containing a deleted version of the mitochondrial genome are unable to form pseudohyphae [53], [54]; however, a role for Mdm32p in enabling invasive growth remains to be identified.
To determine the effect of MDM32 allelic variation on pseudohyphal growth, we replaced the SK1-encoded MDM32-C546/T787 allele, specifying Mdm32p-S182/F262, with BY4741-encoded MDM32-G546/A787, specifying Mdm32p-C182/L262, in the SK1 genetic background. This allelic swap decreased invasive growth in SK1, and agar invasion was rescued upon reintroduction of the native SK1-encoded MDM32 allele (Fig. 5A). The SK1 mutant containing the BY4741-encoded allele of MDM32 exhibited a more rounded cell morphology, with the percentage of cells displaying a cell length:width ratio of less than 1.5 increasing from 17% in wild-type SK1 cells to 68% in the SK1 mutant (Fig. 5B). Reintroduction of SK1-encoded MDM32 recovered levels of cell elongation similar to wild type.
To assess the impact of this allelic variation on mitochondrial function, we grew the SK1 mutant with the BY4741 allele of MDM32 on medium containing non-fermentable glycerol as the sole carbon source. As shown in Fig. 5C, the allele-swapped SK1 mutant grows poorly in glycerol-containing media, indicating that oxidative phosphorylation is impaired. The structure of the mitochondrial network is also perturbed upon introduction of the BY4741 allele of MDM32 in the SK1 background. Using the mitochondrion-specific MitoTracker fluorescent dye, which diffuses passively across the plasma membrane and concentrates in active mitochondria by membrane potential, we can visualize a compact and collapsed mitochondrial network in SK1 cells containing the BY4741-encoded Mdm32p-C182/L262 variant, similar to that observed in mdm32Δ (Fig. 5D). MDM32 is a paralog of MDM31, and the encoded proteins have been found to interact, albeit transiently and weakly, as components of protein complexes at the mitochondrial inner membrane [52]. We, therefore, assessed the effect of allelic variation at MDM32 on Mdm31p binding; however, we observed no difference in the recovery of Mdm31p by co-immunoprecipitation between the respective Mdm32p variants (Figure S3). In sum, MDM32 is a determinant of invasive growth, and replacement of the native SK1 allele of MDM32 with the BY4741-encoded allele yields a mutant filamentous growth phenotype.
The linkage analysis presented here identifies a broad gene set contributing to pseudohyphal growth (Fig. 6). The pattern of allelic linkage indicates a large number of determinant loci underlying invasive growth, consistent with results from systematic single-gene deletion and overexpression studies [28], [31], [32]. Within this gene set, components of the polarisome and mitochondria play important roles in enabling invasive growth. We report here that Pea2p residue 409 impacts bud site selection in haploid cells undergoing invasive growth, although the effect is less pronounced in determining initial distal-versus-proximal budding in virgin mother cells. Pea2p residue 409 lies distinct from the Pea2p coiled-coil region between residues 236 and 327 and is important for Spa2p binding. Spa2p interacts with Ste11p and Ste7p from the Kss1p MAPK pathway, providing a mechanism for polarisome-mediated regulation of Kss1p MAPK activity [41]. Our results further indicate that nuclear-encoded Mdm32p is required for invasive growth in SK1, and that residues 182 and 262, located outside of the mitochondrial pre-sequence (AA 1–102) and at the boundary or outside of a transmembrane domain (AA 161–184 and 636–653), are important in enabling invasion, as well as in determining aerobic respiratory function and mitochondrial morphology. Mdm32p is proposed to function cooperatively with other inner membrane proteins and components of the outer mitochondrial membrane in the maintenance of mitochondrial morphology [52], potentially through cytoskeletal interactions that may be affected by variation at these sites.
This analysis highlights two additional points. First, the non-filamentous BY4741 background is not uniformly repressive with respect to pseudohyphal growth. From the Σ1278b and SK1 crosses, we identified three and five blocks of allelic variation, respectively, in BY4741 linked with the invasive growth phenotype; full listings of the encompassed alleles with respect to each cross are presented in Tables S5 and S6. Alleles in S288C-derived strains that promote pseudohyphal growth are antagonized by alleles that repress filament formation, such as the pseudogene form of FLO8; similarly, alleles in SK1 linked with the non-invasive phenotype may be offset by alleles that promote invasion, such as the SFL1 allele containing a premature stop codon. Second, the identified allelic variation in PEA2 and MDM32 and the allele-swapping experiments performed here indicate that within a given genome, functionally interacting genes coevolve to impact phenotype. The majority of genetic variation linked with invasive phenotype in this study involves site-specific changes that do not create pseudogenes. Alleles of these genes yield functional proteins within the respective genomic contexts; however, a given allele results in a hypomorphic phenotype when introduced into a non-native strain. It should be noted that the BY4741-encoded allele of PEA2 may be viewed as being pseudohyphal competent, as Liu et al. [36] reported that the introduction of Σ1278b-encoded FLO8 in a S288C-derived strain is sufficient to enable at least some degree of pseudohyphal growth. The data here suggest that partner genes have likely co-evolved with genes such as PEA2 and MDM32, and the resulting protein complexes are, thus, an important determinant of cell phenotype. These findings highlight the utility in studying these complexes as a whole, in supplement to individual proteins, in order to accurately identify the functions and properties that specify phenotype.
It is interesting that the studies here indicated very little overlap between alleles linked with invasive growth in the Σ1278b and SK1 strains with respect to BY4741 (Fig. 6). Studies mapping quantitative trait loci (QTL) in a cross of the laboratory strain BY4716 and the vineyard strain RM-11 identified hotspots impacting gene expression, protein abundance, and small molecule-dependence [55]–[58]. These hotspots were principally due to alleles in the BY4716 background, leading Ronald and Akey [59] to suggest that the causative polymorphisms may occur at low frequency. The non-overlapping allele sets identified in our analysis are not suggestive of hotspots, but rather highlights the substantial importance of epistatic interactions in determining the sum filamentous phenotype resulting from variant alleles in the haploid segregants. These epistatic interactions likely represent instances of gene coevolution, which has been suggested to occur at an elevated rate for genes encoding proteins of shared biological functions and/or for proteins that have coevolved between species [60], [61]. Clark et al. [60] have analyzed the rate of covariation for pairs of proteins over evolutionary time, and by this analysis, polarisome components as a whole do not exhibit statistically significant evidence of covariation, although many mitochondrial complexes do yield signature indicating evolutionary rate covariation. Further analyses of individual protein pairs from the strains used in our study will be necessary to identify a set of coevolved proteins that drive the filamentous growth phenotype.
In summary, we used pooled segregant whole-genome sequencing to dissect gene networks that determine yeast pseudohyphal growth. This analysis identified allelic variation in the known pseudohyphal growth genes FLO8 and FLO11, while also revealing variation in the negative regulator SFL1, the coding sequence of which contains a premature stop codon in the invasive SK1 background. We further found that amino acid 409 in the polarisome protein Pea2p is a site of allelic variation critical for the protein's ability to signal through the Kss1p MAPK pathway, establish unipolar budding during pseudohyphal growth, and bind the Spa2p polarisome scaffold. Linkage analysis identifies variation in MDM32 as a determinant of invasive growth between S288C derivatives and the SK1 strain; the 182 and 262 residues are sites of variation and contribute to Mdm32 function in aerobic respiration and invasive growth.
A listing of yeast strains and plasmids used in this study is provided in Tables S7 and S8. Haploid deletion mutants were constructed by PCR-mediated gene disruption using pFA6a-KanMX6 or pUG72 [62], [63]. Yeast strains were propagated on rich YPD medium (1% yeast extract, 2% polypeptone and 2% glucose) medium or synthetic medium as described [64]. Yeast invasive growth was assayed on YPD medium.
The statistical modeling used to derive the probabilities of identifying linkage are described by Birkeland et al. [65]. Following mating as indicated in Fig. 1, resulting strains were sporulated and asci were dissected. The dissected spores were grown overnight at 30°C and were individually tested for mating type. Spores resulting from complete meiosis (four viable spores with two each of the α and a mating types) were then used for whole genome sequencing. Each spore was assigned to invasive or non-invasive pools based on its invasive growth phenotype. To ensure equal representation of all segregants in a pooled population, each haploid strain was grown overnight at 30°C in individual 4 ml YPD cultures. The OD600 of the cultures was determined and used to calculate the appropriate volume of each strain so that upon mixing, we would achieve equal numbers of cells. A mate-pair library with 300-bp fragments was prepared for each of the phenotypic pools, and each library sequenced as paired-end reads using the Illumina Genome Analyzer (University of Michigan DNA Sequencing Core). Sequence analysis was performed as described [65]. To obtain an estimate of the recombinant and non-recombinant spore counts in each phenotypic pool for a given observed sequence variant, we multiplied the number of spores in the pool by the fraction of sequence reads from that pool that matched the corresponding allele variant. These values were then used in standard LOD score calculations.
Budding patterns of invasive cells were determined as previously described [44], [45]. In brief, equal concentrations of mid-log phase cultures were spotted onto YPD plates and incubated for 7 days at 30°C; surface cells were subsequently washed off under a gentle stream of water. Residual invaded cells were recovered from the agar using a sterile toothpick. Cells were washed twice in sterile water and were stained with 2 µg/ml calcoflour white. Bud scars were visualized by fluorescence microscopy. Cells with more than three bud scars were examined. Budding patterns were determined by criteria previously described [45]. Budding patterns were divided into four sub-groups: axial, bipolar, unipolar and random. The axial pattern was defined as a long chain of bud scars on the proximal cell pole. Cells with a cluster of scars exclusively at the distal pole were classified as exhibiting unipolar budding. A pattern of medial bud scars was scored as random budding, whereas cells with bud scars equally distributed on both proximal and distal poles were classified as undergoing bipolar budding. For these analyses, 200–250 cells from each strain were scored.
|
10.1371/journal.pntd.0003959 | Is PCR the Next Reference Standard for the Diagnosis of Schistosoma in Stool? A Comparison with Microscopy in Senegal and Kenya | The current reference test for the detection of S. mansoni in endemic areas is stool microscopy based on one or more Kato-Katz stool smears. However, stool microscopy has several shortcomings that greatly affect the efficacy of current schistosomiasis control programs. A highly specific multiplex real-time polymerase chain reaction (PCR) targeting the Schistosoma internal transcriber-spacer-2 sequence (ITS2) was developed by our group a few years ago, but so far this PCR has been applied mostly on urine samples. Here, we performed more in-depth evaluation of the ITS2 PCR as an alternative method to standard microscopy for the detection and quantification of Schistosoma spp. in stool samples.
Microscopy and PCR were performed in a Senegalese community (n = 197) in an area with high S. mansoni transmission and co-occurrence of S. haematobium, and in Kenyan schoolchildren (n = 760) from an area with comparatively low S. mansoni transmission. Despite the differences in Schistosoma endemicity the PCR performed very similarly in both areas; 13–15% more infections were detected by PCR when comparing to microscopy of a single stool sample. Even when 2–3 stool samples were used for microscopy, PCR on one stool sample detected more infections, especially in people with light-intensity infections and in children from low-risk schools. The low prevalence of soil-transmitted helminthiasis in both populations was confirmed by an additional multiplex PCR.
The ITS2-based PCR was more sensitive than standard microscopy in detecting Schistosoma spp. This would be particularly useful for S. mansoni detection in low transmission areas, and post-control settings, and as such improve schistosomiasis control programs, epidemiological research, and quality control of microscopy. Moreover, it can be complemented with other (multiplex real-time) PCRs to detect a wider range of helminths and thus enhance effectiveness of current integrated control and elimination strategies for neglected tropical diseases.
| In the developing world, over 207 million people are infected with parasitic Schistosoma worms. Schistosoma mansoni is one of the most widespread species, and its routine diagnosis is based on microscopic detection of parasite eggs in stool samples. This technique is, however, highly observer-dependent and has suboptimal sensitivity. We compared the performance of stool microscopy with the highly specific real-time polymerase chain reaction (PCR) we recently described for the detection and quantification of parasite–specific DNA. We tested stool samples collected at two different epidemiological settings: a Senegalese population (n = 197) from a high transmission area where S. mansoni and S. haematobium are co-endemic and a Kenyan school population (n = 760) selected from zones with comparatively low S. mansoni transmission. Microscopy mostly missed low intensity infections that PCR was able to detect. Consequently, the PCR may be very useful for the detection of S. mansoni in areas with low levels of infection. Furthermore, being a highly standardized diagnostic procedure, the PCR may improve schistosomiasis control programs, epidemiological research, and quality control of microscopy. Also it can be easily combined with other PCRs to detect a wider range of helminth infections in a single stool sample.
| Schistosomiasis control strategies are currently based on mass drug administration (MDA) with praziquantel to populations at risk [1]. Disease mapping, MDA allocation, and post-MDA monitoring of infection are based on standard microscopy techniques: urine filtration for Schistosoma haematobium, and Kato-Katz on stool for the other Schistosoma spp., including S. mansoni. However, these techniques are laborious and there are recognized deficiencies in their sensitivity, thereby limiting the accuracy of screening and monitoring results, and thus appropriate decision-making [2]. This impairs the efficiency of global efforts to control and eventually eliminate schistosomiasis.
Better diagnostics have great potential to improve the quality of schistosomiasis control programs. For S. haematobium, a good alternative to standard microscopy is already available in the form of hematuria dipstick tests [3]. The diagnosis of S. mansoni however, still heavily relies on the Kato-Katz thick stool smear. Several other detection tools have been proposed, including the circumoval precipitin test on serum samples [4,5], the FLOTAC technique on fecal samples [6], and the point-of-care circulating cathodic antigen assay (POC-CCA) for detection of Schistosoma antigen in urine samples [7,8]. In addition, DNA-based methods, such as real-time polymerase chain reaction (PCR)-based techniques, are increasingly being used for the detection of Schistosoma spp. infections [9–18]. The advantage of microscopy over Schistosoma species-specific antigen tests is that they can detect multiple helminth species, and that they are quantitative. These features make them better apt for large-scale use in integrated neglected tropical disease (NTD) control programs than the single-pathogen tests. PCR, in a multiplex format, has the same above-mentioned advantages as microscopy but has greater flexibility. Indeed, a multiplex PCR can detect all (Schistosoma and other helminth) species at the same time, and at any moment after the stool has been collected. Moreover, PCR is a highly standardized diagnostic procedure and it can also be used to detect parasitic protozoa or other microorganisms that cannot be identified by Kato-Katz.
The aim of the present study was to compare Kato-Katz with PCR for the detection of Schistosoma—and soil-transmitted helminth (STH)—infections in stools from persons living in S. mansoni-endemic areas. To this end, stool samples from ongoing studies in two countries with different endemicity were examined using both tests.
Informed and written consent was obtained from all participants prior to inclusion into the study. For minors, informed and written consent was obtained from their legal guardians and assent was obtained from the children. The Senegalese survey was part of a larger investigation on the epidemiology of schistosomiasis and innate immune responses (SCHISTOINIR) for which approval was obtained from the review board of the Institute of Tropical Medicine, the ethical committee of the Antwerp University Hospital and ‘Le Comité National d’Ethique de la Recherche en Santé’ of Senegal. All community members were offered praziquantel (40 mg/kg) and mebendazole (500 mg) treatment after the study according to WHO guidelines [19]. The Kenyan survey was performed within the framework of the Schistosomiasis Consortium for Operational Research and Evaluation (SCORE). Ethical clearance from this study was obtained from the Scientific Steering Committee of the Kenya Medical Research Institute (KEMRI-SSC no. 1768), the Kenyan Ethical Review Committee, and the Institutional Review Board of the Centers for Disease Control and Prevention in the USA. All children who were positive for Schistosoma infection were treated with praziquantel (40 mg/kg), and those positive for STHs were treated with albendazole (400mg).
Samples were derived from one community-wide study population from a S. mansoni and S. haematobium co-endemic area in northern Senegal with high S. mansoni transmission [20–22], and from a population of schoolchildren living in a S. mansoni mono-endemic area with comparatively low transmission in western Kenya [23]. The Senegalese survey was conducted in Ndieumeul and Diokhor Tack, two neighboring communities on the Nouk Pomo peninsula in Lac de Guiers (Guiers Lake). Details on this study area have been described elsewhere [20–22]. Stool and urine samples were collected between July and October 2009 and stool samples for PCR were stored for each participant. Stool samples from a subsample of 197 individuals with complete parasitological data were analyzed by PCR. The Kenyan survey was conducted in the Asembo division of the Rarieda district along the shores of Lake Victoria in western Kenya, within the framework of a larger study on the distribution of S. mansoni amongst school children. Eight to twelve-year-old children attending public primary schools within a 10km from the lake were included (12km wide transect). In this area, S. haematobium is virtually absent. Stool samples were collected between October 2010 and April 2011, preferentially from the lower prevalence zones [23], and PCR was performed in a subsample of 760 children from 40 schools with complete parasitological data (see also S1 STARD Checklist).
In Senegal, two stool and two urine samples were collected from each participant on consecutive days. From each stool sample, a duplicate 25 mg Kato-Katz slide was prepared for quantitative detection of Schistosoma spp. eggs and qualitative diagnosis of STHs Ascaris lumbricoides and Trichuris trichiura by microscopy [24–26]. Duplicate slides were examined by two different technicians >24h after preparation of the Kato-Katz smear, and for S. mansoni the average egg count was calculated. In addition, filtration of 10 ml of urine was performed using a 12 μm pore-size filter (Isopore, USA) according to standard procedures to detect S. haematobium eggs [25]. Urine filters were read by a single technician. In Kenya, three stool samples were collected on consecutive days, and from each sample, duplicate 42 mg Kato-Katz slides were prepared for microscopy. Schistosoma mansoni was diagnosed quantitatively at least 24h after slide preparation. STHs were diagnosed qualitatively: A. lumbricoides and T. trichiura at 24h after slide preparation, and hookworm within 1h of slide preparation. Each slide was examined by two independent microscopists and the average was recorded. Urine filtration was not performed in Kenya. In both countries, microscopy was performed blinded to previous results, and S. mansoni infection intensity was expressed as the number of eggs detected per gram of feces (epg). Egg-based microscopy results were compared to DNA-based PCR results.
Real-time PCR was performed blinded to previous results. During preparation of the first stool sample, an additional amount of fecal material (~0.7ml) was sieved and diluted in 2ml of 96% ethanol [12]. Samples were frozen, transported to the Netherlands, and stored for weeks to months until PCR analysis was performed at the Leiden University Medical Center. Washing of samples, DNA isolation and the setup of the PCR were performed with a custom-made automated liquid handling station (Hamilton, Bonaduz, Switzerland).
For DNA isolation, 200μl of feces suspension was centrifuged and the pellet was washed twice with 1ml of phosphate-buffered saline. After centrifugation, the pellet was resuspended in 200μl of 2% polyvinylpolypyrolidone (Sigma) suspension and heated for 10 min at 100°C. After sodiumdodecyl sulfate–proteinase K treatment (2h at 55°C), DNA was isolated using QIAamp DNA-easy 96-well plates (QIAgen, Limburg, the Netherlands). In each sample, 103 PFU/mL Phocin Herpes Virus 1 (PhHV-1) was included within the isolation lysis buffer [27,28].
A Schistosoma multiplex real-time PCR (Schisto-PCR) was performed as described previously [29], with some minor modifications [30]. This PCR targets the Schistosoma-specific internal transcriber-spacer-2 (ITS2) sequence of S. mansoni, S. haematobium, and S. intercalatum, as well as PhHV-1 as an internal positive amplification control. The ITS2-based PCR has been validated extensively with a panel of well-defined DNA and stool sample controls and is virtually 100% specific [30]. Amplification was performed by heating samples for 15 minutes at 95°C, followed by 50 cycles, each of 15 seconds at 95°C and 60 seconds at 60°C. Another multiplex real-time PCR, the ANAS-PCR [31], was performed for the detection of STHs Ascaris lumbricoides, Necator americanus, Ancylostoma duodenale and Strongyloides stercoralis. In contrast to the ANAS-PCR, the Schisto-PCR was not designed to differentiate between the different species tested.
Amplification, detection and data analysis were performed with the CFX96 Real-Time System version 1.1 (Bio-Rad, Hercules, CA) [29]. Negative and positive control samples were included in each PCR run. The PCR output from this system consisted of a cycle-threshold (Ct) value, representing the amplification cycle in which the level of fluorescent signal exceeded the background fluorescence. Hence, low Ct values correspond to high parasite-specific DNA loads in the sample tested, and vice versa. The maximum Ct value was 50, and indicated DNA-negative stool samples. The Ct values of the internal PhHV-1 control were within the expected range for all samples, indicating that there was no evidence of inhibition of amplification in any of these samples.
IBM SPSS 22.0 (SPSS, Inc.) was used for statistical analyses (see also S1 Dataset and S1 SPSS Syntax). Results were considered significant when the p-value was <0.05. Kappa (κ) values were calculated as follows to obtain the level of agreement between microscopy and PCR results beyond that which may be obtained by chance:
κ=observed test agreement−expected test agreement1−expected test agreement
Standard cut-off values were used for egg-based infection categories [1]: Schistosoma mansoni infections with 1–99 epg were classified as light-intensity, those with 100–399 epg as moderate, and those with ≥400 epg as heavy-intensity infections. DNA loads as reflected by Ct-values were not normally distributed. Consequently, the Mann-Whitney U test was used to determine differences in DNA loads between S. mansoni egg-negative and S. mansoni egg-positive individuals, and the Kruskal-Wallis test to determine differences in DNA loads between the different egg-based infection categories. Spearman’s rank correlation coefficients were calculated to investigate the correlation between egg- and DNA-based infection intensities, which did not show a linear trend.
In the Senegalese study subjects, we investigated whether PCR outcomes were influenced by S. haematobium infection status. The Pearson Chi² test (with continuity correction) was used to compare PCR positivity between those with and without S. haematobium infection. The Mann-Whitney U test was used to compare DNA loads in stool samples between individuals with and without S. haematobium eggs in urine, as well as between individuals with single S. mansoni and with mixed Schistosoma infections stratified according to S. mansoni infection intensity.
For the analysis of the Kenyan data at the school level, only schools with data on ≥15 children were included (i.e. 24/40 schools). Pearson’s correlation coefficients were calculated to investigate the correlation between egg- and DNA-based infection prevalences in the different schools. Schools were classified into three groups according to their distance from the shore of Lake Victoria: A) the highest prevalence zone ≤1200m from the lake; B) moderate prevalence zone 1200-3800m from the lake; and C) lowest prevalence zone >3800m away.
When only the first stool sample was taken into account, microscopy detected S. mansoni infections in 57.4% of subjects in Senegal and in 19.2% of subjects in Kenya (Table 1) whilst PCR detected Schistosoma DNA in 72.6% and 32.4% of subjects, respectively. Thus, in Senegal, the Schisto-PCR detected 15.2% ((143–113)/197) more infections than microscopy, and in Kenya, 13.2% ((246–146)/760) more infections than microscopy. When two stool samples were taken into account, 68.5% and 25.9% S. mansoni-positives were detected by microscopy in Senegal and Kenya, respectively. When three stool samples were taken into account in Kenya, 29.5% S. mansoni-positives were detected by microscopy. While the percentages of S. mansoni-positives detected by microscopy increased with an increasing number of stool samples, they were still lower than those detected by Schisto-PCR in a single stool sample, in both countries.
When based on the first stool sample, egg- and DNA-based results corresponded in 76.6% (κ = 0.500) and 81.8% (κ = 0.536) of subjects in Senegal and Kenya, respectively (Table 2). When egg counts were based on all stool samples provided (2 samples in Senegal and 3 in Kenya), test agreement increased to 81.7% (κ = 0.561), and 86.3% (κ = 0.680), respectively. Differences in test agreement between countries were mainly due to the fact that in Senegal, egg-negatives were more often found positive in PCR than in Kenya (>twofold difference). Fig 1 demonstrates that mainly low-intensity infections were missed when egg counts were based on only one stool sample.
People that were classified as having heavy infections by microscopy were always PCR-positive. Percentages of PCR-positives varied from 97% to 83% in the moderate egg count group, and from 79% to 87% in the group with light intensity infections. Median DNA loads were very similar in both countries for the different Schistosoma infection categories (Fig 1).
In both countries, Spearman’s rank correlations between egg- and DNA-based infection intensities were statistically significant (p<0.001) with correlation coefficients ranging from -0.638 to -0.782. These correlations became stronger with the number of stool samples that were taken into account (ρ = -0.747 and ρ = -0.782 for 1 and 2 stool samples, respectively, in Senegal; ρ = -0.638, ρ = -0.708, and ρ = -0.738 for 1, 2 and 3 stool samples, respectively, in Kenya).
Based on standard microscopy on stool and urine, 80% (157/197) of the Senegalese subjects were infected with either Schistosoma spp. The majority of these infections (92/157) were mixed S. mansoni and S. haematobium infections. Single S. mansoni infections were found in 22%, and single S. haematobium infections in 11% of subjects (Table 3). Table 3 compares Schisto-PCR outcomes according to Schistosoma infection status (by microscopy). DNA-based infection frequencies were highest in those individuals with single S. mansoni and mixed infections and lowest in persons with single S. haematobium infections and those without any schistosome infection. As by definition, no Schistosoma eggs were observed in stools from uninfected people. In people with single S. haematobium infections, one would expect a similar (low) percentage of PCR-positives as in uninfected individuals. However, 59% of the single S. haematobium group was PCR positive, compared to 23% of the microscopy negatives (p = 0.009). Ct-values were comparable. Percentages of PCR-positives were similar in the single S. mansoni and mixed Schistosoma infection groups, but the mixed infection group showed significantly lower Ct-values (p = 0.003), indicative of a higher intensity of infection. No effect of the presence of S. haematobium on Ct-values in mixed as compared to single S. mansoni infections was observed after stratification for egg-based S. mansoni infection intensity (Table 4).
To explore the diagnostic value of PCR on stool samples in identifying high-risk schools and/or communities, Kenyan test results were analyzed at school level. Data for 24 schools with at least ≥15 children per school, representing 688 school children, were aggregated. The median sample size per school was 27 (range 15–47). Fig 2 indicates a strong, linear correlation between the percentage of microscopy- and PCR-positives per school (p<0.001). DNA-based infection frequencies were consistently higher than egg-based infection frequencies at the school level when both were based on the same stool samples, and PCR identified 25% (22/24 versus 16/24) more S. mansoni-positive schools than microscopy. When egg counts from all stool samples were taken into account, microscopy identified more S. mansoni-positive schools (20/24), and also more high-risk schools (infection frequencies ≥50% [1]), as compared to when only the first stool sample was taken into account (Fig 2). In those high-risk schools, egg-based infection frequencies calculated from three stool samples (six slides) were as high as, or higher than DNA-based infection frequencies. In low-risk schools on the other hand (infection frequencies <10%), PCR detected more infections than microscopy on three stool samples, and it detected more positive schools.
In addition to Schistosoma, we investigated the occurrence of STH infections by Kato-Katz and ANAS-PCR. In both study areas, microscopy indicated low prevalences of STH infections and this was confirmed by PCR (Table 5). The two techniques detected similar percentages of A. lumbricoides-positives in both countries. Hookworm was only present in Kenya, and the ANAS-PCR showed that these infections only involved N. americanus. Interestingly, PCR detected more than threefold the number of hookworm infections than microscopy.
There are only a handful of studies that compared PCR outcomes with the reference method that is routinely used in endemic areas, i.e. microscopy on Kato-Katz smears [9]. Moreover these studies used different PCR targets [9]. A real-time PCR targeting the cytochrome c oxidase subunit I (cox1) of S. mansoni found similar percentages of S. mansoni-positives as standard microscopy in a Senegalese population [12]. The sensitivity of this PCR was found to be suboptimal because the cox1 region shows considerable genetic variation [32]. PCRs based on the 121-bp tandem-repeat sequence showed more promising results with 7 to 28% higher percentages of S. mansoni infections detected than standard microscopy [10,13,17,33–37]. In contrast to the ITS2-based real-time PCR used in the present study however [29], this PCR cannot quantify DNA loads. The present study was the first to compare standard microscopy to an improved Schisto-PCR targeting the conserved ITS2 sequence.
The ITS2-based PCR detected 13–15% more Schistosoma-positive individuals than microscopy when both tests were performed on the same stool sample. These trends were very similar in the north of Senegal where S. mansoni prevalences are high [20], and in the west of Kenya using stools from schools that had considerably lower S. mansoni prevalences [23]. In Kenya, 25% more schools with S. mansoni-infected children were identified based on PCR as compared to microscopy. We observed that the number of egg-positive individuals increased as more stool samples were taken into account. It is indeed well-known that the sensitivity of microscopy increases as more consecutive stool samples are included in the analysis [38]. This is likely due to the variability of egg counts for an individual with a given worm load [39,40]. More S. mansoni egg-negatives tested positive in PCR in Senegal than in Kenya. This between-country difference may be due to methodological differences between the two studies, such as the amount of fecal material examined per stool sample. In Senegal, 2x25mg fecal material was examined per stool sample while in Kenya 2x42mg was examined per stool sample and this may have resulted in relatively more false negative microscopy results for S. mansoni in Senegal. In addition, the co-occurrence of S. haematobium in the Senegalese population may have resulted in ‘false-positive’ PCR results, as the PCR may pick up some occasional S. haematobium DNA present in the stools.
Trends for S. mansoni infection intensities were very similar to those of infection frequencies. While both PCR and microscopy proved adequate to detect S. mansoni infections with higher egg loads and, consequently higher fecal DNA loads, light infections were more likely to be missed by microscopy. People with light infections often showed low Schistosoma DNA levels in stool, and were egg-negative when one stool sample was considered. When more stool samples were tested, these people tended to shift from the negative egg-based infection category towards the light-intensity infection group. Likewise, comparison of the two techniques in Kenya showed that S. mansoni infections in children from schools with low prevalence and intensity were more likely to be missed by microscopy than those from schools with higher prevalence and intensity. It is indeed known that the sensitivity of microscopy is especially low in light-intensity infections, and in low-transmission areas [40]. Apparently, PCR does not suffer (as much) from this problem and may therefore be particularly useful in such situations. The strong correlation between egg counts and DNA loads in Senegal and Kenya, as well as between egg- and DNA-based infection frequencies in Kenyan schools suggests that DNA loads and DNA-based prevalences can be linked with egg counts and egg-based prevalences, respectively. This implies that the cut-offs which are based on S. mansoni egg counts and that are currently used for the allocation of control interventions (e.g. for MDA [1]), may be conveniently translated into cut-offs based on fecal Schistosoma DNA loads. More studies are needed to assess this into more detail and in more geographical areas [41].
We found the performance of the Schisto-PCR to be very similar in Senegal and in Kenya, despite differences in the level of Schistosoma transmission, geographic S. mansoni strains, co-infecting helminths, and demographic composition as well as genetic background of the study population. An additional advantage of PCR is that it is more objective and uniform than microscopy. It does not suffer from methodological variations (e.g. number and volume of stool samples, calculation of average egg count, quality of microscopy), or inter-observer variation, and it is less error-prone. Moreover, stool samples can be stored for later analysis by PCR and if needed, in a central laboratory. Hence, the Schisto-PCR may be particularly useful as an epidemiological tool to reliably compare levels of infection between geographical areas and between studies [42]. In addition, PCR can be used as a reference standard to assess the quality of locally used (reference) methods, and to compare the accuracy of diagnostic procedures between different study sites [43].
Multiplex PCR allows the detection of multiple helminth species, and this spectrum can be further expanded by combining different multiplex PCRs such as the Schisto-PCR and ANAS-PCR. In the present study, the ANAS-PCR confirmed microscopy results showing relatively low levels of STH infections. While microscopy and PCR gave similar results for A. lumbricoides, PCR was more sensitive in the detection of N. americanus than microscopy. These results are in accordance with previous studies that suggested multiplex PCR to be more sensitive than, or as sensitive as, microscopic techniques for the detection of hookworm and A. lumbricoides in areas of low STH transmission [44–46]. Additional advantages of the ANAS-PCR are that it can also detect S. stercoralis and that it can differentiate between the two common hookworm species N. americanus and A. duodenale. Very recently, our group further extended the Schisto- and ANAS- multiplex PCRs to include T. trichiura. In the near future, it will thus be possible to detect not only Schistosoma spp. but also the other most important intestinal helminths–A. lumbricoides, N. americanus, A. duodenale, S. stercoralis, and T. trichiura [47]–in one single analysis. This is not possible by microscopy.
In this study, we extensively evaluated the ITS2-based Schisto-PCR on stool samples for the detection of S. mansoni and showed that it outperforms standard microscopy on Kato-Katz smears. The Schisto-PCR was more sensitive in detecting S. mansoni than standard microscopy, which makes it particularly useful in low transmission areas, and consequently, in post-control settings. As such, it can be used in the context of schistosomiasis control and elimination, but also for epidemiological research, and for quality control of microscopy. Moreover, it can be complemented with other PCRs such as the ANAS-PCR to detect a wider range of helminths. In this way, DNA-based diagnostic tools may aid in enhancing effectiveness of current integrated NTD control and elimination.
|
10.1371/journal.pntd.0006451 | Seasonal temperature variation influences climate suitability for dengue, chikungunya, and Zika transmission | Dengue, chikungunya, and Zika virus epidemics transmitted by Aedes aegypti mosquitoes have recently (re)emerged and spread throughout the Americas, Southeast Asia, the Pacific Islands, and elsewhere. Understanding how environmental conditions affect epidemic dynamics is critical for predicting and responding to the geographic and seasonal spread of disease. Specifically, we lack a mechanistic understanding of how seasonal variation in temperature affects epidemic magnitude and duration. Here, we develop a dynamic disease transmission model for dengue virus and Aedes aegypti mosquitoes that integrates mechanistic, empirically parameterized, and independently validated mosquito and virus trait thermal responses under seasonally varying temperatures. We examine the influence of seasonal temperature mean, variation, and temperature at the start of the epidemic on disease dynamics. We find that at both constant and seasonally varying temperatures, warmer temperatures at the start of epidemics promote more rapid epidemics due to faster burnout of the susceptible population. By contrast, intermediate temperatures (24–25°C) at epidemic onset produced the largest epidemics in both constant and seasonally varying temperature regimes. When seasonal temperature variation was low, 25–35°C annual average temperatures produced the largest epidemics, but this range shifted to cooler temperatures as seasonal temperature variation increased (analogous to previous results for diurnal temperature variation). Tropical and sub-tropical cities such as Rio de Janeiro, Fortaleza, and Salvador, Brazil; Cali, Cartagena, and Barranquilla, Colombia; Delhi, India; Guangzhou, China; and Manila, Philippines have mean annual temperatures and seasonal temperature ranges that produced the largest epidemics. However, more temperate cities like Shanghai, China had high epidemic suitability because large seasonal variation offset moderate annual average temperatures. By accounting for seasonal variation in temperature, the model provides a baseline for mechanistically understanding environmental suitability for virus transmission by Aedes aegypti. Overlaying the impact of human activities and socioeconomic factors onto this mechanistic temperature-dependent framework is critical for understanding likelihood and magnitude of outbreaks.
| Mosquito-borne viruses like dengue, Zika, and chikungunya have recently caused large epidemics that are partly driven by temperature. Using a mathematical model built from laboratory experimental data for Aedes aegypti mosquitoes and dengue virus, we examine the impact of variation in seasonal temperature regimes on epidemic size and duration. At constant temperatures, both low and high temperatures (20°C and 35°C) produce small epidemics, while intermediate temperatures like 25°C and 30°C produce much larger epidemics. In seasonally varying temperature environments, epidemics peak more rapidly at higher starting temperatures, while intermediate starting temperatures produce the largest epidemics. Seasonal mean temperatures of 25–35°C are most suitable for large epidemics when seasonality is low, but in more variable seasonal environments epidemic suitability peaks at lower annual average temperatures. Tropical and sub-tropical cities have the highest temperature suitability for epidemics, but more temperate cities with high seasonal variation also have the potential for very large epidemics.
| Over the last 30–40 years, arboviral outbreaks have dominated the public health landscape globally [1]. These viruses, most notably dengue (DENV), chikungunya (CHIKV), and Zika (ZIKV), can cause symptoms ranging from rash, arthralgia, and fever to hemorrhagic fever (DENV), long-term arthritis (CHIKV), Guillain-Barré syndrome and microcephaly (ZIKV) [2–4]. DENV, which historically spread worldwide along shipping routes [5], places 3.97 billion individuals at risk worldwide [6] and causes an estimated 390 million cases annually, including 96 million symptomatic cases [7]. CHIKV was introduced into the Americas in December 2013 after an outbreak in St. Martin Island [8]. Since then, autochthonous transmission has been reported in 45 countries [9], and 1.3 billion people worldwide are at risk of contracting CHIKV [10]. More recently, the ZIKV epidemic in the Americas captured global attention after the World Health Organization (WHO) designated it a Public Health Emergency of International Concern in February 2016 in response to its association with neurological disorders. Following the first reported case in Brazil in May 2015, ZIKV has spread to 48 countries and territories where it is transmitted autochthonously [11]. Because DENV, CHIKV, and ZIKV are mostly transmitted by Aedes aegypti mosquitoes, they may have similar geographic distributions and risk factors.
Informed public health decisions to limit the spread and magnitude of these arboviral epidemics depend on a robust understanding of transmission dynamics. One mechanistic modeling framework, the Susceptible–Infected–Recovered (SIR) model, has been implemented successfully to model the dynamics of outbreaks of influenza, measles, and vector-borne diseases such as CHIKV and ZIKV [12–14]. This approach tracks virus population dynamics by compartmentalizing individuals by their state in an epidemic (i.e., Susceptible (S), Infected (I), Recovered (R)). This framework can be extended to include additional compartments, such as a latency stage, or to incorporate the dynamics of the mosquito population for vector transmission.
Arbovirus dynamics are strikingly seasonal and geographically restricted to relatively warm climates [6,7]. This arises because several life history traits of the mosquitoes that transmit DENV, CHIKV, and ZIKV are strongly influenced by temperature and seasonality [15–22]. For simplicity, many existing models assume static life history traits [14], and those that address seasonal forcing tend to incorporate sinusoidal variation as a single transmission parameter, β [23]. Treating seasonal temperature variation as a sinusoidal forcing function on the transmission parameter implies a monotonic relationship between temperature and transmission, such that transmission is maximized at high temperatures and decreases at low temperatures. However, decades of experimental work have demonstrated strongly nonlinear (often unimodal) relationships between mosquito and pathogen traits and temperature that are not well captured in a single sinusoidal forcing function [24]. Efforts by Yang et al. [25,26] addressed the need to include seasonal variation by adopting an SEI-SEIR compartmental framework with time-varying entomological parameters and fitting the model to DENV incidence data in Campinas, Brazil. Other previous work has integrated the effects of temperature on mosquito and parasite traits into temperature-dependent transmission models for DENV, CHIKV, and/or ZIKV, and revealing a strong, nonlinear influence of temperature with peak transmission between 29–35°C [27–34]. However, we do not yet have a mechanistic estimate for the relationship between seasonal temperature regimes and transmission potential, incorporating the full suite of transmission-relevant, nonlinear thermal responses of mosquito and parasite traits.
Here, we expand on previous work with three main advances: (1) we incorporate the full suite of empirically-derived, unimodal thermal responses for all known transmission-relevant mosquito and parasite traits; (2) we examine the influence of seasonal temperature mean and variation (in contrast to constant temperatures or daily temperature variation); and (3) we use a dynamic transmission framework to explore the impact of different seasonal temperature regimes on the epidemiologically-relevant outcomes of epidemic size, duration, and peak incidence (in contrast to R0, or vectorial capacity, which are difficult to measure directly). To do so, we incorporate previously estimated and independently validated thermal response functions for all vector and parasite traits [24] into a dynamic SEI-SEIR model [25,26]. We explore field-relevant temperature regimes by simulating epidemics across temperature means (10–38°C) and seasonal ranges (0–17°C) from across the predicted suitable range for transmission. Specifically, we use the model to ask: (1) How does final epidemic size vary across constant temperatures? (2) Under seasonally varying temperatures, how does the temperature at the start of the epidemic affect the final epidemic size and duration? (3) How do temperature mean and seasonal range interact to determine epidemic size? (4) Which geographic locations have high epidemic suitability based on climate?
We first examined how epidemic dynamics varied across different constant temperatures. Here, we did not introduce seasonal forcing into the model but rather assumed static life history traits for Aedes aegypti for the simulation period. We simulated the model under default starting conditions (see Appendix) at four different constant temperatures: 20°C, 25°C, 30°C, and 35°C. These temperatures were chosen to span the range of temperatures at which arbovirus transmission is likely to be possible [24].
Using the model that included seasonal variation in temperature, we examined how the dynamics of an epidemic varied due to the temperature at which the epidemic began, under two temperature regimes. First, we set Tmax = 40.0°C, Tmean = 25.0°C, and Tmin = 10.0°C in the time-varying seasonal temperature model under default parameters (see Appendix) and varied the temperature at the start of the epidemic from 10.0°C to 40.0°C in increments of 0.1°C. We examined the response of final epidemic size, epidemic length, and maximum instantaneous number of infected individuals. We then repeated this process for a regime with a lower magnitude of seasonal temperature variation: Tmax = 30.0°C, Tmean = 25.0°C, and Tmin = 20.0°C. By comparing these temperature regimes, we can examine how epidemics respond to starting temperatures that are outside the range of plausible temperatures of arbovirus transmission (regime 1) versus restricted to the plausible temperatures for transmission (regime 2) [24].
Using the compartmental modeling framework with the default starting conditions, we examined the variation in final epidemic size as a result of seasonal forcing. To do so, we simulated over a wide range of temperature mean and seasonal variance regimes. The mean annual temperature varied from 10.0°C to 38.0°C in increments of 0.1°C, while the seasonal variation about the mean (i.e., Tmax−Tmin2) ranged from 0.0°C to 17.0°C in increments of 0.1°C. Many of these temperature regimes are unlikely to be observed empirically. However, the simulated temperature regimes spanned the full range of feasible temperature conditions. We recorded the final epidemic size, measured as the number of individuals in the recovered compartment at the end of the simulation, for each unique combination of mean annual temperature and seasonal variation. In addition, we examined the effect of epidemic starting temperature on final epidemic size across the same seasonal temperature regimes. We ran the model under default starting conditions, but allowed the starting temperature to equal Tmin, Tmean, or Tmax.
To observe the interaction of population immunity with the seasonal temperature regime, we simulated the model assuming that 0, 20, 40, 60, or 80% of the population was initially immune. Each simulation began with the introduction of the infected individual occurring at the mean seasonal temperature.
We then compared simulated climate regimes with actual climates in major cities, to measure relative epidemic suitability of the following cities: São Paulo, Brazil; Rio de Janeiro, Brazil; Salvador, Brazil; Fortaleza, Brazil; Belo Horizonte, Brazil; Recife, Brazil; Bogotá, Colombia; Medellín, Colombia; Cali, Colombia; Barranquilla, Colombia; Cartagena, Colombia; Tokyo, Japan; Delhi, India; Manila, Philippines; Shanghai, China; Beijing, China; New York City, USA; Guangzhou, China; Kobe, Japan; and Buenos Aires, Argentina, given 0, 20, 40, 60, and 80% population immunity. These cities were chosen because they represent some of the most populous urban areas across South America and throughout the world.
To characterize uncertainty in the model, we sampled 50 joint posterior estimates for c, Tmin, and Tmax for each life history trait provided by Mordecai et al. [24]. We examined the variability in epidemic dynamics with starting temperatures under each parameterization and report the 95% credible interval for the epidemiological indices. We similarly characterize uncertainty in our estimates of the final epidemic size as a function of the seasonal temperature regime by simulating under each parameterization and reporting the 95% credible interval.
Holding temperature constant, we examined variability in epidemic dynamics across four temperatures: 20°C, 25°C, 30°C, and 35°C. As temperature increased from 20°C to 30°C, the number of susceptible individuals depleted more rapidly (Fig 2, SH). At 20°C and 35°C, the epidemics were small (1.33% and 5.92% of the population infected, respectively) and burned out rapidly. Although simulations run at 25°C and 30°C produced final epidemic sizes of 94.73% and 99.98% of the population infected, respectively (Fig 2, RH), the epidemic peaked much faster at 30°C.
Next, we examined variability in epidemic dynamics due to the temperature at which the epidemic began, given two seasonal temperature regimes (25°C mean and a seasonal range of 10°C to 40°C or 20°C to 30°C, respectively). Given that an epidemic occurred, epidemic length monotonically decreased as a function of starting temperature for the first temperature regime (Fig 3A): warmer temperatures at the start of the epidemic produced shorter epidemics, and vice versa. In the second temperature regime, epidemic length monotonically decreased as a function of starting temperature until ~29°C. When temperature varied from 10°C to 40°C, the longest epidemic simulated was 137.8 days and occurred at starting temperatures of 11.2°C, and the shortest epidemic lasted 16.82 days and occurred when the temperature at the epidemic start was 35.7°C. When the temperature was 35.8°C or higher or 10.2°C or lower, no epidemic occurred. When temperature was constrained between 20°C and 30°C, the longest epidemic simulated was 253.64 days at a starting temperature of 20°C, and the shortest epidemic lasted 136.1 days at a starting temperature of 28.9°C.
In contrast to epidemic length, the response of final epidemic size and maximum number of infected individuals to the temperature at epidemic onset depended on the amount of seasonal temperature variation. When temperature varied widely, from 10°C to 40°C, both final epidemic size and the maximum number of infected individuals responded unimodally to starting temperature, with peaks at 23.9°C and 24.1°C, respectively (Fig 3A). By contrast, when temperature varied more narrowly from 20°C to 30°C, the final epidemic size and the maximum number of infected individuals were insensitive to starting temperature (Fig 3B). Taken together, these results show that epidemics introduced at different times within identical seasonal temperature regimes can produce very similar final epidemic sizes and maximum infection rates, provided that the temperature range is sufficiently constrained. If temperature variation is large, dramatically different final epidemic sizes and maximum infection rates may result.
To address how mean temperature and seasonal variance combined to influence the final epidemic size, we simulated over a wide range of temperature regimes that accounted for variation in the mean and temperature range over a calendar year. We calculated relative epidemic suitability, defined as the final epidemic size as a proportion of the human population, for twenty major cities worldwide (Table 4).
In a low-variation thermal environment, a band of mean temperatures between approximately 25°C and 35°C supports the highest epidemic suitability (Fig 4). As the seasonal temperature range increases, lower mean temperatures are capable of supporting large epidemics. However, outside this narrow band of temperature regimes, epidemic suitability rapidly diminishes, and most temperature regimes did not produce epidemics.
Of the focal 20 major cities, those with high mean temperature and small average temperature variation exhibited the highest epidemic suitability. For instance, Manila, Philippines, which has a monthly mean temperature of 29°C and average seasonal amplitude in mean temperature of 1.50°C, had an epidemic suitability of 0.9998. Cartagena and Barranquilla, Colombia had epidemic suitability of 0.9997. On the other hand, areas with low average temperature and greater temperature variation, such as Beijing and New York, exhibited lower—but still non-zero—epidemic suitabilities of 0.5268 and 0.04088 respectively. Notably, Guangzhou and Shanghai, China have high epidemic suitability (0.9996 and 0.9966, respectively) despite moderate mean temperatures (22.9 and 17.6°C, respectively) due to high seasonal variation in temperature. By contrast, high seasonal variation reduced suitability to 0.9537 in Delhi, India, which has a high mean temperature of 26.3°C (Fig 4).
The relationship between epidemic suitability and seasonal temperature regime was consistent across varying levels of population immunity. Locations with high mean temperatures and small average temperature variation had higher epidemic suitability, regardless of the level of population immunity (S8–S10 Figs). However, as the level of immunity increased from 20% to 80%, the epidemic suitability at given seasonal temperature regime decreased (Table 4).
Epidemic suitability also varied by starting temperature, depending on the seasonal temperature regime. The epidemic suitability of cities with high mean temperature and small average temperature variation—such as Manila, Philippines and Cartagena and Barranquilla, Colombia—did not depend on starting temperature (Table 5). However, areas with low to moderate mean temperature and large average temperature variation (e.g., Kobe, Japan and Shanghai, China) exhibited low epidemic suitability (both 0.0001000) at the minimum starting temperature and moderate-to-high epidemic suitability at the maximum starting temperature (0.6890 and 0.8905, respectively) (Fig 5). The opposite occurred in regimes with high mean temperature and large temperature variation, though these temperature regimes are rarer.
Estimated epidemic suitability is close to one in the most suitable temperature regimes because we assumed that: (i) the population was fully susceptible at the start of the epidemic; (ii) mixing was homogeneous among humans and mosquitoes; (iii) all cases of infection are included regardless of whether or not they are symptomatic; and (iv) no other environmental or social drivers are limiting transmission. As a result, the epidemic suitability metric should be considered an upper bound on the proportion of the population that could become infected based on temperature alone.
Final epidemic size was not sensitive to life history trait parameterization (S8–S10 Figs), using samples from the posterior distribution of thermal response fits for each temperature-dependent trait.
There was uncertainty in the specific numerical values of the epidemiological indices across starting temperatures (S1 Fig). However, the overall functional response of the final epidemic size, maximum number of infected individuals, and the epidemic length to starting temperature was consistent across the samples from the joint posterior distribution.
Recent outbreaks of DENV, CHIKV, and ZIKV in Latin America and across the globe have captured the attention of the public health community and underscore the importance of preparation for future outbreaks. As temperatures rise, the global landscape suitable for such outbreaks will expand and shift geographically, potentially placing a larger proportion of the world’s population at risk [24,29,31]. Understanding how local temperature regimes govern epidemic dynamics is increasingly important for determining resource allocation and control interventions [41]. While previous work has investigated the effects of temperature on DENV, CHIKV, and/or ZIKV transmission, until now we have lacked comprehensive, mechanistic, and dynamic understanding of the effects of seasonally varying temperature on transmission via its (nonlinear) effects on mosquito and parasite traits [27–34]. With our model, which expands on [24] and [25], we show that seasonal temperature mean and amplitude interact with the temperature at epidemic onset to shape the speed and magnitude of epidemics.
At constant temperature, epidemics varied substantially in the rate at which susceptible individuals were depleted. Epidemics simulated at 25°C and 30°C reached similar sizes but the epidemic at 25°C proceeded at a much slower rate (Fig 2). This “slow burn” phenomenon occurs because slower depletion of susceptible individuals can produce epidemics of similar size to epidemics that infect people very rapidly. This phenomenon also occurs in more realistic, seasonally varying temperature regimes.
The temperature at which an epidemic started affected dynamics only under large ranges of temperature variation. When temperature ranged from 10°C to 40°C, the final epidemic size peaked at intermediate starting temperatures (24°C; Fig 3A). However, in highly suitable seasonal environments, final epidemic size was large regardless of the starting temperature (Fig 3B).
At mean starting temperatures, epidemic suitability was sensitive to the interaction between annual temperature mean and seasonal variation. Under low seasonal temperature variation, a narrow band of annual mean temperatures (approximately 25–35°C) had the highest epidemic suitability (Figs 4 & S2–S5). Outside this band of temperature regimes, suitability diminishes rapidly. Larger seasonal variation in temperature lowers the range of optimal annual mean temperatures (i.e., suitability is high in cooler places with larger seasonal variation in temperature; Fig 4).
The relationship between epidemic suitability and the seasonal temperature regime also depended on the temperature at the epidemic onset. Three distinct relationships emerged (Figs 5 & S6 and S7). At intermediate annual mean temperatures of ~25–35°C and low seasonal temperature variation (~0–10°C), epidemic suitability is insensitive to starting temperature because temperature is suitable for transmission year-round. At lower annual mean temperatures (~10–25°C) and higher seasonal temperature variation (~10–15°C), epidemic suitability is highest when epidemics start in moderate to warm seasons, and lower when epidemics start during cooler seasons. Finally, at high annual mean temperatures (> 35°C) and low seasonal temperature variation (~0–10°C), epidemic suitability is high only when epidemics start at the coldest period of the year, because otherwise the temperature is too warm for efficient transmission. The interaction between temperature mean, annual variation, and starting point sharply illustrates the unimodal effect of temperature on transmission. Models that do not include unimodal effects of temperature (e.g., those with sinusoidal forcing on a transmission parameter) may fail to capture the limits on transmission in warm environments.
With rising mean annual temperatures and increasing seasonal temperature variation due to climate change, the landscape of epidemic suitability is likely to shift. Importantly, areas with previously low epidemic suitability may have increasing potential for transmission year-round. By contrast, warming temperatures may drive epidemics in cities with high current suitability (e.g., Manila, Philippines, Barranquilla, Colombia, and Fortaleza, Brazil) to shift toward cooler months. Thus, climate change may alter not only epidemic size and duration but also seasonal timing globally, as it interacts with other important drivers like rainfall and human behavior.
It is important to note that model-estimated epidemic suitability should be treated as an upper bound on the potential for large epidemics because within highly suitable climate regimes, epidemics can vary in magnitude due to human population size and movement dynamics [28], effective vector control, and other mitigating factors. Likewise, our estimates are conditioned on Aedes aegypti presence and virus introduction to support an outbreak.
Although seasonal temperature dynamics provide insight into vector-borne transmission dynamics, other factors like mosquito abundance, vector control, and rainfall also determine transmission dynamics. Thus, temperature must be considered jointly with these factors. Moreover, accurately describing epidemic dynamics of emerging and established vector-borne pathogens will ultimately require integrating realistic models of environmental suitability, as presented here, with demographic, social, and economic factors that promote or limit disease transmission [42,43]. Conversely, we show that the interaction between temperature and the availability of susceptible hosts alone can drive epidemic burnout even in the absence of other limiting factors like vector control and seasonal precipitation. This suggests that correctly representing the nonlinear relationship between temperature and epidemic dynamics is critical for accurately inferring mechanistic drivers of epidemics and, in turn, predicting the efficacy of control interventions.
|
10.1371/journal.pbio.3000068 | Identification and characterization of a mosquito-specific eggshell organizing factor in Aedes aegypti mosquitoes | Mosquito-borne diseases are responsible for several million human deaths annually around the world. One approach to controlling mosquito populations is to disrupt molecular processes or antagonize novel metabolic targets required for the production of viable eggs. To this end, we focused our efforts on identifying proteins required for completion of embryonic development that are mosquito selective and represent potential targets for vector control. We performed bioinformatic analyses to identify putative protein-coding sequences that are specific to mosquito genomes. Systematic RNA interference (RNAi) screening of 40 mosquito-specific genes was performed by injecting double-stranded RNA (dsRNA) into female Aedes aegypti mosquitoes. This experimental approach led to the identification of eggshell organizing factor 1 (EOF1, AAEL012336), which plays an essential role in the formation and melanization of the eggshell. Eggs deposited by EOF1-deficient mosquitoes have nonmelanized fragile eggshells, and all embryos are nonviable. Scanning electron microscopy (SEM) analysis identified that exochorionic eggshell structures are strongly affected in EOF1-deficient mosquitoes. EOF1 is a potential novel target, to our knowledge, for exploring the identification and development of mosquito-selective and biosafe small-molecule inhibitors.
| Mosquito-borne pathogens infect millions of people worldwide, and the rise in insecticide resistance is exacerbating this problem. A new generation of environmentally safe insecticides will be essential to control insecticide-resistant mosquitoes. One potential route to such novel insecticide targets is the identification of proteins specifically needed for mosquito reproduction. Female mosquitoes feed on blood to produce eggs, which are covered with an eggshell; using RNA interference screening of mosquito-specific genes in Aedes aegypti (the mosquito that transmits yellow fever), we identified the eggshell organizing factor 1 (EOF1) protein that plays an essential role in eggshell melanization and embryonic development. Nearly 100% of eggs laid by EOF1-deficient females had a defective eggshell and were not viable. Bleach assays on eggs further confirmed that mosquito-specific EOF1 is required for embryonic development in A. aegypti. Additional experiments revealed that EOF1 also plays an essential role in eggshell formation in Aedes albopictus (the tiger mosquito, a carrier of Zika virus and dengue fever). We hypothesize that EOF1 has evolved within the Culicidae family to effect eggshell formation and therefore maximize egg survival. The results provide new insights, to our knowledge, into mosquito egg maturation and eggshell synthesis and could lead to key advances in the field of mosquito vector control.
| Developing new strategies for vector control is becoming critical because worldwide cases of Aedes aegypti-transmitted dengue and Zika virus infections have risen dramatically in the last decade [1–3]. Researchers have been investigating metabolic regulation of blood meal metabolism in A. aegypti mosquitoes as a strategy for identifying novel protein targets that could be exploited for vector control [4–14]. The approach has been focused on biochemical processes that are likely to be required for completion of the gonotrophic cycle in blood-fed mosquitoes, based on what is known about mosquito biology and metabolic regulation in other organisms. Specific genes in these chosen pathways were then systematically knocked down by microinjection of double-stranded RNA (dsRNA), and the resulting phenotypes were characterized in detail by molecular and biochemical approaches.
The insect eggshell is important as a protective layer for embryonic development. Follicle development and eggshell formation in the A. aegypti mosquito are tightly regulated in response to blood feeding [15–20]. Once female mosquitoes acquire blood, follicle development is initiated via accumulation of vitellogenin yolk proteins. Mosquitoes contain approximately 100 ovarioles per ovary, which are composed of primary and secondary follicles and a germarium, and the ovarian follicles develop synchronously throughout oogenesis (S1 Fig). A single layer of follicular epithelial cells surrounding the oocyte is mainly responsible for secreting a majority of eggshell structural components. The mosquito eggshell is made from different types of proteins ranging from structural proteins, enzymes, odorant binding proteins, and uncharacterized proteins of unknown function. A. aegypti eggshell melanization proteins were identified more than 20 years ago [21], and several key eggshell enzymes have been well characterized [22–28]. Moreover, proteomic studies have been performed on purified mosquito eggshells to identify most of the abundant protein components [29,30]. However, these descriptive studies have not identified essential eggshell proteins that are required for successful embryonic development and larvae viability.
Genomic sequences of Drosophila melanogaster [31], Anopheles gambiae [32], A. aegypti [33], and Culex quinquefasciatus [34] have been completed. Not surprisingly, many predicted putative proteins identified in the genome of mosquitoes are homologous to proteins of known function studied in other organisms. Proteins that are conserved in a wide variety of organisms are not ideal target molecules as vector control agents because of deleterious effects on nontarget organisms such as vertebrates, pollinating agricultural insects, and beneficial predators. We reasoned that if small-molecule inhibitors could be designed to exclusively target mosquito-lineage–specific proteins, they could be used as biosafe vector control agents. The objective of this study is to identify mosquito-specific proteins that are essential for supporting embryonic development in A. aegypti mosquitoes by RNA interference (RNAi) screening. Our findings indicate that eggshell organizing factor 1 (EOF1, AAEL012336) is necessarily required for mosquito eggshell formation and melanization.
We performed data mining and bioinformatic analysis using the GenBank database to identify putative protein-coding sequences that are only present in the genomes of Aedes, Culex, and Anopheles mosquitoes using a cutoff for expected value threshold of 1 × 10^(−15). Importantly, the mosquito-lineage–specific genes we identified (S1 Table) were found to be completely absent in evolutionarily closely related organisms, such as phantom midges, true midges, the crane fly, and sandflies within the suborder Nematocera, and thus these genes are not present in other known animals, plants, fungi, and bacteria species. In order to focus on genes that are expressed and likely to encode proteins that could potentially serve as vector control targets, we excluded genes without corresponding messenger RNA (mRNA) in A. aegypti expressed sequence tags (ESTs) or expressed orthologs in the Aedes albopictus transcriptome shotgun assembly (TSA) database. We also excluded mosquito-lineage–specific genes that appear to be members of a multigene family because RNAi knockdown phenotypes may not be immediately obvious because of possible functional redundancy with other gene family members. This highly selected subset of hypothetical mosquito-lineage–specific proteins may have therefore evolved independently and advantageously within the family Culicidae. Systematic RNAi screening of mosquito-specific genes was performed by directly microinjecting the corresponding dsRNA into female A. aegypti mosquitoes 3 days prior to blood feeding (Fig 1A), and the blood-fed female mosquitoes were individually analyzed for their egg phenotypes, fecundity, and viability (Fig 1B). 40 mosquito-specific genes were screened (S1 Table), and utilizing this experimental approach led to the identification of EOF1, which, upon RNAi knockdown, plays an essential role in the strength and structural integrity of the forming eggshell, as well as its melanization. We hypothesize that EOF1 has evolved within the family Culicidae to affect eggshell formation and melanization and therefore maximize egg survival. We did not observe any defective eggshell in mosquitoes microinjected with dsRNA against 39 other putative genes in A. aegypti mosquitoes, resulting in the production of viable eggs. We also did not observe any significant mortality in response to RNAi against these genes. Thus, we chose to focus on EOF1 in subsequent analyses.
EOF1 sequences found in Aedes, Culex, and Anopheles mosquito species contain an F-box functional motif, and members of the F-box protein family are in general characterized by an approximately 50 amino acid F-box motif that interacts with a highly conserved SKP1 protein in the E3 ubiquitin ligase SCF complex [35], suggesting that EOF1 may function to regulate intracellular protein turnover. A further RNAi study on proteins that participate in the SCF complex was not pursued since these proteins that participate in the complex are relatively highly conserved across taxa, and therefore these proteins may not be ideal target proteins to further characterize in mosquitoes. In addition, RNAi knockdown against proteins in the SCF complex may have effects on other proteins that contain the F-box motif. Recent proteomic analysis has identified over 100 mosquito eggshell proteins [29–30], and some of these proteins identified are enzymes that may be involved in catalyzing eggshell melanization and cross-linking reactions [22–28]. However, EOF1 was not previously identified in these mosquito eggshell proteomic studies, indicating that EOF1 may be an upstream regulatory factor of eggshell proteins. As shown in Fig 1, injection of dsRNA-EOF1 had a significant adverse impact on eggshell formation and egg viability. Single-mosquito analysis showed that phenotypes associated with RNAi-EOF1 range from totally nonmelanized and collapsed to truncated and melanized eggs, while untreated and RNAi-firefly luciferase (Fluc) control mosquitoes laid eggs that exhibit uniformly elongated and melanized patterns (Fig 1C–1E). Approximately 60% of eggs laid by EOF1-deficient mosquitoes did not show any melanization (Fig 1C), while 30% of them had mixed melanization levels ranging from nonmelanized and partially melanized to completely melanized eggshells (Fig 1C). On the other hand, 10% of eggs had completely or partially melanized eggshells (Fig 1C). Overall, nearly 100% of eggs from any melanization levels did not reach the larval stage. Single-mosquito analysis also showed that fecundity and viability from eggs of RNAi-EOF1 females were strongly affected by reduced EOF1 function through RNAi (Fig 1F and 1G). Bleaching experiments on eggs further confirmed that mosquito-specific EOF1 is required for embryonic development in A. aegypti mosquitoes. Under a light microscope, we observed eggs throughout the 2 h eggshell dechorionation experiment period. In the majority of eggs laid by EOF1-deficient mosquitoes, embryos failed to complete embryogenesis and reach the first larval instar (S2 Fig). It has been shown that when a mosquito embryo advances to form a serosal cuticle within the eggshell, bleach treatment was found to only remove the eggshell, but it cannot digest the serosal cuticle, leaving an intact embryo [36]. Thus, if a serosal cuticle has been formed during embryogenesis, the embryo or developed larva should be resistant to bleach treatment. A recently colonized A. aegypti Tucson strain from wild populations [37] also exhibited similar defective egg and embryo phenotypes associated with RNAi-EOF1 (Fig 1H). We analyzed 15 mosquitoes from each group, and the fecundity (mean ± standard error [SE]) of eggs was 33.3 ± 3.0, 26.4 ± 2.8, and 17.5 ± 1.8 (unpaired Student's t test; p < 0.01) for untreated, RNAi-Fluc, and RNAi-EOF1, respectively. The lower number of eggs laid by the Tucson strain compared to the Rockefeller strain of A. aegypti could likely be due to reduced blood ingestion. The viability of eggs (mean ± SE) was 88.4 ± 2.1, 87.0 ± 1.7, and 2.4 ± 0.9% (unpaired Student's t test; p < 0.001) for untreated, RNAi-Fluc, and RNAi-EOF1, respectively. Thus, EOF1 protein is essential for complete eggshell formation and embryonic development in A. aegypti mosquitoes.
Anautogenous female mosquitoes can undergo multiple gonotrophic cycles by repeating blood feeding, vitellogenesis, and oviposition events. Because EOF1 plays an essential role in eggshell formation, we wondered how long the RNAi knockdown effect of EOF1 lasts from a single dsRNA microinjection. We examined the effect of EOF1 deficiency on eggs in three consecutive gonotrophic cycles in individual containers as designed in Fig 2A and Fig 1B. Eggshell melanization (Fig 2B and 2C), fecundity (Fig 2D), and viability (Fig 2E) phenotypes are profoundly altered in EOF1-deficient mosquitoes during the first three gonotrophic cycles. Therefore, our data demonstrate that the RNAi-EOF1 effect from a single dsRNA injection remains substantial for the second and even the third gonotrophic cycles. Furthermore, we found that the timing of dsRNA microinjection is important. The dsRNA has to be microinjected a few days prior to blood feeding in both the first and second gonotrophic cycles in order to induce RNAi-mediated EOF1 depletion and produce defective egg phenotypes (S3 Fig).
Since little is known about this mosquito-specific EOF1 gene except for the phenotypes associated with RNAi, we determined the expression pattern of EOF1 at the mRNA level in untreated A. aegypti by quantitative real-time PCR (qPCR). Five tissues including thorax, fat body, midgut, ovaries, and Malpighian tubules were dissected from sugar-fed female mosquitoes at 3 days posteclosion and blood-fed female mosquitoes at 24 and 48 h post-blood meal (PBM). We also examined the mRNA expression in whole bodies of mixed-sex samples of fourth-instar larvae and pupae and adult male mosquitoes. EOF1 is predominantly expressed in ovaries, and the expression is up-regulated in the ovaries by blood feeding (Fig 3A). qPCR results also indicate that mRNA encoding EOF1 is not strongly detected from larvae, pupae, and adult male mosquitoes. We then examined the pattern of EOF1 expression during the first gonotrophic cycle in detail. Ovaries were isolated at various time points PBM. In ovary samples after 36 h PBM, the primary follicles were carefully isolated from ovaries to exclude nonfollicle ovarian cell types such as muscles and trachea. qPCR data show that EOF1 mRNA expression is up-regulated in response to blood feeding, and the levels remain high even at 14 days PBM (Fig 3B). Since follicular epithelial cells and nurse cells in the primary follicles undergo apoptosis by around 72 h PBM (S4 Fig), mRNAs encoding EOF1 may likely originate from the unfertilized mature oocytes. EOF1 mRNA distribution in primary follicles was further determined using whole-mount fluorescent in situ hybridization (FISH). FISH analysis shows that while three vitelline envelope genes [38] were exclusively expressed in the follicular epithelial cells, EOF1 mRNA transcripts are present in oocyte and nurse cells of primary follicles and weakly expressed in the secondary follicle and germarium (S5 Fig). Western blot analysis showed that EOF1 expression is induced in ovaries in response to blood feeding (Fig 3C). RNAi knockdown level of EOF1 mRNA and protein was confirmed by qPCR and western blot, respectively (S6 Fig).
In A. aegypti, EOF1-deficient female mosquitoes had low fecundity (Fig 1F) and laid eggs that were defective in eggshell formation, leading to the embryonic lethal phenotype (Fig 1G). Similar reproductive phenotypes associated with RNAi-EOF1 were found in A. albopictus (S7 Fig). EOF1 is required for proper eggshell formation, fecundity, and viability. We hypothesized that primary follicles of EOF1-deficient mosquitoes undergo cell death, removing severely affected follicles within the ovaries. We examined ovarian follicle phenotypes associated with EOF1 gene suppression by RNAi in A. aegypti mosquitoes. Representative ovaries at 36 h PBM showed that RNAi-EOF1 ovaries contain follicles that undergo caspase-mediated apoptosis indicated by the increase in red-labeled caspase inhibitor, while these dying follicles were not observed in untreated or RNAi-Fluc control ovaries (Fig 4). Approximately 40% of primary follicles in RNAi-EOF1 ovaries showed caspase-mediated apoptosis. Differential interference contrast and confocal images at a higher magnification showed that the caspase activity was more concentrated in the oocytes than in the follicular epithelial cells from RNAi-EOF1 mosquitoes (Fig 4G and 4H).
Mature ovaries were dissected from the abdomen of dsRNA-injected mosquitoes and photographed at 96 h PBM (Fig 5A–5D). While not all RNAi-Fluc control mature follicles in ovaries have initiated melanization, we frequently observed that some follicles isolated from RNAi-EOF1 are already partially melanized in the ovaries. The partially melanized phenotype in EOF1-deficient ovaries is accompanied by a loss of structural integrity, and thus we hypothesized that decreased chorionic osmotic control results in this alteration of egg shape. To determine whether the water permeability of the mosquito eggshells was affected in response to EOF1 knockdown, we employed two chemical markers, rhodamine B and neutral red, to stain ovarian follicles. Significant differences in the permeability of both markers in ovaries were observed (Fig 5E–5J). While the follicles from both untreated and RNAi-Fluc mosquitoes were only slightly stained, the majority of follicles from RNAi-EOF1 mosquitoes were strongly stained with the markers. Since follicular epithelial cells have been already shed around 72 h PBM (S4 Fig), there is a single oocyte present in each follicle at this developmental stage (96 h PBM). The reduction of EOF1 expression in female mosquitoes resulted in defective eggshells, leading to increased permeability of water into oocytes (Fig 5G and 5J) and altered follicular shape.
EOF1-deficient mosquitoes oviposited eggs with different degrees of eggshell melanization phenotypes that include nonmelanized, partially melanized, and melanized eggs (Figs 1E and 2C). Since nearly 100% of eggs oviposited from RNAi-EOF1 females did not undergo complete embryogenesis (Figs 1G and 2E), we hypothesized that the defective eggshell might be the primary cause of embryonic death. Light microscopy images of eggs from A. aegypti RNAi-Fluc and RNAi-EOF1 mosquitoes revealed that EOF1 may be involved in the specification of the outer chorionic area (OCA) surrounded by the porous nature of the exochorionic network (EN) (Fig 6A–6D). Next, we examined the effect of RNAi-EOF1 on the ultrastructure of eggs in detail by scanning electron microscopy (SEM). We observed a very similar A. aegypti eggshell ultrastructure (Fig 6E and 6G) to other SEM studies [39,40]. An exochorion outermost layer of the eggshell is characterized by the presence of a single protruding central tubercle (CT) and several minute peripheral tubercles (PTs) in the OCA (Fig 6G). However, SEM images showed that the OCA in RNAi-EOF1 eggs is about 6 times larger than eggs of control mosquitoes (Fig 6E–6H), suggesting that EOF1 may be involved in specifying the size of the OCA. A majority of eggshell proteins are likely secreted into the perivitelline space from follicular epithelial cells during follicle development in response to blood feeding. However, it is not well known whether the size of the OCA is strictly determined by surrounding follicular epithelial cells. We also observed that each OCA contains multiple miniaturized CT-like structures also surrounded by EN-like structures instead of one predominant CT. The SDS-PAGE analysis demonstrated that the enriched eggshell protein extracts from EOF1-deficient females showed slightly different patterns from those from RNAi-Fluc controls (S8 Fig). These differences could account for aberrant exochorionic structure in RNAi-EOF1 eggs. Thus, EOF1 may act as an upstream factor to control eggshell surface patterning in A. aegypti.
Data mining was performed using all A. aegypti protein sequences available at the GenBank database in early 2015 to identify putative protein-coding sequences that are only present in the genomes of Aedes, Culex, and Anopheles mosquitoes. Through RNAi screening of putative mosquito-specific genes in A. aegypti, we identified EOF1 as an essential protein for eggshell formation and melanization. During this course of study, a whole-genome analysis in 37 dipteran species, including midges and sandflies, was performed [41]. We performed a TBLASTN search against Nematocera in the Whole-Genome-Shotgun contigs database. A global alignment between mosquito EOF1 and hypothetical proteins detected in culicoid and chaoborus midges shows only about 26%–30% identity, suggesting that it is very difficult to determine whether these highly diverged genes show orthologous relationships. The conserved amino acid residues are present in putative F-box functional domains of these proteins. Taken together, EOF1 may be uniquely evolved in mosquito lineages to play roles in eggshell formation.
We observed that EOF1-deficient mosquitoes lay eggs with different melanization levels, ranging from nonmelanized to completely melanized eggshells (Fig 1C). Different level of eggshell melanization in response to RNAi-EOF1 could be possible because of genetic variation. Alternatively, a differential dsRNA uptake by follicles within the ovary could explain the variation in the levels of eggshell melanization. A single layer of follicular epithelial cells surrounding the oocyte is mainly responsible for secreting a majority of eggshell structural components. Since eggshell components are directly secreted into the extracellular space between the oocyte and the surrounding follicular epithelial cells (S1 Fig), intimate communication between these cells within each ovariole may exist throughout follicle maturation, eventually leading to follicular epithelial cell shedding (S4 Fig), ovulation, and oviposition. In general, mature follicles from mosquitoes do not undergo premature melanization within the ovaries (Fig 5), and gravid females can hold their mature follicles for a long period of time under adverse environmental conditions and still lay viable eggs, which become melanized after oviposition. Thus, the timing of eggshell melanization may likely be tightly regulated and catalyzed by specific enzymes, and their synthesis, secretion, and activation may be critical for proper melanization and thus survival of embryos [22–28]. A possible explanation for aberrant partial melanization of follicles within the EOF1-deficient ovaries prior to an oviposition event is that a loss of EOF1 function may alter hemolymph permeability of the eggshell, affecting a delicate chemical balance within the oocytes, which in turn trigger other eggshell components to prematurely initiate eggshell melanization processes. In addition to A. aegypti, we also confirmed that EOF1 plays an essential role in eggshell formation in A. albopictus (S7 Fig). Given that A. aegypti embryogenesis completes by 3 days postoviposition [36], two lines of our experimental evidence suggest that EOF1 may be necessary for complete embryogenesis in A. aegypti mosquitoes. First, we frequently observed that a majority of eggs deposited by EOF1-deficient mosquitoes collapsed within 16 h, rupturing the oocyte plasma membrane and leaking intracellular contents, including yolk, likely because of incomplete eggshell formation. Second, our bleach assay demonstrates that nearly 100% of eggs from EOF1-deficient mosquitoes with any melanization levels did not reach the larval stage (S2 Fig). Thus, female mosquitoes without EOF1 produce inviable eggs likely because of incomplete embryogenesis. However, it remains to be determined which specific embryogenesis stage was affected in eggs deposited by EOF1-deficient mosquitoes. SDS-PAGE analysis shows that enriched eggshell proteins from EOF1-deficient eggs slightly differ from those from RNAi-Fluc controls (S8 Fig), and thus an identification of the downstream EOF1-dependent eggshell proteins may lead to a better understanding of molecular mechanisms for mosquito eggshell formation. Although it is beyond the scope of this study to screen and identify small molecules that specifically target mosquito EOF1, such molecules may have great promise for controlling the mosquito population. Eggshell proteins specifically affected by EOF1-deficient mosquitoes may also be ideal proteins if they exhibit a high degree of sequence divergence to other insect taxa.
Based on the presence of a conserved F-box motif in EOF1, one possibility is that EOF1 is required in the ubiquitin pathway for controlled degradation of one or more proteins that regulate proper timing of eggshell development. Dysregulation of stage-specific ordered events in RNAi-EOF1–injected mosquitoes could lead to collapse of the developmental program at all downstream control points. The finding that RNAi-EOF1 phenotypes are observed three gonotrophic cycles beyond the time of injection (Fig 2) indicates that the EOF1 protein may not be resynthesized at the onset of each gonotrophic cycle but rather establishes the eggshell development program when the reproductive phase is initiated in the female mosquito. The data in Fig 3 support this proposal in that EOF1 mRNA remains at elevated levels in the ovaries of blood-fed mosquitoes, even out to 14 days. Specifically, if EOF1 mRNA was resynthesized with each gonotrophic cycle, then one would expect EOF1 mRNA to be degraded at completion of the gonotrophic cycle in order to restart the process after the next blood feeding, and that does not appear to be the case. Another possibility is that RNAi effects are particularly long-lasting in ovary tissues and continue to abrogate EOF1 synthesis at each gonotrophic cycle. Additional studies are currently underway to distinguish between these alternative hypotheses. Results from these studies could lead to the development of mosquito control applications using novel biosafe mosquitocides that target mosquito-specific proteins required for embryonic development. It may also be possible to use clustered regularly interspaced short palindromic repeats (CRISPR) and CRISPR-associated protein 9 (CRISPR-Cas9) gene-drive genetic manipulation [42–46] for the same purpose. If successful, such approaches would eventually result in a decrease in the mosquito population and thus lower the transmission of mosquito-borne viral infections.
Most of the experiments were carried out using A. aegypti mosquitoes (Rockefeller strain) and reared as previously described [5]. For comparison, A. aegypti mosquitoes (Tucson strain) were colonized from Tucson, Arizona [37]. A. albopictus (Gainesville strain, MRA-804) mosquitoes were obtained from CDC/MR4. Mosquitoes were maintained at 28°C, 72% relative humidity with a photoperiod of 16 h light and 8 h dark cycle in a CARON 6010 Insect Growth Chamber (Caron Products & Services, Marietta, OH, USA). The larvae were fed on a diet consisting of dog food, Tetramin fish food, and liver powder (10∶10:1 ratio). Male and female adults were maintained on 10% sucrose and kept together throughout all experiments until transferred to oviposition containers. Using an artificial glass feeder, female mosquitoes were allowed to feed on expired human blood donated by the American Red Cross (approved protocol #2010–017). Only fully engorged female mosquitoes were used.
Data mining and bioinformatic analysis were carried out using all A. aegypti GI numbers available at the GenBank database. To identify putative protein-coding sequences that are only present in the genomes of Aedes, Culex, and Anopheles mosquitoes, we use NCBI BLASTP with default search methods and a cutoff for expected value threshold of 1e-15. Proteins selected were further subjected to NCBI TBLASTN search using A. aegypti ESTs or expressed orthologs in the A. albopictus TSA database to determine whether the genes are expressed. Mosquito-specific putative genes without corresponding mRNA in A. aegypti ESTs or A. albopictus TSA database were excluded as candidate genes for RNAi screening. We also excluded genes that appeared to be members of a multigene family because of possible functional redundancy with other gene family members. We also used a TBLASTN search from NCBI against Nematocera (taxid: 7148) in the Whole-Genome-Shotgun contigs database.
RNAi was carried out to knock down A. aegypti mosquito genes. Each gene-specific forward and reverse oligonucleotide primer was designed using a NetPrimer web-based primer analysis tool. A T7 RNA polymerase promoter sequence, TAATACGACTCACTATAGGGAGA, was added to the 5′ end of each primer (S1 Table). All primers were purchased from Eurofins Genomics (Louisville, KY, USA). PCR was performed using the Taq 2X Master Mix (NEB, Ipswich, MA, USA) with mosquito whole-body complementary DNA (cDNA) as a template, and the amplified PCR products were cloned into the pGEM-T easy vector (Promega Madison, WI, USA) for DNA sequence verification using an ABI 377 automated sequencer (Applied Biosystems, Foster City, CA, USA). dsRNA was synthesized by in vitro transcription using a HiScribe T7 Quick High Yield RNA Synthesis Kit (NEB). The purified dsRNA was resuspended with HPLC-grade water (Thermo Fisher Scientific, Waltham, MA, USA) at 7.3 μg/μL concentration. Cold-anesthetized female mosquitoes were injected with 2.0 μg dsRNA (276 nL) using a Nanoject II microinjector (Drummond Scientific Company, Broomall, PA). Injected mosquitoes were maintained on 10% sucrose throughout the experiments.
Using TRIzol reagent (Thermo Fisher Scientific), total RNA was extracted from larvae, pupae, and male adults as well as five tissues including thorax, fat body, midgut, ovaries, and Malpighian tubules dissected from sugar-fed female mosquitoes at 3 days posteclosion and blood-fed female mosquitoes at 24 and 48 h PBM. First-strand cDNA was synthesized from pools of total RNA using an oligo-(dT)20 primer and reverse transcriptase. qPCR was carried out with the corresponding cDNA, EOF1, or ribosomal protein S7 gene-specific primers (S2 Table), PerfeCTa SYBR Green FastMix, and ROX (Quanta BioSciences, Gaithersburg, MD, USA) on the 7300 Real-Time PCR System (Applied Biosystems).
Mosquito ovaries were dissected in 1× PBS under a dissecting microscope and homogenized in lysis buffer (12 mM sodium deoxycholate, 0.2% SDS, 1% triton X-100, complete mini EDTA-free protease inhibitor; Roche Applied Science, Indianapolis, IN, USA). Protein extracts were resolved on SDS-PAGE using a 12% acrylamide separation gel and a 3% stacking gel. The resolved proteins were either stained with GelCode Blue reagent (Thermo Fisher Scientific) or electrophoretically blotted to a nitrocellulose membrane (LI-COR, Lincoln, NE, USA) for western blot analysis. The membranes were blocked with 4% nonfat dry milk and incubated with each primary antibody in 4% nonfat milk in PBS containing 0.1% Tween 20. The EOF1 rabbit polyclonal antibody was generated by GenScript Corporation (Piscataway, NJ, USA) based on an antigenic peptide (LAPNSPSKEDEPAH). The anti-α-tubulin monoclonal antibody from Developmental Studies Hybridoma Bank (University of Iowa, Iowa City, IA, USA) was used as loading controls for ovaries. The dilutions of the primary antibodies were as follows: EOF1 (1:3,000) and α-tubulin (1:2,000). The secondary antibodies were either IRDye 800CW goat anti-rabbit secondary antibody (1:10,000; LI-COR) or IRDye 800CW goat anti-mouse secondary antibody (1:10,000; LI-COR). The protein bands were visualized with an Odyssey Infrared Imaging System (LI-COR).
Knockdown efficiency of RNAi was verified by real-time qPCR using gene-specific primers (S2 Table). cDNA was synthesized from DNase-I–treated total RNA isolated from ovaries of individual dsRNA-injected mosquitoes at 48 h PBM. Normalization was done using the ribosomal protein S7 transcript levels as an internal control, and the knockdown efficiency of RNAi-EOF1 was compared using Fluc-dsRNA–injected mosquitoes as a control. The RNAi knockdown level of EOF1 protein was also determined by western blot analysis using an EOF1-specific polyclonal antibody. Ovarian protein extracts were isolated from 8 individual mosquitoes from RNAi-Fluc or RNAi-EOF1 mosquitoes at 48 h PBM. α-tubulin was used as an internal control.
Eggs laid on oviposition papers remained wet for 3 days before drying at 28°C. Eggs (about 7 days old) on oviposition paper were submerged in water, vacuumed using a Speed Vac for 10 minutes, and allowed to hatch for 2 days. First-instar larvae were counted.
A bleach assay was performed to determine viability of 4-day-old A. aegypti eggs from RNAi studies. Eggs on oviposition paper were soaked in 12% bleach (sodium hypochlorite concentration at 0.72%) at room temperature. A gradual progress of dechorionation of eggshell was observed under a microscope. Light microscopic images of eggs deposited from RNAi-Fluc and RNAi-EOF1 females prior to and after the addition of bleach were taken at 49× magnification (Nikon, SMZ-10A; Nikon, Tokyo, Japan).
mRNA distribution of EOF1 and vitelline envelopes (15a1, 15a2, and 15a3) in A. aegypti primary follicles was determined using whole-mount FISH. Primary follicles were isolated from ovaries of untreated female mosquitoes at 36 h PBM fixed with 4% paraformaldehyde. After washing with 1× PBS, the follicle samples were dehydrated with ethanol (ETOH) in water through a graded series for 10 min each in 10%, 30%, 50%, 70%, and 90% ETOH and 3 times 30 min each in 100% ETOH at room temperature. The samples were hydrated with 1× PBS in ETOH through a graded series for 20 min each in 25%, 50%, 75%, and 100% 1× PBS at room temperature. The follicles were permeabilized with proteinase K, postfixed with 4% paraformaldehyde with 0.1% Tween 20, and treated with DEPC (0.1%) to inactivate RNase. The follicles were then hybridized with digoxigenin-labeled antisense or sense RNA probes overnight at 65 oC. Probe DNA templates were PCR amplified by gene-specific primers (S3 Table), and the RNA probes were synthesized by in vitro transcription as described in dsRNA synthesis above with DIG RNA Labeling Mix (Sigma-Aldrich, St. Louis, MO, USA). After washing, the follicles in PBS were stained for actin cytoskeleton using Acti-stain 488 phalloidin-labeled (Cytoskeleton, Denver, CO, USA) at room temperature and incubated with rhodamine-B–conjugated anti-digoxygenin antibody (1:500 dilution; Jackson ImmunoResearch Laboratories, West Grove, PA, USA) in blocking buffer (LI-COR) to detect the hybridized probes at 4 oC. The follicles were mounted on glass slides and viewed on a spinning disc confocal microscope (Intelligent Imaging Innovations, Denver, CO, USA) at the Keck Imaging Center at the University of Arizona. Images were obtained by using excitation 488 and 561 nm lasers and recorded using identical exposure times (100 ms).
Female A. aegypti mosquitoes were microinjected with dsRNA at 1 day post-adult emergence, and ovaries at 36 h PBM were removed in 1× PBS under a dissecting microscope and immediately incubated with tissue culture media (Medium 199, Thermo Fisher Scientific) containing a caspase inhibitor (SR FLICA Poly Caspase Assay Kit; ImmunoChemistry Technologies, Bloomington, MN, USA) at 37 oC in the dark for 1 h. The ovaries were washed with 1× PBS, fixed with 4% paraformaldehyde, quenched with 25 mM glycine, permeabilized with 0.5% Triton X100, and stained with Acti-stain 488 phalloidin overnight at 4 oC. After washing with 1× PBS, the whole ovaries were mounted on a glass slide using ProLong Gold Antifade reagent (Thermo Fisher Scientific). Immunofluorescence, differential interference contrast, and light microscopic images of the ovaries were captured using a Spinning Disk Confocal Laser Microscope in the Keck Imaging Center at the University of Arizona.
The assay has an advantage in that it can quickly assess whether follicles within the ovaries may contain defective eggshell prior to oviposition. Individual follicles of untreated, RNAi-Fluc, or RNAi-EOF1 mosquitoes at 96 h PBM were dissected and gently separated from the ovaries and transferred to glass scintillation vials. Rhodamine B (final concentration of 1 mM in H2O, Sigma-Aldrich) and neutral red (0.5%, Sigma-Aldrich) were used to stain primary follicles for 10 min on a rocking shaker and thoroughly rinsed with H2O. The stained primary follicles were photographed with a Coolpix 4300 (Nikon).
The ovaries were dissected from mosquitoes injected with Fluc control dsRNA and EOF1 dsRNA at 96 h PBM. Each follicle was carefully separated from the ovaries in 1× PBS under a dissecting microscope. The mature follicles were fixed in 2.5% glutaraldehyde in 0.1 M PIPES for 1 h at room temperature and washed twice in PIPES. The follicles were then postfixed in 1% osmium tetroxide in PIPES for 1 h and washed twice in deionized water for 10 min each. The follicles were dehydrated with ETOH in water through a graded series for 10 min each in 10%, 30%, 50%, 70%, and 90% ETOH and 3 times 30 min each in 100% ETOH at room temperature. The samples were dried with hexamethyldisilazane (HMDS; Electron Microscopy Sciences, Hatfield, PA, USA) in ETOH through a graded series for 20 min each in 25%, 50%, 75%, and 100% HMDS at room temperature. The follicle samples were air dried under a fume hood overnight at room temperature for SEM analysis. The dried samples were metallized with gold using Hummer 6.2 Sputter System (Anatech USA, Union City, CA, USA). Inspect-S scanning electron microscope (FEI, Hillsboro, OR, USA) was used to compare the ultrastructural characteristics of the ovarian follicles of females injected with Fluc and EOF1 dsRNA.
Female mosquitoes were injected with dsRNA at 1 day post-adult emergence, and ovaries were dissected in 1× PBS at 96 h PBM. The dissected ovaries were thoroughly homogenized (40 strokes) in ice-cold 1× PBS using Dounce homogenizers (B pestle). The eggshells were allowed to settle down to the bottom of the homogenizer on ice. The top cloudy fraction was gently aspirated, and the washing step with ice-cold 1× PBS for the eggshells was repeated four times or until the solution was completely cleared. Subsequently, the eggshell was homogenized (20 strokes) in ice-cold 1× PBS using A pestle. The eggshell proteins were subjected to SDS-PAGE and stained with GelCode Blue reagent.
Statistical analyses were performed using GraphPad Prism Software (GraphPad, La Jolla, CA). Statistical significance for fecundity, viability, and RNAi knockdown efficiency was analyzed using an unpaired Student's t test. p values of ≤0.05 were considered significantly different. All experiments were performed from at least three independent cohorts.
|
10.1371/journal.pntd.0004620 | Use of Molecular Diagnostic Tools for the Identification of Species Responsible for Snakebite in Nepal: A Pilot Study | Snakebite is an important medical emergency in rural Nepal. Correct identification of the biting species is crucial for clinicians to choose appropriate treatment and anticipate complications. This is particularly important for neurotoxic envenoming which, depending on the snake species involved, may not respond to available antivenoms. Adequate species identification tools are lacking. This study used a combination of morphological and molecular approaches (PCR-aided DNA sequencing from swabs of bite sites) to determine the contribution of venomous and non-venomous species to the snakebite burden in southern Nepal. Out of 749 patients admitted with a history of snakebite to one of three study centres, the biting species could be identified in 194 (25.9%). Out of these, 87 had been bitten by a venomous snake, most commonly the Indian spectacled cobra (Naja naja; n = 42) and the common krait (Bungarus caeruleus; n = 22). When both morphological identification and PCR/sequencing results were available, a 100% agreement was noted. The probability of a positive PCR result was significantly lower among patients who had used inadequate “first aid” measures (e.g. tourniquets or local application of remedies). This study is the first to report the use of forensic genetics methods for snake species identification in a prospective clinical study. If high diagnostic accuracy is confirmed in larger cohorts, this method will be a very useful reference diagnostic tool for epidemiological investigations and clinical studies.
| Snakebite is an important medical problem in sub-tropical and tropical regions, including Nepal where tens of thousands of people are bitten every year. Snakebite can result in life-threatening envenoming, and correct identification of the biting species is crucial for care providers to choose appropriate treatment and anticipate complications. This paper explores a number of methods, including molecular techniques, to assist care providers in identifying the species responsible for bites in rural Nepal. Out of 749 patients with a history of snakebite, the biting species could be identified in 194 (25.9%). Out of these, 87 had been bitten by a venomous snake, most commonly cobras (n = 42) and kraits (n = 22). This study is the first to report the use of molecular techniques for snake species identification. The diagnostic accuracy of this method appears high but needs to be confirmed in larger studies.
| In rural Nepal snakebite is an important public health problem. A survey conducted in the 1980s showed that about 20’000 people were bitten each year, resulting in over 1’000 deaths [1]. These official figures, however, significantly underestimate the true burden [2–7]. Annual incidence and mortality figures of 1’162/100’000 and 162/100’000, respectively, have been reported in some regions of Nepal [4]. Children are among the primary victims [2,8,9]. Most snakebites occur in the southern plains of Terai, a region with a hot tropical climate and high population density. Most bites occur during the rainy season, from June to September, which corresponds to the peak period for agricultural work.
Eighty nine snake species have been recorded in Nepal, of which 17 are known to be venomous [10]. The Elapidae family includes two species of cobra (Naja naja and Naja kaouthia), the king cobra (Ophiophagus hannah), one species of coral snake (Sinomicrurus macclellandii) and six species of krait (Bungarus bungaroides, Bungarus caeruleus, Bungarus fasciatus, Bungarus lividus, Bungarus niger, and Bungarus walli). The Viperidae are represented by seven species. The most dangerous of these, Russell’s viper (Daboia russelii), appears to be rare in Nepal. Pitvipers (Gloydius himalayanus, Himalayophis tibetanus, Ovophis monticola, Protobothrops spp. Trimeresurus albolabris and Trimeresurus septentrionalis), on the other hand, are widespread from the lowlands to the high mountains. Non-venomous species are also very common and may be involved in bites. Some of these non-venomous species are easily mistaken for venomous ones [10]. For example, rat snakes (Ptyas and Coelognathus species) may be confused with cobras, while wolf snakes, which are common inside and around houses, have a colour pattern similar to that of kraits [11–13]. Bites can therefore be inflicted by a variety of species, in all kinds of environments. Neither the geographical distribution of these species nor their contribution to snakebite mortality and morbidity have been systematically studied in Nepal.
Snakebite envenoming can be life-threatening, and recognizing early signs of systemic envenoming is crucial for the optimal management of patients. Depending on the species of snakes, different organs and tissues can be affected [14]. In envenoming following the bites of elapid snakes, neurotoxicity with progressive descending paralysis is characteristic, and patients usually die of respiratory failure if not adequately ventilated [15–17]. Clinical prognosis and response to antivenom depend on the species, hence knowing which snake is responsible for the bite is of primary importance [14]. In South Asia appropriate tools for species identification are not available. The snake is rarely seen, and if it is, its description by the victim is often misleading [18]. Hospital personnel are generally not trained to properly identify the biting species, even when it is killed and brought by the victim [19–21]. The morphological resemblance of venomous by non-venomous species complicates this task.
Several approaches have been investigated to improve snake species identification in peripheral health centres. Immunodiagnosis of circulating venom antigen can be used to identify the biting species or ‘immunogroups’ of antigenically similar, cross-reacting venoms [22–25], but commercial point-of-care venom detection kits have been marketed only in Australia, for species occurring on that continent [26]. In Sri Lanka, syndromic approaches have been proposed [20] and clinical scores developed based on a systematic analysis of features of envenom envenoming in cases where snake specimens brought to hospital were expertly identified [27]. However, the extent to which these approaches can also be applied to the Nepal setting is not known. The present study aimed at investigating the accuracy, feasibility and usefulness of different, complementary approaches to species identification in patients bitten by snakes in rural Nepal. The main objectives were (1) to clarify the contribution of different snake species causing envenoming and non-envenoming bites through morphological analysis of preserved snake specimens and the use of forensic methods for DNA-based identification, and (2) to explore the utility, diagnostic performance and feasibility of DNA-based identification by forensic methods.
The study was conducted in three centres, namely the Snake Bite Treatment Centre of Damak Red Cross Society and the Snake Bite Management Centre of Charali, both in Jhapa district, and Bharatpur District Hospital, Bharatpur, Chitwan district. All three centres are located in the Terai lowland region of Nepal.
All snakebite victims presenting to the Damak Red Cross Centre and to the Charali Snake Bite Management Centre during the study period were offered to participate in the study, irrespective of the presence and nature of envenoming signs. Hence, victims of bites by elapids, viperids and non-venomous snakes were included. Patients were excluded from the study if they were below 5 years of age or if they had already received antivenom prior to admission. Patients unable or unwilling to give consent were also excluded. In Bharatpur District Hospital, additional exclusion criteria were applied due to the contemporary conduct of a randomized controlled trial on neurotoxic envenoming (clinicaltrials.gov number NCT01284855). In this centre, only snakebite patients presenting with signs of systemic neurotoxic envenoming were included in the study. Those presenting more than 24 hours after the bite, pregnant or breastfeeding women and patients with proven viper bite envenoming were excluded. The recruitment period in Bharatpur was shorter and extended from 1st April 2011 to 31st October 2012.
Dead snakes brought by bite victims were systematically labelled with patient number, initials, date of birth and date of admission, and preserved in 70% ethanol. Morphological identification was conducted by taxonomic experts (UK, DP) who remained blind to the circumstances of bites and to envenoming status of the victims. For each preserved specimen, morphological characters of scalation and dentition were analysed by comparison to relevant reference specimens in museum collections and an existing database [28,29]. Additionally, genetic information from tissue samples of killed snakes were obtained in the form of DNA ‘barcodes’ of the mitochondrial cytochrome b gene [28,29].
Whenever the bite site could be located, trace DNA of the biting snake was collected by rubbing the cotton swab of a Prionix evidence collection tube on the bite site (see Standard Operating Procedure in supplementary material). The sample was left to dry at room temperature. For DNA extraction, one half of the cotton bud was cut off and subjected to lysis in CTAB (cetyl trimethylammonium bromide) buffer with proteinase K at 56°C, followed by one extraction with phenol-chloroform-isoamyl alcohol and then with chloroform-isoamyl alcohol. DNA was precipitated with 98% ethanol and 3M sodium acetate (pH 5,2) and stored overnight at -20°C before washing with 70% ethanol, drying and dissolving in TE-buffer. For PCR we used primers flanking a 400 bp sequence of the mitochondrial cytochrome b gene (cytb) which preferentially amplify snake rather than human cytb sequences. For nested PCR we designed primers with cytb binding sites upstream and downstream of the PCR primers. PCR was performed in 30 μl volumes containing 6 μl of template DNA solution, 0.3 pM of each primer, 83 pM of each dNTP, 3 μl 10 × PCR buffer without MgCl2 (Fermentas), 3.25 mM MgCl2, 1 unit TrueStart Hot Start Taq polymerase (Fermentas), and 13.7 μl H2O. PCR products were visualized by SYBRGreen (Invitrogen) staining in 1% agarose gels and ultraviolet (302 nm) light illumination. Their lengths were estimated using a 100 bp ladder (Fermentas). Cycle sequencing of both strands was performed with 0.16 μl of BigDyeTerminator 3.1 reaction mix (Applied Biosystems), 1 μl primer, 1.92 μl 5 × sequencing buffer (Applied Biosystems) and 5.92 μl milliQ H2O (Millipore). Products of the sequencing reaction were separated on an ABI 3730 sequencer (Applied Biosystems) operated with a 50 cm capillary, 8 sec injection time and 1500 V injection voltage. After visual checks of electropherograms, correction of base-calls and comparison of complementary strands using BioEdit and SequenceScanner software, sequences were submitted to BLAST searches for comparison with the sequences in the GenBank, ENA and DDBJ nucleotide sequence database to determine the snake species. New DNA sequences obtained in the course of this study were deposited in this database. A full description of the methods is included in Melaun and Kuch [30].
Local envenoming was defined as the presence of one or more of the following: (1) necrosis, (2) bullae or blisters, (3) enlarged regional lymph nodes plus either local bleeding or ecchymosis or swelling and (4) swelling extending at least halfway between two articulations. Snakebite victims presenting with moderate local swelling (i.e., not extending farther than one articulation) were not considered as locally envenomed. Systemic envenoming was defined as the presence of one or more of the following: (1) Incoagulable blood as indicated by the 20 Minutes Whole Blood Clotting Test (20’WBCT), (2) spontaneous and continuous bleeding from the bite site, IV line, gums or old wound, (3) gastrointestinal bleeding or blood in urine, (4) neurotoxic sign(s) including inability to frown, bilateral ptosis, inability to open the mouth, inability to protrude the tongue beyond incisors, inability to clear secretions, broken neck sign, skeletal muscle weakness, gag reflex loss and paradoxical breathing. Neurotoxic envenoming was defined as the presence of one or more of the above-mentioned neurotoxic signs. Envenoming status was determined retrospectively, by one of the authors (EA) who was blinded to the results of snake species identification.
The clinical management of snakebite victims followed the Nepal national protocol and WHO SEARO guidelines [14]. All patients presenting with systemic envenoming received antivenom, as did patients presenting with severe local envenoming (e.g., edema extending over full limb). Antivenoms available in Nepal are all manufactured in India and target four species, the Indian spectacled cobra (Naja naja), the common Indian krait (Bungarus caeruleus), the saw-scaled viper (Echis carinatus) and Russell’s viper (Daboia russelii). Additional treatments included anticholinesterases and assisted ventilation for those patients experiencing respiratory paralysis. Patients who were asymptomatic on admission were kept under observation for 24 hours.
A standard Case Report Form (CRF) was designed to prospectively collect data on the circumstances of the bite, the participants’ demographic characteristics, and the clinical features on admission. First aid measures, whether appropriate or not, were recorded, as well as features of clinical management after admission. To ensure harmonization of data collection, study staff were trained on CRF completion and were instructed to follow Standard Operating Procedures (SOP). Separate forms were used for the results of morphological identification and molecular analysis. As these analyses were performed well after recruitment was over, their results did not influence the assessment of snakebite victims by the care-providers. The researchers performing the PCR and DNA sequence analyses were also blind to the results of morphological identification and patient data.
Characteristics on admission were described using percentages for categorical variables and by calculating means and standard deviations or median and Inter Quartile Ranges (IQR) for continuous variables. The Case Fatality Rate (CFR) was calculated as the percentage of patients who died among those showing signs of envenoming. Descriptive analyses were restricted to participants with snake species identified by PCR or morphology. Determinants of PCR positivity were analysed for all patients for whom a PCR was done, as follows: continuous variables were compared using Student’s t test, or the Wilcoxon- rank sum test for non-normally distributed variables. Categorical variables were compared using the Χ2 test with continuity correction or the Fisher exact test, as appropriate. For categorical variables with more than one category, a test for trend was used. Associations were examined at the p<0.05 level of significance. To assess the probability of a positive PCR, Risk Ratios (RR) with 95% Confidence Interval (95% CI) were calculated.
The study was conducted in accordance with the Declaration of Helsinki 1964, as revised in Seoul, 2008, and in compliance with the protocol, Good Clinical Practices (GCP) and Nepal regulatory requirements. The B.P. Koirala Institute of Health Sciences (BPKIHS) Ethics Committee, the Nepal National Health Research Council (NHRC) and Geneva University Hospital Ethics Committee approved the study prior to its start. All participants provided written informed consent before being included in the study.
Between the 1st of April 2010 and the 31st of October 2012, a total of 749 patients were found eligible to be included in the study (Fig 1).
The responsible snake species could be ascertained for 194 (25.9%) patients. In total 722 snakebite victims presented with fang marks, and swabs were collected and analysed from 565 of them. PCR and DNA sequencing yielded a positive result in 153 out of these (27.1%). Sixty two snakes were brought by the victim and all were morphologically identified. Collectively the snake species could be determined in 87 (32.9%) of 264 envenomed patients and in 107 (22.1%) of 485 non-envenomed patients. Table 1 lists the species identified in all three centers. Among the non-venomous species, the checkered keelback (Xenochrophis piscator) was responsible for the majority of bites. Among the venomous species identified the most common were the spectacled cobra (N. naja) and common krait (B. caeruleus). There was a very clear difference between study sites. Most B. caeruleus specimens (20 out of 22) were brought in Bharatpur, while most N. naja (35 out of 42) came from either Charali or Damak. In 21 cases, both morphological identification and a snake DNA sequence were available. In these cases there was a 100% agreement between the two identification methods.
The baseline characteristics and circumstances of the bites of the 194 snakebite victims for whom a species could be ascertained, are summarized in Table 2. The majority of snakebite victims were male (52.6%) and the median age was 30.8 years (IQR = 18–40). There were 28 (14.4%) children (5–15 years). A clear seasonal pattern was observed in all three study sites, with the incidence of bites being higher during the rainy season (June to September).
The median time needed to reach the treatment centre was 85 minutes (IQR = 50 to 120 minutes), with a longer time observed in Bharatpur (194 minutes; IQR: 11 to 269 minutes). Only four (2.1%) snakebite victims visited a traditional healer before reaching the study centre. Use of first aid measures was more common in Damak (n = 91/93; 97.8%) and Charali (n = 72; 98.6%) compared to Bharatpur (n = 11; 39.5%). Tourniquets were by far the most common first aid method applied.
S1 Table compares baseline characteristics, circumstances of the bite and first aid measures between snakebite victims with identified and unidentified snake species.
Among snakebite victims presenting to the treatment centres in Damak or Charali during the study period (n = 676), 191 (27.4%) showed signs of envenoming. Twenty three (24.7%) patients in Damak and four (5.5%) in Charali presented with neurotoxic signs. As expected by local inclusion criteria, all 73 patients recruited in Bharatpur had signs of systemic neurotoxic envenoming on admission, and among those, seven also had local signs (see S1 Text).
The signs and symptoms presented on admission by the 194 snakebite victims for whom a species identity could be ascertained are described in Table 3. Among 87 victims for whom a venomous species was identified, 16 (18.4%) had not developed any sign of envenoming. Among the 73 patients for whom an elapid snakebite could be ascertained, 64 (87.7%) developed signs of envenoming. As expected, none of the 107 patients bitten by non-venomous species exhibited local or systemic signs of envenoming.
Swelling of the bitten limb was present in 43 out of 194 patients (22.2%), and bleeding from the bite site was present in 60 (30.9%). Of note, both signs also occurred in patients bitten by non-venomous species, in particular 25 out of 88 X. piscator bite cases (28.4%) presented with bleeding from the bite site.
The 20 Minutes Whole Blood Clotting test was performed on all victims presenting to one of the study centres during the study period. Out of those with incoagulable blood on admission, the snake species could be determined in 7 cases: 5 were N. naja and 2 were white-lipped pit vipers (Trimeresurus cf. albolabris). No snakebite victim presented with gum bleeding, gut bleeding, or blood in urine on admission.
S2 Table compares clinical features on admission of snakebite victims with identified and unidentified snake species.
As the number of cases in which the snake species could be identified by PCR sequencing was low, we investigated baseline characteristics and circumstances of the bite that may influence the sensitivity of this method. Results are summarized in Table 4. The median time to reach the center was significantly longer in patients with a negative PCR. The probability of a positive PCR result was significantly lower among patients who had used first aid measures. In particular, applying local remedies (e.g., herbs, honey, etc.) was associated with a 2-fold decrease in the probability of a positive PCR. Patients bitten on the upper limb were 60% less likely to have a positive PCR compared to those bitten on the lower limb. Local bleeding also increased the probability of a positive PCR by 1.45 fold. Whereas bite by a venomous species did not affect the chances of a positive PCR (RR = 1.004, 95%CI: 0.893–1.128), showing signs of envenoming significantly increased the chances of a positive result (RR = 1.444, 95% CI: 1.091–1.912, p = 0.017).
A total of 67 (34.5%) snakebite victims received antivenom. Eight were put under mechanical ventilation and nine were transferred to a tertiary care centre. Among patients treated with antivenom the median dose was 2 vials (inter-quartile-range = 1–3).
In total there were 5 deaths (Case Fatality Rate = 5/194 = 2.6%): one in Damak, one in Charali and three in Bharatpur). Of these 5 deaths, 4 had been bitten by a B. caeruleus and 1 by a N. naja.
Between the 1st of April 2010 and the 31st of October 2012, 749 individuals with a history of snakebite presented to one of three study centres in southern Nepal. In 194 (25.9%) patients, the responsible snake species could be ascertained, by either morphological identification or PCR plus DNA sequencing. Most species identified were non-venomous ones. The non-venomous checkered keelback (X. piscator) was the most frequently identified species, followed by the spectacled cobra (N. naja) and the common krait (B. caeruleus). Other venomous species contributing to the snakebite burden in this study comprised several pitvipers (O. monticola, Trimeresurus sp., T. albolabris, and T. popeiorum) as well as various additional elapid snakes, including the first cases of envenoming by the greater black krait (Bungarus niger) and the king cobra (O. hannah) ever reported in Nepal.
As few victims (11.6%) brought the dead snakes to the study centres, the use of PCR to amplify snake trace DNA from bite-site swabs increased to 25.9% the proportion of victims for whom the snake species could be ascertained. Morphological identification of preserved specimens by a qualified herpetologist is the gold standard for species identification. However, this method is seldom used, as snakes are rarely captured and preserved, and as care-providers working in snakebite treatment centres generally lack the appropriate expertise [19–21]. Alternative approaches must therefore be developed to complement morphological identification. Molecular techniques have shown promising results in animal models [30–33], and the present study shows that PCR amplification of a mitochondrial gene region from snake trace DNA is feasible in field setting. Sampling is straight forward and requires minimal training (see SOP in supplementary material), and the storage and transport conditions for the Prionix evidence collection tubes (room temperature and protected from light) can easily be met.
The PCR yielded a positive result in 27.1% of the cases only. This could probably be improved if several factors shown in this study to be associated with a lower sensitivity, such as inappropriate first-aid measures (use of tourniquet or application of local remedies on the bite site) or prolonged time to reach the treatment centre, were corrected or improved by public health interventions. Although the sensitivity of the PCR did not seem to be affected by whether the species was venomous or not, envenoming status did have an effect. This may be due to confounding effects of venom injection. Bites associated with venom injection may go deeper in tissues, and snakes may deposit more trace DNA on bite sites during envenoming bites compared to ‘dry’ bites or bites by non-venomous snakes e.g., snake DNA can also be recovered from snake venom [33].
In all cases where both snake morphological identification and DNA sequence from the bite-site swab were available (n = 21), a 100% agreement was observed between the two methods, suggesting a high specificity of molecular identification. An on-going study conducted by the authors is expected to complement this preliminary data and validate PCR-aided sequencing of snake trace DNA on a larger cohort of patients. Although molecular tools are not yet appropriate for point-of-care (POC) testing and hence cannot be used to guide clinical management, the encouraging results presented here, if confirmed in larger studies, suggest that they could be used as reference tests in future epidemiological and clinical studies. Progress in the development and validation of POC tests for snake species has indeed been hindered by the difficult implementation of the diagnostic gold standard (morphological identification) in rural regions where most bites occur. Besides, molecular tools could be very useful in clarifying the contribution of different snake species to the snakebite burden, and help identify new medically important species.
The species found to have caused neurotoxicity in this study were the two species of cobra (N. naja and N. kaouthia) and three different species of krait (B. caeruleus, B. lividus, and B. niger). Of these, only two (N. naja and B. caeruleus) are included in the production of the Indian polyvalent antivenoms that are available in Nepal. No pre-clinical data exist on the efficacy of Indian polyvalent antivenoms to neutralize the venoms of these two species in Nepal, and no clinical data have been published so far. In this study, patients bitten by a krait had significantly higher chances of being put under mechanical ventilation and being transferred to an intensive care unit compared to those being bitten by a cobra (see supplementary information). This is consistent with published literature which suggests that krait bites often result in poorer outcomes for patients, and high mortality rates [34,35]. The efficacy of the available Indian antivenom in reversing envenoming by kraits is increasingly being questioned, and several case series have reported little or no benefit of immunotherapy [17,36–38].
The frequency of neurotoxicity observed in our study is consistent with elapid snake species being most commonly involved in envenoming bites in Nepal. The fact that we did not observe late-appearing signs such as broken neck sign, muscle weakness and loss of gag reflex may be due to the early presentation of victims to the health centre and the prompt initiation of antivenom therapy. Interestingly, among those patients who presented with incoagulable blood on admission were five victims of N. naja bites. This is not the first report of apparent coagulopathy following bites by this species [39], and a few in vitro studies have reported anticoagulant activities in N. naja venoms [40–42], however, these results need to be interpreted with caution. In fact, the 20 Minutes Whole Blood Clotting Test can give erroneous results if performed incorrectly, in particular if the tubes used bear traces of detergent [14].
The present study has several limitations, the principal one being that the study population differed between study sites. Snakebite victims admitted to Bharatpur District Hospital were only included if they presented with signs of neurotoxic envenoming. It is therefore not surprising that all snakes identified in this centre were venomous, resulting in an overestimation of the contribution of venomous species (and in particular elapids) to the snakebite burden. When Bharatpur was excluded from the analysis, venomous species accounted for only 35.6% of identified bites. The checkered keelback and the spectacled cobra remained the most common species identified (see supplementary information). Another limitation relates to the geographical coverage of the study. The list of species identified here is not representative of all snakes causing bites in Nepal. The three study centres are located in the lowlands of the Terai region where most snakebites occur, but their catchment areas do not cover all of Nepal’s great biogeographical diversity. In particular, species found in mountain regions (although present, e.g., O. monticola) were probably under-represented in our study.
The fact that the biting species could be identified only in a relatively small proportion of patients (25.9%) could in theory lead to bias. Although we cannot exclude that some selection bias occurred in the present study, its impact is likely to be minimal. We compared bite circumstances and baseline characteristics of snakebite victims with or without identification of snake species (S1 and S2 Tables). Differences were seen with regard to season of bite, location and activity at the time of bite and consultation of a traditional healer. However, the magnitude of these differences was minimal. Moreover, the epidemiological characteristics of our study population are consistent with other published reports [2–5,8,43,44], further ruling out the possibility of selection bias and reaffirming the external validity of our findings.
Finally, morphological identification and molecular analysis results were both available in only 21 cases, limiting our ability to evaluate the diagnostic performance of the molecular diagnosis method. Findings presented here thus need to be interpreted with caution. A follow-up prospective validation study is ongoing in Nepal and Myanmar to address this issue.
Snakebite envenoming is an important health problem in Nepal, accounting for up to 39.7% of poisoning cases admitted to emergency units in certain regions [6]. Neurotoxicity following the bites of elapid snakes is of particular concern. This study for the first time addresses the distribution and medical importance of snake species contributing to the burden of snakebite in Nepal. It provides crucial information for clinicians and health workers involved in the management of snakebite victims in Nepal. It notably highlights that the majority of bites are caused by non-venomous snakes, and that the diversity of venomous snake species involved in bites is greater than previously believed. Finally, this study provides initial evidence on the utility of forensic DNA-based methods in the identification of biting snake species.
|
10.1371/journal.ppat.1002124 | Divergent Effects of Human Cytomegalovirus and Herpes Simplex Virus-1 on Cellular Metabolism | Viruses rely on the metabolic network of the host cell to provide energy and macromolecular precursors to fuel viral replication. Here we used mass spectrometry to examine the impact of two related herpesviruses, human cytomegalovirus (HCMV) and herpes simplex virus type-1 (HSV-1), on the metabolism of fibroblast and epithelial host cells. Each virus triggered strong metabolic changes that were conserved across different host cell types. The metabolic effects of the two viruses were, however, largely distinct. HCMV but not HSV-1 increased glycolytic flux. HCMV profoundly increased TCA compound levels and flow of two carbon units required for TCA cycle turning and fatty acid synthesis. HSV-1 increased anapleurotic influx to the TCA cycle through pyruvate carboxylase, feeding pyrimidine biosynthesis. Thus, these two related herpesviruses drive diverse host cells to execute distinct, virus-specific metabolic programs. Current drugs target nucleotide metabolism for treatment of both viruses. Although our results confirm that this is a robust target for HSV-1, therapeutic interventions at other points in metabolism might prove more effective for treatment of HCMV.
| Enveloped viruses draw on cellular machinery and materials to generate copies of their genome, structural proteins, and membrane. These biosynthetic processes use the host metabolic network to provide energy and small-molecule precursors. We have investigated how two important enveloped viruses, human cytomegalovirus and herpes simplex virus-1, alter host metabolism to provide materials for viral replication. We show that rather than passively relying on basal host cell metabolic activity, both viruses actively redirect host cell metabolism, implementing divergent metabolic programs that are robust to host cell type and virus strain. Human cytomegalovirus enhances lipid biosynthesis, while herpes simplex-1 gears central carbon metabolism toward the synthesis of pyrimidine nucleotides. Consistent with these changes, human cytomegalovirus is more sensitive to inhibition of fatty acid synthesis and herpes simplex virus-1 to inhibition of central metabolic reactions leading towards pyrimidine synthesis. As these two closely related viruses have divergent metabolic strategies, and since the metabolic perturbations point to potential drug targets, an important priority is defining the metabolic programs of other viruses.
| Herpesviruses are large, enveloped, double-stranded DNA viruses, capable of both lytic infection and life-long latency in mammalian hosts [1]. They are major causes of human disease. A majority of adults are infected with herpes simplex virus 1 (HSV-1) and/or human cytomegalovirus (HCMV). An alpha-herpesvirus, HSV-1 infects a wide range of organisms and cells types, causing symptoms ranging from cold sores to encephalitis. The prototypical beta-herpesvirus, HCMV, selectively infects non-transformed human cells. Although frequently asymptomatic, HCMV causes severe disease in neonates and immunocompromised adults. All herpesviruses encode metabolic enzymes in their genomes, primarily ones involved in nucleotide metabolism. The HSV-1 genome encodes a viral thymidine kinase, ribonucleotide reductase, dUTPase and uracil DNA glycosylase, while HCMV encodes a functional form of uracil DNA glycosylase [2]. Like all viruses, however, they rely primarily on the metabolic capabilities of their cellular hosts for replication. Specifically, the host provides the energy, amino acids and lipids, as well as most nucleotides, required by the virus.
Improved technologies for measuring both enzymes and metabolites is enabling for the first time in-depth analysis of virus-host cell metabolic interactions. Liquid chromatography coupled to mass spectrometry (LC-MS) facilitates direct measurement of a large number of cellular metabolites [3], [4]. Combined with isotope tracers, metabolic flows (fluxes) can also be determined. These new tools have revealed that, rather than passively relying on basal host cell metabolic activity, many viruses actively redirect host cell metabolism [5], [6], [7]. For example, hepatitis C virus up-regulates host cell glycolysis and modulates concentrations of specific lipids [8]. Similarly, hepatitis B virus replication perturbs cholesterol metabolism by inducing increased 7-dehydrocholesterol levels [9].
Among herpes viruses, the metabolic effects of HCMV have been the most extensively studied. Infection of a human fibroblast cell line with HCMV leads to two-fold increases in glycolytic activity and nucleotide synthesis, as well as yet greater increases in citric acid cycle flux and lipid biosynthesis [10]. Consistent with HCMV's reliance on the metabolic fluxes that it induces, inhibition of the committed step of fatty acid synthesis and elongation, acetyl-CoA carboxylase, blocks HCMV replication [10]. The virus also induces an increased dependence on glutamine that serves to drive the TCA cycle [10], [11]. These metabolic changes are only partially accounted for by increased levels of enzyme transcripts, indicating the involvement of multiple regulatory mechanisms [6].
A limitation of studies of virus-host metabolic interactions to date is that they have focused on single virus-host cell pairs. Moreover, they have often employed transformed host cells that differ markedly from the cells usually infected in vivo. This has precluded understanding whether the observed metabolic effects of viruses are relevant in their natural host cells, preserved across host cell types, and conserved within families of related viruses. To address these issues, here we compare and contrast the metabolic effects of HCMV and HSV-1, across both fibroblast and epithelial host cells. Specifically, we studied the laboratory-adapted AD-169 strain of HCMV, which is restricted to growth in fibroblasts, and whose metabolic effects have been previously studied [6], [10]. In addition, we examined the epitheliotropic clinical isolate strain TB40/E, which grows in many cell types, to study the infection of epithelial cells [12]. For HSV-1 infections, we chose the highly-passaged, non-neuroinvasive KOS 1.1 strain and a prototypical neuroinvasive strain, the F strain [13], [14]. Both primary human foreskin fibroblasts and the MRC5 fibroblast cell line were analyzed after infection by both viruses. HSV-1 infection was also studied in the Vero African green monkey renal epithelial cell line which is traditionally used for its growth. Given HCMV's propensity to cause retinitis [15], it was studied in the ARPE-19 retinal pigment epithelial cell line.
Using LC-MS to probe core metabolite concentrations and fluxes, we find that HCMV and HSV-1 both trigger major metabolic changes in their cellular hosts, and that these changes are similar across different host cell types and for different strains of the same virus. In contrast, the effects of HCMV and HSV-1 diverge markedly. HCMV most greatly impacts pathways generating substrates for lipid metabolism, whereas HSV-1 most greatly impacts deoxypyrimidine metabolism.
We examined the metabolic changes triggered by infection of fibroblast and epithelial host cells with HCMV and HSV-1. Fibroblasts (HFF and MRC5) were held at confluence for 3–5 days then serum-starved for 24 hours prior to infection, while epithelial cells (ARPE19 and Vero) were infected at 80–90% confluence and maintained in the presence of dialyzed serum at all times. As a consequence the fibroblast host cells were growth arrested at the time of infection, while the epithelial cells continued to replicate after mock inoculation [16]. Consistent with their different growth states, there were substantial differences in the metabolome of the host cells prior to infection, with compounds directly involved in proliferative processes such as carbamoyl-aspartate (pyrimidine biosynthesis), dTTP (DNA synthesis), and S-methyl-5′-thioadenosine (polyamine synthesis) markedly higher in the growing epithelial cells than the quiescent fibroblasts (Figure S1). Other compounds, such as those involved in mitochondrial fatty acid oxidation (carnitine and acetyl-carnitine), were higher in growth arrested fibroblasts compared to growing epithelial cells.
Infections were performed at a multiplicity of 3 pfu/cell to ensure near complete exposure of the cell population. Cultures infected with one of the virus strains, or treated with a virus-free mock inoculum, were grown in parallel and sampled at regular time intervals from the beginning of infection until peak virus yields were achieved. Medium was changed every 24 h to ensure a consistent nutrient supply to the cells; lack of media changes in earlier work [6] resulted in some different metabolite patterns from those observed here. In particular, we find that citrate and malate levels increase >10-fold during HCMV infection, compared to the 2-fold change seen in previously published work [6]. Maximum virus output was reached at around 24 h post infection (hpi) in HSV-1 infected cells, and around 96 hpi HCMV infected cells (Figure S2).
Over 80 metabolites were identified and detected in all experiments. Relative concentrations of these species, between infected and mock-infected cells, are shown in Figure 1 (for blue/yellow version of the heat map, see Figure S3; for source data, see Table S1). A third of the compounds were measured by both high resolution mass spectrometry (orbitrap) and triple quadrupole mass spectrometry (QQQ). The profile of any single metabolite detected in multiple LC-MS methods was found to be similar, as indicated by co-clustering of the associated data in almost all cases (Figure 1).
Both viruses triggered >4-fold changes in the levels of roughly half of the metabolites assayed. Among the metabolites changing markedly, those increasing outnumbered those decreasing roughly two-to-one. Although the magnitude of the changes in compound levels depended on the host cell, typically being smaller in the growing epithelial cells, the majority of the trends were host cell invariant. This is remarkable given the differing initial growth and metabolic states of the host cells, and it indicates a robust ability of the viruses to re-program metabolism.
Extracting major trends from the dataset by singular value decomposition [17] resulted in two characteristic vectors that accounted for >10% of the information in the dataset (Figure S4A). These vectors represent prototypical metabolite response patterns. The first vector accounts for 16% of the variation in metabolite levels over the time courses. In this vector, the signal as a function of time shows a similar trend in each of the time courses, thus representing a generic metabolite concentration response to herpesvirus infection (Figure S4B). The smaller signal in the first and last columns corresponding to the infections of epithelial cells reflects the smaller fold-changes in metabolite levels induced by viral infection in the growing epithelial cells compared to growth arrested fibroblasts. The strongest contributor to the generic infection response is dTTP, whose upregulation is consistent with the shared need of both viruses to replicate their DNA. The second vector, accounting for 12% of the variation in the dataset, represents a virus-specific response with opposing patterns for the HCMV and HSV-1 infection time courses (Figure S4B). Key contributors to this virus-specific response include TCA cycle intermediates, consistent with their rise during HCMV but not HSV-1 infection, and the nucleotides dUMP and dTMP, consistent with their rise during HSV-1 but not HCMV infection.
The third most significant vector, which accounts for 6% of the information in the dataset, represents a metabolic response characteristic of Vero cells (Figure S4B). While most of the changes proved to be host cell-independent, the third vector draws attention to the impact of different host cell types on the metabolic effects of viruses. The strongest contributors to this vector are citrate/isocitrate and N-carbamoyl-L-aspartate, due to their depletion in infected Vero cells in contrast to their accumulation in all other cell types. The remaining characteristic vectors account for the other 66% of the information. This large amount of residual information reflects a myriad of metabolite, virus, and cell-type specific dynamics. For example, proline and glycine-betaine showed cell type-specific upward or downward trends. Other metabolites, such as dTTP, showed different dynamic response patterns across different host cell types.
In all cell types tested, HCMV infection induced phosphoenolpyruvate, deoxypyrimidine triphosphates, CDP-choline, and acetylated amino acids, as well as a striking and coordinated increase in citrate, malate and other TCA cycle intermediates (Figure 1). Depleted compounds included glycerophosphoinositol, taurine, and a number of pentose phosphate pathway metabolites. On the other hand, HSV-1 triggered increased levels of pentose phosphate pathway intermediates, as well as glycolytic intermediates, and deoxypyrimidines (Figure 1). Notably depleted compounds included glycine betaine, taurine, creatine, and NAD+. The conserved decrease in the osmolyte, taurine, in both HCMV and HSV-1 likely reflects a host cell response to virus-induced increases in cell volume [18]. Glycolysis, the citric acid cycle, and pyrimidine biosynthesis are discussed in greater detail below.
Glycolysis and the TCA cycle form the backbone of central carbon metabolism in mammalian cells. Through these two pathways glucose is either oxidized to produce energy in the form of NADH and ATP, or converted to precursors of amino acids, lipids and nucleotides. The levels of glycolytic intermediates are altered in a strikingly different manner during HCMV and HSV-1 infections (Figure 2A). The concentrations of metabolites in lower glycolysis increase during HCMV infection, while levels of upper glycolytic intermediates drop. Conversely, in response to HSV-1 infection the opposite occurs. While these concentration measurements are informative, it is not possible to deduce whether changes in influx, efflux or a combination of both are responsible for the perturbations of the metabolite levels. Neither the turnover rate of a metabolic intermediate, nor the material flow through a pathway, can be predicted based on metabolite pool sizes alone. To understand how material flow, i.e., flux, is altered, further assays must be employed.
In cultured mammalian cells, the enzyme-catalyzed reactions of glycolysis convert the bulk of glucose imported from the extracellular environment to lactate, which gets excreted. Thus, changes in the rate of material flow through glycolysis can be approximated by measuring the rate of glucose consumption and lactate production. We determined the glucose uptake and lactate excretion rates in infected and mock treated HFFs by directly measuring the amount of glucose and lactate in the extracellular medium over time (Figure 2B). HCMV increased the uptake of glucose (p = 0.02) and the excretion of lactate (p = 0.0006), in agreement with previously published results on HCMV-infected fibroblasts [10], [19], [20] (Figure 2B). On the contrary, in HSV-1 infected cells the glucose uptake (p = 0.21) and lactate excretion (p = 0.002) rates decreased to a modest extent.
In addition to glucose from the medium, glycolysis can also be fueled by glucose acquired from the breakdown of stored glycogen. Moreover, decreased lactate production can reflect increased glycolytic efflux to the TCA cycle, rather than decreased glycolytic flux. To confirm our conclusions based on the glucose and lactate measurements, we also measured the rate of incorporation of isotope-labeled nutrients into downstream metabolites. Following a switch to labeled media, metabolite pools become progressively more labeled, with the unlabeled fraction exhibiting an exponential-type decay. Flux through a metabolite is the product of the rate of this decay and the total pool size of the metabolite [21]. To reliably estimate this decay rate, it is important to obtain samples at early time points where the fractional labeling is changing rapidly. Because label from glucose gets incorporated very quickly into glycolytic intermediates, measurements at later time points are likely to reflect steady-state labeling fractions, not labeling rates per se. At steady state, the amount of labeled metabolite reflects the total metabolite pool size and the fraction of its production from the labeled substrate, but not the rate of labeling.
To characterize glycolytic flux, we switched HCMV and HSV-1 infected cells, as well as their mock-treated counterparts, to 13C-labeled glucose containing media, and used LC-MS to monitor the labeled forms of downstream metabolites over time. HCMV infection increased the fractional labeling rate of glycolytic intermediates fructose-1,6-bisphosphate and dihydroxyacetone phosphate, while HSV-1 decreased it (Figure 2C). The decrease in the rate of HSV-1 labeling was complemented by a corresponding increase in metabolite concentration. Thus, we can conclude that HCMV significantly increases flux through glycolysis and HSV-1 does not.
Interestingly, in HSV-1-infected cells the metabolites upstream of phosphoenolpyruvate build up, while the ones downstream drop (Figure 2A). This suggests a bottleneck in glycolytic efflux at the step catalyzed by pyruvate kinase, the enzyme that converts phosphoenolpyruvate and ADP to pyruvate and ATP. The buildup of glycolytic metabolites upstream of pyruvate is accompanied by increased levels of pentose phosphate pathway intermediates, thus increasing the availability of ribose-phosphate for the synthesis of nucleotides. During hepatitis C infection the levels of most glycolytic enzymes were shown to be elevated, with the notable exception of pyruvate kinase [8]. Such changes in enzyme levels may lead to a similar metabolic outcome as observed in HSV-1 infected cells. However, as the activity of glycolytic flow is under tight allosteric control [22], direct metabolic analysis of hepatitis C is warranted to confirm this.
The metabolites of the TCA cycle showed a particularly interesting difference in labeling patterns when HCMV- and HSV-1-infected fibroblasts were supplied with uniformly labeled 13C- glucose. In the uninfected, growth arrested fibroblasts, citrate was only minimally labeled over a 2 h time period (Figure 3D, top panel). On the other hand, HCMV-infected fibroblasts produced a significant amount of citrate with two labeled carbon atoms (13C2-citrate) (Figure 3D, center panel), while their HSV-1-infected counterparts generated citrate with three labeled carbons (13C3-citrate) (Figure 3D, bottom panel). These two forms of citrate are produced by different pathways, which are selectively up-regulated in a virus-specific manner.
Labeled carbon atoms derived from 13C-glucose can enter the TCA cycle via two routes (Figure 3A). In one, pyruvate dehydrogenase and citrate synthase incorporate two carbons from glucose into citrate via acetyl-CoA (Figure 3B). The labeling pattern of citrate during HCMV infection indicates increased influx of glycolytic carbon to the TCA cycle via this route (Figure 3D). This pathway indicates a catalytic use of the TCA cycle, with the two-carbon units originating from glycolysis either oxidized to produce energy by complete turning of the TCA cycle, or diverted from the mitochondria to the cytosol through the citrate shuttle, where the acetyl group is freed for fatty acid synthesis and/or elongation. Global flux analysis on HCMV infection showed that both of these uses of glycolytic carbon are up-regulated by HCMV in MRC5 cells [10]. Our results indicate that HCMV infection of HFFs leads to the same up-regulation.
Carbon from glycolysis can also enter the TCA cycle via pyruvate carboxylase, which converts pyruvate to oxaloacetate (Figure 3C). All three labeled carbons in pyruvate are retained in oxaloacetate, which is converted to 13C3-citrate, malate, or aspartate. The labeled forms of TCA cycle intermediates observed in HSV-1-infected cells indicate an up-regulation of carbon influx via pyruvate carboxylase as reflected by the labeling of citrate (Figure 3D, right panel) and malate (Figure S5) when cells are supplied with 13C6-glucose. Furthermore, no citrate is detected with two or five labeled carbons in these cells. Thus, unlike in HCMV-infected cells, the use of glucose to drive the citrate shuttle and ensuing fatty acid synthesis is minimal during HSV-1 infection. The previous metabolic analysis of HCMV infection led to the recognition of a potential new drug target by showing that de novo fatty acid biosynthesis is essential for HCMV replication [10]. Pharmacological inhibitors of enzymes in fatty acid biosynthesis were shown to inhibit not only HCMV replication, but also the replication of influenza, an evolutionarily divergent virus [10]. De novo fatty acid biosynthesis does not appear to bear the same importance for the replication of HSV-1 as for HCMV (Figure 3D). This is reflected in the lower sensitivity of HSV-1 replication to 5-tetradecyloxy-2-furoic acid (TOFA) (Figure S6) [10], an inhibitor of acetyl-CoA carboxylase, the first committed enzyme of fatty acid biosynthesis.
The reaction catalyzed by pyruvate carboxylase is an anaplerotic reaction that serves to replenish the intermediates of the TCA cycle as they are removed for biosynthetic purposes. However, in spite of its up-regulation during HSV-1 infection, after an initial elevation, the levels of TCA cycle intermediates drop (Figure 4). This indicates that HSV-1 triggers an even greater increase in TCA cycle efflux. Notably, the concentration of aspartate, which is produced from oxaloacetate, decreases significantly after infection with HSV-1. In addition to being used for protein synthesis, aspartate is a substrate for pyrimidine nucleotide biosynthesis.
Unlike in HCMV infection, in response to HSV-1 infection the rates of total RNA and total protein syntheses drop [23], [24]. At the same time, viral DNA synthesis increases the demand for deoxyribonucleotides. The nucleotide precursors essential for DNA synthesis can be acquired through salvage reactions or de novo synthesis [25], [26]. When replicating in quiescent cells as opposed to actively dividing ones, viruses face a greater challenge in acquiring nucleotides for viral DNA replication, because the de novo nucleotide biosynthesis pathways are less active [27]. HSV-1 encodes a set of enzymes addressing this problem and their impact is reflected in increased concentrations of the intermediates of the pyrimidine nucleotide biosynthesis pathway (Figure 5). HCMV employs an alternative mechanism whereby the host cell is driven from quiescence to the G1/S boundary of the cell cycle [28], stimulating host cell nucleotide biosynthesis but preventing host DNA replication. Interestingly, in HSV-1 infected serum-starved fibroblasts dTTP levels are not observed to peak and drop after 6 hpi as reported in Vero cells (Figure 1) [26], and BHK cells [29]. In growth arrested fibroblasts the dTTP pool continues to rise throughout the infection (Figure 5). Such a trend was previously observed in mutant BHK cells that lack thymidine kinase and deoxycytidine kinase activities [29]. Confluent, serum-starved fibroblasts may present a similar cellular environment, with very low basal activity of DNA-biosynthetic enzymes.
Uracil can occur in DNA as a result of cytosine deamination or misincorporation of dUTP [30]. The UL50 and UL2 genes of HSV-1 encode enzymes that address these problems. The viral dUTPase (UL50) serves to reduce incorporation of uracil into viral DNA by decreasing dUTP levels and producing dUMP. Uracil-DNA glycosylase (UL2) participates in base excision repair of the HSV-1 genome, removing uracil from viral DNA [31], [32]. These two viral enzymes are likely responsible for the increased dUMP and uracil levels during HSV-1 infection (Figure 5).
While there is no known HSV-1 gene that causes the increased production of carbamoyl-aspartate, evidence for the regulation of aspartate transcarbamoylase during adenovirus infections has been presented in the past [33], [34]. Furthermore, carbamoyl-aspartate levels are observed to rise dramatically in both HCMV and HSV-1 infections (Figure 1) [10]. Carbamoyl-aspartate is produced by the multifunctional CAD protein, which catalyzes the first three steps of de novo pyrimidine biosynthesis in mammalian cells. CAD is highly regulated by growth state-related signaling molecules, such as the epidermal growth factor [35], [36]. Epidermal growth factor receptor has been shown to play a role in the entry of several different viruses, and it or related signaling pathways might contribute to virally-induced increases in carbamoyl-aspartate levels [37], [38], [39].
To confirm that flux from aspartate to pyrimidine nucleotides is up-regulated in HSV-1 infection, we analyzed the labeling pattern of the pathway intermediates after switching cells to medium containing uniformly labeled 13C-glutamine (Figure 6A). As glutamine contributes to anapleurosis in both mock and infected cells, this resulted in labeling of aspartate in both cases, and thus enabled direct comparison of pyrimidine synthesis between these two conditions. Significantly faster labeling of the pyrimidine end-product UTP was observed in infected cells (Figure 6B). As the concentration of UTP is also elevated in HSV-1 infected cells, flux from aspartate to nucleotide synthesis is markedly increased.
Taken together, the above observations indicate an upregulation of flux in HSV-1 infected cells from glucose to de novo pyrimidine nucleotide biosynthesis via the pyruvate carboxylase-catalyzed anaplerotic and the aspartate transaminase 2 catalyzed cataplerotic reactions of the TCA cycle (Figure 7A). In agreement with this, small interfering RNA (siRNA) mediated knockdown of pyruvate carboxylase and aspartate transaminase 2 inhibit HSV-1 replication, but not HCMV (Figure 7B–D).
Viral replication depends on the energy and biosynthetic precursors supplied by host cell metabolism. Using a mass spectrometry-based metabolomic approach we demonstrate that two closely related viruses, HCMV and HSV-1, implement divergent metabolic programs (Figure 1 and Table S1). These programs are robust to host cell type and virus strain. While HCMV enhances glycolytic flux and the delivery of carbon from glucose to the TCA cycle to fuel fatty acid biosynthesis, HSV-1 gears central carbon metabolism toward the production of pyrimidine nucleotide components (Figure 8). The focus of HSV-1, but not HCMV, on nucleotide metabolism is interesting in light of nucleoside analogues (acyclovir and ganciclovir, respectively) being more effective treatments for HSV-1 than for HCMV [40]. Both compounds depend on phosphorylation by viral kinases for their activation, and the metabolic profile of HSV-1 infected cells reflects the activity of the virally encoded thymidine kinase. On the other hand, the only functional HCMV kinase (UL97) is a protein kinase and has little to no nucleotide kinase activity [41]. This difference is reflected in the metabolome and in the lower efficacy of the nucleoside analogues for HCMV. In contrast, we show that TOFA, an inhibitor of the committed step of fatty acid synthesis and elongation, preferentially targets HCMV over HSV-1 (Figure S6).
The viruses also induce robust changes outside of core metabolism. For example, HCMV, but not HSV-1, induces a striking increase in acetylated amino acids (Figure 1). After HSV-1 infection, NAD+ levels dropped by a factor of about 10, but little decline was evident after HCMV infection. We have recently discovered that this NAD+ depletion is due to elevated poly-ADP-ribose polymerase activity (L. Vastag, unpublished work). The activation of poly-ADP-ribose polymerase has also been observed in HIV-1 and Sindbis Virus infected cells [42], [43], [44]. Understanding the significance of such observations requires further study.
Why do these two related viruses induce markedly different changes in host cell metabolism? Both must synthesize viral proteins and nucleic acids and both produce enveloped virions. Perhaps the difference results in part from the markedly different speeds at which the two viruses progress through their replication cycles. HSV-1 produced maximal yields in fibroblasts or epithelial cells within about 24 h, whereas HCMV did not achieve maximal yields until about 96 hpi (Figure S2). One might speculate, then, that HSV-1, which accumulates its DNA fairly rapidly, must elevate nucleotide biosynthesis; in contrast, HCMV, which accumulates its DNA over a much longer time frame, does not require such a strong induction (Figure 1). It is more difficult to suggest why HCMV depends on de novo fatty acid biosynthesis more strongly than HSV-1 (Figure S6). It is conceivable that HCMV induces the production of new membranes to serve as a source for the virion envelope, while HSV-1 virions are built from pre-existing membranes. Consistent with this view, HCMV-infected cells develop a well-defined, membranous assembly compartment during the late phase of infection [45], [46], [47], but no equivalent structure has been described within HSV-1-infected cells.
The metabolic program induced by herpes viruses could be implemented in several ways. One potential strategy involves perturbation of general host biochemical milieu. For example, the HCMV UL37x1 protein elevates free intracellular calcium levels [48], which could potentially activate glycolysis through the action of calcium-sensitive kinases [19]. Alternatively, virus-coded gene products could modify or interact with pivotal regulators of host cell metabolism, e.g., the HCMV UL38 protein [49], or with metabolic enzymes themselves to alter their activity. Yet other strategies could involve modulation of host cell enzyme concentrations through mechanisms involving transcription, translation, or protein stability. A comprehensive systems level analysis, incorporating transcriptomic [50], [51], [52], [53], proteomic [8], and metabolic data should help clarify the relative significance of these latter mechanisms.
In addition to elucidating the mechanisms underlying host cell metabolic hijacking, an important priority is defining the metabolic programs of other viruses. Among herpes viruses, it will be interesting to see whether most fit either the HSV-1 or HCMV prototype, or whether alternative programs exist. For smaller viruses, it will be interesting to see whether their yet more precious genome space includes instructions for extensive host cell metabolic reprogramming. Such work holds substantial practical value, given overarching importance of enzyme inhibitors as antivirals and the utility of metabolomics for identifying new antiviral targets.
Primary human foreskin fibroblasts (HFFs) were collected previously [54] and stored in liquid nitrogen. We used them at passages 8–13. ARPE19 human retinal pigment epithelial cells, Vero green monkey kidney epithelial cells and MRC5 human embryonic lung fibroblasts were purchased from the American Type Culture Collection. Cells were grown in Dulbecco's modified Eagle Medium (DMEM) with 10% fetal bovine serum, 100 µg/mL penicillin and streptomycin (Invitrogen), and 4.5 g/L glucose. HSV-1 strain F [55] was kindly provided by B. Roizman (University of Chicago), the HSV-1 KOS 1.1 strain [56] was a gift from D. Hargett (Princeton University), and both viruses were grown in Vero cells [57]. BADwt-GFP is a phenotypically wild-type HCMV laboratory strain that was generated from a bacterial artificial chromosome (BAC) clone of strain AD169 [58] engineered to express green fluorescent protein [59]. TB40/E-eGFP is a phenotypically wild-type HCMV clinical isolate that was derived from a bacterial artificial chromosome termed TB40-BAC4 [60] containing a green fluorescent protein marker gene under control of the SV40 promoter between US34 and TRS1. HCMV strains were grown in MRC-5 cells. To prepare virus stocks for both HSV-1 and HCMV, the media of infected cells was layered over a sorbitol cushion (20% sorbitol, 50 mM Tris-HCl, pH 7.2, 1 mM MgCl2) and virus was pelleted by centrifugation (20,000 rpm, 1 h, 4°C, Beckman SW28 rotor). Virus stocks were prepared in DMEM with 0.5% bovine serum albumin and without fetal bovine serum, to avoid serum stimulation of the growth arrested fibroblasts during inoculation.
For analysis of metabolites, fibroblasts (HFFs or MRC5) were grown to confluence and maintained in the presence of serum for 5 d. Cells were then washed with serum-free DMEM and maintained in serum-free DMEM for 24 h before infection or mock treatment. Epithelial cells (ARPE19 or Vero) were grown to 80–90% confluence in DMEM with 10% dialyzed serum (Gemini Bio-Products) before infection. At the time of infection cells were inoculated with virus resuspended in DMEM with or without serum. Mock treated cells were inoculated with equivalent, virus-free DMEM. After a 1 h inoculation fresh DMEM was added to the cells, following two washes with the appropriate medium. For each time point in every experiment an additional mock treated and infected plate was processed for packed cell volume measurement. Approximately 5×105 cells were added to packed cell volume tubes (Techno Plastic Products), which were centrifuged at 2000×g for 5 min before reading [61]. Packed cell volume measurements were used to normalize the metabolite levels between samples.
At various times post infection or addition of 13C-labeled glucose- or glutamine-containing DMEM, the media of infected and mock cells was aspirated and −80°C, 80∶20 methanol∶water (v/v) was immediately added to quench metabolism. There were no washing steps prior to metabolism quenching, as such steps risk metabolic alterations. Metabolites were then extracted as described previously [21]. The extract was dried under nitrogen and metabolites were resuspended in HPLC-grade water and centrifuged at 15000×g speed for 5 min to remove particulate matter before analysis. To minimize complications due to excessive sample concentration and associated ion suppression during LC-MS analysis, samples were diluted substantially prior to analysis: metabolites collected from 106 cells (a confluent 60 mm plate of fibroblasts) were resuspended in 500 µL water.
To quantitatively measure the levels of metabolites in extracts prepared from infected or mock treated cultured mammalian cells, two different mass spectrometry methods were employed. Liquid chromatography-tandem mass spectrometry (LC-MS/MS) in selective reaction monitoring (SRM) mode was used to assay for ∼200 metabolites of confirmed identity from a wide range of metabolic pathways [62]. A Finnigan TWQ Quantum Ultra mass spectrometer was used in the positive ionization mode, and a TSQ Quantum Discovery MAX mass spectrometer in the negative mode, each equipped with an electrospray ionization source (Thermo Fisher Scientific). The SRMs were constructed with parameters acquired through optimizing the collision induced fragmentation of purified standards of the given metabolites. The LC method in positive mode employed an aminopropyl column for separation [62], while in negative mode the metabolite extracts were passed through a C18 column using tributylamine as an ion pairing agent to achieve longer retention of polar compounds [63], [64]. In addition, the LC-MS/MS method was complemented with untargeted analysis using liquid chromatography coupled to a stand-alone orbitrap mass spectrometer (Thermo Fisher Scientific Exactive instrument) which performs full scans from 85 to 1000 m/z at 100,000 mass resolution [65]. In this system, identification of compounds is based on two parameters: the retention time on the LC column and the compound mass measured with less than 2 ppm mass accuracy. Peaks were identified and peak heights exported with the Metabolomic Analysis and Visualization Engine (MAVEN) [66].
For glucose uptake and lactate excretion measurements, media samples were collected every 3 h between 45 and 57 hpi for HCMV, and every 2 h between 6 and 18 hpi for HSV-1. The concentrations of lactate and glucose were measured using a YSI 7100 Select Biochemistry Analyzer (YSI Incorporated). Uptake and excretion rates were determined as the rate of concentration change of these compounds in the media. The values were corrected using the packed cell volume of the infected and mock cells.
For experiments involving monitoring the rate of incorporation of 13C-labeled nutrient into downstream metabolites, cells were switched to fresh media 1 h before addition of the labeled nutrient. This minimized the perturbation to the cells when their medium was replaced with isotope containing medium. Cells were then maintained in medium containing the labeled nutrient for different lengths of time. Metabolites were extracted and various isotopically labeled forms quantified by mass spectrometry. The values were corrected for the natural abundance of 13C as described previously [10]. Labeled DMEM was prepared from glucose and glutamine-free DMEM with the addition of U-13C-glucose or U-13C-glutamine (Cambridge Isotope Laboratories). All media were equilibrated to the incubator temperature and gas composition before use.
Double stranded siRNA molecules directed against pyruvate carboxylase (5′-GACUGUACGCGGCCUUCGATT), aspartate transaminase 2 (5′-CUAUUGAGAGCUUCACACATT), and a Universal Negative Control (SIC001) were purchased from Sigma. Subconfluent MRC5 cells seeded into 96-well plates were transfected with 10 pmol of siRNA using Oligofectamine transfection reagent (Invitrogen) according to the manufacturer's instructions. For HCMV experiments, the siRNA transfected cells were incubated for 24 hours and then infected with HCMV strain BADwt-GFP at a multiplicity of 0.1 pfu/cell. The cells were further incubated for 96 hours and media containing the infectious virus were harvested. Since HSV-1 replicates with a faster kinetics than HCMV, the transfected cells were incubated for 3 days to allow efficient knockdown of target gene. The cells were then infected with HSV-1 strain F at a multiplicity of 0.02 pfu/cell and media were harvested 24 hours after infection. The yield of HCMV and HSV in the media was determined by infectious focus assay. Briefly, fresh MRC5 cells were infected with different dilutions of viruses and fixed 24 hours after HCMV or 4 hours after HSV-1 infection with methanol at −20°C. Foci were identified using mouse monoclonal primary antibodies to HCMV immediate early IE1 protein (1B12) [54] or HSV-1 immediate early ICP4 protein [67] and a goat anti-mouse Alexa Fluor 488-conjugated secondary antibody (Invitrogen).
MRC5 cells were seeded into 6-well dishes and transfected at 70% confluence as described above. HFF cells were grown to confluence, serum starved for 24 hours and infected at 3 pfu/cell with HSV-1 (F strain). At selected times post transfection of MRC5 cells and infection of HFFs, cells were washed with phosphate-buffered saline (PBS), harvested and stored at −80°C. Cells were lysed in RIPA-light buffer (50 mM Tris-HCl, pH 8.0, 1% NP-40, 0.1% SDS, 150 mM NaCl, 0.1% Triton X-100, 5 mM EDTA) with protease inhibitors (Roche Applied Science), and protein concentrations were determined by Bradford assay. Proteins were separated by 10% SDS-containing polyacrylamide gel electrophoresis and transferred to nitrocellulose membranes. Membranes were probed with a primary rabbit polyclonal antibody directed against pyruvate carboxylase (NBP1-49536, Novus) at a dilution of 1∶1000 in PBS-T and 1% nonfat milk. After washing with PBS-T, membranes were probed with goat anti-rabbit HRP-coupled secondary antibodies diluted 1∶5000 in PBS-T containing 1% milk. Proteins were visualized by chemiluminescence using the ECL detection system (Amersham).
All p-values were calculated by two-tailed, non-paired T-test.
Pyruvate carboxylase (PC): P11498, aspartate transaminase 2 (GOT2): P00505, carbamoyl-phosphate synthetase 2, aspartate transcarbamylase, and dihydroorotase (CAD): P27708, HSV-1 dUTPase (UL50): P10234, HSV-1 uracil-DNA glycosylase (UL2): P10186, HSV-1 thymidine kinase (UL23): P03176.
|
10.1371/journal.ppat.1005185 | Modulation of the Host Lipid Landscape to Promote RNA Virus Replication: The Picornavirus Encephalomyocarditis Virus Converges on the Pathway Used by Hepatitis C Virus | Cardioviruses, including encephalomyocarditis virus (EMCV) and the human Saffold virus, are small non-enveloped viruses belonging to the Picornaviridae, a large family of positive-sense RNA [(+)RNA] viruses. All (+)RNA viruses remodel intracellular membranes into unique structures for viral genome replication. Accumulating evidence suggests that picornaviruses from different genera use different strategies to generate viral replication organelles (ROs). For instance, enteroviruses (e.g. poliovirus, coxsackievirus, rhinovirus) rely on the Golgi-localized phosphatidylinositol 4-kinase III beta (PI4KB), while cardioviruses replicate independently of the kinase. By which mechanisms cardioviruses develop their ROs is currently unknown. Here we show that cardioviruses manipulate another PI4K, namely the ER-localized phosphatidylinositol 4-kinase III alpha (PI4KA), to generate PI4P-enriched ROs. By siRNA-mediated knockdown and pharmacological inhibition, we demonstrate that PI4KA is an essential host factor for EMCV genome replication. We reveal that the EMCV nonstructural protein 3A interacts with and is responsible for PI4KA recruitment to viral ROs. The ensuing phosphatidylinositol 4-phosphate (PI4P) proved important for the recruitment of oxysterol-binding protein (OSBP), which delivers cholesterol to EMCV ROs in a PI4P-dependent manner. PI4P lipids and cholesterol are shown to be required for the global organization of the ROs and for viral genome replication. Consistently, inhibition of OSBP expression or function efficiently blocked EMCV RNA replication. In conclusion, we describe for the first time a cellular pathway involved in the biogenesis of cardiovirus ROs. Remarkably, the same pathway was reported to promote formation of the replication sites of hepatitis C virus, a member of the Flaviviridae family, but not other picornaviruses or flaviviruses. Thus, our results highlight the convergent recruitment by distantly related (+)RNA viruses of a host lipid-modifying pathway underlying formation of viral replication sites.
| All positive-sense RNA viruses [(+)RNA viruses] replicate their viral genomes in tight association with reorganized membranous structures. Viruses generate these unique structures, often termed “replication organelles” (ROs), by efficiently manipulating the host lipid metabolism. While the molecular mechanisms underlying RO formation by enteroviruses (e.g. poliovirus) of the family Picornaviridae have been extensively investigated, little is known about other members belonging to this large family. This study provides the first detailed insight into the RO biogenesis of encephalomyocarditis virus (EMCV), a picornavirus from the genus Cardiovirus. We reveal that EMCV hijacks the lipid kinase phosphatidylinositol-4 kinase IIIα (PI4KA) to generate viral ROs enriched in phosphatidylinositol 4-phosphate (PI4P). In EMCV-infected cells, PI4P lipids play an essential role in virus replication by recruiting another cellular protein, oxysterol-binding protein (OSBP), to the ROs. OSBP further impacts the lipid composition of the RO membranes, by mediating the exchange of PI4P with cholesterol. This membrane-modification mechanism of EMCV is remarkably similar to that of the distantly related flavivirus hepatitis C virus (HCV), while distinct from that of the closely related enteroviruses, which recruit OSBP via another PI4K, namely PI4K IIIβ (PI4KB). Thus, EMCV and HCV represent a striking case of functional convergence in (+)RNA virus evolution.
| Picornaviridae is a large family of positive-sense RNA viruses [(+)RNA viruses] comprising many clinically relevant human and animal pathogens. Members of the genus Enterovirus include important human viruses like poliovirus (PV), the causative agents of poliomyelitis, coxsackieviruses (CV), causing meningitis and myocarditis, and rhinoviruses (RV), responsible for the common cold and exacerbations of asthma and chronic obstructive pulmonary disease. Perhaps the best-known non-human picornavirus is foot-and-mouth-disease virus (FMDV, genus Aphtovirus), which can cause devastating outbreaks in cattle leading to severe economic loss. Closely related to the Apthovirus genus is the genus Cardiovirus, composed of three species: Theilovirus (TV), encephalomyocarditis virus (EMCV) and the more recently discovered Boone cardiovirus. The species Theilovirus includes, among others, Theiler’s murine encephalomyocarditis virus (TMEV) and Saffold virus (SAFV), a human cardiovirus. While TMEV is known to cause enteric infections and sometimes more severe encephalitis or chronic infection of the central nervous system [1], as yet, SAFV has not been firmly associated with a clinical disease [2]. EMCV can infect a wide range of animals, of which rodents are considered the natural reservoir. Of all domesticated animals, pigs are most prone to EMCV infection, which can lead to fatal myocarditis [3], reproductive failure in sows or sudden death of piglets [4–6].
Like other (+)RNA viruses—such as hepatitis C virus (HCV), dengue virus (DENV), chikungunya virus (ChikV) and coronavirus (CoV)—picornaviruses replicate their genomic RNA on specialized, virus-modified intracellular membranes. These remodeled membranes termed replication organelles (ROs) arise from the concerted actions of both viral nonstructural proteins and co-opted host factors. Enteroviruses, for instance, hijack members of the secretory pathway for replication and formation of ROs [7,8]. Among the viral nonstructural proteins, 2B, 2C, 3A as well as their precursors 2BC and 3AB contain hydrophobic domains which confer them membrane-modifying properties [9–11]. Considerable interest has been given to the study of the small viral protein 3A, which is the key viral player involved in membrane rearrangements. 3A interacts with and recruits secretory pathway components GBF1 (Golgi-specific brefeldin A-resistance guanine nucleotide exchange factor 1) and PI4KB (phosphatidylinositol-4 kinase type III isoform β) to ROs [12–16]. Despite intensive investigation, the role of GBF1 in enterovirus replication is not yet elucidated (reviewed in [8]). Recruitment of PI4KB to ROs leads to a significant local increase of membranes in its enzymatic product PI4P [15]. This PI4P-rich environment serves to further recruit other essential viral and host factors to replication sites, such as the viral polymerase 3Dpol, which is able to specifically bind PI4P in vitro. Recently, we and others revealed that PI4P plays a central role in enterovirus replication by recruiting the oxysterol-binding protein (OSBP) to ROs [17–19]. In uninfected cells, OSBP bridges the ER and Golgi membranes by binding to the ER integral membrane protein VAP-A and to PI4P and Arf1-GTP at the trans-Golgi [20]. Through its sterol-binding domain, OSBP shuttles cholesterol from ER to the Golgi and PI4P from the Golgi to the ER, thereby generating a lipid counterflow at ER-Golgi membrane contact sites (MCSs). In enterovirus infection, OSBP exchanges PI4P for cholesterol most likely at ER-RO MCSs [18]. The unique lipid and protein composition of enterovirus ROs determines their particular 3D architecture, which consists of a complex tubulo-vesicular network, as shown in cells infected with PV and coxsackievirus B3 (CVB3) [21,22].
The lipid transfer function of OSBP at membrane contact sites is not only vital for enteroviruses, but also for HCV [23]. HCV genome replication occurs in association with an ER-derived network of specialized membrane vesicles called the membranous web (MW). Like enterovirus ROs, the HCV MW is enriched in PI4P lipids and cholesterol [23–25]. In the case of HCV, PI4P are generated through recruitment and activation of the ER-localized enzyme PI4KA (phosphatidylinositol-4-phopshate kinase type III isoform α) by the viral protein NS5A [24,26].
Thus far, information regarding virus-host interactions that govern the formation of cardiovirus ROs remains scarce. In a report by Zhang et al, it was suggested that autophagy supports EMCV replication [27]. The study showed that EMCV infection triggered an accumulation of autophagosome-like vesicles in the cytoplasm and that EMCV 3A colocalized with the autophagy marker LC3. However, inhibition of autophagy exerted only minor effects on virus replication [27], which argues against a strong implication of the autophagy pathway in cardiovirus genome replication and/or formation of ROs. Evidence for a role of autophagy in virus replication also exists for enteroviruses and flaviviruses, but rather related to non-lytic virus release or modulation of host innate immune responses than viral genome replication [28–31].
Based on observations that cardioviruses do not require GBF1 or PI4KB for replication [32–34], it is generally believed that cardiovirus replication strategies are distinct from those of enteroviruses. Here, we set out to elucidate whether cardiovirus replication depends on another PI4K isoform. By siRNA-mediated knockdown, we identified PI4KA as a key player in the replication of EMCV. EMCV 3A interacts with and recruits PI4KA to ROs, which increases local PI4P synthesis, eventually leading to downstream recruitment of OSBP. We show that the cholesterol-PI4P shuttling activity of OSBP is important for the global distribution of the ROs and for virus genome replication. Our data reveal that, by exploiting the same cellular pathway, the cardiovirus replication strategy profoundly resembles that of the distantly related HCV and is dissimilar to those of other characterized picornaviruses and flaviviruses in this critical aspect. Thus, the similarity between EMCV and HCV is a striking case of functional convergence in virus-host interactions, indicating that diverse RNA viruses might have a limited choice of pathways in the remodeling of host membrane network for virus replication.
Unlike enteroviruses, cardioviruses do not require PI4KB for replication [34]. To investigate whether other PI4Ks might be involved in cardiovirus replication, we depleted each of the four distinct cellular PI4Ks by siRNA-mediated gene knockdown using a set of siRNA sequences (Ambion) which we previously tested for efficiency and toxicity [35], and monitored the subsequent effects on replication of EMCV. We observed inhibitory effects on EMCV replication when silencing PI4KA, but not upon silencing of the other PI4Ks (Fig 1A). To confirm the importance of PI4KA for EMCV replication, we performed another series of knockdown experiments using another set of siRNA sequences (Qiagen). Depletion of PI4KA, but not PI4KB, significantly reduced EMCV infection, measured by end-point titration of progeny virus production (Fig 1B).
We next wondered which step in the virus life cycle is dependent on PI4KA. To omit the step of virus attachment and cell entry, EMCV RNA was in vitro transcribed and subsequently transfected in cells depleted of PI4KA by siRNAs. Virus replication was strongly inhibited upon PI4KA silencing, as measured by end-point titration of progeny virions (Fig 1C). This indicated that PI4KA is involved in a post-entry step in the virus life cycle. To elucidate whether EMCV requires PI4KA for viral genome amplification, we infected cells with a Renilla luciferase-encoding EMCV (RLuc-EMCV) and quantified the luciferase activity as a measure of viral RNA replication. EMCV RNA replication was severely impaired in cells lacking PI4KA, but not PI4KB (Fig 1D). We excluded that inhibition of EMCV replication by PI4KA silencing was due to cytotoxic effects by a cell viability assay (Fig 1E) and verified the knockdown efficiency by western blot analysis (Fig 1F). Altogether, these results showed that PI4KA plays a key role in EMCV genome RNA replication.
Next, we investigated whether EMCV required the enzymatic activity of PI4KA using AL-9, a PI4K inhibitor that also blocks PI4KB, but at 5-fold higher concentration [36]. Cells were infected with EMCV or RLuc-EMCV at MOI 0.1 and treated with increasing concentrations of AL-9 for 8 h. Coxsackievirus B3 (CVB3), as well as all other enteroviruses, has been previously shown to hijack the Golgi-localized PI4KB for replication [15,34] and was included as a control. As measured by end-point titration and analysis of the luciferase activity (Fig 1G and 1H), EMCV replication was efficiently inhibited by AL-9 in a dose-dependent manner with complete inhibition detected at 10 μM, while CVB3 replication was hampered only at 50 μM (Fig 1G), which is in line with the 5-fold preference of AL-9 for PI4KA over PI4KB. Dipyridamole, a well-established inhibitor of EMCV RNA replication, was included here as positive control. Importantly, AL-9 inhibited EMCV replication also when infection was performed at high MOI (S1A Fig, MOI 10). To corroborate that PI4KA activity is required for the step of viral genome replication, we performed a time-of-addition experiment in which AL-9 was added to the cells at different time points after infection with RLuc-EMCV. Similar to dipyridamole, AL-9 strongly inhibited replication when added up to 3 h after infection (S1B Fig), indicating that not entry but rather a step during genome replication was blocked by AL-9.
Next, we tested whether other members of the cardiovirus genus also depended on PI4KA for replication. Similar to EMCV, replication of the human cardiovirus Saffold virus 3 (SAFV3) (species Theilovirus) was also sensitive to AL-9 treatment (Fig 1I). The cell viability assay demonstrated that AL-9 treatment only exerted slight cytotoxic effects at the highest concentration tested (Fig 1J). These results indicated that different cardiovirus species required the enzymatic activity of PI4KA for genome replication.
Soon after infection, the cytoplasm of EMCV-infected cells accumulates an impressive amount of vesicular membranous structures [37,38]. As yet, there is little information available regarding which viral proteins and host factors are associated with these new virus-induced organelles [27,39]. We set out to investigate whether PI4KA was present at EMCV ROs. Despite repeated efforts, we were unable to detect the endogenous kinase by immunofluorescence staining in any of the cell lines tested. As an alternative, we chose to analyze possible changes in the subcellular distribution of ectopically expressed PI4KA upon EMCV infection. In mock-infected cells (Fig 2A, upper panel), GFP-PI4KA was distributed diffusely throughout the entire cytoplasm, as previously reported by others [40,41]. In infected cells visualized by dsRNA staining, we instead observed a clear difference in the localization of the kinase, which was redistributed to discrete cytoplasmic punctae in a perinuclear region (Fig 2A, lower panel).
We next aimed to elucidate whether these PI4KA punctae coincided with the viral ROs. The small picornaviral protein 3A and its precursor 3AB are membrane-associated and play key roles in viral RNA replication and recruitment of essential host factors [15,42–45]. Hence, we considered 3AB as a suitable marker for EMCV ROs and compared the staining of PI4KA to that of 3AB in infected cells. We observed a striking overlap of GFP-PI4KA with 3AB-positive structures (Fig 2B, lower panels) and could confirm this phenotype when analyzing the localization of ectopically expressed PI4KA bearing an HA-tag (S2 Fig). By contrast, the signal for GFP-PI4KB, which was mainly localized at the Golgi in non-infected cells (Fig 2C, top panel), failed to overlap with 3AB (Fig 2C, lower panels). Interestingly, although in close proximity to dsRNA signals (Fig 2A, lower panel), PI4KA did not clearly overlap with dsRNA (Fig 2A, insets). Taken together, these data demonstrated that PI4KA is selectively recruited to EMCV ROs.
Interestingly, we noticed a loss of the typical Golgi localization of PI4KB in EMCV-infected cells (Fig 2C, lower panel), suggesting that Golgi integrity might be affected upon EMCV infection. Prompted by this and our finding that EMCV utilizes the ER-localized PI4KA for replication, we set out to elucidate whether other ER or Golgi components are present at EMCV ROs. In order to be able to use more antibody combinations in immunofluorescence, we constructed a recombinant EMCV bearing an HA-tag in the nonstructural protein 2C. The tag was introduced after the second amino acid, leaving the 2B-2C cleavage site intact (S3A Fig), and did not impair virus replication (S3B Fig). First, we checked whether 2C-HA and 3AB are present on the same membranes by immunofluorescence microscopy. Indeed, 2C and 3AB signals greatly overlapped (S3C Fig), supporting the idea that these proteins occupy the same membranes of the ROs. Using this tagged EMCV, we noticed that the Golgi structure was indeed altered in infected cells, from 4 h p.i. onwards, as revealed by the dispersed pattern of both cis- and trans-Golgi markers GM130 (Fig 3A) and TGN46, respectively (Fig 3B). However, neither TGN46 nor GM130 were present at 2C-HA-positive structures, suggesting that EMCV ROs are not Golgi-derived. ERGIC53, a marker of the ER-Golgi intermediate compartment also appeared scattered throughout the cytoplasm in infected cells, but without overlapping 2C-HA (Fig 3C). We next compared the localization of 3AB with Sec13 (COPII-coatomer complex component), an ER exit site (ERES) marker, and the ER marker calreticulin. While in non-infected cells, Sec13 displayed mainly a typical perinuclear localization, in EMCV-infected cells it appeared dispersed, but without significantly colocalizing with 3AB (Fig 3D, Mander’s colocalization coefficient M2 = 0.14 ± 0.01, fraction of Sec13 overlapping 3AB). We observed a greater degree of overlap between 3AB and calreticulin (Fig 3E, M2 = 0.4 ± 0.02, fraction of calreticulin overlapping 3AB). Images acquired with higher magnification revealed that most of 3AB was in close contact with ER tubules (Fig 3F). Taken together, these data suggested that EMCV possibly replicates on ER-derived membranes.
Based on the extensive overlap between PI4KA and 3AB and the drastic change in PI4KA pattern in infected cells, we hypothesized that PI4KA might be recruited to replication sites by interacting (directly or indirectly) with one or more of the viral nonstructural proteins. To investigate this, we used the stable cell line Huh7-Lunet/T7 that allows ectopic protein expression under the control of a T7 promoter and has been previously optimized and validated as a reliable and reproducible cellular system to study PI4KA-protein interactions by radioactive Co-IP assays [40,46]. Myc-tagged EMCV nonstructural proteins 2A, 2B, 2C, 3A, 3AB, 3C and 3D were individually co-expressed together with HA-PI4KA in Huh7-Lunet/T7 cells, radioactively labeled, and affinity purified from cell lysates using anti-myc specific antibodies. Autoradiography analysis showed that HA-PI4KA was specifically co-purified by 3A and 3AB, but not by the other viral proteins (Fig 4A). To confirm this interaction by co-immunoprecipitation (co-IP) coupled with western blot analysis, myc-tagged EMCV 3A was co-expressed with HA-PI4KA and subjected to affinity purification using either monoclonal or polyclonal anti-myc antibodies. As shown in Fig 4B, HA-PI4KA only interacted with EMCV 3A, but not with CVB3 3A, which interacts with PI4KB [15,16,45] and was included here as a negative control. These data implied that EMCV nonstructural protein 3A is responsible for PI4KA recruitment to ROs. Interestingly, a diffuse band just below 17 KDa appears to co-purify with EMCV 3C when HA-PI4KA is co-expressed (Fig 4A, indicated by *). We reasoned this could be indicative of a temporal regulation of the PI4KA activity during infection via 3C-dependent degradation. To explore this possibility, we performed western blot analysis of endogenous PI4KA during the time course of infection, but did not detect any bands indicative of degradation, neither in Huh7-Lunet/T7 or HeLa R19 cells (S4 Fig). To test if 3A alone can recruit PI4KA, we examined by immunofluorescence the subcellular localization of HA-PI4KA when co-expressed with 3A, 3AB or 2B, which we considered as a negative control. When expressed alone, HA-PI4KA localized throughout the cell in a diffuse pattern (Fig 4C, top panel), as previously described [40]. EMCV 3A- and 3AB-myc were both localized throughout the cytoplasm and at discrete punctate structures, of which a subset was also positive for PI4KA (Fig 4C). 2B-myc was also distributed in punctae throughout the cytoplasm, but failed to recruit PI4KA (Fig 4C). Collectively, these results indicated that EMCV 3A is the viral protein responsible for engaging PI4KA in replication.
While PI4KB produces PI4P at Golgi membranes, PI4KA is responsible for the synthesis of the PI4P pool at the plasma membrane, where it dynamically localizes [41,47–49]. Our finding that PI4KA activity was critical for EMCV RNA replication prompted us to investigate whether PI4P metabolism is altered during virus replication. Given that EMCV replicates on intracellular membranes, we monitored potential changes in the subcellular distribution of both plasma membrane (PM) and intracellular (IC) pools of PI4P in Huh7Lunet/T7 cells following EMCV infection. The two pools of PI4P can be selectively visualized using two different immunocytochemistry protocols previously established by Hammond et al [50]. While the plasma membrane pool of PI4P appeared unaffected in EMCV-infected cells (Fig 5A), the intracellular PI4P distribution changed from a perinuclear, Golgi-like pattern in mock-infected cells to dispersed throughout the cytoplasm in EMCV-infected cells (Fig 5A). We observed similar PI4P phenotypes in HeLa cells (S5 Fig), indicating that the observed effects were not cell line-specific. Notably, quantitative analysis of the fluorescent PI4P signals revealed a marked increase in the level of intracellular PI4P in infected cells (Fig 5B).
To rule out a possible involvement of PI4KB in establishing the elevated PI4P levels observed in EMCV infected cells, we treated cells with the PI4KB inhibitor BF738735 (Compound 1) [34]. For simultaneous detection of PI4P and viral ROs by immunofluorescence, we infected cells with EMCV-2C-HA. Short treatment with BF738735 severely depleted the Golgi PI4P pool in non-infected cells (Fig 5C), thus reflecting an effective inhibition of PI4KB activity. However, the PI4P phenotype remained unaltered in infected cells (Fig 5C), demonstrating that the EMCV-induced accumulation of intracellular PI4P was not mediated by PI4KB. Most PI4P localized in the vicinity of 2C-HA, with at least a small subset of PI4P overlapping with 2C-HA (Fig 5C). These data together with the finding that EMCV requires PI4KA activity suggested that PI4KA-derived PI4P lipids play a central role in EMCV genome replication.
Various cellular proteins carrying a PI4P-binding domain called the pleckstrin-homology domain (PH), such as the ceramide-transfer protein (CERT), four-phosphate-adaptor protein 1 (FAPP1), or the oxysterol-binding protein (OSBP) can sense and specifically bind PI4P lipids [47,51–53]. Recently, we and others showed that enteroviruses generate PI4P-enriched membranes to recruit OSBP, which in turn exchanges PI4P for cholesterol at ROs [17,18,54]. Moreover, we showed that EMCV is sensitive to itraconazole, which we identified to be an OSBP inhibitor [17], and that cholesterol shuttling is important for EMCV replication [54]. We therefore reasoned that in EMCV-infected cells one purpose of PI4P lipids might be to recruit OSBP to replication membranes to support viral RNA replication. To test if OSBP is required for EMCV replication, we efficiently reduced OSBP expression in HeLa cells by siRNA gene silencing (Fig 6A) and evaluated the subsequent effects on EMCV replication by end-point titration analysis. Replication of EMCV was significantly reduced in cells in which OSBP was depleted compared to control-treated cells (Fig 6A), indicating that OSBP is required for efficient replication. We further used OSW-1, an OSBP ligand that interferes with normal OSBP functioning [55], to pharmacologically inhibit OSBP and analyze whether its lipid transfer function is linked to EMCV infection. Using luciferase-encoding EMCV, we observed a complete inhibition of genome RNA replication after 7 h of treatment with OSW-1 at nanomolar concentrations, with no cytotoxicity present (Fig 6B). A similar inhibition by OSW-1 was observed when infection was performed at high MOI (S6 Fig, MOI 10). Furthermore, by performing OSW-1 time-of-addition experiments, we excluded the possibility that OSBP was involved in early steps in the virus life cycle (Fig 6C). Similar results were obtained when using 25-hydroxycholesterol (25-HC, Fig 6C), another established OSBP ligand [20,56].
Next, we wondered whether endogenous OSBP was present at EMCV ROs and if so, whether this localization was dependent on the PI4P pool generated by PI4KA. To this end, cells were infected with EMCV for 5.5 h and then treated with DMSO or AL-9 for 30 min to acutely deplete PI4P, prior to immunofluorescence analysis. While in non-infected cells OSBP localized throughout the cytoplasm and at the Golgi, OSBP was mainly found at ROs in infected cells, where it largely colocalized with 3AB (Fig 6D, Pearson’s correlation coefficient = 0.71). Since other Golgi proteins were not present at the ROs (Fig 3A and 3B), these results suggested that OSBP is specifically recruited by EMCV. Following inhibition of PI4KA by short treatment with AL-9, we observed a strong and significant reduction of OSBP and 3AB colocalization (Fig 6D, Pearson’s coefficient = 0.58). Importantly, the subcellular localization of OSBP in non-infected cells was not affected by AL-9 treatment (Fig 6D), demonstrating that the presence of OSBP at EMCV replication structures is conditioned by PI4KA-produced PI4P.
Given the colocalization of OSBP with 3AB, we sought to verify whether EMCV 3A was responsible for OSBP recruitment. To this end, myc-tagged EMCV 3A-, 3AB- or 2B-myc were ectopically expressed in Huh7-Lunet/T7 cells and recruitment of endogenous OSBP was analyzed by immunofluorescence analysis. In cells expressing 3A and 3AB, OSBP was redistributed in punctate structures throughout the cytoplasm, with some of these punctae colocalizing with 3A/3AB (Fig 6E). By contrast, OSBP remained localized at the Golgi and did not localize at 2C-positive punctae (Fig 6E). These data indicated that during EMCV infection, OSBP is recruited to ROs via 3A.
To test whether OSBP is involved in transferring cholesterol to ROs in a PI4P-dependent manner, cells were infected with EMCV for 4 h, treated with AL-9 or OSW-1 for 2 h to block PI4KA activity or OSBP function respectively, and subjected to immunofluorescence analysis. In non-infected cells, cholesterol mainly localizes at endosomes in the perinuclear area and at the plasma membrane, as visualized by filipin staining [54]. In infected cells treated with DMSO, we detected cholesterol primarily colocalizing with 3AB-positive structures (Pearson’s coefficient = 0.62), while in drug-treated cells this colocalization was markedly reduced (Fig 7, Pearson’s coefficient = 0.32 for AL-9 and 0.39 for OSW-1). This result confirmed that EMCV ROs acquire cholesterol through the actions of both PI4KA and OSBP.
(+)RNA viruses display immense genetic diversity, yet they all rely on remodeled membranes for viral genome replication. A diverse array of cellular organelles can be remodeled into viral replication structures. For instance, picornaviruses from the genera Enterovirus and Kobuvirus are thought to replicate at modified Golgi membranes [15,45,57], while the flavivirus HCV replicates on a membranous web originated from the ER [58]. To do so, viruses rewire host pathways involved in lipid synthesis and transport to generate replication membranes with unique lipid signatures [59–62]. How picornaviruses from the genus Cardiovirus build their ROs is currently unknown. Here, we identified PI4KA and OSBP as essential host factors for genome replication of the cardiovirus EMCV. Our data suggest that EMCV ROs may be derived from the ER and that PI4KA was recruited to ROs by interacting with the viral protein 3A(B). PI4KA recruitment led to a significant increase of intracellular PI4P levels in infected cells, which proved important for the downstream recruitment of OSBP. Finally, data are presented suggesting that the OSBP-mediated exchange of PI4P and cholesterol at RO-MCSs is critical for EMCV genome replication and the global organization of ROs.
Membrane alterations in the cytoplasm of cardiovirus-infected cells were already observed decades ago by electron microscopy [37,38,63]. As also described for enteroviruses, cardiovirus-induced membranes consist of perinuclear clusters of heterogeneous single- and double-membrane vesicles (DMVs). While recent studies greatly contributed to our understanding of the origin and biogenesis of enterovirus ROs [18,19,21,22,64,65], for cardioviruses these details have remained scarce. In a report by Zhang et al it was proposed that EMCV subverts the autophagy pathway to promote virus replication and RO formation [27]. The authors observed induction of autophagy and accumulation of cytoplasmic double-membrane vesicles (DMVs), a hallmark of autophagosomes, upon EMCV infection. However, inhibition of autophagy had stronger effects on extracellular than intracellular virus yields, which pointed towards a role of autophagy in virus release. Indeed, recent studies using enteroviruses and flaviviruses support the hypothesis that autophagy-derived membranes rather serve as means of non-lytic virus release and spread than as a membrane source for the viral ROs [28–31].
As opposed to enteroviruses, members of the Cardiovirus genus are insensitive to GBF1 depletion by siRNA [32] or treatment with BFA [33,39,66], a compound that targets GBF1 and subsequently blocks activation of Arf1, a key regulator of membrane trafficking in the secretory pathway. Furthermore, cardiovirus replication does not require the Golgi-localized PI4KB, which is essential for enterovirus replication [34]. Collectively, these data suggested that cardioviruses do not rely on Golgi components for replication. In line with these previous findings, we here present data suggesting that EMCV may derive its ROs from ER membranes. Confocal microscopy analysis revealed that EMCV nonstructural proteins partially overlapped with the ER marker calreticulin, but not with markers of ERES, ERGIC, cis- or trans-Golgi network, which appeared dispersed in infected cells, even at the earliest stages of infection. Furthermore, PI4KA, which normally resides at the ER network, was redistributed in EMCV-infected cells to discrete cytoplasmic structures that also contained the viral protein 3AB. Interestingly, the majority of PI4KA-positive punctae were detected in close proximity to viral dsRNA, but did not completely overlap with dsRNA, suggesting a spatial segregation of dsRNA from the viral replication membranes, previously also shown for coronaviruses [67,68].
Using a pharmacological inhibitor of PI4KA, we prove that EMCV and SAFV, which belong to distinct cardiovirus species, both require the lipid kinase activity for replication. In agreement with this result, we observed elevated PI4P levels at intracellular membranes in infected cells, suggesting an important role of PI4P lipids in cardiovirus replication. In non-infected cells, OSBP plays a critical role in lipid homeostasis by exchanging cholesterol for PI4P at the interface of ER and Golgi membranes, to which it localizes under normal conditions [20]. In this process, PI4P lipids also serve as a membrane anchor for OSBP. We hypothesized that in cardiovirus infection, PI4P may serve to recruit OSBP and cholesterol to viral replication sites. Indeed, OSBP was present at EMCV ROs, where it colocalized with the viral protein 3AB. This colocalization was markedly reduced upon AL-9 treatment, demonstrating that OSBP is recruited through PI4KA-produced PI4P. OSBP is an essential cardiovirus host factor, since both genetic depletion by siRNA treatment and pharmacological inhibition by OSW-1 and 25-HC blocked viral genome replication. Cholesterol was redistributed to EMCV ROs upon infection, and treatment with AL-9 or OSW-1 resulted in a significantly reduced colocalization of cholesterol with 3AB, arguing that accumulation of cholesterol at ROs is mediated by both PI4KA and OSBP. These data are in agreement with our recent findings that cholesterol shuttling is important for cardiovirus genome replication [54] and that cardioviruses are sensitive to itraconazole, which we recently discovered to be an. inhibitor of OSBP [65].
Our results indicate that PI4P and cholesterol are vital for the global organization of EMCV ROs. However, as these lipids fulfill multiple functions in various cellular processes [53,69–71], other roles in virus replication should be envisaged. A potential task of PI4P in virus replication may be linked to the PI(4,5)P2 synthesis pathway, since PI4P is the major precursor of PI(4,5)P2 lipids, which were recently attributed an important role in HCV replication [72]. Cholesterol homeostasis was recently shown to play an important role in efficient PV polyprotein processing [73]. Whether cholesterol also ensures a proper microenvironment that supports cardiovirus polyprotein processing remains to be determined. Interestingly and in apparent parallel with the distantly related enteroviruses, exploitation of the PI4K-OSBP pathway by HCV correlates with the induction of membranes of positive curvature [58]. By contrast, the flavivirus DENV, although closely related to HCV, does not require PI4K or OSBP [23] and generates membranes of negative curvature [74]. Hence, the interplay between PI4P and cholesterol may dictate the positive curvature of the membranes at which diverse RNA viruses replicate their genomes.
Through co-IP assays, we identified PI4KA as a novel interaction partner of EMCV proteins 3A and its precursor 3AB. EMCV 3A is a small protein (88 amino acids) of unknown structure, containing a predicted hydrophobic domain in the C-terminus half. Expression of 3A alone was sufficient for PI4KA recruitment in intact cells, arguing that in infection, PI4KA is recruited to ROs by this viral protein. Enteroviruses and kobuviruses recruit PI4KB to ROs also via their 3A protein [15,16,45,57], raising the possibility that diverse picornaviruses might use an evolutionary conserved and 3A-mediated mechanism to generate PI4P-enriched membranes. However, the 3A proteins of entero-, kobu- and cardioviruses do not share any apparent sequence similarity, their name simply reflecting the occupancy of the same locus (3A) in the respective viral genomes. With the exception of their catalytic domain, also the PI4KA and PI4KB isoforms do not share any sequence similarity [75]. Furthermore, unlike 3A of most enteroviruses, cardiovirus 3A does not interact with GBF1 nor blocks protein transport in the secretory pathway when expressed alone [76], highlighting the functional diversification associated with these small viral proteins.
Several lines of evidence suggest that the picornavirus EMCV and the distantly related flavivirus HCV have evolved to exploit common host components in assisting virus RNA replication. First, HCV genome replication occurs at the “membranous web” (MW), a network of single and DMVs that, like the EMCV RO, also mainly originates from the ER [58]. Second, both EMCV and HCV express a viral protein dedicated to recruitment of PI4KA, in order to induce a PI4P-rich environment at the replication sites [24,46]. Third, in HCV infection, PI4P lipids were also shown to be important for the recruitment of OSBP, which mediates cholesterol transfer to the MW [23]. Fourth, inhibition of either PI4KA or OSBP induced clear alterations in the global distribution of EMCV ROs, which appeared more “clustered” upon treatment with AL-9 or OSW-1. A similar clustering effect was also observed for replication structures of the HCV MW upon PI4KA or OSBP inhibition [23,24], whereas no obvious disruption of the enterovirus ROs was observed upon PI4KB or OSBP inhibition [18,19,65]. Together, these observations indicate that EMCV and HCV replication structures share critical host components, and possibly also a similar architecture, although the latter still remains to be determined.
Based on at least two lines of emerging evidence in the context of phylogeny of flavi- and picornaviruses, EMCV and HCV have likely converged on rather than retained their functional similarities upon divergence from the common ancestor (Fig 8). First, the observed commonalities between EMCV and HCV are not shared by other characterized viruses in their respective families, indicating that they are not a manifestation of the properties conserved among the two families. For instance, picornaviruses from different genera exhibit different host factor requirements. Members of the Cardiovirus genus hijack the ER-localized PI4KA (this study), whereas members of the Enterovirus and Kobuvirus genera depend on the Golgi-localized PI4KB [15,34,57]. In contrast, equine rhinitis A virus (ERAV, member of Aphthovirus genus, which is prototyped by FMDV) and hepatitis A virus (Hepatovirus genus) seem to replicate independent of both PI4KB and PI4KA (S7 Fig and [34,77]). Likewise, the flaviviruses DENV and WNV, representing a sister genus to that of HCV, do not rely on either PI4KA or PI4KB [23,78]. While DENV was also shown not to require OSBP [23], for WNV this is not known yet. Importantly, cardioviruses targeting PI4KA occupy a lineage that is farther from the root compared to those of entero- and kobuviruses targeting PI4KB (Fig 8). This phylogenetic pattern is indicative of the relatively recent emergence of the EMCV-specific target properties. Second, EMCV and HCV employ apparently unrelated proteins to mediate the interaction with PI4KA, namely 3A and NS5A (although HCV NS5B may contribute as well [24,46]). Both proteins are membrane-bound, albeit through a hydrophobic domain located at either N-terminus (NS5A) or in the C-terminus-half (3A), and each includes another region which is among the least conserved in the nonstructural proteins of the respective families [79,80].
Our study contributes to the hypothesis that viruses may be confronted with powerful constraints that limit the diversity of host pathways recruited for efficient replication. Thus, a common pathway is used by different RNA viruses that either only moderately diverged (e.g. different species of same genus) or converged on a host target while diverging profoundly (different families–e.g. EMCV and HCV). To date, only a small number of (+)RNA viruses have been studied in terms of host lipid requirements. Identification of the lipid pathways used by other viruses will hopefully provide a deeper insight on the constraints that viruses are confronted with during the endeavor to replicate their genome.
Buffalo green monkey (BGM) kidney cells, baby hamster kidney 21 (BHK-21) and HeLa R19 cells were grown at 37°C and 5% CO2 in Dulbecco’s modified Eagle’s medium (DMEM, Lonza) supplemented with 10% fetal bovine serum (FBS). Huh7Lunet/T7 cells (provided by R. Bartenschlager, Department of Molecular Virology, University of Heidelberg, Heidelberg, Germany) [82] were grown in DMEM (Lonza) supplemented with 10% FBS and 10 μg/ml Blasticidin (PAA). BGM cells were purchased from ECACC and BHK-21 cells were purchased from ATCC. HeLa R19 cells were obtained from G. Belov (University of Maryland and Virginia-Maryland Regional College of Veterinary Medicine, US) [83]. AL-9 and OSW-1 were kind gifts from J. Neyts (Rega Institute for Medical Research, University of Leuven, Leuven, Belgium) and M.D. Shair (Department of Chemistry and Chemical Biology, Harvard University, Cambridge, USA) respectively. 25-HC was purchased from Santa Cruz. BF738735 [84] was provided by Galapagos NV. Filipin III and dipyridamole were from Sigma.
Constructs pTM-HA-PI4KA [24], pEGFP-PI4KA (provided by G. Hammond, NICHD, National Institutes of Health, Bethesda, USA) [41,85] and p3A(CVB3)-myc [86] were described previously. pGFP-PI4KB was a kind gift from N. Altan-Bonnet (Laboratory of Host-Pathogen Dynamics, National Institutes of Health, Bethesda, USA). To generate C-terminal myc-tagged EMCV proteins, genes encoding EMCV nonstructural proteins 2A, 2B, 2C, 3A, 3AB, 3C and 3D were amplified by PCR using the plasmid pM16.1 [87] and primers introducing restriction sites BamHI and HindIII. pM16.1 contains the full-length infectious cDNA sequence of EMCV, strain mengovirus. The PCR products were then cloned into the p3A(CVB3)-myc backbone from which the CVB3-3A gene was removed using the same restriction enzymes. To allow ectopic expression of PI4KA under a CMV promoter, the gene encoding HA-PI4KA was amplified by PCR using pTM-HA-PI4KA as template and introduced in the pEGFP-N3 backbone using restriction enzymes SalI and NotI. EMCV-2C-HA infectious clone was generated by introducing the HA coding sequence (YPYDVPDYA) in-frame after codon 2 in 2C of pM16.1 using mutagenesis primers and the Q5 Site-Directed Mutagenesis Kit (New England Biolabs).
EMCV, EMCV-2C-HA and RLuc-EMCV, which contains the Renilla luciferase gene upstream of the capsid coding region [54], were obtained by transfecting BHK-21 cells with RNA transcripts derived from full length infectious clones pM16.1, pM16.1-2C-HA and pRLuc-QG-M16.1, respectively, linearized with BamHI. GFP-EMCV, which contains the EGFP gene upstream the capsid region, was generated similar as RLucEMCV [54]. CVB3 (strain Nancy) was obtained by transfecting BGM cells with RNA transcripts of the full length infectious clone p53CB3/T7 [86] linearized with SalI. Saffold virus (type 3) was described previously [2]. ERAV (NM11/67) was kindly provided by David Rowlands and Toby Tuthill (University of Leeds, United Kingdom). Virus infections were performed by incubating subconfluent cell monolayers for 30 min at 37°C with virus, after which the virus-containing medium was removed and fresh (compound-containing) medium was added to the cells (t = 0). In the time-of-addition experiments, medium without compound was added at t = 0 and replaced by medium with compound at the indicated time points. At the given time points post infection, cells were either fixed for immunolabeling, freeze-thawed to determine virus titers or, in the case of RLuc-EMCV, lysed to determine replication by measuring the intracellular Renilla luciferase activity using the Renilla Luciferase Assay System (Promega). Virus titers were determined by endpoint titration according to the method of Reed and Muench and expressed as 50% tissue culture infective doses (TCID50).
HeLa R19 or Huh7Lunet/T7 cells were grown to subconfluency on coverslips in 24-well plates. Where indicated, cells were transfected with 400 ng of plasmids using Lipofectamine2000 according to the manufacturer’s protocol and/or infected with EMCV at the specified multiplicity of infection (MOI), followed by compound treatment where specified. At the indicated time points, cells were fixed with 4% paraformaldehyde (PFA) for 20 min at room temperature (RT). Permeabilization was done with PBS-0.5% Triton X-100 for 15 min or PBS/0.2% saponin/5% BSA for 5 min, in the case of filipin staining. Cells were incubated sequentially with primary and secondary antibodies diluted in PBS containing 2% normal goat serum (NGS). The following primary antibodies were used for detection: mouse monoclonal anti-GM130 (BD Biosciences), rabbit polyclonal anti-TGN46 (Novus Biologicals), mouse monoclonal anti-ERGIC53 (Enzo Life Sciences), rabbit polyclonal anti-Sec13 (kindly provided by B.L Tang, Department of Biochemistry, The National University of Singapore, Singapore), rabbit polyclonal anti-calreticulin (Sigma), rabbit polyclonal anti-HA (Santa Cruz), mouse monoclonal anti-HA (Abcam), mouse monoclonal anti-C-Myc (Sigma), rabbit polyclonal anti-myc (Thermo Scientific), mouse anti-PI4P IgM (Echelon Biosciences), mouse monoclonal anti-dsRNA (J2, English & Scientific Consulting), mouse monoclonal anti-EMCV 3AB (kind gift from A.G. Aminev) [88] and rabbit polyclonal anti-OSBP (kindly provided by M.A. De Matteis, Telethon Institute of Genetics and Medicine, Naples, Italy) [65]. Alexa Fluor 488-, 594-conjugated IgG and Alexa Fluor 488- or 594-conjugated IgM (Invitrogen, Molecular Probes) were used as secondary antibodies. Cholesterol was stained with 25 μg/ml filipin III (Sigma) for 1 h at room temperature, included during the incubation with the secondary antibody. Nuclei were counterstained with DAPI.
Staining of plasma membrane or intracellular PI4P was performed as described elsewhere [50]. Briefly, for PM staining, cells were fixed at RT in 4% PFA and 0.2% glutaraldehyde. All subsequent steps were performed on ice. Cells were blocked and permeabilized for 45 min in buffer A (20mM Pipes, pH 6.8, 137 mM NaCl, 2.7 mM KCl) containing 5% NGS, 50 mM NH4Cl and 0.5% saponin. Slides were incubated with primary and secondary antibodies in buffer A containing 5% NGS and 0.1% saponin for 1 h. Finally, slides were post-fixed in 2% PFA in PBS for 10 min. The intracellular PI4P staining was performed at RT as follows: cells were fixed with 2% PFA, then permeabilized for 5 min in 20 μM digitonin in buffer A, blocked for 45 min in buffer A with 5% NGS and 50 mM NH4Cl and then incubated sequentially with primary and secondary antibodies in buffer A with 5% NGS, before post fixation in 2% PFA. All coverslips were mounted with FluorSave (Calbiochem). Images were acquired with a Leica SPE-II DMI-4000 confocal laser scanning microscope or a Nikon Ti Eclipse microscope equipped with an Andor DU-897 EMCCD-camera.
PI4P quantification was performed for at least 40 cells for each condition, using the ImageJ software as described elsewhere [46]. To determine colocalization of Sec13 or calreticulin with 3AB, images were first deconvoluted using NIS advanced Research 4.3 software (Nikon) (10 iterations) and further processed using Image J as follows. Individual infected cells were outlined and a mask was created, and all signal outside the mask was cropped to exclude it from the calculations. Manders’ colocalization coefficient was calculated for at least 10 cells for each condition using the JACoP plugin [89] with a manually set threshold. Colocalization of OSBP with 3A in infected cells was analyzed using ImageJ by determining Pearson’s coefficient for at least 15 cells per condition using the Coloc 2 plugin with default settings. To quantify colocalization of filipin with 3AB, images were first deconvoluted using NIS software (20 iterations), then ImageJ was used to select infected cells and the Pearson’s coefficient of colocalization for at least 15 cells per condition was calculated using the Coloc 2 plugin with default settings.
HeLa R19 cells were reverse-transfected with 2 pmoles of siRNA per well of a 96-well plate (2000 cells/well) using Lipofectamine 2000 (Invitrogen) according to the manufacturer’s indications. Scrambled siRNA (AllStars Neg. Control, Qiagen) was used as a control. SiRNA against hPI4KA (cat. no. S102777390) and hPI4KB (target sequence: 5’-UGUUGGGGCUUCCCUGCCCTT-3’) were from Qiagen. siRNA against hOSBP (two siRNAs mixed at 1:1 ratio, target sequences: 5’- CGCUAAUGGAAGAAGUUUA[dT][dT]-3’ and 5’-CCUUUGAGCUGGACCGAUU[dT][dT]-3’)) was from Sigma. 48 h p.t., cells were either infected with virus, transfected with in vitro transcribed RNA derived from the full length infectious clone pM16.1 or harvested to evaluate the knockdown efficiency by western blot analysis.
Cell viability was determined in parallel with virus infection as follows. One day after seeding cells in a 96-well plate, the compounds were added to the cells and incubated for 8 h. Alternatively, cells were transfected with siRNAs and incubated for 48 h. Subsequently, the medium was replaced with CellTiter 96 AQueous One Solution Reagent (Promega) and optical densities were measured at 490 nm. The obtained raw values were converted to percentage of untreated samples or samples transfected with scrambled siRNAs, following correction for background absorbance.
Metabolic labeling of myc-tagged EMCV proteins and HA-PI4KA was performed as described elsewhere [46]. Briefly, Huh7-Lunet/T7 cells seeded in 6-well plates were co-transfected with 2 μg of plasmid encoding EMCV nonstructural proteins and 2 μg of either pTM HA-PI4KIIIa or an empty pTM vector (mock) using Lipofectamine2000 (Invitrogen) according to the manufacturer’s instructions. 7 h later, cells were starved in methionine/cysteine-free medium for 1 h. Radiolabeling of cells was done by overnight incubation in methionine/cysteine-free medium, supplemented with 10 mM glutamine, 10 mM Hepes, and 100 μCi/ml of Express Protein labeling mix (Perkin Elmer, Boston). Cells were then harvested and lysed in lysis buffer (50 mM Tris-Cl [pH 7.5], 150 mM NaCl, 1% Nonidet P-40 and protease inhibitors) for 1 h on ice, followed by centrifugation at 14,000 g for 10 min at 4°C. Supernatants were further subjected to immunoprecipitation by a 3 h incubation at 4°C with anti-c-myc rabbit polyclonal antibody (Santa Cruz). Immunocomplexes were then captured with protein G-sepharose beads (Sigma) by an additional 3 h incubation at 4°C. Beads were washed three times in lysis buffer, followed by elution of immunocomplexes by boiling in sample buffer, separation by polyacrylamide-SDS gel electrophoresis and detection by autoradiography. For co-IP followed by western blot, cells were seeded in 55 cm2 dishes and transfected with 3.5 μg of each plasmid using polyethylenimine (PEI) (Polysciences). Immunoprecipitation was carried out as described above, but using protein A-sepharose beads (GE Healthcare) and mouse monoclonal anti-C-Myc (Sigma) or rabbit polyclonal anti-myc (Thermo Scientific) antibodies.
Samples separated by SDS-PAGE were transferred to nitrocellulose membranes (Bio-Rad). Membranes were incubated with the following primary antibodies: rabbit polyclonal anti-PI4KA (Cell Signaling), rabbit polyclonal anti-PI4KB (Upstate), rabbit polyclonal anti-OSBP (ProteinTech), rabbit polyclonal anti-EMCV capsid (kind gift from Ann Palmenberg) and mouse monoclonal anti-β-actin (Sigma). Secondary antibodies included IRDye 680-conjugated goat anti-mouse or IRDye 800-conjugated goat anti-rabbit (LI-COR). Images of blots were acquired with an Odyssey Fc Imaging System (LI-COR).
Where indicated, unpaired one-tailed Student’s t-test or two-tailed Mann–Whitney test were applied as statistical analyses using the GraphPad Prism software.
|
10.1371/journal.pcbi.1004792 | Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework | The ability to simultaneously record from large numbers of neurons in behaving animals has ushered in a new era for the study of the neural circuit mechanisms underlying cognitive functions. One promising approach to uncovering the dynamical and computational principles governing population responses is to analyze model recurrent neural networks (RNNs) that have been optimized to perform the same tasks as behaving animals. Because the optimization of network parameters specifies the desired output but not the manner in which to achieve this output, “trained” networks serve as a source of mechanistic hypotheses and a testing ground for data analyses that link neural computation to behavior. Complete access to the activity and connectivity of the circuit, and the ability to manipulate them arbitrarily, make trained networks a convenient proxy for biological circuits and a valuable platform for theoretical investigation. However, existing RNNs lack basic biological features such as the distinction between excitatory and inhibitory units (Dale’s principle), which are essential if RNNs are to provide insights into the operation of biological circuits. Moreover, trained networks can achieve the same behavioral performance but differ substantially in their structure and dynamics, highlighting the need for a simple and flexible framework for the exploratory training of RNNs. Here, we describe a framework for gradient descent-based training of excitatory-inhibitory RNNs that can incorporate a variety of biological knowledge. We provide an implementation based on the machine learning library Theano, whose automatic differentiation capabilities facilitate modifications and extensions. We validate this framework by applying it to well-known experimental paradigms such as perceptual decision-making, context-dependent integration, multisensory integration, parametric working memory, and motor sequence generation. Our results demonstrate the wide range of neural activity patterns and behavior that can be modeled, and suggest a unified setting in which diverse cognitive computations and mechanisms can be studied.
| Cognitive functions arise from the coordinated activity of many interconnected neurons. As neuroscientists increasingly use large datasets of simultaneously recorded neurons to study the brain, one approach that has emerged as a promising tool for interpreting population responses is to analyze model recurrent neural networks (RNNs) that have been optimized to perform the same tasks as recorded animals. Complete access to the activity and connectivity of the circuit, and the ability to manipulate them in arbitrary ways, make trained networks a convenient proxy for biological circuits and a valuable platform for theoretical investigation. However, existing RNNs lack basic biological features that are essential if RNNs are to provide insights into the circuit-level operation of the brain. Moreover, trained networks can achieve the same behavioral performance but differ substantially in their structure and dynamics, highlighting the need for a simple and flexible framework for the exploratory training of RNNs. Here we describe and provide an implementation for such a framework, which we apply to several well-known experimental paradigms that illustrate the diversity of detail that can be modeled. Our work provides a foundation for neuroscientists to harness trained RNNs in their own investigations of the neural basis of cognition.
| Computations in the brain are carried out by populations of interconnected neurons. While single-neuron responses can reveal a great deal about the neural mechanisms underlying various sensory, motor, and cognitive functions, neural mechanisms often involve the coordinated activity of many neurons whose complex individual dynamics are not easily explained by tuning to experimental parameters [1–4]. A growing recognition of the importance of studying population-level responses is reflected in the increasing number of studies that use large datasets of simultaneously or sequentially recorded neurons to infer neural circuit mechanisms [5–9]. At the same time, the novel challenges posed by high-dimensional neural data have led to the development of new methods for analyzing and modeling such data [10–12].
One approach that has emerged as a promising tool for identifying the dynamical and computational mechanisms embedded in large neural populations is to study model recurrent neural networks (RNNs) whose connection weights have been optimized to perform the same tasks as recorded animals [5, 7]. In [5], for example, the “trained” network was analyzed to reveal a previously unknown selection mechanism for context-dependent integration of sensory stimuli that was consistent with data obtained from behaving monkeys. RNNs of rate units, which describe biological circuits as a set of firing rates (nonlinearities) interacting through synapses (connection weights) (Fig 1), interpolate between biophysically detailed spiking-neuron models and the wider class of continuous-time dynamical systems: the units of an RNN can be interpreted as the temporal or ensemble average of one or more co-tuned spiking neurons [13], while any nonlinear dynamical system can be approximated by an RNN with a sufficient number of units [14, 15]. The optimization of network parameters typically specifies the desired output but not the manner in which to achieve this output, i.e., the what but not the how. Trained RNNs therefore serve as a source of candidate hypotheses about circuit mechanisms and a testing ground for data analyses that link neural computation to behavior. Complete access to the activity and connectivity of the circuit, and the ability to manipulate them in arbitrary ways, make trained networks a convenient proxy for biological circuits and a valuable platform for theoretical investigation [12, 16, 17].
For many tasks of interest, however, training can result in multiple networks that achieve the same behavioral performance but differ substantially in their connectivity and dynamics. As highlighted in recent work [8], the particular solution that is discovered by the training algorithm depends strongly on the set of constraints and “regularizations” used in the optimization process, so that training RNNs to perform a task is not entirely unbiased with respect to the how. Indeed, for the purposes of modeling animal tasks in systems neuroscience the question is no longer whether an RNN can be trained to perform the task—the answer appears to be yes in a wide range of settings—but what architectures and regularizations lead to network activity that is most similar to neural recordings obtained from behaving animals.
Answering this question is essential if RNNs are to provide insights into the operation of the brain at the level of neural circuits [18], and extends the classical connectionist approach [19, 20]. Doing so requires a simple and flexible framework for the exploratory training of RNNs to investigate the effects of different constraints on network properties, particularly those constraints that render the RNNs more biologically plausible. For instance, many RNNs studied to date have “firing rates” that are both positive and negative. More fundamentally, existing networks do not satisfy Dale’s principle [21], the basic and ubiquitous observation that neurons in the mammalian cortex have purely excitatory or inhibitory effects on other neurons. The analogous constraint that all connection weights from a given unit must have the same sign can have a profound effect on the types of dynamics, such as non-normality [22], that operate in the circuit. Moreover, connections from excitatory and inhibitory neurons exhibit different levels of sparseness and specificity, with non-random features in the distribution of connection patterns among neurons both within local circuits [23–27] and among cortical areas [28–30]. Notably, long-range projections between areas are primarily excitatory. Such details must be included in a satisfactory model of local and large-scale cortical computation.
We address this challenge by describing flexible, gradient descent-based training of excitatory-inhibitory RNNs that can incorporate a variety of biological knowledge, particularly of local and large-scale connectivity in the brain. Several different methods have previously been used to train RNNs for cognitive tasks in neuroscience, including first-order reduced and controlled error (FORCE) [7, 31, 32] and Hessian-free (HF) [5, 33, 34]. Here we use minibatch stochastic gradient descent (SGD) with the modifications described in [35], which remove the major difficulties associated with pure gradient descent training of RNNs. SGD is conceptually simple without sacrificing performance [36, 37] and is particularly advantageous in the present context for the following reasons: Unlike FORCE and like HF, SGD allows us to more easily formulate the problem of training an RNN as one of minimizing an objective function that can be modified to induce different types of solutions [8]. Meanwhile, like FORCE and unlike HF, for many tasks SGD can update parameters on a trial-by-trial basis, i.e., in an “online” fashion. This opens up the possibility of exploring across-trial effects that cannot be studied when large numbers of trials are required for each iteration of learning, as in the HF algorithm. Although none of the learning methods discussed here can at present be considered biological, recent work also suggests that spike-timing dependent plasticity (STDP) [38], which is believed to be a basic rule governing synaptic weight changes in the brain, may correspond to a form of SGD [39, 40]. However, the focus of our approach will be on the results, not the mechanism, of learning.
We provide an implementation of this framework based on the Python machine learning library Theano [41, 42], whose automatic differentiation capabilities facilitate modifications and extensions. Theano also simplifies the use of Graphics Processing Units (GPUs) when available to speed up computations. The implementation was designed to minimize the overhead for each new task by only requiring a specification of the network structure and correct input-output relationship to be learned. It also streamlines the testing and analysis of the resulting networks by using the same (customizable) specification for both training and testing (S1 Code). We demonstrate the application of this framework to well-known experimental paradigms that illustrate the diversity of tasks and details that can be modeled: perceptual decision-making, context-dependent integration, multisensory integration, parametric working memory, and eye-movement sequence generation. Using the resulting networks we perform both single-neuron and population-level analyses associated with the corresponding experimental paradigm. Our results show that trained RNNs provide a unified setting in which diverse computations and mechanisms can be studied, laying the foundation for more neuroscientists to harness trained RNNs in their own investigations of the neural basis of cognition.
In this section we first define the RNNs used in this work, show how constraints can be introduced, then describe training the networks using a modified form of stochastic gradient descent (SGD).
RNNs receive a set of Nin time-varying inputs u(t) and produce Nout outputs z(t), where inputs encode task-relevant sensory information and outputs typically represent a decision variable or probability distribution (Fig 1). Outputs can also relate to the direct motor effector, such as eye position, by which an animal indicates its decision in the behavioral paradigm. We consider RNNs whose N firing rates r(t) are related to their corresponding currents x(t) by the threshold (rectified) linear “f-I curve” [x]+ = max(x, 0), which maps arbitrary input currents to positive firing rates: x if x > 0 and 0 otherwise. The RNNs are described by the equations
τ x ˙ = - x + W rec r + W in u + 2 τ σ rec 2 ξ , (1) r = [ x ] + , (2) z = W out r , (3)
or, more explicitly,
τ d x i d t = - x i + ∑ j = 1 N W i j rec r j + ∑ k = 1 N in W i k in u k + 2 τ σ rec 2 ξ i , (4) r i = [ x i ] + , (5) z ℓ = ∑ i = 1 N W ℓ i out r i (6)
for i = 1, …, N and ℓ = 1, …, Nout. In these equations τ is the time constant of the network units, Win is an N × Nin matrix of connection weights from the inputs to network units, Wrec is an N × N matrix of recurrent connection weights between network units, Wout is an Nout × N matrix of connection weights from the network units to the outputs, and ξ are N independent Gaussian white noise processes with zero mean and unit variance that represent noise intrinsic to the network. It is worth noting that if for some ℓ = 1, …, N′, N′ ≤ N, the output weights W ℓ i out = δ ℓ i where δij = 1 if i = j and 0 otherwise, then the readout is the same as a subset of the network firing rates. This is useful in situations where the aim is to fix a subset of the units to experimentally recorded firing rates.
Without the rectification nonlinearity [x]+ (in which case r = x), Eqs 1–3 would describe a linear system whose dynamics is completely determined by Wrec. Thus, one way to understand the effect of rectification is to consider a linear dynamical system whose coupling matrix Wrec at any given time includes only those columns that correspond to “active” units with positive summed current xi (and hence positive firing rate ri) [43]. This toggles the network between different linear maps, thereby endowing the network with the capacity for more complex computations than would be possible with a single linear network [44, 45]. As a convenient baseline, the recurrent noise in Eq 1 has been scaled so that in the corresponding linear network without rectification each unit is an Ornstein-Uhlenbeck process with variance σ rec 2 when Wrec = Win = 0.
In practice, the continuous-time dynamics in Eqs 1–3 are discretized to Euler form (which we indicate by writing time as a subscript, Xt = X(t ⋅ Δt) for a time-dependent variable X) in time steps of size Δt as [46]
x t = ( 1 - α ) x t - 1 + α ( W rec r t - 1 + W in u t ) + 2 α σ rec 2 N ( 0 , 1 ) , (7) r t = [ x t ] + , (8) z t = W out r t , (9)
where α = Δt/τ and N(0, 1) are normally distributed random numbers with zero mean and unit variance, sampled independently at every time step. In this formulation, the usual discrete-time RNNs used in machine learning applications correspond to α = 1 or Δt = τ. To minimize computational effort we train the network with a value of Δt that is as large as possible such that the same network behavior is recovered in the continuous limit of Δt → 0.
Although the details of the inputs to the network are specific to each task, it is convenient to represent all inputs as a rectified sum of baseline u0, task-dependent signal utask(t), and Gaussian white noise ξ:
u ( t ) = u 0 + u task ( t ) + 2 τ σ in 2 ξ + (10)
in the continuous description, and
u t = u 0 + u t task + 1 α 2 α σ in 2 N ( 0 , 1 ) + (11)
in the discrete-time description. Motivated by the interpretation that the network under study is only one part of a larger circuit, the baseline and noise terms in the inputs can together be considered the spontaneous firing rate of “upstream” units that project to the network.
We note that in Eq 1 the external “sensory” noise ultimately combines with the intrinsic noise, with the difference that input noise can be shared between many units in the network while the recurrent noise is private to each unit. There are many cases where the external and internal noise trade off in their effect on the network, for instance on its psychometric performance in a perceptual decision-making task. However, the two sources of noise can be biologically and conceptually quite different [47], and for this reason it is helpful to separate the two types of noise in our formulation.
Finally, in many cases (the exception being networks that are run continuously without reset) it is convenient to optimize the initial condition x0 = x(0) at time t = 0 along with the network weights. This merely selects a suitable starting point for each run, reducing the time it takes for the network to relax to its spontaneous state in the absence of inputs. It has little effect on the robustness of the network due to the recurrent noise used both during and after training; in particular, the network state at the time of stimulus onset is highly variable across trials.
A basic and ubiquitous observation in the mammalian cortex, known in the more general case as Dale’s principle [21], is that cortical neurons have either purely excitatory or inhibitory effects on postsynaptic neurons. Moreover, excitatory neurons outnumber inhibitory neurons by a ratio of roughly 4 to 1. In a rate model with positive firing rates such as the one given by Eqs 1–3, a connection from unit j to unit i is “excitatory” if W i j rec > 0 and “inhibitory” if W i j rec < 0. A unit j is excitatory if all of its projections on other units are zero or excitatory, i.e., if W i j rec ≥ 0 for all i; similarly, unit j is inhibitory if W i j rec ≤ 0 for all i. In the case where the outputs are considered to be units in a downstream network, consistency requires that for all ℓ the readout weights satisfy W ℓ j out ≥ 0 and W ℓ j out ≤ 0 for excitatory and inhibitory units j, respectively. Since long-range projections in the mammalian cortex are exclusively excitatory, for most networks we limit readout to the excitatory units. It is also natural in most cases to assume that inputs to the network are long-range inputs from an upstream circuit, and we assume all elements of the input weight matrix Win are non-negative. For consistency with the following, we indicate this as Win = Win,+. Once again, this is only meaningful if the inputs themselves are always non-negative, motivating the rectification of inputs in Eq 10.
In order to train RNNs that satisfy the above constraints, we parametrize the recurrent weight matrix Wrec as the product of a non-negative matrix Wrec,+ and a diagonal matrix D of 1’s and −1’s, Wrec = Wrec,+ D. For example, consider a network containing 4 excitatory units and 1 inhibitory unit; the excitatory/inhibitory signature of the network is then D = diag(1, 1, 1, 1, −1) (a matrix with the specified entries on the diagonal and zeros everywhere else), and the full recurrent weight matrix has the form
+ + + - + + + - + + + - + + + - + + + + ︸ W rec = + + + + + + + + + + + + + + + + + + + + ︸ W rec,+ 1 1 1 1 - 1 ︸ D , (12)
where absent matrix elements indicate zeros. Although an individual unit in an RNN does not necessarily represent a single neuron, we typically fix the self-connections represented by the diagonal elements of Wrec to be zero, see below. Similarly, if the readout from the network is considered to be long-range projections to a downstream network, then the output weights are parametrized as Wout = Wout,+ D.
During training, the positivity of Win,+, Wrec,+, and Wout,+ can be enforced in several ways, including rectification [W]+ and the absolute value function |W|. Here we use rectification.
In addition to dividing units into separate excitatory and inhibitory populations, we can also constrain their pattern of connectivity. This can range from simple constraints such as the absence of self-connections to more complex structures derived from biology. Local cortical circuits have distance [48], layer [26, 49, 50], and cell-type [23, 25, 27, 51] dependent patterns of connectivity and different overall levels of sparseness for excitatory to excitatory, inhibitory to excitatory, excitatory to inhibitory, and inhibitory to inhibitory connections [52, 53]. Although the density of connections in a trained network can be either fixed (hard constraint) or induced through regularization (soft constraint) (see Eq 27), here we focus on the former to address the more general problem of imposing known biological structure on trained networks. For instance, in models of large-scale, distributed computation in the brain we can consider multiple cortical “areas” characterized by local inhibition within areas and long-range excitation between areas. These long-range connections can be distributed according to a highly complex topology [28–30]. It is also desirable when testing specific hypotheses about circuit structure to fix a subset of the connection weights to predefined values while leaving others as “plastic,” modifiable by training.
A simple way to impose hard constraints on the connectivity is to parametrize the weight matrices using masks. As an example, suppose we would like to train a subset of the excitatory weights and also fix two of the inhibitory weights to w1 and w2 so that they are not modified during training. We can implement this by writing
W rec,+ = 0 1 1 1 0 1 0 1 0 1 1 1 0 1 1 0 1 1 0 0 1 1 1 1 0 ︸ M rec ⊙ · + + + · + · + · + + + · + + · + + · · + + + + · ︸ W rec,plastic,+ + 0 0 0 0 w 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 w 2 0 0 0 0 0 ︸ W rec,fixed,+ , (13)
where ⊙ denotes the element-wise multiplication of two matrices (not standard matrix multiplication). Here Wrec,plastic,+ is obtained by rectifying the (unconstrained) trained weights Wrec,plastic, so that Wrec,plastic,+ = [Wrec,plastic]+, while Wrec,fixed,+ is a matrix of fixed weights. The elements that are marked with a dot are irrelevant and play no role in the network’s dynamics. Eq 13 has the effect of optimizing only those elements which are nonzero in the multiplying mask Mrec, which ensures that the weights corresponding to zeros do not contribute. Some elements, for instance the inhibitory weights w1 and w2 in Eq 13, remain fixed at their specified values throughout training. Explicitly, the full weight matrix of the RNN is related to the underlying trained weight matrix Wrec,plastic by (cf. Eq 12)
W rec = ( M rec ⊙ [ W rec,plastic ] + + W rec,fixed,+ ) D , (14)
and similarly for the input and output weights.
In networks that do not contain separate excitatory and inhibitory populations, it is convenient to initialize the recurrent weight matrix as W rec = ρ W 0 rec, where W 0 rec is formed by setting a fraction p, 0 < p ≤ 1, of elements to nonzero values drawn from a Gaussian distribution with mean 0 and variance (pN)−1, and the remaining fraction 1 − p to zero [31]. This can be understood as first generating a random matrix W 0 rec, then multiplying by ρ/ρ0 where ρ0 = 1 is the spectral radius of W 0 rec and ρ is the desired spectral radius of the initial weight matrix. Here the spectral radius is the largest absolute value of the eigenvalues.
To initialize an excitatory-inhibitory network with an arbitrary pattern of connections, we similarly first generate a matrix W 0 rec and let W rec = ( ρ / ρ 0 ) W 0 rec where ρ0 is the spectral radius of W 0 rec. Unlike in the case of random Gaussian matrices, the (asymptotically) exact spectral radius is usually unknown and must be computed numerically. Moreover, since the signs of the matrix elements are determined by the excitatory or inhibitory nature of the units, it is more natural to use a distribution over positive numbers to first generate W 0 rec,+ (Eq 12). Many distributions, including the uniform and log-normal distributions, can be used; inspired by previous work [54], here we use the gamma distribution to initialize the recurrent weight matrix W 0 rec,+. The means μE (excitatory) and μI (inhibitory) of the gamma distributions are chosen to balance the excitatory and inhibitory inputs to each unit [55], i.e., ∑j ∈ exc |μj| = ∑j ∈ inh |μj|, with the overall mean set by the imposed spectral radius ρ. We did not use the “initialization trick” of [56], as this requires the existence of self-connections.
For the input weight matrix W 0 in,+ and output weight matrix W 0 out,+, we initialize with small positive numbers drawn from a uniform distribution.
To train an RNN, we assume that at each time step (or subset of time steps) there is a correct set of target outputs z t target that depend on the current and previous history of inputs ut′ for t′ ≤ t, i.e., we only consider tasks that can be translated into a “supervised” form. The goal is then to find network parameters, which we collectively denote as θ, that minimize the difference between the correct output and the actual output of the network. More generally, we minimize an objective function E ( θ ) that includes not only this error but other terms such as an L1-regularization term (for encouraging sparse weights or activation patterns) that influence the types of solutions found by the training algorithm. We begin with the case where the objective function depends only on the error; one possibility for the loss L ( θ ) that measures the difference between the correct and actual outputs is the squared sum of differences averaged over Ntrials trials, Nout outputs, and Ntime time points:
E = 1 N trials ∑ n = 1 N trials L n , (15) L n = 1 N out N time ∑ ℓ = 1 N out ∑ t = 1 N time M t ℓ error ( z t ) ℓ - ( z t target ) ℓ 2 . (16)
For each trial n in Eq 16, (zt)ℓ is the ℓ-th output, at time t, of the discretized network in Eq 9. The error mask Merror is a matrix of ones and zeros that determines whether the error in output ℓ at time t should be taken into account. In many decision-making tasks, for example, this allows us to train networks by specifying only the final, but not the intermediate, time course for the outputs.
In gradient descent training the parameters of the network are updated iteratively according to (for more sophisticated forms of gradient descent see, e.g., [57])
θ ( i ) = θ ( i - 1 ) + δ θ ( i - 1 ) , (17)
where i denotes the iteration. The parameter change, δθ, is taken to be proportional to the negative gradient of the objective function with respect to the network parameters as
δ θ ( i - 1 ) = - η ∇ E ( i - 1 ) , (18)
where η is the learning rate and ∇ E ( i - 1 ) = ∇ E ( θ ( i - 1 ) ) is the value of the gradient evaluated on the parameters from iteration i − 1. Importantly, the required gradient can be computed efficiently by backpropagation through time (BPTT) [58] and automatically by the Python machine library Theano [41, 42]. In component form the parameter update at iteration i is given by
θ k ( i ) = θ k ( i - 1 ) - η ∂ E ∂ θ k ( i - 1 ) , (19)
where k runs over all the parameters of the network that are being optimized. Eqs 17 and 18 are motivated by the observation that, for a small change δθ in the value of the parameters, the corresponding change in the value of the objective function is given by
E ( θ + δ θ ) - E ( θ ) ≃ ∇ E · δ θ = | ∇ E | | δ θ | cos ϕ , (20)
where |⋅| denotes the norm of a vector and ϕ is the angle between ∇ E and δθ. This change is most negative when ϕ = 180°, i.e., when the change in parameters is in the opposite direction of the gradient. “Minibatch stochastic” refers to the fact that the gradient of the objective function E ( θ ) is only approximated by evaluating E ( θ ) over a relatively small number of trials (in particular, smaller than or comparable to the number of trial conditions) rather than using many trials to obtain the “true” gradient. Intuitively, this improves convergence to a satisfactory solution when the objective function is a highly complicated function of the parameters by stochastically sampling the gradient and thereby escaping saddle points [59] or poor local minima, while still performing an averaged form of gradient descent over many stochastic updates.
Even so, SGD with the objective function given in Eqs 15 and 16 often fails to converge to a solution when the network must learn dependencies between distant time points [60]. To remedy this problem, which is due to some gradient components being too large (exploding gradients) and some gradient components being too small (vanishing gradients), we follow [35] in making two modifications. First, the exploding gradient problem is addressed by simply “clipping” the gradient when its norm exceeds a maximum G: instead of Eq 18 for the direction and size of the update, we use
δθ(i−1)={ −η∇ℰ(i−1)×G|∇ℰ(i−1)|if|∇ℰ(i-1)|>G,−η∇ℰ(i−1)otherwise. (21)
Second, the vanishing gradient problem is addressed by modifying the objective function with the addition of a regularization term:
E = 1 N trials ∑ n = 1 N trials ( L n + λ Ω Ω n ) , (22)
Ω n = ∑ t = 1 N time | ∂ L n ∂ x t ∂ x t ∂ x t - 1 | 2 | ∂ L n ∂ x t | 2 - 1 2 . (23)
In Eq 22 the multiplier λΩ determines the effect of the regularization term Ωn, with no effect for λΩ = 0. In Eq 23, the first term in parentheses is the ratio between the squared norms of two vectors, which we would like to be close to 1. The somewhat opaque (row) vector expression in the numerator can be unpacked as (cf. Eq 7)
∂ L n ∂ x t ∂ x t ∂ x t - 1 j = ∑ k = 1 N ∂ L n ∂ ( x t ) k ∂ ( x t ) k ∂ ( x t - 1 ) j (24) = ( 1 - α ) ∂ L n ∂ x t + α ∂ L n ∂ x t W rec ⊙ ( r t - 1 ′ ) j . (25)
Here each component r′(xt) of r′(xt) is the derivative of the f-I curve, i.e., 1 if x > 0 and 0 otherwise in the case of rectification, and ⊙ denotes element-wise multiplication of two vectors. For consistency in notation we treat r t - 1 ′ here as a row vector. One subtlety in the implementation of this term is that, for computational efficiency, only the “immediate” derivative of Ωn with respect to the network parameters is used, i.e., with xt and ∂ L n / ∂ x t treated as constant [35]. The relevant network parameters in this case are the elements of the trained weight matrix Wrec,plastic, which is related to Wrec through Eq 14.
The role of the regularization term Ωn is to preserve the size of the gradients as errors are backpropagated through time. This is accomplished by preserving the norm of ∂xt/∂xt−1, which propagates errors in time [35], along ∂ L n / ∂ x t, which is the direction in which the change in the objective function is greatest with respect to xt. More intuitively, the impact of the regularization term on network dynamics can be understood by noting that if ∂xt/∂xt′ is small for some t′ < t then, by definition, xt does not depend on small changes in xt′, which may occur when x is close to an attractor. Preserving the norm of ∂xt/∂xt−1 through time therefore encourages the network to remain at the boundaries between basins of attraction and thus encourages longer computation times. For instance, this results in perceptual decision networks that can integrate their inputs for a long period of time, before converging to one of the choice attractors. We note that, although the numerator and denominator in Eq 23 appear, by the chain rule, to preserve the ratio of ∂ L n / ∂ x t - 1 to ∂ L n / ∂ x t, this is only approximately true. Specifically,
∂ L n ∂ x t ∂ x t ∂ x t - 1 = ∂ L n ∂ x t - 1 - ∂ L n , t - 1 ∂ x t - 1 , (26)
because L n , t - 1, the component of L n from time t − 1, depends on xt−1 but not on xt.
Finally, additional regularization terms may be included to change either the dynamics or the connectivity. For instance, there are two ways of obtaining sparse recurrent connectivity. First, we can impose a hard constraint that fixes a chosen subset of weights to be nonzero and modifiable by the optimization algorithm as described above. Second, we may apply a soft constraint by adding the sum of the L1-norms of the weights to the objective function:
E = 1 N trials ∑ n = 1 N trials ( L n + λ Ω Ω n ) + λ 1 rec N 2 ∑ j , k = 1 N | W j k rec | . (27)
In addition, we may choose to encourage solutions with small firing rates through regularization of the L2-norms of the firing rates [8]:
E = 1 N trials ∑ n = 1 N trials ( L n + λ Ω Ω n + λ 2 fr R n fr ) + λ 1 rec N 2 ∑ j , k = 1 N | W j k rec | , (28) R n fr = 1 N N time ∑ j = 1 N ∑ t = 1 N time ( r t ) j 2 (29)
where (rt)j is the firing rate of the j-th unit at time t on each trial. Again, we gain flexibility in defining more complex regularization terms because Theano computes the necessary gradients using BPTT. Although BPTT is simply a specialized chain rule for neural networks, automatic differentiation frees us from implementing new gradients each time the objective function is changed. This greatly facilitates the exploration of soft constraints such as those considered in [8].
To demonstrate the robustness of the training method, we used many of the same parameters to train all tasks (Table 1). In particular, the learning rate η, maximum gradient norm G, and the strength λΩ of the vanishing-gradient regularization term were kept constant for all networks. We also successfully trained networks with values for G and λΩ that were larger than the default values given in Table 1. When one or two parameters were modified to illustrate a particular training procedure, they are noted in the task descriptions. For instance, the number of trials used for each parameter update (gradient batch size) was the same in all networks except for the context-dependent integration task (to account for the large number of conditions) and sequence execution task (because of online training, where the number of trials is one). As a simple safeguard against extreme fine-tuning, we removed all weights below a threshold, wmin, after training. We also note that, unlike in previous work (e.g., [5]), we used the same level of stimulus and noise for both training and testing.
Code for generating the figures in this work are available from https://github.com/xjwanglab/pycog. The distribution includes code for training the networks, running trials, performing analyses, and creating the figures.
In this section we present the results of applying the training framework to well-known experimental paradigms in systems neuroscience: perceptual decision-making [61–63], context-dependent integration [5], multisensory integration [64], parametric working memory [34, 65], and eye-movement sequence generation [66]. In addition to establishing the relative ease of obtaining networks that perform the selected tasks, we show several single-neuron and population analyses associated with each paradigm. These analyses demonstrate that trained networks exhibit many, though not yet all, features observed in recorded neurons, and the study of these networks therefore has the potential to yield insights into biological neural circuits. A summary of the tasks can be found in Table 2.
The tasks presented in this section represent only a small sample of the diversity of tasks used in neuroscience. In addition, we have chosen—in most cases arbitrarily—a simple set of constraints that do not necessarily reflect the full biological reality. Nevertheless, our work provides the foundation for further exploration of the constraints, regularizations, and network architectures required to achieve the greatest correspondence between trained RNNs and biological neural networks.
Many experimental paradigms in neuroscience require subjects to integrate noisy sensory stimuli in order to choose between two actions (Fig 1). Here we present networks trained to perform two variants of perceptual decision-making inspired by the two common variants of the random dot motion discrimination task [61–63]. For both versions, the network has 100 units (80 excitatory and 20 inhibitory) and receives two noisy inputs, one indicating evidence for choice 1 and the other for choice 2, and must decide which is larger. Importantly, the network is not explicitly told to integrate—it is instead only required to “make a decision” following the onset of stimulus by holding a high value in the output corresponding to the higher input, and a low value in the other.
In the variable stimulus-duration version of the task (Fig 2A), stimulus durations are drawn randomly from a truncated exponential distribution (we note that this is often called the “fixed-duration” version because the experimentalist sets the reaction time, in contrast to the “reaction-time” version in which the subject chooses when to respond). This minimizes the network’s ability to anticipate the end of the stimulus and therefore encourages the network to continue integrating information as long as the stimulus is present [63]. In the reaction-time version (Fig 2B), the network must respond soon after the onset of an ongoing stimulus. To control the speed-accuracy tradeoff, the target outputs during training did not require the network to commit to a decision immediately but instead after a short delay [62]; the delay determines the cost incurred for answering early but incorrectly versus correctly but at a later time.
All trials begin with a “fixation” period during which both outputs must maintain a low value, requiring the network to react only to the stimulus. The fixation can be enforced during training in several ways, including a variable fixation period whose duration is drawn from another truncated exponential distribution, or by introducing “catch trials” when no stimuli are presented. For simplicity, here we used a small proportion of catch trials mixed into the training, together with an additional, unambiguous start cue that signals the onset of stimulus.
Networks trained for both versions of the task show comparable performance in their psychometric functions (Fig 2C and 2D), which are the percentage of trials on which the network chose choice 1 as a function of the signed coherence. Coherence is a measure of the difference between evidence for choice 1 and evidence for choice 2, and positive coherence indicates evidence for choice 1 and negative for choice 2. In experiments with monkeys the signs correspond to inside and outside, respectively, the receptive field of the recorded neuron; although we do not show it here, this can be explicitly modeled by combining the present task with the model of “eye position” used in the sequence execution task (below). We emphasize that, unlike in the usual machine learning setting, our objective is not to achieve “perfect” performance. Instead, the networks were trained to an overall performance level of approximately 85% across all nonzero coherences to match the smooth psychometric profiles observed in behaving monkeys. We note that this implies that some networks exhibit a slight bias toward choice 1 or choice 2, as is the case with animal subjects unless care is taken to eliminate the bias through adjustment of the stimuli. Together with the input noise, the recurrent noise enables the network to smoothly interpolate between low-coherence choice 1 and low-coherence choice 2 trials, so that the network chooses choice 1 on approximately half the zero-coherence trials when there is no mean difference between the two inputs. Recurrent noise also forces the network to learn more robust solutions than would be the case without.
For the variable stimulus duration version of the decision-making task, we computed the percentage of correct responses as a function of the stimulus duration for different coherences (Fig 2E), showing that for easy, high-coherence trials the duration of the stimulus period only weakly affects performance [63]. In contrast, for difficult, low-coherence trials the network can improve its performance by integrating for a longer period of time. Fig 2G shows the activity of an example unit (selective for choice 1) across all correct trials, averaged within conditions after aligning to the onset of the stimulus. The activity shows a clear tuning of the unit to different signed coherences.
For the reaction-time version of the task, we defined a threshold for the output (here arbitrarily taken to be 1, slightly less than the target of 1.2 during training) that constituted a “decision.” The time it takes to reach this threshold is called the reaction time, and Fig 2F shows this reaction time as a function of coherence for correct trials, while the inset shows the distribution of reaction times on correct trials. In the case of the reaction-time version of the task, it is interesting to consider the activity of single units aligned to the decision time in each trial, which shows that the firing rate of the unit converges to a similar value for all positive coherences (Fig 2H) [62]. This is a nontrivial observation in both experiment [62] and model, as the decision threshold is only imposed on the outputs and not on the recurrent units themselves.
To illustrate the effect of constraints on connectivity structure—but not on performance—we also trained three networks for the fixed stimulus-duration version of the task shown in Fig 2A. For these networks we did not use a start cue. In the first network, no constraints were imposed on the connection weights except for the absence of self-connections (Fig 3A). The second network was required to satisfy Dale’s principle, with a 4-to-1 ratio of the number of excitatory to inhibitory units, and purely excitatory inputs and outputs (Fig 3B). The third network was similar, but with the additional constraint that the inputs that signal evidence for choice 1 and choice 2 project to distinct groups of recurrent units and decisions are read out from the same group of excitatory units (Fig 3C). The two groups of excitatory units send zero excitatory projections to each other, communicating instead only through the inhibitory units and excitatory units that receive no inputs.
In all three cases, a clear structure could be discerned in the connectivity of the trained network by sorting the units by their selectivity index
d ′ = μ 1 - μ 2 ( σ 1 2 + σ 2 2 ) / 2 , (30)
where μ 1 , σ 1 2 are the mean and variance of the unit’s activity, during the stimulus period, on trials in which the network chose choice 1, and similarly for μ 2 , σ 2 2 for choice 2. For the network without separate excitatory and inhibitory units (Fig 3A), clustering manifests in the form of strong excitation among units with similar d′ and strong inhibition between units with different d′. The learned input weights also excite one population and inhibit the other. In the case of the network with separate excitatory and inhibitory populations (Fig 3B), units with different d′ interact primarily through inhibitory units [67]. Importantly, despite the fact that the recurrent weight matrix was initialized with dense, all-to-all connectivity, the two populations send fewer excitatory projections to each other after training. Similarly, despite the fact that the input weights initially send evidence for both choices to the two populations, after training the two groups receive evidence for different choices. Output weights also became segregated after training. In the third network this structure was imposed from the start, confirming that such a network could learn to perform the task (Fig 3C).
In this section and the next we show networks trained for experimental paradigms in which making a correct decision requires integrating two separate sources of information. We first present a task inspired by the context-dependent integration task of [5], in which a “context” cue indicates that one type of stimulus (the motion or color of the presented dots) should be integrated and the other completely ignored to make the optimal decision.
A network trained for the context-dependent integration task is able to integrate the relevant input while ignoring the irrelevant input. This is reflected in the psychometric functions, the percentage of trials on which the network chose choice 1 as a function of the signed motion and color coherences (Fig 4A). The network contains a total of 150 units, 120 of which are excitatory and 30 inhibitory. The training protocol was very similar to the (fixed-duration) single-stimulus decision-making task except for the presence of two independent stimuli and a set of context inputs that indicate the relevant stimulus. Because of the large number of conditions, we increased the number of trials for each gradient update to 50.
Previously, population responses in the monkey prefrontal cortex were studied by representing them as trajectories in neural state space [5]. This was done by using linear regression to define the four orthogonal, task-related axes of choice, motion, color, and context. The projection of the population responses onto these axes reveals how the different task variables are reflected in the neural activity. Fig 4B shows the results of repeating this analysis [5] with the trained network during the stimulus period. The regression coefficients (Fig 4D) reveal additional relationships between the task variables, which in turn reflect the mixed selectivity of individual units to different task parameters as shown by sorting and averaging trials according to different criteria (Fig 4C).
As a proof of principle, we trained an additional network that could perform the same task but consisted of separate “areas,” with one area receiving inputs and the other sending outputs (Fig 5B), which can be compared to the unstructured connectivity of the original network (Fig 5A). Here each area is conceived of as a cortical area containing a group of inhibitory units that only project locally to excitatory and inhibitory units in the same area. Thus there are no interareal connections originating from inhibitory units. The “sensory” area that receives inputs sends dense, “long-range” excitatory feedforward connections to the “motor” area from which outputs are read out, and receives “sparse” (connection probability 0.2) excitatory feedback projections from the motor area. This example illustrates the promise of using RNNs to explore how large-scale function may arise in the brain.
The multisensory integration task of [64] also presents the animal—rats, in this case—with two sources of information. In contrast to the previous task, however, in the multisensory integration task it is advantageous for the animal to integrate both sources of information when they are available. Specifically, visual flashes and auditory clicks were presented at rates between 9 events/sec and 16 events/sec, and the animal was required to determine whether the inputs were below or above the threshold of 12.5 events/sec. When both visual and auditory inputs were present, they were congruent (presented at the same rate). A network trained for this task is also given one or more congruent inputs, and can improve its performance by combining both inputs when they are available (Fig 6A and 6B). The network contains 150 units, 120 of which are excitatory and 30 inhibitory. A third of the units in the network (both excitatory and inhibitory) received only visual input, another third only auditory input, and the remaining third did not receive any input. Outputs were read out from the entire excitatory population.
The training was again mostly similar to the (fixed-duration) single-stimulus perceptual decision-making task, except for the presence of two congruent inputs on multisensory trials. However, in the present task the network must determine whether the given input is larger or smaller than an arbitrary strength, in contrast to the previous integration tasks where two inputs are compared to each other. As a result, giving the network both positively tuned (increasing function of event rate) and negatively tuned (decreasing function of event rate) inputs [68] greatly improved training. Although gradient-descent training can find a solution when the inputs are purely positively tuned, this results in much longer training times and more idiosyncratic unit activities. This illustrates that, while RNN training methods are powerful, they must be supplemented with knowledge gained from experiments and previous modeling studies. As in experimentally recorded neurons, the units of the network exhibit heterogeneous responses, with some units showing selectivity to choice, others to modality, and still others showing mixed selectivity (Fig 6C).
The context-dependent and multisensory integration tasks represent the two end-cases of when two separate sources of information are available for making a decision. It is of great interest for future inquiry how the same network or set of networks may switch from completely ignoring one input to using both inputs to make the optimal decision depending on the task.
One of the most important—and therefore one of the most widely studied—cognitive functions is working memory, the ability to maintain and manipulate information for several seconds during the planning and execution of a task [69, 70]. Working memory has notably been studied in the context of both oculomotor parametric working memory [71] and vibrotactile frequency discrimination [1, 65], and here we trained a network to perform a task based on the frequency discrimination task. In this task, two temporally separated stimuli, represented by constant inputs whose magnitudes are proportional to the frequency (Fig 7A), are presented and the network must determine which of the two is of higher frequency. This requires the network to remember the frequency of the first input f1 throughout the 3-second delay period in order to compare to the second input f2 at the end of the delay period. The network contains a total of 500 units (400 excitatory, 100 inhibitory), with a connection probability of 0.1 from excitatory units to all other units and 0.5 from inhibitory units to all other units; these connection probabilities are consistent with what is known for local microcircuits in cortex [52, 53]. During training only, the delay was varied by uniformly sampling from the range 2.5–3.5 seconds. As in the multisensory integration task, because the network must compare a single input against itself (rather than comparing two simultaneous inputs to each other), it is helpful for the network to receive both positively tuned and negatively tuned inputs.
The network’s performance on each condition is shown in Fig 7B. Based on the experimental results, we trained the network until the lowest percentage of correct responses in any condition exceeded 85%; for most conditions the performance is much higher [34]. Several different types of behavior are observed in the unit activities. For instance, some units are positively tuned for the frequency f1 during both stimulus periods (Fig 7D, left). Other units are positively tuned for f1 during the first stimulus period but negatively tuned during the second (Fig 7D, right); the switch can occur at various times during the delay. Following [34], we performed a simple linear analysis of the tuning properties of units at different times by fitting the firing rate at each time point to the form r(t) = a0(t) + a1(t)f1. The results are presented in Fig 7C, which shows the correlation of a1 between different time points across the population, and Fig 7E, which shows the percentage of significantly tuned (two-sided p-value <0.05) units at different times. The latter shows trends similar to those observed in monkeys.
An experimental paradigm that is qualitatively very different from the previous examples involves the memorized execution of a sequence of motor movements, and is inspired by the task of [66]. An important difference from a modeling point of view in this case is that, unlike in previous tasks where we interpreted the outputs as representing a decision variable between two choices, here we interpret the network’s two outputs to be the x and y-coordinates corresponding to the monkey’s eye position on the screen. After maintaining fixation on the central dot for 1 second, the task is to execute a sequence of three eye movements and hold for 500 ms each (Fig 8A). For each movement, two targets are presented as inputs to indicate the possible moves in addition to the current dot; although the targets could be presented in a more realistic manner—in a tuning curve-representation, for instance—here we use the simple encoding in which each input corresponds to a potential target location. Throughout the trial, an additional input is given that indicates which sequence, out of a total of 8, is being executed (Fig 8B).
For this task we trained a 200-unit (160 excitatory, 40 inhibitory) network on a trial-by-trial basis, i.e., the network parameters were updated after each trial. This corresponds to setting the gradient minibatch size to 1. Moreover, the network was run “continuously,” without resetting the initial conditions for each trial (Fig 8D). During the intertrial interval (ITI), the network returns its eye position to the central fixation point from its location at the end of the third movement, so that the eye position is in the correct position for the start of the next fixation period. This occurs even though the target outputs given to the network during training did not specify the behavior of the outputs during the ITI, which is interesting for future investigation of such networks’ ability to learn tasks with minimal supervision.
During training, each sequence appeared once in a block of 8 randomly permuted trials. Here we used a time constant of τ = 50 ms to allow faster transitions between dots. For this task only, we used a smaller recurrent noise of σrec = 0.01 because the output values were required to be more precise than in previous tasks, and did not limit readout to excitatory units to allow for negative coordinates. We note that, in the original task of [66] the monkey was also required to infer the sequence it had to execute in a block of trials, but we did not implement this aspect of the task. Instead, the sequence was explicitly indicated by a separate set of inputs.
Because the sequence of movements are organized hierarchically—for instance, the first movement must decide between going left and going right, the next movement must decide between going up and going down, and so forth—we expect a hierarchical trajectory in state space. This is confirmed by performing a principal components analysis and projecting the network’s dynamics onto the first two principal components (PCs) computed across all conditions (Fig 8C).
In this work we have described a framework for gradient descent-based training of excitatory-inhibitory RNNs, and demonstrated the application of this framework to tasks inspired by well-known experimental paradigms in systems neuroscience.
Unlike in machine learning applications, our aim in training RNNs is not simply to maximize the network’s performance, but to train networks so that their performance matches that of behaving animals while both network activity and architecture are as close to biology as possible. We have therefore placed great emphasis on the ability to easily explore different sets of constraints and regularizations, focusing in particular on “hard” constraints informed by biology. The incorporation of separate excitatory and inhibitory populations and the ability to constrain their connectivity is an important step in this direction, and is the main contribution of this work.
The framework described in this work for training RNNs differs from previous studies [5, 8] in several other ways. In this work we use threshold (rectified) linear units for the activation function of the units. Biological neurons rarely operate in the saturated firing-rate regime, and the use of an unbounded nonlinearity obviates the need for regularization terms that prevent units from saturating [8]. Despite the absence of an upper bound, all firing rates nevertheless remained within a reasonable range. We also favor first-order SGD optimization over second-order HF methods. This is partly because of SGD’s widely acknowledged effectiveness in current approaches to machine learning, but also because gradient descent, unlike HF, allows for trial-by-trial learning and may ultimately be more easily related to synaptic learning rules in the brain [39, 40].
Eqs 1–3 are a special case of the more general set of equations for RNNs given in S1 Text, which in turn represent only one of many possible RNN architectures. For instance, machine learning applications typically employ a type of RNN known as Long Short-Term Memory (LSTM), which uses multiplicative gates to facilitate learning of long-term dependencies and currently represents one of the most powerful methods for solving sequence-related problems [72]. For reasons of biological interpretation, in our implementation we only consider generalizations that retain the “traditional” RNN architecture given by Eqs 1–3. These generalizations include additive bias terms in recurrent and output units (corresponding to variable thresholds), different time constants for each unit (e.g., faster inhibitory units), correlated noise [73], and other types of nonlinearities besides simple rectification (e.g., supralinear [74] or saturating f-I curves) for either recurrent units or outputs. We found that biases, though not used for the networks in this work, can improve training in some situations by endowing the network with greater flexibility. The choice of output nonlinearity can be particularly relevant when considering the precise meaning of the outputs, such as whether the outputs are considered a decision variable, probability distribution, or eye position. Probability output models are useful, for instance, when the animal’s confidence about its decision is of interest in addition to its actual decision.
Several works [5, 7, 34] have now demonstrated the value of trained RNNs in revealing circuit mechanisms embedded in large neural populations. In addition to the pioneering work on uncovering a previously unknown selection mechanism for context-dependent integration of sensory inputs in [5], work reported in [7] used trained RNNs to reveal possible dynamical implementations of response criterion modulation in a perceptual detection task under temporal uncertainty. Yet, more recent methods for training networks have not been widely available or easily accessible to the neuroscience community. We have endeavored to change this by providing an easy-to-use but flexible implementation of our framework that facilitates further modifications and extensions. For the tasks featured in this work, the amount of time needed for training was relatively short and largely consistent across different initializations (Fig 9), and could be made even shorter for exploratory training by reducing the network size and noise level. Although further improvements can be made, our results already demonstrate that exploratory network training can be a practical and useful tool for neuroscientists. Moreover, while the present learning rule is not biologically plausible, it is of interest whether the behavioral trajectory of learning can be made similar to that of animals learning the same tasks. In particular, the question of how many trials are needed to learn a given task in model RNNs and animals merits further investigation.
Many interesting and challenging questions remain. Although RNNs of rate units often provide a valuable starting point for investigating both the dynamical and neural computational mechanisms underlying cognitive functions, they will not always be the most appropriate level of description for biological neural circuits. In this work we have not addressed the question of how the firing rate description given by RNN training can be properly mapped to the more realistic case of spiking neurons, and indeed it is not completely clear, at present, how spiking neurons may be directly trained for general tasks using this type of approach. In this work we have only addressed tasks that could be formulated in the language of supervised learning, i.e., the correct outputs were explicitly given for each set of inputs. Combining RNN training with reinforcement learning methods [75–77] will be essential to bringing network training closer to the reward-based manner in which animals are trained. Despite limitations, particularly on the range of tasks that can be learned, progress on training spiking neurons with STDP-type rules and reinforcement learning is promising [78–80], and future work will incorporate such advances. Other physiologically relevant phenomena such as bursting, adaptation, and oscillations are currently not captured by our framework, but can be incorporated in the future; adaptation, for example, can be included in phenomenological form appropriate to a rate model [81, 82].
We have also not addressed what computational advantages are conferred, for example, by the existence of separate excitatory and inhibitory populations, instead taking it as a biological fact that must be included in models of animal cognition. Indeed, although our discussion has focused on the distinction between excitatory and inhibitory neurons, the functional role of inhibitory units may only become apparent when the full diversity of excitatory and inhibitory neuronal morphology and physiology, their layer and type-specific distribution and connectivity [49, 50], and domain-specific (e.g., dendritic versus somatic) targeting of excitatory pyramidal cells by interneurons [23, 25, 27, 51] in the brain are taken into account. Some of these phenomena can already be implemented in the framework by fixing the pattern of connectivity between groups of units (corresponding, for example, to different layers in a cortical column), and future work will explore the implications of such structure on network dynamics.
Finally, although we have performed a few basic analyses of the resulting networks, we have not addressed the detailed mechanisms by which networks accomplish their tasks. In this regard, although state-space analyses of fixed and “slow” points [83] are illuminating they do not yet explain how the network’s connectivity, combined with the nonlinear activation functions, lead to the observed neural trajectories. Discovering general methods for the systematic analysis of trained networks remains one of the most important areas of inquiry if RNNs are to provide useful insights into the operation of biological neural circuits. As a platform for theoretical investigation, trained RNNs offer a unified setting in which diverse cognitive computations and mechanisms can be studied. Our results provide a valuable foundation for tackling this challenge by facilitating the generation of candidate networks to study, and represent a fruitful interaction between modern machine learning and neuroscience.
|
10.1371/journal.pcbi.1000063 | Sequence Similarity Network Reveals Common Ancestry of Multidomain Proteins | We address the problem of homology identification in complex multidomain families with varied domain architectures. The challenge is to distinguish sequence pairs that share common ancestry from pairs that share an inserted domain but are otherwise unrelated. This distinction is essential for accuracy in gene annotation, function prediction, and comparative genomics. There are two major obstacles to multidomain homology identification: lack of a formal definition and lack of curated benchmarks for evaluating the performance of new methods. We offer preliminary solutions to both problems: 1) an extension of the traditional model of homology to include domain insertions; and 2) a manually curated benchmark of well-studied families in mouse and human. We further present Neighborhood Correlation, a novel method that exploits the local structure of the sequence similarity network to identify homologs with great accuracy based on the observation that gene duplication and domain shuffling leave distinct patterns in the sequence similarity network. In a rigorous, empirical comparison using our curated data, Neighborhood Correlation outperforms sequence similarity, alignment length, and domain architecture comparison. Neighborhood Correlation is well suited for automated, genome-scale analyses. It is easy to compute, does not require explicit knowledge of domain architecture, and classifies both single and multidomain homologs with high accuracy. Homolog predictions obtained with our method, as well as our manually curated benchmark and a web-based visualization tool for exploratory analysis of the network neighborhood structure, are available at http://www.neighborhoodcorrelation.org. Our work represents a departure from the prevailing view that the concept of homology cannot be applied to genes that have undergone domain shuffling. In contrast to current approaches that either focus on the homology of individual domains or consider only families with identical domain architectures, we show that homology can be rationally defined for multidomain families with diverse architectures by considering the genomic context of the genes that encode them. Our study demonstrates the utility of mining network structure for evolutionary information, suggesting this is a fertile approach for investigating evolutionary processes in the post-genomic era.
| New genes evolve through the duplication and modification of existing genes. As a result, genes that share common ancestry tend to have similar structure and function. Computational methods that use common ancestry have been extraordinarily successful in inferring function. The practice of discerning evolutionary relationships is stymied, however, by modular sequences made up of two or more domains. When two genes share some domains but not others, it is difficult to distinguish a case of common ancestry from insertion of the same domain into both genes. We present a formal framework to define how multidomain genes are related, and propose a novel method for rapid, robust characterization of evolutionary relationships. In an empirical comparison with the current state of the art, we demonstrate superior performance of our method using a large hand-curated set of sequences known to share common ancestry. The success of our method derives from its unique ability to infer evolutionary history from local topology in the sequence similarity network. This represents a departure from the view that protein family classification must be restricted to families with conserved architecture. By exploiting the structure of the sequence similarity network, our approach surmounts this limitation and opens the door to studies of the role of modularity in protein evolution.
| Accurate identification of homologs, sequences that share common ancestry, is essential for accuracy in function prediction and comparative genomics. Homology identification is integral to the annotation of novel genes [1] and prediction of gene function by various methods, including phylogenetic clustering [2], gene fusion analysis [3],[4], phylogenomic inference [5], and genomic context [6],[7]. Homologous genes are used as markers to identify homologous chromosomal regions for comparative mapping [8],[9], analysis of whole genome duplication [10],[11], phylogenetic footprinting [12], and operon prediction [13]–[15]. Pairwise homology detection is also an integral component of clustering approaches to protein family classification ([1],[16], and work cited therein).
All of these applications exploit one or both of the following properties of homologous sequences: genes that share common ancestry tend (1) to have similar structure and function, and (2) be located in homologous chromosomal regions, making them suitable markers for comparative genomics. Because of their prevalence and importance, it is desirable to incorporate multidomain sequences in such analyses: Multidomain proteins represent 40% of the metazoan proteome, with functional roles in signal transduction, cellular adhesion, tissue repair, and immune response [17]. However, multidomain sequences, especially those with promiscuous domains that occur in many contexts, are frequently excluded from genomic analyses due to lack of a theoretical framework and practical methods for detecting multidomain homologs. In this paper, we extend the traditional definition of homology [18] to multidomain sequences that share a common ancestral gene, providing a formalism suitable for modeling multidomain family evolution, design and validation of multidomain homology identification methods, and incorporation of multidomain sequences in genomic analyses.
The original definition of molecular homology [18] does not capture multidomain evolution. Homology traditionally refers to evolution from a common ancestor by vertical descent (e.g., gene duplication and speciation), but multidomain proteins evolve via both vertical descent and domain insertion. For example, Figure 1 depicts two genes, a and b, which share not only a homologous domain but also a common ancestral gene. In contrast, b and c are a domain-only match, a pair of sequences that share similarity due to insertion of the same domain into both sequences but are otherwise unrelated.
Beta platelet-derived growth factor receptor (PDGFRB) and cGMP-dependent protein kinase 1, beta (PRKG1B), in Figure 2A, are enzymes involved in protein amino acid phosphorylation and provide a concrete example of this situation. Phylogenomic and structural evidence [19]–[22], as well as the promiscuity of the Ig and cNMP-binding domains, supports the common ancestry of this pair (see Methods). They have a statistically significant alignment with an E-value of 2.4e−8 that covers 13% of the average of their lengths. While they share a common domain (Pkinase), the Ig domains are unique to PDGFRB and the cNMP-binding domains are unique to PRKG1B. An example of a domain-only match is shown in Figure 2B. Neural cell adhesion molecule 2 (NCAM2) and PDGFRB share two Ig domains, resulting in a significant alignment, also with an E-value of 2.4e−8, and alignment coverage of 24%. However, the genes that encode them are not homologous and they perform different functions: NCAM2 is involved in cell-cell adhesion with no enzymatic function.
The ability to distinguish multidomain homologs from unrelated pairs that share a domain is essential to genomic analysis. The evolutionary relationship between a and b in Figure 1 supports inferences about genome evolution, organization, and function. The same inferences would not necessarily be justified by the evolutionary relationship between b and c. For example, chromosomal regions enriched with homologous gene pairs are likely to be homologous themselves. In contrast, enrichment with homologous domains does not support the inference that a pair of chromosomal regions is homologous. Heuristics based on similarity and alignment coverage (the fraction of the mean sequence length covered by the optimal local alignment) have been proposed to screen out domain insertions. Recently, approaches based on domain architecture comparison have also been proposed [23]–[26]. To our knowledge, despite the prevalence of methods based on sequence similarity and alignment coverage [27]–[37], the accuracy of these heuristics has never been systematically tested. However, the examples in Figure 2 raise doubt about the general effectiveness of these methods. Both pairs have weak sequence similarity, short alignments, and a similar combination of shared and unique domains. Setting a significance threshold to eliminate NCAM2 would also eliminate roughly 240 sequences that are related to PDGFRB, since more than a quarter of the Kinases that match PDGFRB have E-values less significant than 2.4e−8. Alignment coverage would not help distinguish these two cases: the homologous pair has a shorter alignment than the unrelated pair. Nor could we separate this case by comparing domain content, since PDGFRB and PRKG1B share one domain, while PDGFRB and NCAM2 share two. For this example, sequence similarity, the length of the shared region, and domain architecture comparison all fail to distinguish the homologous pair from the domain-only match.
To determine the extent of this problem, here we evaluate sequence similarity, alignment coverage, and domain architecture comparison on a hand-curated benchmark of 853,465 known homologous pairs. Our results show that these heuristics are all insufficient for consistent, reliable identification of multidomain homologs. Surprisingly, given its widespread use, even a modest alignment coverage requirement dramatically increased the number of mis-assigned homologs in our study. These results challenge two unstated, but widely accepted hypotheses: (1) homologous sequences share similarity along the bulk of their length and (2) the local alignment between homologous sequences usually covers a greater fraction of their mean length than the local alignments of sequences that only share a domain.
These observations suggest to us that sequences alone may not consistently contain enough information to differentiate homology from domain-only matches. We introduce a novel method, called Neighborhood Correlation, that leverages additional information contained in the weighted sequence similarity network to distinguish homologs from domain-only matches. In this network, each vertex corresponds to a sequence. Vertices whose corresponding sequences have significant similarity are connected by an edge with weight proportional to that similarity. The neighborhood of a sequence is the set of vertices adjacent to it; that is, the set of all sequences that match it above a predefined significance threshold. (In this work, “sequence neighborhood” refers to the local context of the sequence in the network and not to the region immediately surrounding it in the genome.) Our analysis demonstrates that the neighborhood structure of gene pairs related through shared domain insertions is characteristically different from that of pairs related through duplication or speciation. These differences in neighborhood organization are detectable and can be exploited to distinguish homology from domain sharing.
A homology detection method for genomic analysis must meet the following criteria: It should correctly predict homologous pairs and reject unrelated pairs, including those that share domains. With a single set of parameter values, it should perform reliably on sequences with a broad range of attributes, including single domain families, multidomain families, families with short regions of conservation, and families with weak sequence homology. Finally, it should be easy to use and fast enough for datasets comprising hundreds of millions of sequence pairs.
In an empirical evaluation, we demonstrate that Neighborhood Correlation meets these criteria. It is highly effective in classifying multidomain homologs and achieves superior performance in comparisons with sequence similarity (BLAST and PSI-BLAST), alignment coverage, and domain architecture comparison. To evaluate performance, we hand-curated a benchmark of 853,465 known homologous pairs of mouse and human sequences, drawn from twenty well-studied families. Our test set includes single-domain families, as well as multidomain families with promiscuous domains that are at risk for domain-only matches. Although comprehensive datasets are available for testing methods for predicting homology of individual domains [38],[39], we are unaware of any other gold-standard dataset of known multidomain families with variable domain architectures. We offer this validation dataset, which is based on published evidence by experts on each of the families, as a resource for future studies.
As a validation of our approach, we applied Neighborhood Correlation to all complete, mouse and human sequences in SwissProt 50.9 to predict homologs. A comparison of our predictions with the euKaryotic Clusters of Orthologous Groups (KOGs) database [40] showed that the set of protein sequences with highly correlated neighborhoods includes the vast majority of pairs that share an orthologous group (i.e., have the same KOG annotation). This is consistent with the fact that orthology is a more restrictive criterion than homology. We also show that most pairs in our set of predictions share at least one domain, according to the Pfam database [41], but many sequence pairs that share a domain are excluded. This is consistent with our goal of identifying gene homology rather than domain homology.
Homology has traditionally been defined in terms of families that evolve by vertical descent [18],[42]; that is, by speciation and gene duplication. However, multidomain sequences evolve by speciation, gene duplication, and acquistion of domains from outside the family [43] (Figure 1). The traditional definition of homology does not apply in this case, as previous authors have pointed out [42],[44]. In the words of Walter Fitch [42], “We must recognize that not all parts of a gene have the same history and thus, in such cases, that the gene is not the unit to which the terms orthology, paralogy, et cetera, apply.” It has been proposed that sub-genic sequence fragments should be the units of interest [44],[45]. However, there are many applications, such as ortholog detection, comparative mapping, and phylogenetic footprinting, for which it is essential to work with a definition of homology where the gene is the basic unit. Moreover, in order to study the evolution of multidomain gene families, it is necessary to focus on genes. The gene is the unit of selection. While domains confer modular function on genes, ultimately it is the functionality of those genes drives their retention.
Here, we propose a model of multidomain homology based on vertical descent and insertion of a sequence fragment into an existing gene. In our model, two sequences are homologous if they are encoded by genes that share an ancestral locus. The rationale for this definition is illustrated in Figure 3, which shows the evolution of genes through vertical descent and domain insertion in the context of the chromosomes in which they reside. When genomic context is taken into account, it is clear that genes g2 and g2′ are homologous, despite the fact that g2 contains a domain not present in g2′ and vice-versa. In contrast, genes g2 and g3′ are not homologous, despite the fact that they share a homologous domain, since g2 and g3′ are not located in chromosomal regions that share common ancestry. For comparative mapping applications, where homologous genes are used as markers for identifying chromosomal regions, this distinction is crucial. For example, phylogenetic footprinting [12] predicts transcription factor binding sites by identifying homologous genes and then searching their flanking chromosomal regions for conserved sequence motifs. In Figure 3, the regions upstream of g2 and g2′ have an elevated probability of sharing conserved motifs since they share common ancestry. However, there is no reason to expect an enrichment of motifs shared between the flanking regions of g2 and g3′.
Our model is applicable to families that evolved through acquisition of a new domain by an existing gene. This can occur through insertion of sequence fragments into the gene or by recruitment of adjacent exons. Formation of a new gene architecture by domain loss is also consistent with our model. Several lines of evidence suggest that acquisition of an auxiliary domain by an existing gene is a relatively common mode of domain shuffling. First, a substantial number of metazoan, chordate, and vertebrate families have been identified that evolved through a pattern of duplication, insertion of domains, and further duplication, a pattern consistent with this model [46],[47]. Second, the existence of promiscuous domains that lend themselves to insertion in new chromosomal environments [48],[49] supports an insertion model. Third, domain insertion is more likely to be successful when a domain is inserted into an existing functional environment, e.g., into the intron of an existing gene. In this case, all regulatory and termination signals required for successful transcription are already present. A fourth line of evidence stems from analyses of the flanking DNA of genes that arose very recently, where traces of the particular domain shuffling mechanism that occurred can still be observed. A number of recently evolved metazoan genes have been discovered that arose through duplication of an existing gene, followed by acquisition of one or more domains by unequal crossing over or by retrotransposition [50]–[54]. Finally, a number of studies have inferred relative rates of various domain shuffling events by applying parsimony models to abstract domain architectures. Their results suggest that the most common domain shuffling scenario involves insertion or deletion of a single domain into an existing multidomain architecture [24],[55],[56].
Our model is not applicable to the case where a new domain architecture is assembled de novo from several unrelated building blocks and subsequently acquires a regulatory region. We consider such a novel architecture to be the progenitor of a new family, since it is not clear that the ancestry of any one constituent is preferred. Similarly, our model does not capture formation of new architectures through fragmentation of more complex ones. However, recent evidence suggests that both of these scenarios occur rarely [24],[55],[57].
Homology detection is the problem of distinguishing between sequence pairs with different types of evolutionary histories: evolution via gene duplication or via domain insertion. Sequence similarity, alignment coverage, and domain architecture comparison have all been considered for this purpose. However, none of these distinguish the homologous pair from the domain-only match given in Figure 2. The empirical results in the following sections confirm that this is not an isolated example. Accurate classification of multidomain homologs requires additional information from another source.
The structure of the sequence similarity network provides a basis for distinguishing pairs related through vertical descent from other pairs. The local network neighborhoods of homologs and domain-only matches differ in both topology and edge weights. In particular, for homologous pairs, the shared neighborhood (i.e., the set of vertices adjacent to both members of the pair) tends to have more vertices and stronger edge weights than their unique neighborhoods (i.e., vertices adjacent to one pair but not the other). This is not true for domain-only matches. We express this distinction quantitatively by the Neighborhood Correlation score of two sequences, defined to be the correlation coefficient of their respective neighborhoods:(1)where S(x,i) is the normalized bit score [58] of the optimal local alignment of query sequence x and database sequence i, N is the number of sequences in the database, and is the mean of S(x,i) over all sequences (see Methods). Note that NC(x,y) increases with the number, weight, and correlation of edges in the shared neighborhoods of x and y and decreases with the number and weight of edges in their unique neighborhoods.
The Neighborhood Correlation score captures properties of the sequence similarity network that are strongly influenced by the evolutionary processes of interest. The number of edges in the shared and unique neighborhoods is influenced by the rates of gene duplication and domain insertion, while edge weights depend on sequence divergence. Immediately following a gene duplication, the two resulting paralogs have identical neighborhoods. The Neighborhood Correlation score of this new pair is initially one and decreases as the sequences diverge. Additional gene duplications in the same family further increase the size of the shared neighborhood and, hence, the Neighborhood Correlation score. In contrast, if a domain is inserted into a single member of the pair, the number of edges in its unique neighborhood increases and the Neighborhood Correlation score decreases. The increase in the number of unique edges is directly related to the promiscuity of the inserted domain, while the weights of these new edges are proportional to the degree of sequence conservation in the domain superfamily. In practice, the impact of insertion of a domain into a single member on the Neighborhood Correlation score is typically small because promiscuity and sequence conservation within domain superfamilies are inversely related. For example, Pfam domains exhibit a highly significant, negative correlation between domain promiscuity (see Methods) and sequence identity (ρ = −0.21, p = 2.08e−30, Spearman test). This can be understood by observing that when a domain is inserted into a new context, it is likely to experience new selective pressures leading to rapid mutational change.
To see how these principles play out in practice, we consider the neighborhoods of PDGFRB, PRKG1B, and NCAM2 in the sequence similarity network derived from our test dataset (Figures 2 and 4). Although the homologous pair, PDGFRB and PRKG1B, and the domain sharers, PDGFRB and NCAM2, have pairwise alignments with similar properties (E-value, alignment length, number of shared domains), their neighborhoods in the weighted sequence similarity network are very different. The shared neighborhood of the Kinase homologs PDGFRB and PRKG1B is substantially larger (779 sequences) than their unique neighborhoods (183 and 142 sequences, respectively). The shared neighborhood consists almost entirely of Kinases. The unique neighborhoods are dominated by domain-only matches, due to Ig in the case of PDGFRB and the cNMP-binding domain in the case of PRKG1B. Sequence similarities within these unique neighborhoods are weak; the Pfam models for the Ig and cNMP-binding domains have average sequence identities of 20% and 18%, respectively. Thus, the edge weights (not shown) in the shared neighborhood are strong and well correlated, while the edge weights in the unique neighborhoods are weak, yielding a Neighborhood Correlation score of NC = 0.65.
Conversely, PDGFRB and NCAM2 are related through domain insertion and have significant sequence similarity due to a shared Ig domain. Their shared neighborhood is relatively small (242 sequences) and comprised primarily of Ig-based matches. These contribute little to the Neighborhood Correlation score of this pair due to low sequence conservation within the Ig superfamily. In contrast, the unique neighborhood of PDGFRB is large (630 sequences), with strong edge weights. For these reasons, PDGFRB and NCAM2 have a Neighborhood Correlation score of 0.29, distinctly smaller than the score for PDGFRB and PRKG1B. Unlike sequence comparison, this clear difference in neighborhood structure can be used to recognize multidomain homology.
Evaluation of classification performance requires a trusted set of positive examples (known homologous pairs) and negative examples (pairs known not to share common ancestry). Although benchmarks are available for detection of remote homology (e.g., SCOP [38], CATH [39]), functional similarity (e.g., the Gene Ontology (GO) [59]), orthology (e.g, COGs [40]), and structural genomics ([16],45,[60], and work cited therein), we are unaware of any gold-standard validation dataset for multidomain homology. Our benchmark is designed to be suitable for testing two classification goals: good overall performance on a large set of sequence pairs and consistent performance on individual families with varying properties. To satisfy these needs, we constructed a test set of 1577 sequences from 20 families of known evolutionary origin (Table 1). The families encompass a broad range of functional categories, summarized in Table 2. The full curation procedure is described in Methods and Text S1.
For each family, we identified two sets of sequence pairs: family (FF) pairs, where both members of the pair are in the family, and non-family (FO) pairs, where only one of the two sequences is in the family. Given a family of size k, we obtain k2 FF pairs (the positive examples) and k(N−k) FO pairs (the negative examples). Individual families, which cover a range of functional properties and domain architecture complexity, can be used for family specific tests. In addition, we constructed a test set (ALL) for general performance evaluation by merging all sets of FF and, respectively, FO pairs, yielding 853,465 positive and 40,459,204 negative examples. Performance measurements obtained with this set could be biased by the Kinase family, which is much larger than the other families. We therefore also considered the set of all sequences excluding the Kinases (ALL-Kin), resulting in 32,629 positive and 17,545,558 negative examples.
Our goal is a method that can correctly identify homologs in multidomain families without degrading performance in other types of families. We therefore devised a benchmark to test a range of homology detection challenges, involving single domain as well as multidomain families. Families with complex and varied domain architectures represent the primary challenge undertaken in this study. Such families result from duplication, domain accretion, and further duplication. Some of these families are defined by a single domain that is unique to the family (e.g., Kinase), while others are characterized by a particular combination of domains (e.g., ADAM) or by a conserved set of domains with variations in domain copy number (e.g. Laminin). Modularity in both single and multidomain families can also arise through the presence of sequence motifs, such as subcellular localization signals, transactivation sequences (e.g., Tbox), and functional components that confer substrate specificity (e.g. USP). These motifs can result in matches to unrelated sequences. In addition, promiscuous domains challenge homology identification because they can result in significant sequence similarity but carry little information about gene homology. Promiscuity can confound reliable detection of homologs even in families with conserved domain architectures.
Remote homology detection is a serious challenge that has received widespread attention. In our dataset, this challenge is represented set by FGF, TNF, TNFR, and USP, families that exhibit low sequence conservation. Finally, we considered homologous pairs with short conserved regions. A minimum alignment coverage criterion is frequently imposed to eliminate domain-only matches, reflecting a widely held, but untested belief that homologous pairs have regions of similarity that cover a substantial fraction of their length. To test the robustness of homology detection methods with respect to alignment length, we included single domain families with short conserved regions such as the Tbox family.
Our selection of test families was limited to those for which it was possible to obtain evidence concerning their evolutionary history. Evolutionary evidence was obtained from published articles and/or curation by a nomenclature committee. In the best cases, direct syntenic evidence of vertical descent can be found. In other cases, indirect evidence such as conserved intron/exon structure is used. Phylogenetic evidence can confirm vertical descent, for example, if all domains in a family have consistent phylogenies. However, phylogenetic disagreement between core and auxiliary domains does not rule out homology according to our model. For each, the evidence used is described in Text S1.
We evaluated Neighborhood Correlation using our benchmark, and compared its performance with other methods currently in use. We considered performance on multidomain homology identification, as well as overall performance on diverse, heterogeneous datasets. We also used Neighborhood Correlation to predict novel homologous relationships.
We compared the performance of Neighborhood Correlation with BLAST [61], alignment coverage [27], and PSI-BLAST [58], methods commonly used for assessing homology, as well as Domain Architecture Comparison (DAC), a recently introduced approach that compares sequences by considering their constituent domains [23]–[26],[55].
BLAST gives a measure of sequence similarity based on the optimal local alignment between two sequences. BLAST does not capture gene structure (e.g., domain organization), nor does it reflect additional information that might be derived from suboptimal local alignments. BLAST is widely used, its behavior is well understood, and its scores are easily compared with those from other studies. A great deal of attention has been devoted to tuning BLAST performance and to developing accurate statistical tests. It represents an attractive balance between rigor and speed.
A significant BLAST score is evidence of similarity greater than that expected by chance, but cannot distinguish whether that similarity stems from vertical descent or domain insertion. In order to eliminate domain-only matches, many analyses combine sequence similarity with alignment coverage to identify homologs [28]–[37]. To be considered homologous, sequence pairs must then satisfy a second criterion in addition to significant sequence similarity: the fraction of the sequence length covered by the optimal local alignment must meet a pre-specified threshold. To our knowledge, alignment coverage criteria have never been empirically evaluated. In this work, we demonstrate that such a requirement is highly detrimental to performance overall, and in nearly all tested families.
In the presence of high sequence divergence, BLAST is limited by the amount of information that can be derived from pairwise comparison. To address this problem, approaches based on multiple sequence alignments (MSAs) have been used to increase sensitivity. PSI-BLAST, one of the most widely used examples of this approach, constructs a Position Specific Scoring Matrix (PSSM) through iterative search and has been shown to dramatically improve sensitivity [62]. MSA-based methods are designed to detect remote homology, not multidomain homology. Since sequences with different architectures cannot be aligned, MSA-based methods are not a natural choice for multidomain homology detection. We included PSI-BLAST in our study because it is widely used as a standard for remote homology detection.
In addition to sequence based methods, we considered direct comparison of domain architectures for multidomain homology detection. Each sequence was represented by a linear sequence of Pfam domains. Linker sequences between domains were ignored, as was sequence variation between instances of a given Pfam domain family. The resulting domain architectures were compared based on their domain composition. In a previous study, we proposed and evaluated 21 different methods for comparing domain architectures [23]. These methods considered properties such as the number of shared domains, domain copy number, total number of domains in a protein, domain order, and domain promiscuity. We included the domain architecture comparison strategy that exhibited the best performance from that study in our current study. This method assigns a score to each pair based on the number of shared domains (see Methods), following the rationale that homologous pairs will have more domains in common than pairs related through domain insertion. In assessing similarity, each domain is assigned a weight inversely proportional to its promiscuity. This reflects the assumption that rare domains convey more information about homology than promiscuous domains.
The performance of each method was assessed via the ROC-n score (Table 3), which represents both false positives and false negatives (see Methods). ROC-n is the area under the Receiver Operating Characteristic (ROC) curve comprised of the top ranking pairs up to the first n false positives. We used n = 100k, where k is family size, corresponding to 100 false positives per query.
In evaluating homology identification methods, we consider two user models. Genome-scale analyses require all-against-all comparison of a large and heterogeneous set of sequences. In order to be suitable for automated, genomic analyses, a method must be robust enough for use without human intervention, deliver consistent behavior on different types of domain architectures, and be fast and easy to use. In this case, the goal is to maximize the total number of homolog pairs that are correctly predicted. A second application is analysis of individual families, where the goal is to obtain good per-family prediction scores over a wide range of families.
To evaluate performance for both user models, we report ROC-100k scores for all pairs (ALL and ALL-Kin), as well as ROC-100k scores for each family. To show how the methods tested behave on proteins with various attributes, we also report the average ROC-100k score per family for single domain families, multidomain families with conserved architectures, and multidomain families with variable architectures.
As a visualization tool, we generated rank plots, which show the scores of all matches to a given query sequence in rank order. Rank plots provide a visual representation of the organizational structure of the network neighborhood of the query sequence, as well as organizational substructure within the family. For example, Figure 5 shows a rank plot for the query sequence PDGFRB, a protein tyrosine kinase. The break in the curve in Figure 5B at NC≈0.8 corresponds to the first match to a Serine/Threonine Kinase, the inflection point at NC≈0.75 corresponds to the first match to a Dual-Specificity Kinase, and the downward plunge at NC≈0.59 corresponds to the first Casein Kinase. Rank plots for each of the 26,197 sequences in our dataset are provided at http://www.neighborhoodcorrelation.org.
When all considered classifiers are applied to the aggregate set of sequence pairs (ALL), Neighborhood Correlation dramatically outperforms the other three methods (Table 3, Figures S1 and S2). In the ALL-Kin dataset, Neighborhood Correlation yields better performance than BLAST and PSI-BLAST, but performs slightly worse than DAC. The superior performance of Neighborhood Correlation on the ALL and ALL-Kin datasets demonstrates that its optimal classification threshold is less sensitive to family specific properties than those of BLAST or PSI-BLAST.
When performance on individual families is considered, Neighborhood Correlation is generally more robust than the other three methods. It perfectly classifies twelve families, more than any other method. In addition, in 16 of 20 families, the discriminatory performance of Neighborhood Correlation is better than or equal to that of all other methods. In particular, Neighborhood Correlation obtains the highest average score for both conserved and variable architectures and performs much better on individual multidomain families except for Myosin and Kinesin. For families with high sequence divergence, including FGF, TNF, and USP, Neighborhood Correlation performs better than BLAST, indicating that neighborhood structure can compensate for a low signal to noise ratio in pairwise comparisons of remote homologs. PSI-BLAST also performs well in such cases.
To demonstrate why Neighborhood Correlation is more effective for complex families, we consider its performance on the Kinase family. Figure 5 shows a rank plot of the results of a query with the Kinase PDGFRB. A robust method is expected to rank all Kinase family members before non-Kinase matches. In particular, we examine pairing between the Kinase PRKG1B and the non-Kinase NCAM2, the genes depicted in Figure 2. Neighborhood Correlation exhibits no difficulty separating these pairs. The match with PRKG1B scores substantially higher than NCAM2 (indicated by magenta and green circles, repectively, in Figure 5). In contrast, the BLAST scores for these sequences are indistinguishable, and the PSI-BLAST scores for these sequences are reversed: The match to NCAM2 obtains θ = 3.65e−40, while the match to PRKG1B is much less significant (θ = 1.26e−25). How typical are these examples? As shown in Figure 6, the sequence similarity distributions of FF and FO pairs overlap completely for BLAST and partially for PSI-BLAST. In contrast, the Neighborhood Correlation score distributions for family and non-family matches are largely distinct, with only a limited overlap in the tails of the distributions.
Neighborhood Correlation also delivers robust performance when sensitivity (Sn) and specificity (Sp) are considered independently. For example, when matches to the query sequence PDGFRB are ranked by Neighborhood Correlation score (Figure 5A), a cutoff of NC = 0.3 results in three false positives with only ten false negatives. In contrast, a BLAST threshold of E<3e−10 results in three false positives and 630 false negatives (Figure 5B). The number of false negatives obtained with PSI-BLAST at this specificity is even greater (Figure 5C). More generally, the ROC-n curves for the Kinase family in Figure 7 demonstrate that Neighborhood Correlation achieves both higher sensitivity and higher specificity than BLAST, except at very high specificity, and always outperforms PSI-BLAST by both measures. Neighborhood Correlation simultaneously achieves Sn≈0.85 and Sp≥0.999. At this specificity, Sn≈0.7 for PSI-BLAST and Sn≈0.55 for BLAST.
While the other methods considered have strengths specific to particular challenges, Neighborhood Correlation delivers the most reliable and consistent performance on large, heterogeneous datasets. Neighborhood Correlation is, therefore, particularly well suited to automated genome-scale analyses, which require that a single classification threshold be suitable for the vast majority of sequence pairs in a genomic dataset. Moreover, Neighborhood Correlation is robust. The distribution of Neighborhood Correlation scores for all sequence pairs in our dataset (Figure S3) has a flat trough ranging from 0.4 to 0.8. Within this range, the prediction quality will be relatively insensitive to the choice of threshold. A putative set of mouse and human homologs imposed by a threshold of NC≥0.6 on all sequence pairs in our dataset is available at http://www.neighborhoodcorrelation.org.
As expected, PSI-BLAST excels at families with low sequence conservation, such as TNF and USP, and generally performs well on single domain families. However, PSI-BLAST falters on complex multidomain families and on sequences with promiscuous domains. PSI-BLAST's average ROC-100k scores for both conserved and variable multidomain families are inferior to those of both Neighborhood Correlation and BLAST. This is exemplified by PSI-BLAST's poor performance (Figure 5B) when querying with PDGFRB, which has two copies of the highly promiscuous Ig domain. PSI-BLAST's iterative profile construction algorithm incorporates matches to the highly promiscuous Ig domain in the growing alignment, even when a very stringent inclusion threshold (E<10−13) is used. As a result, unrelated sequences that contain Ig domains match the resulting profile with better scores than Kinases without Ig. PSI-BLAST performs better on the Kinase family as a whole than it does on PDGFRB (Table 3) because many Kinases are single domain proteins.
When classification of heterogeneous data is considered, PSI-BLAST's performance is inferior to Neighborhood Correlation on the ALL dataset and to both Neighborhood Correlation and BLAST on the ALL-Kin dataset. This demonstrates that no single PSI-BLAST cutoff is suitable for all families. Indeed, inspection of PSI-BLAST output on individual queries (data not shown) indicates that PSI-BLAST scores tend to vary widely from family to family. PSI-BLAST introduces a clear tradeoff between sensitivity and generality, to the particular detriment of large-scale studies. Moreover, PSI-BLAST is characterized by greater instability and running time than BLAST or Neighborhood Correlation.
Domain architecture comparison performs well on single domain families and on multidomain families with conserved domain architectures (e.g., DVL, Notch, Laminin, and WNT). Like PSI-BLAST, DAC can recognize distant homology because domain architectures are recognized by MSA-based models. The performance of DAC on other families is mixed, however, because it faces a number of challenges that do not arise with the other classification methods.
First, all domain architecture comparison methods are substantially restricted by the limitations of domain detection. In our dataset, 12.7% of sequences do not have domain annotations, resulting in low ROC-100k scores for many families. This explains why single domain families, such as Tbox, which have identical domain architectures, do not achieve perfect ROC-100k scores, contrary to expectations. An additional shortcoming is that domain architecture comparison methods do not capture information in linker sequences or sequence variation within a domain family. Therefore, domain architecture comparison tends to assign the same score to pairs that actually differ in sequence divergence. This explains the long plateaus in the ROC curve for DAC in Figure 7.
A particularly challenging problem for domain architecture comparison is how to effectively distinguish domains that proliferated through gene duplication from promiscuous domains that proliferated through domain shuffling. The number of domain partners, used here, is a typical measure of promiscuity, based on the assumption that this measure reflects the frequency of domain insertion [48]. This measure of promiscuity will inappropriately down-weight a domain that characterizes a family, if the domain happens to be the target of insertions of many other domains. Consider, for example, a sequence with a single domain A that sustains repeated duplication, followed by insertion of different domains into the resulting copies, yielding AB, AC, AD, and so on. Domain A will have a high promiscuity score, although it is never inserted into new contexts. As a concrete example, the Pkinase domain partners with more than 100 different domains. However, the resulting high promiscuity score may be inappropriate since Pkinase lacks many of the other characteristics of promiscuous domains, such as small size and 1-1 phase [17], and is important in defining the Kinase family. This explains why domain architecture comparison performs poorly on the Kinase family.
To assess the effectiveness of alignment coverage in eliminating domain-only matches, we compared ROC-100k scores for sequence similarity alone and combined with alignment coverage (α, see Methods). We considered three alignment coverage thresholds, α≥0.3, α≥0.6, and α≥0.8, that span the range of length cutoffs used in the literature (e.g. [32],[34]). The results (Table 4) show that the addition of an alignment coverage criterion does not improve the performance of sequence similarity. For example, a cutoff of α≥0.3 reduces the ROC-100k score by 25% in the ALL dataset and 23% in the ALL-Kin dataset. When families are considered individually, a cutoff of α≥0.3 decreases the ROC-100k score by at least 10% in one-third of the families. Increasing the cutoff to α≥0.6 or α≥0.8 does not increase performance in any family. Note that although the ROC-100k score for KIR when α≥0.6 is higher than the score for sequence similarity alone, this difference is not significant (p = 0.69).
Alignment coverage is based on the assumption that non-homologous pairs have shorter regions of similarity than homologous pairs, yet Table 4 suggests this is not universally true. To assess the extent to which the region of similarity in homologous pairs extends over the bulk of their length, we calculated Precision and Recall (see Methods) for α≥0.3, α≥0.6, and α≥0.8. The results, shown in Tables 5 and Table S1, suggest that full length alignments are not a characteristic property of homologous families, at least in our dataset. In the ALL-Kin dataset, a cutoff of α≥0.3 eliminates 40% of true positives, specifying α≥0.6 eliminates 70% of true positives, and α≥0.8 eliminates 83% true positives. The loss in Recall is even more extreme in the ALL dataset.
To better understand these results, we plotted histograms of α for individual families (Figures 8, S4). While some families do have long regions of similarity, long conserved regions are not a persistent characteristic of most families in our dataset. Several different trends in domain organization can cause this. Some families are characterized by a short, conserved domain, such as the DNA binding domain in the FOX family, and little conservation elsewhere (Figure 8A). Multidomain families exhibit a range of alignment lengths for a variety of reasons. In families characterized by a single defining domain partnered with a variety of auxiliary domains, alignment lengths depend upon the number of domains a given pair has in common. For example, the histogram for the PDE family (Figure 8B) has a small peak near α = 1.0, corresponding to pairs with identical domain architectures, and a much larger peak between α = 0.2 and α = 0.7 that represents pairs of family members with different auxiliary domains. Families can also demonstrate wide variation in due to differences in copy number (e.g., Laminin, Figure 8C). Finally, a broad α distribution can be caused by variation in sequence length within the family. Even when the length of the conserved region is constant, alignment coverage, expressed as a fraction of total length, may vary widely, confounding homology prediction methods based upon alignment coverage.
Given the widespread use of alignment coverage criteria, we were surprised by this poor performance. We examined the possibility that our failure to observe a consistent pattern of long alignments was due to the fact that we considered the length of the optimal alignment, only. To investigate whether including sub-optimal alignments would result in different conclusions, we implemented a simple heuristic (see Methods) that identifies and combines a consistent set of high-scoring local alignments; i.e., alignments that appear in the same order in both sequences and do not overlap. Surprisingly, including suboptimal alignments in the alignment coverage calculation has little impact on our results. The distributions of the combined alignment lengths, shown in turquoise and brown in Figures 8 and S4, differ little from the distribution of optimal alignment length distributions (shown in blue and red). Nor do the values of Precision and Recall obtained with combined alignments differ greatly from those obtained with the optimal alignment (see Table 5 and Table S2). In summary, analysis with combined alignments confirms that full length similarity is not a general characteristic of homologous families.
Protein modularity allows the evolution of diverse function through combinatorial rearrangement of functional building blocks. This versatile evolutionary mechanism played a transformative role in key evolutionary transitions, including the emergence of multicellular animals and the vertebrate immune system. Identification of multidomain homologs is essential to studying the evolution of modular families, as well as to many genomic applications that exploit evolutionary information.
Two obstacles have impeded research on multidomain homology: the absence of formal models and a lack of curated datasets of multidomain homologs for evaluation of proposed methods. In the current paper, we offer preliminary solutions to both problems: We propose an evolutionary model and an associated definition of homology suitable for multidomain proteins. We further provide a curated test set of homologous mouse and human sequence pairs from twenty well-studied families for which there is unambiguous evidence that member sequences are derived from a common ancestor. Our benchmark encompasses various challenges for homology identification methods, including both conserved and variable multidomain architectures, promiscuous domains, single domain families with short regions of conservation, and families with weak sequence conservation. It differs from other available benchmarks in that it seeks to represent evolutionary, rather than structural (e.g., SCOP [38]) or functional (e.g., GO [59]) information. This benchmark is available to the community through the Neighborhood Correlation website.
Using our curated benchmark, we demonstrate that the most widely used homology identification methods, BLAST, PSI-BLAST, domain architecture comparison, and alignment coverage, all face serious limitations in their ability to recognize multidomain homologs. In response, we introduce Neighborhood Correlation, a method that uses a fundamentally different approach to homology identification by deriving evolutionary signal from the local structure of the sequence similarity network. Following a discussion of our model within the historical framework of models of homology, we place our results in the perspective of similar problems and approaches. We discuss Neighborhood Correlation in relation to other evolutionary classifications, the needs of genomic applications and multiple sequence alignment methods, and conclude by reviewing the potential of networks in molecular evolution.
Although models of gene family evolution have been proposed and debated for more than three decades [18], models of multidomain evolution are in their infancy. Gene homology is a yes/no question: genes either share common ancestry or they do not. With this in mind, Fitch [42] argued that when subsequences of genes have distinct evolutionary histories, it is not possible to determine gene homology. Rost and colleagues [45],[63] further proposed that “dissecting proteins into structural domain-like fragments” [45] is the only reasonable way to study relationships in such proteins. We suggest an alternative: By considering the genomic context of genes that encode multidomain proteins, it is possible to define homology for multidomain sequences without violating the tenet that homology is an indivisible property.
We propose a model of multidomain evolution in which the set of events by which sequences diverge is expanded to include domain insertion and deletion as well as mutation. Recent evidence from studies of young genes [50]–[53], as well as indirect evidence of sequence shuffling [17],[24],[49],[55],[56], suggests that our model is consistent with a significant fraction of metazoan multidomain families. This model permits discrimination between genes related by vertical descent and those related by domain insertion alone, which is the basis for our definition of multidomain homology. This in turn enlarges the scope of inquiry from domain family homology to gene family homology, providing a broader context in which to study the evolutionary processes by which modular families are formed. Our model does not describe families that evolved through other domain shuffling processes such as gene fission, the fusion of adjacent genes resulting from read-through errors, or de novo formation of novel architectures through independent insertions in intergenic regions. Extending the model to capture a broader range of domain shuffling scenarios and testing it on other datasets and applications are important directions for future work.
Evidence supporting the validity of our model can be obtained by comparing Neighborhood Correlation with related classifications, such as orthology and domain homology. The success of Neighborhood Correlation in recapitulating homologous relationships in our benchmark empirically supports Neighborhood Correlation as a predictor of homologous genes; that is, sequences derived from a common ancestor by vertical descent, whether by duplication or speciation. Since orthologs, sequences that diverged by speciation in their most recent common ancestor, are by definition homologs, our model predicts that known mouse and human orthologs will have high Neighborhood Correlation scores. To test this prediction, we compared Neighborhood Correlation with KOGs [40]. As expected, 90% of sequences in our dataset with the same KOG annotation have a Neighborhood Correlation Score greater than 0.6 (Figure 9A). However, only 12% of pairs with NC≥0.6 share the same KOG annotation. This is consistent with the observation that gene homology is a necessary but not sufficient condition to establish orthology.
Domain homology, on the other hand, is a less stringent criterion than gene homology. Homologous genes, by definition, share at least one homologous domain. Of pairs with Neighborhood Correlation scores above 0.6, 88% of pairs share at least one Pfam [41] code (Figure 9B), consistent with the assertion that gene homology is a more stringent requirement than domain homology. That the remaining 12% do not share a domain is primarily due to missing annotations. Recall that 12.7% of sequences in our dataset do not contain a recognizable Pfam domain.
Since only some sequences that share a domain are encoded by homologous genes, our model predicts that a significant fraction of sequence pairs that share homologous domains will not have high Neighborhood Correlation scores. In fact, with NC≥0.6, only half of sequence pairs in our dataset share a Pfam domain. These results are consistent with the expectation that gene homology is a less restrictive condition than orthology but more restrictive than domain homology. This analysis provides additional evidence, independent of our curated dataset, that Neighborhood Correlation can predict homologous genes according to our model.
Insight into the evolutionary processes responsible for the development of novel function are of greatest value when considered in the context of entire genomes. To accommodate studies of such scale, a method must be suitable for robust, automated analyses. For the current application, this requires speed, ease of use, and consistent behavior across varied domain architectures.
Neighborhood Correlation displays excellent performance across an array of families with a range of sequence patterns and evolutionary histories. Neighborhood Correlation is able to correctly classify complex families, while maintaining accuracy on simpler families. Further, it displays a classification threshold that is robust with respect to family, yielding good performance on individual families as well as on aggregate datasets in which families may not be known or readily discernible. Since Neighborhood Correlation can be computed easily with existing computing resources and data stores, it is easy to add to a computational workflow. These qualities demonstrate that Neighborhood Correlation is well suited to large-scale genomic analysis.
Empirical evaluation of existing homology detection methods revealed limitations in their applicability, often contrary to common expectations. Meticulous tests of BLAST and PSI-BLAST performance have been carried out on well-characterized datasets [58],[62],[64], but, to our knowledge, performance on multidomain proteins with promiscuous domains and low complexity regions has not been considered empirically. Our tests on datasets with multidomain sequences, promiscuous domains, and low complexity regions show that while BLAST represents an attractive balance between speed and accuracy on conserved, single-domain families, additional screening is needed for correct multidomain classification.
Since Huynen and Bork [27] proposed that alignment length could be used to reduce false positives in ortholog prediction, the practice of pre-screening using an alignment coverage criterion has become widespread in genomic analyses [28]–[37]. To determine the effectiveness of this approach, we investigated the two hypotheses underlying the use of alignment coverage:
Surprisingly, the imposition of an alignment coverage requirement, in addition to sequence similarity, actually decreased the accuracy of homology identification, suggesting that the above hypotheses are not generally true. To our knowledge, this is the first rigorous evaluation of alignment coverage.
Our study suggests that PSI-BLAST, while first-rate for detecting remote homology, is ill-suited to large scale automated analyses on datasets with complex multidomain architectures, promiscuous domains, and low complexity sequences due to its running time, instability, and family dependent score thresholds. The same iterative strategy that confers PSI-BLAST's increased sensitivity leads to a lack of robust behavior when PSI-BLAST is run in an automated manner. Even at extremely stringent inclusion thresholds, false positives are incorporated in during model construction when the query sequence contains promiscuous domains or low complexity regions. Once a false positive is included, PSI-BLAST rapidly degrades the MSA used in subsequent iterations, leading to both incorrect results and excessively long running times. PSI-BLAST required 208 CPU days for our dataset, a 300-fold increase in time over basic BLAST. This slowdown is associated with the large fraction of promiscuous, multidomain, and low complexity sequences in our dataset. When PSI-BLAST is used interactively, the user can eliminate potentially troublesome matches by inspection; however, human intervention is not possible for genome-scale studies. The additional computational cost of calculating Neighborhood Correlation scores once a BLAST search has been performed is negligible. Though PSI-BLAST does offer accuracy improvement over Neighborhood Correlation on families with conserved domain architectures, these issues suggest that PSI-BLAST is impractical for this or larger genomic studies.
Domain architecture comparison performs well on families with low sequence conservation due to the discrimintatory power of multiple alignment based domain models, yet our empirical evaluation of DAC reveals several areas for improvement. Domain architecture comparison can be compromised by faulty or incomplete domain annotation. Failure to capture sequence variation within domain and linker sequences results in an inability to resolve family substructure. A model of promiscuity that better captures domain mobility is needed to correctly classify families defined by a single domain with many partners. Because the sequence similarity network reflects both domain architecture and sequence variation, Neighborhood Correlation avoids many of these difficulties, including unresolved family substructure and sensitivity to domain annotation. Neighborhood Correlation captures modular organization on a range of scales, including sequence motifs as well as structural domains, regardless of whether these subunits are encoded in a database. In addition, Neighborhood Correlation's success on kinase classification, relative to DAC, suggests that it may be possible to derive accurate promiscuity measures from the network.
Neighborhood Correlation differs fundamentally in both goals and approach from Position Specific Scoring Matrices, Profile hidden Markov models, PSI-BLAST, and similar methods that exploit multiple alignments to detect distant homology. MSA-based approaches are not suitable for detecting multidomain homologs with varied architectures. These rely upon full length alignments that are not possible with multidomain sequences. The objective of multiple alignment methods is to identify related sequence motifs when the signal to noise ratio is low. In contrast, the goal of Neighborhood Correlation is to identify homologs that have sustained domain insertions and deletions since their divergence.
Neighborhood Correlation also differs from methods based on multiple alignment in its computational approach. Although both approaches derive information from neighboring sequences, only Neighborhood Correlation exploits the topology of the network. MSA-based methods synthesize a model from a set of neighbors in the sequence similarity network and then use the resulting composite model in pairwise comparisons. Such models reflect aggregate properties of the network neighborhood, but not the underlying topological structure of the network. In contrast, Neighborhood Correlation compares the edge weights for each pair of shared neighbors separately, capturing not only neighborhood membership, but also specific information about how individual sequences in the neighborhood are related. Finally, Neighborhood Correlation derives information from neighborhood difference as well as from neighborhood similarity, taking advantage of the fact that sequences that match one member of the pair and not the other are informative.
Neighborhood Correlation complements a recent set of studies relating multidomain evolution to the global topological properties of the domain similarity network [65]–[69]. Unlike these methods we focus on local network structure as evidence of the evolutionary history of specific sequence pairs and families. In an early use of local network structure, Koonin and colleagues [40] argued that orthologous groups correspond to cliques in the sequence similarity network. In a similar vein, Przytycka and colleagues [70],[71] used a different aspect of local structure (chordality) to test whether domain insertion and intron acquisition are evolving in a parsimonious manner in a given family. In a recent study of protein families in prokaryotes, Medini et al. [72] consider local network structure, but do not relate it to evolutionary processes. In their study, they developed a scoring system based on sets of nearest neighbors in an unweighted network and used these pairwise scores to identify core sets of proteins associated with secretion systems in prokaryotes.
Neighborhood Correlation links local network structure to both domain architecture and evolutionary process. The similarities and differences in domain architecture are reflected in the neighborhoods of adjacent sequences. The number and weights of edges in the shared neighborhood is influenced by the number and conservation of their shared domains. Their unique neighborhoods are similarly influenced by their unique domains. The Neighborhood Correlation score, therefore, is an implicit measure of both sequence similarity and domain architecture comparison.
The history of gene duplication and domain insertion in gene family evolution is also recorded in network topology. Neighborhood Correlation is able to elucidate multidomain homology because it can decipher the traces of this history in the network. In particular, Neighborhood Correlation relies on the hypothesis that the neighborhoods of genes related through duplication are more similar to each other than the neighborhoods of genes related through domain insertion. This hypothesis in turn assumes that
There is concrete evidence to support the latter assertion as indicated by the negative correlation between the promiscuity and sequence identity of Pfam domains, discussed in Results. We are not aware of any studies predicting the relative rates of gene duplication and domain insertion. However, the success of Neighborhood Correlation in classifying multidomain homologs provides indirect evidence that the assertion is true, at least in the dataset studied here. If, contrary to this hypothesis, domain insertions occurred as or more frequently than gene duplications, the Neighborhood Correlation scores of multidomain homologs would not be distinctly higher than those of domain-only matches.
More generally, the success of Neighborhood Correlation has demonstrated that information about the interplay of the processes of gene duplication, domain shuffling, and sequence divergence lies hidden in the local structure of the sequence similarity network. This success suggests that mining network structures is a promising direction for extending bioinformatics methodology, as well as for asking basic questions about evolutionary processes. For example, it has been argued that the increased complexity of multidomain families in metazoans is directly related to the advent of multicellular animals. Multicellularity has evolved several times ([73] and work cited therein). In each case, Nature has had to evolve novel solutions to the problems of coordinated cellular communication and control. It is an intriguing question whether the same patterns of gene duplication and domain insertion that prompted the evolution of metazoan signal transduction families also dominate in other lineages. Future work will determine whether we can further exploit local organization of the sequence similarity network to investigate such questions.
We extracted all complete mouse and human protein sequences from SwissProt Version 50.9 [74], yielding 11,553 mouse protein sequences and 14,644 human protein sequences. Sequence fragments were excluded from this set of sequences by rejecting sequences annotated with a description field containing “(fragment”. We chose SwissProt, a high quality, curated protein sequence database, as opposed to GenBank, which would have resulted in a larger, but less reliable, dataset. KOG annotations were obtained from the Clusters of Orthologous Groups database [40], available from ftp://ftp.ncbi.nih.gov/pub/COG/KOG/. KOG annotations were mapped to SwissProt identifiers by exact matching of KOG FASTA protein sequences with those in SwissProt.
The analysis was carried out on the combined set of mouse and human sequences. In a preliminary study, we compared the performance of Neighborhood Correlation on a smaller, combined set of mouse and human sequences with its performance on separate sets of mouse and human sequences [75] to determine whether Neighborhood Correlation performs differently on comparisons within and across genomes. The mouse-only and human-only data test the ability to classify paralogs within a single mammalian species, as opposed to the combination of orthologs and paralogs seen in the combined dataset. The basic trends in the mouse-only and human-only datasets were the same as the combined dataset for all tests performed. This suggests that Neighborhood Correlation performance is not highly sensitive to the degree of sequence divergence, since paralogous and orthologous sequences in these species exhibit different patterns of divergence.
For each family, we derived a list of designated gene symbols, Pfam [41] and/or InterPro [76] codes from publications by family experts, and reports from the Human Genome Nomenclature Committee (http://www.gene.ucl.ac.uk/nomenclature/genefamily.html). These lists were used to generate a preliminary roster for each family, then confirmed by referring to recent analyses of gene family evolution in the literature. A detailed account of the curation procedure for each family with specific identification criteria and references is given in Text S1. SwissProt accession numbers for all sequences in the twenty families are provided in Dataset S1.
We conducted all-against-all BLAST (Version 2.2.15) [61] and PSI-BLAST (Version 2.2.16) [58] searches for the sequences in our dataset, using the BLOSUM 62 matrix, an affine gap penalty of −(11+k) for a gap of length k, and low complexity filtering. For both searches, the size of the search space was set to Y = n2 and the significance threshold to E = 10N, where n is the size of the database in residues and N is the number of sequences in the dataset.
The combined dataset has N = 26,197 sequences, 11,553 mouse and 14,644 human sequences, corresponding to a total of n = 14,073,417 residues. For PSI-BLAST, four passes were executed with an inclusion threshold of E<10−13 for inclusion in the multiple alignment used to search in the next pass. Although this cutoff is much more stringent than the default, we found it essential to obtain correct results with sequences containing low complexity regions. Less stringent thresholds resulted in the inclusion of unrelated sequences in the intermediate PSSM. Asymmetries (i.e., E(x,y)≠E(y,x)) that occur due to low complexity filtering [77], which is applied only to the query sequence but not to database sequences, were corrected by assigning the better of the two values to both matrix entries. The resulting dataset had 4,864,226 significant BLAST pairs and 10,854,626 significant PSI-BLAST pairs.
The parameter values used in this study embody the view that an all-against-all BLAST search is a single experiment. This approach is roughly equivalent to conducting N single query BLAST searches with E = 10 and Y = mx n, where mx is the length of query sequence x. Treating the all-against-all BLAST comparison as a single experiment results in symmetric E-values in the absence of low complexity filtering. We define θ(x,y) = E(x,y)/10N to be the expected number of chance hits per sequence in the dataset with a score equivalent to, or better than, that of the alignment of query sequence x with matching sequence y. The significance threshold of E = 10N corresponds to θ = 10 chance hits per sequence, in expectation.
We calculated the Neighborhood Correlation scores for all sequence pairs in our dataset from Equation 1 using the similarity score,(2)where ς(x,i) is the normalized bit score [58] of the alignment of x and i and ςmin(x,i) = log2(n2/10N)*0.95 = 28.019, which is 5% less than the bit score corresponding to θ = 10 for a dataset of the size used in this study.
The effectiveness of Neighborhood Correlation depends strongly on how the similarity score, S(x, i), is defined. We considered three measures of similarity: S(x,i) = log ς(x,i), S(x,i) = ς(x,i) and an unweighted comparison of neighborhood membership defined as S(x,i) = 1 if there is a significant match between x and i, and zero otherwise. Although the other two measures performed well on some families, only S(x,i) = log ς(x,i) gave consistent, good performance on a wide range of families. This suggests two factors that may be important to Neighborhood Correlation performance. First, the relatively poor performance of the unweighted score indicates that it is necessary to capture differences in the degree of similarity to sequences in the neighborhood to capture complete evolutionary information. Second, the improved performance obtained with S(x,i) = log ς(x,i) can be understood by recalling that the correlation coefficient captures only linear associations. The use of the logarithm compresses the range of ς(.,.), resulting in scores that more closely approximate linearity.
The choice of ςmin, the score assigned to pairs without significant similarity, may influence Neighborhood Correlation performance in homology identification. We experimented with values of ςmin corresponding to significance thresholds ranging over two orders of magnitude. The results (data not shown) suggest that varying ςmin has little impact on Neighborhood Correlation.
Promiscuity refers to the tendency of domains to be inserted into many different contexts. Typically, promiscuity of a domain is defined as the number of distinct partners associated with it, where two domains are partners if they co-occur in at least one sequence [3]. We obtained the set of Pfam codes associated with all sequences in our dataset from the SwissProt database. For each Pfam domain, we determined the number of distinct Pfam codes that co-occur with it in any of the 26,197 sequences in our dataset.
We further obtained percent sequence identity for each Pfam identifier from the Pfam website. The Spearman ranked correlation coefficient of domain promiscuity and sequence identity was calculated to evaluate whether promiscuity and sequence identity were related.
We conducted an all-against-all domain architecture comparison using the Pfam identifiers provided by SwissProt. Similarity of each pair of sequences, x and y were calculated as follows:(3)where w(di,x) is the weight of domain di in sequence x. Domains are assigned weights inversely proportional to their promiscuity. Promiscuous domains may occur in many unrelated sequences, and so are less useful than relatively rare domains in determining homology. The weight of a domain not contained in a given sequence is zero. As a result, pairs of sequences which share no domains are assigned a similarity of zero. This domain architecture comparison function corrects for the bias of proteins with many domains. Proteins with numerous domains have an elevated probability of sharing a domain with other proteins. Of the 21 domain architecture comparison methods we evaluated in a previous study [23], this was shown to have the best performance.
For every pair of sequences, x and y, with significant similarity, we calculated the alignment coverage, defined as α(x,y) = 2la/(lx+ly), where lx and ly are the length of sequences x and y, and la is the length of the optimal local alignment, define to be the number of columns needed to represent it; that is, it includes gapped positions. The length of the optimal alignment between query x and match y will not, in general, be the same as the length of the optimal alignment between query y and match x. We forced the alignment coverage to be symmetric by setting both α(x,y) and α(y,x) to the maximum of the two values.
By considering only the optimal alignment, we risk underestimating the extent of similarity between homologous sequences. To take suboptimal alignments into account, we used a simple heuristic method for selecting a set of high-scoring local alignments that do not conflict. Two alignments conflict if they overlap or do not appear in the same order in both sequences (see Text S1).
Classifier performance was evaluated using Receiver Operating Characteristic (ROC), which captures the tradeoff between sensitivity (Sn) and specificity (Sp) as a function of the classifier threshold. A ROC curve is a plot of Sn as a function of 1−Sp, where Sn = TP/(TP+FN) and Sp = TN/(TN+FP). TP, FP, TN, and FN refer to the number of True Positives, False Positives, True Negatives, and False Negatives, respectively. In the context of our test, TP is the number of sequence pairs that have common ancestry and have been correctly identified by the classifier. FP represents the number of pairs that are classified as homologs, but are not family pairs. TN and FN refer to the number of non-homologous pairs that are correctly ruled out and incorrectly included, respectively.
The area under the ROC curve provides a single measure of classification accuracy, corresponding to the fraction of correctly classified entities given the best possible choice of threshold. We used the ROC-n score, defined to be the area under the ROC curve truncated after the first n false positives or(4)where ti is the number of FF pairs observed before the ith FO pair and T is the total number of FF pairs in the dataset. When the number of negative examples far exceeds the number of positive examples, as is the case here, the ROC score approaches one, resulting in an unjustifiably optimistic assessment of classifier performance. Rn is a more sensitive figure of merit than the untruncated ROC score in this case [78]. We selected n = 100k, where k is the number of FF pairs. This is equivalent to 100 false positives per query. We found that 100k was sufficiently large so that few FF pairs were missed in most tests but not so large so as to obscure the differences in performance between classifiers.
The statistical significance of the difference between the ROC-n scores obtained by Neighborhood Correlation and sequence similarity was estimated using p-values calculated using the method described in Schaffer et al. [62]. This method tests the null hypothesis that the difference in ROC-n scores is due the sampling process used to obtain the test data. Rejection of the null hypothesis indicates that the difference in ROC-n scores represents a true difference in the performance of the classifiers.
Precision and Recall are also used for evaluation. In the context of our test, Recall denotes the fraction of homologous pairs retrieved and is equivalent to sensitivity. Precision refers to the fraction of protein pairs retrieved that are actually homologous pairs.
http://www.neighborhoodcorrelation.org
The accession numbers used in this paper are from Swiss Prot (http://www.ebi.ac.uk/swissprot): human PDGFRG (P09619), human PRKG1B (P14619), and mouse NCAM2 (O35136). Accession numbers for all 1577 sequences in the twenty families in our benchmark are given in Dataset S1. |
10.1371/journal.ppat.1003699 | Deletion of IL-4 Receptor Alpha on Dendritic Cells Renders BALB/c Mice Hypersusceptible to Leishmania major Infection | In BALB/c mice, susceptibility to infection with the intracellular parasite Leishmania major is driven largely by the development of T helper 2 (Th2) responses and the production of interleukin (IL)-4 and IL-13, which share a common receptor subunit, the IL-4 receptor alpha chain (IL-4Rα). While IL-4 is the main inducer of Th2 responses, paradoxically, it has been shown that exogenously administered IL-4 can promote dendritic cell (DC) IL-12 production and enhance Th1 development if given early during infection. To further investigate the relevance of biological quantities of IL-4 acting on DCs during in vivo infection, DC specific IL-4Rα deficient (CD11ccreIL-4Rα-/lox) BALB/c mice were generated by gene targeting and site-specific recombination using the cre/loxP system under control of the cd11c locus. DNA, protein, and functional characterization showed abrogated IL-4Rα expression on dendritic cells and alveolar macrophages in CD11ccreIL-4Rα-/lox mice. Following infection with L. major, CD11ccreIL-4Rα-/lox mice became hypersusceptible to disease, presenting earlier and increased footpad swelling, necrosis and parasite burdens, upregulated Th2 cytokine responses and increased type 2 antibody production as well as impaired classical activation of macrophages. Hypersusceptibility in CD11ccreIL-4Rα-/lox mice was accompanied by a striking increase in parasite burdens in peripheral organs such as the spleen, liver, and even the brain. DCs showed increased parasite loads in CD11ccreIL-4Rα-/lox mice and reduced iNOS production. IL-4Rα-deficient DCs produced reduced IL-12 but increased IL-10 due to impaired DC instruction, with increased mRNA expression of IL-23p19 and activin A, cytokines previously implicated in promoting Th2 responses. Together, these data demonstrate that abrogation of IL-4Rα signaling on DCs is severely detrimental to the host, leading to rapid disease progression, and increased survival of parasites in infected DCs due to reduced killing effector functions.
| Leishmaniasis is a parasitic infection caused by protozoan parasites of Leishmania species and is transmitted by the sandfly. Disease in humans ranges from localized cutaneous lesions to disseminated visceral Leishmaniasis. Mouse models of Leishmania major infection have demonstrated that a “healing” response in C57BL/6 mice requires the secretion of protective T helper (Th) 1 cytokines, including IFN-γ, which mediates parasite killing by inducing nitric oxide production. Conversely, “non-healer” BALB/c mice are unable to control infection and develop a Th2 immune response characterized by the production of IL-4 and IL-13 cytokines. Although IL-4 is the main inducer of Th2 responses, it has been shown that IL-4 can instruct dendritic cell (DC)-derived IL-12 production and Th1 development if administered during DC activation. To further investigate the role of DCs, a DC specific IL-4Rα-deficient mouse model was established. L. major studies demonstrated hypersusceptibility to infection and strikingly increased parasite loads in peripheral organs of mice lacking IL-4Rα on DCs. Moreover, increased parasite burdens were observed in host cells, including DCs, which showed reduced killing effector functions. In summary, this study demonstrates that IL-4Rα-mediated instruction of DCs occurs in vivo and is necessary to avoid rapid progression of disease in the host.
| Leishmania spp. are protozoan parasites that are transmitted by Phlebotomus spp. sandflies and can cause several forms of disease in humans, ranging from localized cutaneous lesions to visceral Leishmaniasis, where parasites invade internal organs such as the spleen and liver. The incidence of disease is approximately 1.5 million per annum for cutaneous Leishmaniasis, and 500 000 per annum for visceral Leishmaniasis, which is usually fatal if left untreated [1]. Currently there is no vaccine. To identify correlates of immune protection, which may aid in vaccine design and therapeutic strategies, experimental models of cutaneous Leishmaniasis have been established in which disease is induced by infecting mice subcutaneously with L. major. Susceptible BALB/c mice show progressive lesion development with dissemination of parasites to visceral organs, while resistant C57BL/6 mice are able to control infection and heal lesions [2]–[4]. Lack of healing in BALB/c mice is associated with a T helper (Th) 2 response characterized by secretion of interleukin (IL)-4, IL-5, IL-9 and IL-13 [3], [5]–[8], high anti-Leishmania antibody titres [8], [9] and alternative activation of macrophages [9], [10]. In contrast, resistant C57BL/6 mice develop protective Th1 responses with production of IL-12 and IFN-γ, associated with classical activation of macrophages and killing of parasites by effector nitric oxide production [9], [11]–[14]. IL-4 and IL-13, both of which signal through a common receptor chain, the IL-4 receptor alpha (IL-4Rα) are known to be important susceptibility factors in L. major infection [3], [6], [8], [15], [16]. Both BALB/c and C57BL/6 mice secrete IL-4 early after infection however, production of IL-4 is sustained in susceptible BALB/c mice and transient in resistant C57BL/6 mice [17], [18]. It appears that resistant mouse strains redirect the early Th2 response in an IL-12-dependent mechanism, while in susceptible mice the Th2 response persists and dominates the disease outcome by suppressing effector mechanisms needed for parasite killing [3].
While IL-4 is the primary inducer of Th2 responses [19], paradoxically it has also been shown that IL-4 promotes IL-12 production by bone marrow-derived dendritic cells (BMDCs) stimulated with CpG or LPS [20]–[23]. Furthermore, administration of 1 µg of recombinant IL-4 at 0 and 8 hours after infection with L. major led to increased IL-12 mRNA expression by dendritic cells (DCs) in vivo, promoted Th1 responses and rendered mice resistant to infection [21]. It has also been shown that global abrogation of IL-4Rα renders mice resistant to L. major only in the acute phase of infection, with mice continuing to develop necrotic footpad lesions during the chronic phase [15]. However, specific abrogation of IL-4Rα on CD4+ T cells does lead to resistance, indicating a protective role for IL-4Rα signalling on non-CD4+ T cells [24].
A candidate for this protective role may therefore be DCs. These sentinels of the immune system are specialized antigen-presenting cells, proficient at uptake of antigen, migration to the lymph nodes (LN) and activation of lymphocytes. Consequently, they play a critical role in the initiation and differentiation of the adaptive immune response [25], [26]. To investigate the role of IL-4Rα signaling on DCs in resistance to Leishmania, CD11ccreIL-4Rα-/lox mice, deficient in IL-4Rα signaling on DCs, were generated and infected with L. major LV39 and IL81 strains. CD11ccreIL-4Rα-/lox mice were hypersusceptible to both strains of L. major, with increased footpad swelling and necrosis and substantially increased parasite burdens in peripheral organs, including the brain. Hypersusceptibility in CD11ccreIL-4Rα-/lox mice was associated with an upregulation of Th2 responses, impairment in iNOS production by macrophages and inflammatory DCs and increased parasite loads in LN and spleen DCs. Therefore, it is clear that IL-4Rα signaling has important effects on DC phenotype during cutaneous L. major infection, and is necessary to avoid rapid disease progression in the host. This study therefore expands our knowledge on the role of dendritic cells during cutaneous Leishmaniasis and on the effects of IL-4Rα signaling on dendritic cells.
Mice expressing cyclization recombinase (Cre) under control of the cd11c locus [27] were backcrossed to BALB/c for 9 generations, then intercrossed with global IL-4Rα (IL-4Rα-/-) [15] BALB/c mice to generate CD11ccreIL-4Rα-/- BALB/c mice. These mice were subsequently intercrossed with floxed IL-4Rα (IL-4Rαlox/lox) BALB/c mice (exon 6 to 8 flanked by loxP) [28] to generate CD11ccreIL-4Rα-/lox BALB/c mice (Figure 1A). CD11ccreIL-4Rα-/lox mice were identified by PCR genotyping (Figure 1B). Analysis of IL-4Rα surface expression on different cell types by flow cytometry demonstrated that IL-4Rα was efficiently depleted in DCs of the lymph nodes, spleen, skin and lungs, when compared with IL-4Rα-/lox littermate controls and IL-4Rα-/- mice (Figure 1C). As expected CD11c+ alveolar macrophages also had abrogated IL-4Rα surface expression. Other cell types such as T cells, B cells and macrophages had comparable IL-4Rα expression to IL-4Rα-/lox littermate controls. Cre-mediated IL-4Rα deletion in DCs was confirmed at the genomic level by performing PCR for IL-4Rα exon 8 (absent in IL-4Rα-deficient cells) normalized to IL-4Rα exon 5 (present in all cells) using DNA from CD11c+MHCII+ DCs sorted from the spleens of naïve mice (Figure 1D).
To assess functional impairment of DCs in CD11ccreIL-4Rα-/lox mice, we generated bone marrow-derived dendritic cells and stimulated them with LPS in the presence or absence of IL-4 or IL-13. IL-4 is known to enhance DC production of IL-12 in an IL-4Rα dependent manner, so called “IL-4 DC instruction” [21]–[23]. As expected, BMDCs derived from IL-4Rα-/lox mice and BALB/c wildtype controls had significantly increased IL-12 production after the addition of IL-4 (Figure 1E). In contrast, LPS/IL-4 stimulated BMDCs derived from CD11ccreIL-4Rα-/lox mice or from global IL-4Rα-/- mice showed similar levels of IL-12 to those stimulated with LPS alone, with IL-4 having no effect. This demonstrates functional impairment of IL-4Rα signaling on DCs from CD11ccreIL-4Rα-/lox mice. In fact, after the addition of LPS alone, BMDCs with a functional IL-4Rα already showed a trend towards increased IL-12p40 levels, suggesting that endogenous levels of IL-4 found in the culture could influence these BMDCs. IL-13 did not increase levels of IL-12, confirming previous DC stimulation studies [22]. As previously reported [29], IL-4 and IL-13 had no significant effect on BMDC maturation, as shown by similar expression of MHCII, CD86, CD80, CD83 and CD40 (data not shown). Total yield of BMDCs per precursor cell seeded was similar in CD11ccreIL-4Rα-/lox mice and littermate controls and survival after maturation was not significantly different (data not shown).
In order to investigate the role of IL-4Rα signaling on DCs during cutaneous Leishmaniasis, CD11ccreIL-4Rα-/lox mice were infected subcutaneously with 2×106 stationary phase metacyclic promastigotes of L. major LV39 (MRHO/SV/59/P; Figure 2A, 2B and 2C) or with a more virulent GFP-expressing L. major IL81 (MHOM/IL/81/FEBNI; Figure 2D, 2E and 2F) strains into the hind footpad. As previously shown [15], [24], C57BL/6 mice and IL-4Rα-/- deficient BALB/c mice controlled lesion development during acute infection with both L. major strains (Figure 2A and 2D), which correlated with low parasite numbers in infected footpads (Figure 2B and 2E) and draining lymph nodes (Figure 2C and 2F). Susceptible WT BALB/c and IL-4Rα-/lox littermate control mice developed progressive footpad swelling after infection with both strains (Figure 2A and 2D), with increased parasite burdens in the infected footpads (Figure 2B and 2E) and draining LN (Figure 2C and 2F). Hemizygous (IL-4Rα-/lox mice) had slightly reduced footpad swelling compared to BALB/c mice in IL81 infection. The greater virulence of IL81 is reflected in more rapid disease progression, with footpad swelling and parasite burden reaching similar levels by 4 weeks to those obtained with LV39 in 8 weeks. Of importance, CD11ccreIL-4Rα-/lox mice were hypersusceptible to acute L. major infection compared to heterozygous littermate controls and BALB/c mice, showing considerably worsened disease progression when infected with either strain (Figure 2A and 2D), with earlier and dramatically larger footpad lesions, and development of early necrosis (Figure 2A and 2D). Increased disease progression was accompanied by significantly higher parasite numbers in the footpads (Figure 2B and 2E) and LN (Figure 2C and 2F) of infected animals. In addition, infection with a 10-fold lower dose of L. major LV39 also resulted in a hypersusceptible phenotype in CD11ccreIL-4Rα-/lox mice (Supplementary Figure S1 A–C). Histopathological analysis of CD11ccreIL-4Rα-/lox footpads at week 4 after infection with the virulent IL81 revealed severe destruction of epidermis, connective tissue and bone as a result of footpad necrosis, accompanied by increased inflammatory infiltrates and a high load of extracellular L. major amastigotes (Figure 2G). In contrast, infected footpads of IL-4Rα-/lox revealed moderate dermal inflammatory infiltrates with mostly intact epidermis, connective tissue and bone. Together, these data reveal that IL-4Rα signaling on DCs play an important role in host protection against acute L. major infection.
Th1/Th2-type responses were investigated in CD11ccreIL-4Rα-/lox mice and controls during acute cutaneous leishmaniasis (IL81). Antigen-specific restimulation of CD4+ T cells sorted from the LN of infected mice and co-cultured with fixed antigen-presenting cells and soluble Leishmania antigen (SLA) revealed a significantly reduced IFN-γ response in CD11ccreIL-4Rα-/lox mice in comparison to the resistant C57BL/6 or IL-4Rα-/- strains as well as to the susceptible IL-4Rα-/lox littermate controls (Figure 3A). Conversely, the levels of IL-4, IL-13 and IL-10 were significantly higher in CD11ccreIL-4Rα-/lox mice compared to IL-4Rα-/lox, IL-4Rα-/- and C57BL/6 mice (Figure 3B, 3C and 3D). The observed shift in cytokine responses was confirmed in LN cells, stimulated with anti-CD3 or SLA (data not shown) and systemically in the quality of Leishmania-specific antibody immune responses. Sera of week 4 infected mice revealed a predominant type 1 antibody response in IL-4Rα-/- mice, as shown by elevated levels of Leishmania-specific IgG2a (Figure 3E). In contrast, CD11ccreIL-4Rα-/lox mice displayed a predominant type 2 antibody response shown by marked production of IgG1 and total IgE, which was significantly higher than that observed in littermate IL-4Rα-/lox mice (Figure 3F and 3G). A shift towards Th2-type responses also occurred in CD11ccreIL-4Rα-/lox mice in a 10-fold lower dose L. major LV39 infection (Supplementary Figure S1 D–H).
As IFN-γ-induced nitric oxide synthase (iNOS) production by classically activated macrophages (caMphs) is a key control mechanism in L. major infection [14], the activation state of macrophages was determined in the infected footpad at week 4 after infection. Inflammatory macrophages (CD11b+MHCII+CD11c−) from CD11ccreIL-4Rα-/lox mice had significantly reduced iNOS expression compared to those of littermate IL-4Rα-/lox control mice (Figure 3H). Conversely, expression of arginase 1, a marker of alternatively activated macrophages (aaMphs), was higher in macrophages of CD11ccreIL-4Rα-/lox mice (Figure 3I). This altered phenotype was confirmed in iNOS and arginase activity assays performed on total footpad cells stimulated with LPS (Figure 3J and 3K). Together, these results demonstrate a shift towards Type 2 responses and reduced macrophage effector functions in CD11ccreIL-4Rα-/lox mice.
In L. major LV39 infection, parasites were present only in footpads and the draining lymph nodes at week 3, whereas by week 8 parasites had disseminated to the spleen and liver in both CD11ccreIL-4Rα-/lox mice and littermate controls (Figure 4A and 4B). Parasite burdens were much higher in the organs of infected CD11ccreIL-4Rα-/lox mice, compared to littermate control mice. Moreover, in some CD11ccreIL-4Rα-/lox mice, but not in control mice, L. major parasites had disseminated as far as the brain by week 8 after infection (Figure 4B). Similar disease progression was observed after infection with L. major IL81 (Figure 4C), where CD11ccreIL-4Rα-/lox mice already displayed noticeable splenomegaly at 4 weeks post infection (data not shown), and had strikingly increased parasite burdens in all organs analyzed, including the brain (Figure 4C). Histological analysis confirmed the increased presence of disseminated parasites in the spleen and liver of CD11ccreIL-4Rα-/lox mice (IL81, week 4), as shown by the high load of extracellular L. major amastigotes (spleen) and the prevalence of inflammatory foci and leishmanial bodies in mononuclear cells (liver) (Figure 4D). The presence of parasites in brains of perfused CD11ccreIL-4Rα-/lox mice (IL81, week 4) was also confirmed by confocal microscopy (Figure 4E). Parasites were not visible in the brains of littermate controls (data not shown). These results demonstrate a drastic increase in numbers of disseminated parasites in peripheral organs of infected CD11ccreIL-4Rα-/lox mice. Although it has been reported that dissemination could occur within hours after high-dose parasite inoculation [30], infection with GFP+ L. major IL81 and analysis by flow cytometry demonstrated that GFP+ parasites was not detectable in the spleen at 1 or 3 days post infection, whereas at week 4 there was an increase in GFP+ cells compared to day 0 (Supplementary Figure S2).
In order to determine if dendritic cells could harbor L. major parasites, GFP-expressing L. major parasites (IL81) were used to track infected cell populations in different organs by flow cytometry at different time-points (day 3, day 7 and week 4) after infection. Parasite replication occurred in GFP+ cell populations that were sorted and plated out for limiting dilution assays, indicating that GFP positivity was a good marker for viable parasites associated with cells (Supplementary Figure S3). At day 3 after GFP-L. major IL81 infection, plasmacytoid DCs (pDCs), macrophages and neutrophils had infiltrated the infected footpad (Figure 5A). By 4 weeks post infection, numbers of infiltrating cells had increased substantially, with conventional DCs (cDCs) also now present in high numbers (Figure 5B). The number of infiltrating cells was significantly higher in CD11ccreIL-4Rα-/lox mice compared to IL-4Rα-/lox mice (Figure 5B). At the early time point in FP, macrophages were infected with GFP+ Leishmania, with similar numbers in CD11ccreIL-4Rα-/lox mice and littermate controls (Figure 5C). This was in contrast to the draining lymph node, where conventional and plasmacytoid DCs were infected, with higher numbers of DCs infected in CD11ccreIL-4Rα-/lox mice compared to controls (Figure 5D). Similar results were obtained at day 7 post-infection (data not shown). At week 4 post infection, the footpad harbored a pool of infected cells, namely macrophages, cDCs and neutrophils (Figure 5E), while in the draining lymph node, the cDCs were still infected compared to the other cell types (Figure 5F). Again the number of infected DCs was significantly higher in CD11ccreIL-4Rα-/lox mice (Figure 5E and 5F) compared to littermate controls. However, overall numbers of DCs infiltrating the LN at week 4 after L. major IL81 infection were similar in both CD11ccreIL-4Rα-/lox mice and littermate control mice (data not shown), suggesting that differences in parasite killing and not DC migration were responsible for the increased number of infected DCs in CD11ccreIL-4Rα-/lox mice.
Infected DCs were also found in the spleen, with significantly increased numbers of infected cells in CD11ccreIL-4Rα-/lox mice compared to controls (Figure 6A). However, overall numbers of DCs infiltrating the spleen were also increased to a similar degree in both CD11ccreIL-4Rα-/lox mice and littermate controls at week 4 (data not shown), again suggesting that differences in parasite killing and not DC migration were responsible for the increased parasite loads in CD11ccreIL-4Rα-/lox mice. Although it is well known that iNOS-mediated NO production in classically-activated macrophages drives intracellular killing of L. major parasites, a recent study has now implicated a population of iNOS+ – producing inflammatory DCs in controlling Leishmania infection [31]. We therefore examined iNOS production by DCs in CD11ccreIL-4Rα-/lox and littermate control mice using intracellular FACS. In hypersusceptible CD11ccreIL-4Rα-/lox mice, a significantly reduced percentage of CD11chighMHCIIhigh DCs produced iNOS compared to DCs from IL-4Rα-/lox littermate control mice (Figure 6B). This was confirmed at the level of intracellular NO expression, which was also reduced in DCs from CD11ccreIL-4Rα-/lox mice (Figure 6C). Together, these data demonstrate that DCs from CD11ccreIL-4Rα-/lox mice have reduced NO killing effector functions, further explaining the increased parasite burdens in the DCs of these mice.
Previous studies using BMDCs found that IL-4-mediated instruction results in reduced IL-10 production that is responsible for increased IL-12p40 production by DCs upon stimulation with IL-4 plus CpG or LPS [21], [23]. To test whether endogenous amounts of IL-4 could mediate DC instruction in vivo, CD11ccreIL-4Rα-/lox mice and controls were infected with L. major IL81. At 4 weeks post infection, total LN cells were restimulated with SLA and cytokines were measured in the supernatant. Lymph node cells from infected CD11ccreIL-4Rα-/lox mice produced significantly reduced IL-12p40 but increased IL-10 compared to littermate IL-4Rα-/lox mice (Figure 7A). Moreover, intracellular cytokine staining revealed that DCs from CD11ccreIL-4Rα-/lox mice produced less IL-12p40 and more IL-10 than those from littermate IL-4Rα-/lox controls (Figure 7B and Figure S4). Quantification of mRNA found decreased expression of the Th1-promoting cytokine genes for IL-12p40 (Figure 7C) and IL-18 (Figure 7D) in sorted LN DCs from CD11ccreIL-4Rα-/lox mice compared to controls. In contrast, there was a trend towards increased mRNA expression of IL-10 as well as significantly increased mRNA expression of IL-23 and activin A, cytokines which are involved in inducing Th2 responses by promoting Th17 and alternative activation of macrophages, respectively [32], [33] (Figure 7E–G). In addition, differences in IL-12p70 production were detected in vitro. L. major/IL-4 stimulated BMDCs derived from IL-4Rα-/lox mice showed increased IL-12p70 production, whereas IL-4 had no additive effect on IL-12p70 production in BMDCs from CD11ccreIL-4Rα-/lox mice (Figure 7H). IL-13 did not increase IL-12p70 production, as previously shown [22].
Understanding mechanisms of immune control in cutaneous Leishmaniasis is critical for the design of effective therapeutics and vaccines. Although several studies have clearly established that IL-4 is a key cytokine in the development of non-healing disease in BALB/c mice [8], [19], [34], [35], apparently contradictory evidence also suggests that IL-4 has the ability to instruct protective Th1 responses [21], [36]–[41]. The term “instruction theory” was coined when IL-4 was found to promote increased production of IL-12 by BMDCs [20]–[22]. IL-4, but not IL-13, enhances the production of IL-12 induced by pathogen products via signalling through the type 1 IL-4 receptor [21], [22]. The mechanism behind instruction was found to be inhibition of IL-10 by IL-4, leading to higher levels of IL-12 and increased protective Th1 responses [23]. Several studies also indicate that IL-4 and IL-13 may play a role in promoting DC maturation [22], [42]. However, most in vitro and in vivo studies on the effects of IL-4 and IL-13 on DCs have been conducted with exogenously administered IL-4 or IL-13, and thus the relevance of biological quantities of IL-4 signalling through IL-4Rα on DCs during disease in vivo has not been demonstrated. To address these issues, dendritic cell-specific (CD11ccreIL-4Rα-/lox) BALB/c mice were generated using the cre/loxP recombinase system under control of the cd11c locus. These mice were found to have abrogated IL-4Rα expression on DCs and alveolar macrophages, with other cell types maintaining IL-4Rα expression and functioning.
Infection of CD11ccreIL-4Rα-/lox mice with L. major LV39 and IL81 revealed IL-4Rα signaling on DCs to be highly important in protection against cutaneous Leishmaniasis. Compared to IL-4Rα-/lox littermate controls, CD11ccreIL-4Rα-/lox mice showed dramatically worsened disease progression, with increased footpad swelling and necrosis, and significantly higher parasite burdens both locally and in visceral organs such as the spleen and liver. As expected, genetically resistant C57BL/6 mice effectively controlled infection, as did global IL-4Rα-/- mice, which have been shown to be resistant during the acute phase of L. major infection, with disease progression in the chronic phase only [15], [24]. Progressive disease during L. major infection in BALB/c mice has been attributed to the predominance of Th2 cytokines and type 2 antibody immune responses [8], [9], [11], with a previous study by our laboratory showing that CD4+ T cell specific IL-4Rα deficient mice were highly resistant to L. major infection [24]. Analysis of CD4+ T cell cytokine responses in CD11ccreIL-4Rα-/lox mice revealed a decrease in IFN-γ accompanied by a marked increase in IL-4, IL-13 and IL-10, while increased secretion of IgG1 and IgE by B cells confirmed a shift towards a Th2-type immune phenotype. Aside from its role in instruction, IL-10 is known to be a susceptibility factor for L. major infection, being produced at higher levels in susceptible BALB/c mice and capable of suppressing Th1-mediated effector functions [3], [43]. In humans, IL-10 is strongly associated with persistent infection [44].
IFN-γ plays an important role in mediating protective immunity during L. major infection by classically-activating macrophages to induce nitric oxide synthase-mediated NO production for intracellular killing of parasites [9], [14], [45], [46]. Latent Leishmaniasis is reactivated in chronically infected healthy C57BL/6 mice by inhibition of endogenous NOS-2, indicating that iNOS expression is crucial for the sustained control of L. major infection [9], [31], [47]. Induction of iNOS-mediated NO production is counter-regulated by IL-4/IL-13 and IL-4Rα, which promote the development of alternatively activated macrophages and arginase 1 production through depletion of L-arginine as a substrate for iNOS. Interestingly, IL-10 has also been shown to suppress intracellular killing of pathogens in macrophages by suppressing IFN-γ responses [48]–[50] and can induce an alternatively activated macrophage type phenotype in the absence of IL-4 and IL-13 [51]. Parasites such as Leishmania can utilize polyamines generated by arginase 1 activity for their own growth, making alternatively activated macrophages a favorable environment for their survival [52]–[54]. Accumulating reports have demonstrated a role for alternative macrophage activation and arginase 1 expression in influencing susceptibility to L. major infection [7], [9], [55], [56]. LysMcreIL-4Rα-/lox mice which lack IL-4/IL-13 induced alternative activation of macrophages were found to have increased resistance to infection [9], while neutralization of endogenous arginase 1 with N-hydroxy-nor-L-arginine leads to complete healing in BALB/c mice [55].
Macrophages from the footpads of CD11ccreIL-4Rα-/lox mice were found to have reduced iNOS expression and increased arginase 1 expression compared to those from littermate control IL-4Rα-/lox mice, demonstrating a shift in macrophage effector function most likely as a consequence of increased IL-4, IL-13 and IL-10. Recently it has been shown that DCs can also become alternatively activated by upregulating markers such as Ym-1 and RELM-α after administration of IL-4 [29]. In our study, the data suggest that IL-4Rα-independent alternative activation of DCs is also possible, as DCs from CD11ccreIL-4Rα-/lox had decreased iNOS expression, possibly a consequence of reduced IFN-γ and/or increased IL-10 and activin A, and had higher parasite loads than those from littermate controls. Previous studies have revealed that iNOS-producing DCs constitute a major Th1-regulated effector cell population and contribute to resistance to infection by L. major [31], L. monocytogenes [57] and Brucella spp. [58]. The reduced ability of both macrophages and DCs to initiate NO-mediated killing of L. major in CD11ccreIL-4Rα-/lox mice is therefore likely to play a role in the uncontrolled parasite replication observed both in the footpad and at peripheral sites.
In susceptible BALB/c mice, L. major parasites can disseminate within 24 hours from the site of infection in the footpad to the popliteal lymph nodes, spleen, liver, lungs and bone marrow [30], [59]. However, L. major parasites were not detected at early time points during IL81 infection (day 1 and day 3) but were detected at week 4, and were also detected at week 8 but not at week 3 during LV39 infection, suggesting that parasite dissemination may have occurred at a later stage of infection. Dissemination is inhibited by the administration of recombinant IL-12 and resistant mouse strains restrict the spread of the parasites [30]. While several susceptible mouse strains have been reported to show some increase in dissemination [60]–[62], disseminated parasite loads in CD11ccreIL-4Rα-/lox mice were unusually dramatic, with relatively higher parasite burdens in the spleens and footpads compared to other susceptible strains. Unexpectedly, parasites were even identified within the brain of some of the CD11ccreIL-4Rα-/lox mice. This suggests that the L. major parasites managed to cross the immunological blood-brain barrier, which has only rarely been reported for this cutaneous strain with very low levels of parasites detected [63]. However, dissemination of parasites to the central nervous system (CNS) has been frequently observed in visceral Leishmaniasis in both humans and dogs [64]–[67]. It has been suggested that parasites arrive in the CNS via infected leukocytes [65] and/or disruption to the blood brain barrier caused by inflammation [67]. Studying the mechanisms by which other pathogens, such as bacteria, invade the CNS may lend insights into Leishmania dissemination. Many intracellular organisms such as Mycobacterium tuberculosis, Listeria monocytogenes, Brucella spp. and Salmonella spp. appear to make use of the “Trojan-horse” mechanism, using phagocyte facilitated invasion for entry into the CNS [68]. After infection with intracellular pathogens, phagocytes undergo phenotypical changes, such as increased migratory activity and increased expression of adhesion molecules and proinflammatory cytokines, all of which could aid in dissemination and crossing of the blood-brain barrier [68], [69]. Whether infected phagocytes are recruited to the CNS by specific or non-specific means is unknown [69]. In order to determine which cells were infected by L. major, mice were infected with GFP-IL81 parasites and cell populations containing parasites were identified by flow cytometry.
At day 3 and 7 after infection, macrophages harbored L. major in the footpad, while pDCs and cDCs were found to be infected in the lymph node. Similar to other reports, this indicates that DCs were responsible for transporting parasites to the lymph node [70]. At week 4, L. major parasites were still detected in macrophages in the footpads, as well as in DCs and neutrophils, but in the LN they were primarily found in DCs. The number of infected DCs in both footpad and LN was significantly higher in CD11ccreIL-4Rα-/lox mice. A previous study also reported that DCs were the primary infected cell population in the draining LN of L. major infected mice [70]. DCs were also infected with L. major parasites in the spleen, with CD11ccreIL-4Rα-/lox mice again showing a greater number of infected DCs. Numbers of DCs infiltrating the LN and spleen were equivalent in both CD11ccreIL-4Rα-/lox mice and littermate controls during infection. This suggests that the increased survival and/or growth of parasites in DCs, as a consequence of significantly reduced DC iNOS production, was responsible for the increase in infected cell numbers in CD11ccreIL-4Rα-/lox mice. Interestingly, a recent study found that infected DCs, which are monocyte-derived CD11b+ inflammatory DCs expressing Ly6C, F480, Ly6G and iNOS, showed a unique ability to disseminate to peripheral sites in M. tuberculosis infection [71]. Furthermore, CD11b+Ly6C+ cells were found to be the principal phagocytic cells harboring L. monocytogenes in circulation [69], [72]. We hypothesize that dendritic cells may therefore play a role in disseminating L. major parasites to peripheral sites and that their killing effector responses could be important in controlling disease.
The reduced Th1 and increased Th2 responses in CD11ccreIL-4Rα-/lox mice suggests that instruction theory is relevant in vivo, and more importantly, that biological quantities of IL-4 acting through DCs can promote resistance to Leishmania infection. DCs from lymph nodes of CD11ccreIL-4Rα-/lox mice produced more IL-10 and less IL-12 than those from IL-4Rα-/lox mice. Quantification of mRNA expression also revealed interesting differences in DCs from CD11ccreIL-4Rα-/lox mice. Expression of the Th1-promoting genes for IL-12p40 and IL-18 was decreased compared to DCs from littermate control mice, while expression of the Th2-promoting genes for IL-23p19 and activin A were significantly increased. IL-23 production by DCs has been shown to promote Th17 [32], leading to increased neutrophils that enhance susceptibility to L. major by acting as Trojan horses [73]. Activin A is a pleiotropic cytokine belonging to the TGF-beta superfamily, and has previously been found to promote alternative activation of macrophages by inducing Arginase 1 and decreasing IFN-γ-induced expression of iNOS [33]. The absence of IL-4Rα signalling on DCs therefore appears to have a more complex influence on the dendritic cells than just affecting IL-12 production during cutaneous Leishmaniasis in vivo.
Dendritic cell instruction may not be restricted to Leishmaniasis, since other disease models have also demonstrated a protective role for IL-4. Experimental infections with Candida albicans in IL-4 deficient mice led to impaired development of Th1 responses [38], and a Th1 promoting effect of IL-4 has also been observed in autoimmunity [36], [40], [74], tumor immunity [39], [75], [76] and contact sensitivity reactions [41], [77]. There is also evidence to suggest that IL-4 may promote Th1 development in humans, since both human and mouse DCs produce increased levels of bioactive IL-12 after stimulation with IL-4 [20]. A similar effect was observed in human peripheral blood mononuclear cells (PBMCs) treated with IL-4 plus lipopolysaccharide or Staphylococcus aureus [78]. Incorporating exogenous IL-4 as an adjuvant for enhancing strong Th1 responses could therefore be utilised to boost vaccine efficiency against cutaneous Leishmaniasis. Accordingly, parallel studies examining the efficacy of IL-4 as an adjuvant during BMDC-mediated vaccination against L. major, found that IL-4 instruction of DCs was critical in eliciting protective immune responses [79]. The role of IL-4Rα signalling on DCs in eliciting immunity to other intracellular pathogens is therefore of interest to vaccination strategies, and an exciting avenue to be explored.
CD11ccre mice [27] were crossed with IL-4Rαlox/lox BALB/c mice [28] and complete IL-4Rα-/- BALB/c mice [15] to generate hemizygous CD11ccreIL-4Rα-/lox mice. Mice were backcrossed to a BALB/c background for 9 generations to generate CD11ccreIL-4Rα-/lox BALB/c mice. Hemizygous littermate controls (IL-4Rα-/lox) were used as controls in all experiments. Mice were genotyped as described previously [28]. All mice were housed in specific-pathogen free barrier conditions in individually ventilated cages. Experimental mice were age and sex matched and used between 8–12 weeks of age.
This study was performed in strict accordance with the recommendations of the South African national guidelines and University of Cape Town of practice for laboratory animal procedures. All mouse experiments were performed according to protocols approved by the Animal Research Ethics Committee of the Health Sciences Faculty, University of Cape Town (Permit Number: 009/042). All efforts were made to minimize suffering of the animals.
Genomic DNA was isolated from spleen DCs (CD11c+MHCII+) sorted using a FACS Vantage flow cytometer (BD Immunocytometry systems). Purity was determined by flow cytometry and checked by cytospin and staining with the Rapidiff Stain set (Clinical Diagnostics CC, Southdale, South Africa) and was at least 99%. A standard curve was prepared from serial 10-fold DNA dilutions of cloned IL-4Rα exon 5 and exon 8 DNA and RT-PCR was performed using the following primers; exon 5: forward 5′ AACCTGGGAAGTTGTG 3′ and reverse 5′ CACAGTTCCATCTGGTAT 3′, exon 8: forward 5′ GTACAGCGCACATTGTTTTT 3′ and reverse 5′ CTCGGCGCA CTGACCCATCT 3′.
The following antibodies were used for flow cytometry: SiglecF-PE, CD11c-APC, MHCII-APC, F480-PE, CD11b-FITC, CD3-FITC, CD19-PE, PDCA-APC, SiglecH-PE, CD11b-PE, CD11c-PE, CD4-PerCP, CD8-PE, GR-1-PE, CD3-PerCP, anti-CD124-PE, rat anti-mouse IgG2a-PE, CD11c-biotin, CD103-biotin, CD124-biotin and rat-anti-mouse IgG2a biotin with streptavidin-APC (all BD Bioscience, Erembodegem, Belgium) and MHCII-biotin with PerCP streptavidin (BD Bioscience). For intracellular cytokine staining, popliteal lymph node cells from L. major infected mice were seeded at 2×106 cells/well and stimulated at 37°C for 4 hours with phorbal myristate acetate (Sigma-Aldrich) (50 ng/ml), ionomycin (Sigma-Aldrich) (250 ng/ml) and monensin (Sigma-Aldrich) (200 µM) in DMEM/10% FCS. Dendritic cells were stained with CD11c-PE-Cy7 (BD Bioscience) and MHCII-APC, fixed and permeabilized, and intracellular cytokines were stained with anti-IL-10, anti-IL-12 and isotype controls (BD Bioscience) (all PE-labelled). Cells were acquired on a FACS Calibur machine (BD Immunocytometry systems, San Jose, CA, USA) and data were analyzed using Flowjo software (Treestar, Ashland, OR, USA).
BMDCs were generated from bone-marrow progenitors of CD11ccreIL-4Rα-/lox and littermate control mice using 200 U/ml recombinant mouse granulocyte-macrophage colony-stimulating factor (GM-CSF) (Sigma-Aldrich) as previously described [80]. On Day 10, non-adherent cells were harvested and 5×105 BMDCs were stimulated with LPS (Sigma-Aldrich; 1 µg/ml) or Leishmania major IL-81 promastigotes (MOI: 10 parasites/cell) in the presence or absence of 1000 U/ml recombinant mouse IL-4 or IL-13 (rIL-4/rIL-13, BD Biosciences) for 48 h. Following incubation, levels of IL-12p40, IL-12p70 and IL-10 were measured in culture supernatants by ELISA as previously described [15].
Cytokines in cell supernatants were measured by sandwich ELISA as previously described [15]. For antibody ELISAs, blood was collected in serum separator tubes (BD Bioscience, San Diego, CA). Antigen-specific IgG1, IgG2a and IgG2b were quantified by ELISA, as previously described [15]. Detection limits were 5 ng/ml for IgG1 and IgG2b and 0.1 ng/ml for IgG2a and IgG3. Total IgE was determined as described [15]. The detection limits was 8 ng/ml for total IgE.
L. major LV39 (MRHO/SV/59/P) and GFP-expressing L. major IL81 (MHOM/IL/81/FEBNI) (kind gift from Prof. Heidrun Moll, University of Würzburg, Germany) strains were maintained by continuous passage in BALB/c mice and prepared for infection as described previously [15]. Anaesthetised mice were inoculated subcutaneously with 2×106 or 2×105 stationary phase metacyclic promastigotes into the left hind footpad in a volume of 50 µl of HBSS (Invitrogen). Swelling of infected footpads was monitored weekly using a Mitutoyo micrometer calliper (Brütsch, Zürich, Switzerland).
Footpads, spleens and livers were fixed in 4% formaldehyde in phosphate buffered saline and embedded in wax. Tissue sections were stained with either haemotoxylin and eosin or Giemsa.
Following infection of mice with GFP-L. major IL81 parasites for 4 weeks, isolated brain tissue was immediately embedded in OCT (Tissue-Tek; Sakura, Zoeterwoude, Netherlands) medium. Pre-fixing of tissues was avoided to minimize background staining from the fixative. OCT-embedded brain tissue were cut into 10 µm frozen sections and mounted on 3-aminopropyltriethoxysilane-coated slides. Following acetone fixation of tissue, sections were stained with nuclear stain Hoechst. Coverslips were then mounted on sections using Mowiol 4–88 mounting medium (Calbiochem) with anti-fade (Sigma). Images were acquired and analyzed by Ziess LSM 510 confocal microscope (Jena, Germany).
Infected organ and tissue cell suspensions were cultured in Schneider's culture medium (Sigma). Prior to removal of mouse brain tissue for detection of parasite burden, animals were perfused with 20 ml sterile saline solution. Detection of viable parasite burden was estimated by two-fold limiting dilution assay as previously described [15].
CD4+ T cells were positively selected using anti-CD4 MACS beads (Miltenyi Biotec) according to the manufacturer's instructions (purity >95%). Thy1.2-labeled splenocytes were T cell depleted by complement-mediated lysis to enrich antigen presenting cells (APCs). APCs were fixed with mitomycin C (50 µg/ml, 20 min at 37°C) and washed extensively in complete IMDM. A total of 2×105 purified CD4+ T cells and 1×105 APCs were cultured with SLA (50 µg/ml). After 72 h incubation at 37°C, supernatants were collected and cytokine production analysed as previously described [28].
Muscle tissue was separated from infected footpads and digested in DMEM medium supplemented with Collagenase IV (Sigman-Aldrich; 1 mg/ml) and DNase I (Sigma-Aldrich; 1 mg/ml) at 37°C for 60 min. Following incubation, single cell suspensions were isolated by straining through 40 µM cell-strainers. Spleen cells were isolated by pressing through 70 µM cell-strainers, red blood cell lysis was performed and white blood cells were washed and resuspended in 10% DMEM (Gibco).
Total lymph node or footpad cells were labeled with specific mAbs and populations isolated by cell sorting on a FACS Vantage machine. Macrophages from the footpad were gated as CD11bhighMHCIIhighCD11c− cells and DCs, macrophages, neutrophils and B cells from the lymph node were gated as CD11chighMHCIIhigh, CD11bhighMHCIIhighCD11c−, GR-1highSSChighFSChighCD11c− and CD19+CD3−CD11c− cells, respectively. Cells were >98% pure and used for further analysis.
Dendritic cells were stained with specific mAb and sorted from the LN of infected mice. Total RNA was extracted from dendritic cells using Tri reagent (Applied Biosystems, Carlsbad, Calif) and mini-elute columns (Qiagen) according to the manufacturer's protocol. cDNA was synthesized with Transcriptor First Strand cDNA synthesis kit (Roche), and real-time PCR was performed by using Lightcycler FastStart DNA Master PLUS SYBR Green I reaction mix (Roche) on a Lightcycler 480 II (Roche). Primers for IL-12p40: forward 5′ CTGGCCAGTACACCTGCCAC 3′ and reverse 5′ GTGCTTCCAACGCCAGTTC 3′, IL-18: forward 5′ TGGTTCCATGCTTTCTGG 3′ and reverse 5′ TCCGTATTACTGCGGTTGT 3′, IL-10: forward 5′ AGCCGGGAAGACAATAACTG 3′ and reverse 5′ CATTTCCGATAAGGCTTGG 3′, IL-23p19: forward 5′ CAGCTTAAGGATGCCCAGGTT 3′ and reverse 5′ TCTCACAGTTTCTCGATGCCA 3′ and βA subunit (Activin A): 5′ GAGAGGAGTGAACTGTTGCT 3′ and reverse 5′ TACAGCATGGACATGGGTCT 3′. Values were normalized according to the expression of the housekeeping genes HPRT or rS12.
Lymph node and footpad cells collected at week 4 after infection were restimulated with LPS (Sigma-Aldrich; 10 ng/ml). Supernatants were collected at 48 hours for quantification of nitric oxide [81] while arginase activity was measured in cell lysates [81]. Expression of intracellular iNOS and arginase was analyzed in CD11bhighMHCIIhighCD11c− macrophages and CD11chighMHCIIhigh DCs by flow cytometry using rabbit anti-mouse iNOS (Abcam) with goat anti-rabbit PE (Abcam) and goat anti-mouse arginase (Santa Cruz Biotechnology) with donkey anti-goat PE (Abcam). Purified goat IgG and rabbit IgG were used as controls.
Data is given as mean ± SEM. Statistical analysis was performed using the unpaired Student's t test or 1-way Anova with Bonferroni's post test, defining differences to IL-4Rα-/lox mice as significant (*, p≤0.05; **, p≤0.01; ***, p≤0.001) unless otherwise stated. (Prism software: http://www.prism-software.com).
|
10.1371/journal.pbio.1000092 | Structural Studies of the Giant Mimivirus | Mimivirus is the largest known virus whose genome and physical size are comparable to some small bacteria, blurring the boundary between a virus and a cell. Structural studies of Mimivirus have been difficult because of its size and long surface fibers. Here we report the use of enzymatic digestions to remove the surface fibers of Mimivirus in order to expose the surface of the viral capsid. Cryo-electron microscopy (cryoEM) and atomic force microscopy were able to show that the 20 icosahedral faces of Mimivirus capsids have hexagonal arrays of depressions. Each depression is surrounded by six trimeric capsomers that are similar in structure to those in many other large, icosahedral double-stranded DNA viruses. Whereas in most viruses these capsomers are hexagonally close-packed with the same orientation in each face, in Mimivirus there are vacancies at the systematic depressions with neighboring capsomers differing in orientation by 60°. The previously observed starfish-shaped feature is well-resolved and found to be on each virus particle and is associated with a special pentameric vertex. The arms of the starfish fit into the gaps between the five faces surrounding the unique vertex, acting as a seal. Furthermore, the enveloped nucleocapsid is accurately positioned and oriented within the capsid with a concave surface facing the unique vertex. Thus, the starfish-shaped feature and the organization of the nucleocapsid might regulate the delivery of the genome to the host. The structure of Mimivirus, as well as the various fiber components observed in the virus, suggests that the Mimivirus genome includes genes derived from both eukaryotic and prokaryotic organisms. The three-dimensional cryoEM reconstruction reported here is of a virus with a volume that is one order of magnitude larger than any previously reported molecular assembly studied at a resolution of equal to or better than 65 Å.
| Mimiviruses are larger than any other known virus, yet despite their size, the capsid has been shown to be a regular icosahedron. Using cryo-electron microscopy and atomic force microscopy, we show that the icosahedral symmetry is only approximate, in part because one of the 5-fold vertices has a unique “starfish-shaped” feature and because a better three-dimensional reconstruction was obtained by assuming only 5-fold symmetry. Contrary to expectations, the arrangement of the capsomers on the Mimivirus surface is not as that in many other large icosahedral dsDNA viruses. Instead, the faces of Mimivirus have systematic vacant sites that are surrounded by six capsomers with alternative orientations which differ by about 60°.
| Mimivirus, Acanthamoeba polyphaga Mimivirus, is the largest known virus [1–3] and a putative human pneumonia agent [4]. It has an icosahedral shape with a 0.75-μm diameter [3] and a ∼1.2-Mbp genome that contains most of the genes found in small bacteria [5]. The external morphology of Mimivirus had initially led to its false identification as a bacterium [1,4]. Initial cryo-electron microscopy (cryoEM) studies [3] had shown that Mimivirus has a diameter of about 5,000 Å, with multiple layers of proteins and lipid membranes that surround a nucleocapsid. In addition, there is a dense layer of 1,250-Å-long fibers that cover the viral surface, making the total diameter of the particles about 7,500 Å. The outermost layer of the capsid is about 70 Å thick and corresponds to the major capsid protein (MCP). There is an irregularly shaped nucleocapsid, which itself is enveloped by a 70-Å-thick layer, and is separated from the capsid by a distance that varies from 300 to 500 Å. Thus, the large size of Mimivirus, its gene content, and its functional complexity as described here and elsewhere [2–6] stretch the definition of a virus [7].
The capsomer structures of some large double-stranded DNA (dsDNA) viruses—including adenovirus [8], Paramecium bursaria Chlorella virus 1 (PBCV1) [9,10], the bacteriophage PRD1 [11], Sulfolobus turreted icosahedral virus [12], and the marine bacteriophage PM2 [13]—have been determined by x-ray crystallography and shown to be similar. Although these viruses infect a wide variety of hosts covering the prokaryotic, eukaryotic, and archaeal domains of life, the similarity of their MCP structures suggest that they have, in part, evolved from a common precursor [9,12,14]. Each monomer in the trimeric capsomers consists of two successive “jelly-roll” folds, producing a pseudo-hexameric structure with a thickness of ∼75 Å and a diameter varying between 74 Å in PBCV1 [9,10] and about 85 Å in adenovirus [8]. One or other of the two jelly-roll motifs within the monomer often has a large insertion in the DE and FG loops (the β strands along the polypeptide of each jelly-roll are named A to H) (Figures 1 and 2), creating a “tower” on top of each of the three monomers within a capsomer. These towers give capsomers a triangular appearance on the surface while maintaining a pseudo-hexagonal shape below the towers, appropriate for packing into hexagonal arrays [8,15,16]. The Mimivirus MCP is homologous to the MCP of PBCV1 with 31% amino acid identity (Figure 2). Therefore, it is highly likely that the structure of Mimivirus capsomers are similar to the aforementioned capsomers in large dsDNA icosahedral viruses [9,12,14]. However, in Mimivirus, there are about 190 additional amino acids inserted into the DE loop of the second jelly-roll motif that are similar to the large tower insertions in adenovirus (Figures 1 and 2).
The forest of long fibers on the Mimivirus surface increases the ice thickness, creating difficulties for cryoEM [3]. The random scattering of the electrons by the additional ice thickness and by the disordered fibers reduced the signal-to-noise ratio. Here we report that we were able to partially overcome this problem by digesting the fibers with lysozyme and proteases. Both atomic force microscopy (AFM) and cryoEM were then used to analyze the structure of untreated as well as defibered Mimiviruses. The viral capsid surface was found to have a hexagonal array of depressions separated by about 140 Å. These were interpreted as systematic vacancies within a hexagonal array of double jelly-roll capsomers, accounting for the absence of one-third of all capsomers. Furthermore, the previously recognized starfish-like feature [17] was well-resolved and associated with a unique pentameric vertex on mature particles below the forest of surface fibers. We also show that the Mimivirus nucleocapsid has a defined shape surrounded by an envelope, which is separated from the viral capsid by a space whose size and dimensions are conserved in all particles.
A large number of cryoEM particle images were collected to improve the previously computed [3], icosahedrally averaged, three-dimensional reconstructions of Mimivirus. However, increasing the number of images beyond about 30,000 failed to show the anticipated hexagonal arrays of capsomers as found in PBCV1 and other related dsDNA viruses [12,18,19]. Hence, AFM was used in an endeavor to obtain better-resolved structural information. AFM images of defibered Mimivirus showed hexagonal arrays of depressions covering the surface of the virus (Figure 3A) and a starfish-like structure associated with one of the vertices on many of the particles (Figure 4), as had also been observed on some previous EM micrographs [17]. The presence of a structural feature on only one of the 12 vertices demonstrated a significant departure from icosahedral symmetry. Therefore, further reconstructions were based on only 5-fold, rather than icosahedral symmetry (see Materials and Methods), using about 700 lysozyme- and protease-treated defibered virus particles. This reconstruction clearly showed a unique pentameric vertex with a starfish-like attachment, but failed to visualize the hexagonal arrays of depressions seen on the AFM images. Because about 31,000 images of the mature fibered particles had been collected, a further reconstruction was calculated—using these particles and assuming only 5-fold symmetry—which was initialized with the newly reconstructed model from the defibered particles. The resultant 65 Å resolution cryoEM map of Mimivirus showed that surface depressions, separated by 140 Å, were arranged in hexagonal arrays (Figure 3B), which was consistent with the AFM observations. Each equilateral triangular face of the virion consisted of 19 rows of depressions parallel to each edge, with each row containing one less depression than the previous row.
The Mimivirus genome [5] contains four genes, including L425 and R441, that are homologous to the double jelly-roll PBCV1 Vp54, and to the MCPs of other large dsDNA viruses [14]. Although a homology model of the R441 gene product was built by Benson et al. [14], the actual MCP of Mimivirus was found to be the gene product of L425 [5]. The limited resolution of the cryoEM reconstruction barely resolves individual capsomers, but the array of large depressions suggests that these are missing capsomers (vacancies) in the hexagonal arrays of PBCV1-like capsomers. There is one 190-amino-acid-long insertion in the DE loop of the second jelly-roll along the polypeptide of the Mimivirus MCP (Figures 1 and 2), which is located on the external edge of each of the three monomers in a capsomer. The systematic vacancies in Mimivirus could arise as a consequence of steric conflict between these insertions in three neighboring capsomers and would be relieved by creating the systematic vacancies.
Each of the depressions on the Mimivirus surface is surrounded by six barely resolved triangular shapes (Figure 3C–3E), which are similar in appearance to the triangular external surface of capsomers in other viruses with double jelly-roll MCPs [8,15,16]. The orientations of neighboring trimeric capsomers surrounding each depression differ by about 60°, thus generating a 6-fold symmetry axis in the center of each depression (Figure 3C–3F). However, the trimeric shapes are barely resolved from each other so that each of the three “towers” that form the triangular shape at the top of a capsomer merges with the towers of the neighbouring capsomers (Figure 3G). A simulation using the known PBCV1 capsomer structure [9], assembled into hexagonal arrays as found for Mimivirus, demonstrated that the proposed arrangement mimics the observed pattern of depressions with poorly resolved surrounding trimeric capsomers at the resolution attained for the Mimivirus reconstruction (Figure 3F). Given that the distance between depressions is 140 Å, the center-to-center distance between adjacent triangular capsomers will be 81 Å (Figure 3G). This is in the range expected for trimeric capsomers assembled from double jelly-roll monomers [10].
To our knowledge, the arrangement of protein subunits in an icosahedral capsid was first discussed by Crick and Watson [20]. Their concepts were extended by Caspar and Klug, who suggested that arrays of hexagonal capsomers could be interspersed with pentameric capsomers at the icosahedral 5-fold vertices, resulting in only quasi-equivalent environments for monomers at the 5-fold vertex compared with those in hexagonal arrays [21]. The organization of pseudo-hexameric capsomers in large dsDNA icosahedral viruses, for which the triangulation number (T) expresses the number of jelly-rolls rather than monomers in the icosahedral asymmetric unit, is, therefore, a further extension of the concept of quasi-symmetry.
If all the depressions were filled by capsomers, and as there are 19 depressions between neighboring pentameric vertices, the coordinates of the nearest vertex would be h = 19 ± 1 and k = 19 ± 1, where h and k are the number of capsomers along the hexagonal axes of the array (Figure 3G). The uncertainty arises because it is not clear whether there is a depression or a capsomer on each pentameric vertex. Thus, the triangulation number, given by T = h2 + hk + k2 [21], would be 3 × (19 ± 1)2 or have one of nine possible values in the range of 972 ≤ T ≤ 1200. The previously predicted value of around 1,180 jelly rolls [3] was based on an estimate for the center-to-center distance between capsomers being 75 Å. The above observations show that this distance is 81 Å, which would have given T = 1,012, which is still within the range of the above determination. However, this statement further extends the definition of T, because it not only considers the depressions being filled by capsomers, but also tacitly assumes that all the capsomers are similarly oriented. The p6 plane group arrangement of capsomers in Mimivirus allows 3/2 times as much area per capsomer—compared with a completely filled hexagonal array of capsomers in a p3 plane group—as, for instance, in PBCV1 (Figure 3G). Thus, the actual number of jelly-rolls will be 2T/3, and the number of capsomers per icosahedral asymmetric unit will be T/9 or about 120 for Mimivirus.
The p6 plane group organization of the capsomers in Mimivirus is essentially the same as that of trimeric “packing units” observed by cryoEM for infectious bursal disease virus (IBDV) [22,23] which has a T = 13 (h = 1, k = 3) surface lattice. The structure of the IBDV major capsid protein has been determined [24] and shown to have three domains (B, S, and P), of which the S and P domains have jelly-roll folds. However, the domain organization within the IBDV trimeric capsomers is different to the pseudo-hexagonal capsomer structures found in PBCV1 and some other large dsDNA viruses. Thus, although the p6 organization of capsomers in Mimivirus resembles the capsid of a dsRNA virus, the amino acid sequence of the major capsid protein of Mimivirus has greatest similarity to other dsDNA viruses such as PBCV1.
CryoEM studies of Mimivirus recognized that some particles had a special vertex [3]. More recently, transmission electron microscopy (TEM) of sectioned Mimivirus-infected amoeba found a starfish-shaped feature associated with one vertex on many Mimiviruses [17]. Starfish-shaped density features were also observed with cryoEM on some fiberless immature Mimivirus particles that occurred in purified samples [17]. Here we show, using AFM, that a starfish-shaped feature can be seen on many defibered Mimivirus particles (Figure 4). Furthermore, the 5-fold–averaged cryoEM reconstruction of Mimivirus was initiated with a simplified model (see Materials and Methods) that did not have a starfish-shaped feature. However, the resulting reconstruction (Figure 5) showed a starfish-shaped feature similar to what was observed with AFM (Figure 4C) and confirmed the existence of the starfish-shaped feature on each virus. The 5-fold–averaged cryoEM results showed that the arms of the starfish have a thickness of about 400 Å, a width of about 500 Å, and extend about 2,000 Å almost all the way towards the neighbouring 5-fold vertices. The exceptional clarity of the starfish-shaped feature in cryoEM reconstructed map (Figure 5) demonstrated that it must exist on almost every fibered particle. Both AFM and cryoEM showed that the arms of the starfish are inserted and open a gap between the neighbouring faces that are associated with the special vertex (Figure 4D). The five faces associated with the special vertex are inclined by about 5° to what would be expected if the virus were completely icosahedral, accounting for the gap between faces (Figure 5E). The arms of the starfish-shaped feature do not show the hexagonal arrays of depressions (Figure 4D), suggesting that the starfish-like feature is not assembled from the MCP. Evidence for the starfish-shaped feature being a separate entity was also found in cryoEM images of defibered Mimivirus samples in which there were objects that had five arms of appropriate size radiating from a common center (Figure 6A).
CryoEM images of thin sectioned samples [17] and AFM images of mature Mimivirus (Figure 4B) showed that there are star-shaped crevices between the long surface fibers, implying that the starfish-shaped feature is not covered by fibers. It had been suggested that the “starfish”-associated vertices might be the portal for DNA release based on its location further from the associated virus factory [17,25]. Thus, if the long cross-linked fibers of Mimiviruses [3] were to cover the complete viral surface, they would be an obstacle for genome delivery into a host. However, the star-shaped crevice between the fibers could provide an exit portal for the genome.
Scanning electron microscopy [17], traditional TEM of thin sections [25], and cryoEM studies (Figure 6B) show that defibered particles missing the starfish-shaped feature are associated with membrane-like “puffs” at their special vertices. Furthermore, cryoEM showed that particles that had lost their genome (Figure 6C) had also lost the starfish-shaped feature. In addition, AFM showed that the ejected DNA is unprotected by any surrounding proteins (Figure 6D). Thus, the starfish-shaped feature might be acting as a seal to hold together the five faces associated with the special vertex. Therefore, the first step of genome delivery would be the release of the starfish-shape feature, allowing the DNA to exit through the special vertex. Special vertices for genome delivery have also been observed in some other large dsDNA viruses [26,27], in tailed bacteriophages [28–31], and in herpes virus [32]. The presence of a special vertex in tailed bacteriophages or in herpes virus whose MCPs have a HK97-like fold [33] or in viruses that have a double jelly-roll fold in their capsids, suggests convergent evolution to a common solution for genome delivery.
AFM images show that a number of external fibers of Mimivirus are frequently attached to a single central feature at one end with their free end being associated with a globular terminus (Figure 7A and 7B). However, there is no indication where the fibers attach to the capsid on the viral surface. The surface fibers are resistant to proteases unless first treated with lysozyme, suggesting that the fibers are protected by peptidoglycan (as previous suggested [5]), which is consistent with Mimivirus being Gram-positive [1,4]. CryoEM images of Mimivirus that had been partially treated with bromelain show successive rings of density on the fibers separated by 200–500 Å, representing different structural segments along their lengths (Figure 7C). AFM images show murky material surrounding the fibers (Figure 7B) that might be peptidoglycan cross-linking neighboring fibers. Fibers with peptidoglycan components perhaps act as a decoy for attracting amoeba [34].
The central slice of the cryoEM reconstruction, perpendicular to the unique 5-fold axis, showed that the genome is surrounded by a membrane-like envelope (Figure 5F). A central slice, containing the unique 5-fold axis, showed that the nucleocapsid had a concave depression facing the “starfish”-associated vertex (Figure 5E), which suggests a specialized organization that might be required for host infection. The clarity of these features after five-fold averaging implies that the nucleocapsid has a defined shape and also a fixed position relative to the external capsid. Unlike many other viruses in which the genome is closely surrounded by the capsid, Mimivirus has a 300–500 Å gap between the enveloped genome and the outer capsid. Thus, there must be supports across the gap that accurately position the genome relative to the viral capsid and internal membrane, although apparently they are too few or lack symmetry to make them visible in the cryoEM reconstruction. Long internal fibers were observed by AFM after applying mechanical force to the virus that broke the outer capsid layers (see Materials and Methods). These internal fibers have a diameter of about 60 Å with repeat units at intervals of about 70 Å (Figure 7D). The nucleocapsid might be supported by these fibers but, at this time, there is no further evidence for this suggestion.
The enveloped genome within the larger viral capsid, perhaps supported by fibers (Figure 7D), has some similarity to eukaryotic cells. In contrast, the external peptidoglycan component mimics bacterial cell walls (Figure 7A–7C). In addition, the existence of a unique vertex in Mimivirus, possibly for genome delivery [17,25], is reminiscent of tailed bacteriophages. These observations are consistent with other results [2,35], implying that Mimiviruses and some other large icosahedral dsDNA viruses have gathered genes from eukaryotic, prokaryotic, as well as archaeal origins.
The three-dimensional cryoEM reconstruction reported here, which was made possible in part by relaxing the icosahedral symmetry, is of a virus whose volume is an order of magnitude larger than has previously been reported. Thus, the detection of a unique vertex may have been missed in other structural studies in which strict icosahedral symmetry had been imposed [36].
The Mimivirus fibers were digested by sequential application of lysozyme and bromelain. The Mimivirus was pelleted by centrifuging at 1,000g for 30 min. Each volume of pelleted virus was incubated with four volumes of 10 mg/ml lysozyme in TES buffer (0.05 M N-[Tris(hydroxymethyl)methyl]-2-aminoethanesulfonic acid, pH 7.5, 0.01% NaN3) at room temperature for at least one day. The sample was washed twice with TES and digested with five volumes of 14 mg/ml bromelain from pineapple stem (Sigma) in TES buffer (0.035M TES, pH = 7.5, 0.3M KCl, 0.02M DTT) at room temperature for at least one day.
CryoEM data of untreated and defibered Mimivirus were collected as described previously [3]. Micrographs were scanned on a Nikon Coolscan 9000 with a final pixel size of 15.9 Å. The cryoEM reconstruction was performed assuming 5-fold symmetry using programs FREALIGN [37] and a modified version of XMIPP [38] (V.A. Kostyuchenko et al., unpublished data). The reconstruction was initiated with a model in which the density of the five faces around one pentameric vertex were pushed outwards along the associated icosahedral 5-fold axis by about 300 Å. Of a total of 1,378 boxed defibered Mimivirus particles, 691 were selected to produce a map of 120 Å resolution. The resolution was determined using a Fourier shell correlation threshold of 0.5. The map shows a clear starfish-shaped feature. This map was used as a starting model for reconstruction of untreated, fibered Mimivirus. Of a total of 53,640 boxed fibered Mimivirus particles, 30,919 were selected to achieve a 5-fold-averaged reconstruction with a resolution of 65 Å. The cryoEM map has been deposited with the EBI and has been given the accession number of EMDB 10623.
Mimivirus particles, both native and those treated with enzymes, were spread on freshly cleaved mica that was coated with poly-l-lysine and scanned under buffer. Capsids, which were pretreated with lysozyme and bromelain, were, in some experiments, further exposed to 1 mg/ml solutions of proteinase K and 1% SDS at 37 °C for 30 min to 2 h, washed with water, and then imaged. No fixation of any kind was used. Two methods were used to expel the DNA and other internal structures from the virus. In the first method, virus solution was dried on mica, rehydrated with a small amount of water, and then pressed between two surfaces of mica. In the second method, very concentrated virus solution was placed in small wells, crushed with a glass stick, diluted in water, and then deposited on mica. The Mimivirus DNA was recognized by comparing the AFM images with DNA extracted from PBCV1 [16], T4 phage, vaccinia viruses [39], plasmid DNA, and thymus DNA. All these images had the same tangled appearance and had the same height (thickness) above substrate. Furthermore, the material could not come from the host or some other source, because almost all the images show these fibers to be closely associated with isolated virions, not just spread out randomly on the substrate.
AFM analysis was carried out using a Nanoscope III multimode instrument (Veeco Instruments). Samples were scanned at 25 °C using oxide-sharpened silicon nitride tips in a 75-μl fluid cell containing buffer or in air. For scanning in air, silicon tips were used. The images were collected in tapping mode [40] with an oscillation frequency of 9.2 kHz in fluid and 300 kHz in air, with a scan frequency of 1 Hz. Procedures were fundamentally the same as described for previous investigations of viruses [16,41]. In the AFM images presented here, height above substrate is indicated by increasingly lighter color. Thus, points very close to the substrate are dark and those well above the substrate are white. Because lateral distances are distorted due to an AFM image being a convolution of the cantilever tip shape with the surface features scanned, quantitative measures of size were based either on heights above the substrate or on center-to-center distances on particle surfaces. The AFM instrument was calibrated to the small lateral distances by imaging the 111 face of a thaumatin protein crystal and using the known lattice spacings [42] as standard.
|
10.1371/journal.ppat.1002722 | Polydnavirus Ank Proteins Bind NF-κB Homodimers and Inhibit Processing of Relish | Recent studies have greatly increased understanding of how the immune system of insects responds to infection, whereas much less is known about how pathogens subvert immune defenses. Key regulators of the insect immune system are Rel proteins that form Nuclear Factor-κB (NF-κB) transcription factors, and inhibitor κB (IκB) proteins that complex with and regulate NF-κBs. Major mortality agents of insects are parasitoid wasps that carry immunosuppressive polydnaviruses (PDVs). Most PDVs encode ank genes that share features with IκBs, while our own prior studies suggested that two ank family members from Microplitis demolitor bracovirus (MdBV) (Ank-H4 and Ank-N5) behave as IκB mimics. However, the binding affinities of these viral mimics for Rel proteins relative to endogenous IκBs remained unclear. Surface plasmon resonance (SPR) and co-immunoprecipitation assays showed that the IκB Cactus from Drosophila bound Dif and Dorsal homodimers more strongly than Relish homodimers. Ank-H4 and –N5 bound Dif, Dorsal and Relish homodimers with higher affinity than the IκB domain of Relish (Rel-49), and also bound Relish homodimers more strongly than Cactus. Ank-H4 and –N5 inhibited processing of compound Relish and reduced the expression of several antimicrobial peptide genes regulated by the Imd signaling pathway in Drosophila mbn2 cells. Studies conducted in the natural host Pseudoplusia includens suggested that parasitism by M. demolitor also activates NF-κB signaling and that MdBV inhibits this response. Overall, our data provide the first quantitative measures of insect and viral IκB binding affinities, while also showing that viral mimics disable Relish processing.
| Central to the study of host-pathogen interactions is understanding how the immune system of hosts responds to infection, and reciprocally how pathogens subvert host defenses. In the case of insects, understanding of how the immune system responds to infection greatly exceeds understanding of pathogen counterstrategies. Parasitoid wasps are key mortality agents of insects. Thousands of wasp species have also evolved a symbiotic relationship with large DNA viruses in the family Polydnaviridae whose primary function is to deliver immunosuppressive virulence genes to the insect hosts that wasps parasitize. The function of most PDV-encoded virulence genes, however, remains unknown. In this article, we investigated the function of two ank gene family members from Microplitis demolitor bracovirus (MdBV). Our results indicate that Ank-H4 and Ank-N5 function as mimics of IκB proteins, which regulate a family of transcription factors called NF-κBs that control many genes of the insect immune system. IκBs and NF-κBs also function as key regulators of the mammalian immune system. Our results thus suggest that viral Ank proteins subvert the immune system of host insects by targeting conserved signaling pathways used by a diversity of organisms.
| The innate immune system defends insects against a diversity of potential pathogens [1]. As part of this system, the Toll and Imd pathways activate Nuclear Factor-κB (NF-κB) transcription factors, which regulate the expression of antimicrobial peptides (AMPs) and many other genes [2]–[6]. Both pathways have also been implicated in defending insects against both microbes (viruses, bacteria, fungi, protozoans) and multicellular parasites (nematodes, parasitoid wasps) [2], [7]–[13].
All NF-κBs are homo- or heterodimers of Rel proteins, which share a Rel homology domain (RHD) essential for dimerization and DNA binding [14]. In the absence of immune challenge, most NF-κBs form inactive complexes with Inhibitor κB (IκB) proteins that bind the RHD through an ankyrin repeat domain (ARD) [14]–[16]. In Drosophila melanogaster, activation of the Toll pathway by pathogen recognition signals causes NF-κBs comprised of Dif and/or Dorsal to dissociate from the IκB Cactus and translocate to the nucleus [17]–[19]. Activation of the Imd pathway in contrast induces caspase 8-mediated cleavage of the compound protein Relish (Rel-110), which results in its N-terminal, RHD-containing fragment (Rel-68) forming NF-κBs that translocate to the nucleus, and its C-terminal IκB fragment (Rel-49) remaining in the cytoplasm [20], [21].
Reciprocally, pathogens often evolve sophisticated counterstrategies for overcoming host immune defenses [22]–[28]. Among insects, thousands of parasitoid wasp species depend upon large DNA viruses in the family Polydnaviridae to parasitize hosts [28]. All parasitoid wasps lay their eggs into or on the body of another arthropod (the host), and their offspring develop by feeding on host tissues. Most polydnavirus (PDV)-carrying wasps parasitize larval stage Lepidoptera (moths and butterflies), with each wasp species carrying a genetically unique PDV and naturally parasitizing only one or a small number of host species [29]. PDVs persist in wasps and are transmitted to offspring as stably integrated proviruses [28]. Replication in contrast only occurs in the reproductive tract of females where virions accumulate to high densities. Wasps inject a quantity of these virions into hosts when laying eggs, which rapidly infect hemocytes, the fat body, and other tissues. Viral gene products thereafter prevent host immune defenses from killing the wasp's progeny, yet no viral replication occurs in hosts because the encapsidated form of the viral genome lacks essential genes required for virion formation [28], [30], [31]. PDVs are thus beneficial symbionts of wasps that function as replication-defective vectors for delivery of virulence genes to hosts.
Microplitis demolitor parasitizes the non-model lepidopteran Pseudoplusia includens and carries M. demolitor bracovirus (MdBV) whose 190 kb genome encodes 51 genes for proteins larger than 100 amino acids [32]. Most of these genes are expressed in P. includens hemocytes and fat body within 2 h of infection [33], and functional studies implicate several of these genes in disrupting encapsulation, phagocytosis, and melanization [34]–[38]. Some of these genes also belong to a multimember family called ank genes that share an IκB-like ARD but lack the phosphorylation and ubiquitination domains that regulate the dissociation and degradation of insect IκBs after immune challenge [32], [39]. Comparative genomic data indicate that Rel proteins and other components of the Toll and Imd pathways are conserved among insects including Lepidoptera. However, in the absence of any data on NF-κB/IκB binding interactions in P. includens, we previously used Drosophila Rel proteins to assess whether MdBV Ank proteins function as IκB mimics. Co-immunoprecipitation experiments indicated that two family members, Ank-H4 and Ank-N5, complex with Dif, Dorsal, and Relish. Gel shift assays further showed that these Ank proteins prevent NF-κBs containing Dif or Dorsal from binding to the κB site in the drosomysin promoter and also prevent NF-κBs containing processed Relish from binding to the κB site in the cecropinA1 promoter [39]. Taken together, these findings indicate that Ank-H4 and –N5 disrupt both Toll and Imd pathway signaling. However, these data provide no insight on the relative affinity of these Ank proteins for different Rel protein dimers in comparison to endogenous IκBs. They also provide no insight on whether Ank proteins disable Imd signaling by disabling Relish function before or after processing. Here, we show that Ank proteins compete with endogenous IκBs for binding to Relish, block processing of Rel-110, and reduce the expression of AMP genes regulated by the Imd pathway. Our results also reveal that M. demolitor induces the expression of AMP genes in P. includens that are likely regulated by NF-κB signaling, but MdBV inhibits this response.
Rel proteins from mammals require the N-terminal RHD plus a downstream NLS for IκB binding [15], [16], [40]. In contrast, neither dimerization nor NF-κB/IκB binding requires any post-translational modifications or regions outside the RHD and NLS [16], [41]–[50]. We therefore used E. coli to express truncated forms of Dif, Dorsal, and Relish from Drosophila that contained the RHD, NLS and 20 additional C-terminal residues with a C-terminal StrepTagII tag (Figure 1A). These products were used in surface plasmon resonance (SPR) assays. We also produced truncated Rel proteins as N-terminal thioredoxin fusion constructs where the increased size allowed us to more easily distinguish them from IκBs in co-immunoprecipitation experiments (Figure 1A). Since only the ARD is required for IκB binding to NF-κBs [15], [16], we expressed a truncated form of Cactus that consisted of its ARD plus an N-terminal His tag (Figure 1A). In the absence of any information about binding of the IκB domain of Relish (Rel-49), we expressed a full-length version of Rel-49 with an N-terminal StrepTagII tag, and a C-terminal His tag (Figure 1A). Since MdBV Ank-H4 and –N5 consist of only an ARD [39], we expressed full-length versions of these proteins with N-terminal His tags (Figure 1A). Proteins were purified to greater than 90% purity as measured by loading at least 15 µg of protein on SDS-PAGE followed by Coomassie staining. Loading 1 µg of each recombinant protein on SDS-PAGE gels followed by Coomassie staining also confirmed that their size fully agreed with predicted masses (Figure 1B). The quaternary state of each purified recombinant protein was also analyzed by gel filtration, which as expected showed that each Rel protein formed homodimers as determined by comparison with molecular mass standards.
Understanding of NF-κB/IκB binding interactions derives primarily from the study of mammalian Rel (p65, RelB, c-Rel, and the compound proteins p100, and p105) and IκB (IκBα, IκBβ, IκBε, Bcl-3, and C-terminal domains of p105 and p100) family members. This literature indicates that Rel proteins form different homo- and heterodimers and that IκB family members exhibit a gradient of binding preferences for different Rel complexes. For example, IκBα and IκBβ preferentially bind p50-p65 and p50-c-Rel heterodimers, IκBε binds homo and heterodimers containing p65, and Bcl-3 binds p50 and p52 homodimers [40], [45], [48], [51], [52]. Although the IκB domain of p105 binds its corresponding Rel domain (p50) after cleavage, it remains unclear whether binding occurs as part of a compound protein, after cleavage, or both [53]. Current understanding of NF-κB/IκB binding interactions in insects in contrast is both more limited and restricted to family members from Drosophila. Co-immunoprecipitation experiments and transgenic assays indicate that Dif, Dorsal, and Relish form all combinations of homo- and heterodimers [4], [54], [55]. Dif and Dorsal co-immunoprecipitate Cactus [18], [19], [56], but it remains unknown whether Cactus binds Rel protein dimers containing Relish. It also remains unknown whether unprocessed Relish (Rel-110) or its C-terminal IκB domain (Rel-49) bind any Rel protein dimer [57].
Given the limited literature for insects, we conducted SPR assays that measured binding of Ank-H4, Ank-N5, Cactus, and Rel-49 to Dif, Dorsal and Relish homodimers. We also used the kinetic titration method to determine kinetic and thermodynamic constants and circumvent potential problems with non-specific binding and regeneration [58]. Recombinant IκB or Rel homodimers were immobilized on CM5 chips by amine coupling, followed by five sequential injections of doubling concentrations of a given Rel protein or IκB, which served as the analyte. We then generated sensograms by subtracting the response of a reference cell with no IκB from the response of the cell with the immobilized IκB. Our results indicated that each Ank and IκB bound the three Rel protein homodimers we assayed with the exception of Ank-N5, which did not bind Dif (Table 1, Figure 2). The strongest binding interaction we measured was between Cactus and Dif with a Kd of 110 nM, (Table 1). This value reflected a modest association rate (ka = 5.25×104 M−1 s−1) and a very slow off rate (kd = 5.8×10−3 s−1). The affinity of Cactus for Dorsal (Kd = 195 nM) was slightly lower than for Dif and was much lower for Relish (Kd = 2.19 µM). Rel-49 modestly bound Dorsal (Kd = 783 nM) but weakly bound Relish (Kd = 2.75 µM). Compared to Cactus, Ank-H4 displayed a much higher binding affinity for Relish (Kd = 345 nM) and lower binding affinities for Dif (Kd = 581 nM) and Dorsal (Kd = 858 µM). With the exception of Dorsal, Ank-H4 also displayed higher binding affinities for each Rel homodimer than Ank-N5 (Table 1).
Since the strongest binding interaction was between Cactus and Dif, we assessed whether traditional co-immunoprecipitation assays yielded similar trends by adding 3-fold molar excess of Cactus to Dif, Dorsal, and Relish followed by addition of an anti-thioredoxin antibody and protein A beads. Our results indicated that Cactus bound each Rel protein under these conditions, while dilution experiments suggested that Cactus bound Dif and Dorsal more strongly than Relish (Figure 3A). We then asked whether recombinant Ank-H4, Ank-N5, or Rel-49 could compete the binding of Cactus to different Rel homodimers. Rel-49, Ank-H4, and Ank-N5 could not compete the binding of Cactus to Dif or Dorsal under our reaction conditions when present at 200-fold molar excess (data not shown). In contrast, 15-fold molar excess of Ank-H4 reduced Cactus binding to Relish, and fully competed the binding of Cactus to Relish when present at 90-fold molar excess (Figure 3B). Despite exhibiting a lower binding affinity for Relish than Cactus or Rel-49 in our SPR assays, Ank-N5 also competed with Cactus for binding to Relish above 40-fold molar excess (Figure 3C). Rel-49 in contrast did not compete the binding of Cactus to Relish over the same range of concentrations (data not shown). Overall, these data indicated that Cactus bound Dif and Dorsal homodimers more strongly than Relish homodimers. They also indicated that Ank-H4 and –N5 bound each Rel homodimer with higher affinity than Rel-49, and bound homodimeric Relish more strongly than Cactus.
As previously noted, gel shift assays showed that Ank-H4 and –N5 inhibited binding of both Dif/Dorsal-containing NF-κBs to the to the κB site in the promoter of the drosomysin gene and Relish-containing NF-κBs to the κB site in the promoter of the cecropinA1 gene [39]. These findings together with the preceding binding studies collectively suggest that Ank-H4 and –N5 disable Toll and Imd pathway signaling by binding to Dif/Dorsal- and Relish-containing NF-κBs. However, these data do not indicate which form of Relish Ank proteins interact with in vivo. Insect IκBs are thought to primarily bind NF-κBs in the cytoplasm of cells [1]. Studies from mammals, however, yield a more complicated picture with some IκB family members primarily localizing and binding Rel proteins in the cytoplasm (IκBε others binding NF-κBs in the cytoplasm and nucleus (IκBα and ß), and others still preferentially localizing to the nucleus and binding NF-κBs bound to DNA (Bcl-3) [45], [48], [59]. Thus, viral Ank proteins could bind compound Relish (Rel-110) in the cytoplasm, processed Relish (Rel-68) in the nucleus, or both.
We therefore transfected the expression constructs pIZT/Ank-H4, pIZT/Ank-N5, or pIZT (empty vector control) into Drosophila mbn-2 cells that have a functional Imd pathway. This pathway is also activated by commercial LPS which contains PGN [3], [20], [57]. We then prepared cytosolic and nuclear extracts from resting-state and LPS/PGN-challenged cells, followed by SDS-PAGE and immunblotting using an anti-V5 antibody to detect each Ank protein, and antibodies that detected the cytoplasmic protein ß Tubulin and nuclear protein Histone H1. Similar quantities of Ank-H4 were detected in the cytoplasmic and nuclear fractions of cells prior to (0 min) and 60 min after LPS/PGN challenge (Figure 4A). We also detected Ank-N5 in both fractions although its abundance was greater in the nuclear fraction (Figure 4A). The presence of these viral proteins in both fractions, however, was not due to sample preparation because we only detected ß Tubulin in our cytoplasmic fractions and Histone H1 in our nuclear fractions (Figure 4A).
Using total cell extracts and an anti-Rel-68 antibody [20], time course experiments showed that control and Ank protein-expressing cells contained full-length Relish (Rel-110) but little or no processed Relish (Rel-68) prior to LPS/PGN challenge (Figure 4B). We also detected a second 100 kDa band, which based on earlier studies corresponded to a full-length Relish variant (Rel-100) with a shorter N-terminus [20], [60]. Thereafter, we detected processing of Rel-110/-100 in control cells 5 min after exposure to LPS/PGN as evidenced by the appearance of Rel-68 (Figure 4B). In contrast, we detected no processing of Rel-110/-100 in cells expressing Ank-H4 or Ank-N5 over a 90 min assay period (Figure 4B). Examination of cytoplasmic and nuclear extracts from control cells at 0 and 60 min post-exposure to LPS/PGN confirmed that Rel-110/-100 remained in the cytoplasm, whereas Rel-68 was detected in both the cytoplasm and nucleus (Figure 4C). Rel-110/-100 also remained cytoplasmic in cells expressing Ank-H4 and –N5 (Figure 4C).
Combined with our SPR and co-immunoprecipitation data, these findings suggested that Ank proteins bind Rel-110/-100 in the cytoplasm, which in turn blocks formation and translocation of Rel-68 to the nucleus. An alternative explanation, however, could be that Ank proteins directly or indirectly inhibit the processing enzyme Dredd, which is a caspase-8 homolog [20], [21], [61], [62]. We compared Dredd activity in control cells and cells expressing Ank-H4 and –N5 using the substrate Ac-LETD-pNA. We readily detected caspase-8 activity but no differences were detected among treatments, which suggested that Ank proteins did not affect Relish processing activity (Figure S1).
In addition to cecropinA1, other AMP genes activated by the Imd pathway and/or the Imd and Toll pathways include diptericin, metchnikowin and defensin [63]–[66]. To assess whether Ank proteins also reduced the expression of these read-out genes, we transfected mbn-2 cells with the aforementioned Ank expression constructs and then measured transcript abundance of each AMP gene after LPS/PGN challenge. As expected, transcript abundance of diptericin and metchnikowin increased greatly and defensin increased modestly in control cells transfected with the empty vector. However, transcript abundance of each AMP increased significantly less in cells expressing Ank proteins (Figure 5).
As previously noted, our decision to use Drosophila Rel proteins as binding targets for MdBV Ank proteins was driven by a lack of functional data on NF-κB/IκB binding interactions in Lepidoptera generally and the natural host of M. demolitor (P. includens) in particular. We likewise used bacterial cell wall components (LPS/PGN) as an elicitor and AMP gene expression as read-outs in the preceding experiments, because the former is a known activator of the Imd pathway while the latter are well-characterized target genes. Of obvious interest though is whether these findings are relevant to natural parasitism. In the absence of MdBV infection, P includens mounts a potent immune response against M. demolitor that culminates in the encapsulation and death of wasp eggs 24–36 h after parasitism [34], [37], [67]. The pattern recognition receptors (PRRs) that recognize parasitoid wasps are unknown from any insect including P. includens, and it also remains unclear whether parasitism activates NF-κB signaling. Studies in the silkmoth Bombyx mori, however, indicate that bacterial elicitors induce the expression of AMP genes including cecropin B1 and lebocin 4. Similar to Drosophila, the ortholog of Relish (BmRel2) from B. mori also binds κB sites in the promoters of these and other AMP genes [68]–[70]. We also previously identified cecropin and lebocin orthologs from P. includens and showed that immune challenge by heat-killed bacteria induces their expression [71].
Taken together, these data suggest that cecropin and lebocin are potential read-out genes for activation of the Imd pathway in Lepidoptera. We therefore asked if MdBV infection disrupts cecropin and lebocin expression in P. includens after immune challenge by bacteria. Consistent with prior results, transcript abundance of both AMPs rapidly increased in the fat body of P. includens larvae following bacterial challenge relative to our wounding control (Figure 6). Bacterial challenge, however, did not induce the expression of these AMP genes if larvae had been infected 12 h earlier with a physiological dose of MdBV (Figure 6). We then assessed whether immune challenge by M. demolitor eggs and/or MdBV itself also induced the expression of these AMP genes. Similar to bacteria, wasp eggs and inactivated MdBV strongly stimulated the expression of lebocin, while pre-infection with MdBV near fully disabled this response (Figure 7). In contrast, wasp eggs and inactivated MdBV did not induce the expression of cecropin.
Vertebrate pathogens produce several virulence factors that target the innate immune system of hosts by mimicking proteins with essential signaling functions [23], [25], [26], [72], [73]. In some cases these mimics derive from host genes that the pathogen acquired and modified, while in others they share no significant homology with host proteins but through convergence have evolved similar structural features summarized by [26], [73]. Many invertebrate pathogens also subvert host immune defenses but in most cases the identity, function and origins of the virulence factors involved remain unknown [1], [24], [25], [74], [75].
NF-κB signaling is a key potential target for immune subversion in insects because the Toll and Imd pathways are widely conserved, respond to a diversity of infectious organisms, and regulate large numbers of immune genes [69], [76], [77]. Parasitoid wasps are among the most important mortality agents of insects, and more than 40,000 of these wasp species depend upon symbiotic PDVs for successful parasitism of hosts [78]. Strikingly, almost all PDV isolates studied to date encode ank genes [28], [31], while our own previous studies with MdBV indicated at least some ank genes function as IκB mimics [39].
Here we report the first kinetic measurements of insect IκB/NF-κB binding. Consistent with our own and earlier co-immunoprecipitation data [18], [19], [56], our SPR results indicate that Cactus binds the RHD and NLS domains of Dif and Dorsal with higher affinity than the same domains from Relish. Our SPR results also indicate that Ank-H4 binds Relish, Dif and Dorsal homodimers with similar affinity, while our competition experiments indicate that Ank-H4 and –N5 compete with endogenous IκBs to Relish. Among vertebrate family members, detailed kinetic studies have been conducted with recombinant IκBα and its NF-κB binding partners (p50/p65 or p65/p65) using SPR, isothermal titration calorimetry (ITC), and fluorescence polarization competition assays [40], [52], [79]. As we observe, these studies reveal very low dissociation rates for IκBα/NF-κB complexes, which are consistent with the long half-life these complexes exhibit in vivo [40]. The Kd values we determined are also broadly similar to those determined for IκBα/NF-κB homodimers (3–180 nM) but much lower Kd values have been determined for IκBα/NF-κB heterodimers (30–40 pM) than we detected. This suggests the possibility that binding interactions between IκBs and NF-κB homodimers may be weaker than those between IκBs and NF-κB heterodimers. However, the aforementioned vertebrate studies also indicate that the strength of IκB/NF-κB binding interactions in vitro is highly sensitive to salt concentration, temperature, and other testing conditions. Thus, the conditions we used in our binding studies could also be suboptimal, which could also explain why Ank-N5 competed Cactus binding to Relish in vivo but exhibited lower binding affinities for Relish than Cactus in vitro.
Our finding that Cactus most strongly binds Dif and Dorsal homodimers is fully consistent with the known role of these Rel proteins in regulating Toll signaling. In contrast, the significance of Cactus also binding Relish is less clear. Prior studies indicate that Relish processing is not be affected by RNAi knockdown of Cactus [80]. However, Relish does co-immunoprecipitate with Dif and Dorsal as a presumptive heterodimer, which form after processing [4], [55]. Given evidence from crystal structures of mammalian IκB/NF-κB complexes that IκBs contact both members of the dimer [15], it is thus possible that Cactus is functionally important in regulating Dif-Relish or Dorsal-Relish heterodimers. Another interesting feature of our binding data in regard to endogenous IκBs is that Rel-49 binds Dorsal, Dif and Relish homodimers, which parallel studies from mammals indicating that the IκB domain of p105 also binds its corresponding Rel domain after cleavage [50]. In contrast, it was not technically possible for us to assess whether Rel-49 also binds the Rel domain of Relish prior to processing. Thus, further studies will be required to understand the importance of Rel-49 in regulating the activity of compound versus processed Relish. Additional studies will also be needed to measure and understand the binding affinities of Cactus and Rel-49 for the Rel protein heterodimers that form in vivo [4], [15].
Our previous results [39] together with the binding data of this study collectively indicate that Ank proteins suppress both Toll and Imd pathway signaling by binding to Dif, Dorsal, and Relish-containing NF-κBs. Results of the current study further reveal that Ank proteins localize to both the cytoplasm and nuclei of mbn-2 cells, and disrupt Imd signaling by blocking processing of Rel-110/-100 in the cytoplasm rather than by interfering with Rel-68 in the nucleus. Notably, the inhibitory activities of vertebrate IκB family members also correlates more strongly with the efficiency that a given family member sequesters its target NF-κB in the cytoplasm than with its ability to inhibit binding of NF-κBs to DNA in the nucleus [59].
In Drosophila and mosquitoes, processing of compound Relish depends upon cleavage by the caspase-8 homolog Dredd while Dredd itself is negatively regulated by the FAF1 homolog Caspar [21], [62], [81]. We think it unlikely, however, that Ank proteins affect either Caspar or Dredd after immune challenge given we detected no reduction in Dredd/caspase 8 activity cells expressing Ank proteins. As noted above, these findings also raise important but unresolved questions about the role of the IκB domain in compound Relish in resting cells and whether Rel-49 functions as an IκB after processing. Although Rel-49 bound Rel protein homodimers in our SPR assays, the apparent ability of Ank proteins to bind and block processing of compound Relish in vivo suggests that the Rel-49 domain does not strongly interact with RHD and NLS of compound Relish either before or after immune challenge.
Our own unpublished transcriptome data identifies Rel gene homologs and most other components of the Toll and Imd pathways in P. includens. However, in the absence of any background studies on IκB/NF-κB interactions at the protein level, we currently are not able to directly determine whether MdBV Ank proteins bind to and disable NF-κBs in this natural host of M. demolitor. However, prior studies do indicate that ank-H4 and –N5 are rapidly and persistently expressed in the fat body and hemocytes of P. includens after MdBV infection [33], while results of the current study show that MdBV infection inhibits the expression of two AMP genes. Studies from B. mori suggest these AMP genes are likely regulated by the Imd pathway, although in P. includens we recognize the possibility they could also be regulated fully or in part by Toll signaling. We also show that immune challenge with M. demolitor eggs or inactivated MdBV strongly induces the expression of one of these AMP genes (lebocin) and that MdBV infection blocks this response. These results are fully consistent with our results in Drosophila cells, which show that Ank-H4 and –N5 disable both Imd and Toll signaling [39]. Results of the current study also suggest parasitism activates NF-κB signaling in the natural host but MdBV subverts this response.
Other viruses and parasitoids are known to also activate NF-κB signaling [8], [9], [11], [12], [81], [82], but no studies to our knowledge indicate that AMPs are important effector molecules in defense against these entities. However, NF-κBs regulate many other genes in response to infection whose function remains unknown [11], [76], [83], [84]. Thus, while MdBV disables expression of AMP genes, it is likely that other genes with roles in anti-parasitoid or anti–viral defense underlie the benefits to M. demolitor of subverting NF-κB signaling. Given that MdBV encodes other virulence genes that disable hemocyte function and the phenoloxidase (PO) cascade [35]–[38], it is also likely that ank genes interact with other MdBV gene products to disable both cellular and humoral defense responses of hosts.
The Imd and Toll pathways are thought to also play important roles in defending insects against opportunistic microbes, which most commonly infect insects by oral ingestion [1]. Thus, a possible cost to suppressing NF-κB signaling could be that it renders hosts and a developing parasitoid more susceptible to infection by other organisms. However, infection of hosts by PDVs also induces profound alterations in behavior including a near complete cessation of feeding, which likely reduces the risks of infection by another pathogen before the wasp's progeny complete their development [28].
Studies of two other PDVs implicate Ank proteins in inhibition of NF-κB dependent transcriptional activity [85], [86], while comparative data show that some Ank protein family members localize to the cytoplasm of insect cells, others localize to nuclei, and others still localize to both [85]–[87]. PDVs like MdBV belong to the genus Bracovirus, which evolved more than 100 million years ago from another taxon of viruses that infect insects called nudiviruses [30], [31], [88], [89]. Comparative genomic data further indicate the largest and most conserved genes encoded by bracoviruses are the ank and ptp gene families. The absence of any genes with significant homology to ank genes among known nudiviruses suggests this gene family originated from a eukaryote where it potentially functioned as an IκB. However, the ancient origins of the ank family together with rapid rates of evolution make it unclear whether this eukaryote was a wasp, an insect host, or another organism that predates the evolution of the Hymenoptera [31], [90], [91].
All studies were approved by the Biological Safety and Animal Care and Use Committee of the University of Georgia and were performed in compliance with relevant institutional policies, National Institutes of Health regulations, Association for the Accreditation of Laboratory Animal care guidelines, and local, state, and federal laws.
M. demolitor and P. includens were reared as previously described [92]. MdBV and MdBV genomic DNA were isolated from adult female M. demolitor as outlined by [93]. MdBV was transcriptionally inactivated by UV light treatment [94], while M. demolitor eggs were collected aseptically from female wasps in sterile phosphate-buffered saline (PBS) [37].
For bacterial expression of Dif and Dorsal, ORFs containing the RHD, nuclear localization signal (NLS), and 20 amino acids downstream of the NLS were polymerase chain reaction (PCR)-amplified using gene specific primers with sequence extensions for cloning into pET-LIC vectors (Table S1). The plasmids pSHhis-Dif and pSHhis-Dorsal respectively served as templates [4]. Full length Relish and Rel-49 were similarly amplified using gene specific primers and the plasmid pSHflag-Relish as template [4]. Each of the aforementioned plasmids was obtained from T. Ip (University of Massachusetts). The ARD of Cactus was amplified using specific primers and a full-length cDNA clone of Cactus (LD10168) from the Drosophila Genomics Resource Center as template, while Ank-H4 and Ank-N5 were amplified using specific primers and MdBV genomic DNA as template (Table S1). Briefly, 1 ng of template, 250 nM of each primer and 1.2 Units of KOD HiFi DNA Polymerase (Novagen) were combined in a 50 µl volume and amplified using the following conditions: 25 cycles of 98°C for 15 sec, 61°C for 2 sec, and 72°C for 20 sec. Rel products were then cloned into either pET32-EK-LIC, which encodes an N-terminal Thioredoxin and 6× histidine (His) affinity tag or pET51-EK-LIC, which encodes an N-terminal StrepTagII affinity tag. The IκB domain of Relish (Rel-49) was cloned without a stop codon into pET51-EK-LIC resulting in a C-terminal His tag. Cactus, Ank-H4, and Ank-N5 in contrast were cloned into pET-30-EK-LIC, which encodes an N-terminal His tag. Each construct was confirmed by DNA sequencing, and then expressed by transforming into Escherichia coli strains BL21 (DE3) cells. Transformed E. coli were grown in 2 L Luria Broth containing 100 µg/ml ampicillin (pET32 and pET51) or 50 µg/ml kanamycin (pET30) at 37°C with shaking at 275 rpm until the A600 reached 0.8–1.0. The cultures were cooled to room temperature and then induced with 0.1 mM isopropyl-β-d-thiogalactopyranoside (IPTG) for an additional 4–24 h at 20°C. Bacterial cells were harvested by centrifugation at 5000× g for 10 min and used immediately or stored at −80°C.
Bacterial pellets from 0.8 L cultures were resuspended in 40 ml of lysis buffer (50 mM Tris-HCl pH 8.0, 300 mM NaCl, 10 mM imidazole). After addition of lysozme (1 mg/ml) in 50 mM Tris-HCl (pH 8.0), cells were incubated on ice for 1 h followed by sonication with six, 10 sec bursts at 200 W using a Branson 450 Sonifier. For the constructs containing His affinity tags, the soluble recombinant proteins were purified from the clarified supernatant by incubating with 2 ml Ni-NTA Supreflow (Qiagen) agarose beads for at least 2 h and mixing by tumbling end over end at 4°C. The beads were then pelleted by centrifugation at 500× g for 5 min, and the supernatant (flow through) was removed. The beads were then resuspended with an equal volume of Buffer A (20 mM Tris–HCl, pH 8.5, 0.5 M KCl, 5 mM β-mercaptoethanol, 10% glycerol, 20 mM imidazole) at room temperature and quantitatively transferred to a 15 ml column at room temperature. The column was then packed at 1.5 ml/min with buffer A until the bed volume was constant, then washed with 10 volumes of buffer A at 1 ml/min, followed by two volumes of buffer B (buffer A, containing 1 M KCl and no imidazole). The column was then washed with two more volumes of buffer A, and the protein was eluted with buffer C (buffer A with 100 mM imidazole). Fractions (1.0 ml) were analyzed by SDS–PAGE and immunoblotting using an anti-His monoclonal antibody (Sigma or Qiagen) and stored at 4 or −80°C. When necessary, remaining contaminating proteins were removed by gel filtration using a Superdex75 column (Amersham). Proteins with an N-terminal StrepTagII were isolated using Streptactin agarose (Novagen). Briefly, bacterial lysates were prepared as described above and applied to a 2 ml packed and equilibrated Streptactin column at 0.5 ml/min at 4°C. The column was washed with 40 ml of wash buffer (50 mM Tris, pH 8, 150 mM NaCl, 5 mM 2-mercaptoethanol) at 0.5 ml/min at 4°C. Proteins was eluted with 10 ml of elution buffer (wash buffer with 2.5 mM desthiobiotin) and stored at 4° or −80°C. Proteins were desalted with PD-10 columns (Amersham) and washed into appropriate buffers using spin filtration. Protein concentrations were determined using the Pierce Coomassie Plus Bradford assay.
Biosensor experiments were run on a Biacore 3000 instrument (GE Healthcare) at room temperature. Recombinant ligands (usually IκBs) were immobilized on research grade CM5 sensor chips by amine coupling as follows. The carboxymethyl surface of the chip was activated for 8 min at 5 µl/min with a 1∶1 mixture of 0.4 M N-ethyl-N′-(3-dimethylaminopropyl) carbodiimide (EDC) and 0.1 M N-hydroxysuccinimide (NHS). Recombinant IκBs diluted to 10 µg/ml in 10 mM sodium acetate, pH 4.5 were injected using quickinject in 5 sec pulses until a surface density of 5500 response units was achieved. Excess activated succinyl groups were then blocked by injecting 1 M ethanolamine, pH 8.5 for 8 min at 10 µl/min.
Kinetic titration experiments were performed by serially diluting recombinant analytes (usually NF-κBs) in running buffer (10 mM HEPES, pH 7.4, 150 mM NaCl, 3.4 mM EDTA, 0.005% (v/v) plus surfactant P20), and sequentially injecting doubling concentrations for 60 or 90 sec, allowing 120 sec dissociation after each injection. Injections were made across both the ligand bound cell and a reference cell, in which the surface had been activated with EDC: NHS, and then immediately deactivated with ethanolamine. Sensorgrams were recorded by automatic subtraction of the blank reference cell from the experimental cell to remove non-specific binding affects and to correct for drift. Typically, five injections of 160 nM through 2.56 µM NF-κB were measured. The response profiles were fit to the kinetic titration model (provided by Biacore) assuming simple 1∶1 Langmuir binding to generate kinetic and thermodynamic binding constants. The high surface densities used were necessary to produce clean responses above the noise of the machine. To determine whether mass transport effects significantly influenced results, the Dif-Cactus interaction was analyzed using a surface density of 15,000 response units (RU). The constants measured were within the standard error of the experiments using a ligand surface density of 5000 RU. Therefore, mass transport effects were deemed negligible.
All incubations were performed at room temperature on a rotator. Rel protein homodimers were diluted into binding buffer (50 mM Tris, pH 8, 150 mM NaCl, 0.1% BSA, 0.1% Triton X-100) to a concentration of 4.8 nM, and incubated with varying concentrations of competing IκBs for 1 h. Cactus was initially added to 14.3 nM, a 3 fold molar excess of the Rel protein, and incubated for 1 h. Rabbit anti-thioredoxin antibody (0.5 µl, Sigma T 0803) was then added and incubated for 1 h. Protein A beads (BioRad Affigel) were equilibrated in binding buffer and 20 µl of equilibrated, packed beads were added to the reactions and incubated for 1 h. The beads were pelleted by centrifugation at 1000× g, the supernatant was discarded, and the beads were washed 3× with binding buffer. The beads were then washed a fourth time with binding buffer with no BSA or Triton X-100 and the supernatant discarded followed by suspension in 50 µl of 1.5× SDS-sample buffer plus 2-mercaptoethanol and boiled for 5 min. The resulting supernatants were then subjected to SDS-PAGE and immunoblot analysis as described below.
The coding sequences for ank-H4 and ank-N5 were previously cloned into the expression vector pIZT/V5-His (Invitrogen), which uses the OpIE2 promoter from Orgyia pseudotsugata baculovirus for constitutive expression of the gene of interest and incorporates a C-terminal V5 epitope tag [39]. Drosophila mbn-2 cells were maintained in Schneider's medium (Sigma) supplemented with 10% fetal bovine serum (Atlanta Biologicals) [20]. Mbn-2 cells were transfected by adding cells to 6 well culture plates (Corning) (1×106 cells per well in 1 ml of complete medium). Twenty-four h later, 2 µg of each construct (pIZT/Ank-H4, pIZT/Ank-N5, or pIZT/V5-His (empty vector)) was diluted into 1 ml of Schneider's medium without serum followed by addition of 16 µl of Cellfectin (Invitrogen). After a 20 min incubation period, complete medium was removed from the cells in each well and the transfection medium was added. The transfection medium was then removed after 6 h and replaced with 1 ml of complete medium. Cells were immune challenged 48 h post-transfection with 10 µg/ml of commercial lipopolysaccharide (LPS) that contained peptidoglycan (PGN) (Sigma) for 2–24 h. Following collection and centrifugation, cell pellets were washed 3× in PBS (pH 7.2). Whole cell lysates were prepared by resuspending cell pellets in lysis extraction buffer (20 mM HEPES, pH 7.5, 100 mM KCl, 0.05% Triton X-100, 2.5 mM EDTA, 5 mM DTT, 5% glycerol, and protease plus phosphatase inhibitor cocktail (Roche). Cytoplasmic and nuclear extracts were prepared using NE-PER Nuclear and Cytoplasmic Isolation Kit (Pierce) plus protease and phosphatase inhibitor cocktail. Protein concentrations were determined by Bradford assay. Total RNA was isolated from mbn-2 cells using the Hi-Pure RNA extraction kit (Roche) and quantified using a Nanodrop spectrophotometer (Thermo Scientific).
For analysis of recombinant proteins, samples were electrophoresed on 1 mm PageR precast minigels (Lonza) and transferred to PVDF (Immobilon) by tank transfer. The membranes were blocked for 1 h in 5% dry milk in TPBS (0.1% Tween 20), followed by detection using a mouse anti-His monoclonal antibody (1: 2000) (Qiagen) and a goat anti-mouse horseradish peroxidase-conjugated secondary antibody (Jackson Laboratory) (1∶20,000). Bands were visualized using 3, 3-diaminobenzidine. For analysis of cell extract proteins, samples (20 µg of protein per lane) from mbn-2 cells were electrophoresed on precast 4–20% gradient gels (Lonza) followed by transfter to PVDF membranes and blocking as described above. Ank-H4 and -N5 were detected using a murine anti-V5 antibody (Invitrogen) (1: 10,000) and a goat anti-mouse horseradish peroxidase-conjugated secondary antibody (Jackson Laboratory) (1∶20,000). β-tubulin and Histone H1 were detected using a goat anti-β tubulin polyclonal antibody (Abcam) (1∶5000) or mouse anti-Histone H1 antibody (Santa Cruz) (1∶1000) followed by incubation with a goat or mouse anti-rabbit horseradish peroxidase-conjugated secondary antibody (1∶10,000 or 1∶5000). Relish was detected using a rabbit anti-Rel-68 antibody (S. Stoven, University of Umea) and anti-rabbit horseradish peroxidase-conjugated secondary antibody (1∶10,000. Bands were visualized by chemiluminescence using the ECL Advance Western blotting detection kit (Amersham Biosciences) and a bio-imaging system (Syngene).
First-strand cDNA was synthesized from mbn-2 cell total RNA using random hexamers and Superscript III (Invitrogen) [33]. rqRT-PCR reactions were run using a Rotor-Gene 3000 Cycler (Corbett) with 10 µl reaction volumes containing 1 µl of cDNA, 5 µl of iQ SYBR Green Supermix (Bio-Rad) and 250 nM of forward and reverse primers specific for the AMP genes diptericin, metchnikowin, and defensin, or the Drosophila 18 s ribosomal gene (Table S1). Cycling conditions were: initial denaturation at 94°C for 3 min, followed by 45 cycles with denaturation at 94•C for 10 sec, annealing at 50/55•C for 15 sec, and extension at 72°C for 20 sec. Data were acquired during the extension step, and analyzed with the Rotor-Gene application software. For every amplicon, reactions were carried out in quadruplicate, from which mean threshold cycle (CT) values plus standard deviations were calculated. All data were normalized to internal 18 s rRNA levels from the same sample. To compare transcript abundance for a given gene among treatments, we calibrated each ΔCT value against 0 h control, generating a ΔΔCT value, followed by transformation using the expression 2−ΔΔCT to obtain relative transcript abundance values (RA) [95]. In some cases these data were non-normally distributed. We therefore used a natural log transformation of each RA followed by ANOVA and pairwise t-tests to assess differences among treatments [33].
Dredd activity in mbn-2 cells transfected with pIZT/Ank-H4, pIZT/Ank-N5, or empty vector was assessed using a commercially available caspase-8 assay (Caspase-Glo) and the luminogenic substrate Ac-LETD-pNA (Promega) according to the manufacturer's protocol. All assays were performed in duplicate using independent samples and a BioTek Synergy 4 plate reader. Relative luminescence units (RLU) were determined 10 min after addition of substrate with the resulting data thereafter analyzed by ANOVA.
P. includens fifth instars (day 2) were immune challenged by injecting larvae with heat killed E. coli (1×106 cell in 1 µl of PBS), a physiological dose of inactivated MdBV (1×109 virions ( = 0.1 wasp equivalents) in 1 µl PBS [93], or 3–5 M. demolitor eggs in PBS using a glass needle mounted on a micromanipulator. Larvae injected with sterile PBS alone served as a wounding control. Other larvae were first injected with 0.1 wasp equivalents of viable MdBV followed 12 h later by immune-challenge using the above elicitors. Fat body was dissected from individual larvae in sterile PBS either before challenge with each elictor (0 h) or 2 and 8 h after. Isolation of total RNA, first-strand cDNA synthesis, and rqRT-PCR reactions were then run using primers specific for the P. includens cecropin, lebocin or 18 s ribosomal RNA gene (Table S1) as described above.
|
10.1371/journal.pcbi.1001114 | CAERUS: Predicting CAncER oUtcomeS Using Relationship between Protein Structural Information, Protein Networks, Gene Expression Data, and Mutation Data | Carcinogenesis is a complex process with multiple genetic and environmental factors contributing to the development of one or more tumors. Understanding the underlying mechanism of this process and identifying related markers to assess the outcome of this process would lead to more directed treatment and thus significantly reduce the mortality rate of cancers. Recently, molecular diagnostics and prognostics based on the identification of patterns within gene expression profiles in the context of protein interaction networks were reported. However, the predictive performances of these approaches were limited. In this study we propose a novel integrated approach, named CAERUS, for the identification of gene signatures to predict cancer outcomes based on the domain interaction network in human proteome. We first developed a model to score each protein by quantifying the domain connections to its interacting partners and the somatic mutations present in the domain. We then defined proteins as gene signatures if their scores were above a preset threshold. Next, for each gene signature, we quantified the correlation of the expression levels between this gene signature and its neighboring proteins. The results of the quantification in each patient were then used to predict cancer outcome by a modified naïve Bayes classifier. In this study we achieved a favorable accuracy of 88.3%, sensitivity of 87.2%, and specificity of 88.9% on a set of well-documented gene expression profiles of 253 consecutive breast cancer patients with different outcomes. We also compiled a list of cancer-associated gene signatures and domains, which provided testable hypotheses for further experimental investigation. Our approach proved successful on different independent breast cancer data sets as well as an ovarian cancer data set. This study constitutes the first predictive method to classify cancer outcomes based on the relationship between the domain organization and protein network.
| It is widely known that cancer is a complex process in which a large number of genes appear to be involved. Through experimental approaches, some oncogenes and tumor suppressors have been identified as playing important roles in the signaling and the regulatory pathways. However, we have not fully understood the complete mechanism of how cancer develops and how it leads to different disease outcomes (aggressive/dangerous or non-aggressive/less-dangerous). In order to identify a list of gene signatures and better predict cancer outcome, we developed an integrated and systematical approach by investigating gene expression profiling alternation caused by disruptions between protein-protein interactions and domain-domain interactions in the human interactome. Our approach achieves the favorable predictive performance if tested on a set of well-documented breast cancer patients, which suggests that the disrupted interactome is important to determine patient prognosis. Our approach is robust if tested on other independent data sets. This work provides a promising prognostic tool to classify different cancer outcomes.
| Cancer development is a complex process driven by multiple genetic and environmental factors [1], [2], [3]. Understanding the underlying mechanism of this process and identifying related markers to assess the outcome of this process could lead to better management and treatment of this complex disease. For example, the majority of breast cancer patients are currently over-treated [4] due to the lack of accurate assessment of the risk of metastasis. As a result, a substantial proportion of patients are receiving the otherwise avoidable aggressive adjuvant therapy in accordance to the current guidelines [5]. Although the importance of identifying prognostic signatures that could accurately predict cancer outcomes is widely appreciated, it has remained a challenging task. With the emergence of large amounts of DNA microarray-based tumor gene expression profiles, molecular diagnostics and prognostics have begun to provide solutions to this challenge [6]. Several predictive tools [7], [8], [9], [10] were reported to classify different cancer outcomes primarily based on the identification of gene expression signatures observed in these outcomes. However, the predictive performance of these approaches was limited. For instance, in two large-scale expression studies [9], [10], approximately 70 gene markers were identified that could be used in the prediction of the metastasis in breast cancer, but only with an accuracy of 60–70%. This relatively low accuracy could be explained by some intrinsic shortcomings of the microarray data, as different experiment and analysis designs could yield inconsistent results due to systematic errors [11] and by the heterogeneity of carcinogenesis resulting from multiple factors such as specific samples and cancer types [6]. Recently, the prognostic predictive performance has been improved by integrating the gene expression profiles and the human interactome data, based on the notion that disruption of protein interaction network might affect disease outcomes [12]. Protein-protein interactions (PPIs) play an important role in the process of carcinogenesis. At the molecular level, any genetic alternation such as somatic mutations, translocations, deletions and insertions that modify expressed protein-coding genes could cause changes in a PPI-based regulatory mechanism that governs normal cell function. This could lead to aberrant or uncontrolled cell growth and eventually to cancer [1]. For example, mutations in the zinc finger domain presented in the oncoprotein MDM2 can disrupt the interaction of MDM2 with ribosomal proteins L5 and L11 and mediate p53 degradation [13]. The recent availability of large-scale PPI networks has made it possible to identify better gene signatures by combining the gene expression measurements with the perturbed protein interaction networks in the cell. Chuang and colleagues [14] developed a method to find subnetwork-based signatures by incorporating PPI networks and gene expression profiles. The resultant subnetworks with their gene expression profiles were used as markers to predict the prognosis of breast cancer patients. This study yielded an accuracy of 70–72% in determining a breast cancer as metastatic versus non-metastatic. Their study revealed the usefulness of the PPI network in conjunction with the gene expression profiles and provided a starting point to future studies. More recently, Taylor and colleagues [12] proposed a new methodology to predict breast cancer outcome based on the correlation of gene expression profiles between hub proteins and their interacting partners in the PPI network. This approach showed improved predictive performance at an accuracy of 76% when tested on a different set of gene expression profiles from breast cancer patients. These studies demonstrated that the topology of a PPI network could be a helpful line of biological evidence in differentiating cancer outcomes. In the meantime, however, there are other important biological elements that might be involved in the development of cancer genome and phenotype. To further strengthen the power of novel predictive tools, these lines of biological evidence need to be investigated and incorporated if proven useful.
In an alternative approach, we focused on the prediction of cancer outcomes within the context of domain interaction network. Domains are defined as independent structure and/or functional blocks of proteins. It is clear that protein-protein interactions are mediated by the interactions between protein domains [15]. For example, SH2 domains mediate many critical protein interactions in signal transduction [16], [17]. Disrupted domain-domain interactions (DDIs) have been shown to stop the chain reaction of biological pathways at any point [18], [19], thus lead to various diseases [20], [21], [22]. This fact has motivated us to investigate the disruptions in a PPI network that are caused by DDIs, which might be a defining feature of tumor phenotype and thus could be used to determine patient prognosis. In the context of DDIs, we can categorize a given interacting protein into one of the two types based on the relationship of this protein and its neighboring proteins in the protein interaction network (Figure 1). We call a protein a ‘singlish-interface’ protein if it interacts with its neighboring proteins through the same domain-domain interaction; therefore, those domain-domain interactions are mutually exclusive (Figure 1A). Conversely, we call this protein a ‘multiple-interface’ protein if it interacts with its neighboring proteins through different domain-domain interactions, as those interactions are simultaneously possible (Figure 1B). It has been demonstrated that singlish-interface proteins evolve faster than multiple-interface proteins and are more likely to interrupt protein interactions and disturb the protein interaction network [23]. Therefore, we hypothesize that singlish-interface proteins are also more likely to be involved in the process of tumor progression than multiple-interface proteins. Meanwhile, DDIs could be interrupted by genomic variations located within interacting domains. One type of these genomic variations is somatic mutation. Somatic mutations are genetic alternations in DNA that are neither inherited nor passed to offspring. Some of these are thought to be driving the cancer process and have been refereed to as “driver mutations”, which can contribute to the development of the cancers or other diseases [24]. Therefore, we sought to investigate the perturbation of the protein interaction network in cancerous cells caused by the presence of somatic mutations, and to examine whether somatic mutation data could provide help in the prediction of cancer outcome. In summary, in addition to PPI data and gene expression data, we looked into incorporating two other types of data that might be functionally associated to the disturbance of the PPI networks: domain-domain interactions (DDIs) and somatic mutations.
In this study, we propose an integrated approach, named CAERUS, to predict the likelihood of cancer outcomes in unknown cancer patients provided the gene expression profiles of these patients are available. To implement CAERUS, we first developed a model to score each protein present in the expression profiles based on the domain connections to their interacting partners and the somatic mutations located in the domains. Next, gene signatures defined as proteins whose scores are above a preset threshold were identified. Then we computed the correlation of gene expression profiles of the gene signatures and their neighboring proteins. A modified naïve Bayes classifier was used to predict cancer outcome based on this correlation. Compared to previous studies, our study has several advantages. First, apart from the PPI network and the gene expression profiles, the DDI network and the somatic mutations within domains were integrated into our predictive model, which has improved the prediction performance to an accuracy of 88.3%, sensitivity of 87.2% and specificity of 88.9%. Second, our results compiled a list of cancer-associated gene signatures and domains, which provided testable hypotheses for further experimental investigation. Third, our approach is not specific to a specific cancer dataset and can thus be applied to different independent cancer data sets.
We tested whether our identified gene signatures are good indicators to differentiate a set of two groups of sporadic and non-familial breast cancer patients [25]. We defined patients who were disease free after extended follow-up as patients with ‘good outcome’ and those who died of disease as patients with ‘poor outcome’. The patient data was filtered to remove patients that were still alive with disease or dead from other reasons, as reported by Taylor [12]. The resultant dataset contained 179 patients with ‘good outcome’ and 74 patients with ‘poor outcome’. For each patient, a profile was computed based on the difference of the gene expression value between the gene signatures and their neighboring proteins. For the identification of gene signatures, we applied a scoring procedure to the protein domains present in each gene products based on the number of mutually exclusive DDIs they participated in (see methods). Using this approach we found that only one parameter needed to be tuned: the threshold (c) of domain index scores (Sd). The threshold (c) was tuned by testing our approach on the breast cancer data set using different Sd values (see methods). We then evaluated the performance of our approach by calculating three performance measurements: accuracy, sensitivity and specificity. In this study, accuracy = (TP+TN)/(TP+FP+TN+FN); sensitivity = TP/(TP+FN); specificity = TN/(TN+FP). A true positive is defined as the case that a “poor outcome” patient was successfully predicted as having the “poor outcome” and a true negative is defined as the case a “good outcome” patient was correctly predicted as having the “good outcome”. From the observation of the performance plot based on different Sd (Figure 2), we concluded that our approach achieved the best performance with the accuracy of 85.8%, the sensitivity of 87.1% and the specificity of 82.6% when the threshold (c) of domain index scores (Sd) were set as 50. We also found that with higher threshold (c), a smaller set of gene signatures were generated, and consequently lower the performance was. On the contrary, with lower threshold (c), the gene signature list contained higher noise and generated more false positives and negatives. Next, we did survival analysis to prove the ability to predict survival of our approach under this setting and observed the significantly different 10-year survival (Mantel-Cox Log Rank test, nominal P-value = 2.19×10−8) (Figure 3) between two groups of patients.
A total of 171 gene signatures were identified in a breast cancer data set [25] using our approach at the threshold (c) of 50 as described in the above section. These gene signatures mainly are involved in 5 major cancer-related biological processes: transcription (P-value = 9.3×10−10), DNA repair (P-value = 3.8×10−5), signal transduction (P-value = 7.9×10−13), cell cycle (P-value = 1.1×10−9) and protein phosphorylation (P-value = 2.9×10−26) if we performed GO Term enrichment analysis using FuncAssociate [26] (Figure 4A). The complete list of over-represented GO terms associated with identified gene signatures is in the supplementary materials (Table S1). In addition, 36 human biological pathways can be derived when we mapped the gene signatures to the Reactome database that contains manually curated human biological pathways [27] (P-value<0.001) (Figure 4B). For instance, the well-known oncogenic transcription factors such as FOS, JUN and NFκB were identified as gene signatures by this study. We also identified some DNA repair genes including XRCC5, MSH, PCNA and others as gene signatures. These genes were demonstrated to cause cancer because mutations in those genes disable the ability of DNA repairing, which subsequently leads to the accumulation of mutations [28], [29], [30]. Genes involved in signal transduction, an important type of pathways in cancer development, such as MARK14, VAV1 and PIK3R1 were also identified as gene signatures in this study. Besides, a group of cyclin-dependent kinases (CDK2, CDK3, CDK4, CDK6) that control cell proliferation [31] and genes (SRC, ABL1) related to protein phosphorylation [32] were also identified by our approach. In summary, there were 38% (65 out of 171) of the identified gene signatures found to be the genes associated with cancers in Online Mendelian Inheritance in Man (OMIM; http://www.ncbi.nlm.nih.gov/omim/). This percentage is significantly greater than what could be found purely by chance (Adjusted P-value<10−12, by Fisher's Exact Test), indicating the capability of our approach to identify disease genes. Interestingly, only 15% (26 out of 171) of the identified gene signatures were known cancer susceptibility genes compared to a list of 410 genes downloaded from The Cancer Gene Census (http://www.sanger.ac.uk/genetics/CGP/Census/), whose mutations had been causally implicated in cancer, but the small overlap is still statistically significant at P-value of 7.7×10−6 by Wilcoxon Test. This result was consistent with those of the previous studies, which yielded 21% and 16%, respectively [12], [14]. In order to examine the importance that the cancer susceptibility genes contribute to cancer prognosis, we employed these 410 known cancer susceptibility genes as signature genes to predict breast cancer outcomes, we observed a relatively low accuracy of 72.6%, sensitivity of 72.9% and specificity of 71.4% if tested on the same breast cancer set (Figure S1). Taken together, the low percentage of known cancer susceptibility genes present in our gene signature list suggests that the mutations in not only these genes, but also other genes, might collectively affect the process of tumor-aggressiveness and response to therapy in various ways by disrupting the modularity of the PPI network. Among other genes in our gene signature list but not in the list of known cancer susceptibility genes, 32% (46 out of 145) of genes can be mapped to the human biological pathways in which known cancer susceptibility genes anticipate in the Reactome database (P-value = 2.1×10−8 by Z-test). Therefore, we speculated that the other genes could be the downstream effectors of the cancer susceptibility genes and the changes in their expression value could reflect the disruption of the PPI network caused by the mutations in the cancer susceptibility genes. In order to investigate what types of domains tend to exist in ‘singlish-interface’ proteins and disrupt protein interactions, we calculated the number of involved domain-domain interactions of each domain in ‘singlish-interface’ proteins against the whole genome and compared it to that expected by chance (P<0.01, Z-test) (see Figure 1). We identified a list of 29 over-represented domains within 171 gene signatures (Table 1). Interestingly, 93% (27 out of 29) of the domains were annotated as cell signaling domains such as SH2, Pkinase and Ras according to the SMART database [33] indicating that these domains were likely to play a critical role in carcinogenesis through disrupting the protein interactions within signaling pathways. For example, the SH2 domain of the oncoprotein Src interacts with 86 domains within 57 proteins. It has been demonstrated that SH2 domain regulates intracellular signalling cascades by interacting with high affinity to phosphotyrosine-containing target peptides [34], [35] and is related to cancer cell migration and proliferation [36]. Another example is that the Pkinase domain contains the catalytic function of protein kinases that are essential in the process of phosphorylation [37], [38]. Many diseases including cancer are caused by dysfunction of phosphorylation [39].
It is widely accepted that genetic changes such as somatic mutations are implicated in cancer development [40]. Also, some somatic mutations reveal the role of functional domains in hereditary disorders and complex diseases [41]. For example, tumors highly sensitive to epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors often contain dominant mutations in exons that encode a portion of the tyrosine kinase (TK) domain of EGFR [42]. To investigate the possibility that somatic mutations within domains represent another type of important signal to differentiate two classes of patients, we incorporated the somatic mutation data compiled from the COSMIC database to our scoring model (see methods) by searching for the genes having mutually exclusive domains that harbor somatic mutations. We hypothesized that these mutations could disrupt DDIs and PPIs and consequently change the modularity of the human protein interaction network. By employing the modified domain index function that incorporates the somatic mutation data, we tuned again the threshold (c) using different Sd values. At the threshold of Sd = 80, our approach identified 126 gene signatures and achieved the accuracy of 88.3%, the sensitivity of 87.2% and the specificity of 88.9% when tested on the breast cancer outcome data (Figure 2). All of 126 gene signatures belong to a list of 171 gene signatures identified by the CAERUS approach without integrating the somatic mutation data, which indicates that 45 gene signatures failed to pass a preset threshold after the somatic mutation data were used. To test weather the slight improvement on predictive performance (0.038 difference in the area under the ROC curve) is statistically significant, we tested CAERUS on randomized 126 genes from the list of 171 gene signatures and repeated this procedure 100 times (Figure S2). We found that this improvement is indeed statistically significant at the P-value of 2.8×10−5 by Wilcoxon Test. Compared to the performance of CAERUS' that does not incorporate the somatic mutation data, the improvement on CAERUS' performance by integrating the somatic mutation data suggests that the somatic mutation data can be used to supplement our accuracy to predict cancer survival outcome. However, the capability of using the mutation data appears limited due to the fact that not all mutations are driving the development of the cancer, the so-called “driver mutations” [43]. Minor performance improvement could be explained by the incompleteness of currently available somatic mutation data or the bias introduced by “passenger mutations”. With the help of the numerous Cancer Genome Projects [44], [45], the size of the somatic mutations data in human will grow in the near-future possibly providing us with even better indications from mutation data.
Identifying novel prognostic markers to classify different cancer outcomes has been widely studied with the increasingly available gene expression profiles. The approaches described in previous publications can be categorized into three classes: 1) gene expression pattern-based method, in which markers are selected based on whether their expression profiles can differentiate different groups of patients [9], [10]; 2) PPI subnetwork-based method, in which each marker represented as a subnetwork in the PPI network was identified by maximizing the mutual information measuring the association between the expression value of each gene in the subnetwork and the types of patients [14]; 3) PPI modularity-based method in which each gene signature was identified by comparing the difference of the gene expression value between a hub gene and their interacting partners in the PPI network [12]. In this study, we employed a novel approach based on finding genes in the PPI network with mutually exclusive domains and somatic mutations located in these domains as the markers. Wang et al [10] and van de Vijver et al [25] reported 63% and 62% accuracy, respectively, for the prediction of metastasis using gene expression pattern-based methods. Using the PPI subnetwork-based method, Chuang et al [14] yielded the accuracy of 72.2% and 70.1% using the same data set as Wang et al and van de Vijver et al did. Using the PPI modularity-based method, Taylor et al [12] reported the accuracy of 76% tested on the breast cancer patient data set. We first applied our approach on the same data set as Chuang et al [14] used and adopted the identical training and testing strategy (five-fold cross-validation) and observed that our approach achieved the accuracy of 83.2%, the sensitivity of 84.6% and the specificity of 82.5%. Next, we applied our approach on the same data set as Taylor et al [12] used and adopted the identical training and testing strategy (five-fold cross-validation) and observed that our approach achieved the accuracy of 87.3%, the sensitivity of 87.2% and the specificity of 88%, which indicates that our method outperforms other approaches and provides a promising solution to predict cancer outcome (Figure 5).
To test the robustness of our approach on different independent data sets or different types of cancer, we first applied our approach to a data set that included 236 primary invasive breast tumors [46]. Using five-fold cross-validation, our approach achieved the accuracy of 92.4%, the sensitivity of 94% and the specificity of 90.2%. Our approach also revealed significantly different 10-year survival (Mantel-Cox Log Rank test, nominal P-value = 1.8×10−25) (Figure 6A). Another independent data set that includes 117 primary breast tumors was utilized to evaluate the performance of our approach [47]. Using the leave-one-out cross-validation (LOOCV) strategy due to insufficient sample size, our approach achieved the accuracy of 89.8%, the sensitivity of 85.7% and the specificity of 91.6% with the significantly different 10-year survival (Mantel-Cox Log Rank test, nominal P-value = 7×10−4) (Figure 6B). These results indicate that our predictive approach has good performance in predicting breast cancer outcome when tested on different independent data sets. Next, we compiled a set of 110 patients with advanced-stage ovarian cancer that contains the gene expression profiles of 34 patients without disease recurrence and 76 patients with disease recurrence [48]. We applied our approach to this data set using the five-fold cross-validation strategy. We observed that our approach achieved the accuracy of 90.1%, the sensitivity of 90.4% and the specificity of 88.6%, further validating the robustness of our predictive approach when tested on different types of cancer data sets. The good predictive performance is also demonstrated by the 10-year survival curve (Mantel-Cox Log Rank test, nominal P-value = 3.12×10−12) (Figure 6C).
Biological network information has been proven to be a useful feature to improve prognosis performance [12], [14]. In this context, our study constitutes the first predictive method to classify cancer outcomes based on the information of protein interaction interfaces in the protein interaction networks. Compared to previous predictive approaches, the most outstanding feature of CAERUS is that we investigated biological network disruptions linked to cancer outcomes at the protein domain level. The favorable predictive performance of our approach suggests that association exists between cancer outcome and the alteration in the protein interaction network, and more importantly, that the alteration is probably caused by the genetic variations within interacting domains. These genetic variations are capable of interrupting the physical interactions between proteins and thus causing abnormal biological functions associated with cancer progression. In this study, we applied CAERUS primarily on breast cancer data sets and achieved favorable predictive performance. However, the strength of CAERUS is not restricted to a certain type of cancer; other types of cancer such as ovarian cancer can be analyzed in a similar manner. It is worth noting that the potential of the approach described in this study is restrained by the limitations of currently available data sources, as these data sources, such as the protein interaction data, the domain interaction data, the gene expression data are incomplete and also contain biases. The currently available somatic mutation data is also limited and not individual-based. With the growth in the size and better quality of these data sets, our study would lead to a more reliable and robust prognosis tool to access cancer outcome. Furthermore, this study could be optimized with the integration of additional types of data. For instance, we could achieve better predictive performance by integrating the patients' transcriptome data obtained via the RNA-seq technology which measures gene expression levels more accurately compared to the microarray approach [49]. With patient-specific somatic data, it will become possible to fine-tune the CAERUS approach and we would be able to achieve better performance. In addition, the effects of protein post-translation modifications such as phosphorylation, methylation and acetylation could also be potentially integrated into our model to reflect the influence of these types of modifications on the organization of the protein-protein interaction network during cancer development. In conclusion, we presented a novel and integrated approach to predict different cancer outcomes, which could be of significant clinical application.
We downloaded 108,307 unique PPIs in human from the iRefIndex database (ftp://ftp.no.embnet.org/irefindex/data) version of June 4, 2009. The iRefIndex database [50] provides a non-redundant list of protein interactions derived from several major protein interaction databases including BIND, BioGRID, CORUM, DIP, HPRD, IntAct, MINT, MPact, MPPI and OPHID. We also used a set of DDIs downloaded from the iPfam database [51], a DDI database based on RCSB Protein Data Back (PDB) crystal structures (http://www.pdb.org), which consists of 3,020 DDIs and 914 domains. For somatic mutations involved in cancer, a list of 88,641 somatic mutations was retrieved from the COSMIC database (version 43) that contains the mutation data and associated information extracted from the primary literature [52].
A set of gene expression profiles of 295 breast cancer patients and clinical results was collected from the work of van de Vijver and colleagues [25]. This data set was applied to test the performance of CAERUS. We defined patients who were disease free after extended follow-up as patients with ‘good outcome’ and those who died of disease as patients with ‘poor outcome’. The data was filtered to remove patients that were still alive with disease or dead from other reasons, as reported by Taylor [12]. The resultant dataset contained 179 patients with ‘good outcome’ and 74 patients with ‘poor outcome’. The mean duration of follow-up was 7.5 years for ‘good outcome’ patients and 2.8 years for ‘poor outcome’ patients. Two independent breast cancer data sets were employed for the validation purpose. The first data set consists of gene expression profiles of 236 patients with primary invasive breast tumors that derived from oligonucleotide arrays and the corresponding survival data of these patients were collected based on the patient records accompanying with the paper [46]. In this data set, 134 patients were classified as ‘good outcomes’ and 102 patients with ‘bad outcomes’ using the same abovementioned criteria. The mean duration of follow-up was 10.9 years for ‘good outcome’ patients and 4.9 years for ‘poor outcome’ patients. The second data set was obtained from the gene expression profiles of a cohort of 117 patients with breast tumors, of which 83 patients had ‘good outcomes’ and 34 patients had ‘bad outcomes’ derived from each patient’s survival duration and disease recurrence information included in the paper [47]. The mean duration of follow-up was 7.2 years for ‘good outcome’ patients and 2.1 years for ‘poor outcome’ patients. In addition, we compiled the data from a set of 110 Japanese patients who were diagnosed with advanced-stage serous ovarian cancers [48]. The gene expression profiles and the clinical characteristics of each patient were extracted from the supporting materials of the paper, in which 34 patients were labeled as ‘good outcomes’ and 76 patients as ‘bad outcomes’ using the same criteria described in previous data sets. The mean duration of follow-up was 3.3 years for ‘good outcome’ patients and 1.2 years for ‘poor outcome’ patients.
Given a gene expression data set and a gene signature x, we computed a score to measure the difference in co-expression of the gene signature and its neighboring proteins P = {p1, …., pn) in the PPI network between two types of cancer outcomes (“good/disease-free” vs. “poor/recurrent disease”) using the following equation:where n is the number of interactors of the gene signature x; rx,pi,good and rx,pi,poor is the Pearson correlation coefficient of expression values of protein x and its interactors P = {p1, …., pn} in different groups of patients (good or poor). The Pearson correlation coefficient of expression values of protein x and its interactors in the different groups is calculated by the following equation:
As a probabilistic model based on the Bayes' theorem, the naïve Bayes classifier has been widely applied to the classification problem in different fields of the biological sciences such as inferring cellular networks [53], modeling protein signaling pathways [54] and the prediction of protein-protein interaction interfaces [55]. Given the training dataset and testing dataset in which each data sample is represented as an n-dimensional vector (, …, ), 2 classes (Cgood, Cpoor). Here, n is the number of gene signatures; is the difference in co-expression of the gene signature i and its neighboring proteins in the PPI network in patient x. The prediction procedure follows as:
According to the Bayes theorem, we can get the highest posterior probability of each cancer patient sample x based on the following equation:where the class prior probabilities P(Cgood) is calculated by Xgood/X, the value of the number of training samples of class Cgood divided by the total number of training sample. P(|Cgood), P(|Cgood), …, P(|Cgood) can be easily calculated by , where is the number of training samples of class Cgood having the gene expression difference score falling into one certain bin/category, and Xgood the number of training samples belonging to Cgood. In this study, we divided the gene expression difference score into 20 bins as it ranges from 0 to 1.
In order to classify cancer patient samples in the testing dataset, we calculated the P(x|Ci)P(Ci) for each class Ci. Sample/patient x was then predicted as belonging to class Cgood if and only ifIn other words, it is assigned to the class Cgood for which P(x|Cgood)P(Cgood) is the maximum.
The method has been implemented in Perl and is available for downloading from http://www.oicr.on.ca/research/ouellette/caerus. It is distributed under the terms of GPL (http://opensource.org/licenses/gpl-2.0.php)
|
10.1371/journal.ppat.1003397 | Host Defense and Recruitment of Foxp3+ T Regulatory Cells to the Lungs in Chronic Mycobacterium tuberculosis Infection Requires Toll-like Receptor 2 | Acute resistance to low dose M. tuberculosis (Mtb) infection is not dependent on Toll-like receptor (TLR) 2. However, whether TLR2 contributes to resistance in chronic Mtb infection has remained uncertain. Here we report that, following low dose aerosol infection with Mtb, mice lacking TLR2 (TLR2KO), in comparison with wild type (WT) mice, exhibit enhanced cellular infiltration and inflammation in the lungs, and fail to stably control bacterial burden during chronic infection. IFNγ and IL-17 was expressed at equivalent levels in the two groups; however, the characteristic accumulation of Foxp3+ T regulatory cells (Tregs) in pulmonary granulomas was significantly reduced in TLR2KO mice. Nonetheless, this reduction in Tregs was independent of whether Tregs expressed TLR2 or not. To directly link the reduced number of Tregs to the increased inflammation present in the TLR2KO mice, we used a macrophage adoptive transfer model. At seven weeks post-Mtb infection, TLR2KO mice, which were adoptively transferred with WT macrophages, displayed enhanced accumulation of Tregs in the lungs and a concomitant reduction in inflammation in contrast with control mice that received TLR2KO macrophages. However, the pulmonary bacterial burden between the two groups remained similar indicating that TLR2's role in modulating immunopathology is functionally distinct from its role in restricting Mtb growth in chronic infection. Together, these findings unequivocally demonstrate that TLR2 contributes to host resistance against chronic Mtb infection and reveal a novel role for TLR2 in mediating the recruitment of Foxp3+ Tregs to the lungs to control inflammation.
| Tuberculosis (TB) is an important cause of mortality in many parts of the world. Infection with Mycobacterium tuberculosis (Mtb), the causative agent of TB, is usually acquired via inhalation of airborne droplets containing the bacteria. Following inhalation, Mtb interacts with specialized receptors, called Toll-like receptors (TLRs), on phagocytic cells present in the lung. In this study, we examine the contribution of TLR2 in activating the body's natural defenses against Mtb. Wild type mice infected with Mtb by the aerosol route are able to control bacterial replication in the lung and maintain it at a steady level during chronic infection. However, in genetically modified mice that do not express TLR2 (TLR2KO), Mtb infection leads to increased inflammation in the lung and inability to control Mtb growth. Here, we identify that the increased inflammation present in the lungs of Mtb-infected TLR2KO mice is due to the diminished ability of a type of regulatory cell (Foxp3+ Tregs) to accumulate in the lungs. The ability to recruit Tregs to the lungs is restored in TLR2KO mice if they are adoptively transferred with macrophages from wild type mice. In summary, we demonstrate that TLR2 functions in protection against chronic Mtb infection by controlling Treg accumulation in the lung to limit inflammation and tissue damage.
| Mtb expresses a large diversity of TLR2 ligands, including several types of lipoproteins and glycolipids, and also a trehalose dimycolate [1]–[3]. Interaction of these ligands with TLR2 expressed on macrophages and dendritic cells has multiple downstream effects. Several studies have reported that Mtb-derived TLR2 ligands produce a pro-inflammatory response [4]–[6], and consistent with these findings, mice deficient in TLR2 have diminished IL-17 response [7]. TLR2 signaling also induces direct antimicrobial activity in Mtb-infected human macrophages [8] by vitamin D3-dependent up-regulation of anti-microbial peptides [9]. Additionally, TLR2 signaling leads to suppressive effects on the functions of antigen-presenting cells (APCs). For example, TLR2 signaling in APCs induces IL-10 secretion [10] and prolonged signaling inhibits MHC class II expression [11], [12], antigen processing [3], [13], and IFNγ responsiveness [14].
Despite the extensive remodeling of macrophage functions following TLR2 signaling, TLR2-deficient mice are able to resist acute Mtb infection [15]–[17] and develop an appropriate secondary immune response [18] following a low dose aerosol infection. However, TLR2-deficient mice infected with a high dose of Mtb are more susceptible than WT to chronic infection and display an exaggerated immune inflammatory response, characterized by pneumonitis and enhanced cellular infiltration [15], [17]. These findings implicate a potential role for TLR2 in controlling inflammation during chronic infection.
Risk of developing tuberculosis has been shown to be associated with polymorphisms within the TLR2 gene, particularly within the TIR domain [19]–[21]. Analysis of a polymorphic guanine-thymine (GT) repeat located upstream of the TLR2 translational start site correlated shorter GT repeats with development of tuberculosis (TB) and lower TLR2 expression [22]. In addition, a TLR1 transmembrane domain polymorphism was shown to regulate the innate immune response to triacylated lipopeptides as well as extracts of mycobacteria [23]. Although the mechanism behind how these polymorphisms affect the immune response to Mtb is unclear, these correlations suggest an important role for TLR2 in host defense against Mtb. The aim of this study was to identify the mechanism by which TLR2 signals control inflammation and contribute to host resistance against Mtb. Here, we report that TLR2 functions in protection against chronic Mtb infection by keeping bacterial replication in check and limiting inflammation through recruitment of Foxp3+ Tregs to the lungs.
We first evaluated the role of TLR2 in host resistance against chronic Mtb infection. WT and TLR2KO mice were aerosol-infected with approximately 100 CFU of Mtb and disease progression was monitored for 15 weeks. As shown in Figure 1A, TLR2KO mice exhibited a significantly increased bacterial burden at weeks 7 and 10, and, by week 12, there was more than a log increase in bacterial burden in the TLR2KO mice as compared with WT. Beginning at 10 weeks following infection, the TLR2KO mice also began to succumb to infection. The WT mice, as expected, were able to control infection and maintain a steady bacterial load. A repeat experiment using a similar infectious dose of Mtb (around 150 CFU) demonstrated consistent findings (Fig. 1B).
Previous studies have indicated that the magnitude of the immune response leading to control of Mtb infection can be dependent on infectious dose [24]. To ensure that the reduced resistance of TLR2KO mice was not dependent on Mtb inoculum size, disease progression in response to a very low inoculum was also followed. During aerosol infection with a low dose of Mtb (approximately 10 CFU), TLR2KO mice again demonstrated a log increase in pulmonary bacterial burden, although increased morbidity was not observed until around 18 weeks post-infection, at which point the difference in bacterial burden between WT and TLR2KO mice was around 2 logs (Fig. 1C). Consistent with the differences observed in bacterial loads, acid-fast staining of infected lung tissue demonstrated increased bacilli within lung macrophages in the TLR2KO mice (S1A). The absence of TLR2 did not affect Mtb dissemination and replication outside of the lung since differences were not observed in the spleen (S1B).
The development of pulmonary pathology in H&E stained sections of infected TLR2KO and WT mice was next characterized (Figure 1, D–I). The murine granulomatous lesion is a collection of peripheral lymphocytic aggregates with B cell follicles juxtaposing areas with macrophages and other inflammatory cells types. These lesions lack the heterogeneity exhibited by human granulomas; although the granulomatous reaction is progressive, necrosis is not exhibited until an exorbitant bacillary load is achieved, and caseous necrosis and cavitation do not occur in C57BL/6 mice [25]. At 4 weeks post-infection, tissue architecture was similar between both groups, with distinct areas of granulomatous cellular infiltration (arrows) surrounded by unaffected lung areas apparent in WT and TLR2KO (Figure 1, D and E). As infection progressed, the granuloma architecture in the TLR2KO began to deviate from what is normally observed in WT mice. At 7 weeks post-infection, WT mice developed typical compact granulomatous lesions (arrow) containing macrophages and lymphocytic infiltrates (Figure 1F). In comparison, lungs of the TLR2KO mice exhibited increased inflammation with disrupted granuloma architecture (Figure 1G). Similarly, during the chronic stage at 12 weeks post-infection, the characteristic granulomatous structure of macrophages and densely compact lymphocytes was apparent in the WT lungs (Figure 1H) while extensive cellular infiltration and markedly reduced alveolar spaces were observed in TLR2KO lungs (Figure 1 I). The TLR2KO lungs displayed poorly formed granulomas, with loosely aggregated lymphocytes dispersed amongst macrophages (Figure 1 I).
Overall, the TLR2KO mice exhibited an early inflammatory response to Mtb similar to WT. However, as the infection progressed towards the chronic stage, the granuloma architecture in the WT lungs stabilized, while the TLR2KO lungs became consolidated with infiltrates which disrupted the granuloma morphology and progressively spread to comprise a majority of the lungs. These data confirm and extend the findings reported by Drennan and colleagues [17] that Mtb infection of TLR2KO mice leads to an exaggerated inflammatory response in the lungs.
Flow cytometric analysis of the cell surface markers CD11c, CD11b, and Gr-1 demonstrated significant increases in the numbers of CD11c−CD11b+ recruited macrophages, CD11c+CD11b+ myeloid DCs, CD11c+CD11b− alveolar macrophages, and Gr-1hiCD11b+ neutrophils infiltrating the lungs at a late stage of infection (10 and 13 weeks) (Figure S2B). Differences in NK1.1+ NK cells and CD19+ B cells were not observed (data not shown). Overall, these observations of increased recruitment of inflammatory cells to the lungs in the absence of TLR2 are consistent with the severe inflammatory lung pathology in TLR2KO mice during late stages of infection.
The enhanced inflammatory response in the lungs of Mtb infected TLR2KO mice led to the hypothesis that a regulatory T cell population may be lacking in the absence of TLR2. Natural Foxp3-expressing regulatory T cells (natTregs) are present in Mtb granulomas [26]. Therefore, the presence of natTregs in the lungs based on expression of Foxp3 was investigated. Lungs were harvested at the indicated time points post-infection, and flow cytometric analysis was performed to determine the percentage of Foxp3+ cells out of the CD4+ T cell population (Figure S3). The percentage of CD4+ cells expressing Foxp3 in the lungs was similar between WT and TLR2KO mice prior to Mtb infection. However, following infection, the TLR2KO mice displayed decreased frequencies of Foxp3+ Tregs in the lungs compared to WT (Figure 2A). While the percentage of Foxp3+ cells was lower in TLR2KO lungs than WT lungs, Tregs in both groups had equivalent expression of glucocorticoid-induced TNF receptor-related protein (GITR) and cytotoxic T-lymphocyte antigen-4 (CTLA-4), indicating a true natTreg phenotype of the Foxp3+ cells present in TLR2KO lungs (data not shown). Of note, it was also observed that both groups displayed decreased percentages of Foxp3+ cells out of the CD4+ population in the lungs compared to their naïve counterparts, although to a greater extent in the TLR2KO. This decrease is probably reflective of a greater expansion of CD4+ effector T cells compared to CD4+ Tregs following Mtb infection. Enumeration of cell infiltrates showed that total cell numbers, and CD4+ and CD8+ T cell numbers in TLR2KO mice were similar to WT at 7 weeks following infection (Figure S2A), a time when the percentage of Treg cells was lower in TLR2KO. This supports that the decreased percentage of Tregs in TLR2KO mice is not merely due to the greater expansion of T effector cells.
However, total cell numbers and T cell numbers continually increased over time in TLR2KO mice, while they stabilized in WT mice (Figure S2A). Therefore, we further investigated differences in Foxp3+ cells in the WT and TLR2KO by immunohistochemical staining of lung sections for the presence and localization of Foxp3-expressing cells. At 4 weeks post-infection, there were few Foxp3-expressing cells present in the granulomatous lesions of either WT or TLR2KO mice, although more were apparent in WT. At this time point, Foxp3+ cells were present in perivascular and peribronchiolar regions in both groups (Figure S4, Panels G and H). By 12 weeks post-infection, examination of WT (Figure 2C) and TLR2KO lung lesions (Figure 2F) by immunochemistry showed accumulation of high numbers of Foxp3+ cells in lesions of WT mice (Figure 2D). In contrast, very few Foxp3+ cells were observed in the affected areas of the lungs of TLR2KO mice at this time point (Figure 2G). Individual Foxp3+ cells or small clusters of Foxp3+ cells were dispersed randomly in TLR2KO lungs (Figure 2H), although they were not aggregated to the same extent as in WT (Figure 2E). Together, these findings show that TLR2 signals are necessary for the accumulation of regulatory T cells in the lungs during Mtb infection.
Other studies using murine models of Mtb infection have demonstrated that depletion of CD4+CD25+Foxp3+ natTregs or the complete absence of this population results in increased frequencies of IFNγ-producing CD4+ effector T cells [27], [28]. Given the importance of IFNγ in the protective immune response against Mtb, the possibility that decreased frequencies of Foxp3+ Tregs in the lungs of TLR2KO mice may correlate to enhanced Th1 responses was investigated. ELISPOT assay was performed with cells isolated from the lungs during chronic stages of infection and re-stimulated with Mtb-pulsed bone marrow-derived DCs serving as APCs. As shown in Figure 3A, there were no differences in the numbers of Mtb-specific IFNγ secreting cells in the lungs of WT and TLR2KO mice. Further, the decreased accumulation of Foxp3+ Tregs did not correlate to enhanced IL-17 gene expression, as there were no significant differences in IL-17 gene expression between WT and TLR2KO lungs (Figure 3B). Overall, these results indicate that the decreased accumulation of Foxp3+ Tregs in the lungs of TLR2KO mice is not associated with enhanced Mtb-induced Type I T cell responses.
It has been reported that TLR2 signaling on natTregs promotes their expansion and survival [29]–[31]. Therefore, the decreased number of Foxp3+CD4+ cells observed in the lungs of TLR2KO mice could result from decreased expansion of this population in peripheral lymphoid organs. To address this, the levels of Foxp3+ Tregs in the spleens were monitored following Mtb infection. Consistent with a previous report [32], decreased Foxp3+CD4+ cells were observed in the spleens of naïve TLR2KO mice as compared to WT levels. However, during Mtb infection this difference was no longer observed (Figure 2B), suggesting that differences in peripheral expansion of Foxp3+ Tregs between WT and TLR2KO mice do not account for the decreased frequencies of these cells in the lungs in the absence of TLR2.
Given the possibility that TLR2 signals are important for natTreg survival, we determined if reconstitution of TLR2KO mice with WT Tregs would allow for Treg accumulation in TLR2KO lungs. CD4+CD25+ Tregs from naïve WT mice were transferred to both TLR2KO and WT mice one day prior to and 4 weeks post-Mtb infection. TLR2KO and WT mice that did not receive Tregs were infected at the same time as controls. Flow cytometric analysis of Foxp3+CD4+ cells at 6 and 10 weeks post-infection indicated that the transfer of WT Tregs did not increase the frequency of this population in the lungs of TLR2KO mice, which were significantly lower than WT controls at 6 weeks (Figure 4A). Consistent with this, total pulmonary cell numbers at 10 weeks were significantly increased in both recipient and non-recipient TLR2KO as opposed to WT controls (Figure 4B). Also, the transfer of Foxp3+Tregs did not affect bacterial burden in the lungs. Total CFU in the lungs was increased by 1 log in the TLR2KO groups at both 6 and 10 weeks, although this increase was only significant at 6 weeks within the non-recipient group (Figure 4C). Together, these data demonstrate that the transfer of WT Foxp3+ Tregs failed to restore Treg numbers to WT levels in the lungs of TLR2KO hosts.
To definitively address whether the absence of TLR2 on natTregs is responsible for their decreased accumulation in the lungs of Mtb-infected TLR2KO mice, an adoptive transfer model using T cell-deficient Rag2−/− mice was used (described in Figure S5A). Conventional naïve CD4+CD25− T cells (which will be referred to as Tconv) and CD4+CD25+ T cells (Treg) were purified from WT (CD45.1+) and TLR2KO (CD45.2+) mice. By flow cytometry, greater than 99% of sorted CD25+ cells reacted with anti-Foxp3 antibody (data not shown). Rag2−/− mice were reconstituted one day prior to Mtb infection with combinations of 2×106 Tconv cells and 2×105 Treg cells. Group I received WT Tconv (CD45.1) and TLR2KO Treg (CD45.2), and Group II received TLR2KO Tconv (CD45.2) and WT Treg (CD45.1). These combinations allowed for investigation of the effects of TLR2 signaling on Tconv and Treg populations separately, on an otherwise WT background for myeloid and stromal cells. At 4 and 9 weeks post-infection, Treg accumulation and bacterial burden was analyzed.
Single cell suspensions were prepared from the lungs and spleens derived from the two groups of mice at the indicated time points after Mtb challenge and stained with antibodies against CD4, CD45.1, CD45.2 and Foxp3 for flow cytometric analysis. Lymphocytes were gated on, followed by gating on CD4+ cells. For Group 1, the frequencies of Treg and Tconv were determined by gating on CD45.2 and CD45.1 populations, respectively, within the CD4+ gate, and are presented as percentage out of CD4+ T cells. Similarly, for Group II, the frequencies of Treg and Tconv were determined by gating on CD45.1 and CD45.2 populations, respectively, within the CD4+ gate. The dot plot analysis for lung and spleen is presented in Supplementary Figure 5B and 5C, respectively. Equivalent numbers of CD4+ cells were recruited to the lungs in the two groups of mice at both 4 and 9 weeks post-infection (Figure 5A). Similarly, analysis of the Treg and Tconv populations out of the CD4+ cells showed that both groups had similar frequencies of these populations in the lungs (Figure 5A). Therefore, accumulation of Tregs in Group I mice that received TLR2KO CD4+CD25+ T cells was similar to that of Group II mice that had received WT CD4+CD25+ T cells. Analysis in the spleens showed a similar result as in the lungs, with no differences observed in CD4+ cell numbers or in the frequencies of Tregs between the two groups (Figure 5B). Immunohistochemistry for Foxp3+ cells also demonstrated that Foxp3+ Tregs could be detected in the lungs in both groups (Figure 5C). Consistent with the equivalent Treg/Tconv frequencies observed, Mtb growth kinetics in the lungs were comparable in both groups of mice (Figure 5D). Further flow cytometric analysis of the CD45.1 and CD45.2 populations in the lungs and spleen (Figure S5D–G) of the two groups of mice demonstrated that, in both groups, Foxp3 expression was retained at a similar level and was limited to the congenic marker of the original CD4+CD25+ injected population. These findings confirm that the Tregs which accumulated in the lungs of Group I originated from the TLR2KO CD4+CD25+ T cells injected and were not due to conversion of the injected WT CD4+CD25− population. It is important to note that the immunoregulatory functions of B cells [33] are lacking in the reconstituted Rag2−/− mice. Nonetheless, the findings from this experiment demonstrate that TLR2 signaling on CD4+Foxp3+ Tregs is not necessary for their expansion and subsequent recruitment into Mtb-infected lungs.
Since TLR2 signaling on Tregs themselves does not affect their accumulation into Mtb-infected lungs, we considered that TLR2 signals on pulmonary myeloid cells may play a role in Treg recruitment to granulomatous areas. To directly address this potential role of TLR2 on myeloid cells, we investigated whether providing TLR2-expressing WT myeloid cells to a TLR2KO host would restore normal accumulation of Foxp3+ Tregs in the lung and protect from inflammatory pathology. Macrophages purified from WT or TLR2KO mice were adoptively transferred directly into the lungs by intra-tracheal instillation one day prior to Mtb infection. At seven weeks post-infection, lungs from the TLR2KO mice adoptively transferred with WT macrophages (WT→TLR2KO mice) or TLR2KO macrophages (TLR2KO→TLR2KO mice) were harvested and evaluated for Treg accumulation, cellular infiltration, granulomatous inflammation, and bacterial burden. Flow cytometric analysis of single cell suspensions of lungs showed a significantly higher percentage of Foxp3+ Tregs in the WT→TLR2KO mice than the TLR2KO→TLR2KO mice (Figure 6A) while the percentage of CD4+ T cells was similar in the two groups (Figure 6B). Immunohistochemical staining of lung sections demonstrated high levels of Foxp3+ cell accumulation in the granulomatous lesions of WT→TLR2KO mice (Figure 7A), while very few Foxp3+ cells were observed in the affected areas of the lungs of TLR2KO→TLR2KO mice (Figure 7D). In the latter group, the sparsely recruited Foxp3+ cells were mainly observed in perivascular and peribronchiolar regions (Figure 7D). No background was observed in serial sections stained with isotype control (Figure 7B and E). Consistent with higher Foxp3+ Treg cell accumulation, WT→TLR2KO mice exhibited significantly less cellular infiltration than the TLR2KO→TLR2KO mice (Figure 6C). Histopathological evaluation of lung tissue revealed that the characteristic granulomatous structure with compact aggregation of cells observed in WT mice was restored in the WT →TLR2KO mice (Figure 7C), while lung tissue from the TLR2KO→TLR2KO mice exhibited loosely aggregated lymphocytes and increased inflammation (Figure 7F) with a significantly greater area of lung involvement as expected in a TLR2KO host (Figure 6D). While WT→TLR2KO mice exhibited improved Treg accumulation and decreased inflammatory pathology, the bacterial burden in the lungs (Figure 6E) and spleen (Figure 6F) was similar in both groups, indicating that the role of TLR2 in controlling bacterial burden may be distinct from its role in controlling inflammation. Overall, these findings indicate that TLR2 signaling from macrophages promotes Treg recruitment to the lungs and decreases inflammatory pathology during Mtb infection.
In this study, we conclusively demonstrate that Mtb infection in the absence of TLR2 results in poorly formed granulomas, progressive pulmonary pathology, and increased lung bacterial burden during chronic infection. Our data also indicate that Treg dysfunction may, in part, underlie the immune pathogenesis observed in the lungs of TLR2KO mice. Moreover, we have demonstrated that transfer of WT macrophages significantly enhanced the accumulation of Foxp3+ Tregs within pulmonary granulomatous lesions in TLR2KO mice and concomitantly alleviated pulmonary inflammation. This clearly establishes a causal role for Tregs in controlling the immunopathology of TB. Finally, our data that bacterial burden in the lungs of TLR2KO mice was not affected by the transfer of WT macrophages suggests that the immunoregulatory function of TLR2 can be uncoupled from its antibacterial function. TLR2 ligation activates a multitude of MAPK signaling pathways [34]; it is likely that these distinct pathways regulate the dual functions of TLR2 in chronic infection.
The finding that Tregs expand independently of direct TLR2 signals was surprising given the evidence that TLR2 ligation on Foxp3+ Tregs enhances their proliferation and survival [29]–[31]. These prior reports, however, investigated the effects of TLR2 signaling on Tregs through direct ligation by synthetic TLR2 ligands, not in the context of TLR2 signaling that may occur during an infection. Our study demonstrates that, during Mtb infection, TLR2 signaling on Tregs does not significantly contribute to the expansion and recruitment of Tregs to sites of infection. Instead, TLR2 activation on myeloid cells is necessary to induce the accumulation of Tregs in the lungs. The precise mechanism by which myeloid cells contribute to Treg accumulation remains to be determined, but it is likely dependent on the initiation of an appropriate chemokine axis in the microenvironment of the granuloma. For example, Tregs express the chemokine receptor CCR4 and are highly chemotactic towards its ligands, macrophage-derived chemokine (MDC/CCL22) and thymus and activation regulated chemokine (TARC/CCL17) [35], [36]. While the secretion of CCL22 and CCL17 has been implicated in Treg recruitment to the tumor microenvironment in several studies [37]–[41], the role of these chemokines in the recruitment of Tregs to inflammatory sites during Mtb infection has yet to be addressed. It has also been shown that a subset of Foxp3+ Tregs expressing the Th1 transcription factor T-bet as well as the chemokine receptor driven by T-bet, CXCR3, accumulate at sites of Th1-mediated inflammation [42]. Therefore, a deficiency in any of the IFNγ-inducible CXCR3 ligands may affect Treg recruitment to Mtb granulomas. It is also possible that activation of TLR2 on macrophages may be important for supporting Treg proliferation and maintenance in the lung following recruitment. Future experiments using specific antibodies and gene-deficient mice are necessary to analyze these various possibilities.
Mycobacterial lipids, including TLR2 ligands such as LpqH, are released from Mtb-infected macrophages via exocytosis [43], [44]. Also, Mtb releases membrane vesicles within macrophages that stimulate cytokine and chemokine release in a TLR2-dependent fashion [45]. Therefore, it is conceivable that during chronic infection, Mtb-derived TLR2 ligands are released into the granuloma microenvironment where they can interact with macrophages and perhaps dendritic cells to initiate the chemokine axis required to direct Tregs towards areas of granulomatous inflammation in the lungs and, subsequently, into the granuloma. Our findings indicate that Tregs are necessary to control granulomatous inflammation and maintain a stable granuloma. However, other studies found that natTregs undergo expansion in the blood and at disease sites, and their removal from circulation improved cytokine production from T cells [46]–[48]. Also, Tregs were shown to delay Th1 cell activation in the murine model [49]. Together, these data suggest that the temporal removal of Tregs may be beneficial to the host in enhancing a protective immune response, but, because of the persistent nature of Mtb, it is critical that sufficient numbers of Tregs are recruited to the lungs to mitigate immunopathology. This idea is supported by studies in non-human primates which found that higher frequencies of Tregs correlated to the development of latent TB over active TB disease [50] and that Tregs and Teffector cells acted together to control inflammation without enhancing Mtb replication [51].
The current findings that TLR2 is required for controlling chronic Mtb infection support and extend a previous study [17]. However, these results contradict studies by Holscher and colleagues who reported that TLR2/4 double- and TLR2/4/9 triple-deficient mice are able to efficiently control low dose aerosol infection with Mtb [52], [53]. Reasons for these variable outcomes to Mtb infection are not clear, but they could be related to different experimental conditions, dose of infection, Mtb strain, or perhaps to differences in commensal microbiota. A recent study by Iwasaki and colleagues [54] showed that gut microbiota, critical for maintaining immune homeostasis in the gut mucosa [55], [56], can also influence immunity to infection at a distant site, such as the respiratory mucosa. The authors demonstrated that the commensal microbiota were required for optimal activation of the adaptive immune response against influenza virus infection by providing signal 1 for the expression of mRNA for pro–IL-1β and pro–IL-18 at steady state. This requirement for intact commensal bacteria to generate an appropriate adaptive immune response was found to be restricted to pathogens that are dependent on inflammasomes for immune cell priming, and not to all respiratory pathogens. For example, T cell and B cell responses to Legionella pneumophila were not affected in antibiotic-treated animals. Given that NLP3 inflammasome activation is linked to exacerbated pathology in the lungs of Mtb-infected mice [57], it is possible that microbiota may differentially influence the outcome of Mtb infection in mice bred at different facilities. Our finding that the transfer of WT macrophages restores Treg accumulation and results in improved control of inflammation in TLR2KO mice reduces the likelihood that our observations were influenced by differences in the microbiota of WT and TLR2KO mice that were bred at different sites. Nevertheless, it does not rule out the possibility that differences in microbiota could be a contributing factor to the discrepant results seen between our studies and those of Holcher and colleagues [52]. Depending on the composition of the microbiota, the steady state expression of pro-IL-1β and IL-18 may vary, thus necessitating differences in the requirement of regulatory components to control pulmonary inflammation during Mtb infection. This is an area rich for future investigations.
There are several possible pathways that may be activated by Tregs to prevent immunopathology and associated tissue damage. There is now accumulating evidence to indicate that beyond its anti-microbial function, IFNγ also limits immunopathology in Mtb-infected hosts. IFNγ down-modulates IL-17 production and the subsequent accumulation of pathogenic neutrophils [57]. IFNγ signalingalso dampens the production of the pro-inflammatory cytokines, IL-1α and IL-1β, from myeloid cells [58], via NOS2-mediated inhibition of assembly and activation of the NLRP3 inflammasome [59]. Conspicuous B cell aggregates with characteristic germinal center features are present in the lungs of Mtb-infected mice [60] and emerging evidence indicates a regulatory role for these cells during Mtb infection. Mtb-infected mice deficient in B cells display an exaggerated immunopathology associated with enhanced neutrophil recruitment to the lungs [33]. The enhanced inflammation observed in the TLR2KO mice in our study, however, was not associated with alterations in IFNγ-secreting cells, IL-17 gene expression, or B cell numbers during chronic infection. Despite this, the increased mononuclear cell infiltration suggests that Tregs may function to restrain the influx of monocytes and neutrophils to the lung. Although neutrophils provide protection during acute infection [61], their presence in the lung in chronic infection is associated with pathology ([62] and reviewed in [63]). Thus, by controlling neutrophil recruitment, Tregs may limit inflammation and pulmonary pathology. Also, it is possible that by regulating cellular recruitment, Tregs serve to limit the availability of intracellular niches for Mtb to replicate. Indeed, inhibition of recruitment of new macrophages to the granuloma reduces Mtb numbers in the lung [64], [65]. Although we have delineated that TLR2 controls Mtb growth and inflammation via distinct mechanisms, it is possible that an interplay could exist between the two mechanisms. Future studies will address whether the increased bacterial burden observed in TLR2KO hosts is due to the absence of TLR2-mediated antimicrobial activity within the granuloma or as a consequence of the enhanced inflammation. In sum, our study supports that the TLR2/Treg axis is one of several regulatory circuits that are activated during chronic Mtb infection to safeguard the host against exaggerated inflammation and damage induced by the persistence of Mtb in the lungs.
All animal experiments described in this study conform with the UMDNJ Newark IACUC Guidelines, NIH and USDA policies on the care and use of animals in research and teaching, and the policies of the Guide for the Care and the Use of Laboratory Animals. Animal protocols used in this study were approved by the UMDNJ Institutional Animal Care and Use Committee. (Assurance number A3158-01). Every effort to eliminate animal pain and distress through the use of anesthesia, analgesics or tranquilizers was made.
C57Bl/6 mice and B6-Ly5.2 congenic mice were purchased from The Jackson Laboratory (Bar Harbor, ME). Rag-2-deficient (Rag2−/−) mice were purchased from Taconic Farms, Inc. TLR2-deficient (TLR2KO) mice were developed by S. Akira and colleagues [66]. TLR2KO mice were bred and maintained under pathogen-free conditions at the transgenic animal facility at the UMDNJ-NJMS. M. tuberculosis-infected mice were housed in the BSL3 facility at the Public Health Research Institute at UMDNJ-NJMS. Animal protocols used in this study were approved by the UMDNJ Institutional Animal Care and Use Committee.
The virulent Erdman strain (Trudeau Institute, Saranac Lake, NY) of M. tuberculosis was used for all infections. Bacterial stocks were generated by initial passage in C57Bl/6 mice. Bacterial colonies obtained from lung homogenates were grown in 7H9 media until mid-log phase, and the culture was stored in aliquots at −80°C. The stock titer was determined by plating 10-fold serial dilutions on Middlebrook 7H11 selective medium (Difco). Female mice (age 6–8 weeks) were infected via the respiratory route using a closed-air aerosolization system from In-TOX Products or the Inhalation exposure System from Glas-Col. Mice were exposed for 20 minutes to nebulized bacteria at a density optimized to deliver a standard low dose of around 100 CFU (unless otherwise indicated). For all infections, the actual infection dose was determined by plating total lung homogenates from a minimum of 2 mice on Middlebrook 7H11 plates at 24 hours after aerosol exposure.
Lungs and spleens were harvested at indicated time points post infection. The right superior lobe of the lung was used for determining bacterial burden. The right lower lobe was reserved for histological studies. The right middle lobe was reserved for protein determination. The postcaval lobe was reserved for tissue gene expression. The remaining lung tissue was cut into small pieces and digested with 2 mg/ml collagenase D (Roche) for 30 minutes at 37°C. The digestion was stopped by adding 10 mM EDTA. The digested tissue was forced through a 40 µm cell strainer (BD Falcon) to obtain single cell suspensions. Spleen tissues were processed similarly, but without collagenase digestion. Red blood cells were lysed using ACK lysing buffer (Quality Biological, Inc). The number of viable cells obtained per tissue was determined by trypan blue dye exclusion.
Lung tissue was homogenized in PBS containing 0.05% Tween-80. Total CFU per lung was determined by plating 10-fold serial dilutions on Middlebrook 7H11 plates. CFU were counted after 21 days of incubation at 37°C.
Lung tissue was fixed in 4% paraformaldehyde in PBS for four days, followed by paraffin embedding. For histopathological analysis, five to seven micrometer sections were cut and stained using a standard H&E protocol. For visualization of acid-fast bacilli, tissue sections were stained using the Ziehl-Neelsen method.
For immunohistochemical detection of Foxp3+ cells, tissue samples were de-paraffinized with xylene and rehydrated with ethanol gradations and water. The samples were subjected to heat-induced antigen retrieval by microwave warming using 10 mM citrate buffer (pH 6.0). Endogenous peroxidase activity was blocked using 0.3% hydrogen peroxide and then subsequently blocked with 1× PowerBlock (BioGenex). PBS containing 0.05% Tween-20 was used to wash tissues in between steps. For each sample, serial sections were incubated with the primary anti-mouse/rat Foxp3 antibody (clone FJK-16s; eBioscience) at a 1∶250 dilution or with isotype control (BioLegend) at the same concentration. Sections were subsequently incubated with biotinylated secondary antibody (1∶100 Vector Laboratories). Streptavidin horseradish peroxidase (BioGenex) was used to label the secondary antibody for immunodetection by DAB chromogen (BioGenex). After counterstaining with Mayer's hematoxylin (BioGenex), the samples were dehydrated with ethanol gradations, dipped in xylene, and mounted using Permount (Fisher Scientific).
For quantitation of involved lung area, photomicrographs of H&E stained lung sections were captured using a 5× objective lens. A 546 point grid overlay was superimposed onto each image using Image-Pro Discovery Software, and the numbers of points hitting areas of granulomatous infiltration were counted. The percentage of affected lung tissue was calculated as number of involved points/total points per section ×100.
Single cell suspensions were stained at saturating conditions using specific monoclonal antibodies. All mAbs were directly conjugated to one of the following fluorochromes: Alexa Fluor 488, FITC, PE, PerCpCy5.5, PE-Cy7, APC, or Alexa Fluor700. Isotype controls were included for each. The following mAbs were used in the studies: CD4 (clone RM4–5), CD8 (clone 53-6.7), Foxp3 (clone FJK-16s), CD11c (clone HL3), CD11b (clone M1/70), Gr-1 (Ly-6G and Ly-6C; clone RB6-8C5), CD45.1 (clone A20), CD45.2 (clone 104). Abs to Foxp3, CD45.1, and CD45.2 was purchased from eBioscience. The remaining Abs were purchased from BD Biosciences. For surface staining, cells were re-suspended in FACS buffer (1× PBS +2% fetal calf serum and 0.09% sodium azide) containing a cocktail of mAbs against proteins of interest. For Foxp3 staining, surface staining was performed, followed by fixation, permeabilization, and intracellular staining of Foxp3 according to the manufacturer's protocol (Ebioscience). Following surface or intracytoplasmic staining, samples were fixed in 4% paraformaldehyde for 30 minutes and then acquired on a FACSCalibur or LSRII (BD Biosciences). Analysis was performed using FlowJo software (Tree Star, Inc.). Gating was based on isotype controls.
ELISPOT assay to detect the frequency of Mtb-specific IFNγ producing cells was performed as described previously [67]. 96-well MultiScreen HTS filter plates (Millipore) were coated with 8 µg/ml anti-IFNγ antibody (clone R4-6A2, BD Biosciences). Single cell suspensions from lungs were seeded at 0.25×105, 0.5×105, and 1×105 cells per well. Cells were restimulated with Mtb-infected bone marrow-derived dendritic cells (3 MOI, overnight) at a ratio of 1∶2, or uninfected BMDCs as a control. The cultures were supplemented with IL-2 at 20 U/ml. The cells were co-cultured for 40 hr at 37°C. The plates were subsequently washed with PBS containing 0.05% Tween-20 and treated sequentially with biotinylated secondary antibody (Clone XMG1.2, BD Biosciences), ELISPOT streptavidin-HRP (BD Biosciences), and the HRP substrate 3-amino-9-ethyl-carbazole (Sigma). Spot-forming units were enumerated using an ELISPOT plate reader (Cellular Technology).
Bone marrow-derived dendritic cells (BMDCs) were prepared as described previously [68]. Briefly, bone marrow cells were flushed from the femurs and tibiae of mice with PBS containing penicillin and streptomycin (100 U/ml each). Red blood cell lysis was performed using ACK lysing buffer. 2×106 bone marrow cells were seeded into 10 cm Petri dishes in 10 ml RPMI-1640 media (Mediatech, Inc.) containing 10% defined FBS (HyClone Laboratories, Logan, UT) and supplemented with penicillin (100 U/ml), streptomycin (100 µg/ml), glutamine (2 mM), β-ME (50 µM), and 10% conditioned medium from murine GM-CSF-secreting X63 cells. On Day 3, an additional 10-ml complete medium containing GM-CSF was added to the cultures. On Day 6, the cultures were fed by changing fifty percent of the culture medium. Non-adherent cells were harvested on day 8.
The MACS Regulatory T cell isolation kit (Miltenyi Biotec) was used to separate CD4+CD25− and CD4+CD25+ lymphocytes from spleens and peripheral lymph nodes (axillary and inguinal) of naïve WT congenic B6-Ly5.2 mice (CD45.1+) and TLR2KO mice (CD45.2+). For adoptive transfer, a mixture of 2×106 CD4+CD25− (naïve conventional T cells) and 2×105 CD4+CD25+ (natural regulatory T cells) in PBS were transferred into Rag2−/− mice via retro-orbital injection. Group I received WT CD4+CD25− and TLR2KO CD4+CD25+ cells, and group II received TLR2KO CD4+CD25− and WT CD4+CD25+ cells. One day after transfer, recipient mice were infected with a low dose aerosol of Mtb. At 4 and 9 weeks post-infection, recipient mice were euthanized and lungs and spleens were used for analysis.
Lung lobes were homogenized in 1 ml of TRIzol reagent in lysing matrix D tubes (MP Biomedicals) using a FastPrep homogenizer (MP Biomedicals). Samples were immediately stored at −80°C following lysis in TRIzol. Total RNA was extracted via the manufacturer's TRIzol/chloroform method and purified using RNeasy columns (Qiagen). Total RNA was reverse transcribed using Superscript II RT (Invitrogen). Real-time PCR was performed using the Mx3000P system (Stratagene). TaqMan gene expression assay (Applied Biosystems) for IL-17A and β-actin were used to determine relative IL-17 expression. Relative gene expression was determined by the ΔΔCt calculationt, where ΔCt = Ct (gene of interest) – Ct (normalizer = β-actin) and the ΔΔCt = ΔCt (sample) – ΔCt (calibrator). Total RNA from uninfected lungs was used as calibrator. Baseline gene expression from uninfected WT and TLR2KO was equivalent.
Peritoneal exudate macrophages (PEM) were prepared and adoptively transferred as described previously [69]. Briefly, PEM from WT and TLR2KO mice (5 mice/group) were elicited by intra-peritoneal injection of 2 mls of sterile thioglycollate broth 5 days before peritoneal lavage. PEM from each group of mice were pooled and TLR2KO mice received 2.5×106 WT or TLR2KO macrophages via the intra-tracheal route. One day after transfer, recipient mice were infected with a low dose aerosol of Mtb.
|
10.1371/journal.pgen.1004582 | PRO40 Is a Scaffold Protein of the Cell Wall Integrity Pathway, Linking the MAP Kinase Module to the Upstream Activator Protein Kinase C | Mitogen-activated protein kinase (MAPK) pathways are crucial signaling instruments in eukaryotes. Most ascomycetes possess three MAPK modules that are involved in key developmental processes like sexual propagation or pathogenesis. However, the regulation of these modules by adapters or scaffolds is largely unknown. Here, we studied the function of the cell wall integrity (CWI) MAPK module in the model fungus Sordaria macrospora. Using a forward genetic approach, we found that sterile mutant pro30 has a mutated mik1 gene that encodes the MAPK kinase kinase (MAPKKK) of the proposed CWI pathway. We generated single deletion mutants lacking MAPKKK MIK1, MAPK kinase (MAPKK) MEK1, or MAPK MAK1 and found them all to be sterile, cell fusion-deficient and highly impaired in vegetative growth and cell wall stress response. By searching for MEK1 interaction partners via tandem affinity purification and mass spectrometry, we identified previously characterized developmental protein PRO40 as a MEK1 interaction partner. Although fungal PRO40 homologs have been implicated in diverse developmental processes, their molecular function is currently unknown. Extensive affinity purification, mass spectrometry, and yeast two-hybrid experiments showed that PRO40 is able to bind MIK1, MEK1, and the upstream activator protein kinase C (PKC1). We further found that the PRO40 N-terminal disordered region and the central region encompassing a WW interaction domain are sufficient to govern interaction with MEK1. Most importantly, time- and stress-dependent phosphorylation studies showed that PRO40 is required for MAK1 activity. The sum of our results implies that PRO40 is a scaffold protein for the CWI pathway, linking the MAPK module to the upstream activator PKC1. Our data provide important insights into the mechanistic role of a protein that has been implicated in sexual and asexual development, cell fusion, symbiosis, and pathogenicity in different fungal systems.
| The specific response to environmental cues is crucial for cell differentiation and is often mediated by highly conserved eukaryotic MAP kinase (MAPK) pathways. How these pathways react specifically to huge numbers of different cues is still unclear, and current literature about adapter and scaffolding proteins remains scarce. However, gaining fundamental insight into molecular signaling determinants is pivotal for combating diseases with impaired signal transduction processes, such as Alzheimer's disease or cancer. Importantly, signal transduction can easily be studied in lower eukaryotes like filamentous fungi that are readily genetically tractable. The fungus Sordaria macrospora has a long history as an ideal model system for cell differentiation, and we show here that the proposed cell wall integrity (CWI) MAPK module of this fungus controls differentiation of sexual fruiting bodies, cell fusion, polar growth and cell wall stress response. We further discovered that developmental protein PRO40 binds the MAPK kinase kinase (MAPKKK), the MAPK kinase (MAPKK) and upstream activator protein kinase C (PKC1) of the CWI pathway and is required for MAK1 activity, thereby providing evidence that PRO40 is a scaffold protein. Collectively, our findings reveal a molecular role for a protein implicated in development, cell fusion, symbiosis, and pathogenicity in different fungi.
| Mitogen-activated protein kinase (MAPK) cascades are central components of signaling networks in all eukaryotic organisms [1]–[3]. They consist of a three-tiered module containing a MAPK kinase kinase (MAPKKK), a MAPK kinase (MAPKK), and a MAPK, each activating the subsequent one via phosphorylation. MAPK signaling has been extensively studied in the yeast Saccharomyces cerevisiae, in which five MAPKs have been reported (reviewed in [4]). Three MAPK modules have also been identified in most filamentous fungi, including Aspergillus fumigatus, Magnaporthe grisea, Neurospora crassa, and Sordaria macrospora [5], [6]. Based on homology to S. cerevisiae, they supposedly constitute a cell wall integrity (CWI), pheromone signaling (PS), and osmosensing cascade. Notably, CWI pathway components have been studied in various fungi and are known to be not only responsible for cell wall stress response. For example, in A. fumigatus, the CWI pathway is involved in pyomelanin and gliotoxin formation, response to reactive oxygen species and siderophore biosynthesis [7], [8]. The CWI pathway of N. crassa is necessary for polar growth, conidiation, fusion of conidial germling protrusions (CATs; conidial anastomosis tubes), and fruiting body formation [9]–[11]. CWI MAPKs from Cochliobolus heterostrophus, Coniothyrium minitans, Fusarium graminearum, and Magnaporthe oryzae have further been shown to be involved in female fertility, heterokaryon formation, mycoparasitism and pathogenicity [12]–[15].
Scaffold and adapter proteins are vital for the spatiotemporally correct assembly and signaling output of MAPK pathways and are involved in decision making, thereby enabling highly specific adaptation of signaling pathways [16]–[19]. However, despite many studies already addressing the role of the CWI pathway in fungal development, little is still known about the coordination of this pathway as well as the regulation of specific responses. Further, in S. cerevisiae, the polarisome component Spa2p is known to act as a scaffold-like protein for the CWI MAPKK and MAPK during polar growth [20]. However, information about scaffold or adapter proteins of the CWI pathway is still lacking for filamentous fungi.
In this study, we analyzed the CWI pathway of the model fungus S. macrospora. This filamentous ascomycete has four major advantages over other fungal systems for the study of sexual development (reviewed in [21], [22]). First, it rapidly forms mature fruiting bodies (perithecia) within 7 days. Second, it is self-fertile (homothallic) and thus does not need a mating partner. Moreover, the sexual phenotypes are immediately recognizable. Third, it does not form aerial hyphae with vegetative spores (conidia), which overgrow small pre-fruiting structures like the ascogonial coils (10–20 µm) or the spherical protoperithecia (20–90 µm) and thus prevent their observation. Forth, a collection of developmental S. macrospora mutants is available and well characterized showing defects at different stages of sexual development. Recent analyses have specifically focused on ‘pro’ mutants generating only protoperithecia, a stage the wildtype reaches after 3–4 days. Characterizing these ‘pro’ mutants has led to the identification of several developmental proteins [23]–[28].
Using a forward genetic approach, we identified mutant pro30 as a CWI pathway mutant and generated three CWI kinase deletion strains for functional studies regarding sexual development, cell fusion, vegetative growth, and cell wall stress response. Affinity purification of CWI MAPKK MEK1 revealed that this kinase interacts with the developmental protein PRO40, a protein essential for sexual development and cell fusion [24]. Using further affinity purification-mass spectrometry and yeast two-hybrid analyses, we show that PRO40 binds to MAPKKK MIK1, MAPKK MEK1, and the upstream activator protein kinase C (PKC1). Phosphorylation studies revealed that PRO40 is required for correct signaling via the CWI pathway. Here, we propose a new model in which PRO40 acts as a scaffold protein for the CWI pathway, linking the MAPK module to its upstream activators.
To identify regulators of fruiting body formation, we previously generated a large collection of developmental S. macrospora mutants [22], [25]. One class of mutants was named with the prefix ‘pro’, because these mutants develop only protoperithecia. In this study, we analyzed the underlying mutation in mutant pro30 by next-generation sequencing as described recently [25] (Table S1, SRX483430). SNP analysis revealed a C to T transition at position 904 of the SMAC_03673 gene in the pro30 mutant genome, resulting in a Q302stop substitution at the protein level (Figure S1). Progeny from a cross of pro30 to fus were analyzed, and the mutation was found to strictly co-segregate with the sterile phenotype (Figure S1). SMAC_03673 encodes a 1714 amino acid protein that exhibits 88.4% identity to N. crassa CWI MAPKKK MIK1 (NCU02234, XP_959647.2) and 21.3% identity to S. cerevisiae CWI MAPKKK Bck1p (EWG95039.1, Figure S9), as revealed by BLAST searches [29]. SMAC_03673 was therefore renamed mik1. To confirm that the C to T transition in mik1 is responsible for the mutant phenotype, we transformed pro30 with a full-length copy of mik1. As can be clearly seen from Figure 1, pro30 transformants regained the ability to form perithecia. Thus, a functional MIK1 MAPKKK is required for sexual development in S. macrospora.
The finding that the MIK1 MAPKKK is required for sexual development in pro30 prompted us to analyze the role of CWI pathway kinases in sexual development in more detail. BLAST searches [29] against the N. crassa genome sequence (http://www.broadinstitute.org/annotation/genome/neurospora/MultiHome.html) [30] and the S. cerevisiae genome sequence (http://www.yeastgenome.org/) [31] revealed that MAPKK SMAC_02183 (CCC11961.1) and MAPK SMAC_05504 (CCC12327.1) are homologous to N. crassa MEK-1/S. cerevisiae Mkk1p and Mkk2p (Figure S10) and N. crassa MAK-1/S. cerevisiae Mpk1p (Figure S11), respectively. We subsequently renamed SMAC_02183 and SMAC_05504 mek1 and mak1, respectively. Deletion strains for mik1, mek1, and mak1 were generated by homologous recombination as described previously using S. macrospora Δku70 as host [32]. For generating deletion strains devoid of the Δku70 background, PCR-verified primary transformants were crossed to spore color mutant fus or to sterile mutant pro40 [24], [25]. Subsequent single spore isolation led to Δmik1, Δmek1 and Δmak1 single deletions as verified by PCR and Southern blot analysis (Figure S2, Figure S3, and Figure S4).
We compared sexual development of the three different kinase deletion strains to wildtype. Figure 2A shows sexual structures generated after 7 days of growth on BMM. Like pro30, the kinase deletion mutants did not develop further after protoperithecium formation. Note that mature perithecia were never observed, even after prolonged incubation (Figure 2A). However, sexual development in the deletion strains was re-established by introducing wildtype copies of the deleted genes. Recently, the sexual structures of N. crassa Δmik1, Δmek1, and Δmak1 mutants were described as difficult to detect due to early-onset autolysis [9]. Although we observed autolysis in S. macrospora Δmik1, Δmek1, and Δmak1, protoperithecia were found frequently in Δmik1 and Δmek1 (>200 protoperithecia per microscope slide). It should be noted, however, that we rarely found protoperithecia in Δmak1 (2 protoperithecia on average per microscope slide; data not shown).
A general observation made by ourselves and others is that a defect in sexual development is often linked to a defect in hyphal fusion [23], [33]–[36]. We therefore examined pro30, Δmik1, Δmek1, and Δmak1 for the occurrence of fusion events between vegetative hyphae. Fusion bridges were frequently observed in the wildtype by light microscopy (Figure 2B). However, we were unable to detect fusion bridges in the three kinase deletion mutants as well as in pro30, although hyphae frequently made contact (Figure 2B).
To evaluate the role of MIK1, MEK1, and MAK1 in cell wall stress response, we performed growth tests on medium containing Calcofluor White (CFW). CFW is a common agent used to test fungal mutants for cell wall stress-related defects [37]. We assessed growth during 7 consecutive days in race tubes on synthetic SWG medium ± CFW. Vegetative growth of Δmik1, Δmek1, and Δmak1 was severely impaired even without CFW and was reduced by 70–80% in comparison to wildtype. Figure 3A shows mean values of average growth rates from three independent experiments. Growth of the wildtype on SWG + CFW (gray bar) was reduced by 19% compared to growth on SWG. A much more drastic effect of CFW on vegetative growth was observed in the kinase deletion strains. Growth in these strains was reduced by 62–91% in the presence of CFW (Figure 3A, gray bars), compared to growth on SWG (Figure 3A, black bars). Integration of wildtype copies of mik1, mek1, and mak1 into the respective deletion strains partially complemented the growth defect. This result may be explained by mis-expression of the kinase genes from the constitutive A. nidulans gpd promoter (mik1 and mek1) and the inducible S. macrospora Smxyl promoter (mak1) [38], [39].
Signal transduction via a MAPK module requires several subsequent phosphorylation events, eventually leading to phosphorylation of the MAPK. Western blot analysis with deletion mutants and complemented strains showed that MIK1 and MEK1 are required for MAK1 phosphorylation, confirming the composition of the three-tiered MAPK module (Figure 3B). We further analyzed the localization of MIK1, MEK1, and MAK1 by generating GFP fusions. Functionality of the fusion proteins was confirmed by complementation of the corresponding deletion strains. Fluorescence microscopy showed that MIK1 and MEK1 reside in the cytoplasm and are clearly absent from spherical organelles (Figure 4A). Co-localization experiments using strains with MEK1-GFP and tdTomato-tagged histone H2B identified these organelles to be nuclei (Figure 4B). MAK1-GFP localized to the cytoplasm and the nucleus. This localization pattern is consistent with previously published data, e.g. from the fission yeast Schizosaccharomyces pombe, where MAPKKK Mkh1 and MAPKK Pek1 localize to the cytoplasm, whereas MAPK Pmk1 shuttles between the cytoplasm and the nucleus [40].
As mentioned above, little is known about the regulation of the fungal CWI pathway through scaffolds or adapters. We surmised that these regulators might be identified by searching for interaction partners of the kinases. In a recent transcriptomics analysis, we found the mek1 transcript among the most abundant transcripts in protoperithecia [41], and therefore were prompted to search for MEK1 interaction partners.
MEK1 was fused to an N-terminal tandem affinity purification (TAP) tag consisting of protein A, a TEV protease cleavage site and the calmodulin-binding peptide [42], [43]. TAP in combination with mass spectrometry using multi-dimensional protein identification technology (MudPIT) [44], [45] facilitates the detection of low-abundant proteins. This approach has been applied successfully to the identification of S. macrospora proteins from complex mixtures, and yields a huge number of peptides even after tandem purification [43].
Functionality of NTAP-MEK1 was shown by complementation analysis (Figure 2A), and single spore isolate E292 that expressed high levels of the fusion protein (Figure S5) was chosen for TAP-MudPIT. Proteins that were identified with at least two different peptides in at least two out of four replicate experiments are listed in Table S2. Notably, the most abundant protein detected by MudPIT was MEK1 itself. We further detected other members of the CWI pathway, namely MIK1 and upstream components protein kinase C (PKC1, SMAC_04666; CCC11683.1) and small GTPase RHO1 (SMAC_06239; CCC07244.1). Strikingly, one of the most abundant proteins detected in the MEK1 TAP-MudPIT analysis was PRO40 (CCC06426.1; Table S2). This protein has previously been shown to be involved in sexual development and hyphal fusion [24]. Fungal PRO40 homologs have been connected to sexual and asexual development, cell fusion, symbiosis, and pathogenicity, but their exact molecular function is still currently unknown [35], [46]–[48].
To verify the MEK1-PRO40 interaction, we performed affinity purification of FLAG-tagged PRO40 (FLAG-AP) followed by MudPIT. The PRO40-FLAG fusion construct has already been shown to complement pro40 [24]. For FLAG-AP, we used strain T184.2NS11 (Δpro40::pro40-3xFLAG) yielding detectable amounts of PRO40-FLAG in the eluate (Figure S5).
FLAG-AP in combination with MudPIT was performed three times with T184.2NS11 and twice with wildtype (control). A full list of proteins detected with at least two different peptides in at least two out of three (PRO40-FLAG) or two out of two (wildtype) replicate experiments is given in Tables S3 and S4, respectively. Due to the single FLAG purification step, the number of identified proteins in the PRO40-FLAG dataset was even larger (444 proteins) than the number of proteins identified in the TAP-MEK1 dataset (308 proteins). However, five proteins were identified consistently in all three experiments with PRO40-FLAG, but not with wildtype control experiments. Of these five proteins, three have been assigned functions in cell differentiation. PKC1 is an upstream activator of the CWI module, COP9-2 (SMAC_01284; CCC07717.1) is a subunit of the COP9 signalosome, which has been described to regulate sexual development in A. nidulans, and PRO4/LEU1 (SMAC_07082; CCC13458.1) is an enzyme that is involved in the leucine biosynthesis pathway, which is essential for fruiting body formation in S. macrospora [49]–[51]. Since we detected PRO40 in one of the wildtype control experiments (Table S4), we calculated ratios between the spectral counts in PRO40-FLAG and control experiments to evaluate the specificity of such proteins detected in both experiment types (see Materials and Methods, Table S5). This approach reduced the number of high-confidence hits for PRO40 interaction partners to 123 proteins, including MEK1 and RHO1.
In addition to TAP-MudPIT with NTAP-MEK1 in a Δmek1 background, we also performed TAP-MudPIT in a Δpro40 background (strain E2544; Δpro40::NTAP-MEK1; Figure S5). Proteins that were identified with at least two different peptides in at least two out of three replicate experiments were considered as high-confidence interactors (Table S2). Among these, we identified MIK1 and RHO1, but not PKC1. Interestingly, MEK1 seems to interact with the Woronin body protein HEX1 (SMAC_01601; CCC08037.1) independent of whether or not PRO40 is present, since HEX1 was detected in both the Δmek1::NTAP-MEK1 and the Δpro40::NTAP-MEK1 datasets (Table S2).
From our results, we subtracted known background (a list of proteins considered background derived from numerous unrelated affinity purification-MS experiments is provided in Table S6), and by comparing the three datasets (Δpro40::PRO40-FLAG, Δmek1::NTAP-MEK1, and Δpro40::NTAP-MEK1) found overlaps between the PRO40 and MEK1 interaction networks (Figure 5, Table 1). Specifically, we found 12 proteins in all three datasets. From the CWI pathway, this group contains MEK1 and RHO1. Another 17 proteins appeared in the Δpro40::PRO40-FLAG and Δmek1::NTAP-MEK1 datasets, but not in the Δpro40::NTAP-MEK1 dataset (Figure 5, Table 1), indicating that interaction of MEK1 with these 17 proteins depends on the presence of PRO40. Among these 17 proteins was PKC1, ATP citrate lyase ACL1 (SMAC_06775; CCC07573.1), previously found to be involved in S. macrospora sexual development [52], and a putative regulatory subunit of protein phosphatase PP2A, RTS1 (SMAC_02633; CCC10054.1). This regulatory subunit is of high interest because of a recently described fungal striatin-interacting phosphatase and kinase (STRIPAK) complex, containing protein phosphatase 2A (PP2A) and several ‘PRO’ proteins [43].
As mentioned above, we recently performed a transcriptomics analysis of protoperithecia by laser microdissection and RNA-seq [41]. When we superimposed data from this study on the Venn diagram in Figure 5, we found that the group of interaction partners shared by PRO40 and MEK1 contains proteins whose transcripts belong to the top500 genes (with respect to read counts) in either vegetative growth conditions (blue) or protoperithecia (magenta). For example, besides the mek1 transcript, the rho1 and acl1 transcripts are in the top500 list in protoperithecia (Figure 5; [41]). These data highlight the significance of the MEK1-PRO40 interaction network for fungal sexual development.
We further searched for PRO40 interaction partners by performing yeast two-hybrid assays with full-length PRO40 as bait. For prey, we generated two S. macrospora Matchmaker libraries using cDNA derived from different cultures to include cDNA from vegetatively and sexually propagating mycelia. Screening of 107 yeast cells yielded 1600 clones with ade and his reporter gene activity, and 96 clones additionally showing lacZ reporter gene activity in two subsequent assays were subjected to DNA isolation and sequence analysis, which led to the identification of 13 different genes (Table S7). 11 of the 96 clones carried mek1 sequences. To assess the strength of the PRO40-MEK1 interaction, we performed quantitative β-galactosidase assays with a strain carrying BD-PRO40 and AD-MEK1_v01. Note that MEK1-v01 corresponds to an N-terminally truncated MEK1, which is due to an annotation error in the S. macrospora genome version 01 vs. version 02 [5], [41]. Mean values of β-galactosidase activity of three independent experiments were 86.7±16.9 U/mg protein compared to 5.0±2.4 U/mg protein for the control experiment (BD-PRO40 and AD), indicating a strong interaction between PRO40 and MEK1.
To gain further insight into the protein interactions within the deduced multi-subunit complex, we tested reciprocal interaction between the six proteins, PRO40, RHO1, PKC1, MIK1, MEK1, and MAK1, in a yeast two-hybrid assay. We used constitutively active (RHO1_CA) and inactive (RHO1_CI) versions of RHO1 to determine interactions dependent on RHO1 activity. Protein structures of all tested proteins are given in Figure 6A. Mating of yeast strains carrying GAL4-AD and -BD translational fusions with the abovementioned proteins resulted in diploid cells that were tested for reporter gene activity. As can be seen from growth of yeast colonies on SD medium lacking adenine and histidine, PRO40 interacted with PKC1, MIK1, and MEK1, but not RHO1 and MAK1 (Figure 6B; A growth control of yeast colonies is shown in Figure S6). MEK1 showed interaction with PRO40 and MIK1, as seen in the TAP-MudPIT analysis, and with MAK1 only in the yeast two-hybrid analysis. However, we did not detect interactions between MEK1 and RHO1 or MEK1 and PKC1. Formation of homodimers was observed for PRO40, PKC1, MIK1, and MAK1. Figure 6C displays a schematic overview of signal transduction and protein-protein interactions in the PRO40-CWI complex.
To map PRO40 domains mediating interaction, we generated yeast two-hybrid vectors containing cDNA fragments of pro40 (Figure 6A). PRO40 contains a WW domain, which is implicated in mediating protein-protein interactions [53], and several regions enriched for certain amino acids such as asparagine and glycine. Further, the PRO40 N-terminal half, enriched for glutamine and proline, is predicted to be highly disordered by IUPred [54], [55], as was also described for the N. crassa PRO40 homolog SOFT [56]. For yeast two-hybrid analysis, we generated five overlapping fragments PRO40a-e, containing different domains. As can be seen from Figure 6B, the N-terminal proline-rich and disordered PRO40 derivative, PRO40a, interacted with PRO40 itself as well as MEK1. PRO40c interacted with all full-length PRO40 interaction partners, namely PRO40, PKC1, MIK1, and MEK1. Since fragment PRO40c contains the WW domain, we hypothesized that binding of PRO40 to these interaction partners might be mediated by this domain. To test this hypothesis, we generated a modified PRO40, PRO40AAA (W575A, W598A, P601A), inserting mutations described to render the WW domain non-functional [57]. Surprisingly, PRO40AAA showed the same interactions as full-length PRO40 in a yeast two-hybrid assay (Figure 6B). PRO40AAA was further able to interact with itself.
We next generated MEK1 derivatives for yeast two-hybrid analysis (Figure 6A). As shown in Figure 6B, MEK1d comprising the kinase domain was sufficient for interaction with PRO40, MIK1, and MAK1. Moreover, interaction between MEK1 and MAK1 was accomplished via MEK1a and MEK1c, both comprising the proline-rich region. For interaction of MEK1 and PRO40, the MEK1 kinase domain (MEK1d) and either proline-rich PRO40a or central PRO40c were sufficient (Figure 6B). Figure 6D summarizes the results from interaction studies with PRO40 and MEK1 derivatives. Taken together, our data indicate that PRO40 connects the CWI MAPK module to its upstream activator PKC1.
To gain insight into the biological function of the MEK1-PRO40 interaction, we generated a Δmek1/pro40 double mutant. Single spore isolates from crosses of the single mutants were subjected to sequencing of the pro40 ORF and Southern blot analysis (Figure S7). The Δmek1/pro40 double mutant was analyzed with regard to sexual development and hyphal fusion and showed the same developmental phenotype as the Δmek1 and Δpro40 single deletion strains (Figure 7A, B, compare Figure 2A, B). To test whether Δpro40 shares the stress response phenotype of the CWI kinase deletion strains, we performed growth tests on SWG ± CFW. As can be seen from Figure 3A, Δpro40 was less impaired in vegetative growth than the kinase deletion strains, and the growth reduction in the presence of CFW was also less distinctive. The Δmek1/pro40 double mutant was much more impaired than Δpro40 and showed the same growth defect as the Δmek1 single deletion strain. From these observations, we conclude that PRO40 does not play a major role in the cell wall stress response.
Since PRO40 interacts with three members of the CWI kinase pathway, we investigated whether the localization of MIK1, MEK1, and MAK1 was altered in the pro40 mutant or the Δpro40 deletion strain. Localization of the kinases in vegetative hyphae was identical in mutants and wildtype, with MIK1 and MEK1 residing in the cytoplasm and MAK1 localizing to the cytoplasm and the nucleus (Figure 7C, compare Figure 4A). Since PRO40 and the kinases are required for sexual development, we also investigated the localization of MIK1, MEK1, and MAK1 in protoperithecia of pro40, Δpro40, and wildtype. As illustrated in Figure 7D, GFP-MIK1 and MEK1-GFP showed a uniform distribution in protoperithecia of all investigated strains, most likely due to cytoplasmic localization. Similarly, the MAK1-GFP signal was found in the cytoplasm, and additionally in patches that resemble nuclei. In summary, the kinases showed the same localization in the presence and the absence of PRO40, in vegetative as well as in sexual tissues.
To clarify the putative scaffolding role of PRO40 for the CWI pathway in more detail, we compared MAK1 phosphorylation levels in wildtype to MAK1 phosphorylation levels in Δpro40 and the pro40 overexpression strain T182.4NS11 (Figure 8). First, we examined MAK1 phosphorylation during a developmental time course. As can be seen from Figure 8A, MAK1 activity is strongly reduced in Δpro40 in comparison to wildtype at all investigated time points. In contrast, MAK1 is hyper-phosphorylated in the pro40 overexpression strain T184.2NS11. Second, we assayed MAK1 phosphorylation under stress conditions (Figure 8B). In the wildtype, MAK1 phosphorylation was strongly induced by H2O2. Again, MAK1 activity was strongly reduced in Δpro40; however, a response to 15 minutes H2O2 stress was still evident in the mutant. The pro40 overexpression strain displayed MAK1 activity similar to wildtype. Thus, we concluded that PRO40 acts as a scaffold protein for CWI pathway components during sexual development, hyphal fusion, and stress response (Figure 8C).
In this study, we investigated the S. macrospora CWI kinase pathway and showed that the three kinases MIK1, MEK1, and MAK1 are required for the transition from protoperithecia to perithecia. Although the three deletion strains reach the same level of protoperithecia development, we observed that Δmak1 displayed significantly fewer protoperithecia than Δmik1 and Δmek1. This observation indicates that MAK1 obtains input not only from the upstream CWI kinases, but also from other pathways. Crosstalk between the CWI pathway and other stress response pathways has been observed in a number of fungi (reviewed by [58]). For example, the N. crassa PS and CWI pathways have both been described to control the cell wall stress response, hyphal fusion, and sexual development [10]. Here, we performed a large-scale analysis of MEK1 interactions in two different background strains. Our data establish a basis to functionally analyze further interaction partners of MEK1 for a regulatory role in CWI signaling.
In this study, we revealed the developmental protein PRO40 as a scaffold protein for the CWI pathway. S. macrospora PRO40 was previously identified by complementation of sterile mutant pro40, harboring a transition in the pro40 gene that leads to an early translational stop [24]. Like its homolog SOFT from N. crassa, PRO40 is important for cell fusion [34], [35]. PRO40/SOFT homologs have further been shown to be important for cell fusion in Alternaria brassicicola, Epichloë festucae, F. graminearum, and F. oxysporum [46]–[48], [59]. In addition, PRO40/SOFT has been connected to pathogenicity in A. brassicicola, F. graminearum, and F. oxysporum and to symbiosis in E. festucae [46]–[48], [59]. Recently, N. crassa SOFT was shown to display an oscillatory localization at the tips of CATs approaching cell fusion, which alternates with the PS pathway MAPK MAK2 [60]. It was also proposed that SOFT and MAK2 act in two different signaling pathways with one sending and one receiving a yet unknown signal eventually leading to fusion of two CATs. Our data reveal that PRO40 of S. macrospora is a scaffold protein for another signaling pathway, namely the CWI pathway. Thus, a conceivable mechanism of PRO40 during CAT fusion is the regulation of the CWI kinases. All CWI kinase mutants of N. crassa have been described to be CAT fusion deficient [10], [61], and it should be highly elucidating to analyze the localization of the kinases during CAT fusion.
PRO40/SOFT homologs have been found in stress granules (A. oryzae), at septal pores in response to various stresses or hyphal injury (A. oryzae, N. crassa, S. macrospora), as well as associated with Woronin bodies (S. macrospora) [24], [62]–[64]. Woronin bodies are peroxisome-derived organelles found only in filamentous ascomycetes and contain a crystalline core assembled from the HEX1 protein [65]. An association of S. macrospora PRO40 with Woronin bodies was already found by co-localizing PRO40 with GFP-tagged HEX1 [24]. HEX1 was not identified as a significant PRO40 interaction partner in our experimental setup (Table S5); however, MEK1 was found to interact with HEX1 and Leashin, the Woronin body tether [66]. A connection of Woronin body function to the CWI pathway has already been observed in A. oryzae. There, PKC1 is required for HEX1 phosphorylation and subsequent HEX1 self-assembly [67]. The main function of Woronin bodies is the plugging of septal pores after hyphal injury, but they also play a role in plant infection and survival of nitrogen depletion in M. grisea [65], [68]. The Woronin body protein HEX1 is involved in asexual reproduction and virulence in F. graminearum, and hex1 has been found to be regulated by PS MAPK MAK2 and downstream transcription factor PP-1, homologous to yeast Ste12p, in N. crassa [69], [70]. These findings are in agreement with a possible developmental role of Woronin bodies and the HEX1 protein.
By yeast two-hybrid analysis, we found two regions of PRO40 to be important for the observed interactions with CWI pathway components, namely the N-terminal proline-rich part and the central region encompassing the WW domain. Most interestingly, the N-terminal region of PRO40, including the proline-rich part, is also highly disordered. Protein disorder has been recognized as an important feature in signaling, since conformational fluctuations in disordered regions allow highly specific binding to multiple interaction partners in a regulated manner, thereby increasing functional capability [71]. Further analysis is needed to ascribe such versatile functions to PRO40 disordered regions. Another protein-protein interaction domain, the WW domain, has been found in PRO40. However, our yeast two-hybrid studies show that it is dispensable for interaction with PRO40, PKC1, MIK1, and MEK1. This indicates that PRO40 contains further unrecognized protein-protein interaction motifs within the region encompassing the WW domain.
Scaffold proteins are defined as proteins that not only bind to different signaling proteins, but that also attune signaling outputs [17], [72]. Although PRO40 does not affect the localization of MIK1, MEK1, and MAK1, it affects MAK1 phosphorylation, both during development and during stress response. Thus, PRO40 is a scaffold protein of the CWI pathway. We further attempted to address the question how PRO40 affects signaling via the CWI pathway by complementation analysis with constitutively active MAK1 and MEK1, inserting previously described mutations [73], [74]. However, these MAK1 and MEK1 versions were unable to reinstate wildtype morphology in the corresponding deletion mutants and thus were inept for further studies (our unpublished results). Since pro40 mutants did not display the same general growth defect as the CWI kinase mutants and were not as impaired as these mutants on media containing CFW, we conclude that PRO40 does not act as a scaffold of the CWI pathway during all CWI pathway functions. Recently, the presence of different CWI pathways has been suggested for N. crassa. There, different membrane sensors, WSC-1 and HAM-7, activate signaling via the CWI module, leading to cell wall stress response and hyphal fusion, respectively [75]. Our data strongly suggest that PRO40 acts as a scaffold protein for the CWI pathway during fungal development, hyphal fusion, and stress response.
In conclusion, we have identified PRO40 as a new scaffold protein of the highly conserved CWI pathway, linking the MAPK module to upstream activator PKC1. Collectively, our findings provide important insights into the mechanistic role of a fungal protein that has been implicated in sexual and asexual development, cell fusion, symbiosis, and pathogenicity in diverse fungal systems.
Cloning and propagation of recombinant plasmids was performed using standard laboratory conditions [76] and Escherichia coli strain XL1 Blue MRF' [77] as host for plasmid amplification. Alternatively to restriction-ligation-mediated cloning, recombinant plasmids were generated by homologous recombination in S. cerevisiae strains PJ69-4a, AH109 or Y187 [78], [79; Clontech, Palo Alto, CA, USA] as described previously [43], [80]. Recombinant yeasts were selected by prototrophy to leucine, tryptophan or uracil. Yeast experiments were carried out according to standard protocols (Clontech Yeast Protocol Handbook, PT3024-1), and plasmid isolation was performed as described by Bloemendal et al. [43].
The wildtype strain (S91327) of S. macrospora was obtained from our laboratory collection. Details for all S. macrospora strains used in this study are given in Table S8. Unless stated otherwise, standard growth conditions and DNA-mediated transformation were performed as described previously [81], [82]. Transformants were selected on medium containing either nourseothricin (50 µg/ml) or hygromycin B (80 U/ml). Sensitivity to Calcofluor White (CFW; Sigma Aldrich, St. Louis, MO, USA) was measured in 30 cm race tubes containing 15 ml solid SWG medium (derived from synthetic crossing medium according to Nowrousian et al. [83])±250 µg/ml CFW. For each strain, two race tubes were measured in each experiment and the growth front was marked every 24 hrs for 7 consecutive days. Preparation of DNA and Southern hybridization were performed as described [81].
S. macrospora cDNA libraries SmI and SmII were generated using the Matchmaker Library Construction & Screening Kit (Clontech, Palo Alto, CA, USA). RNA was extracted from S. macrospora wildtype according to Pöggeler et al. [84] from 3 and 6 days old floating cultures (inducing sexual development) and 3 days old shaking cultures (repressing sexual development), and mRNA was isolated with the polyATtract mRNA isolation kit (Promega, Madison, WI, USA). For SmI, a mixture of 15 mRNA extractions and for SmII, a mixture of 7 mRNA extractions was used to generate cDNA according to the Matchmaker Library Construction & Screening Kit manual (PT3955-1, Clontech, Palo Alto, CA, USA). Each cDNA mixture was co-transformed with pGADT7-Rec into yeast strain AH109 and plated on SD medium lacking leucine. Colonies were harvested after 6 days of growth. Library titers were 4.3×107/ml and 1.5×107/ml for SmI and SmII, respectively.
All plasmids and oligonucleotides used in this study are listed in Tables S9 and S10, respectively. For yeast two-hybrid analyses, PCR was performed on S. macrospora cDNA and PCR fragments cloned into pGBKT7 and pGADT7 as follows:
For RHO1, constitutively active (RHO1_CA) and constitutively inactive (RHO1_CI) versions were generated by inserting mutations G15V/C191S and E41I/C191S, respectively [85]. Specifically, a RHO1_CA fragment was generated from S. macrospora cDNA using primers rho1_CA-for/rho1_CA-rev and ligated EcoRI/BamHI into pGADT7 and pGBKT7 to generate pA-RHO1_CA and pB-RHO1_CA, respectively. For pA-RHO1_CI, two PCR fragments generated with primer pairs rho1_CI-for x rho1_CIint-rev and rho1_CI-rev x rho1_CIint-for, were co-transformed into yeast with SmaI-digested pGADT7. pB-RHO1_CI was generated by ligating a 0.6 kb EcoRI/BamHI fragment from pA-RHO1_CI into EcoRI/BamHI-digested pGBKT7.
To generate pA-PKC1, yeast recombination was performed with SmaI-digested pGADT7 and three cDNA fragments amplified by PCR using primer pairs 4666-01-AD/4666-02, 4666-03/4666-04, and 4666-05/4666-06-AD. For pB-PKC1, a similar strategy was employed with SmaI-digested pGBKT7 and three PCR fragments produced with primer pairs 4666-01-BD/4666-02, 4666-03/4666-04, and 4666-05/4666-06-BD.
To generate pA-MIK1, yeast recombination was performed with EcoRI/BamHI-digested pGADT7 and five PCR fragments produced with primer pairs 3673-1-AD/3673-2, 3673-3/3673-4, 3673-5/3673-6, 3673-7/3673-8, and 3673-9/3673-10-AD. For cloning of pB-MIK1, yeast recombination was again employed with SmaI-digested pGBKT7, two PCR fragments produced with primer pairs 3673-1-BD/3673-2 and 3673-9/3673-10-BD, and a 5254 bp NdeI/BamHI fragment from pA-MIK1.
For mek1 vectors, a 1458 bp mek1 cDNA fragment was amplified with primers 6419-9/6419-10 and ligated EcoRI/BamHI into pGBKT7 and pGADT7 to generate pB-MEK1_v01 and pA-MEK1_v01, respectively. For full-length cDNA vector pA-MEK1, pA-MEK1_v01 was digested with EcoRI and transformed into yeast together with a 425 bp BglII/BamHI fragment of pA-MEK1a (see below). Likewise, pB-MEK1 was generated by yeast recombination using EcoRI-linearized pB-MEK1_v01and a 653 bp XhoI/BamHI fragment of pB-MEK1a.
For mak1 vectors, a PCR-fragment produced with primers HR-mak1-for/HR-mak1-rev as well as BamHI-digested pGADT7 were transformed into yeast, generating pA-MAK1. pB-MAK1 was generated by ligation of a 1426 bp EcoRI/PstI fragment from pA-MAK1 into EcoRI/PstI-digested pGBKT7.
For pB-PRO40, full-length cDNA was amplified using primers Y2H-05/Y2H-06neu and ligated EcoRI/PstI in pGBKT7. pA-PRO40 was generated by yeast recombination of two PCR fragments produced with primer pairs AD-40-for1/AD-40-rev1 and AD-40-for2/AD-40-rev2, and a 3961 bp EcoRI/PstI fragment from pB-PRO40 into pGADT7/BamHI.
To generate pA-MEK1a and pB-MEK1a, PCR was performed on S. macrospora cDNA using primer pair mek1_F1-fw/mek1_F1-rv, the PCR fragment subcloned into pDrive, cut EcoRI/BamHI and ligated into EcoRI/BamHI digested pGADT7 and pGBKT7, respectively. pA-MEK1b/pB-MEK1b, pA-MEK1c/pB-MEK1c and pA-MEK1d/pB-MEK1d were generated accordingly, using primer pairs mek1_F2-fw/mek1_F2-rv, mek1_F1-fw/mek1_F2-rv and mek1_F4-fw/6419-10, respectively.
To generate yeast two-hybrid vectors encoding PRO40 derivatives, five pro40 fragments were amplified from cDNA and subsequently ligated into EcoRI and BamHI sites of pGADT7 and pGBKT7. Primers used were Y2H-13/Y2H-07 for PRO40a, Y2H-08/Y2H-09 for PRO40b, Y2H-01/Y2H-12 for PRO40c, Y2H-10/Y2H-11 for PRO40d, and Y2H-03/Y2H-04 for PRO40e. pA-PRO40e was generated by ligating a 1007 bp EcoRI fragment from pB-PRO40e into the pGADT7 EcoRI site.
Vector pB-PRO40AAA encoding PRO40 with a mutated WW domain (PRO40 W575A, W598A, P601A) was generated by yeast recombination of a PCR fragment (primers 40-6/40-7) into ScaI-digested pB-PRO40. To generate pA-PRO40AAA, a 3961 bp EcoRI/PstI-fragment from pB-PRO40AAA and a 10322 bp SalI fragment from pA-40 were recombined in yeast.
The S. macrospora pro40 cDNA was used as bait to screen both S. macrospora cDNA libraries for interacting proteins using the Matchmaker System (Clontech, Palo Alto, CA, USA). Yeast Y187 cells were transformed with pB-PRO40, mated with 1 ml cDNA library and plated on selective media (SD-trp-leu-ade and SD-trp-leu-his-ade). Colonies were re-inoculated on selective media lacking histidine or adenine and histidine. Growing yeast cells were subjected to two subsequent lacZ filter tests (Clontech Yeast Protocol Handbook, PT3024-1). 96 randomly chosen colonies showing reporter gene activity were chosen for PCR amplification of cDNA inserts using primer pair pAD-2/pAD-FPneu and PCR products were directly sequenced with primer pADfor96er. Quantitative measurements of β-galactosidase activity were carried out as described previously [86].
To test interactions between full-length proteins as well as derivatives of MEK1 and PRO40, strains carrying single plasmids were generated by electroporation [87] using matα strains (Y187, PJ69-4α) and mata strains (AH109, PJ69-4a) as recipients for BD and AD fusion constructs, respectively. Diploid strains were generated and tested for reporter gene expression as previously described [88]. For drop plating, yeast colonies were resuspended in 200 µl SD medium and 5 µl were spotted on SD supplemented with histidine and adenine as well as SD lacking histidine and adenine. Due to transactivation, pB-MEK1 was exchanged for pB-MEK1_v01, and pB-MEK1b, pB-MEKc and pA-RHO1_CA were omitted from the analysis.
Deletion vectors for mik1 and mak1 were generated by yeast recombination as described [43]. For pKO-MIK1, 5′ (1000 bp) and 3′ (1000 bp) flanking regions of mik1 were PCR-amplified using S. macrospora genomic DNA and primer pairs 3673-5fw/3673-5rv and 3673-3fw/3673-3rv, respectively. Flanking regions were transformed into yeast together with an hph cassette cut EcoRI from plasmid pDrivehph [89], and EcoRI/XhoI-linearized vector pRS426 [90]. Plasmid pKO-MAK1 was generated accordingly, using mak1 5′ (1000 bp) and 3′ (1039 bp) flanking regions amplified with primer pairs 5504-5fw/5504-5rv and 5504-3fwIT/5504-3rvIT, respectively.
To generate a mek1 deletion, 5′ (832 bp) and 3′ (913 bp) flanking regions of mek1 were PCR-amplified using primer pairs KO-mek-1/KO-mek-2 and KO-mek-3/KO-mek-4, respectively, and subcloned into pDrive. Due to annotation changes concerning mek1 in genome version 02 of S. macrospora [25], a 5′-truncated version (mek1_v01, nt280–1858, encoding amino acids 35–519) was used as basis for generating mek1 deletion and TAP vectors. The 5′ and 3′ regions were cut SnaBI/BamHI and XbaI/ApaI and successively ligated into the corresponding sites of vector pDrive-Hyg (I. Godehardt and U. Kück, unpublished data). Linearized pKO-MIK1, pKO-MEK1 and pKO-MAK1 were transformed into S. macrospora Δku70 [32] and transformants were selected for by hygromycin resistance. Single-spore isolates in which mik1, mek1 or mak1 had been replaced by the hph cassette and which had the wildtype genetic background were obtained as described previously through crosses against spore color mutant fus or mutant pro40 [24], [25], [32].
For pNTAP-mik1, yeast recombination was employed. Fragments used for transformation were BamHI-digested pDS21 [91] and five PCR products generated with S. macrospora genomic DNA and primer pairs NTAP-mik-fw/3673-2, 3673-3/3673-4, 3673-5/3673-6, 3673-7/3673-8, and 3673-9/NTAP-mik-rv. pRSnat-gfp-mik1 was generated by amplification of egfp from pDS23 (M. Nowrousian, unpublished) using primers Pgpd_egfp_for/mik1_egfp_rev and subsequent recombination in yeast with HindIII-digested pNTAP-mik1. For pGFP-MIK1_NA, 5′ and 3′ mik1 sequences were amplified from S. macrospora genomic DNA with primer pairs 3673-5fw/3673-5rv-gfp and 3673-11/3673-3rv, respectively, and subsequently recombined into pRSnat-gfp-mik1, replacing the gpd promoter and trpC terminator.
For complementation and TAP, mek1 was amplified from genomic DNA with primers mek1-BamHI-fw/NTAP-mek-BamHI-rv, subcloned into pDrive, cut with BamHI and cloned into BamHI-digested pDS21 [91], generating pNTAP-MEK1. Vector pRSnat-mek1-gfp_V3 was generated by amplifying mek1 from S. macrospora genomic DNA with primer pair Pgpd-mek1_V3/gfp-mek-rev and recombination into linearized pDS23 in yeast. For pMEK1-GFP_NA, 5′ and 3′ mek1 sequences were amplified from S. macrospora genomic DNA with primer pairs 2183-5fw_IT/mek1_F1-rv and 2183-3fw-gfp/2183-3rv_IT, respectively, and transformed into yeast together with a 2.8 kb PvuII-SpeI fragment from pRSnat-mek1-gfp_V3 and linearized pRSnat [92].
A mak1 complementation vector was constructed by amplifying mak1 using primers Pxyl-mak-for/NTAP-MAK-rv, and recombining the PCR fragment into NotI/BamHI-digested pNpX-GFP [38], yielding pNpX-MAK1. Vector pRSnat-mak1-gfp was generated by amplifying mak1 from genomic DNA with primer pair CTAP-mak1-fw/GFP-mak1-rv and transforming the PCR fragment in yeast together with HindIII-linearized pDS23. For pMAK1-GFP_NA, 5′ (5504-5fw/GFP-mak1-rv) and 3′ (5504-3fw-gfp/5504-3rv_IT) mak1 sequences were amplified and transformed in yeast together with a 0.8 kb BamHI fragment from pRSnat-mak1-gfp and linearized pRSnat.
To search for PRO40 interaction partners, pC-FLAG-PRO40 [24] was transformed into Δpro40 and single spore isolate T184.2NS11 was used for further analysis. For FLAG-AP, dried mycelium was ground in liquid nitrogen, suspended in FLAG extraction buffer (50 mM Tris-HCl pH 7.4, 250 mM NaCl, 10% glycerol, 0.05% NP-40, 1 mM PMSF, 0.2% protease inhibitor cocktail IV (Calbiochem), 1 mM benzamidine, 1 µg/ml leupeptin) and centrifuged for 30 min at 16000 rpm. 50 ml crude protein extract was incubated with 300 µl anti-FLAG M2 affinity gel (A2220, Sigma Aldrich, St. Louis, MO, USA) overnight at 4°C on a rotator. Bound complexes were collected by centrifugation and washed twice in 45 ml and once in 1 ml cold washing buffer (50 mM Tris-HCl pH 7.4, 150 mM NaCl, 0.05% NP-40, 1 mM PMSF, 1 mM benzamidine, 1 µg/ml leupeptin) with rotation at 10 min intervals. The affinity gel was transferred to a 1.5-ml centrifuge tube and incubated in 500 µl of cold washing buffer containing 2 µl protease inhibitor cocktail IV (Calbiochem) and 0.5 mg/ml 3× FLAG peptide (F4799, Sigma Aldrich, St. Louis, MO, USA) for 6 hr at 4°C on a rotator. After centrifugation, the supernatant was transferred to a new 1.5-ml tube, the gel briefly washed in 500 µl cold washing buffer, centrifuged and the supernatant was combined with the first supernatant. Purified complexes were subjected to trichloroacetic acid precipitation and directly used for mass spectrometry.
For TAP analysis, pNTAP-mek1 was transformed into S. macrospora Δmek1 and Δpro40, and transformants were selected on media with nourseothricin. Primary transformants expressing NTAP-MEK1 were used for single spore isolation. For protein extraction, S. macrospora strains E292 (Δmek1::NTAP-MEK1) and E2544 (Δpro40::NTAP-MEK1) were grown in P-flasks with BMM liquid medium for 3 d at 27°C. TAP analysis was performed as described previously [43].
Tryptic digestion of proteins and MudPIT analysis [44], [45] were performed as described previously [43] using an Orbitrap Velos ion trap mass spectrometer coupled to an Accela quaternary U-HPLC pump (Thermo Fisher Scientific). Proteome Discoverer software version 1.2 was used for MS/MS data interpretation, and data were searched against the S. macrospora database (smacrosporapep_v4_110909) with tryptic peptides, mass accuracy of 10 ppm, fragment ion tolerance of 0.8 Da, and with oxidation of methionine as variable modification allowing 4 missed cleavage sites. All accepted results had a high peptide confidence with a score of 10. Proteins identifies with at least two different peptides in at least two of three to four independent experiments were considered for further analysis.
To identify contaminants in TAP-MudPIT data, we used an extended background list from Bloemendal et al. [43] (Table S6). For validation of PRO40-FLAG-MS data, we performed FLAG-AP experiments with wildtype protein extract as control. Since these experiments yielded a large number of proteins, a previously described quantification approach was employed to evaluate proteins that were identified with at least two different peptides in 2–3 PRO40-FLAG-AP experiments (Table S3), but also in wildtype control experiments (Table S4) [93], [94]. This procedure was necessary, because the PRO40 bait protein was identified in one of the control experiments. Therefore, spectral counts for each identified protein were first divided by the sum of spectral counts for each MS run (PRO40_1, 9559; PRO40_2, 5010; PRO40_3, 7999; wt_1, 5633; wt_2, 5901). Then, values for each experiment type were added and used to calculate a ratio between PRO40 and wildtype control data (Table S5). 79 proteins showing a ratio ≥2 were considered significant hits. From these proteins, PRO40 (ratio 18.63) and MEK1 (ratio 2.15) were verified as direct PRO40 binding proteins by yeast two-hybrid analysis, and RHO1 (ratio 2.23) was verified as indirect PRO40 binding protein via the interaction with PKC1, showing the applicability of this approach.
Immunodetection of TAP-tagged proteins was performed as described using a polyclonal anti-calmodulin binding peptide antibody (1∶2000, Merck Millipore, Billerica, MA, USA) and an anti-rabbit HRP-linked secondary antibody (Cell Signaling Technology, Danvers, MA, USA) [43]. FLAG-PRO40 was detected as described [24] using a monoclonal anti-FLAG antibody (1∶2000, Sigma Aldrich, St. Louis, MO, USA) and an anti-mouse HRP-linked secondary antibody (Cell Signaling Technology, Danvers, MA, USA).
For analysis of MAK1 phosphorylation status, strains were pre-cultured in liquid BMM for 2 days at 27°C. Three standardized inoculates were transferred into liquid HEPES (50 mM) -buffered BMM and cultivated for an additional three to six days at 27°C and 30 rpm. For induction of cell wall stress, cultures were subjected to 0.01% H2O2 for 15, 30, or 45 minutes. Mycelia were harvested by filtration, ground in liquid nitrogen, and resuspended in FLAG extraction buffer with phosphatase inhibitors (1% Phosphatase-Inhibitor-Cocktails II and III, Sigma Aldrich, St. Louis, MO, USA). After centrifugation at 15000 rpm for 30 min, equal amounts of total protein were subjected to SDS PAGE and Western Blotting according to standard protocols [76]. Phosphorylated MAK1 was detected using a polyclonal anti-phospho-p44/42 antibody (Cell Signaling Technology, Inc., USA) and an anti-rabbit HRP-linked secondary antibody (Cell Signaling Technology, Danvers, MA, USA) according to the manufacturer's protocol. Chemiluminescence was detected using a ChemidocXRS system (Biorad) and Clarity Western ECL substrate (Biorad). For an internal standard, an anti-α-tubulin antibody (Sigma Aldrich, St. Louis, MO, USA, T9026) was used in combination with an anti-mouse HRP-linked secondary antibody (Cell Signaling Technology, Danvers, MA, USA).
Microscopy was performed with an AxioImager microscope (Zeiss, Jena, Germany). For characterization of sexual development by DIC microscopy, strains were grown on BMM-coated glass slides for 2–7 days as described previously [24]. Hyphal fusion was investigated in 2 days old cultures grown on cellophane-covered MMS plates as described [43]. Localization of fluorescently labeled proteins in vegetative hyphae was investigated on BMM-covered glass slides as described previously [24]. For fluorescence microscopy of protoperithecia, strains were grown on cellophane-covered BMM plates for 3 days, and pieces of cellophane were fixed in 0.2% formaldehyde in PBS (58 mM Na2HPO4, 17 mM NaH2PO4, 68 mM NaCl; pH 7.4). Fluorescence was observed using filter sets (Chroma Technology) 41017 (HQ470/40, HQ525/50, Q495lp) or 49002 (ET470/40×, ET525/50m, T495lpxr) for EGFP and filter set 49008 (ET560/40×, ET630/75m, T585lp) for tdTomato.
Mutant pro30 from our laboratory collection was back-crossed several times to wildtype or brown-spored fus [25] and finally crossed to fus (Figure S8). DNA was extracted from 40 sterile progeny as described previously [25]. 40 fertile strains were collected from three crosses of mutants pro30, pro32, and pro34 to fus (Figure S8). Mutants pro32 and pro34 are described elsewhere [82; Teichert and Kück, unpublished]. 5 µg of pooled genomic DNA for pro30 and wt, respectively, was subjected to 50 bp paired-end Illumina/Solexa sequencing with a HiSeq2000 at GATC Biotech (Konstanz, Germany). Cleaning of raw data, mapping to the S. macrospora reference genome [5], [41], and analysis of sequence variants was performed as described [25] using the Burrows Wheeler Alignment tool [95], SAMtools [96] and custom-made Perl scripts, with minor modifications (Text S1). Genome sequencing data have been deposited at the sequence read archive (SRA; acc. no. SRX483430 and SRR1046323 for pro30 and wildtype (wt_3) [82], respectively).
|
10.1371/journal.pntd.0005820 | Knowledge, attitudes and practices towards yaws and yaws-like skin disease in Ghana | Yaws is endemic in Ghana. The World Health Organization (WHO) has launched a new global eradication campaign based on total community mass treatment with azithromycin. Achieving high coverage of mass treatment will be fundamental to the success of this new strategy; coverage is dependent, in part, on appropriate community mobilisation. An understanding of community knowledge, attitudes and practices related to yaws in Ghana and other endemic countries will be vital in designing effective community engagement strategies.
A verbally administered questionnaire was administered to residents in 3 districts in the Eastern region of Ghana where a randomised trial on the treatment of yaws was being conducted. The questionnaire combined both quantitative and qualitative questions covering perceptions of the cause and mechanisms of transmission of yaws-like lesions, the providers from which individuals would seek healthcare for yaws-like lesions, and what factors were important in reaching decisions on where to seek care. Chi-square tests and logistic regression were used to assess relationships between reported knowledge, attitudes and practices, and demographic variables. Thematic analysis of qualitative data was used to identify common themes.
A total of 1,162 individuals participated. The majority of individuals (n = 895, 77%) reported that “germs” were the cause of yaws lesions. Overall 13% (n = 161) of respondents believed that the disease was caused by supernatural forces. Participants frequently mentioned lack of personal hygiene, irregular and inefficient bathing, and washing with dirty water as fundamental to both the cause and the prevention of yaws. A majority of individuals reported that they would want to take an antibiotic to prevent the development of yaws if they were asymptomatic (n = 689, 61.2%), but a substantial minority reported they would not want to do so. A majority of individuals (n = 839, 72.7%) reported that if they had a yaws-like skin lesion they would seek care from a doctor or nurse. Both direct and indirect costs of treatment were reported as key factors affecting where participants reported they would seek care.
This is the first study that has explored community knowledge, attitudes and practices in relation to yaws in any endemic population. The belief that ‘germs’ are in some way related to disease through a variety of transmission routes including both contact and dirty water are similar to those reported for other skin diseases in Ghana. The prominent role of private healthcare providers is an important finding of this study and suggests engagement with this sector will be important in yaws eradication efforts. Strategies to address the substantial minority of individuals who reported they would not take treatment for yaws if they were currently asymptomatic will be needed to ensure the success of yaws eradication efforts. The data collected will be of value to the Ghana Health Service and also to WHO and other partners, who are currently developing community mobilisation tools to support yaws eradication efforts worldwide.
| Yaws, a bacterial skin infection, is endemic in Ghana. WHO has launched a campaign to eradicate yaws based on community mass treatment with the antibiotic azithromycin. Community perceptions of disease are an important contributor to participation in mass treatment interventions. This study used questionnaires to understand beliefs about yaws amongst individuals living in endemic communities in Ghana. Most individuals reported that ‘germs’ were the cause of yaws although the route of transmission was less well understood with many individuals reporting that dirty drinking or washing water was responsible for transmission. Participants reported they would normally seek care within the formal healthcare sector although many individuals reported they would visit traditional healthcare providers or pharmacies. Cost of care was the key factor for many participants. A majority of individuals reported they would be happy to take an antibiotic to prevent infection but a large minority (38.8%) reported that they would not. This study provides the first data on community beliefs about yaws in Ghana and will be valuable in helping the Ghana Health Service and partners develop community mobilisation tools to support yaws eradication efforts worldwide.
| Yaws, caused by Treponema pallidum subsp. pertenue, is endemic in Ghana. The disease is reported in all districts but predominantly in the south of the country. The majority of clinical disease is seen in young children, with the organism being transmitted by skin to skin contact with infectious lesions. In 2012, the World Health Organization (WHO) launched a new global eradication campaign based on total community mass treatment (analogous to mass drug administration—MDA) with azithromycin [1]. This involves the treatment of all individuals in an endemic community regardless of the presence or absence of clinical yaws. Achieving high coverage of mass treatment will be fundamental to the success of this new strategy [2]; coverage is dependent, in part, on appropriate community mobilisation. The success of interventions may be jeopardized by non-participation within endemic communities [3], because it may leave reservoirs of infection from which disease may re-emerge.
Studies in Ghana have established that amongst residents of rural communities there is a strong conceptual connection between lack of cleanliness and disease causation [4–6]. In a study conducted within Bono (Akan) society, sickness and disease may be attributed to many causes, including water (nsuo), dirt (efi,) lack of personal hygiene, cleanliness, the environment in which one sleeps, and bad air or wind (mframa) [6]. Another study reported that beliefs about dirty bodies (shoefi) are central to the understanding of disease for the Akan, contributing to the importance attached to daily bathing [5]. Many traditional belief systems incorporate a strong correlation between microorganisms (mmoa) and dirt (efi), in which efi is “synonymous with contamination and disease, and implies the necessity of cleansing and purification” [6]. Community perceptions of disease causation may play an important role in access to or utilization of health services [7,8].
There are no publications on community knowledge, attitudes and practices (KAP) related to yaws in Ghana or other endemic countries. However, an understanding of these issues will be vital in designing effective community engagement strategies to support the yaws eradication effort. This study was conducted alongside a randomised control trial investigating the optimal dose of azithromycin for treatment of yaws in Ghana. The study was conducted in three districts (total population 319,898) which had reported 1,505 cases of yaws in the preceding 36 months, and aimed to collect both quantitative and qualitative data on local knowledge, attitudes and practices concerning the causation and treatment of yaws.
A verbally administered questionnaire was developed, based on existing information about yaws and previous studies on community beliefs around skin disease in West Africa [9,10]. Prior to commencing this study, the questionnaire was pre-tested with a small number of participants from a group of communities not included in recruitment of the definitive sample of respondents. The questionnaire was refined based on this experience to ensure clarity of questioning.
This study was conducted in 3 districts in the Eastern region of Ghana namely Ayensuanor, West Akyem and Upper West Akyem in 2015. These contiguous districts were selected out of 4 districts where a randomised trial on different doses of azithromycin for the treatment of yaws was being conducted (ClinicalTrials.gov identifier NCT02344628). The fourth district was 400 km away and could not be included in the KAP study because of distance and costs reasons. None of these districts had previously received azithromycin mass treatment for yaws or trachoma. Within the chosen districts, multi-stage sampling was used to select participants for the adjunctive investigations described here. The first level of sampling was communities in the study districts which were selected using simple random sampling. In each selected community, an estimate of the total adult population was obtained from the district directorate of health services. In selected communities, systematic sampling was used to select participating households by randomly selecting an initial house and then utilising a quasi-random clockwise walking method [11,12]. In each house, either the oldest male or female resident was interviewed. If no adults were present at the time of visit, the house was excluded. In order to obtain a gender-balanced survey, the selection of respondent was alternated by gender between houses where possible.
Questionnaires were administered by local staff from each district. Participants were shown photos of typical yaws lesions as examples of the disease being discussed. Data were collected on demographics, including gender, ethnicity, religion, age, highest level of education achieved and occupation. The questionnaire combined both quantitative and qualitative questions covering several themes, including perceptions of the cause and mechanisms of transmission of yaws-like lesions, the providers from which individuals would seek healthcare if they themselves had a yaws-like lesions, and what factors were important in reaching decisions on where to seek care. Prior to conducting fieldwork, all staff received training on qualitative techniques and on administering the questionnaire, with training provided by the lead member of the KAP study team (MMA). Questionnaires were administered in local languages, and recorded on a standard form. Data were entered into a study database in EpiInfo by one of two members of the study team (either RD or BO).
Quantitative data were analysed with the use of descriptive statistics. Categorical variables were summarized using absolute numbers and percentages. Chi-square tests and logistic regression were used to assess relationships between reported knowledge, attitudes and practices, and demographic variables, including age, ethnicity, religion and level of education. These analyses were performed using STATA 13.1 (Statacorp). For the purposes of qualitative analysis, we considered knowledge to be participants’ reported beliefs around the causation and transmission of yaws, attitudes to reflect what individuals believed could or should be done to prevent disease, and practices to encompass the healthcare seeking behaviour of individuals with yaws and the underlying reasons for presenting. Qualitative data were assessed across each of these domains. Data were manually codified to identify common themes. Representative examples of major themes were identified in each domain.
The studies were approved by the World Health Organization (RPC 720), London School of Hygiene & Tropical Medicine (LSHTM 8832) and Ghana Health Service (GHS 13/11/14) ethics committees. Written informed consent was obtained from all participants.
A total of 1,162 individuals participated. Slightly more women (n = 600, 51.6%) were included than men, and the median age of participants was 36 years (IQR 27–46 years). The majority of participants were Christians (n = 966, 83.1%) and of Akan ethnicity (n = 593, 51.0%); other reported religious affiliations and ethnicities are shown in Table 1. A minority of participants reported a personal history of a lesion consistent with yaws (n = 214, 18.4%).
The majority of individuals (n = 895, 77%) reported that “germs” were the cause of yaws lesions. Contact with an individual with a similar lesion was reported to be an important cause by 532 participants (45.8%). Many individuals (n = 816, 70.2%) incorrectly believed that washing in or drinking dirty water were possible routes of transmission (Table 2). There was substantial overlap in beliefs with the majority of those who believed contact with an infected individual was a route of transmission also believing that washing or drinking water played a role (n = 457, 85.9%). In open-ended questions many respondents (38.1%) also suggested transmission was more broadly related to hygiene, contact with dirty water or a lack of washing. 13%(n = 161) of respondents believed that the disease was caused by supernatural forces such as witchcraft, curses or a punishment from god. Belief in a supernatural cause was not associated with gender, religion or ethnic group (p > 0.05 for all comparisons).
Participants frequently mentioned lack of personal hygiene, irregular and inefficient bathing, and washing with dirty water as fundamental to both the cause and the prevention of yaws.
More broadly, many participants asserted that bathing and personal hygiene were closely associated to wider notions of longevity of life and health, stating that bathing was fundamental to healthy living.
Contact with individuals and/or sharing of sponges or other items used in bathing were also reported as playing an important role in disease transmission, while some individuals reported a belief that transmission could be caused by airborne spread of germs.
The over-riding theme emerging from the interviews was that dirt, dirtiness and germs were inextricably connected to notions of cleanliness in understanding the mechanism of transmission, and that personal hygiene was believed to be the key intervention to protect individuals from germs and dirt—and therefore from infection.
A majority of individuals reported that they would want to take an antibiotic as part of a mass treatment campaign to treat yaws, even if they themselves were asymptomatic (n = 689, 61.2%), but a substantial minority reported they would not want to do so (31.8%). After adjustment for confounders, neither religion nor level of education were associated with individuals reporting that they would accept treatment for yaws (p = 0.06 and p = 0.09 respectively). After controlling for other factors individuals of Ga-Adangbe were slightly more likely to report willingness to accept treatment (OR 1.6 95% CI 1.05–2.45, p = 0.01). After controlling for demographic factors, belief in supernatural causation of yaws-like lesions was associated with a non-significant lower likelihood of accepting medication (OR 0.53–95% CI 0.07–3.99, p = 0.541).
A majority of individuals (n = 839, 72.7%) reported that if they had a yaws-like skin lesion they would seek care from a doctor or nurse. A substantial minority of individuals reported that they would seek care through the private sector, most commonly a pharmacy (n = 426, 37%) or a private shop (n = 28, 2.4%) (Table 2). Many individuals also reported that they would seek care from a traditional healer or witchdoctor (n = 496, 42.7%). Belief in a supernatural cause for yaws was strongly associated with the individual reporting that they would seek care from a traditional healer or witchdoctor (OR 2.93–95%CI 2.1–4.2, p <0.0001). Many individuals (n = 276, 36.8%) reported that they would seek care both from the formal healthcare sector (doctor, nurse, hospital) and a traditional healer or witchdoctor.
Cost was the factor most commonly reported by respondents as being important in determining where they sought care (n = 1,080 92.9%), with both the direct cost of treatment (n = 1,029, n = 88.6%) and the cost of travel (n = 646, 55.6%) reported as considerations. The opinions of family members or an individual’s previous experience were relatively infrequent considerations (n = 118, 10.2%) in guiding treatment seeking behaviours (Table 2). The qualitative data confirmed that cost was a major driver in decision making about where to seek care:
Participants reported receiving a wide range of different treatments for yaws-like lesions. The most commonly reported treatments were tablets (n = 505, 43.6%) or injectable antibiotics (46.2%) but many people also reported receiving topical treatments (Table 2).
We have explored community knowledge, attitudes and practices with regards to yaws in southern Ghana, a region of the country where yaws is endemic. Whilst most participants reported that infection was in some way related to “germs”, they reported a variety of perceived routes of transmission, including both correct (contact with infected individuals) and incorrect (e.g, contact with dirty water) routes of transmission. These findings are in keeping with studies undertaken in Ghana concerning knowledge, attitudes and practices in relation to Buruli ulcer, which have emphasised the perceived role of contact with infected individuals and drinking dirty water as important routes of transmission for that disease [9,10]. In the study reported here, almost 15% of participants believed that witchcraft or curses could be responsible for skin ulcers. Attributing an illness to supernatural forces such as witchcraft and curses is a common phenomenon and is often due to a lack of knowledge of the true aetiology of illness, particularly when the illness is prolonged [9]. Many individuals in Buruli ulcer endemic areas of Ghana report similar beliefs about causation of disease [9,10].
Whilst the majority of individuals reported that they would seek care from a doctor or a nurse for the management of yaws-like lesions, the prominent role of private healthcare providers is an important finding of this study. In Ghana, most individuals are eligible for free health care via a national health insurance programme and most individuals have access to a local primary health care facility. However indirect costs and time to travel to state providers may still represent a barrier to accessing ‘free’ health-care and may be a reason for seeking alternative care provider if that is more readily available locally [13]. As in other parts of sub-Saharan Africa private pharmacies are also common. There individuals may purchase a wide-range of medications directly without the need for medical consultation. More than a third of individuals reported that they would seek care from private providers: specifically staff at private pharmacies or traditional healers. Acceptance of a given intervention can depend on a variety of factors, including understanding of causation, perceived effectiveness of the intervention and both direct and indirect cost factors [7]. Our findings that cost and distance represent barriers to accessing healthcare are in keeping with other studies from Ghana [14,15]. Whilst community-based MDA programmes should address cost issues, health education will be needed to address beliefs around causation and the effectiveness of azithromycin against yaws. Yaws eradication programmes should also consider how best to engage with these other providers, who operate outside of the formal healthcare system, and who may be an essential means for detecting cases of yaws in endemic communities as the eradication endpoint is approached. They should be considered both in Ghana and in other settings where such informal providers play important roles alongside formal healthcare systems. As the majority of yaws cases occur amongst children, consideration could also be given to training school teachers to identify signs and symptoms of yaws, and engaging them in the yaws eradication effort and further studies exploring this possibility should be undertaken.
A sizeable minority of individuals in our study reported that they would not be willing to take medicine for yaws if they were asymptomatic at the time that it was offered to them. This is potentially worrisome, as estimates suggest that for each symptomatic case of yaws, there are 5–10 individuals with asymptomatic infection, and that latent yaws can reactivate up to a decade after primary disease has resolved in an individual; this, in fact, is part of the basis for the use of an initial strategy of total community treatment, rather than one involving identification and treatment of those with disease [16]. If our unwillingness rate was translated into non-participation rates in yaws eradication campaigns, it would represent a significant challenge for interrupting transmission in Ghana and likely increase the number of rounds of treatment required to reach that goal [2].
Studies of the acceptability of azithromycin MDA for trachoma have reported associations with gender and beliefs around disease causation [7]. In our respondents, we did not identify any clear demographic factors associated with the likelihood of treatment refusal when asymptomatic. The majority of cases of yaws occur in children (who were not included in this survey) but in the context of MDA for trachoma, the decision of the household head is often important in determining whether any members of a household receive azithromycin [3]. Interventions that increase the acceptability of treatment to household adults might also result in improved uptake of treatment amongst children.
Individuals who reported a supernatural belief about causation, such as witch-craft or a curse, were more likely to report that they would not accept treatment, but this association did not reach statistical significance. More in-depth qualitative work should be considered to explore this issue in greater detail.
This study has a number of limitations. First, we did not conduct in-depth interviews alongside the questionnaires. This may have limited our ability to explore issues arising from the questionnaire in more detail, or to fully explore the nuances of individual or community beliefs around yaws. Second, the clinical lesions of yaws are similar to many other skin diseases and we cannot be certain of the extent to which the beliefs expressed by community members are specific to yaws rather than applying more broadly to locally endemic skin diseases. Clinically important differentiations between diseases such as scabies, Buruli ulcer, yaws and leprosy may not be easily discernible to community members, although Buruli ulcer is not known to be endemic in any of the districtsin which this study was conducted. Although MDA with azithromycin may have ancillary benefits in treating some other skin diseases [17,18] it will not affect many common skin infections, such as Buruli ulcer, scabies or tinea. Health education programmes will need to clearly emphasise that whilst azithromycin will have a significant impact on yaws and some other diseases, it is not a panacea for all skin complaints. Finally we cannot be certain that individuals would always seek care for their children given how common skin disease is in many communities, nor that they would seek care for their children in the same place that they stated they would seek care for themselves. Data from other studies suggests that the attitude of the adult or head of household dictates engagement with healthcare services for all household members [3] so we believe that the viewpoints of adults would be key in determining healthcare entry points for children.
Despite these limitations, this is the first study of which we are aware that has explored community knowledge, attitudes and practices in relation to yaws in any endemic population. We were able to collect quantitative and qualitative data from a broad cross-section of individuals across a range of ages, genders, and ethnic and religious groups in three districts in Ghana. Our data highlight the need to address barriers to accessing care including both direct and indirect healthcare costs. The data collected will be of value to the Ghana Health Service and also to WHO and other partners, who are currently developing community mobilisation tools to support yaws eradication efforts worldwide.
|
10.1371/journal.pgen.1002159 | Transportin-SR Is Required for Proper Splicing of Resistance Genes and Plant Immunity | Transportin-SR (TRN-SR) is a member of the importin-β super-family that functions as the nuclear import receptor for serine-arginine rich (SR) proteins, which play diverse roles in RNA metabolism. Here we report the identification and cloning of mos14 (modifier of snc1-1, 14), a mutation that suppresses the immune responses conditioned by the auto-activated Resistance (R) protein snc1 (suppressor of npr1-1, constitutive 1). MOS14 encodes a nuclear protein with high similarity to previously characterized TRN-SR proteins in animals. Yeast two-hybrid assays showed that MOS14 interacts with AtRAN1 via its N-terminus and SR proteins via its C-terminus. In mos14-1, localization of several SR proteins to the nucleus was impaired, confirming that MOS14 functions as a TRN-SR. The mos14-1 mutation results in altered splicing patterns of SNC1 and another R gene RPS4 and compromised resistance mediated by snc1 and RPS4, suggesting that nuclear import of SR proteins by MOS14 is required for proper splicing of these two R genes and is important for their functions in plant immunity.
| Plant immune receptors encoded by Resistance (R) genes play essential roles in defense against pathogens. Multiple R genes are alternatively spliced. How plants regulate the splicing of these R genes is unclear. In this study, we identified MOS14 as an important regulator of two R genes, SNC1 and RPS4. Further analysis showed that MOS14 functions as the nuclear import receptor for serine-arginine rich (SR) proteins, which play diverse roles in RNA metabolism. Loss of the function of MOS14 results in altered splicing patterns of SNC1 and RPS4 and compromised resistance mediated by snc1 and RPS4, suggesting that nuclear import of SR proteins by MOS14 is required for proper splicing of these two R genes and is important for their functions in plant immunity.
| In eukaryotes, the nuclear envelope forms a barrier between the cytoplasm and the nucleus. Trafficking of macromolecules across the nuclear envelope occurs through the nuclear pore complex (NPC) [1]. Previous studies on MOS3 [2], MOS6 [3], MOS7 [4] and MOS11 [5] have revealed the importance of nucleocytoplasmic trafficking in plant immunity. Mutations in MOS3, MOS6, MOS7 and MOS11 suppress the constitutive defense responses of snc1 (suppressor of npr1-1, constitutive 1), a gain-of-function mutant carrying a mutation in a Toll/interleukin-1 receptor-Nucleotide Binding-Leucine Rich Repeat (TIR-NB-LRR) R protein [6]. MOS3 encodes the nucleoporin Nup96 [2], whereas MOS11 encodes a putative RNA binding protein [5]. Both MOS3 and MOS11 are required for mRNA export. MOS6 encodes a putative importin-α [3], whereas MOS7 encodes another nucleoporin, Nup88, which is required for nuclear accumulation of snc1 and two general defense regulators, Enhanced Disease Susceptibility 1 (EDS1) and Nonexpresser of PR genes 1 (NPR1) [4].
Nuclear import receptors play essential roles in transferring proteins from the cytoplasm to the nucleus. The largest group of nuclear import receptors belong to the importin-β super-family. Members of the importin-β super-family have rather low overall sequence similarity but they all have a conserved N-terminal RAN-binding domain [7], [8]. The import receptors recognize the nuclear localization sequence (NLS) of target proteins to facilitate their transport through the NPC. Upon RAN-GTP binding to importin-β, the importin-β complex is dissociated and the cargo is released into the nucleus.
The importin-β super-family can be divided into several sub-families according to the direction and the cargo type they transport [9]. Among them, the transportin-SR (TRN-SR) subfamily functions as nuclear import receptors for serine-arginine rich (SR) proteins. TRN-SR was originally identified as an interactor of SR domains of ASF/SF2 [10] and papillomavirus E2 [11]. In humans, the C-terminus of TRN-SR interacts with SR proteins and the interaction can be disrupted upon RAN-binding to its N-terminus [10].
SR proteins are a highly conserved family of nuclear proteins that play important roles in splicing [12]–[14]. They contain RNA recognition motifs (RRM) at the N-terminus and an arginine-serine rich (RS) domain at the C-terminus. The NLS is located in the RS domain. SR proteins not only function as splicing factors for constitutive splicing [15], [16], they also regulate alternative splicing through splice site selection in a concentration-dependent manner [17], [18].
Several plant R genes including the tobacco N gene [19], the barley Mla6 [20], Arabidopsis SNC1 [21] and RPS4 [22]–[24] are alternatively spliced. For example, six transcript variants (TV) have been identified for RPS4 [23], [24]. Compromised RPS4-mediated resistance resulting from a lack of the TVs suggests that alternative splicing of RPS4 is required for its function [23]. However, it is unclear how alternative splicing of these R genes is controlled and why it is necessary for immunity. In this study, we report that Arabidopsis MOS14 encodes a TRN-SR that is required for proper splicing of SNC1 and RPS4, suggesting that SR proteins may play important roles in the control of the splicing of these two R genes.
Arabidopsis snc1 constitutively activates defense responses and displays enhanced resistance to pathogens. snc1 mutant plants exhibit dwarf morphology with curly leaves. Suppressor screens of snc1 have previously been carried out using fast neutron and T-DNA insertional mutagenesis [2], [25]. To identify additional suppressor mutants of snc1, we treated snc1 npr1 seeds with ethane methyl sulfonate (EMS) and screened the M2 plants for mutants that suppressed snc1 dwarfism. From this population, we identified mos14-1 snc1 npr1 (Figure 1A).
In snc1 npr1, defense marker gene PR1 and PR2 are constitutively expressed. As shown in Figure 1B and 1C, constitutive activation of PR1 and PR2 is suppressed in mos14-1 snc1 npr1. Analysis of SA levels also showed that the elevated SA levels in snc1 npr1 are suppressed by mos14-1 (Figure 1D and 1E). To test whether enhanced pathogen resistance in snc1 npr1 is affected by mos14-1, mos14-1 snc1 npr1 seedlings were challenged with the virulent oomycete pathogen Hyaloperonospora arabidopsidis (H.a.) Noco2. As shown in Figure 1F, resistance to H. a. Noco2 is lost in mos14-1 snc1 npr1.
To map the mos14-1 mutation, we crossed mos14-1 snc1 npr1 (in the Columbiaecotype background) with Landsberg erecta (Ler)-snc1 [2]. In the F2 mapping population, about a quarter of the progeny showed morphology similar to the triple mutant. Crude mapping using 24 F2 plants revealed that mos14-1 is linked to the lower arm of chromosome 5 (Figure 2A). Further analysis indicated that mos14-1 is flanked by marker MMN10 and MUB3. Fine mapping using about 1200 F2 plants narrowed mos14-1 to a 60 kb region between marker K19B1 and MRG21. To identify the mos14-1 mutation, PCR fragments covering this 60 kb region was amplified directly from mos14-1 snc1 npr1 and sequenced. A single G to A mutation was found in At5g62600 (Figure 2B), which is located at the junction of the 13th intron and 13th exon of the gene. RT-PCR analysis using primers flanking the mutation showed that splicing of At5g62600 was affected by the mutation (Figure 2C). The RT-PCR fragments were cloned into the pGEM-T vector. Subsequent sequence analysis of cDNA clones from mos14-1 revealed that they fell into six different classes. All of them represent transcript variants that were incorrectly spliced. An alignment of wild type cDNA and the cDNA variants from mos14-1 are shown in Figure S1.
To confirm that the mutation in At5g62600 is responsible for the suppression of snc1 npr1 mutant phenotypes, a genomic clone containing At5g62600 was constructed and transformed into mos14-1 snc1 npr1. Transgenic plants from five independent lines carrying the wild type At5g62600 displayed snc1-like morphology (Figure 2D). Further analysis of a representative transgenic line showed that the expression of PR1 and PR2 was similar to snc1 npr1 (Figure 2E and 2F). In addition, resistance to H. a. Noco2 was also restored in the transgenic line (Figure 2G), confirming that At5g62600 complemented mos14-1 and MOS14 is At5g62600.
To obtain the mos14-1 single mutant, we backcrossed mos14-1 snc1 npr1 with wild type plants. The mos14-1 single mutant was obtained by genotyping the F2 plants. The mos14-1 single mutant flowers late and has reduced fertility. Besides, it exhibits small stature (Figure S2). When the genomic clone of At5g62600 was introduced into the mos14-1 single mutants, it reverted the size and fertility of the mutant to wild type-like and also suppressed the late flowering phenotype, showing that the developmental phenotypes observed in mos14-1 are caused by the mos14-1 mutation.
MOS14 is a single copy gene in Arabidopsis. It encodes a protein with 25% identity and 45% similarity to the TRN-SR in Drosophila, suggesting that MOS14 may be a transporter for SR proteins. MOS14 and its animal homologs are highly conserved at their N-terminus (Figure S3), which contain the importin-β N-terminal domains.
To determine the subcellular localization of MOS14, transgenic plants expressing MOS14 under its native promoter with a C-terminal GFP tag were generated in both wild type and mos14-1 backgrounds. Expression of MOS14-GFP in mos14-1 suppresses the developmental phenotypes of mos14-1 (Figure S4), suggesting that the fusion protein is functional. Confocal fluorescence microscopy analysis of transgenic plants expressing the MOS14-GFP fusion protein showed that the GFP signal is found exclusively in the nucleus (Figure 3), indicating that MOS14 is a nuclear protein. In the nuclei of root cells, GFP fluorescence was excluded from a large part of the nucleus, probably the nucleolus. We did not observe similar exclusion of MOS14-GFP from parts of the nuclei in epidermal cells, probably because these nuclei are much smaller than those in root cells.
In animals, TRN-SR binds SR proteins via its C-terminus and transport SR proteins through the nuclear envelope. Binding of RAN-GTP to the N-terminus of TRN-SR in nucleus results in the release of SR proteins. To test whether MOS14 is able to interact with SR proteins, the N-terminus (1–281) and C-terminus (282–958) of MOS14 were expressed in the bait vector and four selected Arabidopsis SR proteins (AtRS2Z33, AtRSZ21, AtRS31 and AtSR34) were expressed in the prey vector for yeast two-hybrid assays. As shown in Figure 4A, the C-terminus, but not the N-terminus of MOS14 interacts with the SR proteins. We also tested the interactions between MOS14 and AtRAN1. As shown in Figure 4B, the N-terminus, but not the C-terminus of MOS14 interacts with AtRAN1.
To test whether the mos14-1 mutation affects the nuclear import of Arabidopsis SR proteins, we made constructs expressing four SR genes AtRS2Z33, AtRSZ21, AtRS31 and AtSR34 with a C-terminal GFP tag. These constructs were transformed into protoplasts of wild type and mos14-1 plants to check for the localization of the SR-GFP proteins. A construct expressing the SARD1-GFP fusion protein was included as the control [26]. As shown in Figure 5A, in both wild type and mos14-1 protoplasts, SARD1 was localized in the nucleus. Consistent with previous studies [27], the SR-GFP proteins were clearly localized in the nucleus of wild type protoplasts. However, in mos14-1 protoplasts, the SR-GFP proteins were mainly localized in the cytoplasm (Figure 5A and Table 1), suggesting that MOS14 is required for the nuclear localization of SR proteins.
Unlike GFP expressed under 35S promoter which is distributed throughout the whole cell, the SR-GFP proteins were localized to discrete foci in the cytoplasm of mos14-1 protoplasts. The pattern of these foci resembles that of P-bodies, which are distinct foci in the cytoplasm of eukaryotic cells containing many enzymes involved in mRNA turnover. Because of the diverse roles of SR proteins in RNA metabolism, it would not be surprising if they also function in P-bodies. The effect of mos14-1 on the localization of AtRSZ21 and AtSR34 was further confirmed in transgenic plants expressing the AtRSZ21-GFP and AtSR34-GFP fusion proteins. As shown in Figure 5B and 5C, AtRSZ21-GFP and AtSR34-GFP were localized in discrete foci in the cytoplasm of guard cells in mos14-1 background. The GFP fusion proteins were also observed in the cytoplasm of leaf pavement cells in mos14-1. Taken together, these experiments indicate that MOS14 is a transporter for SR proteins.
Multiple SNC1 transcripts with intron 2 and intron 3 removed or retained have previously been detected [21]. Because none of the transgenic plants expressing the snc1 cDNA exhibit dwarf morphology like snc1 mutant plants (Figure S5), alternative splicing is probably required for the function of SNC1. Since mos14-1 affects the nuclear localization of SR proteins and SR proteins participate in pre-mRNA splice site recognition and spliceosome assembly, we tested whether splicing of SNC1 was affected in mos14-1. Primers flanking the introns of SNC1 were designed to evaluate its splicing pattern of SNC1 (Figure 6A). Consistent with the previous report, we detected transcripts with either intron 2 or 3 retained (Figure S6). As shown in Figure 6B, we detected another transcript that contains both intron 2 and 3 (TV1) in addition to the regular transcripts with both intron 2 and 3 removed (TV4) in mos14-1 snc1 npr1. In wild type plants, the amount of TV2 and TV3 is small compared to that of TV4. Both TV2 and TV3 increased dramatically in the mos14-1 snc1 npr1 mutant plants (Figure 6B). Similar alteration of SNC1 transcript patterns was also observed in the mos14-1 single mutant (Figure S8B). Since PCR reaction using the RNA samples showed no amplification, the DNA fragments from RT-PCR represent SNC1 transcripts rather than genomic DNA contamination. Further analysis of SNC1 transcript variants in mos14-1 and mos14-1 snc1 npr1 lines carrying the wild type MOS14 transgene showed that the splicing patterns of SNC1 in the transgenic lines are similar to those in the wild type plants (Figure S8A and S8B). These data indicate that mos14-1 affects the splicing of the SNC1 transcript.
The R gene RPS4 was also reported to be alternatively spliced [23]. We designed primers to detect the transcript variants for RPS4 by RT-PCR. As shown in Figure 6C, the levels of TV1 are similar in wild type and mos14-1. However, TV2+TV3 increased considerably and TV4 was significantly reduced in mos14-1, indicating that mos14-1 also affects the splicing pattern of RPS4 transcripts. The altered RPS4 transcript patterns in mos14-1 snc1 npr1 and mos14-1 can be complemented by the MOS14 transgene (Figure S8C and S8D).
To determine whether MOS14 has a general role in RNA splicing, we analyzed splicing of two housekeeping genes Actin1 and β-tubulin4 by RT-PCR using primers that flank introns. ROC1 was used as the control because it contains no intron. We found that splicing of Actin1 and β-tubulin4 was not affected in mos14-1 (Figure S7). We also analyzed the splicing patterns of U1-70K, AtSR30 and AtSR34, three genes reported to be alternatively spliced [28], [29]. As shown in Figure S7, the splicing of AtSR30 and AtSR34, but not U1-70K was clearly affected by mos14-1. Alteration of the transcription patterns of AtSR30 and AtSR34 in mos14-1 further supports the role of MOS14 in alternative splicing. Since the splicing of Actin1, β-tubulin4 and U1-70K is not affected by mos14-1, there may be a certain level of specificity in MOS14-mediated pre-mRNA processing.
To test whether the splicing defect in mos14-1 leads to a decrease in snc1 and RPS4 transcripts, real-time RT-PCR was carried out using primers to amplify an unspliced region at the 3′ end of the two genes. As shown in Figure 6D and 6E, expression levels of both snc1 and RPS4 decreased significantly in the presence of the mos14-1 mutation.
Since mos14-1 altered the splicing pattern of RPS4 and reduced its expression, we tested whether RPS4-mediated immunity is affected by mos14-1. As shown in Figure 7A, growth of Pseudomonas syringae pv. tomato (P.s.t.) DC3000 avrRps4 in mos14-1 is about ten-fold higher than that in wild type, suggesting RPS4-mediated immunity is compromised in mos14-1. We also tested whether MOS14 is required for basal resistance by challenging the mos14-1 plants with the virulent pathogen P.s.t. DC3000. As shown in Figure 7B, bacterial growth is about ten-fold higher in mos14-1 compared to wild type, indicating that MOS14 is also required for basal resistance.
Previous studies on snc1 suppressor mutants revealed that multiple components are involved in the regulation of plant immunity. In particular, pathways involved in mRNA export, protein import and protein export were found to contribute to immune responses. Here we report the identification of MOS14 as a novel component of nucleocytoplasmic trafficking required for plant immunity. Loss of MOS14 function suppresses the constitutive defense responses of snc1, compromises resistance mediated by RPS4 and impairs basal resistance against P.s.t. DC3000. These findings show that MOS14 plays a critical role in plant immunity.
MOS14 encodes a nuclear protein with high sequence similarity to TRN-SR proteins in animals. TRN-SR proteins have been shown to function as nuclear import receptors for both phosphorylated SR proteins as well as the splicing repressor protein RSF1 which antagonizes SR proteins in the nucleus [11], [30]. Since their discovery, TRN-SR proteins have not been extensively studied [10]. MOS14 is a single-copy gene, while the Arabidopsis genome has 18 genes belonging to six subfamilies of SR proteins, of which three are plant-specific [31]. There is no close homolog of RSF1 in Arabidopsis. Like the TRN-SR proteins in animals, the N-terminus of MOS14 interacts with AtRAN1 and the C-terminus interacts with SR proteins. In addition, localization of several SR proteins to the nucleus was impaired by mos14-1. These data indicate that the mechanism of nuclear import of SR proteins is conserved between plants and animals.
Very limited studies have been performed on the genetic characterization of TRN-SR proteins. In C.elegans, RNAi of the MOS14 homolog Transporter of SR-1 (TSR-1) leads to embryonic lethality, suggesting TRN-SR proteins can be essential for viability [32]. Intriguingly, the mos14-1 mutation is not lethal, although it does cause multiple development phenotypes such as reduced stature and fertility. In addition to its functions in development, our genetic analysis of MOS14 revealed that it plays important roles in both R gene-mediated resistance as well as basal defense, suggesting that nuclear import of SR proteins is important for plant immunity. The reasons why mos14-1 leads to these pleiotropic defects and not lethality awaits further investigation.
SR proteins play important roles in general RNA splicing, alternative splicing, as well as other processes of RNA metabolism. Consistent with the function of MOS14 in the nuclear import of SR proteins, the mos14-1 mutation affects the splicing of SNC1 and RPS4. Several R genes including SNC1, RPS4 and tobacco N gene are alternatively spliced, and alternative splicing of RPS4 and N gene are required for their function [23], [33]. In mos14-1, alternative splicing of both SNC1 and RPS4 are altered. This effect probably contributes to the suppression of snc1 mutant phenotypes by mos14-1 and compromised RPS4 function in the mos14-1 single mutant. In addition to the altered ratio of transcript variants, the expression levels of snc1 and RPS4 were also reduced. The reduced expression of snc1 and RPS4 is probably caused by splicing defects resulting from the reduced nuclear localization of SR proteins.
In mos14-1 snc1 npr1, the SNC1 TV-4 transcript level is only modestly reduced, suggesting that reduced accumulation of TV-4 may not be the only factor that contributes to the complete suppression of snc1 mutant phenotype. In addition to reduced accumulation of TV-4, levels of SNC1 TV-1, TV-2 and TV-3 are considerably increased in mos14-1 snc1 npr1. These transcripts are predicted to produce truncated snc1 proteins because of introduction of early stop codons. It is possible that these truncated proteins may interfere with the function of the full-length snc1. Because snc1 and RPS4 are not the only genes whose splicing are affected by mos14-1, altered splicing of one or more unknown positive regulators of plant defense could also contribute to the suppression of snc1 mutant phenotypes.
In addition to the compromised resistance responses mediated by snc1 and RPS4, basal resistance against P.s.t. DC3000 is also compromised in mos14-1. It remains to be determined how mos14-1 affects basal resistance. One possibility is that MOS14 is required for the splicing of one or more R genes that contribute to basal resistance against P.s.t. DC3000. Alternatively, mos14-1 may cause splicing defects in defense regulators required for basal resistance.
In summary, we have identified MOS14 as a nuclear transporter of SR proteins. Our data suggest that regulation of R gene splicing by SR proteins is critical for plant immunity. Future studies on individual SR proteins will help us better understand how SR proteins regulate the splicing of R genes.
All plants were grown at 23°C under 16 hr light/8 hr dark in plant growth rooms or chambers, if not specifically mentioned. To identify mutations that suppress the mutant phenotypes of snc1, snc1 npr1 seeds were mutagenized with EMS. About 30,000 M2 plants representing about 1,500 M1 families were screened for suppression of the dwarf morphology of snc1 npr1-1. Mutants lacking the dwarf phenotype were further analyzed for suppression of the constitutive defense responses in snc1 npr1.
About 0.1 g tissue was collected and RNA was extracted by Takara RNAiso reagent. The RNA was treated with Promega RQ1 RNase-Free DNase to remove contaminating genomic DNA. Reverse transcription was subsequently carried out using oligo-dT and the M-MLV RTase cDNA synthesis kit from Takara. About 200 ng of total RNA was included in each RT reaction. For semi-quantitative and real-time PCR , one fiftieth of the cDNA was used in each reaction. A total of 40 cycles were performed for semi-quantitative RCR except 28 cycles for ROC1. Real-time PCR was carried out using Takara SYBR® Premix Ex Taq™ II. The primers for real-time PCR analysis of PR1, PR2, SNC1 [34] and ROC1 (also called cyclophilin) [35] were described previously. ROC1 is a housekeeping gene without introns. The sequences of primers used for SNC1 and RPS4 transcript variants analysis are shown in Table S1. Primers to amplify U1-70K [28], AtSR30 and AtSR34 [29] were described previously.
For infections with H. a. Noco2, three-week-old soil-grown plants were sprayed with H. a. Noco2 at 5×104 spores/ml. The inoculated seedlings were subsequently kept in a growth chamber with high humidity (>80%) at 18°C under 12 hr light/12 hr dark cycle for seven days before growth of H. a. Noco2 was quantified, as previously described [36].
For infections with P.s.t. DC3000 or P.s.t. DC3000 avrRps4, five-week-old soil-grown plants were infiltrated with bacterial suspensions (OD600 = 0.001) in 10 mM MgCl2. Samples were taken at day 0 and day 3.
To analyze the SA levels in the mutant plants, SA was extracted using a previously described procedure [37] and measured by high-performance liquid chromatography.
For the transgenic complementation test, three PCR fragments, F12R37 (3.9K), F14R38 (3.8K) and F23R19 (2.9K) covering the 11 kb region where MOS14 is located were amplified from wild type genomic DNA. The primers used for amplification of F12R37, F14R38 and F23R19 are F12, R37, F14, R38, F23 and R19 respectively, and their sequences are provided in Table S1. These fragments were sequentially sub-cloned into pBluescript SK+. The complete 11 kb fragment was subsequently cloned into a modified pGreen0229 vector containing the NOS terminator to obtain the construct pMOS14:MOS14. The final construct containing MOS14 was transformed into mos14-1 snc1 npr1 through Agrobacterium-mediated transformation.
For the subcellular localization study of MOS14, PCR fragments F12R37 (3.9K), F14R38 (3.8K) and F23R20 (2.9K) were sequentially sub-cloned into pBluescript SK+. The primers used for amplification of F23R20 are F23 and R20 and their sequences are listed in the Table S1. The 11 kb fragment described above was cloned into a modified pCambia1305 vector expressing C-terminal tagged GFP to obtain pMOS14:MOS14-GFP.
For transient expression of SR proteins in protoplasts, full-length cDNAs of AtRS2Z33, AtRSZ21, AtRS31 and AtSR34 without the stop codons were amplified by PCR and cloned into the modified pUC19 vector pUC19-35S-cmGFP4 that expresses GFP under the 35S promoter.
To obtain transgenic plants expressing snc1 cDNA, full-length snc1 cDNA was amplified from total cDNA of snc1 and cloned into a modified pGreen0229 vector. The cDNA clone was sequenced to make sure the sequence is correct and no intron was retained.
To obtain transgenic plants expressing AtSR34-GFP and AtRSZ21-GFP, full-length cDNAs of AtSR34 and AtRSZ21 without the stop codons were amplified by PCR and cloned into a modified pCambia1300 vector expressing C-terminal tagged GFP under 35S promoter. The constructs were transformed into Col-0 and mos14-1 through Agrobacterium-mediated transformation.
To make constructs for the yeast two hybrid assays, an SfiI restriction site was first introduced to the multiple cloning site of pGBKT7 and pGADT7 to obtain pGBKT7a and pGADT7a, respectively. cDNA expressing the N-terminal or C-terminal region of MOS14 and AtRAN1 were amplified by PCR and cloned into pGBKT7a. Full-length cDNAs of AtRS2Z33, AtRSZ21, AtRS31 and AtSR34 were amplified by PCR and cloned into pGADT7a. cDNA expressing the N-terminal or C-terminal region of MOS14 were also cloned into pGADT7a. The plasmids expressing the MOS14 fragments were co-transformed with the vectors expressing AtRAN1 or one of the SR proteins into the yeast strain PJ694α for yeast two-hybrid analysis.
For confocal fluorescence microscopy analysis of MOS14-GFP, the roots or leaves of six-day-old seedling grown on MS plates were first stained with propidium iodide (PI) for 1 min and then washed in ddH2O for at least three times. The concentration of PI used for staining the roots was 10 µg/ml, whereas the concentration of PI used for the leaves is 10 mg/ml. The stained sample was observed using a Zeiss Meta 510 confocal microscope. Excitation wavelengths for GFP and PI were 488 nm and 543 nm, respectively. For root samples, the emission filter used for PI was LP560 nm. For leaf samples, the emission filter used for PI was BP560 nm-615 nm. For both root and leaf samples, the emission filter for GFP was BP505 nm-530 nm.
Plasmids used for protoplast transfections were purified with Invitrogen PureLink™ HiPure Plasmid Filter Purification Kit. Transformation of wild type or mos14-1 protoplasts was performed as previously described [38]. After transformation, protoplasts were kept in the dark for about 16 hours. The transformed protoplasts were examined using a Zeiss Axiovert 200 fluorescence microscope. The pictures of representative protoplasts were taken using confocal fluorescence microcopy. For autofluorescence, the emission filter used was 650 nm-740 nm. Confocal fluorescence microscopy analysis of transgenic plants expressing AtSR34-GFP and AtRSZ21-GFP was performed on three-week-old seedlings using a procedure described in the analysis of MOS14-GFP localization.
Sequence data from this article can be found in the Arabidopsis Genome Initiative or GenBank/EMBL databases under the following accession numbers: At5g62600 (MOS14), At2g14610 (PR1), At3g57260 (PR2), At4g38470 (ROC1), At2g37620 (Actin1), At5g44340 (β-tubulin4 ), AAD38537 (hTRN-SR1), CAB42634 (hTRN-SR2), NP608708 (dTRN-SR), AF025464 (TSR1) and CAA99366 (MTR10a).
|
10.1371/journal.pbio.0050052 | FMRP Mediates mGluR5-Dependent Translation of Amyloid Precursor Protein | Amyloid precursor protein (APP) facilitates synapse formation in the developing brain, while beta-amyloid (Aβ) accumulation, which is associated with Alzheimer disease, results in synaptic loss and impaired neurotransmission. Fragile X mental retardation protein (FMRP) is a cytoplasmic mRNA binding protein whose expression is lost in fragile X syndrome. Here we show that FMRP binds to the coding region of APP mRNA at a guanine-rich, G-quartet–like sequence. Stimulation of cortical synaptoneurosomes or primary neuronal cells with the metabotropic glutamate receptor agonist DHPG increased APP translation in wild-type but not fmr-1 knockout samples. APP mRNA coimmunoprecipitated with FMRP in resting synaptoneurosomes, but the interaction was lost shortly after DHPG treatment. Soluble Aβ40 or Aβ42 levels were significantly higher in multiple strains of fmr-1 knockout mice compared to wild-type controls. Our data indicate that postsynaptic FMRP binds to and regulates the translation of APP mRNA through metabotropic glutamate receptor activation and suggests a possible link between Alzheimer disease and fragile X syndrome.
| Alzheimer disease (AD) and fragile X syndrome (FXS) are devastating neurological disorders associated with synaptic dysfunction resulting in cognitive impairment and behavioral deficits. Despite these similar endpoints, the pathobiology of AD and FXS have not previously been linked. We have established that translation of amyloid precursor protein (APP), which is cleaved to generate neurotoxic βamyloid, is normally repressed by the fragile X mental retardation protein (FMRP) in the dendritic processes of neurons. Activation of a particular subtype of glutamate receptor (mGluR5) rapidly increases translation of APP in neurons by displacing FMRP from a guanidine-rich sequence in the coding region of APP mRNA. In the absence of FMRP, APP synthesis is constitutively increased and nonresponsive to mGluR-mediated signaling. Excess APP is proteolytically cleaved to generate significantly elevated βamyloid in multiple mutant mouse strains lacking FMRP compared to wild type. Our data support a growing consensus that FMRP binds to guanine-rich domains of some dendritic mRNAs, suppressing their translation and suggest that AD (neurodegenerative disorder) and FXS (neurodevelopmental disorder) may share a common molecular pathway leading to the overproduction of APP and its protein-cleaving derivatives.
| Alzheimer disease (AD) is a neurodegenerative disorder characterized by senile plaques and neurofibrillary tangles. The plaques are predominantly composed of beta-amyloid (Aβ), a 39–42 amino acid peptide cleaved from the amyloid precursor protein (APP). APP is likely important for synapse formation in the developing brain [1], while excess Aβ causes impaired synaptic function [2]. Disordered synaptic transmission is also a hallmark of other neuronal disorders, such as epilepsy and fragile X mental retardation syndrome (FXS).
FXS is the most prevalent form of inherited mental retardation, affecting one in 4,000 men and one in 8,000 women. This X chromosome–linked disorder is characterized by moderate to severe mental retardation (overall IQ <70), autistic-like behavior, seizures, facial abnormalities (large, prominent ears and long, narrow face) and macroorchidisim [3]. At the neuroanatomic level, FXS is distinguished by an overabundance of long, thin, tortuous dendritic spines with prominent heads and irregular dilations [4,5]. The increased length, density, and immature morphology of dendritic spines in FXS suggest an impairment of synaptic pruning and maturation.
In the majority of cases, FXS results from a trinucleotide (CGG) repeat expansion to >200 copies in the 5′-UTR of the fmr-1 gene (located at Xq27.3) [6]. The CGG expansion is associated with hypermethylation of the surrounding DNA, chromatin condensation, and subsequent transcriptional silencing of the fmr-1 gene, resulting in the loss of expression of fragile X mental retardation protein (FMRP) [7].
FMRP is an mRNA-binding protein that is ubiquitously expressed throughout the body, with significantly higher levels in young animals [8]. The protein has two heterogeneous nuclear ribonucleoprotein (hnRNP) K homology domains and one RGG box as well as nuclear localization and export signals. FMRP interacts with BC1 RNA as well as a number of RNA-binding proteins, including nucleolin and YB1 and the FMRP homologs FXR1 and FXR2 [9]. FMRP has been implicated in translational repression [10–15], and in the brain, cosediments with both translating polyribosomes [16] and with mRNPs [12]. The RGG box of FMRP binds to intramolecular G quartet sequences in target mRNAs [17], while the KH2 domain has been proposed to bind to so-called kissing complex RNAs based on in vitro selection assays [18]. In addition, FMRP binds to uridine-rich mRNAs [19,20]. In aggregate, more than 500 mRNA ligands for FMRP have been identified, many with the potential to influence synaptic formation and plasticity [10,17].
FMRP is required for type 1 metabotropic glutamate receptor (mGluR)–dependent translation of synaptic proteins, including FMRP and postsynaptic density 95 (PSD-95) [21,22]. Both PSD-95 and FMRP mRNAs contain putative G-quartets in their 3′-UTR and coding sequence, respectively [22,23]. Database searches revealed that APP mRNA possesses a G-quartet–like motif in the coding region (position 825–846 of the mouse sequence) embedded within a guanine-rich domain (694–846) containing several DWGG repeats. APP mRNAs (70% of APP695 and 50% of APP751/770) are associated with polyribosomes in rat brain [24], suggesting that translational regulation could play an important role in APP production. Indeed, APP contains a 5′-UTR iron response element previously implicated in translation control [25]. Therefore, we hypothesized that APP mRNA translation would be regulated by FMRP.
We now show that after stimulation with the mGluR agonist DHPG, APP levels increased significantly in wild-type (WT) but not synaptoneurosomes (SNs) or cultured neurons from knockout (KO) animals. In KO SNs or neurons, APP was constitutively elevated. APP mRNA coimmunoprecipitated with FMRP in WT, resting SNs, but this interaction was lost with DHPG treatment. FMRP monomer bound to the 5′ end of the G-rich sequence in the coding region of APP mRNA. Our data indicate that FMRP represses the translation of APP through mGluR-dependent interactions with APP mRNA. Consistent with constitutively elevated APP levels, the proteolytic products Aβ40 and Aβ42 are elevated in the brains of fmr-1 KO mice compared to WT.
Our laboratory has shown that FMRP and PSD-95 mRNAs are rapidly translated in mouse primary cortical neurons in response to the type 1 mGluR agonist DHPG [22]. Normal regulation was lost in fmr-1 KO-derived neurons, implicating FMRP in this process. These and other FMRP-regulated mRNAs contain G-quartets, which have been proposed as at least one site of mRNA/FMRP interaction [17]. Database searches of brain mRNAs revealed that the coding region of human, mouse, and rat APP mRNAs contained a G-quartet–like sequence (Figure 1A) within a G-rich domain containing several DWGG repeats (Figure 1B). The putative G-quartet motif in APP mRNA has the potential to form a stable structure containing three guanine planes (Figure S1). FMRP binds to G-rich sequences (so-called G-quartets; consensus site: DWGG-N(0–2)-DWGG-N(0–1)-DWGG-N(0–1)-DWGG, where D is any nucleotide except C and W is A or U) [17] arranged in a planar conformation and stabilized by Hoogsteen-type hydrogen bonds. In the human APP mRNAs, the G-rich region containing the putative G-quartet motif is found in all three splice variants (APP695, APP751, and APP770; 87 nucleotides upstream of the sequence coding for the Kunitz-type protease inhibitor domain, which is missing in APP695). FMRP also binds to kissing complex sites [18], but APP mRNA lacks such a site. Therefore, we prepared cortical lysates as well as SNs from WT mice and immunoprecipitated FMRP. Contrary to a previous report utilizing a different protocol for the preparation of SNs [26], APP mRNA is present in SNs, and reverse transcription (RT)–PCR revealed that the message was brought down with specific, but not control, antisera in cortical lysates as well as SNs (unpublished data). Thus, APP mRNA is a potential target of FMRP, presumably via the coding region putative G-quartet.
APP is highly expressed in neurons and dendrites and may promote synaptic maturation [1]. Conversely, overexpression of APP and its proteolytic product, Aβ, have been implicated in the synaptic losses seen early in the development of AD [27]. Therefore, we asked if APP translation was regulated by dendritic FMRP. We utilized fmr-1 KO mice, a rodent model for FXS, that display dendritic spine anomalies similar to that in the human disorder [28–30]. Cortical SNs were prepared from both WT and fmr-1 KO mice, and overall protein synthesis was analyzed in response to DHPG (100 μM)–induced mGluR activation. SNs from either animal were metabolically active with equivalent total 35S-Met incorporation (Figure 2). Therefore, FMRP was not required for basal protein synthesis, which is in agreement with a prior report [31]. However, we did not observe an increase in overall protein synthesis in response to DHPG, whereas Weiler and colleagues [31] observed a 1.3-fold increase in 35S-Met incorporation after 5 min of stimulation.
To assess de novo synthesis, 35S-labeled WT or KO SNs were immunoprecipitated with anti-APP. After 15 min of incubation, untreated WT SNs translated modest amounts of APP, which rapidly increased by 2.7-fold with DHPG treatment. After 1 hr, APP remained elevated in stimulated SNs over the control, but the difference was less (1.6-fold) than at 15 min, suggesting more persistent translation in the unstimulated controls, slowing of new synthesis after stimulation, and/or compensatory protein turnover in the DHPG-treated samples (Figure 3A and 3B). In KO SNs, APP synthesis was less than in WT SNs and showed a minimal response to DHPG. The translational inhibitor anisomycin blocked DHPG-mediated synthesis of APP, as did the specific mGluR5 inhibitor MPEP (Figure 3C and 3D).
In order to assess changes in steady-state levels, rather than new protein synthesis, APP was measured in WT and KO SNs in response to DHPG by Western blot analysis (Figure 4). In WT SNs, there was a rapid increase in total APP levels within 5 min of DHPG treatment (1.6-fold, n = 3), which was completely absent in KO SNs. Regardless of treatment, APP levels remained nearly constant over time in KO SNs, as did β-actin. In the absence of DHPG, steady-state levels of APP were substantially higher in KO SNs compared to WT SNs. Within 20 min of DHPG treatment, APP levels in WT SNs approached those seen in unstimulated KO SNs (Figure 4). Protease inhibitors increased steady-state levels of APP in WT SNs to those seen in KO SNs (unpublished data). These data suggest that APP mRNA is translationally repressed by FMRP in unstimulated WT SNs. mGluR activation rapidly derepresses APP synthesis as shown for FMRP and PSD-95 [21,22]. APP levels during maximal derepression approach those seen constitutively in fmr-1 KO cells. After the cessation of mGluR signaling, APP levels presumably drop due to degradation, which appears more robust in WT than KO cells.
SNs are a relatively crude preparation of pre- and postsynaptic densities that are contaminated with other cell types, such as astrocytes, which form synapses with neurons. Thus, we prepared primary embryonic day–18 cortical neuron cultures from WT and fmr-1 KO brains and assessed dendritic APP levels by immunofluorescence. APP was found in the cell body as well as dendritic puncta of both WT and fmr-1 KO neurons (Figure 5A). There was a 21% increase in the basal level of APP in untreated fmr-1 KO neurons compared to WT (Figure 5B). Neurons stimulated with DHPG for 10 and 20 min prior to cell fixation showed a 18%–25% increase in dendritic APP levels in WT but no increase in fmr-1 KO cultures (Figure 5B). These data confirm our findings in SNs that (1) fmr-1 KO mice have higher basal synaptic levels of APP, and (2) DHPG increases APP levels selectively in WT samples. These data also demonstrate that FMRP and mGluR activation regulate APP synthesis in both FVB and C57BL/6 mice, as the SNs were prepared from the former strain, and the primary cortical neurons from the latter strain.
FMRP and homologs have been implicated in the control of mRNA decay. There are increased APP mRNA levels in the cerebral cortex, hippocampus, and cerebellum in a FXS mouse model [32], and FXR1P, an FMRP homolog, is an AU-rich element–binding protein that binds to and regulates TNFα mRNA stability and translation [33]. APP mRNA contains two 3′-UTR cis-elements within 200 bases of the stop codon that mediate message stability. Hence, we analyzed APP mRNA and 18S rRNA decay in SNs by real-time PCR. APP mRNA did not decay over 120 min regardless of mGluR activation in WT and KO SNs (Figure S2). These data indicate that mGluR-dependent APP translation was independent of mRNA stabilization. APP mRNA has a half-life of approximately 5 h in resting immune cells, which is prolonged in activated cells [34,35] or rat PC12 (Westmark and Malter, unpublished data). Thus, APP mRNA decays with comparable kinetics in SNs and mammalian cells.
The mechanism underlying FMRP-mediated translational repression is controversial [36]. Alterations in the association of FMRP with polyribosomes, small nontranslated RNAs, or other proteins have all been proposed [9,12,37,38]. We asked if the APP mRNA/FMRP interaction changed after DHPG. Thus, FMRP was immunoprecipitated from WT SNs (60 min after DHPG), and the pellet was reverse transcribed and analyzed by real-time quantitative PCR (qPCR). APP mRNA was readily detected in anti-FMRP pellets in untreated WT SNs (Figure 6A). However, APP mRNA associated with FMRP could not be detected in DHPG-stimulated WT SN immunopreciptates (IPs) or in the KO with or without DHPG within 40 cycles of real-time PCR. The negative controls for this experiment were duplicate IPs in the absence of 7G1–1 FMRP antibody, which also did not produce any real-time PCR Ct values for APP mRNA within 40 cycles (data not shown). The >60-fold difference in FMRP-associated APP mRNA was highly significant. Evaluation at earlier times revealed that the APP mRNA–FMRP complex was lost within 5 min of DHPG treatment (unpublished data).
Immunoprecipitation of FMRP from WT SNs followed by Western blotting (Figure 6B) or 35S-Met incorporation analysis (unpublished data) demonstrated that DHPG treatment does not interfere with the ability of anti–7G1–1 antibody to bind to FMRP. In fact, in both assays there was more FMRP immunoprecipitated from the DHPG-treated WT SNs, which is in agreement with previous reports that DHPG stimulates the dendritic translation of FMRP [22,39]. Our data suggest that physical interactions between FMRP and APP mRNA underlie translational repression, with mGluR activation rapidly moderating these events. Presumably, the loss of FMRP/APP mRNA interaction results in rapid, pulsatile protein expression in dendrites.
FMRP is a component of large RNP complexes [38]. The data presented here demonstrate that APP mRNA is also associated with this RNP. To determine the likely interaction site, in vitro RNase protection assays were performed on FMRP IPs from whole-cortex lysates. Residual APP mRNA was mapped by RTqPCR with primers immediately surrounding the predicted G-quartet (Figure 7A). Surprisingly, the G-rich area immediately preceding the G-quartet (nt 699–796) was approximately 4-fold more protected from nuclease digestion than fragments containing the predicted G-quartet (825–846). Although this protected area does not contain a canonical G-quartet motif, the sequence is very G-rich and contains several closely spaced DWGG repeats. The smallest amplicon (nt 774–871) containing the predicted G-quartet motif amplified a 98-base fragment, of which 46 nucleotides were guanines (47% G-rich; Table S1). Although this is the most G-rich amplicon of those tested, and T1 ribonuclease cuts 3′ of single-stranded G-residues, the 98-nt protected fragment (amplicon 699–796) was 40% G-rich, providing nearly equivalent numbers of targets for digestion. Thus, nucleotides 699–796 in the coding region of APP mRNA possess a G-rich sequence that is protected from nuclease digestion by an RNP complex containing FMRP.
The FMRP-containing RNP complex likely protects other cis-elements in APP mRNA as well. Our laboratory has defined a 29-base element located 200 nucleotides downstream of the stop codon in APP mRNA that regulates message decay [35]. We have also identified two proteins, nucleolin and hnRNP C, that bind to this 29-base element [40]. Since nucleolin interacts with FMRP to form multiprotein complexes [38], we would expect the 29-base element to be protected from T1 ribonuclease digestion of anti-FMRP IPs, as shown in Figure 7A (nt 2318–2416). Our data suggest that multiple cis-regulatory elements of APP mRNA interact with the FMRP-containing RNP complex.
Despite the presence of an RNase-protected, G-rich sequence, APP mRNA might not associate directly with FMRP. To determine if this was the case and to further characterize the interaction, we performed a modified CLIP assay [41]. SNs were cross-linked with ultraviolet light, immunoprecipitated with anti-FMRP, digested with T1 ribonuclease, and analyzed by SDS-PAGE. FMRP immunoreactive material (80 kDa) was excised and analyzed by RTqPCR. The amplicon encompassing the G-rich sequence (nucleotides 699–796) of APP mRNA gave a positive signal that was approximately 5-fold greater than that of the predicted G-quartet motif–containing sequence (nt 774–871) immediately downstream (Figure 7B). Thus, our data define the G-rich region immediately preceding the predicted G-quartet as the binding site between FMRP and APP mRNA. The loss of FMRP binding at the G-rich region presumably derepresses APP translation, as it was contemporaneous.
Increased translation of APP provides more targets for cleavage by β- and γ-secretases. Therefore, we would expect fmr-1 KO mice to have exacerbated Aβ production with aging. Right-brain hemispheres from middle-aged FVB mice (11–13 mo old) were homogenized in protein extraction buffer containing 1% Triton X-100 and protease inhibitors and the soluble material was analyzed by enzyme-linked immunosorbent assay (ELISA) for Aβ40 and Aβ42. The fmr-1 KO mouse brain contained 1.6 times more Aβ40 and 2.5 times more Aβ42 than WT controls (Figure 8A). We also tested Aβ40/42 levels in C57BL/6 mice (12–14 mo old) to ensure that this was not a strain-specific event. We did not observe an increase in soluble Aβ40 or Aβ42 levels in fmr-1 KO C57BL/6 brain samples (unpublished data), but guanidine-soluble fractions showed a 2.8-fold increase in Aβ40 and a 1.2-fold increase in Aβ42 (Figure 8B). Therefore, the brains of two distinct murine strains lacking fmr-1 both showed increased APP and accumulated pathogenic Aβ species over time.
Synaptic plasticity is required for normal learning and memory and is impaired in FXS. High dendritic spine density is normal for young mice, but synapse pruning during postnatal development is absent in the KO, resulting in increased spine density in adulthood [42]. The molecular basis for defective pruning in fmr-1 KO mice is unknown, but likely reflects the loss of FMRP-regulated translation of synaptic mRNA. FMRP regulates group 1 mGluR-dependent translation of mRNA targets important in diverse neuronal functions [36]. For example, FMRP normally represses the translation of microtubule-associated protein 1B (MAP1B) mRNA during synaptogenesis. In FXS, MAP1B expression is constitutively elevated, leading to abnormally increased microtubule stability [43]. Therefore, it is of great interest to identify FMRP-dependent synaptic mRNAs that contribute to dendritic structure and function.
Herein, we show that APP mRNA is a previously unappreciated target for FMRP-mediated translational repression at the synapse. The normal physiologic role of APP remains ill defined, but increasing evidence suggests an important role in synapse formation [44,45] and maturation [1]. APP localizes to postsynaptic densities, axons, dendrites, and neuromuscular junctions [1,46]. APP/APP-like protein 2 double-KO mice exhibit defective neuromuscular junctions, excessive nerve terminal sprouting, and defective synaptic transmission [47]. APP is developmentally regulated with maximal expression during synaptogenesis and subsequently declines when mature connections are completed [48,49]. Therefore, synaptic overexpression of APP during early development may contribute to the immature dendritic spines and inadequate synaptic pruning characteristic of FXS.
We have identified a G-rich region located within nucleotides 699–796 in the coding region of APP mRNA as an FMRP-binding site. The G-quartet–like sequence immediately downstream of this G-rich region was not protected from nuclease digestions of FMRP IPs. This result was surprising because the intramolecular G-quartet motif has been identified by in vitro RNA selection assays as the site of interaction with FMRP [17]. As expected from the FMRP interaction site mapping results, alignment of the G-rich region (nt 699–796) and DWGG repeats of mouse APP mRNA is highly conserved with both the human (86%) and rat (93%) sequences. Our data suggest that there may be flexibility in the spacing of the DWGG repeats for G-quartet formation and agrees with previous findings that the presence of a G-quartet does not ensure binding by FMRP [17].
We have determined that FMRP associates with APP mRNA in SN preparations, and that translation of APP mRNA increases in response to DHPG. DHPG-upregulated translation of APP can be blocked by the translational inhibitor anisomycin or the mGluR5-specific inhibitor MPEP. mGluR-mediated translation is concurrent with FMRP dissociation from APP mRNA and is independent of mRNA decay. The rapid dissociation of FMRP from APP mRNA, in response to mGluR activation, suggests that post-translational modifications, such as phosphorylation, may regulate FMRP binding activity. Ceman and colleagues have shown that FMRP is phosphorylated N-terminal to the RGG box and that phosphorylation/dephosphorylation status of the protein is correlated with binding to stalled versus active polyribosomes [50]. Our data support a model developing in the literature whereby FMRP acts as an immediate early-response protein that regulates translation at the synapse. When FMRP is bound to APP or other synaptic mRNAs, translation is repressed. Upon mGluR activation, FMRP is released from the nontranslating RNP, resulting in prompt protein synthesis. In FXS, high levels of protein are constitutively produced that are normally translationally repressed by FMRP.
We would predict that constitutively upregulated APP would lead to increased processing to Aβ. Indeed, increased Aβ40 and Aβ42 are present in two mouse models for FXS. To date, the only abnormal neurpathologic observations in the human FXS brain have involved impaired synaptic pruning and maturation [51]; however, a very limited number of aged FXS brains have been studied [4,29,52], so other neuropathologies, such as increased amyloid burden and synaptic degeneration normally associated with AD, cannot be excluded. It is difficult to measure cognitive decline in mentally retarded individuals; however, in support of our prediction, fragile X–associated tremor/ataxia syndrome in males is associated with dementia [53].
The normal physiologic function(s) of APP are not well understood, albeit the protein is likely important for synapse formation in the developing brain [1]. A recent report demonstrates that children with severe autism and aggression express >2-fold more secreted βAPP (1,200 pg/ml) than children without autism (500 pg/ml) [54]. Many people with FXS (67% of men and 23% of women) are also autistic [55]. Interestingly, the highest levels of secreted βAPP were found in two children with FXS [54]. Thus, overproduction of secreted βAPP may contribute to the developmental disabilities observed in patients with FXS and autism. In addition, FMRP mRNA and protein expression are downregulated as a function of aging in the mouse brain [56], suggesting that repressed transcripts, such as APP, would be upregulated with aging, a well-known phenomenon in animals and humans.
In conclusion, FMRP represses translation of APP mRNA in dendrites, suggesting a link between two neurodevelopmental disorders, FXS and autism, and a neurodegenerative disease, AD.
The anti-FMRP antibody (mAb7G1–1) [10] was obtained from the Developmental Studies Hybridoma Bank, University of Iowa (http://www.uiowa.edu/~dshbwww). The anti-APP polyclonal antibody (catalog number 51–2700) was purchased from Zymed Laboratories (http://www.invitrogen.com), and the anti-mouse β-actin antibody (catalog number A5441), protease inhibitor cocktail (catalog number P2714), ribonuclease T1 (catalog number R1003), and poly(D)-lysine (catalog number P6407) were purchased from Sigma Chemical Company (http://www.sigmaaldrich.com). The anti-rabbit and anti-mouse HRP-conjugated secondary antibodies, percoll, Redivue Pro-Mix-L [35S] (catalog number AGQ0080) and enhanced chemiluminescence detection reagents were obtained from Amersham Pharmacia (http://www5.amershambiosciences.com). Anti-22C11 APP antibody (mAB348) was acquired from Chemicon (http://www.chemicon.com). The rabbit polyclonal Aβ40 (catalog number 9131), Aβ42 (catalog number 9134), and rodent Aβ (catalog number 9154) antibodies were purchased from Signet Laboratories (http://www.signetlabs.com). DHPG (catalog number 0805) was obtained from Tocris Cookson (http://www.tocris.com). Omniscript RT was acquired from Qiagen (http://www.qiagen.com). The MagnaBind Protein A beads, PAGEprep advance kit, and micro BCA protein assay reagent kit were obtained from Pierce Biotechnology (http://www.piercenet.com). DNA oligonucleotides were synthesized by Integrated DNA Technologies (http://www.idtdna.com), and SYBR Green PCR master mix was obtained from Applied Biosystems (http://www.appliedbiosystems.com). NeuroBasal medium, B27 supplement, goat anti-mouse rhodamine-conjugated antibody, and ProLong Gold Antifade with DAPI were from Invitrogen (http://www.invitrogen.com). TRI-Reagent was purchased from Molecular Research Center (http://www.mrcgene.com). MPEP was purchased from Tocris Cookson or synthesized by Technically (http://www.technically.com) and provided by FRAXA Research Foundation (http://www.fraxa.org).
The WT and fmr-1 KO mice in the FVB and C57BL/6 backgrounds were a generous gift from Aaron Grossman and Dr. Bill Greenough (University of Illinois at Urbana-Champaign). The fmr-1 KO mice were originally developed by Frank Kooy and backcrossed >11 times to FVB mice, albeit these FVB mice have the genes for pigmentation and normal vision [28]. Mice were housed two to four per microisolator cage on a 6am–6pm light cycle with ad libitum food (Purina 5015 mouse diet; http://www.purina.com) and water. The cages contained seeds and a neslet as the only sources of environmental enrichment. All animal husbandry and euthanasia procedures were performed in accordance with the National Institutes of Health and an approved University of Wisconsin–Madison animal care protocol through the Research Animal Resources Center. fmr-1 genotypes were determined by PCR analysis of DNA extracted from tail biopsies. The FVB strain was used for all experiments described herein except for preparing the cultured neuronal cells (Figure 5) and the Aβ ELISAs (Figure 8B).
SNs were prepared from WT and fmr-1 KO mouse cortical tissue [57,58]. Briefly, mouse pups aged 14–17 d were killed by carbon dioxide asphyxiation followed by removal of the brain cortices. The cortices were washed in ice-cold gradient medium (GM buffer: 0.25 M sucrose, 5 mM Tris [pH 7.5], and 0.1 mM EDTA), transferred to a glass dounce homogenizer containing ice-cold GM buffer, and gently homogenized with five strokes of the loose pestle followed by five strokes of the tight pestle. The homogenate was spun at 1000g for 10 min at 4 °C in round-bottom tubes to pellet cellular debris and nuclei. The supernatant (2 ml aliquots) was applied to percoll gradients (layers = 2 ml each of 23%, 15%, 10%, and 3% isomotic percoll) and spun at speed (32,500g) for 5 min at 4°C. The third band from the top of the gradient (the 23%/15% interface) containing intact SNs was removed and pooled for the experiments. The two higher-molecular-weight bands at the 15%/10% and 10%/3% interfaces contain broken membranes. The salt concentration of the SNs was adjusted by adding one-tenth volume of 10× stimulation buffer (100 mM Tris [pH 7.5], 5 mM Na2HPO4, 4 mM KH2PO4, 40 mM NaHCO3, 800 mM NaCl). To suppress nonspecific excitation, 1 μM tetrodotoxin was added. The protein concentration of the SNs was determined by Bradford assay and ranged from 200–500 ng/μl.
SNs were equilibrated to room temperature by rotation on a nutator mixer for a minimum of 10 min. DHPG was dissolved in 1× stimulation buffer immediately prior to use and added to the SNs (100 μM final concentration). Samples were mixed at room temperature in 1.5 ml Eppendorf tubes for the indicated times.
WT and KO SNs (450 μl) were mixed with 25 μl Redivue Pro-Mix-L [35S] for 5 min prior to stimulation with 25 μl 2 mM DHPG. Samples were flash frozen at the indicated times. To analyze new protein synthesis, SN lysates were cleared of free isotope, percoll, and sucrose by purification with the PAGEprep Advance kit per the manufacturer's directions. Protein concentrations were determined by the BCA assay, and 15 μg protein was denatured and loaded per lane on 12% SDS gels. The gels were dried and exposed to a phosphorimager screen.
To specifically analyze APP synthesis, WT and KO SN lysates (500 μl) were immunoprecipitated with APP antibody. Briefly, SN lysates were precleared with protein A magnetic beads in 1 ml volumes containing 500 μl SNs, 500 μl 2× IP buffer (20 mM HEPES [pH 7.4], 400 mM NaCl, 60 mM EDTA [pH 8], and 2% Triton X-100), protease inhibitor cocktail, and 100 μl packed fresh protein A magnetic beads. For the immunoprecipitations, 10 μg anti-APP antibody (Zymed catalog number 51–2700) and fresh protein A magnetic beads were added and mixed overnight at 4 °C. The beads were washed three times with IP buffer, and the final, washed pellets were suspended in 40 μl 2× SDS sample buffer and boiled for 5 min; the proteins were then separated on 12% SDS gels. The gels were transferred to nitrocelluose membrane, dried, exposed to a phosphorimager screen, and scanned on a STORM 860 phosphorimager (Molecular Dynamics, http://www6.amershambiosciences.com). The 120-kDa APP bands were quantitated with ImageQuant software (GE Healthcare Life Sciences, http://www4.amershambiosciences.com).
For the inhibitor studies, SNs (425 μl) were preincubated with 25 μl anisomycin (40 μM final concentration) or MPEP (10 μM final concentration) for 10 min prior to the addition of 25 μl 35S-Met for 5 min and stimulation with DHPG (100 μM final concentration) for 15 min. Samples were processed as described in the preceding paragraph.
Aliquots of SNs were collected at 5, 10, and 20 min after DHPG treatment, quenched with an equal volume of 2× SDS sample buffer (8% SDS, 24% glycerol, 100 mM Tris [pH 6.8], 4% β-mercaptoethanol, 0.02% bromophenol blue, 2% Triton X-100, 2% deoxycholate, 2% NP-40 alternative, and 2% sarkosyl) and boiled for 5 min prior to analysis by 12% SDS-PAGE. The separated proteins were transferred to 0.45 μm nitrocellulose membrane in Towbin buffer containing 20% MeOH with a Criterion Blotter (Bio-Rad, http://www.bio-rad.com; 100 V at 4 °C for 75 min). The membranes were blocked in 5% nonfat dry milk and hybridized with anti-rabbit APP antibody (dilution, 1 μg/ml) and anti-mouse β-actin antibody (dilution, 1:20,000) followed by hybridization with anti-rabbit or anti-mouse HRP-conjugated secondary antibodies (dilution, 1:2000). Proteins were visualized by enhanced chemiluminescence on a STORM 860 phosphorimager.
Pregnant females (embryonic day 18) were anesthetized with halothane prior to decapitation and transfer of the uterine sac to ice-cold HBSS. Cortices were removed, washed with ice-cold HBSS, lysed with 0.5 mg/ml trypsin for 25 min at 37 °C, washed with HBSS, suspended in NeuroBasal medium (supplemented with 2% B27 supplement, penicillin/streptomycin, and 0.5 mM glutamine), triturated 70× with a 10-ml pipet, and passed through a 70-μm cell strainer. Cells were counted by trypan blue dye exclusion and plated at 1.25 × 105 cells/ml on poly(D)-lysine–coated glass coverslips in 12-well tissue-culture dishes and cultured for 11 d at 37 °C/5% CO2. Half of the culture medium was removed and replaced with fresh, warm medium on day 4.
Neuronal cells were treated with 100 μM DHPG, washed with PBS containing 2% FBS, fixed in 4% PHA for 10 min at room temperature, and permeabilized with MeOH (−20 °C) for 15 min. Fixed, permeabilized cells were stained with anti-22C11 against the amino-terminus of APP (Chemicon number mAB348; 1:2000 for 1 h) and visualized with goat anti-mouse rhodamine-conjugated secondary antibody (Invitrogen; 1:500 for 30 min in the dark). All washes and antibody dilutions were in PBS containing 2% FBS. Coverslips were fixed to slides with 12 μl ProLong Gold Antifade with DAPI (Invitrogen) and dried overnight.
Images were acquired with a Nikon C1 laser-scanning confocal microscope with EZ-C1 v2.20 software (Nikon, http://www.nikon.com) at 60× magnification. APP levels in the puncta of four to seven dendrites per sample were quantitated with IMAGE J software using the Analyze Particles function (minimum of 205 puncta analyzed per treatment) (Rasband, W.S., Image J, U.S. National Institutes of Health, http://rsb.info.nih.gov/ij; 1997–2006). Figures were assembled with Adobe Photoshop 8.0 (Adobe Systems, http://www.adobe.com). All DHPG-treated and fmr-1 KO samples were highly statistically different from untreated WT samples by t-test analyses (p <0.001) and expressed as SEM.
Aliquots of SNs were collected at the indicated timepoints and flash frozen at −80 °C. The samples were thawed and vortexed to prepare SN lysates. To directly reverse-transcribe RNA from SN lysates without an RNA purification step, a modified method for the detection of mRNA in single neurons was utilized [59]. Briefly, 2 μl SN lysate was added per standard RT reaction containing RNase-free DNase I and random nonamer primer. The reactions were incubated at 37 °C for 15 min to destroy any contaminating genomic DNA, 65 °C for 5 min to inactivate the DNase I, and 20 °C for 10 min to anneal the random primer. Omniscript RT was added and reverse transcription proceeded at 37 °C for 60 min before inactivation at 93 °C for 5 min. The RT reactions were diluted 5-fold with DEPC water prior to real-time PCR analysis. For the statistical analysis, APP mRNA levels from triplicate experiments were determined, normalized to 18S rRNA, and plotted as a percentage of total APP mRNA. Error bars depict SEM.
The PCR primers were designed with Primer Express software from Applied Biosystems, and BLAST homology searches of the amplicons revealed that the primers were gene specific. PCR reactions were optimized for primer and template concentrations and contained 500 nM APP primers (forward: 1701-ccgtggcacccttttgg-1717; and reverse: 1774-gggcgggcgtcaaca-1760) or 300 nM 18S primers (forward: 98-cattaaatcagttatggttcctttgg-123; and reverse: 181-tcggcatgtattagctctagaattacc-155), 10.5 μl 1:5 diluted RT reaction and 12.5 μl SYBR green PCR mix in a 25 μl reaction volume. The cycle conditions were as follows: 2 min at 50 °C and 10 min at 95 °C (40 cycles: 15 s at 95 °C, 1 min at 60 °C), followed by a dissociation stage for 15 s at 95 °C, 1 min at 60 °C, and 15 s at 95 °C. The average PCR efficiencies for the APP and 18S primers over a 200-fold concentration range were 100% (APP) and 101% (18S) (n = 9 experiments each), with a delta slope of 0.079. As the difference in slopes between the sample PCR (APP) and the normalization control (18S) was less than 0.1, the comparative CT method was utilized to calculate the relative concentration of APP mRNA normalized to 18S rRNA. SNs are void of nuclei; however to ensure there was no genomic DNA contamination, control RT reactions on SN templates in the absence of reverse transcriptase were analyzed by real-time PCR and found void of APP PCR product. The final APP and 18S PCR products were analyzed by agarose gel electrophoresis and were single bands of the correct molecular weight by EtBr staining (74 bp for APP; 84 bp for 18S).
SN lysates were precleared with protein A magnetic beads and immunoprecipitated with 10 μl RNasin, 10 μg 7G1–1 FMRP antibody (or no antibody controls), and 100 μl packed fresh protein A magnetic beads for 3 h at 4 °C. The IPs were washed with IP buffer (10 mM HEPES [pH 7.4], 200 mM NaCl, 30 mM EDTA [pH 8], and 0.5% Triton X-100) and suspended in 1 ml TRI-Reagent. Total RNA was isolated and precipitated in the presence of 2 μg tRNA. The final pellet was suspended in DEPC water, solubilized 10 min at 60 °C, and reverse transcribed with Qiagen Omniscript and random nonamer primer (60 min at 37 °C, 5 min at 93 °C). The cDNA was diluted 5-fold and analyzed for APP by qPCR as described immediately above.
The cortices from six WT mice (13 d old) were torn into pieces and homogenized in cold immunoprecipitation buffer (10 mM HEPES [pH 7.4], 200 mM NaCl, 30 mM EDTA [pH 8], and 0.5% Triton X-100) containing 2× protease inhibitor cocktail and 0.4 U/μl RNasin. The homogenate was spun at 1,000g for 10 min at 4 °C to remove nuclei and unlysed cells, and the pellet was discarded. The cleared lysate was flash frozen in aliquots at −80 °C.
Pellets from anti-FMRP immunoprecipitations of whole-cortex lysate were washed once with immunoprecipitation buffer and once with DPBS before digestion with ribonuclease TI (0.8–4.0 U) in a 100-μl reaction volume for 30 min at 37 °C with occasional mixing to disperse the magnetic protein A beads. The digested samples were washed twice with DPBS to remove RNA fragments. Protected RNA was isolated with TRI-Reagent and analyzed by RTqPCR. The primer sequences for the real time PCR are listed in Table S2. The delta Ct between undigested and T1-digested samples was calculated and plotted as a percentage of APP699–796 mRNA.
For the modified CLIP assay [41], cleared cortical lysate was cross-linked with 400 mJ/cm2 ultraviolet light in an UV Stratalinker 2400 (Stratagene, http://www.stratagene.com), immunoprecipitated with anti-FMRP, and digested with ribonuclease TI. The washed pellets were suspended in 40 μl SDS loading buffer containing no reducing agents, heated for 10 min at 70 °C, applied to 12% SDS/PA gels, and transferred to 0.45 μm nitrocellulose membrane in Towbin buffer. Western blotting of a duplicate membrane indicated that FMRP migrates at 80 kDa. A band encompassing approximately the 75–85 kDa molecular weight range was excised, transferred to TRI-Reagent, and vortexed vigorously for 15 min at 37 °C. RNA was isolated and analyzed by RTqPCR.
For soluble brain lysates, right hemispheres from four WT (aged 11, 13, 13, and 13 mo) and three KO mice (aged 11, 12, and 12 mo; FVB strain) and four WT (aged 13.5, 13.5, 12, and 12 mo) and four KO mice (aged 14, 14, 12, and 12 mo; C57BL/6 strain) were homogenized in 500 μl protein extraction buffer (10 mM Tris [pH 7.6], 2 mM EDTA, 150 mM NaCl, 1% Triton X-100, 0.25% NP-40, and 1× protease inhibitor cocktail). Insoluble material was removed by centrifugation at 12,000 rpm for 10 min, and aliquots of the soluble fraction were flash frozen. For total brain lysates, left hemispheres were homogenized in cold, 5 M GnHCl, mixed for 3–4 h at room temperature, and frozen at −80 °C. Sandwich ELISAs with the Signet Aβ40/9131 and Aβ42/9134 capture antibodies and the rodent Aβ/9154 reporter antibody were performed as previously described [60]. Aβ levels were quantified based upon standard curves run on the same ELISA plate and then expressed as a percentage of Aβ compared to WT controls.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the gene products mentioned in this paper are human APP mRNA (NM_000484), mouse APP mRNA (X59379), rat APP mRNA (X07648), and 18S mRNA (M27358).
|
10.1371/journal.pcbi.1000453 | A Generalized Framework for Quantifying the Dynamics of EEG Event-Related Desynchronization | Brains were built by evolution to react swiftly to environmental challenges. Thus, sensory stimuli must be processed ad hoc, i.e., independent—to a large extent—from the momentary brain state incidentally prevailing during stimulus occurrence. Accordingly, computational neuroscience strives to model the robust processing of stimuli in the presence of dynamical cortical states. A pivotal feature of ongoing brain activity is the regional predominance of EEG eigenrhythms, such as the occipital alpha or the pericentral mu rhythm, both peaking spectrally at 10 Hz. Here, we establish a novel generalized concept to measure event-related desynchronization (ERD), which allows one to model neural oscillatory dynamics also in the presence of dynamical cortical states. Specifically, we demonstrate that a somatosensory stimulus causes a stereotypic sequence of first an ERD and then an ensuing amplitude overshoot (event-related synchronization), which at a dynamical cortical state becomes evident only if the natural relaxation dynamics of unperturbed EEG rhythms is utilized as reference dynamics. Moreover, this computational approach also encompasses the more general notion of a “conditional ERD,” through which candidate explanatory variables can be scrutinized with regard to their possible impact on a particular oscillatory dynamics under study. Thus, the generalized ERD represents a powerful novel analysis tool for extending our understanding of inter-trial variability of evoked responses and therefore the robust processing of environmental stimuli.
| When Hans Berger described the human EEG in the 1920s, a pivotal finding was the demonstration of prominent oscillations in the frequency range between 8 and 12 Hz, which he called alpha wave rhythm. He also described for the first time the so-called “alpha blockade,” i.e., the suppression of the ongoing alpha activity when the subject opens his eyes. Based on these early findings, induced changes of macroscopic EEG oscillations have been reported for diverse physiological manipulations and processing of sensory information. The magnitude and the latency of these induced changes are, however, subject to variations, even if identical stimuli are processed. In order to enable investigations of the underlying neural mechanisms of these variations, we here establish a mathematical framework which allows one to scrutinize candidate explanatory factors with regard to their possible impact on the characteristics of the induced oscillatory dynamics.
| When Hans Berger [1] described the human EEG in the 1920s, a pivotal finding was the demonstration of prominent oscillations in the frequency range between 8 and 12 Hz, which he called alpha wave rhythm. He also described for the first time the so-called “alpha blockade”, i.e., the suppression of the ongoing alpha activity when the subject opens his eyes. In the 1970s Gert Pfurtscheller and colleagues [2] introduced the term event-related desynchronization (ERD) for this kind of frequency specific changes of ongoing EEG activity. Based on these findings induced changes of oscillations have been reported for diverse physiological manipulations and processing of sensory information. For instance voluntary movement results in a circumscribed desynchronization in the upper alpha and lower beta bands, localized close to sensorimotor areas [3],[4]. A desynchronization localized to the auditory cortex following auditory stimuli was reported in MEG recordings [5]. Moreover, the alpha band rhythms demonstrate a relatively widespread desynchronization in perceptual, judgement and memory tasks [6],[7]. In contrast the upper alpha band desynchronization is often topologically restricted, e.g., it develops during the processing of semantic information over the left hemisphere, where the degree of desynchronization is closely linked to semantic memory processes [8]. In addition to oscillations in the alpha and lower beta band, induced oscillations were also reported for the frequency band around 40 Hz with visual stimulation [9] and in movement tasks [10],[11] (for a comprehensive review on ERD cf. [12],[13]).
Beside ERD, EEG correlates of stimulus processing comprise evoked event-related potentials (ERPs); these are commonly assessed by averaging over many instances of stimulus presentations to reduce unrelated EEG activities which can dominate the single-trial responses. To comprehend the interrelationship between evoked and ongoing rhythmic activity various studies have examined the impact of ongoing cortical activity on the latency and the magnitude of ERP components [14]–[19]. Notably, however, the inter-trial variability of ERD itself is not yet fully understood as there exist only a few investigations on the influence of exogenous factors such as stimulus intensity or interstimulus interval on the characteristics of ERD (see, e.g., [20]–[22]) and even less studies on endogenous factors such as attention or the phase and magnitude of EEG rhythms (see, e.g., [23]–[26]). Basically, an adequate data analytical framework for a “state-conditional ERD” is missing which could capture the impact of fluctuating brain states on inter-trial ERD variability. As we will illustrate the customary ERD measure impedes the analysis of state-conditional dependencies of the ERD on endogenous or exogenous factors. Specifically, we will identify the constant baseline, as it is incorporated as reference in the conventional ERD model, as the main cause which hampers a reliable analysis of the ERD variability. In particular, we will show that the use of a constant baseline as reference can lead even to spurious observations of ERD and event-related synchronization (ERS). Based on this result, we generalize the ERD concept by first substituting the constant baseline by a dynamic reference and then derive a reliable measure for state conditional ERD.
To this end the paper is organized as follows: First, we briefly analyze the conventional ERD framework and derive a generalized ERD concept. Second, we extend both, the conventional and the generalized ERD measure towards the analysis of state dependencies. With the application in section “Results” we first comparatively study the capabilities of the two alternative concepts in retrieving known state dependencies by means of artificially generated data. Afterwards, on the basis of a case study, we will outline how our novel framework can be used to investigate the impact of three endogenous factors on the latency and magnitude of the ERD response in the somatosensory system. A discussion along with an outlook concludes the paper.
One of the authors (SL), who had previous experiences with the acquisition of somatosensory evoked potentials (SEP) as a risk-free routine clinical procedure, served as volunteer subject for the proof-of-concept SEP recording.
To prepare for the introduction of the generalized ERD concept, we first present a brief outline of the conventional ERD measure. The standard measure of ERD quantifies a change in signal band power as difference between a baseline period prior to the event and an post-event period. Typically, the ERD is evaluated as the averaged response over a set of single trials. Up to now, two - basically similar - methods for estimating the ERD have been established, namely the power method [3] and the inter-trial variance method [27]. The advantage of the latter lies in the fact, that it compensates for the spectral bias which is introduced by phase-locked components. However, as the inter-trial variance method requires a slightly more complicated notation, but can be straightforwardly derived from the power method, we will for sake of simplicity introduce the conventional as well as the novel generalized ERD framework solely along the lines of the power method.
In order to attain a mathematical expression of the customary ERD, let denote the instantaneous signal power in a narrow frequency band during the event-related period . Moreover, let denote the averaged power in the reference period , that is(1)Denoting the expectation value, i.e., the average across trials, by , the traditional ERD at time is defined as(2)By convention an ERD corresponds to a negative value, i.e., a decrease in power, while ERS refers to an increased signal power [3]. Note that the changes of the signal power are quantified only with respect to the deviation from the fixed, constant baseline level . The conventional view on ERD is illustrated in Fig. 1-A.
We start with the following consideration: if an unperturbed dynamics is stationary it follows, that the expectation value is a constant function and therefore independent of . Thus any point in time could be used to empirically estimate this constant value, just by averaging across trials (independent realizations of ). However, if the dynamics is non-stationary, e.g., exhibits a deterministic negative trend, then the expectation value is not necessarily constant and therefore depends on .
Consequently, in order to quantify event-related changes of a non-stationary dynamics, an appropriate baseline should reflect the deterministic portion of the unperturbed non-stationarity dynamics. Hence, instead of using a fixed, static reference value, the generalized ERD measure uses the expected unperturbed dynamics as dynamic reference and therefore contrasts the expected dynamics of the instantaneous signal power between an unperturbed and an event-related condition. In order to get a reliable estimate of the expected unperturbed dynamics, we propose the use of so-called catch trials, which can be drawn from a continuous EEG measurement during time periods without the occurrence of the event under study (e.g., without somatosensory stimulus or a self-paced movements). This enables us to contrast event-related and reference dynamics directly. Therefore, we define the generalized ERD as the relative difference between both dynamics. Mathematically speaking, let be a binary variable, that distinguishes between the two types of single trials, i.e., between catch and event-related trials . Then we define the generalized ERD as(3)Here and denote the conditional expectation of the band power at time for the event-related and the unperturbed condition, respectively. Complying with the notation of the conventional framework a desynchronization corresponds to negative values, i.e., a decrease in power, while an increase in signal power indicates an event-related synchronization (ERS).
By means of a customized example of somatosensory induced ERD/S Fig. 1 illustrates the two different notions of measuring ERD. In this example the ERD/S is induced at a non-stationary cortical state, that is characterized by a prominent negative drift in the signal band power (readily identifiable from the unperturbed dynamics in the right panel). Consequently, the conventional and the generalized ERD/S yield significantly different results. Most conspicuously, the conventionally measured ERS lasts for a much shorter period and its peak would also be reduced in magnitude. Moreover, relative to the static baseline, the event-related dynamics drops below this level for a second time subsequent to the ERS period. According to the conventional interpretation this would indicate a second ERD phase. However, the cause of this spurious second ERD can be directly attributed to the non-stationary cortical state at stimulus onset. In contrast, the generalized framework which directly compares against the dynamic reference, which captures the deterministic trend, can deal with this phenomenon and yields the familiar ERD-ERS complex.
Note that if the unperturbed dynamics is stationary, then the expected reference dynamics is a constant and is equal to the conventional baseline . Therefore the conventional and the generalized measure of ERD will coincide with each other in case of analyzing stationary dynamics. In this sense the proposed framework constitutes a generalization of the conventional ERD towards the analysis of spectral perturbations in the presence of dynamical cortical states. Accordingly, the difference between the two approaches only becomes evident when analyzing non-stationary dynamics. One particular field of application of the generalized ERD measure is the analysis of state conditional dependencies of ERD, where the conditional dynamics are not necessarily stationary.
To enable investigations of the influence of arbitrary factors, such as the reaction time in a behavioral response paradigm or the magnitude of a particular EEG eigenrhythm, on the characteristic of the ERD (e.g., the ERD latency or magnitude), we incorporate an additional conditional variable into the ERD measures. To this end, let be the explanatory variable representing the factor to be investigated, e.g., representing the level of cortical occipital alpha activity. The conditional gERD, given a particular state (e.g., low, medium or high level of alpha activity), is defined as(4)In this formula the denominator and the enumerator represent the state conditional reference and event-related dynamics, respectively. Note, the state variable is not necessarily limited to discrete values, such as low, medium and high alpha activity, but can also be continuous valued, e.g., representing the amplitude value itself. For computational aspects of estimating conditional gERD, however, we refer to the Supplementary Methods section in Text S1. Moreover, Matlab code is available at http://bbci.de/supplementary/conditionalERD/.
Remark: The conventional ERD measure as given in Eqn 2 can be extended in an identical fashion by means of conditional expectations values.(5)However, in section “Results” will show that this simple extension of the standard measure yields spurious observations of ERD/S. For a detailed description of the empirical estimators of the state conditional ERD please refer to the Supplementary Methods section in Text S1.
The following applications will serve as a proof of concept of the proposed framework. We will illustrate the potential of the proposed generalized ERD framework for the analysis of state conditional ERD and uncover the limitations of the conventional methods. Initially we conduct a comparative evaluation of both frameworks by means of artificially generated data with known truth. The application in such a controlled, artificial environment will reveal that the conventional ERD can give rise to observations of spurious ERD/S. Afterwards we investigate the state dependencies of the characteristics of somatosensory induced ERD on three local cortical states.
In order to compare the capabilities of the generalized and the conventional conditional ERD framework properly, we generate three sets of surrogate ERD data that exhibit different kinds of dependency on an explanatory variable . To this end, we use two simple models for the power envelope of unperturbed dynamics on the one hand and for the dampening process on the other hand. Moreover, both models will provide the opportunity to control their dependency on the explanatory factor.
The human perirolandic sensorimotor cortices show rhythmic macroscopic EEG/MEG oscillations with spectral peak energies around 10 Hz (localized predominantly over the postcentral somatosensory cortex) and 20 Hz (over the precentral motor cortex) [28]. These so-called exhibit fast inherent fluctuations as they are limited to brief periods (spindles) of 0.5–2 s duration [29], which appear to occur in the absence of overtly processing sensory information or motor commands. ERD/S of the have been reported for different physiological manipulations, e.g., by motor activity, both actual and imagined [3],[30],[31], as well as by somatosensory stimulation [25]. In this context standard trial averages of power typically reveal a sequence of attenuation followed by a rebound which often overshoots the pre-event baseline level [31],[32]. In the following we will present a case study, investigating the impact of three endogenous factors on the characteristics of somatosensory induced ERD.
We presented the novel data analytical framework of gereralized ERD that allows for a reliable analysis of ERD also in the presence of dynamical cortical states. To this end, we started from the observation that the conventional ERD measure can give rise to spurious detection of ERD, when analyzing non-stationary dynamics (Fig. 1-A). We then identified the constant baseline as the limiting factor of the conventional ERD measure. Accordingly, we generalized the conventional ERD framework with respect to the choice of reference. In particular, we substituted the constant baseline by a reference dynamics and derived a novel generalized measure for the quantification of ERD, by defining ERD/S as the relative deviation of the event-related dynamics from this reference dynamics. In particular, we proposed the use of the natural relaxation dynamics of the unperturbed EEG rhythm as a reference. In this context we also discussed how the acquisition of this reference dynamics can be incorporated into the experimental design by means of catch trials. Afterwards, we validated the ability of the generalized ERD measure to afford a reliable quantification of induced spectral perturbations even in the presence of non-stationary dynamics (Fig. 1-B). Moreover, we pointed out that the conventional and the generalized ERD measure yield identical results in case of stationary dynamics. Consequently, due to the lower effort in designing and conducting the experiment as well as in analyzing the data, if stationarity holds for the dynamics under study, then the conventional measure is preferred. However, we also emphasized, that stationarity cannot be assumed for investigations of state conditional dependencies.
Following the introduction of the generalized ERD framework, we extended both, the generalized and the conventional ERD measure in order to afford the quantification of state conditional ERD. Here, the application of a reliable state conditional measure can be used to scrutinize candidate explanatory factors, such as the level of activity of a particular EEG eigenrhythm or the stimulus intensity, with respect to their possible impact on a the oscillatory dynamics under study. As a proof of concept, we compared the respective capabilities of the conventional and the generalized state conditional framework first on simulated and afterwards on real ERD data. Here, in the well controlled scenario of artificially generated data, the comparison of the results of the conventionally estimated with the true analytically obtained state conditional ERD, clearly revealed the limitations of the conventional framework in retrieving the given functional relationship of the ERD on the explanatory variable. Furthermore, the conventional conditional ERD measure gave rise to spurious observations of ERD and ERS which were not even modelled in the artificial data (see Fig. 3). Unlike the conventional method, which failed, the novel generalized measures performed well at retrieving the true underlying functional relationship of the conditional ERD on an explanatory variable from the surrogate data (see Fig. 3). Finally, we illustrated the potential of the proposed novel framework for neurophysiological investigations by analyzing ERD data from a median nerve stimulation paradigm. In particular, we applied the novel estimator of generalized conditional ERD to analyze the impact of three explanatory factors on the inter-trial variability of the contra-lateral mu-rhythm ERD induced by somatosensory stimulations. Specifically, we investigated the impact of the magnitude of local prestimulus mu-rhythm activity, the magnitude of occipital alpha and the magnitude of the ERS response to the preceding stimulus on the ERD magnitude and latency. As a result, we found that the mu amplitude dynamics is a strictly local phenomenon, both in time and in space. Moreover, the application of the gereralized conditional ERD measure revealed that lower occipital alpha, possibly indicative of system-wide increase of arousal, can be linked to a faster mu rhythm ERD at pericentral cortices. Therefore, the proposed framework was able to provide evidence for the existence of a sensible physiological dynamics related to the interaction between ongoing activity and stimulus-induced responses.
In principle, the three given examples represent just a small sample of new possibilities: comparable analyses could be envisioned for the impact of various external factors such as: the inter-stimulus interval (ISI) [22], where short ISI results in stimulus presentation, while the processing of the previous event is still going on; the duration of the experiment, where the effects of fatigue on both, the event-related and the unperturbed dynamics can introduce variability of the ERD response; the simultaneous processing of multiple stimuli that potentially have a masking effect [34]; but also the influence of endogenous factors such as: the phase of a particular EEG eigenrhythm [17]; the synchronization level between adjacent cortical areas [35]; or causal coupling of various brain rhythms [36].
Moreover, recently the interest in inter-trial variability of ERD responses was sparked by the presentation of an alternative mechanism contributing to the generation of evoked responses [37]. In particular, the authors presented theoretical and empirical evidence that the amplitude fluctuations of neuronal alpha oscillations can be associated with changes in the mean value (baseline shift) of ongoing activity. Furthermore, they proved, when stimuli modulate the amplitude of alpha oscillations, these baseline shifts become the basis of a novel mechanism for the generation of evoked responses. Consequently, combining the two kinds of analysis, i.e., the analysis of ERD variability with the interpretation of ERD as a mechanism for the generation of ERP, may result in an additional explanation of inter-trial variability of ERPs.
Another important direct application area is brain-computer interfacing [38]–[40] which could benefit from this generalized conditional ERD framework: here, classifiers that discriminate between, e.g., imaginary left and right hand movements, could possibly yield an improved accuracy when considering state dependent behavior of ERD.
While there are a series of advantages and potentials, the application of the generalized framework comes at the expense of an experimental paradigm which has to comprise both, event-related and catch trials. Additional demands for a reliable estimation of state conditional ERD originate from the greater number of required trials compared to the estimation of unconditional ERD.
Notably, EEG scalp recordings mainly measure excitability fluctuations of superficial cortical layers, with minimal or no information on subcortical relays of the neural network supporting a given rhythm, e.g., an increased thalamic excitability may result in a low amplitude desynchronized cortical EEG [41]. Therefore, a cortical ERD is to be conceived as an electrophysiological index of an activated thalamo-cortical system involved in the processing of sensory or cognitive information or in the production of motor behavior [42]. While the analysis of cortico-subcortical interaction is naturally limited when based on scalp EEG data only, the modelling of inter-trial variability of evoked responses can improve the understanding of cortico-cortical interactions on a macroscopic scale [43] and it is here that the generalized conditional ERD represents a useful tool for such analyses with respect to accompanying ERD/S responses.
|
10.1371/journal.ppat.1004553 | Discovery of Replicating Circular RNAs by RNA-Seq and Computational Algorithms | Replicating circular RNAs are independent plant pathogens known as viroids, or act to modulate the pathogenesis of plant and animal viruses as their satellite RNAs. The rate of discovery of these subviral pathogens was low over the past 40 years because the classical approaches are technical demanding and time-consuming. We previously described an approach for homology-independent discovery of replicating circular RNAs by analysing the total small RNA populations from samples of diseased tissues with a computational program known as progressive filtering of overlapping small RNAs (PFOR). However, PFOR written in PERL language is extremely slow and is unable to discover those subviral pathogens that do not trigger in vivo accumulation of extensively overlapping small RNAs. Moreover, PFOR is yet to identify a new viroid capable of initiating independent infection. Here we report the development of PFOR2 that adopted parallel programming in the C++ language and was 3 to 8 times faster than PFOR. A new computational program was further developed and incorporated into PFOR2 to allow the identification of circular RNAs by deep sequencing of long RNAs instead of small RNAs. PFOR2 analysis of the small RNA libraries from grapevine and apple plants led to the discovery of Grapevine latent viroid (GLVd) and Apple hammerhead viroid-like RNA (AHVd-like RNA), respectively. GLVd was proposed as a new species in the genus Apscaviroid, because it contained the typical structural elements found in this group of viroids and initiated independent infection in grapevine seedlings. AHVd-like RNA encoded a biologically active hammerhead ribozyme in both polarities, and was not specifically associated with any of the viruses found in apple plants. We propose that these computational algorithms have the potential to discover novel circular RNAs in plants, invertebrates and vertebrates regardless of whether they replicate and/or induce the in vivo accumulation of small RNAs.
| Viroids are a unique class of subviral pathogens found in plants, and they are difficult to identify since they are free circular non-coding RNAs and often replicate to low levels in host cells. We previously described the computational algorithm PFOR that discovers viroids by analyzing total small RNAs of the infected plants obtained by next-generation sequencing platforms. However, the algorithm written in PERL language is very slow, and viroid identification depends on the in vivo accumulation of extensively overlapping sets of small RNAs to target viroids. Here we report the development of PFOR2 that adopted parallel programming in the C++ language and was significantly faster than PFOR. We also describe a simple computational program that after incorporation into PFOR2 is capable of identifying viroids from deep sequencing of long RNAs instead of small RNAs. Moreover, we report the identification of Grapevine latent viroid (GLVd) and Apple hammerhead viroid-like RNA by the computational approach. Since our new algorithms do not depend on the analysis of viroid-derived small RNAs produced in vivo, it is possible to discover viroids in a wide range of host species including plants, invertebrates and vertebrates.
| Viroids and a group of satellite RNAs (satRNAs) have single-stranded circular RNA genomes that range in size from 220 to 457 nucleotides (nt) [1]–[4]. These subviral pathogenic RNAs lack protein-coding capabilities and thus depend on either host-encoded DNA-dependent RNA polymerase (viroids) or helper virus-encoded RNA-dependent RNA polymerase (circular satRNAs) for replication [5], [6]. Viroids and circular satRNAs have been proven to be excellent biological models for studying non-coding RNAs (ncRNAs) and basic biology [3], [4]. The most notable examples include the discovery of RNA-directed DNA methylation (RdDM) in viroid-infected plants [7] and of the hammerhead ribozymes in viroids [8] and circular satRNAs [9] of plant viruses. Interestingly, recent studies have revealed the production of thousands of non-replicating circular RNAs (circRNAs) across species from Archaea to humans [10], [11]. These circRNAs are largely generated from back-spliced exons, in which splice junctions are formed by an acceptor splice site at the 5' end of an exon and a donor site at the downstream 3' end [10], [12], [13]. The functions of circRNAs are largely unknown, although a few circRNAs have recently been shown to play regulatory roles as, for example, microRNA sponges [10], [12], [13].
Viroids infect many crops and cause severe symptoms in susceptible hosts that result in economically important diseases [2]. However, the rate of discovery of the replicating circular RNAs is slow compared to the discovery of viruses [14]. For example, fewer than 40 viroid species, all of which infect higher plants, have been identified [15] since the first report in 1971 [1]. The slow rate of viroid discovery is often attributed to the technical difficulty involved in the purification and characterization of the naked non-coding circular RNAs that generally occur at low concentrations in the infected host [16]. We have recently described an approach for sequence homology-independent discovery of replicating circular RNAs by analyzing the total small RNA populations from samples of diseased tissues with a program known as progressive filtering of overlapping small RNAs (PFOR) [17]. The PFOR approach relies on the observations that rolling-circle replication of viroids and some satRNAs produces multimeric head-to-tail dsRNAs [5] and that continuous overlapping sets of small interfering RNAs (siRNAs) processed by Dicer [18]–[21] from the direct repeat dsRNAs accumulate to high levels in infected plant tissues [22], [23]. PFOR retains viroid-specific siRNAs for genome assembly by progressively eliminating non-overlapping small RNAs and those that overlap but cannot be assembled into a direct repeat RNA. Use of PFOR for the analysis of a grapevine small RNA library led to the discovery of a viroid-like circular RNA of 375 nt that encodes active hammerhead ribozymes in both plus and minus polarities [17]. However, it remains unknown whether the identified circular RNA can initiate independent infection.
Two major limitations of the first version of PFOR restrict its application in the discovery of circular RNAs. First, the iterative filtering of small RNAs that are not derived from a replicating circular RNA is a slow process and takes up more than 90% of the PFOR running time. Because PFOR was written in the explanatory PERL language, analyzing complex small RNA libraries using PFOR may take hours or days. Second, circular RNAs are not identifiable by PFOR if they neither replicate nor trigger Dicer-dependent siRNA production in a eukaryotic cell.
In this study, a considerably improved version of PFOR was developed by adopting parallel programming in the C++ language. The use of the new version of PFOR, designated PFOR2, led to the discovery of a new viroid from grapevine and a viroid-like RNA from apple tree. Moreover, a new program was developed and incorporated into PFOR2 for the discovery of distinct classes of circular RNAs, including those that neither replicate nor induce in vivo accumulation of Dicer-dependent siRNAs. We propose that the application of PFOR2 would speed up the discovery of novel circular RNAs and expand the list of known host species that can be independently infected by viroids.
The computational algorithm of PFOR has been developed to identify replicating circular RNAs including viroids by deep sequencing of small RNAs [17]. A key step of PFOR algorithm is to separate terminal small RNAs (TSRs) from internal small RNAs (ISRs) in a small RNA pool. A small RNA is defined as an ISR if it overlaps at least one other small RNA at both ends larger than k-mer in the pool, whereas a TSR overlaps at least one other small RNA larger than k-mer in the pool at only one end of the TSR. The process of PFOR includes two main steps: filtering all non-overlapping small RNAs and terminal small RNAs (TSRs) with overlapping k-mers and assembling the remaining internal small RNAs (ISRs) predicted to derive from circular RNAs (Fig. 1C). Filtering TSRs derived from linear non-repeat precursor RNAs takes up more than 90% of PFOR running time. Therefore, to shorten the computing time required for filtering TSRs and to improve the performance of PFOR, PROR2 was developed by converting the previous algorithm written in the explanatory PERL language into the C++ language and adopting the parallel programming technology of OpenMP [24]. Because multiple shared memory filtering processes were performed concurrently in PFOR2, the TSR filtering process was expected to be faster than PFOR (Fig. 1A).
To test the performance of PFOR2, two publicly available small RNA libraries, in which both known viroids and viroid candidates have been verified by RT-PCR and Northern-blot hybridization, were used to compare the running time between PFOR and PFOR2. Hop stunt viroid (HpSVd), Grapevine yellow speckle viroid (GYSVd) and Grapevine hammerhead viroid-like RNA (GHVd RNA) were each identified, and their full-length genomic RNA sequences were obtained by both PFOR and PFOR2 from the grapevine tree sRNA library, which contains 4,701,135 reads of 18–28 nt in length (GEO accession no. GSM458928). However, PFOR2 required only 67 seconds and was 3.3 times faster than PFOR. The second sRNA library was from a peach tree infected with Peach latent mosaic viroid (PLMVd) and contained 7,862,905 reads of 18–28 nt in length (GEO accession no. GSM465746). Both PFOR and PFOR2 were again able to identify PLMVd and to recover the complete sequence. Instead of 2 hours by PFOR using a default k-mer of 17, PFOR2 required only 22 minutes. These results demonstrated that PFOR2 was indeed faster than PFOR in viroid discovery.
PFOR2 was next applied to determine whether an apple plant with typical symptoms of apple scar skin disease contained new circular RNAs. The sRNA library was constructed from this apple plant, and 15,321,500 clean reads of 18–30 nt in length with a predominant size of 21 nt were obtained after deep sequencing (S1 Figure). Two putative circular RNAs were identified from the apple tree sRNA library by both PFOR and PFOR2, although PFOR2 was six times faster. The first RNA species was 333 nt and shared 96% sequence identity with a variant of Apple scar skin viroid (ASSVd) (accession no. KC110858) isolated previously from apple in China and was hence considered to be a new isolate of ASSVd. The second RNA species was 434 nt in length and showed no sequence similarity to any of the known entries in GenBank. Interestingly, the second predicted circular RNA also contained the conserved sequences found in hammerhead ribozymes as shown previously for GHVd RNA [17]. Thus, the second RNA identified from the apple sRNA library by PFOR and PFOR2 was tentatively designated as apple hammerhead viroid-like RNA (AHVd-like RNA).
To verify the predicted sequence and the circular nature of AHVd-like RNA, total RNAs from the diseased apple were isolated for divergent RT-PCR analysis. According to the sequence of AHVd-like RNA assembled by PFOR2, two pairs of adjacent primers with opposing polarities (AHVd-13F/12R and AHVd-88F/87R, sequences of primers are shown in S1 Table) were designed for RT-PCR so that the full-length AHVd-like sequences would be amplified only when AHVd-like RNA existed in a circular form. We found that RT-PCR analysis of the apple RNA sample with either primer pair yielded a single DNA species with the expected size, demonstrating the circular nature of AHVd-like RNA from the apple tree (Fig. 2A). Moreover, direct DNA sequencing of the RT-PCR products confirmed the sequence of AHVd-like RNA assembled by PFOR2.
We noted that the full-length AHVd-like RNA could be amplified when either of the primer pairs was used in RT reactions, indicating the existence of both plus and minus circular RNA molecules in the infected tissue. To further investigate the in vivo properties of AHVd-like RNAs, nucleic acid preparations were analyzed by denaturing PAGE and Northern-blot hybridization with a probe either corresponding or complementary to the assembled full-length AHVd-like RNA. The characteristic circular and linear forms were detected in the infected tissue but not in the healthy apple plant (Fig. 2B). This result further validated the in vivo circularity of AHVd RNAs. Furthermore, the AHVd-like RNAs with opposing polarities appeared to accumulate at different levels (Fig. 2B). Given that the strand accumulating at a higher level is arbitrarily assigned to be the plus polarity, the sequence of AHVd-like RNA obtained by PFOR2 was designated as the plus strand.
Cloning and sequencing of full-length cDNA clones of viroids would supply relevant information on sequence variability in the natural viroid-like RNA populations. The sequenced cDNA clones of AHVd-like RNA were amplified by the two primer pairs, AHVd-13F/12R and AHVd-88F/87R. Therefore, the putative mutations located at the positions of one pair of primers were determined through amplification and sequencing with the second pair of primers. In total, 14 sequences of AHVd-like RNA were obtained. None of these AHVd-like RNA sequences were 100% identical to other 13 sequences. However, one sequence (clone of 1–12 shown in S2 Figure) was identical to the assembled sequence of AHVd-like RNA by PFOR2. The alignment of these 14 sequences revealed the presence of 36 mutations in the population of AHVd-like RNA. Although a high-fidelity DNA polymerase was used for PCR amplification, we were not able to exclude possible errors introduced during RT-PCR. Thus, after 22 mutations detected only in one clone were removed, the remaining 14 mutations found in at least two clones were tentatively considered to be natural variations (Fig. 3A and S2 Figure). The above analyses showed that the clone of 1–12 represented consensus sequences of AHVd-like RNA, a circular molecule of 434 nt consisting of 114 G (26.3%), 116 C (26.7%), 96 A (22.1%), and 108 U (24.9%) with a G+C content of 53% (Fig. 3A).
AHVd-like RNA did not contain the characteristic central conserved region (CCR) found in the viroid family Pospiviroidae [15]. However, both strands of AHVd-like RNA could be folded into the conserved hammerhead ribozyme structure found in the Avsunviroidae and other small catalytic RNAs [25]. The predicted secondary structure of minimal free energy for AHVd-like RNA was of the quasi-rod-like class of viroids and showed two bifurcations at the right terminal part of the molecule (Fig. 3A), which was similar to that of Eggplant latent viroid (ELVd) [26]. The paired nucleotide residues represented 68.2% of the total, including 56.8% G:C, 35.1% A:U, and 8.1% G:U. Interestingly, 11 out of 14 observed mutations either were mapped in the loop regions or did not affect base pairing (Fig. 3A), which indirectly supported the proposed secondary structure of AHV-like RNA existing in vivo.
Both strands of AHVd-like RNA could form natural hammerhead structures (Fig. 3B) containing 11 strictly conserved residues [27] and the adjacent helices flanking the self-cleavage sites of a group of small catalytic RNAs. In the plus and minus hammerhead structures of AHVd-like RNA, helix III was stable and helices I and II were closed by short loops 1 and 2. These features were similar to the hammerhead structures of (i) ELVd [26], (ii) PLMVd [27], (iii) satellite RNAs of the nepoviruses and sobemoviruses [9], [28], (iv) a cherry small circular RNA (csc RNA1) [29], and (v) GHVd RNA discovered very recently in grapevine [17].
The hammerhead structures of AHVd-like RNA were carefully compared with those of other known viroids and circular satRNAs (namely viroid-like RNAs), revealing some common salient features (Fig. 3B). i) In most natural hammerhead structures, positions 10.1 and 11.1 form a G-C pair, and positions 15.2 and 16.2 form a C-G pair (see ref. [30] for nomenclature). AHVd-like RNA hammerhead structures conformed to this rule. ii) A cytidylate residue preceded the predicted self-cleavage sites of AHVd RNA hammerhead structures, as occurs in most other known hammerhead structures. iii) The residue of position 7 between the conserved CUGA and GA sequences was a U in both RNA hammerhead structures of AHVd-like, which also conformed to the examples observed in most natural hammerhead structures wherein this residue is U, C, and, exceptionally, A. However, the hammerhead structures of AHVd-like RNA exhibited some peculiarities. Both hammerhead structures of AHVd-like RNA shared sequence similarities in the helices and loops with the strictly conserved helix II and loop 1 (Fig. 3B). Sequence similarities included 4 base-pairs of CAGG with CCUG, forming helix II in the consensus hammerhead structure of AHVd-like RNA (Fig. 3B), which corresponded to the equivalent positions in the consensus hammerhead structure of ELVd, the plus strand hammerhead structures of GHVd RNA [17], satellite RNAs of Chicory yellow mottle virus (CYMV) and Tobacco ringspot virus (TRSV) [31]. Moreover, loop 2 of the (+) hammerhead of AHVd-like RNA contained 7 nucleotides and was the largest reported among natural hammerheads. Importantly, the base substitutions found in different AHVd-like RNA variants in the region of the hammerhead structures did not affect the stability of–helix III and no mutations were found in helices I and II (Fig. 3B), suggesting that these self-cleaving domains were functional in vivo.
The activity of the predicted ribozymes encoded by AHVd-like RNA was investigated. Full-length monomeric plus and minus AHVd-like RNA transcripts were synthesized in vitro from linearized plasmids and were found to be self-cleaved during transcription and after purification when incubated under standard self-cleavage conditions in a protein-free buffer (Fig. 4). The cleaved fragments (5′F and 3′F) for the plus and minus strands of the transcripts showed the expected lengths based on the predicted self-cleavage sites of the hammerhead structures (Fig. 4). The predicted cleavage sites (Fig. 3A and 3B) were also experimentally confirmed by rapid amplification of 5′-cDNA ends (5′-RACE)-PCR (S3 Figure). We noted that the plus strand full-length AHVd-like RNA transcripts were more stable during transcription than the minus strand transcripts (Fig. 4B), suggesting a higher self-cleavage efficiency of the minus strand hammerhead ribozyme.
Although the above analyses determined the circularity and self-cleavage activity of AHVd-like RNA, it was still unclear whether AHVd-like RNA represented a new viroid or a circular satRNA. If AHVd-like RNA corresponded to a new plant circular satRNA, it was expected that a helper virus would be present in the diseased tissues to support its replication. To this end, the sRNAs from the diseased apple tree were assembled by Velvet program [32]. BLAST analysis identified contigs that showed sequence similarities with Apple chlorotic leaf spot virus (ACLSV), Apple stem grooving virus (ASGV), and Apple stem pitting virus (ASPV). The presence of these three plant viruses was further confirmed by RT-PCR (S4 Figure). However, we noted that none of these three plant viruses have been reported to have satRNAs. A survey of 182 apple tree samples from different cultivars was performed to determine whether AHVd-like RNA co-existed with any of these viruses. We found that AHVd-like RNA was detected in 75 of the 182 apple tree samples. Notably, the incidence of AHVd-like RNA was not associated with ACLSV, ASPV, or ASGV, suggesting that AHVd-like RNA might be a novel viroid (S2 Table). However, the viroid nature of AHVd-like RNA remained to be verified because neither Northern-blot hybridization nor RT-PCR detected the replication of AHVd-like RNA in the apple seedlings free of AHVd-like RNA one year after mechanical inoculation with the dimeric transcripts synthesized in vitro from the full-length cDNA clones of AHVd-like RNA described above.
Given that the size distribution of sRNAs derived from viroids might serve as an indicator of the subcellular localization or replication sites of the viroids [33], [34], we next compared the accumulation and profile of sRNAs derived from AHVd-like RNA with those of ASSVd in the same tissues of the apple tree. Similar to ASSVd, AHVd-like RNA specific sRNAs from different size families were all divided approximately equally into the plus and minus strands and the predominant sRNA species was the 21 nt class (Fig. 5A). However, few of the AHVd-like RNA-specific sRNAs were 24 nt long. A similar size distribution profile was observed for vd-sRNAs from tissues infected by PLMVd [33], [35] and GHVd [17]. In contrast, a notable amount of ASSVd sRNAs belonged to the 24 nt class (Fig. 5B), similar to several viroids that replicate in the nucleus [34], [36]–[38]. These findings suggest that AHVd-like RNA may not replicate in the nucleus. AHVd-like RNA specific sRNA reads of 21 to 24 nt in length were mapped to the corresponding positions on the AHVd-like RNA genomic or anti-genomic RNAs (Fig. 5C). As previously reported for PLMVd sRNAs isolated from infected peach [30], [32] and GHVd RNA-specific sRNAs from infected grapevine [17], the sRNAs of AHVd-like RNA were derived from every position in both the genomic and anti-genomic strands, and their distribution was biased, with a profile of several hotspots (Fig. 5C).
Grapevine is a natural host for many viroids [39]. Although most of these viroids do not induce symptoms in grapevine, cultivated grapevines with latent viroid infections may serve as reservoirs for certain viroids to infect crops and cause severe diseases [40]. The discovery of a novel viroid-like circular RNA from the original ‘Pinot noir’ grapevine by PFOR [17] suggests that more novel viroids or viroid-like RNAs may exist in cultivated grapevines, especially in some old grapevines. Collections of several grapevine stocks of at least 100 years of age in Xinjiang, China [39] allowed us to test this hypothesis. Of these grapevine trees, a ‘Thompson Seedless’ plant grown in Tulufan was selected for sRNA deep sequencing and viroid discovery by PFOR2.
The obtained sRNA library contained 14,033,487 clean reads of 17–30 nt in length, with 21 nt as the most dominant size class (S5 Figure). PFOR2 analysis of the library took 2 hours and 24 minutes and was 7.1 times faster than PFOR analysis. Complete genomes of four known viroids: HpSVd, GYSVd-1, GYSVd-2, and Australia grapevine viroid (AGVd), which have been previously detected by RT-PCR and Northern-blot hybridization in this old grapevine tree [39], were assembled by PFOR2. PFOR2 analysis of the grapevine sRNA library also identified a putative circular RNA molecule of 328 nt in length that shared 79% sequence similarity with Citrus viroid VI (CVd-VI) (accession no. AB019508). Because CVd-VI had not been previously isolated from grapevine and the sequence similarity between CVd-VI and the identified RNA molecule was below the viroid species demarcation criteria of 90% sequence similarities [15], we hypothesized that the circular RNA revealed by PFOR2 represented a new viroid and was tentatively designated as Grapevine latent viroid (GLVd) hereafter.
To confirm the viroid nature of GLVd, we first determined whether GLVd existed in a circular form in vivo. Two sets of adjacent primers of opposite polarity (GLVd-252F/251R and GLVd-141F/140R, shown in S1 Table), derived from the predicted sequence by PFOR2, were synthesized and used for the amplification of the full-length circular GLVd by RT-PCR. As a control, PCR was performed with these primers using the template of total DNA isolated from the old grapevine without the RT step to determine whether GLVd was derived from repeat elements of the host genome. Divergent RT-PCR with either of the two primer pairs yielded a product with the expected size whereas no specific products were amplified by PCR (Fig. 6D), confirming the circular RNA nature of GLVd. The amplified DNA of the expected size was eluted, and four clones from each primer pair were sequenced. Sequence analysis revealed the presence of a master sequence represented by six clones, while the two sequence variants contained a deletion of A at position 54 and a substitution (G/A) at position 125, respectively (Fig. 6A). Importantly, the master sequence of GLVd was identical to the sequence discovered by PFOR2 and was 328 nt in length, with 67 A (20.4%), 70 U (21.3%), 96 G (29.3%) and 95 C (29%), producing a G+C content of 58.3%.
The availability of the full-length GLVd genomic sequence allowed us to synthesize a GLVd-specific riboprobe for detecting various molecular forms of GLVd RNA by Northern-blot hybridization. Total RNAs extracted from the old grapevine were separated by denaturing PAGE followed by Northern-blot hybridization, leading to the detection of the characteristic circular and linear forms (Fig. 6D). These findings together indicated that GLVd existed as a circular RNA in the old grapevine tree.
The minimal free-energy secondary structural prediction revealed a rod-like conformation of GLVd. The predicted secondary structure of GLVd was similar to that proposed for most viroids [15], [41] and contained 63.4% paired nucleotides, including 67.3%, 24.0% and 8.7% of G:C, A:U, and G:U pairs, respectively (Fig. 6A). Notably, the GLVd structure contained the central conserved region (CCR), which is the key structural element and taxonomic criterion for assigning viroids to the family Pospiviroidae. The sequences of upper and lower CCR of GLVd were nearly identical to that of Apple scar skin viroid (ASSVd) [42], the type species of the genus Apscaviroid. The terminal conserved region (TCR) of GLVd was also similar to that found in apscaviroids (Fig. 6A and B). Furthermore, the GLVd structure included a polypurine stretch located in the pathogenicity domain, which is conserved in the family Pospiviroidae [41].
Hairpin I (HPI) formed by the upper CCR strand and the flanking inverted repeat [43], [44] is a conserved structural element in the family Pospiviroidae. A typical HPI was identified in GLVd and included the capping palindromic tetraloop, the adjacent 3-bp stem, and the 7-bp stem interrupted by two opposite-facing nucleotides that were seemingly unpaired (Fig. 6C). However, sequence variations were noted in the HPI of GLVd compared to the known apscaviroids (Fig. 6C and S6 Figure). The nucleotide substitution of U by C at the left terminus of the upper CCR converted a G:U base-pair in the stem of HPI into a G:C base-pair, which was predicted to strengthen the stability of this structure. In contrast, a large internal loop present in GLVd HPI would weaken the stability of HPI (Fig. 6C). Detection of the conserved structural features of viroids such as CCR, TCR, and HPI in GLVd further supports the idea that GLVd is a viroid.
To further verify the viroid nature of GLVd, dimeric head-to-tail transcripts of GLVd were transcribed in vitro from the constructed cDNA clones of GLVd. Virus-free grapevine seedlings (cv ‘Beta’) grown in early spring were mechanically inoculated with the GLVd transcripts by slashing the stems with razor blades. Uninoculated healthy seedlings from the same batch were kept as controls. Because GLVd was undetectable by Northern-blot hybridization 3 and 6 months post inoculation, we re-inoculated the seedlings with a higher dose of GLVd transcripts and detected weak hybridization signals from 4 of the 18 inoculated grapevine plants 3 months after the secondary inoculation. To facilitate GLVd detection in the young tissues, the apical shoots of the inoculated grapevine plants were removed, and the leaves from the young lateral branches that emerged 6 months after the secondary inoculation (or 12 months after the first inoculation) were collected for RNA extraction. We found that GLVd infection became readily detectable in 6 of the 18 inoculated grapevine plants using either Northern-blot hybridization or RT-PCR. The progeny sequence was determined via DNA sequencing of the cloned RT-PCR products and was found to be the same as the inoculated GLVd transcripts. Therefore, GLVd autonomously replicated in its natural host grapevine, fulfilling the most critical criterion to be considered as a viroid.
To determine the taxonomy of GLVd, the sequence of GLVd was aligned with all of the known species in the genus Apscaviroid. The phylogenetic tree constructed using viroids of genus Colviroid as the out-group revealed two subgroups of apscaviroids (Fig. 7A). GLVd was clustered in subgroup-II and most closely related to CVd-VI and a tentative new species of Persimmon viroid 2 (PVd-2) identified very recently from American persimmon (Diospyros virginiana L.) [45] (Fig. 7A). These results indicated that GLVd should be considered as a new member in the genus Apscaviroid. Interestingly, careful inspection of the apscaviroid alignments identified two types of repeat sequences between GLVd and PVd-2 (Fig. 7B). Further study is necessary to determine whether the repeat sequences were involved in host adaptation because simple sequence repeats (SSRs) distribute non-randomly in viroid genomes and might play a significant role in the evolution of viroids [46].
We next developed a simple computational program, Splitting Longer reads into Shorter fragments (SLS), as part of PFOR2 to discover biologically active circular RNAs via the deep sequencing of long RNAs instead of small RNAs. The program cut the sequenced long RNAs into 21-nt virtual sRNAs of 20-nt overlap with their 5′ and 3′ neighboring sRNAs before PFOR2 analysis to retain only 21-nt virtual ISRs for the final assembly of circular RNAs (Fig. 1B). To determine the efficacy of SLS-PFOR2, we sequenced the total RNAs from PSTVd-infected potato seedlings by constructing independent libraries using Not Not So Random (NNSR) library protocol [47] after either depletion of the abundant ribosomal RNAs [48], [49] or enrichment for circular RNAs following the degradation of linear RNAs by RNase R [50] (S7 Figure). The sequencing of the rRNA-depleted library yielded 774,621 reads longer than 100 nt, among which 83 reads were derived from PSTVd with a mean length of 160 nt. A total of 92,093 reads longer than 100 nt were obtained from the RNase R-treated library, with 55 reads from PSTVd. We found that the full-length PSTVd molecule of 354 nt was readily identified by SLS-PFOR2 from both the rRNA-depleted library (k-mer 19 or 20) and the RNase R library (k-mer 17 or 18) with a running time of 3 hours 20 minutes and 103 hours 14 minutes, respectively. These results demonstrated that SLS-PFOR2 is capable of discovering circular RNAs independently of the in vivo production or the deep sequencing of their specific small RNAs.
Next-generation sequencing (NGS) approaches can readily identify viral and subviral pathogens in samples of plant and animal diseased tissues that are related in nucleotide sequence or encoded protein sequence to a known pathogen. The development of PFOR for viroid discovery thus represents a conceptual advance because, unlike NGS and several available classical approaches, PFOR does not depend on sequence homology with a known viroid. The major improvements described in this study overcame the limitations of the published version of PFOR that restrict its potentially widespread applications in pathogen discovery. PFOR2 was 3.3, 5.4, and 7.1 times faster than PFOR in the analysis of the three small RNA datasets from grapevine, peach, and apple, respectively. The enhanced speed is likely to be critical for viroid discovery when targeting hosts with large genome sizes and/or abundant small RNA populations. For example, our analysis of a small RNA library from Areca catechu with 46,637,488 reads took 8 hours and 40 minutes by PFOR2 instead of 110 hours by PFOR (unpublished data). The efficacy of PFOR2 was verified with the discovery of GLVd as a novel viroid that initiates independent infection in its natural host. Moreover, the development of SLS-PFOR2 eliminates the requirement for the in vivo production and accumulation of Dicer-dependent siRNAs to target the circular RNAs to be identified. As a result, small RNA sequencing becomes unnecessary, and RNA-seq libraries depleted of either ribosomal RNAs or linear RNAs can be analyzed by SLS-PFOR2 for the discovery of both replicating and non-replicating circular RNAs in diverse organisms. In principle, SLS-PFOR2 can identify novel viroid circular RNAs in host species that replicate but do not trigger the biogenesis of viroid-specific siRNAs. Therefore, SLS-PFOR2 has the potential to expand the list of both viroids and host species that support viroid infection.
PFOR and FOR2 separate small RNAs in the pool into TSR and ISR groups based on the presence of the minimal overlapping k-mer among reads and remove all TSRs progressively. As a result, the overlapping sets of small RNAs retained after the filtering process might be different when different k-mers are used, leading to the variations in the sequences assembled by PFOR and PFOR2 that may not reflect the natural heterogeneity of viroids. The successful detection of each viroid by PFOR and PFOR2 is dependent on whether the circular RNA has been completely covered by a set of overlapping small RNAs with the minimal length defined by k-mers at both ends after removing all TSRs. Because each ISR is allowed to be used only once during the assembling step, only one viroid would be revealed when two or more viroids share small RNAs with lengths defined by k-mers or longer. For example, although ASSVd was revealed by PFOR2 analysis of the apple library using k-mers in the size range of 18 to 20 nt, AHVd-like RNA was identified using k-mers of 18 or 19, but not of 20. In the grapevine sample, HpSVd was identified by PFOR2 using k-mers of 17 to 21, whereas GYSVd-1 and GYSVd-2 were each revealed with a specific k-mer most likely because the two viroids are 80% identical in sequence and share small RNAs. Therefore, it is necessary to analyze each small RNA library using PFOR2 with k-mers from 17 to 21 and to verify the assembled sequences of viroid candidates by RT-PCR and cDNA sequencing.
The evolutionary origin of viroids remains unknown [51]. However, it has been proposed that most, if not all, present viroid diseases of cultivated plants originated recently by the accidental introduction of viroids from endemically infected wild plants into susceptible cultivated plants [52]. Thus, the identification of the original wild host plants as symptomless viroid carriers may provide additional insight into possible evolutionary scenarios. Cultivated grapevines were assumed to be associated with ‘Etrog’ citron fruit, displaying citrus viroid disease symptoms as depicted in an ancient synagogue from the early 6th century C.E. in Israel [40]. This suggested that cultivated grapevines with latent infections of viroids may serve as reservoirs for viroid spreading and causing diseases in other hosts. Accordingly, the viroids that cause some epidemic diseases at present are likely to come from the originally infected grapevines. This hypothesis is consistent with the finding that the cultivated grapevines asymptomatically infected with HpSVd were considered as the origin of the hop stunt disease epidemic in commercial hops in Japan [53]. It is also possible that grapevines might harbor some unknown viroids that are yet to be identified. The discovery of a novel viroid-like circular GHVd RNA previously [17] and GLVd in this work supports this idea. GLVd is related to both CVd-VI and PVd-2, which were isolated from diseased Etrog citrons (Citrus medica L.) with mild petiole necrosis and leaf bending [54] and American persimmon [45], respectively. Since GLVd and PVd-2 appear to originate from a common ancestor, it will be of interest to determine in future studies if the two repeated sequences detected between GLVd and PVd-2 (Fig. 7B) play a role in host adaptation during transmission from its original host to certain new susceptible hosts.
Our conclusion that GLVd is a novel viroid is supported by the molecular and biological evidence presented here including its circularity, typical structural elements of viroids, and self-replication in grapevine seedlings. The phylogenetic analysis indicates that GLVd is a member of the genus Apscaviroid. Although we found that GLVd was able to independently infect grapevine seedlings (cv ‘Beta’), the accumulation of GLVd was low, and no obvious symptoms were observed in infected grapevine plants. Although future studies on biological properties of GLVd may further differentiate this viroid from those previously reported, the conserved structural elements, the low sequence identity (maximum of 79% with CVd-IV) with other members in the genus Apscaviroid, and the natural host of GLVd, strongly support the possibility of annotating it as a new species in the genus Apscaviroid.
It is currently unclear whether AHVd-like RNA is a viroid or a satellite RNA, in contrast to GLVd. AHVd-like RNA shared no homology with the apple genome and was not amplified by PCR without a RT step, indicating that AHVd-like RNA was exogenous. Given that AHVd-like RNA encoded self-cleavage activities and was not specifically associated with any of the viruses identified in apple trees, we propose that AHVd-like RNA is a viroid. However, we were unable to demonstrate independent infectivity in apple seedlings for either the in vitro transcripts from dimeric AHVd-like RNA cDNA clones or AHVd-like RNA purified from apple tissues. In this regard, AHVd-like RNA may be related to ASSARNA-2, a circular RNA that was previously isolated from diseased apple plants in Japan and China, known to migrate more slowly than the 330-nt ASSVd RNA and unable to establish independent infection in apple seedlings [55]–[57]. Furthermore, our search for the conserved tertiary structure of a kissing loop, which is found in most Avsunviroidae viroids [58], [59] and in GHVd RNA [17], identified only weak kissing loops of 3 base-pairs in AHVd-like RNA (S8 Figure). Therefore, we cannot rule out the possibility that AHVd-like RNA is a novel satRNA. However, we note that virus-derived siRNAs produced by the antiviral Dicer of a fungal host are predominantly within the 20- to 22-nt range with a peak at 21-nt [60]. It is therefore less likely that AHV-like RNA replicates and triggers Dicer recognition in a fungal host since 21- and 22-nt small RNAs derived from AHVd-like RNA were clearly more abundant than 20-nt and the remaining size classes of small RNAs as found for plant viral siRNAs produced hierarchically by Dicer-like 4 (DCL4) and DCL2 [61].
For the initial identification of viroids and viroid-like RNAs from apple, leaves were collected from an apple (Malus pumila Mill. cv. Fuji) plant, the fruits of which showed typical symptoms of apple scar skin viroid disease, in Shandong province China, in July 2012. The grapevine (Vitis vinifera L.) leaf samples used for determination of GLVd were from Tulufan in Xinjiang China. This grapevine plant (cv. Thompson seedless) for sample collections is more than 100 years old. Young leaves of both apple and grapevine were immediately put into RNAlater stabilization solution (Ambion, USA) after collection and sent to a laboratory for deep sequencing analysis. Moreover, approximately 10 g of apple and grapevine leaves were packaged with ice, kept fresh at low temperature, and sent to a laboratory for RNA analysis using PAGE and northern-blot hybridization.
To survey the occurrence of viroid-like apple RNA in China, 182 leaf samples of variant apple cultivars from five provinces were collected from 2012 to 2014 and kept at −80°C.
Total RNAs used for deep sequencing analysis were extracted by TRIzol reagent (Invitrogen) following the manufacturer's instructions. The integrity of the resulting RNA preparations was evaluated before preparing cDNA libraries using an Agilent Technologies 2100 bioanalyzer. For RNA analysis by PAGE and northern-blot hybridization, nucleic acid preparations were obtained with buffer-saturated phenol followed by ethanol precipitation, as reported previously [62]. Methoxyethanol and CTAB were used to remove polysaccharides during purification [62].
To prepare templates of RT-PCR amplification performed for cloning full-length sequence of viroid-like apple RNA, the obtained crude extracts were run on a non-denaturing 5% polyacrylamide gel stained with ethidium bromide, and the region of the gel delimited by the 250-bp and 400-bp DNA markers was excised and eluted as described previously [26].
In the experiment involving the RNA-seq of the potato samples, the extracted total RNAs by TRIzol reagent were purified by depleting ribosomal RNAs and non-circular RNAs. RNA species smaller than 200 nt, such as 5S ribosomal RNA, were first removed using the RNeasy Mini Kit (Qiagen, USA), and 28S and 18S ribosomal RNAs were depleted by hybridization with specific probes following the instructions for the RiboMinus Plant Kit (Invitrogen, USA). To enrich circular RNAs, the total RNAs from the same sample were digested with RNase R (Epicentre, USA) at 37°C for 90 min to remove non-circular RNAs.
The RNA extracts were separated using two-dimensional PAGE (2D-PAGE) under non-denaturing and denaturing conditions and stained with ethidium bromide, as previously described [63]. To determine the circularity of RNAs, the total RNAs were run on denatured PAGE gel containing 8 M urea and then transferred to Hybond N+ nylon membranes by upward capillary transfer in 20×SSC buffer. The hybridization was performed at 68°C overnight by specific probes that were generated by a DIG RNA labeling kit (Roche) according to the manufacturer's instructions. The immunological detection was performed by adding chemiluminescent substrate to the membrane following the manufacturer's instructions.
The small RNA libraries were constructed using Illumina's small RNA sample preparation Kit (Invitrogen, USA) following Illumina's method. The protocols for sRNA purification, adaptor ligation, RT-PCR amplification, library purification and high-throughput DNA sequencing on an Illumina HiSeq-2000 have been reported previously [64], [65]. Two sRNA libraries of an old grapevine plant and an apple tree were sequenced. Raw data from the Illumina platform were first processed to trim adaptor and barcode sequences. Reads of 18-30 nt were extracted from the obtained trimmed reads to generate sRNA libraries for assembly. The sRNA library of a peach tree infected with PLMVd (accession no. GSM465746) and the small RNA library from a grapevine tree cultivar Pinot noir ENTAV115 (accession no. GSE18405) were downloaded from the NCBI Gene Expression Omnibus (GEO) database. All of the prepared sRNA libraries were fed into an in-house pipeline. Briefly, exogenous sRNA was enriched by subtracting sRNA derived from the host genome using the Bowtie2 with default parameters [66]. The highly enriched exogenous siRNA from each sample were assembled de novo using Velvet [32] and PFOR [17]/PFOR2. The resulting contigs were queried against the GenBank nt and nr databases using the BLAST program [67].
The RNA-seq libraries of potato samples were constructed with a modified Not Not So Random (NNSR) sequencing method [47]. Two libraries of potato were sequenced using an Ion-torrent sequencer according to the manufacturer's instructions.
PFOR in the PERL language was converted to PFOR2 initially by using the C++ language. OpenMP is an Application Programming Interface (API) that supports multi-platform shared memory multiprocessing programming [24]. The parallel programming technology OpenMP was employed by PFOR2 to parallelize the filtering process of singletons and TSRs concurrently. Vector was also used in PFOR2 to store all sequences temporarily in the filtering process to simplify OPENMP parallelization. In PFOR, a two-level hash table was built at each iteration process to store all sequences in the pool, whereas in PFOR2, a two-level hash table was only established at the first iteration, and non-ISRs were deleted from the two-level hash table for each subsequent iteration.
The SLS (Splitting Longer read into Shorter fragments) program was developed to cut longer reads into virtual sRNAs. The final number of generated virtual sRNAs was dependent on two metrics: sRNA size and overlap size between neighboring sRNAs (step size). Typically, a longer read was cut into contiguous 21-nt sRNAs covering the whole read, in which each sRNA overlapped 20 nt with its 5′ and 3′ neighboring sRNAs (step size = 1).
The primers of HpSVd, GYSVd-1, GYSVd-2 and AGVd were described previously [39]. The primer sets for amplification of GLVd and viroid-like apple RNA were designed from the corresponding sequences of contigs assembled by PFOR2 and were listed in Supplemental Table 1. The first-strands of cDNAs were synthesized with Mu-MLV reverse transcriptase (Promega) at 42°C for 1 h, and PCR amplification was performed by high-fidelity pfu DNA polymerase (Thermo, USA) to generate full-length sequences of viroids and viroid-like RNA. The products of RT-PCR amplification were ligated with additional adenine (A) at the end using Taq DNA polymerase (Takara, Dalian) and cloned into pGEM-T vectors (Promega) with protruding 3′-terminal thymine (T). The recombinant plasmids were amplified by transforming DH5α Escherichia coli cells, and positive clones were randomly selected for sequencing.
The sequenced recombinant plasmids containing full-length of AHVd-like cDNA amplified with primers of AHVd-88F and AHVd-87R were digested with Nco I or Sal I to generate linear plasmids. RNA transcripts in both orientations were synthesized by T7 and SP6 RNA polymerase as described previously [29], [68]. The products of in-vitro transcription were purified by RNeasy Mini Kit (Qiagen, USA). The purified transcripts were incubated at 37°C for 1 h and then separated by 5% denaturing PAGE containing 8 M urea and visualized by ethidium bromide staining. Full-length of AHVd transcripts and the longer fragments resulting from their in vitro self-cleavage were excised from the gels and eluted, separately. The ribozyme activities of the purified transcripts were assessed according to previously described methods [29], [68]. The purified longer fragments were used to validate the self-cleavage sites of AHVd-like RNA by 5′RACE amplification, which was conducted using the 5′RACE System for Rapid Amplification of cDNA Ends kit (Invitrogen).
Head-to-tail dimmers of the entire sequence of GLVd and AHVd-like RNA were prepared by ligation of unit-length inserts and cloning into pGEM-T vectors (Promega), as described previously [69]. The orientation of the inserts of dimeric cDNAs was validated by sequencing. The resulting recombinant plasmids were digested into linear forms and used to synthesize dimeric transcripts with positive polarity by T7 RNA polymerase (Promega). The dimeric transcripts of GLVd and AHVd-like RNA were mechanically inoculated into grapevine (cv ‘Beta’) and apple virus-free seedlings (cv ‘Fuji’), respectively, by slashing the stems with razor blades. Each seedling was inoculated with at least 500 ng of dimeric transcripts. The inoculated seedlings were grown in a common greenhouse. The infectivity of the infectious clones of GLVd and AHVd-like RNA was examined by northern-blot hybridization every three months.
The secondary structures with minimum free energy for GLVd and AHVd-like RNA were predicted by the circular version of the MFold program [70]. The obtained secondary structures were further edited for print by RnaViz 2 [71]. To search for possible kissing-loops in AHVd-like RNA, the Kinefold web server [72] was used with default parameters.
|
10.1371/journal.pntd.0001337 | Proteomic Analysis of Human Skin Treated with Larval Schistosome Peptidases Reveals Distinct Invasion Strategies among Species of Blood Flukes | Skin invasion is the initial step in infection of the human host by schistosome blood flukes. Schistosome larvae have the remarkable ability to overcome the physical and biochemical barriers present in skin in the absence of any mechanical trauma. While a serine peptidase with activity against insoluble elastin appears to be essential for this process in one species of schistosomes, Schistosoma mansoni, it is unknown whether other schistosome species use the same peptidase to facilitate entry into their hosts.
Recent genome sequencing projects, together with a number of biochemical studies, identified alternative peptidases that Schistosoma japonicum or Trichobilharzia regenti could use to facilitate migration through skin. In this study, we used comparative proteomic analysis of human skin treated with purified cercarial elastase, the known invasive peptidase of S. mansoni, or S. mansoni cathespin B2, a close homolog of the putative invasive peptidase of S. japonicum, to identify substrates of either peptidase. Select skin proteins were then confirmed as substrates by in vitro digestion assays.
This study demonstrates that an S. mansoni ortholog of the candidate invasive peptidase of S. japonicum and T. regenti, cathepsin B2, is capable of efficiently cleaving many of the same host skin substrates as the invasive serine peptidase of S. mansoni, cercarial elastase. At the same time, identification of unique substrates and the broader species specificity of cathepsin B2 suggest that the cercarial elastase gene family amplified as an adaptation of schistosomes to human hosts.
| Schistosome parasites are a major cause of disease in the developing world, but the mechanism by which these parasites first infect their host has been studied at the molecular level only for S. mansoni. In this paper, we have mined recent genome annotations of S. mansoni and S. japonicum, a zoonotic schistosome species, to identify differential expansion of peptidase gene families that may be involved in parasite invasion and subsequent migration through skin. Having identified a serine peptidase gene family in S. mansoni and a cysteine peptidase gene family in S. japonicum, we then used a comparative proteomic approach to identify potential substrates of representative members of both classes of enzymes from S. mansoni in human skin. The results of this study suggest that while these species evolved to use different classes of peptidases in host invasion, both are capable of cleaving components of the epidermis and dermal extracellular matrix, as well as proteins involved in the host immune response against the migrating parasite.
| Human skin is a formidable barrier for much of the microbial world. In addition to the mechanical barrier of structural proteins in the epidermis, basement membrane and dermal extracellular matrix, both the epidermis and dermis are bathed in plasma proteins, including early sentinels of the immune system [1]. In order to successfully breach this barrier, an invading pathogen must degrade protein matrices while minimizing the immune response that it elicits. To this end, many invading organisms utilize insect bites or other mechanical trauma to facilitate their entry into skin, but the multi-cellular larvae of the schistosome blood fluke—the causative agent of the disease schistosomiasis—have the remarkable ability to directly penetrate host skin and gain access to dermal blood vessels [2], [3].
The invasive larva(e)—termed cercaria(e)—is 300 µm long, 70 µm wide and comprised of roughly 1000 cells [4]. Upon direct contact with the surface of human skin, cercariae begin to secrete vesicles containing a variety of proteins and an adhesive, mucin-like substance [5]. Proteomic studies identified the majority of proteins secreted by S. mansoni cercariae. These include histolytic peptidases [6], [7]. The most abundant peptidase in S. mansoni secretions is an S1A serine peptidase, termed cercarial elastase (SmCE) (GenBank: AAC46967.1) that has activity against insoluble elastin and other fibrillar macromolecules of skin [8]. Biochemical and immunolocalization studies have confirmed SmCE activity in cercarial secretions [9], [10]. Moreover, applying an irreversible serine peptidase inhibitor to ex vivo skin before exposure to cercariae blocks the majority of larvae from invading, suggesting that this serine peptidase has an essential role in skin penetration [11].
While the serine peptidase, cercarial elastase, plays a key role in S. mansoni skin invasion, the zoonotic species S. japonicum has no serine peptidases in its larval secretions. S. japonicum, however, encodes a number of isoforms of cathepsin B2 (SjCB2) (GenBank: CAA50305.1), a cysteine peptidase, which are secreted by the invading parasite [12]. Moreover, orthologs of SjCB2 have been identified in the cercarial secretions of other, non-human schistosome species, including members of the genus Trichobilharzia [13]. This led us to the hypothesis that the primary invasive peptidase differs between schistosome species, with S. mansoni, a human-specific schistosome species, utilizing cercarial elastase, and S. japonicum, a zoonotic schistosome species, utilizing cathepsin B2. Given that T. regenti also appears to utilize cathepsin B2 for skin invasion, these observations suggest that the use of a serine peptidase in invasion is the exception, not the rule, among parasitic schistosomes. The use of cercarial elastase may reflect unique properties required by S. mansoni to preferentially infect human hosts.
To confirm that cathepsin B2 is also capable of facilitating skin invasion, we used a proteomic approach to identify potential substrates in host skin, for both S. mansoni cercarial elastase and S. mansoni cathepsin B2 (SmCB2) (GenBank: CAC85211.2), a close homolog of S. japonicum CB2. Although RNAi has been developed as a tool in juvenile and adult schistosome worms, it is currently unavailable for the intramolluscan and cercarial stages of development [14]. We therefore chose to use a proteomic approach to validate the roles of these peptidases in skin invasion.
We found that the vast majority of cleaved proteins resulting from human skin exposure to either purified SmCE or SmCB2 overlap, suggesting that both enzymes are capable of facilitating parasite migration through skin. However, we also identified several potential substrates in skin that appear to be cleaved by only one of the two enzymes. Candidate substrates were further validated by in vitro cleavage of purified human skin proteins with either peptidase. Together, these observations suggest that more than one mechanism of skin penetration may have evolved as an adaptation specific to the schistosome-host relationship.
To determine the number of cercarial elastase and cathepsin B2 protein isoforms in schistosome species, all full-length protein sequences (i.e., those possessing the full catalytic core of the peptidase) were collected from both GenBank (NCBI) and S. japonicum and S. mansoni genome annotation websites (Sanger Institute GeneDB). ClustalW (DNA Databank of Japan), was then used to perform multiple sequence alignments and to construct phylogenetic trees. A Blosum protein weight matrix was used to score the alignment, with a gap open penalty of 10, a gap extension penalty of 0.20, and gap distance penalty of 5. Bootstrapping values were calculated using the p-distance method, with a count of 100. The resulting phylogenetic tree was visualized with the program Dendroscope.
S. mansoni cercariae were shed from Biomphalaria glabrata using a light induction method as previously described [11]. SmCE activity was purified from lysate as previously described with the following modifications [15]. Cercariae shed from approximately 50 snails were pelleted by centrifugation at 100 rcf for 1 minute and stored at −20°C. One milliliter of pelleted cercariae was resuspended in 5 ml 300 mM sodium acetate, pH 6.5, 0.1% Triton X-100, 0.1% Tween-20, 0.05% NP40, and sonicated for 1 minute at 40% output. Soluble protein was harvested by centrifugation for 15 minutes at 7, 500 rcf, followed by 0.2 µ filtration. Fractions were again measured for SmCE activity against AAPF-pNA (Ala-Ala-Pro-Phe-p-nitroanilide), and active fractions were run on 10% bis-TRIS polyacrylamide gels (Invitrogen, Carlsbad, CA) according to the manufacturer's specifications, and silver stained [16]. For confirmation of protein identification, bands corresponding to the correct molecular weight of SmCE were excised from the gel, and subjected to in-gel trypsin digestion, followed by LC-MS/MS peptide sequencing, described below. Active site titration was performed using the synthetic peptide inhibitor AAPF-CMK (Ala-Ala-Pro-Phe-chloromethylketone).
Recombinant SmCB2 was expressed in Pichia pastoris as previously described [17].
Media containing secreted protein underwent 0.2 µ filtration and lyophilization. SmCB2 activity was purified as previously described [17]. Fractions were monitored for SmCB2 activity against 5 µM ZFR-AMC (Z-Phe-Arg-7-amino-4-carbamoylmethylcoumarin) in citrate-phosphate buffer, pH 5.3 supplemented with 4 mM DTT. Enzyme concentration was measured by active site titration using the cysteine peptidase inhibitor CAO74 (N-(L-3-trans-propylcarbamoyloxirane-2- carbonyl)-L-isoleucyl-L-proline).
The human skin sample was taken in compliance with protocols approved by the Committee on Human Research at the University of California, San Francisco. Written informed consent was obtained for the operation and use of tissues removed.
Excised human skin was stored at −80°C. For digestion experiments, skin was thawed, dissected into eight 150–170 mg sections, and placed in 1.5 ml microfuge tubes. To each of these skin sections 100 µl of digestion solution containing either peptidase or inhibited peptidase at 1.8 µM was added, along with corresponding controls. SmCE reaction buffer consisted of 100 mM TRIS-HCl, pH 8; SmCB2 reaction buffer consisted of 100 mM sodium acetate, pH 5.5, 4 mM DTT. Inhibited SmCE was prepared by incubating 1.8 µM SmCE with 2 µM AAPF-CMK for one hour at room temperature; inhibition was monitored against AAPF-pNA, prior to its addition to skin. Similarly, inhibited SmCB2 digestion solution was prepared by incubating 1.8 µM SmCB2 with 2 µM CAO74 for one hour at room temperature, with full inhibition monitored by activity against ZFR-AMC, prior to its addition to skin. Inhibitor alone digestion solutions were prepared to control for human skin peptidase activity using either 2 µM AAPF-CMK in 100 mM Tris, pH 8.0, or 2 µM CAO74 in 100 mM sodium acetate, pH 5.5, 4 mM DTT. After addition of digestion solution to skin samples, the reaction mix was vortexed briefly, and then incubated for 5 hours at 37°C. Following incubation, reactions were centrifuged for 20 minutes at 16,000 rcf at 4°C, and the resulting supernatant was saved as the soluble fraction. Fifteen microliters were removed for analysis on a bis-TRIS 4-20% acrylamide gel. Gels were silver-stained and stored at 4°C.
Proteomic analysis of skin digestion samples was performed by LC-MS/MS on two independent preparations as follows. Representative preparative gels are shown in Figures S1 and S2, and contain replicate lanes of approximately 20 µg total protein for each of the skin digestion solutions. Each pair of sample lanes was cut into ten protein bands, and diced into 1–2 mm cubes, then subjected to in-gel trypsin digestion, following a previously published protocol [6]. The resulting peptides were extracted and analyzed by on-line liquid chromatography/mass spectrometry using an Eksigent nanoflow pump and a Famos autosampler that were coupled to a quadrupole-orthogonal-acceleration-time-of-flight hybrid mass spectrometer (QStar Pulsar or QStar Elite, Applied Biosystems, Foster City, CA). Peptides were fractionated on a reversed-phase column (C18, 0.75×150 mm) and a 5–50% B gradient was developed in 35 min at a 350 nl/min flow rate. Solvent A was 0.1% formic acid in water, solvent B was 0.1% formic acid in acetonitrile. Data were acquired in information-dependent acquisition mode: 1 sec MS surveys were followed by 3 sec CID experiments on computer-selected multiply charged precursor ions. Peak lists were generated using Analyst 2.0 software (Applied Biosystems) with the Mascot script 1.6b20 (Matrix Science, London, UK).
Database searches were performed using ProteinProspector v. 5.7.1 (http://prospector2.ucsf.edu) [18]. Searches were performed using the SwissProt databank (August 10, 2010, 519,348 entries). For false discovery rate estimation, this database was concatenated with randomized sequences generated from the original database [19]. Search parameters included selecting trypsin as the digestion enzyme, allowing one missed cleavage but no non-specific cleavages. Peptide modifications that were searched included carbamidomethyl (Cys) as the only fixed modification, and up to two variable modifications from among the following: oxidation (Met), acetyl (N-term), oxidized acetyl (N-term), pyroglutamate (Gln), Met-loss (N-term), and Met-loss+acetyl (N-term). Mass accuracy settings were 200 ppm for precursor and 300 ppm for fragment masses. Data reported in Table S3 has a Protein Prospector minimum score cutoff of 22 (protein), 15 (peptide) and maximum expectation values of 0.01 (protein) and 0.05 (peptide), resulting in a 2% false discovery rate.
Lyophilized type I human skin collagen (Calbiochem) was resuspended in 17.5 mM acetic acid for a final concentration of 1 mg/ml. Human complement C3 (Calbiochem) was purchased as a 1.2 mg/ml stock. For SmCB2 digestion, 180 nM enzyme was added to 50 µl collagen I or 25 µl complement C3 in 50 mM sodium acetate, pH 5.5, 4 mM DTT and incubated at 37°C for 1–22 hours. For SmCE digestion, 180 nM enzyme was added to 50 µl collagen I or 25 µl Complement C3 in 50 mM Tris, pH 8.0 and incubated at 37°C for 1–22 hours. Both enzymes were also pre-incubated with 1 mM CAO74 (SmCB2) or 1 mM AAPF-CMK (SmCE) for one hour at room temperature prior to their addition to collagen. As a control, collagen was incubated in 50 mM sodium acetate, pH 5.5, 4 mM DTT or 50 mM Tris, pH 8.0 for 22 hours at 37°C. To stop the reaction, 15 µl reduced SDS-PAGE loading dye (Invitrogen) was added, and a sample of each reaction was run on a 4–20% Tris-Glycine SDS PAGE gel (Invitrogen). Bands were then electroblotted onto PVDF membrane (Biorad, Foster City, CA) and visualized by Coomassie Blue staining. N-terminal sequence of selected bands was determined using Edman chemistry on an Applied Biosystems Procise liquid-pulse protein sequenator at the Protein and Nucleotide Facility, Stanford University.
To outline the molecular evolution of larval peptidases in schistosomes, all previously reported orthologs were re-examined (Figure 1). In addition to the previously identified full-length cercarial elastase isoforms in S. mansoni--SmCE1a (GenBank: AAM43939.1), SmCE1b (GenBank: CAA94312.1), SmCE1c (GenBank: AAC46968.1), SmCE2a (AAM43941.1) and SmCE2b (GenBank: AAM43942.1) and Schistosoma haematobium cercarial elastase (GenBank: AAM4394)--sequencing and annotation of the full S. mansoni genome revealed three additional full-length genes [15], [20] (Figure 1A). In marked contrast, the S. japonicum genome contains only a single cercarial elastase isoform (Sjp_0028090). No cercarial elastase genes have been detected in any Trichobilharzia species.
Both S. mansoni and S. japonicum encode a number of cathepsin B genes (Figure S3). We chose to focus on the cathepsin B2 isotype, since a proteomic analysis of S. japonicum cercarial secretions identified a peptide sequence common to this subset [12] (Figure 1B). Notably, while the S. mansoni genome encodes only a single cathepsin B2 isoform, S. japonicum encodes four CB2 isoforms. In one of these isoforms, SjCB(Y)2d (GenBank: CAX71091.1), the nucleophilic cysteine of the active site is mutated to tyrosine, which may diminish, if not eliminate, its catalytic activity. Three of the four SjCB2 isoforms (SjCB2b (GenBank: CAX71088.1), SjCB2c (GenBank: CAX71090.1) and SjCB(Y)2d correspond to the peptide sequence identified in proteomic analysis of S. japonicum cercarial secretions [12]. A full list of schistosome cercarial elastase and cathepsin B isoforms is provided as supplementary material (Tables S1 and S2).
A previous proteomic study generated a list of proteins that were released as soluble peptides from ex vivo human skin upon treatment with live S. mansoni cercariae, indicating that they are actively degraded during cercarial migration through skin [21]. These included many of the structural components of skin, including extracellular matrix proteins, proteins involved in cell-cell adhesion and multiple serum proteins. To identify specific substrates of CE in skin, and to compare these to potential substrates of cathepsin B2, we treated ex vivo skin with either peptidase. Since active, recombinant S. japonicum cathepsin B2 is not currently available, and purifying sufficient amounts of native peptidase from S. japonicum was not feasible, we used S. mansoni cathepsin B2 as model peptidase in our analysis. S. mansoni cathepsin B2 has high homology to the S. japonicum cathepsin B2 (90% sequence identity and 94% sequence similarity for the mature peptidase, see Figure S4), including the active site and substrate binding pocket, and therefore is likely to display highly similar biochemical properties and substrate specificity [17]. SmCE was purified directly from S. mansoni cercariae, and the protein composition of proteolytically active fractions was determined by mass spectrometric analysis as a mixture of SmCE1a, 1b and 2a isoforms, but not SmCE2b. This is consistent with the isoform composition of previous proteomic analysis of S. mansoni cercarial secretions [6], [22]. Active SmCB2 was expressed in recombinant form in P. pastoris and purified as previously described [17]. To ensure that equimolar amounts of active enzyme were added to skin samples, an active site titration was first performed for both SmCE and SmCB2 with respective covalent inhibitors.
In comparison to control samples treated with inhibited peptidase, multiple skin proteins migrated through an SDS-PAGE gel as smaller fragments, i.e. fragments less than the predicted molecular weight of the full-length protein, upon addition of active SmCE or SmCB2. These were thus identified as substrates of the specific enzyme and included multiple extracellular matrix proteins (Table 1). Addition of both SmCE and SmCB2 to skin led to the cleavage of collagen VI, which is found in interstitial tissue, and collagen XII, a collagen located in the basement membrane of the epidermis [23]. Only samples incubated with active SmCB2 showed cleavage of collagens I, III and XVIII. In addition to collagen, several other components of the extracellular matrix were degraded upon treatment with either peptidase, including vitronectin, fibronectin, and galectin. Both vimentin and talin-1, cytoskeletal proteins that are associated with desmosomes, were cleaved upon addition of either peptidase. Two additional extracellular matrix components, tenascin-X and thrombospondin-1, were uniquely cleaved upon addition of SmCB2.
Another subset of extracellular proteins identified as substrates of SmCE and SmCB2 were derived from blood plasma that bathes the dermis. These included components of the coagulation cascade, e. g. fibrinogen, antithrombin-III, as well as proteins involved in the host immune response, e. g. complement C3, complement factor D. Addition of either active SmCE or SmCB2 led to the digestion of gelsolin, an actin assembly protein that exists intracellularly and in plasma. Addition of active SmCB2 also led to the digestion of both kininogen-1 and fibrinogen, both of which are members of the coagulation cascade. Complement C3, an integral component of both the classical and alternative complement activation pathways was cleaved upon addition of either SmCE and SmCB2; complement C4A and complement D proteins, respective members of the classical and alternative complement activation pathways, were cleaved by SmCB2 alone.
In addition to the extracellular proteins identified, many cytosolic proteins were also cleaved by either SmCE or SmCB2. A complete list of peptides identified is provided as a supplementary table (Table S3).
To corroborate proteomic identification of substrates in skin, candidate substrates were selected for in vitro digestion with either SmCE or SmCB2. Type I collagen was of particular interest, given that lower molecular weight peptides of the protein were only found in skin samples treated with SmCB2, suggesting it is cleaved by SmCB2 but not SmCE. To test this with purified protein, type I human collagen was treated with either SmCB2 or SmCE for up to 22 hours at 37°C, and cleavage of the protein was determined by SDS-PAGE analysis (Figure 2). While the majority of collagen I was degraded after 5 hours with SmCB2 (Figure 2A), SmCE treatment resulted in the appearance of discrete lower molecular weight bands only after 22 hours of enzyme treatment (Figure 2B). This confirms that SmCE shows reduced activity against type I collagen relative to SmCB2, even in vitro. To confirm that the two peptidases cleaved collagen at unique sites, candidate lower molecular weight bands resulting from peptidase treatment were submitted for N-terminal sequencing, and the resulting amino acid sequence was mapped onto the full protein to determine cleavage sites (Figure 2C). Consistent with previous analysis of SmCE substrate specificity, in vitro digestion of collagen I revealed that peptide bond cleavage only occurred following a leucine residue (VRGL/TGPI) [15], [24]. In comparison, SmCB2 cleavage occurred following an arginine residue (GER/GGP), which is consistent with its reported activity, including a level of “promiscuity” in its amino acid preference in the P2 substrate binding pocket, relative to other types of cathepsins [17], [25].
Complement C3 was also of particular interest as a potential substrate of both SmCE and SmCB2, given its role in the host immune response against the parasite [26]. Purified complement C3 was treated with SmCB2 or SmCE. Discrete lower molecular weight bands were visible within 1 hour of treatment with either peptidase, in comparison to inhibited peptidase controls (Figure 3A, B). N-terminal sequencing of selected fragments again revealed that both SmCE and SmCB2 digested the protein in a manner consistent with their known specificities, with an arginine in the P1 position (RR/SVQ) for SmCB2 and and a tyrosine in the P1 position (TMY/HAK) for SmCE (Figure 3C).
In S. mansoni, the most abundant peptidase in cercarial secretions is a serine peptidase, termed cercarial elastase (SmCE) for its ability to degrade insoluble elastin [8], [22]. In addition to proteomic analysis, biochemical and immunolocalization studies have detected SmCE activity in cercarial secretions and confirmed that the enzyme is able to cleave such substrates as type IV collagen (basement membrane collagen), fibronectin, laminin and immunoglobulin in vitro [9], [10], [27]. Here, we have shown that SmCE cleaves additional substrates in skin, including several types of collagen, other extracellular matrix proteins, and components of the complement cascade.
Recent sequencing and annotation of the S. mansoni genome suggests a unique role for cercarial elastase. An expanded gene family was identified with ten individual genes that encode multiple isoforms of the peptidase. Even without a complete genome, multiple orthologs of SmCE have been also been found in S. haematobium, a related human-specific species of schistosome common throughout North Africa and the Middle East [15]. This is not the case for the zoonotic S. japonicum, a schistosome species that infects humans and other mammals throughout southeast Asia. The S. japonicum genome contains only a single gene encoding cercarial elastase. This gene corresponds to the cercarial elastase “2b” isoform in S. mansoni, for which minimal transcript is made relative to other CE isoforms (Ingram and McKerrow, unpublished). While one report suggested that CE was detected by immunofluorescence in S. japonicum secretions, no cercarial elastase protein was detected in a high resolution mass spectrometric proteomic analysis of S. japonicum acetabular secretions, and no cercarial elastase-like activity was identified by direct biochemical assays [12], [20]. Trichobilharzia regenti, an avian schistosome that is capable of invading human skin, but not establishing a successful infection in humans, encodes a cysteine peptidase, cathepsin B2 (TrCB2 (GenBank: ABS57370.1)), which has elastinolytic properties and localizes to the acetabular glands of the parasite [13]. S. japonicum also encodes a cathepsin B2 ortholog, and transcript is expressed in the developing larval stage of the parasite. Moreover, proteomic analysis has identified cathepsin B2 as being present in S. japonicum cercarial secretions [20]. Notably, S. japonicum has 40- fold higher cathepsin B activity in its acetabular secretions, relative to S. mansoni secretions [12]. It is therefore likely that in S. japonicum cercariae, cathepsin B2, not cercarial elastase, is the predominant invasive enzyme.
The differential use of these two classes of peptidases raises the question of how their respective pH optima are achieved in schistosome secretions. SmCB2 is maximally active under acidic, reducing conditions [28]. Since the influence of S. japonicum cercarial secretions on the local environment of skin is unknown, SmCB2 incubations were performed under acidic conditions to ensure optimal peptidase activity. SmCE activity is optimal in a slightly alkaline environment, and S. mansoni secretions are also alkaline; therefore all SmCE incubations were performed at pH 8 [29]. Certainly, for S. mansoni, the evolutionary selection is most likely coordination of the pH of the acetabular gland secretions and the pH optimum of the peptidase. The pH optimum of the cercarial elastase is 8, and the pH of the secretions is also alkaline [30]. As S. mansoni cercariae migrate through skin, a microenvironment is created by the secreted material, which allows for optimal activity of the peptidase. The situation is less clear for S. japonicum and the Trichobilharzia cercariae. While some activity of the cathepsin B2 is likely to continue at neutral, or even alkaline pH, the pH optimum is slightly acidic [17]. The situation is reminiscent of the secretion of cathepsin B by macrophages into tissue compartments of vertebrates. Secreted human cathepsin B is known to degrade extracellular matrix proteins in human tissue, where it has been reported to facilitate tumor invasion and metastasis [31]. The pH optimum of mammalian cathepsin B is also slightly acidic [32]. It is not known if the microenvironment around migrating macrophages is acidic or when that enzyme is released; however, it appears that there is sufficient cathepsin B activity to cause tissue degradation.
Given the unavailability of active, recombinant SjCB2 or sufficient amounts of S. japonicum cercariae from which to purify the native enzyme, we chose to perform our proteomic study with SmCB2, which displays high sequence homology (90% amino acid sequence identity for the mature peptidase) to its S. japonicum ortholog. We therefore hypothesized that it is likely to display similar biochemical characteristics, including similar substrate specificity. While we cannot say conclusively that SjCB2 is the protease facilitating S. japonicum cercarial invasion, we believe that our study, along with previous work from other groups, supports the proposed role for cathepsin B2 in host skin protein degradation [12], [13].
This conclusion, that S. japonicum uses a cathepsin B2 peptidase for skin invasion, while S. mansoni uses a serine peptidase (SmCE), has implications for the evolution of the human host-parasite relationship in schistosomiasis. A plausible model is that the cathepsin B2 family first emerged as the functional cercarial peptidase during trematode evolution. In contrast, the “humanized” parasites such as S. mansoni appear to have switched to a serine peptidase for cercarial invasion. This model is supported by the notable expansion of the serine peptidase gene family from the single 2b gene found in S. japonicum to the multiple isoforms expressed in S. mansoni [20], [33]. While the genome of the other “humanized” parasite, S. haematobium, has not been completed, it is already clear from EST analysis that more abundant serine peptidase isoforms are present in that genome [15].
What is the advantage of a larval serine peptidase for the “humanized” schistosomes? It is interesting to note that by BLAST analysis, some of the proteins with highest homology to cercarial elastase are mammalian mast cell peptidases, which are present in skin [12]. It is therefore possible that cercarial elastase evolved by convergence to resemble a human peptidase, in order to evade detection by the host immune system. Previous work shows that S. mansoni cercariae migrate through skin at a much slower rate than their S. japonicum counterparts [34]. Despite this, an inflammatory response to S. japonicum cercariae occurs more frequently than to S. mansoni cercariae [34], [35]. Cathepsin B2 is a likely target of the inflammatory response, given that many cysteine peptidases are allergenic [36]. Perhaps the rapid transit of non-humanized cercariae through skin precludes the need for an invasive enzyme that mimics a host peptidase. Other aspects of immune evasion, such as the elimination of complement factors and immunoglobulin, may be common to both species. C3 and C4 components bind to the tegument of schistosomes, but are degraded by both SmCE and SmCB2 [26], [37].
The results reported here show that S. mansoni cathepsin B2 (a model for S. japonicum cathepsin B2) and S. mansoni cercarial elastase are both capable of degrading proteins in skin that act as a barrier to cercarial invasion. Many skin proteins are substrates for both enzymes, but cathepsin B2 appears to cleave a broader range of substrates, and therefore may be a more effective invasive enzyme than cercarial elastase.
|
10.1371/journal.pgen.1003035 | Biochemical Diversification through Foreign Gene Expression in Bdelloid Rotifers | Bdelloid rotifers are microinvertebrates with unique characteristics: they have survived tens of millions of years without sexual reproduction; they withstand extreme desiccation by undergoing anhydrobiosis; and they tolerate very high levels of ionizing radiation. Recent evidence suggests that subtelomeric regions of the bdelloid genome contain sequences originating from other organisms by horizontal gene transfer (HGT), of which some are known to be transcribed. However, the extent to which foreign gene expression plays a role in bdelloid physiology is unknown. We address this in the first large scale analysis of the transcriptome of the bdelloid Adineta ricciae: cDNA libraries from hydrated and desiccated bdelloids were subjected to massively parallel sequencing and assembled transcripts compared against the UniProtKB database by blastx to identify their putative products. Of ∼29,000 matched transcripts, ∼10% were inferred from blastx matches to be horizontally acquired, mainly from eubacteria but also from fungi, protists, and algae. After allowing for possible sources of error, the rate of HGT is at least 8%–9%, a level significantly higher than other invertebrates. We verified their foreign nature by phylogenetic analysis and by demonstrating linkage of foreign genes with metazoan genes in the bdelloid genome. Approximately 80% of horizontally acquired genes expressed in bdelloids code for enzymes, and these represent 39% of enzymes in identified pathways. Many enzymes encoded by foreign genes enhance biochemistry in bdelloids compared to other metazoans, for example, by potentiating toxin degradation or generation of antioxidants and key metabolites. They also supplement, and occasionally potentially replace, existing metazoan functions. Bdelloid rotifers therefore express horizontally acquired genes on a scale unprecedented in animals, and foreign genes make a profound contribution to their metabolism. This represents a potential mechanism for ancient asexuals to adapt rapidly to changing environments and thereby persist over long evolutionary time periods in the absence of sex.
| Bdelloid rotifers are tiny invertebrates with unusual characteristics: they withstand stresses, such as desiccation and high levels of ionising radiation, that kill other animals, and they have survived over millions of years without sexual reproduction, which contradicts theories on the evolutionary advantages of sex. In this study, we investigate another bizarre feature of bdelloids, namely their ability to acquire genes from other organisms in a process known as horizontal gene transfer (HGT). We show that HGT happens on an unprecedented scale in bdelloids: approximately 10% of active genes are “foreign,” mostly originating from bacteria and other simple organisms like fungi and algae, but now functioning as bdelloid genes. About 80% of foreign genes code for enzymes, and these make a major contribution to bdelloid biochemistry: 39% of enzyme activities have a foreign contribution, and in 23% of cases the activity in question is uniquely specified by a foreign gene. This indicates biochemistry, such as toxin degradation and antioxidant generation, that is unknown in other animals and that is expected to improve the “robustness” of the bdelloid. It also represents a possible mechanism for survival without sex, by diversification of functional capacity and even replacement of defective genes by foreign counterparts.
| Bdelloid rotifers (Rotifera, Bdelloidea) are abundant, ubiquitous microinvertebrates that inhabit aqueous habitats [1]. They possess an extraordinary and unique combination of characteristics among the Metazoa: they have survived for tens of millions of years without sexual reproduction, while speciating similarly to sexual organisms; they can withstand extreme desiccation by undergoing anhydrobiosis; and they display other properties usually associated with extremophiles such as tolerance of high levels of ionizing radiation [2]–[7]. In addition, the bdelloids Adineta vaga and Philodina roseola contain foreign DNA sequences in at least some subtelomeric chromosomal regions of their genomes, and these probably derive from horizontal gene transfer (HGT) [8]. Three of these genes were shown to be transcribed, and Boschetti et al. [9] showed that in a related bdelloid species, A. ricciae, four different foreign genes, out of a set of 36 identifiable foreign and native sequences sampled, were expressed. Of these, three were upregulated by evaporative water loss and were therefore part of the desiccation stress response.
This suggests that horizontal gene transfer (HGT) might contribute significantly to the remarkable biology of the bdelloid rotifer. However, the proportion of the bdelloid genome harbouring foreign sequences, how many of these sequences are expressed, and their contributions to bdelloid physiology, are completely unknown. To address these issues, we present the first global analysis of the transcriptome of a bdelloid rotifer, A. ricciae, which shows that horizontally acquired genes are expressed on a scale unprecedented in animals and that they make a profound contribution to bdelloid metabolism. We suggest this is highly significant in the context of the extremophile nature of bdelloids and their long term evolutionary persistence without sex, which theory suggests should limit their ability to adapt to changing environments [10]–[13].
To capture expression of genes active during the hydrated and dehydrated states, cDNA was prepared and pooled from a laboratory strain of A. ricciae under both conditions, then partially normalised to reduce coverage of abundant transcripts. Paired-end, massively parallel sequencing was performed on cDNA fragments of mean size 200 bp using the Illumina platform; 19.5 million 76-base reads were assembled to give an initial library of 61,219 transcript contigs of size range 118–3674 bp. Of these, 28,922 contigs gave at least one significant blast hit (E-value≤10−5) when compared to the UniProtKB database, allowing the identification of their likely product, and these were used for further analysis.
Those transcripts originating from horizontally acquired genes were identified by assigning each contig an HGT index, hU, defined as the difference between the “bitscore” (i.e. score in bits) of the best non-metazoan match and the bitscore of the best metazoan match in the database. The subscript, U here, signifies the database used, UniProtKB; S for Swiss-Prot is used where appropriate below. A positive hU value for a given transcript means that its translation gives a better alignment to a non-metazoan protein than to a metazoan protein, and vice versa for a negative hU value. For comparison with other invertebrates, we carried out the same analysis with transcript datasets from the monogonont rotifer Brachionus plicatilis (a distinct class within phylum Rotifera, that has both sexual and asexual life phases, and is not considered anhydrobiotic, but can form desiccation-tolerant resting eggs), the nematode Caenorhabditis elegans and the fly Drosophila melanogaster.
Although for hU>0, a non-metazoan origin is indicated, there will be some uncertainty where non-metazoan bitscores are close to those of metazoans. Therefore, a threshold signifying foreign origin needs to be set at some value higher than zero. Figure 1A shows that the bdelloid contains many more foreign transcripts than other invertebrates, regardless of where a threshold might be set, and therefore other species can be used as a reference for ‘background’ levels of HGT in invertebrates. We calculate R(hU), the relative proportion of transcripts with HGT index value greater than a given value of hU, where R = (the percentage of transcripts from species 1 with HGT index≥hU)÷(the percentage of transcripts from species 2 with HGT index≥hU). In comparisons between A. ricciae and other invertebrate species, we notice that, for hU≤0, R is relatively constant since both metazoan and non-metazoan sequences are included. However, as the hU = 0 threshold is passed, R increases with hU as metazoan sequences are excluded, and the greater proportion of foreign sequences in the bdelloid transcriptome becomes apparent. R then plateaus around hU = 25–30 and is approximately constant up to hU∼100 (Figure 1B). This suggests that, as the threshold of hU = 30 is exceeded, the proportion of sequences judged to be foreign decreases, but at a similar rate in both the bdelloid and the comparator species, i.e. the ratio between species remains constant, indicating that increasing stringency above hU = 30 only results in loss of truly foreign genes from the count, and does not give a better test of “foreignness”. Figure 1B also shows that there is approximately 5-fold more HGT in A. ricciae than in either B. plicatilis or C. elegans, since R≈5 for hU≥30. For the comparison of A. ricciae and D. melanogaster, the ratio is appreciably higher at R≈16 (data not shown), in line with the apparently very low levels of HGT in the fly (Figure 1A). A comparison of B. plicatilis with C. elegans (Figure 1B) does not show the inflection between hU = 0 and hU = 30, consistent with these species having a similar proportion of foreign sequences in their transcriptomes. We used linear models to test whether differences in hU were significant among taxa. Results confirmed that A. ricciae had both a significantly higher mean hU score and a significantly higher probability per gene of hU>30 than the other taxa, even when controlling for differences in contig length between the assemblies (all comparisons, p<0.001, details in legend to Figure S1).
Of the identified bdelloid contigs, 9.7% (2,792/28,922) were shown to have hU≥30 and so were considered to be of foreign origin (Figure 1A, Table S1). In B. plicatilis, 1.8% (171/9,685) of transcripts have hU≥30, while in C. elegans and D. melanogaster this figure is 1.8% (206/11,168) and 0.6% (105/18,368), respectively (Figure 1A). This demonstrates that, independent of the dataset dimensions, the level of expressed HGT in bdelloid rotifers is far greater than in other invertebrates tested.
Phylogenetics was used to validate the foreign origins of putative horizontally acquired sequences [14] and this can be performed meaningfully where contigs with hU≥30 have a significant (E-value≤10−5) blast match to at least one metazoan sequence, allowing a phylogenetic tree to be constructed. However, two-thirds (1,884/2,792; 67%) of sequences with hU≥30 do not give a significant metazoan match, which strongly supports a foreign origin. For the remaining (908/2,792) contigs, phylogenetic trees were built in PhyML from amino-acids sequences using a JTT model [15]. Each contig was assigned to a particular group according to the aLRT support for each metazoan or non-metazoan taxon as follows: group 1 contains sequences that are monophyletic with Metazoa (or where there were only metazoan hits from the blast analysis); group 2 contains sequences for which monophyly with Metazoa cannot be strongly rejected; group 3 contains cases where there are too few sequences to define a meaningful clade; group 4 contains cases where monophyly with Metazoa can be strongly rejected; group 5 contains transcripts which are monophyletic with another single (non-metazoan) taxon. Analysis of these data showed that 98% of A. ricciae transcripts with hU≥30 and at least one significant metazoan hit fall into groups 4 and 5 with high node support (summarised in Table 1; Table S1; Figure S2) and therefore are supported as truly non-metazoan. For example, an acetyl-CoA synthetase (Enzyme Commission [EC] number 6.2.1.1) does not cluster with metazoan sequences for this enzyme, instead grouping within the eubacterial clade with high support (aLRT support = 0.86) (Figure 2A; Figure S2C). More than half of foreign transcripts appeared prokaryotic (59% eubacterial, 1% archaeal); the remainder were eukaryotic in origin: 23% fungal, 6% from algae or plants, and 11% from other eukaryotic taxa (largely protists).
A similar analysis can be performed for other invertebrates. For example, there are 206 transcripts from C. elegans with hU≥30 of which 108 give significant blast matches only with non-metazoan sequences. For the remaining transcripts, phylogenetic analysis shows that 92% (90/98) fail to cluster with metazoan examples (summarised in Table 1; Table S2; Figure S3). Therefore, 96% (198/206) of these C. elegans transcripts were verified as foreign by the phylogenetics. Although there are no comprehensive studies in the literature, the frequency of HGT we detect in C. elegans is higher than inferred in an earlier study [16]. One possible confounding factor might be that the phylogenetic placement of individual C. elegans sequences is impaired by filtering out other nematode sequences (see Materials and Methods). To check this, we repeated the evaluation including the top blast hits from nematodes, i.e. homologous and paralogous examples (Table S2; Figure S3). From the phylogenetics, we found that 93% (91/98) of C. elegans sequences did not cluster with the metazoa and therefore 97% (199/206) of the total set of transcripts with hU≥30 are likely to be foreign. This shows that the vast majority still lack a close non-nematode metazoan match when additional nematode sequences are included in the analysis. We interpret this finding as evidence of HGT in an ancestor of nematode species in the sample. However, as our aim here is not to evaluate levels of HGT in other metazoa beyond providing a baseline for comparison with bdelloids, these analyses are meant to illustrate that the results are robust to variations in the method, such as which sequences are included for evaluation.
To confirm that foreign transcripts originated from the bdelloid genome and were not due to contaminating or commensal organisms, several corresponding genomic regions were PCR-amplified and Sanger sequenced, and this showed that foreign genes were linked to a gene of metazoan origin or to another foreign gene from a different taxon (Figure 2B). In some cases (asterisks in Figure 2B), the foreign transcript was close to a gene previously described in a bdelloid rotifer. The sequences were also aligned to an early draft of the A. ricciae genome, where 91% of foreign transcripts aligned for at least 50% of their length, compared to 90% of all transcripts and for metazoan transcripts only (Figure 2C; data not shown). Furthermore, 81% of foreign transcripts were aligned to the same genomic contig as metazoan transcripts or foreign transcripts of a different phylogenetic group, which rules out an origin from contamination for this set (examples given in Table S3 correspond to some foreign sequences in Figure 3). This proportion is likely to rise as genome assemblies improve, but even if 10–20% of foreign genes cannot be shown to be part of the bdelloid genome, and thus represent contamination, this would only reduce the foreign component of the transcriptome to 8–9%, rather than 10%, which is still remarkably high.
Where HGT has been observed between prokaryotes, operational genes encoding, for example, enzymes, predominate over informational genes concerned with transcription and translation [17], [18]. If a similar situation pertains in bdelloids, we would expect to find many foreign genes that encode enzymes, which largely fall into the operational category [18]. Bdelloid transcripts with biochemical functions were identified by alignment to proteins with EC numbers in the Swiss-Prot database. This database was used as the quality of annotation is higher than UniProtKB and the smaller number of proteins should reduce the false positive rate (although it will also increase the number of false negatives). Of the 26,001 transcript contigs with matches in the Swiss-Prot database, 2,947 (11.3%) had hS≥30 and were categorised as foreign, i.e. a similar proportion to the previous analysis using UniProtKB (Figure S4).
Approximately 50% (13,059/26,001) of contigs (irrespective of their hS value) had a match with an assigned EC number. These were then tagged as either metazoan (hS≤0), indeterminate (0<hS<30) or foreign (hS≥30). Therefore, of the foreign transcripts, 79% (2,341/2,947) were annotated with an EC number, showing that a large majority are concerned with metabolism. In fact, when the functions of those without an EC number were analysed, a further 93 sequences could be identified as enzymes that lacked EC numbers due to poor annotation. This increases the proportion of foreign transcripts encoding metabolic functions to 83% (Figure S4).
Transcript contigs (in all categories) with assigned EC numbers were mapped onto the Kyoto Encyclopedia of Genes and Genomes (KEGG) reference pathways (denoted ‘K’ plus a number in the following). In total, 839 EC numbers assigned to the rotifer transcriptome were matched to 152 metabolic pathways (Table 2 and Table S4). Of the 839 EC numbers, 23% (191) were only assigned to foreign transcripts, with a further 16% (138) being assigned to both foreign and metazoan transcripts. This made a total of 39% of identified enzyme activities with a contribution from foreign transcripts, suggesting that HGT has the potential to radically diversify bdelloid biochemistry.
Many pathways containing foreign transcript products specify metabolism found only in micro-organisms and unknown in metazoans (Figure 3, Figure S5, Table 2, Table S4). Several of these are concerned with degradation of toxic compounds, and we give three examples here: 1) breakdown of phenylacetonitrile (benzyl cyanide) is initiated by the products of two genes derived from bacteria (EC 4.2.1.84 or EC 3.5.5.1; K00643; Figure 3A, Figure S5A), and other nitrile compounds, such as benzonitrile, can also be metabolised similarly (K00627, Figure S5B); 2) the organochloride pesticide, 1,3-dichloropropene, is degraded in five steps to the central metabolite, acetaldehyde, and the first of these is exclusively specified by the foreign-encoded enzyme haloalkane dehalogenase (EC 3.8.1.5; K00625; Figure 3B, Figure S5C); 3) branches of the degradation pathways for benzoate (K00362) and bisphenol (K00363) are also represented by foreign gene products (Figure S5D, S5E). Not all steps in these pathways are present in our transcriptome sample. This is partially due to the use of the Swiss-Prot database to assign EC numbers; performing the same analysis using the UniProtKB database adds steps to many pathways. However, there might also be incomplete capture of transcripts during cDNA cloning and sequencing, or bdelloids might only partially metabolise certain compounds. If the latter is correct, such partial metabolism might still be sufficient for detoxification or metabolite utilisation in other pathways.
HGT is also implicated in improved resource acquisition, e.g. two-step pathways to convert the ubiquitous natural phosphonates 2-aminoethylphosphonate (AEP) and 3-phosphonopyruvate into useable metabolites are enabled by foreign transcripts encoding 2-aminoethylphosphonate-pyruvate transaminase (EC 2.6.1.37) or phosphonopyruvate decarboxylase (EC 4.1.1.82) and phosphonoacetaldehyde hydrolase (EC 3.11.1.1) (K00440; Figure 3C, Figure S5F). Furthermore, several foreign transcripts are implicated in utilisation of a range of polysaccharides not normally directly available to metazoans, e.g. cellulose (K00500; Figure 3D, Figure S5G) and polygalacturonate (K00040; Figure 3E, Figure S5H) breakdown; α-N-arabinofuranosidase (EC 3.2.1.55; K00520, Figure S5I), glucan endo-1,3-β-glucosidase (EC 3.2.1.39; K00500, Figure S5G) and fructan β-fructosidase (EC 3.2.1.80; K00051, Figure S5J) are also encoded. Cellulase activity has been described in other invertebrates but, where this does occur, it seems to result from HGT (e.g. ref. [19]).
Other pathways novel to metazoans but represented in the bdelloid transcriptome are biosynthetic, some of which are associated with robustness. These include production of the powerful antioxidant, trypanothione, normally only produced by parasitic protozoa, which is specified by two foreign transcripts: a glutathionylspermidine synthetase (EC 6.3.1.8), and a trypanothione synthase (EC 6.3.1.9; K00480; Figure 3F, Figure S5K). Such antioxidants could play a role in desiccation tolerance, where protection of repair systems against oxidative stress is thought to be crucial [20]–[22]. Foreign gene products can also add extensions or linking steps to existing metazoan metabolism in A. ricciae. Valine and isoleucine are essential amino acids in animals and must normally be accumulated from the diet. However, foreign transcripts encode ketol-acid reductoisomerase (EC 1.1.1.86) and dihydroxy-acid dehydratase (EC 4.2.1.9), allowing completion of biosynthetic routes to these amino acids from pyruvate (K00290; Figure 3G, Figure S5L). A. ricciae also encodes a fungal form of pyruvate decarboxylase (EC 4.1.1.1; K00010), allowing an additional end step to glycolysis for the regeneration of NAD+ from NADH under anaerobic conditions with the production of ethanol (Figure 3H, Figure S5M); animals usually only form lactate from pyruvate under these conditions. A further intriguing possibility highlighted by the transcriptome analysis is that the bdelloid can fix carbon from CO2, using eubacterial forms of phosphoenolpyruvate synthase (EC 2.7.9.2) and phosphoenolpyruvate carboxylase (EC 4.1.1.31; K00720; Figure 3I, Figure S5N), by a route used in plants and bacteria, but not in fungi or animals. Where it is meaningful to do so, i.e. where there are significant metazoan blast matches, phylogenetic trees are shown in Figure S2G–S2M for example transcript contigs representing foreign-encoded activities in Figure 3.
In a few cases, where we would expect to find a metazoan sequence, this is absent from the transcriptome and the activity is instead represented by a foreign counterpart. For instance, a fungal form of stearoyl-CoA delta-9 desaturase (EC 1.14.19.1; K01040; Figure S5O), an essential enzyme for the synthesis of monounsaturated fatty acids, is present, but no metazoan equivalent was discovered in the transcriptome. To control for the possibility that relevant metazoan genes had not been expressed in study samples, we searched a draft A. ricciae genome sequence, but failed to find them, although the gene encoding the foreign transcript was present. While the inability to detect a particular sequence is not proof of its absence, this suggests that the metazoan form of stearoyl-CoA delta-9 desaturase has been lost in the bdelloid, perhaps following a detrimental mutation, and that a foreign gene has been co-opted in this role. Other examples of a foreign sequence potentially replacing a metazoan counterpart include nicotinate-nucleotide diphosphorylase (EC 2.4.2.19; K00760), which catalyses a step in NAD+ biosynthesis, and the antioxidant peptide-methionine (S)-S-oxide reductase (EC 1.8.4.11).
In recent years, there has been increasing interest in HGT, but most investigations have been performed in prokaryotes or in unicellular eukaryotes. In these organisms, HGT is considered a main driver of innovation, often associated with speciation [23], [24]. In multicellular eukaryotes, there has been less emphasis on HGT, partly because it is thought to occur on a much smaller scale [14], [25], and partly because there are fewer well-annotated genome sequences available. Since de novo whole genome assembly is still a significant challenge for complex organisms, particularly for the bdelloid rotifer with its unusual genome characterised by degenerate tetraploidy, divergence of formerly allelic sequences, and gene conversion between gene copies [7], [26], [27], we chose to assess HGT primarily at the transcriptome level. This study represents the first global analysis of the expressed genes in a bdelloid rotifer, A. ricciae, and the contribution of horizontally acquired sequences to its transcriptome. The results reveal a remarkable degree of HGT in the bdelloid, with approximately 10% of identifiable, transcribed sequences deriving mainly from prokaryotes, but also from fungi, plants and algae, and protists.
The method for assessing HGT in the bdelloid transcriptome is novel, but follows principles currently recognised as the most rigorous, where sequence matching is coupled with phylogenetics [14]. There have been relatively few such global analyses among the Metazoa that test for expression of horizontally acquired sequences, one example being in Hydra magnipapillata, where seventy-one “gene models” apparently derive from bacteria, 70% of which were shown to be transcribed [28]. For the bdelloid work, we introduced the HGT index, h, which is calculated as the difference in bitscores between best non-metazoan and best metazoan matches in blast alignments, to give a measure of the “foreignness” of any sequence. We preferred the HGT index to the alien index (AI), developed previously for assessing foreign sequences in bdelloid subtelomeric regions [8] and also used in the Hydra study [28], because h is calculated from bitscores and is therefore not influenced by the sizes of the databases used to perform the blast screen. In contrast, if E-values are used, as for the AI, the score changes with database size. Additionally, an arbitrary constant must be included in the AI formula so that the index does not become infinite with identical matches to database sequences; this adjustment is unnecessary with the HGT index. Although Figure 1A showed that, whatever value of h is chosen, there is a greater proportion of foreign sequences in the bdelloid than in other invertebrates, it is useful to adopt a threshold value to signify a foreign sequence. In principle, any sequence with h>0 is more likely be foreign, but there will be uncertainty at values close to zero where non-metazoan and metazoan sequences have similar degrees of divergence from the test sequence. One technique for identifying a reliable threshold value of h is to normalise the proportion of foreign sequences against the “background” levels found in other invertebrates. The greater proportion of horizontally acquired sequences in the bdelloid then becomes apparent above the minimum threshold level of h required to confidently identify their foreign nature, as shown in Figure 1B. This was validated by phylogenetics, where possible (i.e. where matching metazoan counterparts exist), which showed that the vast majority of bdelloid transcript contigs with hU≥30 did not cluster with metazoan sequences.
There are other technical considerations in any assessment of HGT. For example, we classified sequences as either metazoan or non-metazoan, and therefore any HGT from other animals (including other bdelloids) into the A. ricciae genome would be missed. Of course, there is no reason to believe that bdelloids are unable to acquire genes from other metazoans, or indeed from other rotifers; in fact, this might be more efficient than acquisition from microorganisms, since fewer changes to metazoan genes should be required before they become expression competent. Therefore, our approach is likely to give a minimum estimate of the extent of HGT in the bdelloid. Another factor that might influence this estimate is the approximately half of transcript contigs that showed no match with known sequences and therefore had to be excluded from further analysis. If all these sequences originate from vertical transmission into the bdelloid lineage, then this would reduce the estimate of HGT. However, there is no a priori reason to assume this: the proportion of foreign sequences in this non-matched set could be higher, lower or about the same as in the matched set. How the matched and non-matched sequence sets are defined could also potentially influence the proportion defined as HGT. We used 10−5 as a maximum value for a significant match when blast screening the transcript contigs against the databases and this gave 28,922 contigs in the matched set. If 10−10 or 10−15 is used as a cut-off value, the number of matched contigs decreases to 22,719 and 17865, respectively, but the fraction scored as foreign (i.e. with hU≥30) remains high, at 11.5% and 11.7% of matched sequences, respectively. Which database is used for blast matching also does not seem to be a major factor since both UniProtKB and Swiss-Prot gave similar proportions of foreign transcripts at 9.7% and 11.3%, respectively.
A final technical consideration might be to ask whether the HGT resulting from the endosymbiosis of the mitochondrial precursor affects our results. Endosymbiosis was a primordial event for eukaryotes, with acquisition of mitochondrial precursors taking place in the earliest eukaryotic cells, perhaps two billion years ago [29]. The horizontal gene transfer we describe is very unlikely to have occurred before the divergence of bdelloids from monogonont rotifers (or B. plicatilis would share similarly high levels of foreign genes), and therefore probably took place at most 100, more likely 65–80, million years ago [30]. If horizontal gene transfer has continued throughout bdelloid evolution, many events will be more recent. Consequently, most, perhaps all, gene flux from mitochondrial precursor to nucleus would have occurred before bdelloids arose. Thus, we would not expect significant differences in numbers of nuclear mitochondrial genes between bdelloids and the other major class of rotifers, the monogononts, as exemplified by B. plicatilis in our study. To test what proportion of foreign genes apparently derive from mitochondrial nuclear genes, we blast aligned sequences of 1,098 known nuclear mitochondrial genes from MitoCarta (www.broadinstitute.org/pubs/MitoCarta) against our transcripts. Using a cut-off of 10−5, only 0.7% of transcripts of foreign origin (hU≥30) matched mitochondrial nuclear genes, whereas 2.9% of those of metazoan origin (hU≤0) gave matches. If we adjust the blast cut-off to 10−10 and 10−15, these proportions are approximately the same: 0.7% vs. 3.3%, and 0.8% vs. 3.6%, respectively. This shows that transcripts for nuclear mitochondrial genes are less likely to be found in the foreign sequence set than among metazoan transcripts and therefore will not cause an overestimate of HGT.
The complexity of foreign gene expression observed in the bdelloid rotifer A. ricciae is comparable to that in prokaryotes [31] and is far greater than in other animals where relatively few genes are involved [14], [25], [28]. For example, while in Hydra perhaps 50 foreign genes are active [28], in Drosophila ananassae, which has acquired most of the genome of its endosymbiont, Wolbachia, by HGT, only 28 genes are transcribed; the model fly, D. melanogaster, has not acquired the Wolbachia genome [32], [33]. In pea aphids, red body colour results from the expression of carotenoid genes acquired and diversified from fungal counterparts [34], [35]. In the sea slug, Elysia chlorotica, HGT and expression of the algal psbO gene allows photosynthesis in plastids also acquired from the alga [36]. However, there is a need for more animal studies at the whole transcriptome level. It is surprising, for example, that there are no comprehensive global studies of HGT in C. elegans in the literature [37], as our analysis suggests there are approximately 200 foreign transcripts in the model nematode. The software pipeline developed for this study has the potential to be used more widely where expression data are available to gain a more complete picture of HGT in metazoans.
Nevertheless, the scale of HGT in the bdelloid seems to be unusual among animals and it would be interesting to address the importance of asexuality and desiccation tolerance in this phenomenon. For example, transcriptome data from the nematode Panagrolaimus superbus, which is anhydrobiotic, but gonochoristic (i.e. reproduces only sexually), has recently been published [38]. The authors highlighted one foreign sequence in the P. superbus transcriptome, but did not perform a global analysis for HGT. If this nematode contains low numbers of foreign sequences, it would rule out that desiccation tolerance per se, without asexuality, is associated with extensive HGT. Another characteristic of HGT in A. ricciae is that the source organisms are extremely diverse and include examples that are unlikely to be symbionts or even in the bdelloid's immediate habitat, such as the trypanosome relative from which trypanothione biosynthesis genes derive. Therefore, bdelloids are likely able to readily scavenge and incorporate DNA from the environment, and desiccation, which could lead to both leakiness in cell membranes and double-strand breaks in rotifer chromosomes, might facilitate this.
HGT in the bdelloid has the potential to radically extend and complement metazoan biochemistry, since approximately 80% of foreign sequences are concerned with enzyme activity, much of which is novel in animals. This supports the complexity hypothesis, which states that genes whose products are involved in relatively few protein-protein interactions, such as those specifying enzymes, are more likely to be horizontally transferred than those with a higher degree of connectivity, like transcription factor genes [17], [39], [40]. Thus, although the complexity hypothesis was developed to explain observations in prokaryotes, it also seems to apply to the large scale HGT observed in the bdelloid. It would be interesting to investigate in the bdelloid a more recent suggestion from a study in prokaryotes that highly expressed genes are less likely to be horizontally transferred between organisms [41]. Technically, this might be difficult to achieve, as we estimate there are at least 533 source organisms that have contributed to the bdelloid genome by HGT, but we will explore this in future work.
The novel biochemistry implicated includes the ability to degrade toxins, and indeed to exploit them and a range of otherwise unmetabolisable organic molecules as food sources, and to use novel biosynthetic pathways to generate protective molecules, for example antioxidants, or nutrients that are rare in the environment. This is expected to increase bdelloid stress tolerance and competitiveness, and could be important for anhydrobiosis. Bdelloids do not produce trehalose or other non-reducing disaccharides [42]–[44] and have unusual LEA proteins [26], [44], [45], and therefore mechanisms associated with desiccation tolerance in other anhydrobiotes do not apply. Recently, the bdelloid A. vaga was shown to have high antioxidant capacity; this reduces protein oxidation, which is thought to be a major problem caused by desiccation and the dry state [22]. Antioxidants in bdelloids have not been characterised, but it will be of interest to determine how far HGT plays a role; this is currently under investigation.
It is also tempting to speculate that HGT facilitates long-term persistence in the absence of sex: asexuals are unable to bring together novel gene combinations arising within a population since they lack conventional genetic exchange mechanisms; equally, asexuals cannot eliminate detrimental mutations readily [10], [11]. Uptake and expression of genes from other organisms is a means of diversifying functional capacity, particularly biochemical capacity, and the potential to replace defective genes with foreign counterparts could protect against loss of function through mutation.
The bdelloid rotifer Adineta ricciae was supplied by Claudia Ricci, University of Milan. A clone culture was split into four populations: one was kept hydrated and the other three were dehydrated for 24, 48 and 72 h, as described previously [9]. RNA was extracted separately from each bdelloid population with TRI reagent (Sigma) according to manufacturer's instructions. RNA purity and concentration were measured with a NanoDrop spectrophotometer. Oligo(dT)-primed cDNA from all four sets was prepared with a Clontech/Takara SMART PCR cDNA Synthesis Kit and an Advantage 2 PCR Enzyme System using Invitrogen SuperScript III Reverse Transcriptase. 1 µg cDNA from each preparation was pooled and the resulting mixed cDNA library was normalized with Evrogen Trimmer cDNA normalization kit, according to manufacturer's instructions. About 8 µg of both the normalized and non-normalised cDNA library (each made of the mixed of hydrated and desiccated rotifers) were pooled and a paired-end sequencing library with insert size 200 bp was prepared. Massively parallel Illumina sequencing was performed, resulting in 19.5 million 76-base reads. These were assembled with the CLC-bio (www.clcbio.com) assembler, using a k-mer size of 22, no minimum contig length and all other options at the default settings. The resulting assembly used 9,048,520 of the reads (46.4%) for a total length of 27,227,333 bp giving an average coverage of 25.3 times. This produced a library of 61,219 transcript contigs of size range 118–3674 bp, with median size 341 bp, and mean size 445 bp (standard deviation 295 bp). Transcript contigs have accession numbers HE687510 to HE716431.
Analysis of the bdelloid transcriptome was performed using R (The R Project for Statistical Computing, http://www.r-project.org/) complemented with NCBI-Blast 2.2.23+–2.2.25+ (Basic Local Alignment Search Tool) [46], ClustalW2 (EMBL-EBI) and PhyML 3.0 [47]. Blastx was used to compare the complete set of 61,219 bdelloid transcripts against taxa-specific subsets of UniProtKB, labelled as Metazoa, Plantae, Fungi, Eubacteria, Archea and “Other Eukaryotes” (Eukaryotes which are neither Metazoa nor Plants nor Fungi). The taxa-specific subsets only included sequences from complete proteomes (keyword: KW-0181) in order to reduce the search space and to avoid bias towards specific types of proteins that have been sequenced in many organisms. E-value and bitscores were collected for the best five hits of each transcript against each taxon, and 32,297 sequences that did not have any match with at least one taxon with an E-value≤10−5 were excluded from further analysis. The alien index [8] and the HGT index (hU) were calculated for each of the remaining 28,922 sequences. The HGT index (hU) is calculated as the difference between the highest non-metazoan and the highest metazoan bitscore. Bitscores, being independent of the search space, do not depend on the size of the database used to calculate the blast score, reducing the incorrect determination of sequences. Setting the hU threshold value is explained in the text. Similar analyses were performed for the C. elegans (WormBase release WS226; www.wormbase.org), D. melanogaster (FlyBase release r5.37) and B. plicatilis transcriptomes. In the first two of these cases, proteins from the phylum containing the test organism (i.e. Nematoda/Arthropoda) were excluded from the Metazoan database, as is common practice [8], [28]. For both A. ricciae and B. plicatilis this exclusion was not necessary as there are currently no complete proteomes available for the phylum Rotifera. For B. plicatilis, ESTs with accession numbers FM897377–FM945301 [48] were first assembled with CAP3 [49] into 16024 contigs, which became 9685 contigs after filtering for a blastx E-value≤10−5.
To confirm the non-metazoan origin of the sequences with hU≥30 and at least one significant metazoan hit, each transcript meeting these conditions was translated and aligned using ClustalW2 to the output (the best hits for each of the five taxa) of the previous blastx analysis. Each alignment was then trimmed to exclude regions where only one of the sequences was present, and phylogenetic trees were built in PhyML from amino-acids sequences using a JTT model [15]; branch support was calculated with the aLRT (approximate Likelihood-Ratio Test) method. The transcripts were then assigned to one of five groups according to the aLRT support for each metazoan or non-metazoan taxon: group 1 contains sequences that are monophyletic with Metazoa (or where there were only metazoan hits from the blast analysis); group 2 contains sequences for which monophyly with Metazoa cannot be strongly rejected; group 3 contains cases where there are too few sequences to define a meaningful clade; group 4 contains cases were monophyly with Metazoa can be strongly rejected; group 5 contains transcripts which are monophyletic with another single (non-metazoan) taxon. Analysis of these data showed that 98% of the sequences with at least one significant metazoan hit and hU≥30 are truly non-metazoan as they fall into groups 4 and 5 (Table 1; Table S1). To compare the bdelloid transcriptome to those of other species, the same analysis was performed on the published transcriptomes from the monogonont rotifer B. plicatilis, the nematode C. elegans and the fly D. melanogaster, calculating the percentage of sequences above threshold for a given value of hU as shown in Figure 1A.
To confirm the presence of foreign genes in the bdelloid genome and to assess the possibility of contamination from food, symbionts, parasites and other organisms, we manually sequenced the genomic DNA around some genes of interest. A number of assembled transcript fragments, chosen at random from a subset of foreign sequences encoding biochemical functions that have never been reported in metazoans, were blast screened against a (partial) genome assembly of A. ricciae and the longest genomic DNA contig for each transcript was identified. This was then compared using blastx to the NCBI non-redundant database to find other genes on the same genomic DNA fragment, and primers were designed around these regions. Genomic DNA was extracted from an A. ricciae population derived from the original, and 11 individual regions, were PCR amplified using Finnzymes Phusion High Fidelity Taq polymerase, adding an A overhang after PCR with Advantage 2 PCR Enzyme System. The resulting PCR product was cloned into pCR 2.1 TOPO TA (Invitrogen), inserted into competent E. coli (New England BioLabs) and white-blue colony screening was performed. Ten positive colonies for each PCR product were chosen, and plasmid DNA was purified and restriction digested to check for insert size. One clone for each genomic DNA region was sequenced via primer walks using a standard dideoxy method at the University of Cambridge Department of Biochemistry Sequencing Facility. Of the 11 attempted, eight are shown as Figure 2B. For the remaining three examples, one amplification worked, but sequencing could not be completed since the insert was long and unstable in E. coli: although the sequence of the middle of this fragment could not be determined, we confirmed that one end contained a metazoan gene and the other contained two genes of bacterial origin. Another amplification was not of the target region, and one amplification failed altogether.
The successfully amplified and sequenced genomic DNA regions were then manually aligned in Geneious (www.geneious.com) with the relevant transcripts from the library, then blastx aligned against the non-redundant NCBI database and annotated. Each annotated gene was considered metazoan or not-metazoan according to its best hits in the published database. Figure 2B represents eight genomic regions with the annotated genes colour-coded according to the tree in Figure 2A (metazoa, black; eubacteria, blue; fungi, pink; protists, grey). In two cases, two foreign genes are present on the same genome fragment, but they derive from different taxa. Occasionally, a gene next to a foreign representative has been identified previously in a bdelloid rotifer species, and is annotated with an asterisk in the figure. Figure 2B shows the shortest region including one foreign gene and one metazoan (or another non-metazoan from a different taxon) gene, but in a few examples the actual sequenced region was longer than shown. Accession numbers for these eight genomic regions are HE662868 to HE662875.
Transcripts were aligned to the draft genome using blastn. To determine the total length of alignment of a transcript all matches for that transcript fulfilling the following criteria were used: 1) E-value≤10−3; 2) non-overlapping with any previous matches; 3) longer than 40 bp (a minimum exon length constraint); and 4) on the same genomic contig as a previous match OR within 1000 bp of the start/end of a genomic contig when a previous match was also within 1000 bp of the start/end of a genomic contig (a maximum intron length constraint). The total aligning length (sum of the length of the matches that fulfill these conditions) was then divided by the length of the transcript and this percentage plotted as a histogram for all transcripts. Transcripts were considered to be “on” a genomic contig if they had a match on it fulfilling the above criteria. For each genomic contig with a foreign transcript on it the number of metazoan transcripts and foreign transcripts with a different origin was calculated.
The 28,922 sequences with at least one match with E-value≤10−5 were blastx aligned to the whole of Swiss-Prot (532,146 proteins at time of analysis) and the results were filtered to give only transcripts with at least one match to a protein that was annotated with an EC number. An HGT index for these transcripts was then calculated as before (but denoted hS, to show that the comparison was done with Swiss-Prot rather than with UniProtKB, cf. hU). Based on hS, the transcripts were then subdivided into three groups: horizontally transferred (hS≥30), indeterminate (0<hS<30) and metazoan (hS≤0). For horizontally transferred and metazoan transcripts the EC numbers of their first match were collated and input into the KEGG website (http://www.genome.jp/kegg/tool/map_pathway1.html) to determine which KEGG pathways they occurred in.
EC numbers were also assigned a colour: green (EC number is only annotated to matches of transcripts with metazoan origin), red (EC number is only annotated to matches of horizontally transferred transcripts), grey (EC number is only annotated to matches of transcripts with indeterminate origin), orange (green plus red (and possibly grey)), pink (red plus grey), light green (green plus grey). These were input into the KEGG website (http://www.genome.jp/kegg/tool/map_pathway2.html) to produce the coloured pathway diagrams shown in Figure 3 and Figure S5.
The results were then extracted from the KEGG website and a hypergeometric test performed to calculate which KEGG pathways were enriched for horizontal transfer as compared to the total number of unique EC numbers found for all transcripts and the total number of unique EC numbers found for horizontally transferred transcripts. Benjamini-Hochberg multiple testing correction was performed to reduce the false positive rate (Table S4). The workflow is shown in Figure S6.
|
10.1371/journal.pntd.0000968 | Clinical Presentation of T.b. rhodesiense Sleeping Sickness in Second Stage Patients from Tanzania and Uganda | A wide spectrum of disease severity has been described for Human African Trypanosomiasis (HAT) due to Trypanosoma brucei rhodesiense (T.b. rhodesiense), ranging from chronic disease patterns in southern countries of East Africa to an increase in virulence towards the north. However, only limited data on the clinical presentation of T.b. rhodesiense HAT is available. From 2006-2009 we conducted the first clinical trial program (Impamel III) in T.b. rhodesiense endemic areas of Tanzania and Uganda in accordance with international standards (ICH-GCP). The primary and secondary outcome measures were safety and efficacy of an abridged melarsoprol schedule for treatment of second stage disease. Based on diagnostic findings and clinical examinations at baseline we describe the clinical presentation of T.b. rhodesiense HAT in second stage patients from two distinct geographical settings in East Africa.
138 second stage patients from Tanzania and Uganda were enrolled. Blood samples were collected for diagnosis and molecular identification of the infective trypanosomes, and T.b. rhodesiense infection was confirmed in all trial subjects. Significant differences in diagnostic parameters and clinical signs and symptoms were observed: the median white blood cell (WBC) count in the cerebrospinal fluid (CSF) was significantly higher in Tanzania (134cells/mm3) than in Uganda (20cells/mm3; p<0.0001). Unspecific signs of infection were more commonly seen in Uganda, whereas neurological signs and symptoms specific for HAT dominated the clinical presentation of the disease in Tanzania. Co-infections with malaria and HIV did not influence the clinical presentation nor treatment outcomes in the Tanzanian study population.
We describe a different clinical presentation of second stage T.b. rhodesiense HAT in two distinct geographical settings in East Africa. In the ongoing absence of sensitive diagnostic tools and safe drugs to diagnose and treat second stage T.b. rhodesiense HAT an early identification of the disease is essential. A detailed understanding of the clinical presentation of T.b. rhodesiense HAT among health personnel and affected communities is vital, and awareness of regional characteristics, as well as implications of co-infections, can support decision making and differential diagnosis.
| Sleeping sickness, or Human African Trypanosomiasis (HAT), caused by Trypanosoma brucei rhodesiense is one of the most neglected tropical diseases. It affects mainly rural, poor East African populations and has very high socio-economic impacts. T.b. rhodesiense HAT is an acute disease; patients quickly progress from the first stage, where trypanosomes are detectable in blood and lymph, to the second stage, where parasites penetrate the central nervous system. If left untreated, T.b. rhodesiense HAT is fatal. Disease control is hampered by the absence of sensitive diagnostic tools and safe drugs. Second stage patients can only be treated with melarsoprol, a highly toxic, arsenical drug. It is more difficult to treat patients successfully at advanced stages of the disease, and late onset of treatment should be avoided. Yet, most patients are treated for other conditions prior to HAT diagnosis. Therefore, it is important that health personnel in T.b. rhodesiense endemic regions have a detailed understanding of the clinical presentation of the disease and consider regional characteristics of T.b. rhodesiense HAT for decision making and differential diagnosis.
| Human African Trypanosomiasis (HAT), also known as sleeping sickness, is caused by the protozoan parasites T.b. gambiense (West and Central Africa) and T.b. rhodesiense (East and South Africa). The disease is transmitted by tsetse flies (Glossina ssp.) predominantly in the rural areas of most of sub Saharan Africa. 60 Million people live at risk of infection, but less than 10% are under adequate surveillance [1], reflecting its neglected status. Sleeping sickness caused by either subspecies presents in two disease stages defined as the first, or haemo-lymphatic stage and the second, meningo-encephalitic stage. Diagnosis of HAT is made in blood, lymph and the cerebrospinal fluid (CSF). The second stage of the disease is indicated by the presence of trypanosomes and/or an elevated white blood cell (WBC) count (≥5WBC/mm3) in the CSF. The disease stage and the causative species of infection direct the choice of treatment. T.b. gambiense infections are treated with pentamidine in the first stage and eflornithine, a combination of eflornithine and nifurtimox, or melarsoprol in the second stage [2]–[4]. T.b. rhodesiense first and second stage infections are treated with suramin and melarsoprol respectively [2]. In the field, the trypanosome subspecies is entirely determined by the geographical location of the patient as the distinction of T.b. gambiense and T.b. rhodesiense is only possible in well equipped laboratories through PCR analysis. The detection of the human serum resistance-associated (SRA) gene unequivocally identifies T.b. rhodesiense trypanosomes [5], [6]. In Uganda, the only country where both forms of the disease are present, a potential geographical overlap of the two endemic areas has become likely [7]. This would hamper determination of infective trypanosomes under field conditions and therefore also the identification of the correct treatment.
For first stage infections there are no specific clinical signs and symptoms in both forms of the disease; fever, headache and loss of appetite are common. In T.b. rhodesiense the presence of a chancre at the site of the infective bite may be indicative for a trypanosome infection [8]. Second stage infections show disease-characteristic neuro-psychiatric signs and symptoms: severe endocrinological and mental disturbances and severe motor problems are the main signs [9]. While often considered together, Gambiense and Rhodesiense HAT are clinically and epidemiologically different diseases [10]. T.b. gambiense HAT is a chronic disease, whereas T.b. rhodesiense is characterized by an acute disease progression. If left untreated, both forms of HAT are fatal. The mean time to reach the second stage has been estimated at over one year for T.b. gambiense [11] but only 3 weeks for T.b. rhodesiense HAT [12]. Correspondingly, average times from infection to death are almost 3 years and 6 to 12 months, respectively [11], [12].
A diversity of forms of clinical progression from asymptomatic to acute have been reported for T.b. gambiense infections [13]–[15]. This seems to be even more pronounced for T.b. rhodesiense infections; a wide spectrum of disease severity ranging from a chronic disease pattern in southern countries of East Africa with existing reports of asymptomatic carriers [16] to an increase in virulence towards the north had been described [17]. Even though those differences were already described more than 60 years ago [18] the first comparative study was carried out in 2004: on the basis of the SRA gene polymorphism, trypanosomes isolates from Uganda (acute profile) and Malawi (chronic profile) confirmed to be of different genotypes. However, clinical characteristics of the study groups were limited the presence of a chancre and the self-reported duration of illness [19]. Another hypothesis postulates that the differences in disease severity could be attributed to differences in genetic resistance to trypanosomiasis among host populations [18].
From the estimated 50′000 to 70′00 cases per year [20], over 97% are T.b. gambiense cases and only a few thousand are due to T.b. rhodesiense [1]. Therefore, most literature concentrates on T.b. gambiense HAT. Its clinical picture and related cardiac and endocrinological disorders have been extensively described [21]–[28]. On the other hand, literature on the clinical aspects of T.b. rhodesiense HAT is scarce. We identified four studies (see table 1) describing its clinical presentation. Only one study in 60 patients infected with T.b. rhodesiense was designed prospectively and used a standardized questionnaire [29].
In this paper we describe the clinical presentation of second stage T.b. rhodesiense HAT in 138 patients from two distinct geographical settings in East Africa. We compare our findings to the existing literature and discuss factors that could explain the differences observed.
The Kaliua Health Centre (KHC), a 50-bed missionary hospital in Tanzania (Urambo District) and the Lwala Hospital, a designated 100- bed district hospital in Uganda (Kaberamaido District) participated in the Impamel III program (improved application of melarsoprol).
A proof-of-concept trial (n = 60) followed by a utilization study (n = 78) to assess the safety and efficacy of the abridged, 10-day melarsoprol schedule for the treatment of second stage HAT [30], [31] in T.b. rhodesiense patients.
Eligible for enrolment were second stage patients with a minimum age of 6 years and confirmed second stage HAT. Patients with first stage infections, pregnant women and moribund or unconscious patients were excluded. Patients were passively enrolled at the study sites.
Diagnosis of HAT was made in blood and in CSF. Blood was examined using microscopy and/or the haematocrit centrifugation technique [32]. If trypanosomes were present, a lumbar puncture was performed for disease staging. Analysis of the CSF was done by direct microscopy and/or single modified centrifugation technique and white blood cell (WBC) count using counting chambers. Second stage infections were confirmed by the presence of trypanosomes and/or ≥5 WBC/mm3 in the CSF. The standard assessment of co-infections included malaria, filariasis and voluntary testing for HIV/AIDS.
The local Principal Investigators filled individual case report forms (CRFs). Data used for describing the clinical presentation of the disease were patient demographics, diagnostic findings, self reported duration of illness and clinical signs and symptoms on admission graded by scale of severity (grade 0, 1, 2).
Each participant gave written informed consent. For the participation of children and adolescents (below 18 years) the parents, the legal representative or the guardian gave written informed consent. Ethical clearances were obtained from the Ethics Committees in Tanzania (National Institute for Medical Research), Uganda (Ministry of Health) and Switzerland (Ethics Committee of both cantons of Basel). Before first patient enrolment, the Impamel III program was registered in the database of Current Controlled Trials (ISRCTN40537886).
All data were double entered and verified using Epi Data 3.1 software (www.epidata.dk) and analysis was accomplished with the statistical software package STATA Version IC10.0 (STATA, StataCorp, USA). The statistical analysis was performed comparing proportions with the Pearson Chi Square and means with the Student's t test. Logistic regression was used to test differences between groups of patients with different co-infections.
The use of the abridged 10-day melarsoprol schedule for the treatment of second stage T.b. rhodesiense HAT was highly satisfactory (detailed safety and efficacy data to be published separately). In this paper we describe the clinical presentation of the disease in 138 second stage patients from Tanzania and Uganda. The majority of patients were passively detected. Nine (9) patients from Uganda (13%) were actively identified during a survey of the National Agricultural Research Organisation (NARO) in the HAT endemic region of the country. There was no significant difference between actively and passively recruited patients for the median WBC count in the CSF (actively detected: median WBC = 27, IQR = 24; passively detected: median WBC = 19, IQR = 43, p = 0.067) and the median self-reported duration of illness (actively detected: median = 3 months, IQR = 2, passively detected: median = 2 months, IQR = 4, p = 0.141). 14 patients (11 in Uganda and 3 in Tanzania) could not be examined per protocol as they died or were in a comatose state upon arrival at the study sites which led to an exclusion of those patients from the Impamel III trials.
By molecular analysis of blood samples, the presence of the SRA gene [33] was demonstrated and confirmed T.b. rhodesiense infection in all trial subjects [34].
Data on the demographic and diagnostic baseline characteristics of the study population are shown in table 2. The proportion of male (57.2%) and female (42.8%) patients was comparable. 18.8% (26/138) trial participants were younger than 16 years whereof 88.5% (23/26) were enrolled in Uganda. There were no county-specific differences for the presence of trypanosomes in blood and CSF: 99% (68/69) of patients from Tanzania and 91% (63/69) from Uganda had trypanosomes in blood (p = 0.0524) and 70% (55/69) and 86% (59/69) respectively had trypanosomes in the CSF (p = 0.3690). However, there was a significant difference for the median WBC count in the CSF in Tanzania and Uganda (134 vs. 20 WBC/mm3, p<0.0001). Also, a body mass index (BMI) below 16.5 was more frequent in patients from Uganda (p<0.0001).
Clinical signs and symptoms reported at baseline and the level of significance (95%) are summarized in table 3. Headache, fever, general body pain and joint pains were common in both study populations. Clinical suspicion for cardiac insufficiency was found in both countries: 5.1% (7/138) of the patients had indication for left heart insufficiency (combination of cough and dyspnoe) and 5.8% (8/138) for right heart insufficiency (combination of oedema and hepatomegaly). Patients in Uganda had a more unspecific presentation of the disease whereas specific signs and symptoms for second stage HAT, namely sleeping disorders and aggressiveness were more common in patients from Tanzania.
To look at changes of diagnostic markers and clinical signs and symptoms over time we compared them in patients grouped by self-reported duration of illness (see figure 1). In Tanzania and Uganda 21.7% (15/69) and 36.2% (25/69) respectively were diagnosed with HAT having signs and symptoms for one month or less. 47.8% (33/69) of patients from Tanzania and 31.9% (22/69) of patients from Uganda were diagnosed having signs and symptoms of the disease between 1 and 3 months. Respective percentages for diagnosis of HAT after feeling ill for more than 3 months were 30.4% (21/69) in Tanzania and 31.9% (22/69) in Uganda. In both countries, the presence of trypanosomes in blood and/or CSF and the WBC count in the CSF did not significantly change over time. Also, there was no change over time for most of the clinical signs and symptoms. However, we observed that tremor (p = 0.01), walking difficulties (p = 0.040), sleeping disorders at night (p = 0.029), disturbed appetite (p = 0.044) and aggressiveness (p<0.001) aggravated over time in all patients.
Per protocol, standard assessment of co-infections at baseline included malaria and filariasis. 79.7% (55/69) of the patients from Tanzania and 2.9% (2/69) from Uganda were malaria positive on admission. None were found positive for filariasis. The HIV status was determined on voluntary basis. In Tanzania, 94.2% (65/69) of the patients tested their status and 24.6% (16/65) were found positive. In Uganda, 31.9% (22/69) tested their status and 9.1% (2/22) were found positive. We used the data from Tanzania to study implications of malaria and HIV co-infections on the clinical presentation and treatment outcomes of T.b. rhodesiense HAT. No significant difference either in the clinical appearance or in treatment outcomes for those patients was found. Details are shown in table 4 and 5.
Based on data from the Impamel III trials we describe the clinical presentation of second stage T.b. rhodesiense HAT in Tanzania and Uganda and confirm a wide spectrum of clinical presentation in these two geographically distinct areas in East Africa. In both settings T.b. rhodesiense HAT followed the classical disease pattern, but interestingly the neurological signs and symptoms typical for HAT were seen in a relatively small percentage of patients from Uganda. In patients from Tanzania, however, they were the dominate clinical manifestation. This correlated with the significantly higher reported CSF WBC counts in patients from Tanzania.
Unspecific signs of the disease such as fever, headache, general body pain and joint pains were reported in similar proportions in both study populations. We observed fever (≥37.5) in 29.7% (41/138) of the trial subjects. In the literature, fever was reported in the range of 31–71% in second stage patients from Zambia [18], [29], [35]. In the two study populations we saw high fever (>38.5) on admission only in Uganda (5.8%, 4/69) whereof 50% were children. Fever seems to be more common in T.b. rhodesiense than T.b. gambiense second stage patients in which fever was only occasionally reported (16%) and high fever was mostly seen in children [25]. In the two study populations, oedema was reported in Uganda and Tanzania in 20.3% and in 37.7% of the patients, respectively (p = 0.0244). This was comparable to the reported range of oedema in the literature (21.7–43.3%) [18], [29], [35], [36].
The clinical aspects of T.b. gambiense HAT [21], [25], [26], [37] have been systematically studied and show that the hallmark of second stage disease are neurological signs and symptoms [21], [25]. Unfortunately, this has never been done for T.b. rhodesiense HAT and hampers comparisons. However, published data report sleeping disorders during daytime hours with 63.3–70.5% of patients being affected [29], [35]. We observed sleeping disorders during daytime hours in Uganda and Tanzania in 56.5% and in 95.7% of the patients, respectively (p<0.0001). Similarly, sleeping disorders at night time are reported in the literature in 28.3% of patients [29]. We observed it in 34.8% of the patients from Uganda and in 92.8% of the patients from Tanzania (p<0.0001). Also other neurological signs and symptoms were significantly more frequent in patients from Tanzania; tremor (p = 0.0001), abnormal movements (p<0.0001), inactivity (p = 0.0076) and aggressiveness (p<0.0001). Clearly, the neurological signs and symptoms are more pronounced in Tanzania than in Uganda, and when compared to the literature.
In Uganda, almost 50% of patients were in a poor nutritional status (48% had BMI<16.5) as food security is very poor in this part of the country. This most likely contributes to weakness and, therefore, walking difficulties in the absence of neurological symptoms. Malnutrition is associated with immunodeficiency and higher susceptibility for a wide range of infections such as tuberculosis [38], [39] and pneumonia [40], as well as a poorer response to treatment. Another potential consequence of malnutrition in Uganda is an increased number of patients admitted with severe coma indicating a more rapid progression of the disease. Yet, we assume that many HAT cases from T.b. rhodesiense endemic areas in Tanzania die without ever having had contact with the health system due to geographical isolation.
With regards to treatment outcomes, we did not see any differences in the two study populations. In both countries all patients were free of parasites at end of treatment. Also, there was no apparent difference in parasite clearance rates. Time- and treatment-dependant dynamics of CSF WBC counts in the two study populations will be published separately.
Cardiovascular involvement is typical, but rarely of clinical relevance in T.b. gambiense HAT [41], [42]. We have limited knowledge of the effects of cardiac involvement in T.b. rhodesiense patients, but there is evidence that perimyocarditis seems to play an important role in the clinical course and fatal outcomes [43], [44]. We observed symptoms of cardiac failure such as oedema (swelling of legs) in 29% of the patients. Hepatomegaly occurred in 18%, dyspnoea in 7% and cough in 20% of the patients. However, echocardiography or laboratory testing (i.e. brain natrium peptide) could not be performed to confirm heart failure.
Co-infections with malaria and HIV were studied in detail in the patient population from Tanzania as the majority of the patients were malaria-positive on admission (80%) and agreed to voluntary testing of their HIV status (94.2%). Patients that were malaria-positive on admission more often had pruritus (p = 0.025), sleeping disorders during day time hours (p = 0.026) and disturbed appetite (p = 0.01). Also, they exhibited strange behaviour more often (p = 0.001). However, there is insufficient evidence for profound differences in malaria-positive and malaria-negative subjects, possibly due to asymptomatic carriers.
We identified one study that looked at T.b. rhodesiense and HIV co-infections in 25 patients from Kenya. In terms of treatment outcomes no conclusive results were obtained [45]. Our results indicate that the HIV status of the patient does not change the clinical presentation and/or the treatment outcomes of T.b. rhodesiense HAT. For T.b. gambiense HAT, there seems to be no association between HIV and HAT infection rates [46], [47] but evidence exists for a negative association with treatment outcomes [47], [48]. More research efforts are needed to better understand the complex interactions of co- infections, especially for neglected tropical diseases [49].
Our findings on the different clinical presentation of T.b. rhodesiense HAT in the two study populations could be due to an observation bias, bias in patient selection, or in comparing patients at incongruous time points after infection. Bias due to co-infections or differences in host and/or parasite genetics is also possible.
An observation bias can not be ruled out but is however less likely as the Impamel III program was conducted with a structured case report form (CRF) and one monitoring person. We have seen variability in signs and symptoms with clear definitions (e.g. lymphadenopathy, abnormal movements or tremor) as well as subjective definitions (e.g. insomnia, headache or inactivity). We can not completely rule out a selection bias due to the exclusion of moribund and unconscious patients in which baseline examination per protocol was not possible. However, the number of excluded patients was relatively small (<10%) and the two study populations were similar in regards to self-reported duration of illness. Even though unsuccessful, active case searches were conducted in both countries which reduced a potential selection bias. Central nervous system involvement in T.b. rhodesiense HAT was previously reported within 3 weeks to 2 months of infection [12]. One third of the study population already had clear neurological signs and symptoms within one month of infection which reflects the acuteness of T.b. rhodesiense infections. The WBC count in the CSF as well as most of the clinical signs and symptoms also developed quickly and did not significantly change over time. Disease progression was noticeable by aggravation of tremor, walking difficulties, sleeping disorders at night time, disturbance of appetite and aggressiveness over time, in both study populations.
Based on the results shown we rule out a bias of our findings due to co-infections. Previous infections with trypanosomes and/or host genetics might be determinants for the different clinical presentation of the disease in Tanzania and Uganda. There are speculations that apathogenic forms of the disease could influence immune responses to pathogenic infections [50], [51] supported by the fact that HAT is more acute in white than in the black populations [52], [53]. But we also see a high variability in disease severity among African populations [17], [18], a fact that has been related to the descent of people: people of Nilotic descent, who migrated into the East African region from tsetse-free areas during the past 2,000 years may have less tolerance than people of Bantu descent, whose ancestors have been exposed to human trypanosomes for several thousand years [18]. Our findings do not align with this theory as in Tanzania, the majority of the population is of Bantu origin and in Uganda the majority of the population is of Nilotic origin. Different parasite genotypes could be responsible for the observed spectrum of disease severity, a hypothesis has already been raised 60 years ago [16], [17], [54]. Recent findings on the phylogenetic relationship between different T.b. rhodesiense strains showed that the high variability of the T.b. rhodesiense genome is attributed to multiple and independent evolutions from T.b. brucei [55]. Our data show a clear difference in the clinical presentation of T.b. rhodesiense HAT in Tanzania and Uganda but a detailed assessment of host and parasite genotypes was beyond the scope of this paper.
T.b. rhodesiense HAT is a highly neglected disease and tools for disease control are very limited. There are no sensitive diagnostics at hand and melarsoprol, the only available drug to treat second stage disease, is toxic. An early identification of the disease is vital to prevent late onset of treatment. However, most of the patients are first treated for other conditions such as malaria and pneumonia. A low degree of disease awareness among health personnel is common and aggravated by the low prevalence and the focal distribution of HAT. A detailed understanding of the clinical presentation and regional characteristics of T.b. rhodesiense HAT is important and can support decision making and differential diagnosis at health facility level.
|
10.1371/journal.pcbi.1006485 | Active dendrites regulate the spatiotemporal spread of signaling microdomains | Microdomains that emerge from spatially constricted spread of biochemical signaling components play a central role in several neuronal computations. Although dendrites, endowed with several voltage-gated ion channels, form a prominent structural substrate for microdomain physiology, it is not known if these channels regulate the spatiotemporal spread of signaling microdomains. Here, we employed a multiscale, morphologically realistic, conductance-based model of the hippocampal pyramidal neuron that accounted for experimental details of electrical and calcium-dependent biochemical signaling. We activated synaptic N-Methyl-d-Aspartate receptors through theta-burst stimulation (TBS) or pairing (TBP) and assessed microdomain propagation along a signaling pathway that included calmodulin, calcium/calmodulin-dependent protein kinase II (CaMKII) and protein phosphatase 1. We found that the spatiotemporal spread of the TBS-evoked microdomain in phosphorylated CaMKII (pCaMKII) was amplified in comparison to that of the corresponding calcium microdomain. Next, we assessed the role of two dendritically expressed inactivating channels, one restorative (A-type potassium) and another regenerative (T-type calcium), by systematically varying their conductances. Whereas A-type potassium channels suppressed the spread of pCaMKII microdomains by altering the voltage response to TBS, T-type calcium channels enhanced this spread by modulating TBS-induced calcium influx without changing the voltage. Finally, we explored cross-dependencies of these channels with other model components, and demonstrated the heavy mutual interdependence of several biophysical and biochemical properties in regulating microdomains and their spread. Our conclusions unveil a pivotal role for dendritic voltage-gated ion channels in actively amplifying or suppressing biochemical signals and their spatiotemporal spread, with critical implications for clustered synaptic plasticity, robust information transfer and efficient neural coding.
| The spatiotemporal spread of biochemical signals in neurons and other cells regulate signaling specificity, tuning of signal propagation, along with specificity and clustering of adaptive plasticity. Theoretical and experimental studies have demonstrated a critical role for cellular morphology and the topology of signaling networks in regulating this spread. In this study, we add a significantly complex dimension to this narrative by demonstrating that voltage-gated ion channels on the plasma membrane could actively amplify or suppress the strength and spread of downstream signaling components. Given the expression of different ion channels with wide-ranging heterogeneity in gating kinetics, localization and density, our results point to an increase in complexity of and degeneracy in signaling spread, and unveil a powerful mechanism for regulating biochemical-signaling pathways across different cell types.
| Microdomains that emerge from spatially constricted spread of biochemical signaling components play a central role in defining several neuronal computations, including compartmentalization of neuronal plasticity and localized targeting of membrane components [1–10]. Theoretical and experimental studies have demonstrated that the spread of these microdomains are regulated by several biophysical and biochemical parameters. These parameters include the concentrations, localization profiles, binding and diffusion constants of the signaling components that are part of the signaling network, the morphological structure of the compartment, network topologies and feedback motifs [1–5,8,11]. Most such studies have considered neuronal dendrites, which form the prominent structural substrate for microdomain spread and physiology, to be passive structures that lack active dendritic conductances. Neuronal electrical signaling and physiology, however, is defined by the presence and plasticity of voltage-gated ion channel conductances, some of which are present at higher densities in the dendrites than at the cell body [9,12–24]. Additionally, consequent to their ability to significantly alter calcium influx into neuronal compartments, these active dendritic conductances are well established as critical regulators of synaptic plasticity profiles [12–14,19,25–35]. Given such pivotal role of active dendrites in determining neuronal physiology and plasticity, the question on whether active dendritic conductances regulate the spatiotemporal spread of signaling microdomains is important and has not been addressed.
To address this question, we employed a multiscale, multicompartmental, morphologically realistic, conductance-based model that accounted for the biophysics of electrical signaling [16–18,36,37] and the biochemistry of calcium handling [25,26,38–40] and downstream enzymatic signaling in a hippocampal pyramidal neuron. We chose the calcium–calmodulin–calcium/calmodulin-dependent protein kinase II (CaMKII)–protein phosphatase 1 (PP1) signaling pathway owing to its critical importance to several forms of neuronal plasticity [41–52], and employed physiologically relevant theta-burst stimulation (TBS) or theta-burst pairing (TBP) protocol [12,20,53–58] to initiate a calcium microdomain through N-Methyl-d-Aspartate receptor (NMDAR) activation at a synapse. We studied the spatiotemporal spread of calcium and other downstream microdomains in a dendritic segment compartmentalized to 2000 compartments, each spanning ~97-nm of length.
Using this setup, we assessed the role of two dendritically expressed inactivating conductances [16,18]—one restorative (A-type K+) and another regenerative (T-type Ca2+)—and showed that they respectively suppress and enhance the spread of phosphorylated CaMKII through different mechanisms. We also assessed the cross-dependencies of these two channels with other model components, and demonstrated the heavy mutual interdependence of several biophysical and biochemical properties in regulating microdomains and their spread. Our results provide compelling evidence for a critical role of active dendrites in regulating the spatiotemporal spread of signaling microdomains. These conclusions call for a marked rethink of the complexities associated with subcellular signaling networks, with future experiments focusing on the role of voltage-gated ion channels in tuning location-dependent signaling specificity and spread, in regulating robust information transfer and efficient encoding of afferent inputs in signaling networks, in regulating clustered plasticity of spatially-adjacent synapses on dendritic branches, and in behavioral state- and activity-dependent changes in such signal propagation.
Neuronal excitability is critically regulated by morphological as well as intracellular channel localization profiles [13–15,19,59–61]. As arbitrary choices for morphological properties and channel parameters would preclude extrapolations of our results to physiology, as a first step, we employed a hippocampal pyramidal neuron reconstruction and systematically matched the electrical properties of this model neuron with their electrophysiological counterparts (Fig 1A–1D). This was especially important because a goal of the study was to assess the dependence of signaling spread on neuronal excitability, therefore necessitating the excitability of the model to match measurements from hippocampal neurons across the somatodendritic arbor. The specific signaling pathway that we chose to assess in this study is depicted in Fig 1E. The rationale behind the choice of this pathway was its critical importance to several forms of neuronal plasticity [41–52]. Additionally, from a physiological standpoint, the spread of the CaMKII microdomain in this signaling pathway directly translates to the spread of neuronal plasticity through phosphorylation of several substrates that include 2-amino-3-(5-methyl-3-oxo-1,2-oxazol-4-yl) propanoic acid (AMPA) receptors and several ion channels [4,42,43,51,52,56,62]. As the assessment of microdomains requires the analyses of the spatiotemporal spread of each signaling component across the neuronal structure [4,7,8], we employed partial differential equations to model the reaction kinetics coupled to the diffusion of the individual components in the signaling pathway.
To accommodate the steep spatial decay of calcium, and the consequent requirement for finer spatial discretization when compared to discretization required for electrical simulations [63], we followed the established approach of employing a single oblique dendrite to assess signaling microdomains [8]. Although the analyses associated with signaling microdomains (Fig 2A) was performed in this oblique dendrite, its presence as part of the morphologically realistic electrical model ensured that the branching profiles and channel conductances required for maintaining excitability properties match their experimental counterparts. To accommodate fine spatial discretization, the oblique (of total length 193 μm) highlighted in Fig 2A was compartmentalized to 2000 compartments (making each compartment size to be around 97 nm). With the rest of the somatodendritic arbor discretized to accommodate electrical length constants, this spatial discretization procedure resulted in a total number of 2864 compartments in the entire neuronal structure.
We placed a single synapse containing both AMPA and NMDA receptors, whose kinetics and voltage-dependence properties (of NMDA receptors) were derived from electrophysiological measurements, at the center of this 100-μm region on the oblique. The AMPAR density was set such that the somatic unitary EPSP amplitude was ~0.2 mV, to match with experimental observations [64]. Consistent with our motivations of understanding signaling microdomains that are relevant to plasticity induction in the hippocampus, we stimulated this synapse with the well-established theta burst stimulation (TBS; Fig 2A) protocol that induces synaptic plasticity in hippocampal neurons [53]. As expected, the voltage traces obtained with this stimulation resulted in temporally summating excitatory postsynaptic potentials (EPSPs) that resulted in local dendritic, but not axosomatic spikes (Fig 2A). Although the recorded voltage traces were not very different across a span of 3 μm on either side of the synapse (Fig 2B), the calcium concentration (consequent to influx through NMDARs at the synapse) displayed sharp attenuation with distance, thereby establishing the calcium microdomain induced by TBS (Fig 2C and 2D).
Given the reaction-diffusion framework employed here (Fig 1E), this calcium microdomain propagated along the signaling network, manifesting as localized increases in the concentration of calcium-bound calmodulin, activated and phosphorylated CaMKII (pCaMKII) and as localized reductions in the concentrations of unbound calmodulin and non-activated CaMKII (Fig 2E–2G). As would be expected from the binding kinetics of the reactions, and especially by the autophosphorylation of CaMKII, there was an increase in the spatiotemporal spread of the microdomain associated with pCaMKII compared to that of calcium (compare Fig 2D with Fig 2F; see Fig 2G). These results quantitatively demonstrate that TBS induces localized calcium influx, which, through propagation along an established signaling pathway, results in a microdomain of pCaMKII with a spread larger than that of the calcium microdomain [41,43,47].
In order to look at how several key parameters of the model affect the micorodomain spatiotemporal kinetics, we performed sensitivity analyses involving different values of calcium diffusion constant (DCa), total concentrations of calmodulin ([CaM]T) and CaMKII ([CaMKII]T). We observed no significant change in the electrical response to TBS for different values assigned to each of these three parameters (Fig 3A), which was expected because all these parameters are involved in the signaling pathway downstream of electrical responses. However, the spatiotemporal dynamics of downstream signaling microdomains were sensitive to these parametric values (Fig 3A). Specifically, increase in DCa expectedly enhanced the spatial spread, also resulting in a reduction in peak calcium concentration (Fig 3A and 3B). Consequent to the enhanced spatial spread of calcium, the pCaMKII microdomains showed an enhanced spread, although there was little increase in peak [pCaMKII]max (Fig 3A and 3B). Upon increase in [CaM]T, as more free calcium was now bound to CaM, there was a small reduction in the peak values of [Ca]cmax (Fig 3A and 3B). As a direct consequence of the larger availability of calcium-bound calmodulin, there was a significant increase in the peak values of [pCaMKII] and in the spread of pCaMKII (Fig 3A and 3B). Varying the [CaMKII]T, however, did not alter the calcium or pCaMKII dynamics significantly across the tested range (Fig 3A and 3B). These results demonstrated that within the specific parametric configurations, the spatiotemporal evolution of pCaMKII was more sensitive to the concentration of calmodulin than on calcium diffusion or on total CaMKII concentration.
The presence and plasticity of A-type potassium (KA) channels in hippocampal neuronal dendrites and their roles in regulating dendritic excitability and synaptic plasticity profiles are well established [16,20–22,25–27,29,30,65–68]. Does the presence of or plasticity in KA channels in hippocampal dendrites alter the spread of microdomains in plasticity-inducing enzymes? To address this question, we performed the simulations described in Fig 2 with different densities of KA channels in the oblique specified in Fig 2A (see S1 Table for a figure-wise catalog of oblique parameters). As would be expected from the ability of KA channels to regulate synaptic and action potential amplitudes [16], we found that the voltage response to TBS was lower when the density of KA channels was increased (Fig 4A, 4D and 4E). This difference in voltage response directly translated to changes in the calcium influx through NMDARs (Fig 4B, 4D, 4F, 4H and 4I), and introduced small changes in the spread of the calcium microdomain, especially towards the terminal (Fig 4D and 4F).
Strikingly, this small increase in the spread of calcium microdomain was significantly amplified with propagation along the signaling pathway (Fig 4D, 4F and 4G). Specifically, although the peak pCAMKII response was not very different across different densities of KA channels (Fig 4G and 4J), the spread of the pCaMKII microdomain showed tremendous enhancement with reduction in the density of KA channels (Fig 4G and 4K).
How sensitive are our conclusions to changes in key model parameters? To answer this question, we performed sensitivity analyses to assess the impact of KA channels on pCaMKII microdomain spread with two-fold increase or decrease in several key parameters (Fig 5). Across different parametric combinations, the peak pCaMKII and its spread were consistently larger with lower values of KA channel density, thereby confirming that our conclusions were not restricted by the choice of default parametric values. The sensitivity analyses also demonstrated that the spread of microdomains was critically reliant on several parameters [8], pointing towards robustness of microdomain spread through degeneracy involving several biochemical and biophysical components [8,69]. Specifically, an increase in synaptic AMPAR (Fig 5A) or NMDAR (Fig 5B) densities, or the calcium diffusion constant (Fig 5E), or the autophosphorylation rate of CaMKII (Fig 5G) or the density of R-type calcium (CaR) channels (Fig 5H) enhanced the spread of pCaMKII microdomain. In contrast, an increase in the rate associated with the plasma membrane calcium pump (Fig 5C) or the Vmax of the SERCA pump (Fig 5D) or the total capacity of the calcium buffer (Fig 5F) suppressed the pCaMKII microdomain spread.
Together, these results unveil a pivotal role for dendritic A-type potassium channels in suppressing the spatiotemporal spread of microdomains in plasticity-inducing enzymes.
Low voltage-activated transient T-type calcium (CaT) channels, with their predominant dendritic presence, significantly alter synaptic integration, calcium influx and dendritic spike initiation in hippocampal pyramidal neurons [18,70,71]. Although the diverse roles of CaT channels in regulating neuronal physiology and plasticity have been explored [72–76], it is not known if these channels contribute to the spread of signaling microdomains of plasticity-inducing enzymes. Therefore, as a next step, we repeated the simulations described in Fig 2 with different densities of CaT channels in the oblique specified in Fig 2A. We found that the peak local voltage response to TBS was slightly higher at terminal end. Additionally, the spatiotemporal voltage response profile was not significantly different with various densities of CaT channels (Fig 6A and 6E), although the cell entered spontaneous spiking with very high density of these channels (e.g., Fig 6A; top panel). Although the peak calcium response was not very different with different densities of CaT channels (Fig 6B, 6F and 6H), there was a small increase in the spread of calcium microdomain with increase in CaT-channel density (Fig 6B, 6F and 6I).
Despite the absence of large changes in peak voltage and calcium and pCaMKII (Fig 6C, 6G and 6J) responses at the location of the synapse, we found a significantly large increase in the spread of the pCaMKII microdomain with increase in CaT-channel density (Fig 6C, 6G and 6K). Additionally, and in contrast to the case with KA channels (Fig 4G), this increase in spread was symmetric about the synaptic location, which was a consequence of the manner in which these channels altered the microdomain spread. Specifically, whereas KA channels altered the spread of microdomains by asymmetrically modulating the voltage and calcium responses (Fig 4E and 4F), CaT channels modulated the spread through symmetric changes in calcium spread (Fig 6F) without significant changes in voltage response (Fig 6E).
Next, we performed sensitivity analyses on our model to assess the impact of CaT channels on pCaMKII microdomain spread with two-fold increase or decrease in several key parameters (Fig 7). We found, across different parametric combinations, that the spread of pCaMKII was consistently larger with higher values of CaT channel density, thereby confirming that our conclusions were not restricted by the choice of default parametric values. Results from these sensitivity analyses with reference to individual parameters (Fig 7) corroborated our earlier conclusions (Fig 5) on their specific roles in altering pCaMKII microdomains, also providing further evidence for degeneracy in the spatiotemporal spread of signaling microdomains.
Together, our results provide compelling evidence for a critical role for dendritic T-type calcium channels in enhancing the spatiotemporal spread of microdomains in plasticity-inducing enzymes, brought about by increases in calcium influx through these channels.
As a next step in our analyses, instead of individually varying either KA (Figs 4 and 5) or CaT (Figs 6 and 7) channel densities, we varied both channels together in the oblique to different densities and repeated our simulations to assess the spread of pCaMKII microdomains (Fig 8). We performed these simulations at two different densities of CaR channels to assess the interactions between KA, CaT and CaR channels in regulating microdomains. Although results from this set of simulations were consistent with our overall conclusions that KA and CaT channels respectively suppress and enhance the pCaMKII microdomain spread, the quantitative changes observed were dependent on the other channels present in the dendritic branch. For instance, when CaR channels were absent, the impact of KA channels on pCaMKII and its spread (Fig 8A and 8C) was minimal compared to that when CaR channels were present (Fig 8B and 8D). In contrast, from the same set of figures, it may be noted that the impact of CaT channels on pCaMKII and its spread was minimal when CaR channels were absent than when they were present (especially when KA channels were absent). Together these results unveil a crucial role for active dendritic conductances in regulating the spatiotemporal spread of signaling microdomains. Additionally, in conjunction with the sensitivity analyses presented earlier (Fig 3, Fig 5, Fig 7), our conclusions underscore the heavy mutual interdependence of several biophysical and biochemical properties (that account for synaptic, intrinsic and kinetic parameters of several membrane and cytosolic signaling components) in regulating microdomains and their spatiotemporal spread along a signaling pathway.
Excitatory synapses typically impinge on a dendritic spine, which has been postulated to form a biochemical compartment for calcium and downstream signaling molecules [4,51,77–87]. In order to study the effect of the presence of spine on the signaling microdomains, we incorporated a spine consisting of a head and a neck on the dendritic shaft (Fig 9A) and recorded the spread of voltage, calcium and the downstream signaling molecules in all the three sections. The AMPAR density in the spine was set such that the somatic unitary EPSP amplitude was ~0.2 mV, to match with experimental observations [64]. This implied that the local spine and dendritic voltages were on the order of tens of mV for unitary EPSPs as well as during a TBS input (Fig 9A), which is consistent with large-amplitude spine and local-dendritic voltages recorded during unitary events [88–91]. We also noted that temporal summation during TBS elicited local dendritic spikes in the immediate vicinity of the synapse (Fig 9A), which does not propagate to the soma owing to attenuation during propagation [21].
Consistent with the role of dendritic spines as biochemical compartments, the calcium concentrations which were on the order of 10s of μM at the spine-head dropped significantly with propagation along the spine neck into the dendritic shaft. Specifically, there was almost a 100× attenuation of the calcium levels at the dendritic shaft when compared to that at the spine head (Fig 9B). Whereas the high concentrations of calcium at the spine head could be attributed to the high surface to volume ratio (SVR) of the spine compartments, the significant fall in propagating calcium was consequent to the calcium off mechanisms (i.e., calcium pumps and buffers) expressed in the spine compartments [77,92]. As a consequence of this significant attenuation, the dendritic shaft calcium concentration for a spine-localized synapse was lower than that when the same synapse (with identical receptors and activation dynamics) was located on the dendritic shaft (Fig 9C). Despite the large reduction in the dendritic calcium concentration as a consequence of spine localization, the corresponding reduction in pCaMKII was comparatively lower (Fig 9C). Overall, despite quantitative differences, the spatiotemporal spread of all molecular species across the signaling topography within the dendritic shaft with a spine-localized synapse (Fig 9D) was qualitatively comparable to the signaling spread with a dendrite-localized synapse (Fig 2G).
Our analyses thus far were performed in the absence of any background synaptic activity. Would the presence of spontaneous background synaptic activity alter the spatiotemporal spread of microdomains across the dendritic structure? To address this, we incorporated randomly activated balanced excitatory and inhibitory synapses throughout the dendritic arbor, resulting in fluctuating membrane potential dynamics (Fig 10A). We then compared the dynamics of voltage propagation and the spread of signaling molecules, across the spine and the dendritic shaft, in the presence or absence of the background synaptic activity (Fig 10B and 10D). Quantitatively, owing to the predominant dendritic presence of background excitatory synapses (Fig 10A), there was a small increase in the dendritic response voltage and the consequent calcium influx in the presence of a background voltage fluctuation (Fig 10B). Overall, although this small increase in dendritic calcium resulted in a minor enhancement of pCaMKII levels, the spatiotemporal kinetics of the microdomains and the dynamics of their spread were comparable across the spine and the dendrites, irrespective of the presence or the absence of background activity (Fig 10B–10D).
How do A-type K+ and T-type Ca2+ channels present on the dendrite regulate microdomain spread when synaptic stimulation arrives on a spine? To address this, we repeated our simulations described in Fig 9 with different densities of A-type K+ (Fig 11) and T-type Ca2+ (Fig 12) channels. We found that dendritic A-type potassium channels suppressed the spatiotemporal spread of pCaMKII microdomains (Fig 11), whereas dendritic T-type calcium channels enhanced the spatiotemporal spread of these downstream microdomains (Fig 12). As noted earlier, mechanistically, the A-type K+ channels regulate the spread of downstream microdomains by altering the voltage response whereas T-type Ca2+ channels act through the calcium influx they mediate, without significantly altering the voltage responses. As a consequence of this and because of the tremendous attenuation associated with calcium concentration when the synapse was localized on the spine (Fig 9), the impact of spine localization of the synapse was quantitatively different in the additional presence of these two channel subtypes (Fig 11 vs. Fig 12). Specifically, the impact of the additional presence of A-type K+ channels on pCaMKII spread (Fig 11) was significantly lower compared to that of T-type Ca2+ channels (Fig 12) when the synapse was localized on the spine, because of the respective indirect vs. direct roles of these channels in altering calcium concentration. Together, these results provide further evidence for active dendritic regulation of the spatiotemporal spread of microdomains in plasticity-inducing enzymes through distinct mechanisms and disparate dependencies on synaptic localization profiles.
Our analyses thus far were limited to the presence of a single synapse-containing spine that was placed at the center of the oblique dendrite. How would the microdomain spread and the impact of active dendrites on such spread be affected by the presence of other spine structures on the same dendritic structure? To directly address this question, we randomly placed spines spread across the dendritic oblique under consideration in various numbers (100, 200, 500 and 1000). Each of these additional spines had the same passive and active properties, as well as the calcium handling mechanisms, although none of them received any synaptic connections. With one of the distinct configurations with reference to the total number of spines on the oblique dendrite, we stimulated the central synapse-containing spine with the TBS protocol, and compared the spatiotemporal spread of microdomains at various spine densities. As a direct consequence of the overall increase in surface area and the active nature of the membrane, we found that the spatial spread of voltage increased with increase in spine density (Fig 13A and 13B), especially in the oblique dendrite for propagation towards the trunk. As a consequence of the presence of additional spines, similar to the impact of increased dendritic diameter on microdomain spread [8], we observed a dissipation of calcium and pCamKII microdomains consequent to TBS. Specifically, we noted that the peak values of dendritic [Ca] and [pCaMKII] resulting from TBS progressively decreased, and the corresponding spatial spread gradually enhanced, with systematic increase in spine density (Fig 13C).
How does the presence of active dendritic components alter microdomain propagation in the presence of spines? Do our conclusions on the role of KA and CaT channels change with the incorporation of dendritic spines? To test this, we repeated our analyses presented in Figs 11 and 12 with 1000 additional spines (~5 spines/μm [93]) randomly distributed on the oblique under consideration (Fig 14). We performed these analyses for distinct densities of KA (Fig 14A–14C) and CaT (Fig 14D and 14E) channel densities, with all other parameters (except for the incorporation of spines) remaining the same as the simulations performed to obtain Fig 11 (KA channels) and Fig 12 (CaT channels) respectively. We found our conclusions in the presence of these additional background spines to match with our earlier conclusions, whereby the presence of KA and CaT channels respectively suppressed and enhanced the signaling spread of dendritic calcium and pCaMKII (Fig 14).
Although TBS is a widely used LTP-induction protocol, a more robust version of the protocol involves pairing some or all synaptic stimulations as part of TBS with somatically initiated action potentials. The version of TBS where all stimulations are paired with appropriately timed action potentials has been referred to as theta-burst pairing (TBP), with the backpropagating action potentials invading a significant proportion of the dendritic tree allowing for enhanced calcium influx during the protocol [12,20,54–58]. Would the presence of paired backpropagating action potentials alter our conclusions in terms of the spread of downstream microdomains and on the impact of active dendritic conductances on such spread? To test this, we paired synaptic stimulations within TBS with backpropagating action potentials (Fig 15A), and assessed the impact of such TBP on the spread of downstream microdomains.
As expected from the backpropagation of action potentials, there was a small increase in the voltage responses observed in the oblique dendrite during TBP as compared to responses during TBS (Fig 15B). Although the backpropagating action potential amplitude was large on the dendritic trunk (Fig 1C), the invasion of backpropagating action potentials into obliques was minimal because of several factors including the branching structure of the dendritic arbor where obliques are made of smaller diameters, the expression of A-type K+ channels in obliques, and the slow recovery of dendritic sodium channels from inactivation [21,22,36,67,94]. As a consequence of this small increase in voltage response, we found small increases in the calcium and pCaMKII concentrations downstream in the spine and dendritic compartments (Fig 15B). These small changes in signaling concentrations also reflected in the spread across the different dendritic segment, overall suggesting that the impact of paired backpropagating action potentials on microdomain spread in oblique dendrites to be minimal (Fig 15C–15E).
Next, we assessed the specific contributions of the active dendritic conductances in regulating the spread of downstream microdomains with TBS or TBP as the plasticity induction protocol. Specifically, we virtually knocked out (by setting the corresponding conductance to zero, only in the dendritic and spine compartments under consideration) either A-type K+ or T-type Ca2+ channels and reassessed the spread of microdomains across the signaling pathway (Fig 15C–15E). Consistent with our prior observations (Figs 4–8, Figs 11 and 12), we found that knocking out A-type K+ or T-type Ca2+ channels respectively enhanced and reduced the peak values and the spread of the calcium and pCaMKII microdomains (Fig 15C–15E). The small differences between TBP and TBS observed earlier, in terms of TBP eliciting a slightly larger calcium response were reflected in both knockout simulations as well (Fig 15D and 15E).
Together, our results provide compelling evidence for a critical role for dendritic channels in regulating the spatiotemporal spread of microdomains in plasticity-inducing enzymes, effectuated by changes in excitability and/or calcium influx, in a manner that was invariant to several structural and parametric configurations.
The prime conclusion of this study is that active dendritic conductances play a critical role in regulating the spatiotemporal spread of microdomains associated with plasticity-inducing kinases. We demonstrated this by employing theta-burst synaptic stimulation to a multiscale multicompartmental model that was biochemically and biophysically constrained by experimental measurements. We studied the impact of two inactivating conductances with predominantly dendritic localization profiles, the restorative KA and the regenerative CaT conductances, and showed that they modulate microdomain spread through two distinct mechanisms. Whereas KA channels regulated the spread of pCaMKII microdomains by altering the voltage response to the theta burst stimulus, CaT channels regulated this spread by modulating the calcium influx consequent to TBS without significantly changing the response voltage. Finally, assessing the cross-interactions of KA, CaR and CaT channels (Fig 8) along with their interactions with key structural, biophysical and biochemical parameters (Fig 3, Fig 5, Fig 7, Figs 9–15), we demonstrated the heavy mutual interdependence of several model components in regulating signaling microdomains. Our conclusions unveil a critical role for active dendrites in regulating the spatiotemporal spread of signaling microdomains associated with subcellular molecular networks. Given the physiologically constrained approach that we have employed in this study, our results are also predictions that could be experimentally tested by measuring pCaMKII [43] after TBS or TBP (to different locations along the somatodendritic arbor) in the presence of pharmacological agents to block different channels. Finally, as several cell types express voltage-gated ion channels, and biochemical signaling strength and spread are ubiquitous in their regulatory capacity, our conclusion that voltage-gated ion channels that are present on the plasma membrane could regulate biochemical signaling strength and spread has implications that are not limited only to neurons. In what follows, we discuss the several implications of our conclusions for neuronal physiology and plasticity, and elucidate potential future directions.
Our results clearly demonstrate that signaling spread is an active process that is not just governed by cell morphology and the topological motifs and binding kinetics associated with the signaling network [5,8,95–97], but also by the types, kinetics and localization of ion channels in the dendritic arbor. This active amplification (Fig 6, Fig 11, Fig 15) or suppression (Fig 4, Figs 12–15) of biochemical signals and their spread by dendritic ion channel conductances call for a marked rethink of the complexities associated with subcellular signaling networks. Specifically, the numbers associated with ion channel subtypes, their auxiliary subunits, their subcellular localization profiles, and their local or global modulation through neuromodulatory substances or activity-dependent plasticity are staggeringly combinatorial [9,13–15,23]. The additional regulatory capacity of complex dynamics of subcellular signaling by this complex channel network implies a manifold increase in the complexity of molecular signal transduction. These results also clearly demonstrate several disparate combinations of neuronal parameters—associated with morphology, background spine density, signaling motifs, binding kinetics, diffusion and ion channel densities, for instance—could result in similar signal propagation in a given signaling network, pointing directly to degeneracy in spatiotemporal spread of signaling components [98]. In conjunction with recent literature on degeneracy in cellular-scale physiology and plasticity in hippocampal neurons [24,25,39,99–101], these results point to significant degeneracy in hippocampal physiology spanning different scales.
What are the consequences of such combinatorial complexities and degeneracy associated with signaling networks? First, a growing body of established literature that spans several scales of biology has linked complexity and degeneracy as requirements for robustness in biological systems [69,98,102]. In this context, the degeneracy associated with signaling spread could be postulated as a mechanism for achieving robust signaling transfer and spread in the presence of external and internal noise factors. Our demonstration of active suppression or amplification of signaling strength and spread by specific channels provide an additional mechanism by which noise could be selectively suppressed through channels with specific kinetic and voltage-dependence properties. Second, theoretical frameworks at the cellular scale have argued for efficient coding of incoming information [103–105] through ion channel localization and plasticity [14,28,55,106,107], which at the molecular scale has found reflection in terms of maximizing information transfer by matching signaling dynamics to input source statistics [108–111]. The results described here provide a way to unify these two apparently disparate theoretical frameworks (in two different scales) by showing their convergence towards regulation of signal strength and spread. Future studies should endeavor to holistically unify the systems [103–105], cellular [14,28,55,106,107] and molecular [108–110] versions of the efficient coding hypothesis, accounting for input statistics and neuronal response properties at all scales. Third, tunability of information transfer is a critical requirement in several signaling networks [112,113]. With active conductances modulating signaling strength and spread, it is clear that the specific signal that is transmitted would now be dependent on the postsynaptic channel densities (along the pathway of spatial propagation) as well, and not just on the input stimulus and the signaling motifs, thus providing an additional regulatory mechanism for tuning signaling specificity and spread.
Given the location-dependent expression profiles of different ion channels, the dependence of signaling spread on active dendritic conductances could directly translate to location-dependence of signaling spread. Specifically, the density of A-type K+ channels is higher in distal dendrites implying a significant suppression of the spread of signaling at distal locations. However, as different channel conductances have very different channel localization profiles and neuronal physiology is an emergent outcome of intricate and complex spatial and kinetic interactions between these different channels [9,13–15,23,39,99,114], the spread of downstream signaling molecules would also be determined by these interactions (Fig 3, Fig 5, Figs 7–15). Further, as several of these channels have non-uniform distributions, our results imply that the spread of downstream signaling microdomains might not be symmetric with reference to the synaptic location (the point of origin of the second messenger). Such a scenario provides a putative mechanism for spatiotemporally steering the spread and specificity of downstream signaling by regulating ion channel properties and localization profiles. Finally, given that oblique dendritic branches could have different branch strengths as a consequence of differences in A-type K+ channel expression [22], our results present a testable prediction that CaMKII-dependent plasticity could spread to channels and receptors located over larger distances in branches with lower A-type K+ channel expression (oblique dendrites with higher branch strength). In branches where the A-type K+ channel density is higher, on the other hand, could have the plasticity confined to a much smaller region owing to the constricted spread of the CaMKII microdomains. In this context, systematic analyses of the extent of spatiotemporal influence of different ion channel clusters on signaling microdomains, and of the dependence of such influence fields on the inhomogeneous distribution of different ion channels, the morphology of the dendritic arbor, the direction of propagation of voltage signals, the presence of background synaptic activity and the specific location of the channel in the dendritic arbor (e.g., on the trunk vs. on the thinner obliques) would provide further quantitative insights into the roles of active dendrites on the spread of microdomains [114].
Our study presents the possibility of location-dependent expression profiles of channels and their impact on voltage and calcium signals as potential mechanisms to steer downstream signaling molecules. However, quantitative links between voltage recordings, calcium transients and the spatiotemporal spread of downstream microdomains should not be generalized without specific analyses of the channels and the signaling components expressed in a specific system. First, although voltage transients provide one trigger for cytosolic calcium influx, they are not the only source of calcium, with the ER and other store-operated mechanisms playing a role in regulating calcium influx [26,115–120]. Second, the calcium transients (both amplitude and spread) are critically regulated by several factors (Figs 3–14) including surface-area-to-volume ratio, spine densities in specific dendritic arbors, the densities of channels, receptors, several pumps, transporters and buffer concentrations [26,77,79,81,87,92,121]. Therefore, factors such as altered surface-area-to-volume ratio and nonhomogeneous distribution of any of these components would critically affect calcium amplitude and spread, and alter calcium transients and downstream signaling independent of changes in voltage transients [100]. Third, the signaling spread of downstream molecular species is not a simple function of voltage and calcium transients, but is also critically dependent on several factors including surface-area-to-volume ratio of the compartment, background spine densities, the binding affinities, diffusion and subcellular localization of the different signaling components, the topology of the signaling cascade and on the presence of negative regulators upstream [2,4,8,11]. Finally, all these components—the ones that govern the voltage and calcium transients and those that govern the downstream signaling—critically interact with each other through several routes (Fig 3, Fig 5, Figs 7–15), implying a complex parametric and interactional landscape created by the presence of active dendritic components in regulating signaling microdomains.
Based on these observations, we also postulate that active dendrites constitute a putative mechanism to regulate clustered plasticity, a phenomenon where spatially adjacent synapses undergo concurrent plasticity, in dendritic branches [122–127]. Under such a postulate, activity-dependent plasticity [13–15,19,20,22] and/or state-dependent neuromodulation [65,72,128,129] of active dendritic conductances could control the degree of compartmentalization of experience-dependent synaptic plasticity on specific dendrites, thereby regulating the degree of clustering of functional synaptic inputs [123,124,126,127]. Active dendritic conductances, especially restorative conductances, are therefore very critical in confining the spatiotemporal spread of plasticity, thereby assigning a dendritic branch as a fundamental functional unit of biophysical and biochemical signal integration [122–127]. Finally, such regulation of signaling spread by active dendrites, coupled with well-established plasticity and modulation in these dendritic conductances [13–15,19,20,22,65,72,128,129] imply that signaling spread in any signaling molecule in a dendritic branch is dynamic and state-dependent, and that it would be inappropriate to assign a static picture for such a complex dynamical system. This dynamical spread in signaling microdomains has to be assessed accounting for morphological properties of the structure, the network topology, the binding and diffusion kinetics of each signaling component, active dendritic conductances and their properties, the different substrates for the plasticity-inducing enzymes, the localization of all these components, and importantly behavioral state- and activity-dependent modulation and/or plasticity in each of these components.
In accommodating the significant computational complexity associated with a reaction-diffusion system with stringent requirements on spatial discretization into a morphologically realistic model, we restricted our attention in this study to only a few channel types that are expressed in hippocampal dendrites. Although our analyses provide compelling evidence for a critical role for dendritic ion channels in regulating signaling spread, future studies should systematically characterize the impact of these channel types, including the HCN, calcium-activated potassium and inwardly rectifying potassium channels, on different signaling pathways. Another limitation of our analyses is the absence of metabotropic receptors and calcium-induced calcium release (CICR) and mechanisms associated with the activation of these metabotropic receptors. Specifically, it is established that CICR and other store-related mechanisms significantly interact with plasma membrane ion channels in yielding a complex landscape for the passive and active propagation of calcium within the cytosol [26,115–120]. These CICR mechanisms are critically tied to the activation of specific metabotropic receptors, and have been shown to be involved in certain forms of plasticity through specific signaling cascades [130,131]. Future studies should incorporate these ER-related mechanisms, metabotropic receptors and the complex interactions between dendritic ion channels and ER mechanisms [26,132] in deriving more routes for active-dendritic regulation of signaling microdomains through such interactions.
Although our choice of the signaling pathway was motivated by its physiological relevance to plasticity induction and spread, and by the requirement to reduce computational complexity, the pathway is significantly oversimplified from the standpoint of known complexities in signaling motifs and pathways [4,95–97]. Given the possibility that different ion channels and their spatial and kinetic interactions [39,99,114] could differentially interact with different biochemical network motifs in different morphological structures [7,8], future studies should assess such multiscale interactions towards robustness to internal/external noise, increases in information transmission and storage capacity, efficient encoding of afferent signals and towards tunability of signaling specificity and spread [14,28,39,55,69,98,99,102–114]. From a generic standpoint, the basic conclusion of our analyses on a critical role for voltage-gated ion channels and their structural/functional interactions with the different signaling components in regulating signaling microdomains is extendible to other neuronal structures expressing active dendrites, and even to other cell types expressing voltage-gated ion channels. However, given the critical dependencies on these conclusions on the specific channels and their distributions, on cellular morphologies and on the topology of the signaling cascade, future studies should build cell-specific multiscale models that expand on the analyses presented here. Finally, our reaction-diffusion model is a deterministic system that coupled compartmental modeling with partial differential equations (of continuous concentrations of signaling species) to assess the spread of microdomains. Although such continuous deterministic models have provided significant insights into biological signal transduction, a more realistic approach to the problem would be to employ a stochastic discrete system (involving number of molecules of signaling species) that would mimic the stochastic interactions of individual molecules within biological systems [4,5,11,133–135].
We employed a multiscale, multicompartmental, morphologically realistic, conductance-based model that accounted for the biophysics of electrical signaling and the biochemistry of calcium handling and downstream enzymatic signaling in a hippocampal neuron. Parameters associated with these were derived from electrophysiological and biochemical data from hippocampal pyramidal neurons. The biophysical model for electrical signaling and the models for calcium on and off mechanisms, including diffusion, were adapted from previous literature [25,26,38–40]. The signaling pathway, and the biochemical models for enzymatic signaling downstream of calcium were adopted from [10,41–52,97,136–138].
A morphologically realistic multicompartmental 3D model (Fig 1A) was constructed from a reconstructed CA1 pyramidal neuron morphology (n123) taken from the Neuromorpho database [139,140]. Passive parameters were set as follows: Cm = 1 μF/cm2; Rm and Ra for various compartments along the somato-apical trunk were functions of radial distance of the compartment from the soma, x [55]:
Ra(x)=Rasom+(Raend−Rasom)1+exp(300−x50)Ω.cm
(1)
Rm(x)=Rmsom+(Rmend−Rmsom)1+exp(300−x50)kΩ.cm2
(2)
where Rmsom = 125 kΩ.cm2 and Rasom = 120 Ω.cm were values at the soma, and Rmend = 85 kΩ.cm2 and Raend = 70 Ω.cm were values assigned to the terminal end of the apical trunk (which was ~425 μm away from the soma for the reconstruction under consideration). The non-uniformity of passive properties considered here follows from evidence from the literature that has argued for the necessity of such non-uniformity to match electrophysiological measurements [55,114,118,141–143], and has been specifically employed to match passive input resistance of the somato-apical trunk [39]. The basal dendrites, the axonal compartments, and apical obliques had somatic Rm (Rmsom = 125 kΩ.cm2) and Ra (Rasom = 120 Ω.cm) [39]. Except for the oblique where the signaling microdomains were assessed (Fig 2A) this neuronal model was compartmentalized using the dλ rule [144] to ensure that each compartment was smaller than 0.1 λ100, where λ100 is the space constant, computed at 100 Hz for the section under consideration. This process resulted in 873 compartments for the entire neuronal structure. As eliminating numerical errors in the estimation of Ca2+ signals (whose space constant is on the order of 0.5 μm) requires much smaller compartment sizes than such electrical compartmentalization [26,63,145], the oblique (of total length 193 μm) highlighted in Fig 2A was compartmentalized to 2000 compartments (making each compartment size to be around 97 nm). This spatial discretization procedure resulted in a total number of 2864 compartments in the neuronal structure.
Six different types of voltage gated ion channels (VGIC) were incorporated into these models: a fast Na+ (NaF), a delayed rectifier K+ (KDR), a hyperpolarization-activated cyclic-nucleotide-gated non-specific cationic (HCN), an A-type K+ (KA) and R- (CaR) and T-type Ca2+ (CaT). Biophysically realistic, Hodgkin-Huxley type conductance-based models derived from hippocampal pyramidal neurons were employed for modeling all these channels. The kinetics, voltage-dependencies and subcellular localization profiles of these channels were derived from hippocampal pyramidal neurons, and the details are provided below and in S1 Text.
A canonical glutamate synapse consisting of colocalized NMDARs and AMPARs was placed at the mid point of a proximal oblique represented in Fig 2A. Specifically, the NMDAR current was modeled as a combination of three different types of ionic currents namely Ca2+, Na+ and K+ [29]:
INMDA(v,t)=INMDANa(v,t)+INMDAK(v,t)+INMDACa(v,t)
(7)
where,
INMDANa(v,t)=P¯NMDAPNas(t)MgB(v)vF2RT([Na]i−[Na]oexp(−vFRT)1−exp(−vFRT))
(8)
INMDAK(v,t)=P¯NMDAPKs(t)MgB(v)vF2RT([K]i−[K]oexp(−vFRT)1−exp(−vFRT))
(9)
INMDACa(v,t)=P¯NMDAPCas(t)MgB(v)4vF2RT([Ca]i−[Ca]oexp(−2vFRT)1−exp(−2vFRT))
(10)
where P¯NMDA is the maximum permeability of the NMDAR; PCa = 10.6, PNa = 1, PK = 1 [159,160]. Default values of concentrations were (in mM): [Na]i = 18, [Na]o = 140, [K]i = 140, [K]o = 5, [Ca]i = 50 × 10−6, [Ca]o = 2. The concentrations for sodium set its equilibrium potential at +55 mV and that for potassium at –90 mV. MgB(v) governs the magnesium dependence of the NMDAR current, given as [161]:
MgB(v)=(1+[Mg]oexp(−0.062v)3.57)−1
(11)
with the default value of [Mg]o set at 2 mM. s(t) governed the kinetics of the NMDAR current, and was set as:
s(t)=a(exp(−tτd)−exp(−tτr))
(12)
where a is a normalization constant, making sure that 0 ≤ s(t) ≤ 1,τd is the decay time constant, τr is rise time, with τr = 5 ms, and τd = 50 ms.
Current through the AMPAR was modeled as the sum of currents carried by sodium and potassium ions:
IAMPA(v,t)=IAMPANa(v,t)+IAMPAK(v,t)
(13)
where,
IAMPANa(v,t)=P¯AMPAPNas(t)vF2RT([Na]i−[Na]oexp(−vFRT)1−exp(−vFRT))
(14)
IAMPAK(v,t)=P¯AMPAPKs(t)vF2RT([K]i−[K]oexp(−vFRT)1−exp(−vFRT))
(15)
where P¯AMPA is the maximum permeability of the AMPAR, whose default value was set at 20 nm/s, yielding a unitary voltage response of around 0.3 mV. PNa was taken to be equal to PK [162]. s(t) was the same as that for the NMDA receptor, but with τr = 2 ms and τd = 10 ms [29]. P¯NMDA=NAR×P¯AMPA, with a default value for the NMDAR:AMPAR ratio (NAR) set at 1.5.
We employed theta burst stimulation (TBS), an established protocol for induction of synaptic plasticity [53], to assess the spread of signaling microdomains in the model. For TBS, the synapse was stimulated with a burst of 5 action potentials at 100 Hz, and this burst was repeated 30 times at 200 ms interval (5 Hz; theta frequency) each (Fig 2A). To analyze the effect of back-propagating action potentials (bAPs) initiated during the induction protocol on the spatio-temporal dynamics of microdomains, we used the theta burst pairing (TBP) protocol that has been employed for inducing neuronal plasticity [20,54,55]. In this protocol, each synaptic stimulation pulse during TBS was followed by a current pulse injection at the soma (current clamp amplitude = 2.5 nA; duration of each pulse = 1 ms, each of which initiates an axo-somatic action potential that backpropagates into the dendritic arbor as well), with a time lag of 5 ms (Fig 15A). This led to the TBS to be paired with theta-burst firing (TBF), together resulting in the TBP protocol [20,54,55].
The overall Ca2+ dynamics was modeled as a function of various mechanisms that lead to changes in cytosolic Ca2+, [Ca]c, within a neuron. The partial differential equation governing the cytosolic Ca2+ dynamics was [26,38,163,164]:
∂[Ca]∂t=Dca∇2[Ca]+β(Jleak−JSERCA)+Rbuf+JCC−Jpump
(16)
where DCa represents experimentally determined diffusion coefficient for Ca2+ [165,166] and β constitutes the density of SERCA pumps and leak channels on the ER along the somato-dendritic axis. JCC, JSERCA, Rbuf, Jpump and Jleak represent the Ca2+ flux due to calcium channels, SERCA pumps, buffering, plasma membrane Ca2+ extrusion pumps and the ER leak channels, respectively. Radial diffusion of Ca2+ was incorporated by radial compartmentalization of the neuronal compartments into 4 concentric annuli, and diffusion along longitudinal axis of the neuron was also implemented [144]. The concentrations of individual molecular species (e.g., calcium, calmodulin) are reported for the outermost annulus of different longitudinal compartments. Detailed descriptions for each of the fluxes are presented below, and a list of all parameters employed for modeling calcium dynamics, with values derived from previous studies [26,38,163,164], are listed in Table 1.
ER leak channels and SERCA pump. The rate of Ca2+ influx into the cytoplasm through ER leak channels was modeled as [163]:
Jleak=L(1−[Ca][Ca]ER)mM/ms
(17)
where L was chosen such that there was no net flux of Ca2+ between the ER and the cytosol at resting state.
The Ca2+ uptake by sarcoplasmic endoplasmic reticulum Ca2+ ATPase (SERCA) pump was modeled as [163]:
JSERCA=Vmax[Ca]2[Ca]2+KP2mM/ms
(18)
where Vmax depicts the maximal rate of pump and KP is dissociation constant of Ca2+ binding to the pump.
The signaling pathway that was assessed in this study is pictorially represented in Fig 1E [41–52]. The specific binding reactions are listed below, with all associated parameters tabulated as Table 2 [10,97,136–138]. First, we modeled the binding of cytosolic calcium (Ca) to calmodulin (CaM) to form the calcium-calmodulin complex (CaMCa4) as:
The forward and backward reaction rates were specified as kCaMon and kCaMon×KCaM, respectively, with KCaM defining the dissociation constant related to this binding. Then, the binding of CaMCa4 to CaMKII (CaMKII) to form a complex (CaMKII_CaMCa4) complex was modeled as:
CaMCa4+CaMKII⇋CaMKII_CaMCa4
(25)
The forward and backward reaction rates were specified as kCaMKIIon and kCaMKIIon×KCaMKII, respectively, with KCaMKII defining the dissociation constant related to this binding. Note that CaMKII_CaMCa4 was alternately represented as CaMKII*, with the * representing activation (e.g., Fig 2G). Autophosphorylation of CaMKII was modeled using simple first order kinetics [45,48,169] as:
CaMKII_CaMCa4→pCaMKII_CaMCa4
(26)
with the associated reaction rate specified as kCaMKIIAuto. Note that CaMKII_CaMCa4 was alternately represented as pCaMKII*, with the p representing phosphorylation (e.g., Fig 2G). Finally, the dephosphorylation of CaMKII [50] occurred through protein phosphatase 1 (PP1):
pCaMKII_CaMCa4+PP1→CaMKII_CaMCa4
(27)
with the associated reaction rate specified as kPP1. Apart from the radial and longitudinal diffusion of calcium (Eq 16), signaling components represented in Eqs 24–26 also were subjected to radial as well as longitudinal diffusion:
∂[X]∂t=DX∇2[X]
(28)
where X represented any of CaM, CaMCa4, CaMKII, CaMKII_CaMCa4 or pCaMKII_CaMCa4. As initial conditions, CaM was set to the total calmodulin concentration, [CaM]T, CaMKII was set to the total CaMKII concentration, [CaMKII]T, with CaMCa4, CaMKII_CaMCa4 and pCaMKII_CaMCa4 set to zero. The total concentrations and the diffusion constants for these signaling components are listed in Table 2.
Electrical measurements from the model were recorded employing established procedures [39,55] and are detailed below. To measure the backpropagation of action potentials into dendrites [16,39,147], an action potential was initiated at the soma (2 nA current for 1 ms) and the amplitude of the backpropagating action potential (bAP) was measured at various locations along the somatoapical trunk (Fig 1C). Input resistance (Rin) was measured from the local voltage response to a local injection of a 100-pA hyperpolarizing current pulse. The ratio of the steady-state voltage response to the injected current amplitude was taken as the Rin for that location, and the procedure was repeated for all locations along the somatoapical trunk to construct the Rin functional map (Fig 1D).
To avoid ambiguities with reference to distance representations, distance from the synapse toward the terminal and trunk were designated positive and negative values, respectively (Fig 2A and 2B). The spatiotemporal spread of each signaling component was represented as a kymograph with the X-axis representing time, the Y-axis representing space (distance from the synapse towards the trunk and terminal), with the color code representing the numerical value of the component that was plotted. The spread of signaling microdomains through the analyzed signal pathway was depicted as a flow chart of kymographs [8]. The extent of the microdomain was quantified as the area under the curve (AUC) of the plot depicting the maximum value of the signaling component as a function of distance from the synapse (e.g., Fig 2F). In computing the AUC for individual compartments, whereas the entire spatial stretch of the compartment was employed for the spine head and spine neck, the dendritic AUC was computed over the span of 50 μm on either side of the synaptic (or spine) location. When the signaling spread was quantified for different parametric configurations, these plots were computed for each parametric configuration (e.g., Fig 4G) and the AUC values obtained from these plots were assessed as a function of the parameter that was being varied (e.g., Fig 4K). With such quantification, an increase or a decrease in the computed AUC would be a measure of enhancement or suppression, respectively, in the spatial spread of the corresponding signaling microdomain. In performing sensitivity analyses with reference to several critical parameters, the default value associated with each parameter was either increased or decreased two-fold (e.g., Fig 5) to assess the impact of such a change in the microdomain spread (quantified as the AUC mentioned above).
For simulations where the synapse was localized on a spine (Figs 9–15), a spine neck (length 1 μm × 0.1 μm diameter) connected to a spine head (length 0.5 μm × 0.5 μm diameter) were added at the center of the oblique dendritic shaft (Fig 9A). The spine-head had 10 compartments whereas the spine-neck had 20 compartments, making the size of each compartment to be ~50 nm. The spine had the same passive and active conductances as that of the center of the oblique dendrite from which it originated: Rm = 125 kΩ.cm2, Ra = 120 Ω.cm, g¯Na = 16 mS/cm2, g¯KDR = 10 mS/cm2, g¯KA = 60.55 mS/cm2, g¯h = 68.75 μS/cm2, g¯CaT = 285.7 μS/cm2. A single synapse containing colocalized AMPA and NMDA receptors was placed at the center of the spine head. The AMPAR permeability was set such that the unitary EPSP amplitude at the soma was ~0.2 mV to match experimental observations [64,170,171], and was set at P¯AMPA = 15 nm/s (somatic voltage with spine: 0.22 mV; somatic voltage without spine: 0.25 mV). The NMDAR permeability was set at 1.5× of the AMPAR permeability.
To study the impact of spine density on the spatiotemporal spread of biochemical microdomains, we incorporated several spines throughout the oblique dendrite under consideration. Each of these spines had the same morphology as the synapse-containing spine (Fig 9A): spine-neck (length 1 μm × 0.1 μm diameter) connected to a spine-head (length 0.5 μm × 0.5 μm diameter). Each spine had the same passive and active conductances as that of the oblique dendritic compartment from which it originated, including the calcium handling mechanisms. To compare the effect of different densities of spines on the spatiotemporal spread of biochemical microdomains in active dendrites, we compared simulation outcomes in a control (one synapse-containing spine; ~0 spine/μm otherwise) outcomes where the oblique dendrite was populated with spines at four distinct densities. Specifically, the four other cases were built with 100, 200, 500 and 1000 spines (corresponding to ~0.5 spines/μm, ~1 spine/μm, ~2.6 spines/μm and ~5 spines/μm, respectively) distributed randomly (compartments chosen from a uniform distribution) throughout the 2000 compartments of the 193-μm length of the dendritic oblique. All spines except for the synapse-carrying spine (Fig 9A) were devoid of any synaptic connections. With this morphological configuration that reflected the characteristics of a hippocampal pyramidal neuron dendrite, we applied TBS through the central synapse-containing spine and looked at the effects of the spatiotemporal kinetics of each species in our chosen biochemical pathway with different spine densities. We tested the effects of A-type K+ channel and T-type Ca2+ channel densities for the 1000 spine (density = ~5 spines/μm) case, as this spine density represented the closest approximation to experimental evidence [93].
For simulating background synaptic activity impinging on the neuron, we incorporated balanced excitation and inhibition so as to keep the average resting membrane potential (RMP) at around –65 mV [172]. One excitatory synapse was placed at each compartment of the somato-apical dendritic arbor within a 300 μm radial distance. Similarly, one inhibitory synapse was placed at each compartment within a radial distance of 50 μm perisomatically, including both apical and basal segments. For both the excitatory and inhibitory synaptic populations, independent random spike generators, each firing at an average rate of 4 Hz was used for input stimulation of each synapse. All the synapses were modeled using an Ohmic formulation with the current through the synapse defined as:
isyn(t)=gsyn(t)(V−ER)
(29)
where gsyn (t) defined the time-dependent evolution of each synapse after the onset of an afferent spike, and ER defined the reversal potential for the synaptic receptors (ER = 0 mV for excitatory synapses and ER = –80 mV for inhibitory synapses). gsyn (t) was modeled using a double exponential synaptic formulation:
gsyn(t)=g¯[exp(−t/τd)−exp(−t/τr)]
(30)
where g¯ defined the maximal conductance of each synapse set at 0.1 nS for excitatory synapses and 0.6 nS for inhibitory synapses. τr (= 2 ms) was the synaptic rise time constant and τd (= 10 ms) was the decay time constant for all the synapses. Upon stimulation with such randomized background activity, the mean somatic RMP was found to be –64.33 mV ± 0.74 mV (Fig 10A).
All simulations were performed in the NEURON simulation environment [144]. The resting membrane potential of the model neuron was fixed at –65 mV. For all experiments, the simulation temperature was set at 34°C and ion channel kinetics were appropriately adjusted according to their experimentally determined Q10 coefficients. The integration time step was fixed at 25 μs for all simulations to avoid numerical errors in the solution to the differential equations. Data analysis was performed using custom-built software under the IGOR Pro (Wavemetrics Inc., USA) programming environment. The NEURON codes employed to perform the simulations reported in this article are available at the following URL: https://senselab.med.yale.edu/ModelDB/ShowModel.cshtml?model=244848. An updated version of the code in the website fixes and accounts for a volume-scaling bug (See S1 Fig).
All biochemical reactions involving Ca, CaM and CaMKII, their forward rate constants and dissociation constants are listed in S2 Table.
|
10.1371/journal.pbio.1000068 | Quantifying the Integration of Quorum-Sensing Signals with Single-Cell Resolution | Cell-to-cell communication in bacteria is a process known as quorum sensing that relies on the production, detection, and response to the extracellular accumulation of signaling molecules called autoinducers. Often, bacteria use multiple autoinducers to obtain information about the vicinal cell density. However, how cells integrate and interpret the information contained within multiple autoinducers remains a mystery. Using single-cell fluorescence microscopy, we quantified the signaling responses to and analyzed the integration of multiple autoinducers by the model quorum-sensing bacterium Vibrio harveyi. Our results revealed that signals from two distinct autoinducers, AI-1 and AI-2, are combined strictly additively in a shared phosphorelay pathway, with each autoinducer contributing nearly equally to the total response. We found a coherent response across the population with little cell-to-cell variation, indicating that the entire population of cells can reliably distinguish several distinct conditions of external autoinducer concentration. We speculate that the use of multiple autoinducers allows a growing population of cells to synchronize gene expression during a series of distinct developmental stages.
| Although bacteria are unicellular, the individual cells communicate with each other via small diffusible molecules. This communication process, known as quorum sensing, allows groups of bacteria to track the density of the population they are in, synchronize gene expression across the population, and thereby carry out collective activities similar to those of cells in multi-cellular organisms. Many bacterial species use multiple signaling molecules, but it remains a mystery why multiple signals are required and how the information encoded in them is integrated by bacteria. To explore these questions, we studied a model quorum-sensing bacterium Vibrio harveyi. Using single-cell fluorescence microscopy, we quantified quorum-sensing responses and analyzed the mechanism of integration of multiple signals. Surprisingly, we found that information from two distinct signals is combined strictly additively, with precisely equal weight from each signal. Our results revealed a coherent response across the population with little cell-to-cell variation, allowing the entire population of bacterial cells to reliably distinguish multiple environmental states. We argue that multiple signals and multiple response states could be used to distinguish distinct stages in the development of a bacterial community.
| In a process called quorum sensing, bacteria communicate with one another using extracellular signaling molecules called autoinducers. Quorum sensing allows groups of bacteria to track their cell numbers, synchronize gene expression on a population-wide scale, and thereby carry out collective activities. In quorum sensing, bacteria produce, release, and detect autoinducers that accumulate in a cell-density–dependent manner, and, thus, autoinducer concentration serves as a proxy for cell number. Quorum-sensing systems are widespread in the bacterial world, existing in both Gram-negative and Gram-positive bacteria, and quorum sensing is used to control such diverse functions as bioluminescence, virulence-factor secretion, biofilm formation, conjugation, and antibiotic production [1–3].
Typically, Gram-negative bacteria use acyl-homoserine lactones and Gram-positive bacteria use peptides as autoinducers. To our knowledge, these two kinds of molecules most often promote intraspecies cell–cell communication, because a particular acyl-homoserine lactone or particular peptide can be detected only by the bacterial species that produces it [2]. In addition, a non–species-specific autoinducer called AI-2, which is a family of interconverting molecules all derived from the same precursor 4,5-dihydroxy 2,3-pentanedione, is produced and detected by a large variety of both Gram-negative and Gram-positive bacteria [4,5]. Interestingly, many bacterial species use more than a single autoinducer molecule for quorum sensing. For example, Gram-negative bacteria (e.g., Rhizobium) can use multiple homoserine lactones and likewise, Gram-positive bacteria (e.g., Bacillus) can use several peptides for communication [2,6]. These bacteria have evolved sophisticated quorum-sensing circuits to detect and integrate the information contained in multiple autoinducers.
It remains a mystery how and why bacteria integrate multiple autoinducer signals and what additional information multiple autoinducers reveal about the cells' environment that one autoinducer cannot reveal [7]. Furthermore, while in principle, quorum sensing enables bacteria to act in synchrony, the behavior of the entire population is ultimately dictated by events inside single cells. Recent single-cell studies of gene expression in bacteria have revealed that noise is inevitable even for isogenic cells in essentially homogeneous environments, and that noise can result in heterogeneous phenotypes within a population [8–14]. Likewise, in quorum sensing, noise could make individual cells behave differently from one another even if they receive identical autoinducer inputs. To understand quorum-sensing signal integration and, ultimately, the evolution of cooperative behaviors at the population level, it is imperative to understand how cells behave individually. Specifically, do cells respond in unison or do they maintain population diversity? Bulk measurements—which focus on the population's response—generally mask the behavior of individual cells and thus lose information about cell-to-cell variation. To fully understand the molecular mechanism underlying quorum sensing as well as the general principles underlying bacterial communication and cooperation, we must study this process at the single-cell level.
To begin to explore the above questions, we investigated the network of the model quorum-sensing bacterium Vibrio harveyi, the first bacterium shown to use more than one autoinducer for quorum sensing [15,16]. V. harveyi has a particularly ideal system in which to undertake these studies because the components of the quorum-sensing circuit have been defined (Figure 1A) and the autoinducers are known and in hand. V. harveyi produces and detects three autoinducers: AI-1 (3-hydroxybutanoyl homoserine lactone), CAI-1 ([S]-3-hydroxytridecan-4–1), and AI-2 ([2S,4S]-2-methyl-2,3,3,4-tetra hydroxytetrahydrofuran borate) [6,17,18]. AI-1 is only produced by V. harveyi, CAI-1 is produced by V. harveyi as well as other Vibrios, and as discussed, AI-2 is produced by many bacterial species. Thus AI-1, CAI-1, and AI-2 could provide information about the numbers of V. harveyi, Vibrios, and total bacteria in the vicinity, respectively. The three autoinducers are detected extracellularly by their cognate transmembrane receptors: LuxN, CqsS, and LuxPQ, respectively [19]. Information from the autoinducer-sensing pathways is transduced through shared components LuxU and LuxO [20–22] and five small regulatory RNAs (sRNAs) [23,24] to the master quorum-sensing regulator LuxR [25] (Figure 1A). LuxR activates and represses genes including those required for bioluminescence, siderophore production, type III secretion, and metalloprotease production [2,26–28].
Here we report the quantitative single-cell fluorescence-microscopy studies of V. harveyi quorum sensing, which have allowed us to define the mechanism of quorum-sensing autoinducer signal integration. Our studies revealed highly uniform behavior in individual cells, suggesting that the V. harveyi quorum-sensing circuit is designed to tightly synchronize the population response to autoinducers. This network operates in stark contrast to other regulatory circuits (e.g., such as that underpinning sporulation in Bacillus subtilis), which appear designed to generate diversity among the members of the population [29–32]. We also discovered that information from the different autoinducers is integrated in a strictly additive way, with an unexpected balance between the signaling strengths of the different autoinducers, allowing the population as a whole to distinguish multiple states of autoinducer concentration. These results have important implications for the developmental cycle of V. harveyi and possibly for other bacteria that use multiple autoinducers.
To investigate the mechanism underlying how V. harveyi integrates the information contained in its multiple autoinducers, we engineered strains that allowed us to examine each quorum-sensing signaling pathway in isolation as well as strains that allowed us to analyze the signaling properties of the combined pathways. In the present study, we focused only on integration of signals from autoinducers AI-1 and AI-2 through the LuxN and LuxPQ pathways, respectively. We did not study CAI-1 signaling through CqsA. Our rationale is as follows: First, under our laboratory conditions, the CAI-1 pathway is the weakest of the three signaling pathways, and thus AI-1 and AI-2 are the major inputs influencing quorum-sensing–controlled gene expression; second, we wanted to analyze the simplest possible case of integration of two signals. For this set of experiments, we constructed V. harveyi strains carrying only the LuxN pathway, only the LuxPQ pathway, or both pathways. In each case, the V. harveyi strains lacked the CqsS pathway. To enable quantitative measurements of signaling through the individual and combined pathways, all the strains were engineered to contain a transcriptional fusion of gfp fused to a quorum-sensing responsive promoter. Additionally, all of our strains constitutively produced red fluorescent protein (i.e., mCherry) that we used as an internal standard (Figure 1B) [33].
The strains used are as follows: The LuxN+ strain carries wild-type luxN on the chromosome and is deleted for cqsS and luxPQ. The strain is also deleted for luxM, encoding the AI-1 synthase LuxM, and is therefore exclusively responsive to exogenously added AI-1. Similarly, the LuxPQ+ strain is deleted for luxN and cqsS as well as luxS encoding the AI-2 synthase. This strain is only responsive to exogenous AI-2. The combined LuxN+ LuxPQ+ strain lacks cqsS, luxM, and luxS, and is responsive to exogenously supplied AI-1 and AI-2. In each strain, gfp is fused to the qrr4 promoter, which is one of the genes encoding the quorum-sensing sRNAs that are activated by LuxO-P (Figure 1B). mCherry is driven by the constitutive tac promoter inserted at an intergenic region of the chromosome. Because mCherry is expressed constitutively, it reports on the cell's overall protein level, including variations due to cell size and phase of the cell cycle. Normalizing the reporter green fluorescent protein (GFP) intensity by the internal standard mCherry intensity therefore provides an accurate measure of quorum-sensing receptor activity, and eliminates errors caused by uneven illumination or inaccurate segmentation of cells during image processing. The engineered V. harveyi strains were grown to steady state (Figure S1) in broth medium containing particular autoinducer concentrations. Cells were transferred to glass slides on a microscope, and phase-contrast and fluorescence snapshots were taken. Microscopy images were processed automatically by a custom computer program to obtain fluorescence intensities of individual cells. (For details, see Materials and Methods.)
Each autoinducer-detection pathway contributes uniquely to the overall V. harveyi integrated quorum-sensing response. Thus, to understand how cells communicate, understanding the signaling properties of the individual quorum-sensing pathway is imperative. Toward this end, we measured dose responses of individual cells of the LuxN+ mutant responding to AI-1. LuxN+ mutant cells were grown in series-diluted concentrations of exogenous AI-1, and the distributions of PQrr4-GFP intensities of individual cells at each AI-1 concentration were obtained (Figure 2). A gradual increase in the mean PQrr4-GFP intensity distribution occurred with decreasing AI-1 concentration, reflecting increasing kinase activity of LuxN, and, consequently, increasing LuxO-P concentration. While we observed heterogeneity in PQrr4-GFP expression over the population, the distribution of PQrr4-GFP intensities remained single-peaked with moderate variance around the population average at all AI-1 concentrations (cell-to-cell variation was somewhat smaller after normalizing by mCherry intensity; see Figure S2). This result suggests that all the V. harveyi cells respond identically to AI-1, which promotes well-coordinated cellular behavior across the population. The shift in the mean PQrr4-GFP intensity between zero and saturating AI-1 is obviously larger than the standard deviation within the population at any AI-1 concentration, suggesting that cell-to-cell variation, or noise, in quorum sensing is low enough to allow the cells to reliably mount distinct responses to low and high AI-1 concentrations.
We performed similar individual-cell dose–response experiments on the V. harveyi LuxPQ+ mutant strain to determine the signaling properties of the AI-2 pathway. For comparison, in Figure 3A we show dose–response curves for both the LuxN+ and LuxPQ+ mutant strains to AI-1 and AI-2, respectively. Means and standard deviations over a population of cells are reported for each strain. Similar to the results shown in Figure 2, at all autoinducer concentrations the normalized PQrr4-GFP-intensity distributions are single-peaked, with standard deviation over the mean always smaller than 0.4. For each data point, the population sample consists of 100 individual cells, thus the standard error of the mean is one-tenth of the standard deviation of the population. Each dose–response curve can be described by a simple Hill function αAI + βAI/(1 + [AI]/KAI) with Hill coefficient equal to one. The inhibition constants for AI-1 and AI-2 are KAI-1 = (6.9 ± 0.5) nM and KAI-2 = (6.4 ± 0.5) nM, respectively. Note that a 1 nM concentration is approximately one molecule of autoinducer in the volume of a single V. harveyi cell, indicating an extremely sensitive response of V. harveyi cells to autoinducers. The LuxN+ strain has approximately 50% higher PQrr4-GFP levels than the LuxPQ+ strain at low autoinducer concentrations where LuxO-P and PQrr4-GFP are maximal. However, the two strains have similar residual levels of PQrr4-GFP, which remain measurable above background at saturating autoinducer concentrations.
The above experiments allowed us to determine the signaling response of the LuxN pathway to AI-1 and that of the LuxPQ pathway to AI-2 when each pathway is present alone. We likewise wondered how the cells respond to AI-1 and AI-2 when the two pathways are present together. To examine this, we performed experiments analogous to those above with the V. harveyi LuxN+ LuxPQ+ strain in the presence of combinations of AI-1 and AI-2. Surprisingly, we found that although the amplitudes of the autoinducer responses are different when the two quorum-sensing pathways are present individually (Figure 3A), the amplitudes of the AI-1 and AI-2 responses are nearly identical when the two pathways are present simultaneously (Figure 3B). In particular, the dose–response curves for AI-1 (blue) and AI-2 (red) almost overlap, both in the case when one autoinducer is present alone and in the case when a saturating amount of the other autoinducer is also present. Critically, the overlap of these curves depends on the extremely similar amplitudes of the responses as well as the similar inhibition constants for AI-1 and AI-2 as observed in Figure 3A. The very similar amplitudes of the two autoinducer dose–response curves demonstrate that each autoinducer-sensing pathway contributes approximately half of the total response.
Figure 3B clearly shows that when both pathways are present (e.g., in the LuxN+ LuxPQ+ strain), each autoinducer alone is only capable of partial inhibition of PQrr4-GFP expression. When AI-1 and AI-2 concentrations are increased together, with similar concentrations of each autoinducer present, the resulting dose–response curve of PQrr4-GFP expression covers the entire dynamic range (yellow-green curve). The PQrr4-GFP distribution is always single-peaked, and noise in GFP expression is always moderate, with the standard deviation over the mean no more than 40%. Again, we take this to mean that despite the existence of noise in the quorum-sensing pathway, individual cells are able to discriminate several distinct states. For example, the PQrr4-GFP distributions do not substantively overlap for these three cases: when both AI-1 and AI-2 are below 1 nM, when both are around 10 nM, and when both are above 100 nM. Thus, it appears that individual V. harveyi cells can accurately determine the level of external autoinducers. This result suggests that, in principle, V. harveyi cells can not only detect low and high cell-density states with low and high autoinducer concentrations, but also some intermediate cell-density states represented by intermediate autoinducer concentrations.
To obtain a more comprehensive view of the autoinducer response of the LuxN+ LuxPQ+ strain, we explored a grid of possible combinations of AI-1 and AI-2 concentrations. In this way, the complete dose–response surface was obtained (Figure 3C). This surface, displaying average PQrr4-GFP production, is almost mirror-symmetric with respect to the equal AI-1 and AI-2 diagonal; i.e., the PQrr4-GFP expression is almost invariant with respect to exchange of AI-1 and AI-2 concentrations. Notably, there are at least three distinct states of the output PQrr4-GFP level: high (both AI-1 and AI-2 concentrations are low, indicated by the red area in Figure 3C), intermediate (one autoinducer concentration is low and the other is high, indicated by the two green areas), and low (both AI-1 and AI-2 concentrations are high, indicated by the blue area). This surface confirms that more than two quorum-sensing states can be deciphered by the cells. However, interestingly, under these conditions, high AI-1/low AI-2 is apparently not distinguished from low AI-1/high AI-2 (see Discussion).
For a signal-integration circuit such as the quorum-sensing circuit in V. harveyi that involves multistep, bidirectional, biochemical reactions, one might expect the two signals to be integrated in a complicated nonlinear manner. Surprisingly, however, we found quite the opposite. That is, AI-1 and AI-2 signal integration is simply additive. The dose–response surface of the LuxN+ LuxPQ+ strain can be accurately described by the additive function
where the γ's and K's are fitting parameters. The inhibition constants have the same values as in the individual pathways: KAI-1 = 6.9 nM and KAI-2 = 6.4 nM (Figure 3A and 3B). As shown in Figure 3D, the average PQrr4-GFP expression values obtained from Equation 1 agree with the measured values over the entire dose–response surface. The two noncooperative Hill functions correspond to the individual responses of the LuxN and the LuxPQ pathways, respectively. Therefore, we conclude that LuxN and LuxPQ make independent, additive contributions to GFP levels presumably via additive contributions to LuxO-P.
Although the two autoinducer signals are combined additively with approximately equal weights in their input to the circuit, we find that the two pathways contribute differently to the noise in PQrr4-GFP expression. As shown in Figure 4A, the LuxPQ+ strain (with no LuxN receptor) has significantly larger relative noise, i.e., larger cell-to-cell variation, than does the LuxN+ strain (with no LuxPQ receptor) for the same mean PQrr4-GFP level. Apparently, signaling through the LuxPQ receptor introduces more noise to the circuit than does signaling through the LuxN receptor. This difference is confirmed by the distinct noise levels observed for the LuxN+ LuxPQ+ strain treated with either saturating AI-1 or saturating AI-2 (Figure 4B). In the LuxN+ LuxPQ+ strain, the mean PQrr4-GFP levels are nearly identical under these two conditions, but the relative noise is almost a factor of two larger when only LuxPQ contributes kinase activity (AI-1 saturating) than when only LuxN contributes kinase activity (AI-2 saturating). Indeed, as shown in Figure 4B, noise in the LuxN+ LuxPQ+ strain is at its absolute maximum when only LuxPQ contributes kinase activity.
Our observation that the LuxN and LuxPQ pathways contribute independently and additively to PQrr4-GFP expression implies that the kinase activities of LuxN and LuxPQ must be regulated by the autoinducers. We draw this conclusion from the following simple model for the signaling pathway leading to PQrr4-GFP expression: We assume that LuxN and LuxPQ are the dominant kinases and phosphatases for LuxU, that phosphotransfer between LuxU and LuxO is reversible, and that PQrr4-GFP expression is a linear function of LuxO-P concentration [O-P]. The final assumption follows from the observed additivity of PQrr4-GFP expression with respect to AI-1 and AI-2, which is difficult to understand unless [O-P] is in the linear regime of the qrr4 promoter driving gfp, i.e., the maximal [O-P] is far below the level required to half saturate the promoter activity. The kinetic equations describing this model are
where [U-P] is the LuxU-P concentration, and KN, KPQ, PN, and PPQ are the total cellular kinase and phosphatase activities of LuxN and LuxPQ, respectively. At steady state, the time derivatives in Equation 2 can be set to zero, yielding
where [O]tot is the total concentration of LuxO. To explain the observed broad range of additivity of PQrr4-GFP expression with respect to the autoinducers, Equation 3 must be separable into two terms, one of which depends only on AI-1 and the other only on AI-2. This is possible if the autoinducers regulate the receptor kinase activities KN and KPQ, but not if the autoinducers regulate only the receptor phosphatase activities PN and PPQ, since the latter appear only in the denominator of Equation 3. Indeed, for additivity to be achieved, the denominator of Equation 3 must be approximately constant, which implies one of two scenarios: (1) only the kinase activities of LuxN and LuxPQ are regulated by autoinducers while phosphatase activities are not, and the kinase and phosphatase activities satisfy KN + KPQ << k−/k+ · (PN + PPQ), implying that LuxO-P levels are far from saturation, i.e., [O‐P] << [O]tot; and (2) the kinase and phosphatase activities are both regulated, but their sum is independent of autoinducer concentration such that KN + KPQ + k−/k+ · (PN + PPQ) remains constant. Unlike the first scenario, the second scenario requires fine-tuning of reaction rates and therefore seems less likely. While the signaling pathways leading to LuxO-P are likely to include some processes not considered in our simple model (e.g., intrinsic dephosphorylation of LuxU-P and LuxO-P), our qualitative conclusions—in particular that the kinase activities of LuxN and LuxPQ must be autoinducer regulated—are robust to such quantitative corrections.
Since the amplitudes of the responses to AI-1 and AI-2 are almost identical in the LuxN+ LuxPQ+ strain (Figure 3B), the maximum total kinase activities of the two receptors LuxN and LuxPQ must be nearly the same (i.e., KN KPQ). However, for the strains expressing only a single receptor type, the peak PQrr4-GFP expression is 50% higher for the LuxN+ than for the LuxPQ+ strain (Figure 3A). This apparent discrepancy can be readily accounted for if the total phosphatase activity of LuxPQ is higher than that of LuxN, i.e., PPQ > PN (including possible differences in receptor concentration).
Living cells monitor their environment using a variety of signal-transduction systems, ranging from simple two-component systems in prokaryotes to highly complex signal-transduction networks in mammalian cells. Since environmental cues are always numerous, the ability to integrate multiple signals is indispensable if cells are to behave appropriately. However, the mechanisms and logic by which cells integrate environmental signals remain, by and large, poorly understood. Here we have quantitatively analyzed the integration of multiple autoinducer signals by the model quorum-sensing bacterium V. harveyi using single-cell fluorescence microscopy. Our studies reveal a unified response across the population, with moderate cell-to-cell variation. We find that signals from two distinct autoinducers, AI-1 and AI-2, are combined strictly additively in a single phosphorelay pathway, with each autoinducer contributing nearly equally to the total response. Moreover, the cell-to-cell variation in response is small enough that the entire population of cells can reliably distinguish at least three distinct conditions of external autoinducer concentration.
We used GFP under the control of the chromosomal sRNA Qrr4 promoter as a reporter of the activity of the quorum-sensing signaling pathway (Figure 1). In all our strains, the GFP distribution was always single-peaked at all autoinducer concentrations, with cell-to-cell standard deviation no more than 40% of the mean, suggesting that populations of V. harveyi cells respond coherently to autoinducer signals. By contrast, genes in some other bacterial systems are known to have bimodal (i.e., two-peaked) expression distributions. In many cases, bimodal gene expression is also hysteretic (i.e., cells remain for a long time in one state of expression), which constitutes a form of cellular “memory.” For example, bimodal distributions in gene expression enable sporulation and competence in B. subtilis [29–32], stringent response in mycobacteria [34], and induction of the lac operon in Escherichia coli [35,36]. In all these cases, bimodality and hysteresis are believed to provide advantages to the organism by enabling phenotypic diversity within isogenic populations. In general, hysteresis in gene expression requires some form of positive feedback. The lack of bimodality in our engineered strains of V. harveyi is expected since there is no positive-feedback loop in the circuit controlling Qrr sRNA expression in these cells. Since our engineered strains lack both the downstream transcription factor LuxR and the autoinducer synthases, there exists the possibility that the sRNAs or LuxR could feed back positively to the synthases and produce a bistable circuit in wild-type cells. In quorum sensing, bistability has only been reported for a rewired LuxIR circuit in V. fischeri [37]. In this case, the positive feedback and the resulting bistability and hysteresis occur at the population level and divide the entire population into two separate subpopulations, each with a unique phenotype. Our consistent observation of a narrowly peaked distribution of quorum-sensing responses strongly suggests that V. harveyi cells respond in unison to the presence of autoinducer signals. For quorum-sensing cells, in contrast to bacteria undergoing competence, sporulation, or the stringent response, operating as a coherent population appears to be more important than maintaining phenotypic diversity.
An outstanding question is why V. harveyi and related species use multiple autoinducer signals, but funnel all the information into a single pathway. We can envision two main possibilities (potentially in combination): The multiple autoinducers could reveal information about the community composition (e.g., which species are present and in what abundance), or the multiple autoinducers could reveal information about the development stage of the community (e.g., the growth stage of a biofilm). In support of the first possibility, the three autoinducers used by V. harveyi have distinct ranges of species specificity: intraspecies for AI-1, within Vibrios for CAI-1, and across many species for AI-2 [7]. Thus, different combinations of the three autoinducers could indicate different compositions of a bacterial community. In our experimental conditions, however, we found that cells could not distinguish between high AI-1/low AI-2 and high AI-2/low AI-1 (Figure 3B and 3C). This result argues for the second possibility, namely that different combinations of autoinducers represent different stages of community development. For example, if a growing V. harveyi community typically accumulates AI-2 before AI-1, then the signaling contour in Figure 3C would always be traversed along the right edge, and cells could reliably interpret an intermediate signaling strength as a condition of high AI-2/low AI-1, since the opposite condition of high AI-1/low AI-2 would rarely, if ever, be encountered. In much of eukaryotic development (e.g., embryogenesis), the rate of development is fixed and driven by a clock [38], obviating the need for a signal representing the stage of development. However, without the support of a surrounding organism, the rate of development of a bacterial community depends on unpredictable environmental conditions, such as nutrient availability, and therefore some means of determining the stage of development is required so that cells in the community can behave appropriately. Recent models of biofilm growth suggest that communities may be mixed at early stages, but that at later stages competition for nutrients by overgrowth of neighboring cells can result in large domains of cells descended from a single progenitor, and therefore composed of a single species [39]. If so, generic signals such as AI-2 may be most informative at early stages of biofilm growth, while species-specific signals such as AI-1 may be reserved for later stages. We are currently exploring the order of accumulation of the V. harveyi autoinducers AI-1, CAI-1, and AI-2 to test whether different autoinducer combinations could signal different stages of community development.
Given that the autoinducer signals are combined in one pathway in V. harveyi, why should the signals be combined additively, as we observe for AI-1 and AI-2? Simple alternatives would be for saturating autoinducer levels to be combined in “logic gates,” such as AND, in which both autoinducer signals would be required for a full response, or OR, in which either signal would be sufficient for a full response. However, these logic gates have only two possible output states: on or off. In contrast, the addition of the two autoinducer signals allows for more than two output states of the signaling pathway, and therefore potentially allows for more than two expression states of quorum-sensing regulated genes. Indeed, we discovered three distinct levels of signaling strength, represented by the heights of the plateaus in Figure 3C. Moreover, the standard deviation of PQrr4-GFP expression across the population of cells was sufficiently small (Figure 4B) so that the entire population can apparently distinguish the three distinct plateau heights. This means that, in principle, every cell in the population can distinguish three external autoinducer conditions: both autoinducers low, both autoinducers high, and a third condition in which one autoinducer is high and the other is low. The reliability with which cells can distinguish among these three conditions is increased by the equal spacing of the plateau heights as shown in Figure 3C. Given a uniformly distributed input of autoinducer concentration and the observed level of noise (i.e., cell-to-cell variation in PQrr4-GFP expression), a significantly unequal spacing of the plateau heights would lead to overlapping distributions of PQrr4-GFP expression for the two more closely spaced plateaus. The implication is that noise might then cause some cells to misinterpret external conditions and regulate quorum-sensing genes inappropriately. The need for all cells to reliably distinguish among multiple autoinducer conditions may therefore explain not only the additivity of the quorum-sensing pathway, but also why the contributions of the AI-1 sensor LuxN and the AI-2 sensor LuxPQ to the total kinase activity are so nearly equal—equal kinase activities mean equally spaced plateau heights, which in turn mean that individual cells are less likely to confuse one autoinducer condition with another.
The existence of multiple quorum-sensing output states potentially underpins diverse patterns of quorum-sensing regulated gene expression. For example, in previous studies, the quorum-sensing circuit of V. harveyi was found to act as an autoinducer “coincidence detector” (i.e., requiring both AI-1 and AI-2) for full induction of bioluminescence [19,40]. Thus, in the present context, the three distinguishable levels of signaling output (indicated by Qrr4 promoter activity) appear to be collapsed by downstream signal-processing events to two levels of bioluminescence. More generally, the target genes of quorum sensing could be tuned to different signaling output levels so that only particular classes of genes are switched ON/OFF at early, middle, or late stages of community development. Alternatively, some genes could have graded expression between these different developmental stages. The requirement for multiple distinct output states might also explain our observation of a graded, rather than switch-like, response of the Qrr4 promoter. Specifically, our dose–response data are well described by a noncooperative, n = 1 Hill function response to both autoinducers. Cooperativity would have resulted in an n > 1 Hill function and therefore a more switch-like response of PQrr4-GFP to autoinducers. During the signaling process, cooperativity could in principle have arisen from the binding of autoinducers to receptors, transfer of phosphate among the protein components in the phosphorelay, and/or binding of phosphorylated LuxO to DNA. Our results suggest that in fact all of these steps are noncooperative, despite the fact that the receptors are likely dimers [22] and that LuxO may function as a tetramer or octamer [Tu KC, unpublished data]. Indeed, a graded noncooperative response of Qrr expression to autoinducers is essential for the existence of multiple, distinguishable quorum-sensing states, as a switch-like response of the Qrr expression would have allowed for only two states.
Based on a simple kinetic model for signaling (Equation 2), we have argued that the kinase activities of LuxN and LuxPQ are regulated by autoinducers, whereas for most two-component receptors, it is still an open question whether the kinase or phosphatase or both activities are regulated by input stimuli. Previously, LuxN receptors have been successfully modeled as switching between two states: the ON (kinase dominant) and OFF (phosphatase dominant) states [41,42]. Each receptor has intrinsic kinase and phosphatase rates depending only on the state in which the receptor exists. Extending this model to LuxPQ, the total cellular kinase activities KN and KPQ consist of a major contribution from those receptors in the ON state with little or no contribution from those in the OFF state. From the constraints set by additivity, we conclude that the phosphatase activities PN and PPQ are unregulated (i.e., receptors have the same phosphatase rates in both the ON and OFF states). Note that autoinducer concentrations only affect the thermal balance between ON and OFF states, and therefore the kinase and phosphatase activities are regulated only via the biasing of receptors between states (of course, the total kinase and phosphatase activities also depend on receptor concentrations). The low levels of PQrr4-GFP expression with saturating AI-1 in the LuxN+ strain, saturating AI-2 in the LuxPQ+ strain, and saturating AI-1 plus AI-2 in the LuxN+ LuxPQ+ strain indicate that kinase rates in the OFF states are much smaller than those in the ON states for both LuxN and LuxPQ. By decreasing the fraction of receptors in the ON state, autoinducers reduce the total kinase activity of the quorum-sensing receptors in V. harveyi. (See Text S1 for more details.)
Regulation of the kinase activities of LuxN and LuxPQ appears to be necessary to achieve three equally spaced levels of LuxO-P (Equation 3). The requirement for kinase regulation in V. harveyi quorum sensing therefore appears to stem from the need to combine multiple input signals into more than two distinguishable output levels of LuxO-P. One prediction from this analysis is that the sensor CqsS, which was not present in our strains, is likely to also have its kinase activity regulated by its autoinducer CAI-1. Moreover, CqsS is likely to contribute additively to total kinase activity and with a strength comparable to that of LuxN and LuxPQ, resulting in four maximally distinguishable levels of kinase activity and therefore four distinguishable autoinducer conditions.
The similarity of the responses to AI-1 and AI-2 is striking, not only in the amplitudes but also in the inhibition constants. We speculate that V. harveyi usually encounters similar amounts of AI-1 and AI-2, and the responses of receptors have been optimized to match the natural dynamic range of autoinducer concentrations. It has been demonstrated that single mutations in the receptors LuxN and LuxPQ can result in dramatic changes in their inhibition constants [22,42], so the similar values for AI-1 and AI-2 may represent an evolved optimum.
We also quantified the noise in PQrr4-GFP expression in our three reporter strains. Noise is an inherent feature of signal transduction and gene expression both in prokaryotes and eukaryotes. Due to the low copy number of cellular components and the stochastic nature of biochemical reactions, fluctuations are inevitable. Large fluctuations might be deleterious for processes requiring precise control but beneficial for those providing phenotypic diversity. In quorum sensing, bacterial cells detect population cell density to coordinate their behavior on a community-wide scale. Low noise in quorum-sensing signal transduction might therefore benefit the population of cells by allowing all cells to behave correctly and in unison at each stage of community development. Indeed, we observed low noise in PQrr4-GFP expression in all our strains. At all autoinducer concentrations the standard deviation over the mean was less than or close to 0.4 (Figure 4). In other systems, the dominant source of cell-to-cell variation in gene expression has been attributed to extrinsic noise, e.g., differences among cells in concentrations of general purpose cellular components such as RNA polymerases and ribosomes [8]. In the quorum-sensing circuit we have studied, the noise we observed is also likely due to extrinsic factors rather than to biochemical noise in phosphotransfer or transcription and translation of PQrr4-GFP. The most likely source of the noise we observed is fluctuations in concentrations of the pathway components, such as the receptors LuxN and LuxPQ and the response regulator LuxO. The noisier response in LuxPQ pathway is very likely caused by variations in the copy number of the LuxPQ receptors, which suggests that there could be some additional regulation of receptor expression in the quorum-sensing circuit.
All V. harveyi strains used in this study were derived from the wild-type strain BB120 [43] and grown aerobically at 30 °C in Autoinducer Bioassay (AB) broth. E. coli S17–1λpir was used for general DNA manipulation and grown with aeration at 37 °C in LB (Luria-Bertani) broth. The relevant strains and plasmids are listed in Table S1.
DNA manipulation was performed using standard procedures [44]. Phusion DNA polymerase was used for PCR reactions. dNTPs, restriction enzymes, and T4 DNA ligase were obtained from New England Biolabs. DNA purification kits were provided by Qiagen. E. coli was transformed by electroporation using a Bio-Rad Micro Pulser. Plasmids were introduced into V. harveyi by conjugation [15] and exconjugants were selected using the antibiotic resistances carried on the plasmids together with polymyxin B.
A cat-resistance cassette from pKD3 [45] was cloned into vector pCMW1 [7] downstream of gfp at the BamH1 site, making pTL3. The GFP-Cmr fragment from this construct was subsequently amplified by PCR and recombined using the λ red technique [45] into a cosmid to replace the wild-type qrr4 gene, producing pTL20. Lastly, PQrr4-GFP-Cmr was introduced onto the chromosome to replace qrr4 by allelic recombination. Ptac-mCherry was amplified from the vector pEVS143-mCherry containing an IPTG inducible mCherry gene and cloned into pKD13 [45] at the NheI site, resulting in pTL82. The cosmid, pTL83, was constructed using the λ red technique by recombining the Ptac-mCherry-Kanr fragment into the intergenic region downstream of the entire lux operon. Final insertion of Ptac-mCherry-Kanr onto the V. harveyi chromosome was accomplished by allelic recombination.
To construct the various V. harveyi sensor mutants, pKM780 carrying ΔluxS::Cmr, pJMH291 carrying ΔluxN::Cmr, pDLS100 carrying ΔluxPQ::Cmr, pJMH244 carrying ΔcqsS::Cmr, and pKM705 carrying ΔluxR::Kanr were used to sequentially delete the corresponding wild-type genes by allelic recombination. Following each gene deletion, the plasmid pTL18 containing an IPTG-inducible FLP recombinase, derived from pEVS143 and pCP20 [45], was introduced into the V. harveyi strain to eliminate the antibiotic resistance marker on the chromosome.
For dose–response experiments, V. harveyi strains LuxN+ (TL87), LuxPQ+ (TL88), and LuxN+ LuxPQ+ (TL89) were grown in AB medium for 8∼12 h. Growth was monitored by measuring optical density at 600 nm. Cultures were diluted to OD600 = 10−6 ∼ 10−7, and exogenous autoinducers were added at the specified concentrations. Following growth to steady state (13∼14 h; OD600 = 0.005 ∼ 0.05), cells were concentrated by centrifugation and maintained on ice until measurements were made. One μl of cell culture was spread on a glass slide and covered with a 1% AB agarose pad as well as a coverslip.
Phase-contrast and fluorescent images were taken at room temperature using a Nikon TE-2000U inverted microscope. Custom Basic code was written to control the microscope. Images were acquired using a 100× oil objective and a cooled CCD camera (−65 °C, Andor iXon). Segmentation of individual cells was performed on phase-contrast images. Background and cellular auto-fluorescence values were subtracted from the green and red channels, respectively. Total fluorescence intensity of each cell was obtained by summing all pixels and fractions of pixels in the segmented cell region. Normalized GFP values for each cell were calculated by normalizing total green to total red fluorescence intensity.
|
10.1371/journal.pntd.0002977 | Synergy Testing of FDA-Approved Drugs Identifies Potent Drug Combinations against Trypanosoma cruzi | An estimated 8 million persons, mainly in Latin America, are infected with Trypanosoma cruzi, the etiologic agent of Chagas disease. Existing antiparasitic drugs for Chagas disease have significant toxicities and suboptimal effectiveness, hence new therapeutic strategies need to be devised to address this neglected tropical disease. Due to the high research and development costs of bringing new chemical entities to the clinic, we and others have investigated the strategy of repurposing existing drugs for Chagas disease. Screens of FDA-approved drugs (described in this paper) have revealed a variety of chemical classes that have growth inhibitory activity against mammalian stage Trypanosoma cruzi parasites. Aside from azole antifungal drugs that have low or sub-nanomolar activity, most of the active compounds revealed in these screens have effective concentrations causing 50% inhibition (EC50's) in the low micromolar or high nanomolar range. For example, we have identified an antihistamine (clemastine, EC50 of 0.4 µM), a selective serotonin reuptake inhibitor (fluoxetine, EC50 of 4.4 µM), and an antifolate drug (pyrimethamine, EC50 of 3.8 µM) and others. When tested alone in the murine model of Trypanosoma cruzi infection, most compounds had insufficient efficacy to lower parasitemia thus we investigated using combinations of compounds for additive or synergistic activity. Twenty-four active compounds were screened in vitro in all possible combinations. Follow up isobologram studies showed at least 8 drug pairs to have synergistic activity on T. cruzi growth. The combination of the calcium channel blocker, amlodipine, plus the antifungal drug, posaconazole, was found to be more effective at lowering parasitemia in mice than either drug alone, as was the combination of clemastine and posaconazole. Using combinations of FDA-approved drugs is a promising strategy for developing new treatments for Chagas disease.
| Chronic infection with Trypanosoma cruzi causes progressive damage to the heart and other organs that is fatal in about 30% of cases. Known as Chagas disease, this is a major public health problem in Latin America. The existing medicines were developed over forty years ago and are not widely used because of toxicity and unreliable effectiveness. To discover better treatments, we screened a collection of existing drugs for growth inhibitory activity on Trypanosoma cruzi. Several dozen orally administered drugs were discovered, but when used by themselves they were not strong enough to cure the infection in an animal model. We tested a set of 24 of these drugs in every two-way combination and identified eight synergistic partners. At least two of these combinations were able to substantially lower parasite levels in the mouse model of Trypanosoma cruzi infection. Thus, finding pairs of FDA-approved drugs that can be used in combination may be a pragmatic and effective strategy for designing new therapies for Chagas disease.
| The need for new more effective drugs to treat Chagas disease has not been matched by drug discovery efforts. An estimated 8 million people have chronic infection with the etiologic agent, Trypanosoma cruzi [1]. Existing treatments consist of two nitroaromatic compounds (benznidazole and nifurtimox) that are poorly tolerated and have uncertain efficacy for curing chronic infection [2]. Historically, the pharmaceutical industry has not invested substantially in tropical diseases such as Chagas disease for economic reasons. The rising costs of bringing new drugs to the market exacerbates the situation, despite the recognized expansion of Chagas disease into wealthier parts of the world [3]. No new clinical drugs for Chagas disease have been licensed or evaluated in Phase III clinical trials since the introduction of benznidazole and nifurtimox in the 1960–70's. The barriers to bringing entirely new clinical entities through preclinical and clinical development are formidable, hence, alternative strategies for Chagas disease drug development need to be considered. Repurposing existing drugs is an attractive option for “neglected tropical diseases” because the costs associated with preclinical testing and attrition are avoided and, generally, the safety profiles and pharmacological characteristics are well characterized and can be matched to the particular clinical need. Thus, it may be possible to discover licensed drugs that could be rapidly advanced to clinical trials for neglected diseases such as Chagas disease. To address this question, we combined in vitro screening of compounds for anti-T. cruzi activity with follow-up in vivo studies in a murine model of acute T. cruzi infection. This strategy has been employed by us and others leading to the discovery of various categories of drugs with anti-T. cruzi activity [4]–[6]. For example, antifungal agents (i.e., ergosterol biosynthesis inhibitors), tricyclic antidepressants, and various antipsychotic agents have been discovered in such screens [5]. The drug discovery efforts have led to a phase II clinical trial of the antifungal agent, posaconazole, in Chagas patients in Spain (ClinicalTrials.gov Identifier: NCT01162967), and Argentina (ClinicalTrials.gov Identifier: NCT01377480) with results yet to be published. Although azole antifungal drugs represent a potentially attractive therapeutic alternative to the existing treatment options, their efficacy for treating Chagas disease is not yet established. It is important to continue to try to identify existing drugs in hopes of repurposing them for Chagas disease.
However, with the exception of azoles (and allopurinol) [7], none of the clinical drugs discovered to date has shown enough activity to lead to testing in formal clinical trials. Thus, a different strategy may be necessary to find “off the shelf” drugs that could be used for Chagas disease. In this study, we screened a collection of Food and Drug Administration (FDA)-approved drugs and biologically active compounds, and then systematically evaluated the hits from our screens in combinations searching for synergistic partners (Figure 1). A number of novel drug combinations showed in vitro synergy and improved survival in the mouse model of acute T. cruzi infection, supporting the utility of this strategy for drug development. Additional work will be necessary to establish which drug combinations may be curative in animal models and candidates for possible clinical studies.
The Spectrum Collection of 2000 biologically active, diverse compounds was purchased from MicroSource Discovery Systems, Inc. (Gaylordsville, CT) [8]. The collection includes ∼700 FDA-approved drugs. The compounds were provided as 10 mM DMSO stocks in 96-well plate format. Compounds used in synergy assays and mouse efficacy studies were purchased from Sigma-Aldrich, except as follows. JK-11 corresponds to compound 1 in a previous publication [9] and, benznidazole was acquired as previously described [10].
Compounds were tested against T. cruzi (Tulahuen strain) stably expressing the beta-galactosidase gene as previously described [11]. All in vitro assays were performed on mammalian-stage T. cruzi grown in co-culture with murine 3T3 fibroblasts using RPMI-1640 media (w/o phenol red, w/o L-glutamine) supplemented with 10% heat inactivated fetal calf serum, 100 u/mL penicillin/100 ug/mL streptomycin, 2 mM L-glutamine (final concentrations) [10]. Fibroblasts were plated at a density of 2×103 per well in 96 well tissue culture plates. After 24 hours of incubation, 1×104 trypomastigotes/well were added to the fibroblasts and incubated for 4 hours before addition of the test compounds from the Spectrum Collection (10 µM final concentration). Cultures were incubated at 37°C for 5 days, then developed with chlorophenol red β-D-galactopyranoside as previously described [11]. The percent inhibition is reported with standard deviation of the mean. For the effective concentration causing 50% growth inhibition (EC50) measurements, the compounds were tested in triplicate in serial two-fold dilutions and EC50 (or EC25) values were calculated by non-linear regression using Graphpad Prism (San Diego, CA). Similarly, for measuring the cytotoxicity concentration (CC50) for 3T3 fibroblast cells, cultures were incubated with drugs for 72 hours and developed using Alamar Blue (Alamar Biosciences Inc, Sacramento, CA) as previously described [11]. Z-prime values were calculated for each 96-well plate based on positive (4 wells) and negative controls (4 wells) [12].
Twenty-four compounds were selected for testing in combinations. All two-way combinations were tested (300 experiments). First, EC25 concentrations were determined for the individual compounds against T. cruzi amastigotes as described above. To test for synergy, compounds were evaluated in quadruplicate individually at the experimentally determined EC25, and in combination with other compounds at each respective EC25 concentration (further explained in the Discussion section). For inclusion in downstream analysis, each individual compound in a pair was required to inhibit 25±10% of growth in positive control (untreated) wells. If not in this range, the experiment was repeated. The measured growth of T. cruzi amastigotes was compared to the predicted effect of the combination as follows. Assuming a simple additive effect, the predicted inhibition of the drug pairs was expected to be the product of the percent-growths of each compound when tested alone. For example, if compound A gave 75% growth of the control and compound B gave 80% of the control growth, then the combination would be predicted to be 60% (i.e., 75%×80% = 60%). With this “prediction”, we then evaluated each compound combination for whether it resulted in more or less growth than would be expected by the additive effects, and calculated a proportional effect based on the following equation.
The results were tabulated and displayed in a heat-map format to help visualize the variance away from the predicted effects of the pairs. Cells in green indicate a greater effect than predicted (“synergism”) and the squares in red indicate a lesser effect than predicated (“antagonism”). A few empty boxes remain for experiments that did not meet the quality standard mentioned above despite at least two efforts.
Drug combinations observed to have possible synergism in the screen described above were subjected to formal isobologram analysis using the fixed ratio method [13]. Drug combinations were set up with the highest concentrations in the following proportions of their EC50: 4∶0, 2.67∶1.33, 2∶2, 1.33∶2.67, 0∶4. Serial two-fold dilutions were performed in triplicate. Amastigote cell growth was quantified by colorimetric readout after 5 days of culture. For each ratio, an EC50 was calculated for each of the drugs. The fractional inhibitory concentrations (FIC) were calculated as the [EC50 when in combination]/[EC50 of drug alone]. The sum of the FIC was calculated as follows: ΣFICs = FIC drug A + FIC drug B. The mean sum of the FIC (xΣFIC) was calculated as the average of ΣFIC from the three different fixed ratios. The interactions were considered synergistic for xΣFIC≤0.5, indifferent for xΣFIC between 0.5 and 4, and antagonistic for xΣFIC>4.
Age 8–10 week-old BALB/c female mice were obtained from Harlan Laboratories. Mice were infected with 1×104 tissue culture derived wild-type trypomastigotes of the Tulahuen strain by subcutaneous injection on day 0. They were administered test drugs in groups of five by oral gavage on days 7–11. All drugs were dissolved in vehicle composed of sodium carboxymethylcellulose 0.5% w/v, benzyl alcohol 0.5% v/v, Tween 80 0.4% v/v diluted in 0.9% aqueous NaCl solution. Parasitemia was quantified by examining tail blood specimens at times points indicated in Figures 2, 3, S4, and S5 as previously described [14].
All mouse work for this project was reviewed and approved by the University of Washington Institutional Animal Care and Use Committee under protocol 2154-01. The University of Washington has an approved Animal Welfare Assurance (#A3464-01) on file with the NIH Office of Laboratory Animal Welfare (OLAW), following guidelines of the USDA Animal Welfare Act and Regulations.
The Spectrum Collection of 2000 compounds was screened at 10 µM against intracellular T. cruzi amastigotes in duplicate plates. Thirty-one compounds were not included in the screen due to precipitation. The complete ranked data set is provided in Table S1. The quality of the data was excellent as demonstrated by the Z-prime values averaging 0.65 (Figure S1). Growth inhibition of 3 standard deviations above the no-drug control corresponds with 32.1% inhibition, yielding a hit rate of 40.1% by this criterion (i.e. 791 hits, Table S1). By defining “hits” as compounds causing ≥75% growth inhibition, a subset of 350 compounds (17.8%) was identified, including all compounds above the yellow line in Table S1.
Our goal was to establish a set of compounds for characterization as potential anti-trypanosomal agents. With this in mind, we eliminated compounds that are known to be toxic or lack potential to be developed as drugs (criteria for exclusion are shown in Table 1). By applying these criteria, 94 compounds were readily removed leaving 256 (13.1%) compounds (see Table S1, column labeled “Discarded”). Examples of excluded compounds include phenylmercuric acetate (toxic) and emetine (induces vomiting). The 256 remaining compounds were next tested for growth inhibition on mammalian 3T3 fibroblasts to exclude compounds that inhibited T. cruzi growth due to cytotoxicity to the host cells. The average Z′-value for these assays was 0.855. There were 105 compounds that caused >33% growth inhibition of 3T3 fibroblasts, and considered cytotoxic and excluded from further analysis. The remaining 151 compounds (Table S2) represent 7.7% of the original library and are distributed amongst a variety of chemical/drug classes (Table 2). Selected compounds from Table S2 were subjected to dose response testing against T. cruzi amastigotes with EC50 values shown in Table S2 and Table 3. We prioritized a set of compounds that exhibited potency in the in vitro screen and represented FDA-approved drugs with substantial clinical use (to exclude poorly-characterized candidates with the potential for toxicities).
Most of the compounds had EC50 values in the 1–10 µM range with the exception of clemastine, primaquine, and simvastatin which had high nanomolar EC50s. It was our judgment that the compounds probably lacked sufficient anti-T. cruzi potency to be curative as monotherapies in the animal model of T. cruzi infection. (In vivo data shown below supported this assumption). As a result, we turned to the possibility that some of these compounds (and several additional drugs known to have activity on T. cruzi) might by synergistic with each other and this could lead to combinations for effective chemotherapy. The investigations of this hypothesis are described in the following section.
Twenty-four compounds were selected for synergy testing (Table 3). These included 17 from the Spectrum Collection screen (#1–17) and an additional 7 compounds selected from the literature (#18–24). The compounds were picked for the following reasons: 1) potency in screening assays (EC50<10 µM), 2) orally route of administration (except for pentamidine), 3) diversity of drug class, and 4) established history of safe clinical use (exceptions being JK-11 and Ro 48-8071 which are not registered drugs). The 24 compounds in Table 3 were subjected to testing in every possible combination. The data are shown in a matrix (Figure S2) that is heat-mapped based on the “proportional effect” of the drug pairs as described in the Methods. We obtained usable synergy data for 297 of the 300 drug pairs. Of these pairs 232 (79%) showed positive proportional effects >0% and 63 (21%) showed negative proportional effects (≤0). An example of a pair showing apparent synergism is cloperastine and clemastine (proportional effect of 88%). This was calculated as follows: cloperastine alone resulted in 79% of normal growth, clemastine alone allowed for 74% of normal growth. The predicted growth is the product of these two observations (0.79*0.74 = 0.58). However, the combination actually resulted in 7% of normal growth. Using equation 1 in the Methods, the calculated proportional effect is 88% (with a maximum possible proportional effect of 100%).
Twenty-three combinations that appeared to show the most synergism were next tested in formal isobologram analyses in order to quantify the interactions by this standard method. The sum of fractional inhibitory concentrations (FICs) for various combinations are listed in Table 4. Eight drug combinations were confirmed to be synergistic by having the sum of FICs less than 0.5. Four of these involved the antihistamine compound clemastine and four involved the sterol 14-demethylase inhibitor JK-11. We added another sterol 14-demethylase inhibitor, posaconazole, to these combination studies since it is now of special interest in clinical trials for treatment of Chagas disease. Like JK-11, it was also found to be synergistic with clemastine. However, fourteen of the combinations had sum of FICs above the 0.5 cut-off and thus were merely additive in the interaction rather than synergistic. The isobologram graphs are shown in Supplementary material, Figure S3.
Selected drugs identified in the above screens were tested alone or in combination in the mouse model of T. cruzi infection. In the first experiment, we focused on posaconazole and benznidazole because of their advanced clinical status for treating Chagas disease. Since benznidazole and posaconazole are known to have curative activity as monotherapies, we used sub-curative doses so that additive or synergistic interactions could be detected when used in combinations. The other drugs were administered at doses described in the literature for treating mice. Dosing schedules are listed in Table 5. Briefly, mice were gavaged once or twice daily with a given drug or combination on days 7–11 post-infection. We conducted a second experiment (Figure 3) examining the same drugs with the purpose of confirming and expanding upon the initial results shown in Figure 2.
As intended, posaconazole and benznidazole given alone at the indicated doses cause some attenuation of parasitemia compared to vehicle-treated controls. Clemastine (5 mg/kg or 100 mg/kg) and amlodipine (10 mg/kg) given as monotherapies show no differences compared to the vehicle treated mice. Of the dual therapies tested, the most potent combination was the calcium channel blocker, amlodipine, plus posaconazole, which resulted in a nearly complete suppression of parasitemia and 80–100% survival (Figures 2 and 3). The combination of posaconazole plus clemastine suppressed parasitemia to a lesser extent, whereas the combination of posaconazole and benznidazole was not substantially different from posaconazole alone (Figures 2 and 3). Administering clemastine to the mice twice per day along with posaconazole was marginally better than administering clemastine and posaconazole once per day (Figure 3).
A third experiment shown with supplementary data (Figure S4) demonstrated a similar result in which posaconazole plus amlodipine is the most synergistic combination followed by a modest effect of combining benznidazole and posaconazole. In this experiment, we observed a lower mortality rate with the T. cruzi infection possibly due to variation with preparing or injecting the parasites. A final mouse experiment (Figure S5) investigated additional combinations as suggested by the in vitro experiments such as mefloquine plus clemastine, mefloquine plus amiodarone, and amiodarone plus clemastine. Unfortunately, none of these combinations showed any effect above vehicle treatment.
The Microsource Spectrum collection of 2000 compounds yielded a high hit rate in the primary screen with approximately 40% of compounds causing growth inhibition greater than 3 standard deviations above control levels. This is not surprising considering the nature of the library (known bioactive compounds) and the fact that compounds with toxicity to mammalian cells will necessarily result in inhibition of intracellular T. cruzi growth. We took three steps to eliminate compounds of low interest. First, we required at least 75% inhibition of intracellular growth 10 µM which we considered sufficient potency to be biologically interesting. Next we eliminated compounds that were not candidates for drug development, such as known toxins or drugs with only parenteral routes of administration (Table 1). And third, we rescreened the active compounds against host 3T3 cells to eliminate those with >33% inhibition at 10 uM and thus causing non-specific toxicity. The result was 151 compounds (7.7% of the original set) falling into a variety of categories shown in Table 2. The largest group of compounds (90) was non-drug natural products, which were not further considered for the current purposes since they are not established drugs. These compounds may remain of potential interest for de novo drug development or target identification. Of the remaining 61 drugs/compounds, psychotropic drugs are prominent in the hit list (Table S2). These included several phenothiazines such as thioridazine and chlorpromazine, which have been reported in other studies of trypanosomes [5], [15]–[21]. There is evidence that phenothiazines act on T. brucei by inhibiting trypanothione reductase [22]. Phenothiazines have been shown to cause direct lysis of T. cruzi trypomastigotes [23]. Further development of phenothiazines as antichagasic agents has probably not been rigorously pursued due to concerning side effects of this drug class and the narrow therapeutic window between parasite and host cytotoxicity.
Tricyclic compounds such as nortriptyline and clomipramine also appeared as hits in our screens. As with phenothiazines, these compounds have been previously reported to inhibit growth of T. cruzi [24], including a study showing activity of clomipramine in the mouse model of chronic T. cruzi infection [25]–[27]. The tricyclic antidepressants, similar to phenothiazines in structure, have also been shown to inhibit trypanothione reductase [28]. Finally, amongst psychotropic drugs, three selective serotonin reuptake inhibitors (SSRI) had selective anti-T. cruzi activity: fluoxetine, paroxetine, and sertraline. The EC50 values were fairly modest, in the 2–6 µM range, which suggests that on their own they may not be sufficiently potent to be used as anti-T. cruzi agents since therapeutic blood levels of these drugs in humans are typically in the 0.1–2 µM range and they tend to be highly protein bound (information from package inserts). There is at least one other study reporting an SSRI (fluoxetine) with anti-T. cruzi activity (EC50 = 7 µM) [5].
Among antihistamine drugs some familiar compounds such as azelastine (EC50 = 2.2 µM) and clemastine (EC50 = 0.4 µM) were identified in the screens. Azelastine was also identified in the high-throughput screen by Engel et al. [5]. Such compounds are interesting because of their favorable safety profile (they are used as over-the-counter drugs) although at normal doses blood levels are probably not high enough to mediate potent anti-parasitic activity. The idea of combining antihistamines with anti-T. cruzi activity with drugs such as nifurtimox has appeal since it is common that antihistamines need to be provided to control side effects such as skin reactions.
Several cardiovascular drugs were also identified in the screen, including the dihydropyridine calcium channel blockers nicardipine (EC50 = 5.9 µM) and amlodipine (EC50 = 1.1 µM). These have been previously reported to show inhibitory activity against both Leishmania species and T. cruzi with a selectivity index over mammalian cells around 7–9 [29]. A mechanism of action has not been defined. Prazosin and reserpine also had EC50 values slightly less than 10 µM in our screen. Since therapeutic levels of these drugs in humans are lower than these EC50 values, it is unlikely that they could be effective when used alone for treating T. cruzi infection. Finally, the antiarrhythmic drug, amiodarone, was identified in the screen. This drug was previously reported to have intrinsic anti-Trypanosoma cruzi activity [30], [31], which is particularly fortuitous since amiodarone is frequently used to help manage the arrhythmias that are common in Chagas disease. There is evidence that amiodarone inhibits an enzyme in the ergosterol biosynthesis pathway (oxidosqualene cyclase) and has synergistic activity with posaconazole [30]. With all of these cardiovascular drugs, there needs to be special caution when considering their use in patients with Chagas disease due to the potential to exacerbate underlying cardiac problems.
Not surprisingly, several of the antifungal drugs in the library were the most potent compounds in the screen including ketonazole (EC50 = 0.001 µM) and amphotericin B (EC50 = 0.04 µM). Azole drugs such as ketoconazole bind the sterol C14-demethylase enzyme (CYP51) and inhibit sterol biosynthesis [32]. Amphotericin B is thought to act by binding to ergosterol [33], a sterol that is not present in mammalian cells but is a critical component of the T. cruzi cell membrane. A liposomal preparation of amphotericin B was shown to have suppressive in vivo activity in mice with T. cruzi infection [34], but further development for treating human Chagas disease has not been pursued. As discussed in the Introduction, the repurposing of azole antifungal drugs for Chagas disease, in particular posaconazole, is now in human clinical trials [35].
The following antimalarial drugs were identified in the screen: mefloquine, primaquine, artemisinin, hydroxychloroquine, and pyrimethamine. Mefloquine has been shown to have anti-T. brucei activity in the mouse model [36], but we are unaware of data for T. cruzi. The 8-aminoquinolone compound class (including primaquine) has previously been tested against trypanosomatid parasites, including T. cruzi [37]–[42]. Beyond studies in the mouse model of T. cruzi infection [40], further investigations for use in Chagas disease have not been published. Artemisinins have also been previously shown to have in vitro activity against T. cruzi and T. brucei in the low micromolar range [43], but further development has not been reported. Pyrimethamine (EC50 of 3800 nM) is a known inhibitor of dihydrofolate reductase-thymidylate synthase which has been shown to be essential in the African trypanosome [44]. Pyrimethamine was not particularly potent against T. brucei with an EC50 of 17 µM [44], but due to the lower EC50 on T. cruzi further investigation may be warranted.
Three more compounds from the screen merit further discussion: triamterene, oxyphenbutazone, and minocycline. Triamterene (EC50 of 1660 nM) is a widely used diuretic that blocks the epithelial sodium channel in the renal collecting tubule. It also is an inhibitor of folate metabolism [45] and has been shown to have modest activity (48 µM) against T. brucei but we have not found reports of it being tested against T. cruzi. Oxyphenbutazone (EC50 of 12,000 nM) is an active metabolite of the nonsteroidal anti-inflammatory drug phenylbutazone which is used for veterinary purposes but not in humans due to risk of agranulocytosis. Its activity against T. cruzi has not been previously reported to our knowledge. Finally, the antibiotic minocycline was found to have an EC50 of 9800 nM in our assay. This drug has been described to have activity in the mouse model of T. brucei infection [46], [47]. The related drug, tetracycline, has little or no inhibitory activity on trypanosomes (in the low micromolar concentrations used for the tetracycline inducible genetic systems for studying the trypanosomes). The mechanism of action of minocycline in T. cruzi is unknown, but it could potentially bind the small subunit of the kinetoplast ribosome a similar mechanism to its effects in prokaroytes [48].
From the subset of 53 active drugs (Table S2) we selected 17 for synergy testing (Table 3, #1–17). Another 7 drugs/compounds of particular interest were added to the list (Table 3, #18–24). These included the clinical drug for Chagas disease, benznidazole. Considering the well-described problems with tolerability and efficacy of benznidazole, we were interested in establishing whether a second drug could be combined with a lower dose of benznidazole to improve efficacy. We nominated several drugs that target the sterol biosynthesis pathway, which is a well validated therapeutic target in T. cruzi [49]. These compounds included our preclinical candidate (JK-11) that inhibits CYP51 (sterol C14α-demethylase) [9], as well as the bisphosphonate drug, pamidronate, that inhibits farnesyl pyrophosphate synthase [50], the allylamine antifungal drug, terbinafine, which inhibits squalene epoxidase [51], and the oxidosqualene cyclase inhibitor, Ro 48-8071 [10], [52]. We also included pentamidine in the list. Pentamidine's mechanism of action is not entirely clear, but it is used clinically for African trypanosomiasis and leishmaniasis, and has oral analogs under development for trypanosomiasis [53]. These 24 compounds were tested for synergy in two-way combinations under strict conditions. In this assay, compounds were tested in quadruplicate at their EC25 concentrations both individually (to confirm that the compound was accurately assayed at its EC25) and in combination. We chose to study the selected compounds at the EC25 for two reasons. First, this concentration results in parasite growth inhibition that is substantially greater than the intrinsic (baseline) variance of the assay. Second, the relatively low concentration at the EC25 allows for a large range of growth inhibition to be observed such that synergistic activity can be detected if it is present.
More than 75% of the combinations showed positive interactions (green in the heat map, Figure S2) meaning that the combined effects were more than predicted by the equation shown in the Methods section. This does not necessarily mean that the interaction reached the level of being “synergistic”. Using the relative proportional effects in Figure S2, we ranked compound combinations for synergy potential. Based on the rankings, we were able to prioritize specific drug pairings for isobologram analysis.
Twenty-four combinations were tested on T. cruzi using the fixed-ratio method with results shown in Table 4. Eight drug combinations had an average FIC<0.5, which is considered “synergistic”, and all but four of the 23 had average FIC values <1.0. The four combinations with the lowest average FIC values included JK-11 as one of the paired drugs. Similarly, clemastine also appeared in 4 of the combinations reaching “synergy” levels. Unfortunately, clemastine does not appear to be synergistic with benznidazole with an average FIC of 1.20. In these experiments, we also investigated posaconazole which has the same target of action (the CYP51 enzyme) as JK-11. Both posaconazole and JK-11 were synergistic with clemastine, but the combination of posaconazole and amlodipine did not reach the synergy level (average FIC 0.645) that was observed with JK-11 and amlodipine (average FIC 0.367). Surprisingly, we did not observe synergy between posaconazole and amiodarone (average FIC 1.62), which had previously been shown to be synergistic [30]. This finding may be due to the use of different parasite strains, host cells, incubation times, or other experimental variables. Finally, the combination of posaconazole and benznidazole showed an average FIC of 0.91. Although this is not “synergistic”, the interaction falls in the “additive” range and reinforces the notion of testing these two drugs together as has been reported in mouse model [54] and in a clinical trial underway in Argentina (http://clinicaltrials.gov/show/NCT01377480).
Based on these results, we decided to test various drugs alone and in combinations in the mouse model of acute T. cruzi infection. Aside from benznidazole and posaconazole, none of the drugs had dramatic effects on parasitemia when used alone (although there were slight effects observed with allopurinol and amlodipine, Figure S4). This supported our view that these drugs would need to be tested in combination with other drugs in order to generate significant inhibitory effects in vivo. The most effective combination was posaconazole plus amlodipine (Figure 2), a result we confirmed in additional experiments (Figure 3 and S4). Parasitemia was dramatically suppressed in mice treated with amlodipine plus posaconazole or clemastine plus posaconazole, but was not completely eliminated with these combinations at the doses used. Since posaconazole was dosed well below the maximum tolerated dose, parasitemia was only partially suppressed by the posaconazole part of the combination. In vitro, the combination of posaconazole plus amlodipine was borderline synergistic (average FIC 0.645), thus it is possible that a biological interaction is occurring that results in the favorable combined effect on parasitemia of these two drugs in vivo. However, it is also known that both of these drugs are metabolized by a common liver enzyme, CYP3A4, thus it is also possible that the interaction is pharmacological in that amlodipine may be boosting blood/tissue levels of posaconazole (or vice versa). The strategy of using pharmacological interactions to boost drug activities is being seen more commonly, for example with the use of ritonavir or cobicistat in antiretroviral combination therapies involving protease inhibitors [55]. Further studies will be necessary to better characterize the interaction of posaconazole and amlodipine.
The combination of posaconazole and clemastine boosted suppression of parasitemia (Figure 2 and 3). It is not clear if there is a pharmacological interaction between these two drugs in mice or in vivo synergy on the parasites. This combination was synergistic in vitro with an FIC of 0.46. The combination of posaconazole and benznidazole showed only a modest boost in parasitemia suppression in both experiments (Figure 2, 3, and S4), somewhat less favorable than described in another recent report [54].
Some combinations that were synergistic in vitro did not demonstrate similar effects in vivo such as clemastine + mefloquine and clemastine + amiodarone (Figure S5). It seems most likely that sufficiently high blood and tissue levels are not being achieved or maintained to produce the desired effect, but further investigation is needed. We also looked at some combinations that were merely additive in vitro, but nonetheless seemed like interesting partners to test in vivo, such as allopurinol plus posaconazole. We did not observe a positive interaction with this combination (Figure S4). Similarly, the combination of benznidazole and clemastine did not appear to show a positive interaction in mice (Figure S4) nor did the combination of mefloquine and amiodarone (Figure S5).
These experiments show positive interactions between some well-established drugs in a mouse model of acute T. cruzi infection. Many more combinations that were identified in the in vitro experiments have yet to be tested in vivo, so future studies may reveal even more potent drug combinations. Future studies will also need to focus on whether the combination chemotherapy can lead to parasitological cures in mice. As noted above, we deliberately used low doses of benznidazole and posaconazole in these studies in order to facilitate observing effects on bloodstream parasitemia. In future studies, we plan to determine if combining off-the-shelf drugs can allow us to use shorter courses or lower than maximum doses of benznidazole or posaconazole to cure T. cruzi infected mice. The ultimate goal would be to identify new treatments based on combination therapy that are more effective, better tolerated, and simpler to administer than current regimens for treating Chagas disease.
|
10.1371/journal.pcbi.1004161 | Perturbation-Expression Analysis Identifies RUNX1 as a Regulator of Human Mammary Stem Cell Differentiation | The search for genes that regulate stem cell self-renewal and differentiation has been hindered by a paucity of markers that uniquely label stem cells and early progenitors. To circumvent this difficulty we have developed a method that identifies cell-state regulators without requiring any markers of differentiation, termed Perturbation-Expression Analysis of Cell States (PEACS). We have applied this marker-free approach to screen for transcription factors that regulate mammary stem cell differentiation in a 3D model of tissue morphogenesis and identified RUNX1 as a stem cell regulator. Inhibition of RUNX1 expanded bipotent stem cells and blocked their differentiation into ductal and lobular tissue rudiments. Reactivation of RUNX1 allowed exit from the bipotent state and subsequent differentiation and mammary morphogenesis. Collectively, our findings show that RUNX1 is required for mammary stem cells to exit a bipotent state, and provide a new method for discovering cell-state regulators when markers are not available.
| The discovery of stem cell regulators is a major goal of biological research, but progress is often limited by a lack of definitive markers capable of distinguishing stem cells from early progenitors. Even in cases where markers have been identified, they often only enrich for certain cell states and do not uniquely identify states. While useful in some contexts, such enriching markers are ineffective tools for discovering genes that regulate the transition of cells between states. We present a method for identifying these cell state regulatory genes without the need for pre-determined markers, termed Perturbation-Expression Analysis of Cell States (PEACS). PEACS uses a novel computational approach to analyze gene expression data from perturbed cellular populations, and can be applied broadly to identify regulators of stem and progenitor cell self-renewal or differentiation. Application of PEACS to mammary stem cells resulted in the identification of RUNX1 as a key regulator of exit from the bipotent state.
| Adult stem cells are functionally defined based on their ability to regenerate tissues. This unique regenerative ability can be recapitulated in culture models, where single stem cells, but not differentiated cells, form tissue rudiments in three-dimensional extracellular matrices. These tissue rudiments, or organoids, exhibit many of the topological, functional and phenotypic traits of the corresponding tissue. For example, mammary stem cells form ducts and lobules in collagen matrices that resemble structures present in the breast [1–3], while colon stem cells form mini-crypts in Matrigel that resemble analogous structures in the small intestine [4].
Given their potential for regenerative medicine, there is significant interest in identifying genes that regulate self-renewal or differentiation of stem cells. In systems with well-defined markers of stem, progenitor and differentiated states, this can be accomplished by inhibiting candidate genes and assessing the resulting effects on cell state proportions [5]. However, for many tissues markers of stem cells and early progenitors are not available, and even in cases where such markers are available they often only enrich for states of interest. This lack of defining markers has complicated efforts to screen for cell-state regulators, because changes in the number of cells expressing an enriching marker may not quantitatively reflect changes in the stem or progenitor cell types of interest.
We have addressed this difficulty by developing a new approach that identifies cell state regulators without requiring defining markers of cell state, termed Perturbation-Expression Analysis of Cell States (PEACS). Application of PEACS to mammary stem cells led to the discovery of a novel role for RUNX1 in exit from the bipotent state. We anticipate that PEACS will be useful in the many contexts where defining markers are not available, and have implemented the algorithm as a software tool available to the scientific community.
The analysis underlying PEACS is based on several observations. First, populations of stem cells propagated in culture are heterogeneous, and invariably include early progenitors and other more differentiated cell types. While typically considered a drawback of maintaining stem cells in culture, this heterogeneity is essential for the computational analysis underlying PEACS. Second, experimental conditions that perturb transitions between stem and progenitor states will also perturb the relative proportions of stem and progenitor cells in a heterogeneous population of cells. For example inhibiting a gene required for stem cell self-renewal will reduce the proportion of stem cells in a heterogeneous population, with a concomitant relative increase in progenitors or other more differentiated cell types.
The computational challenge then is to use the population expression vectors—one for each perturbation—to infer which perturbations modulate cell-state proportions. However, without knowing either the cell state proportions or the gene-expression vectors of the individual states, it may appear that there is insufficient information to make such an inference. The solution lies in a third key observation: the gene-expression profiles (vectors) of heterogeneous populations of cells are weighted linear combinations of the expression profiles (vectors) of the component states within the population, with the weights in this linear combination corresponding to cell-state proportions. In other words, the gene-expression signal of the population is a linear mixture of component signals, the latter of which are unknown. The key is to deconvolute this signal (Fig 1).
Several computational algorithms have been designed precisely for this purpose—to infer the constituent components of mixed signals—under the assumption that the mixed signal is a weighted linear combination of constituent components. The most commonly used algorithm to infer linear components, SVD/PCA, iteratively minimizes the reconstruction error of a mixed signal, under the constraint that the component newly identified in a given iteration be orthogonal to all of the previously identified components. Given the immense success of SVD/PCA in solving many problems across diverse fields, we decided to assess its effectiveness for our problem. A second algorithm, NMF, reconstructs mixed signals by identifying components which have only non-negative loadings. Some researchers have found this non-negative constraint to be appealing, since negative loadings of genes can be difficult to interpret biologically; for this reason we also included this method for comparison. A third algorithm, ICA, does not require that the constituent components be orthogonal to one another—and instead identifies components by maximizing their independence in a statistical sense. ICA has proven useful for deconstructing mixed signals (e.g., audio) into their constituent parts.
Although our goal in developing PEACS was to apply it in settings where neither the state expression vectors nor cell-state proportions are known, to assess the effectiveness of the algorithms described above (SVD, NMF, ICA) we needed an idealized context in which cell-state proportions could be experimentally defined. Experimentally defining cell-state proportions would make it possible to assess, for each algorithm, how well it identified changes in cell-state proportions across experimental conditions. To generate such idealized experimental conditions we mixed three different breast cancer cell lines (T47D, SUM159, MDA-MB-231) in defined proportions—for example 1:1:1, 1:2:2, 1:1:0—with 10 mixtures in total. In this idealized experiment the three cancer lines represented different “cell states” that were mixed in defined proportions to create heterogeneous populations (Fig 2A; T47D = State A, MDA-MB-231 = State B, SUM159 = State C). We isolated total mRNA from these heterogeneous populations and profiled the expression of 17 differentiation-related genes and GAPDH, thereby generating a gene-expression profile for each heterogeneous population (S1 Table). Lastly, we applied SVD, NMF and ICA to the gene expression matrix to assess the relative performance of these algorithms in identifying changes in cell-state proportions.
The results of the SVD, NMF and ICA analyses are presented in Fig 2B–2D. SVD/PCA successfully identified components that closely correlated with the proportions of the cell states in our idealized experiment: the first component exhibited a strong negative correlation with the fraction of cells within the population in State A (r2 = 0.92), while the second component correlated with the fraction of cells in State B (r2 = 0.47). Additionally, the replicates for each perturbation clustered closely together in the space spanned by these first two components identified by the SVD/PCA algorithm (Fig 2B right). Moreover, the first two SVD components together explained ~90% of the variation in the gene-expression data (as can be seen by the Scree plot in S1A Fig), which is consistent with the two degrees of freedom inherent in the design of this idealized experiment. In contrast to SVD/PCA, the two components identified by NMF both correlated strongly with the fraction of cells in State A (r2 = 0.92 and 0.92 respectively)—with component 1 correlating negatively with the proportion of cells in State A, and component 2 correlating positively with the proportion of cells in State A (Fig 2C, S1E Fig); neither NMF component 1 nor 2 was correlated with states B or C (all r2 < 0.43; S1E Fig). For this analysis the NMF factorization was performed with parameter k = 2, because the two components together explained over 95% of the variance in the gene-expression data (S1B Fig). As was the case for the SVD/PCA algorithm, the replicates for each perturbation clustered closely in the space spanned by the two components identified by NMF; this strongly suggested that the components identified by the algorithm reflected biological signal rather than experimental noise. Unlike the SVD/PCA and NMF algorithms, the first two ICA components did not correlate with the fraction of cells in any of states A, B or C (all pairwise r2 < 0.13, Fig 2D, S1F Fig). Moreover, in almost all cases the various replicates for a given perturbation did not cluster together in the space spanned by the first two components identified by ICA (Fig 2D right).
Collectively these observations indicated that both the SVD/PCA and NMF algorithms effectively identified components that correlated strongly with cell-state proportions, while ICA failed to do so. Moreover, these observations showed that only the SVD/PCA components spanned the 2 degrees of freedom inherent this idealized experiment, which, by design, involved cellular populations that were mixtures of exactly 3 cell states.
One potential explanation for why the SVD and NMF components tracked cell-state proportions is that the components were identifying genes differentially expressed between cell states. We could directly compare gene loadings in the various components with gene expression in the various states because the gene-expression profiles of the pure states were known in our idealized experimental conditions (Fig 2E). This comparison revealed that genes with the highest loadings in SVD component 1 were uniquely expressed or repressed in state A; this was consistent with the observation that this component tracked with the fraction of cells in state A. Similarly, NMF components 1 and 2—both of which also tracked with state A—identified a very similar set of genes uniquely expressed by state A (Fig 2E). A key difference, however, was that unlike SVD component 1, which included positive and negative loadings corresponding respectively to genes down or up in state A, both of the NMF components had only positive gene loadings—with NMF component 1 having positive gene loadings for the genes down in state A, and NMF component 2 having positive loadings for the genes up in state A (Fig 2E). In contrast, SVD component 2 identified the only two genes that were strongly differentially expressed between states B and C (HOXA5, FOXO1; Fig 2E); these two genes, HOXA5 and FOXO1, were respectively down and up in state B relative to state C, and were expressed near median levels in state A. Thus, the highest loadings of SVD1 in this idealized experiment marked genes differentially expressed between luminal and basal cells, including the established luminal markers GATA3 and STAT5A. More generally, these findings suggested that the highest loadings in the SVD component vectors may serve to identify markers of specific cell states in contexts where such markers are not known.
Since our goal in developing PEACS was to identify perturbations that affect cell state proportions, we needed a method for reducing the SVD component weights to a single score that quantifies the extent of change in cell-state proportions. For this purpose the Euclidian metric, which corresponds to the natural notion of ‘distance’ in 1, 2 and 3-dimensional space, was attractive for several reasons. First, we expect distances in SVD space to scale linearly with the extent of the change in cell state proportions. Consistent with this, analysis of the SVD1 v SVD2 replicate plot for the idealized experiment (Fig 2B right panel) revealed that small perturbations in cell state proportions (e.g. 1:1:1 to 1:2:2) resulted in small distances in component space, whereas large changes in cell state proportions (e.g. 1:1:1 to 0:1:1) resulted in large distances in SVD component space. Second, the Euclidian metric makes it straightforward to quantify how noise in the various dimensions impacts the reliability of multidimensional distance estimates.
We therefore used the Euclidean metric to compare distances between samples in the space spanned by the first k SVD components, where k was chosen using the standard approach of looking for an ‘elbow’ in the corresponding Scree plot. To account for biological variability across replicates (or different shRNAs targeting the same gene), we defined the PEACS score as the Euclidean distance divided by the standard error about the mean for each set of replicates (Fig 3).
Intuitively, this PEACS score—Euclidean distance divided by standard error—can be thought of as a ‘signal-to-noise’ ratio, which scales the magnitude of a change by the error in the distance estimate. Empirical p-values for PEACS scores were determined by Monte Carlo sampling: for a given perturbation with n replicates, a null distribution was obtained by randomly sampling n expression profiles from the experimental data, calculating a PEACS score, and iterating this process 10,000 times to generate a PEACS score null distribution. The empirical p-value was then determined by ranking the PEACS score for the given perturbation relative to the PEACS scores generated by this Monte Carlo procedure.
We next applied PEACS to the MCF10A human stem cell model of mammary morphogenesis [6]. When seeded into a three-dimensional collagen matrix, MCF10A cells form ductal, lobular, and ductal-lobular tissue rudiments (Fig 4A–4C). These tissue rudiments are monoclonal, indicating that they arise from single stem cells, and are morphologically similar to structures present in the human mammary gland (Fig 4C; S2 and S3 Fig; S1 Movie).
As a first step, we used gene-expression profiling to identify 39 developmentally implicated transcription factors (TFs) expressed in MCF10A cells (S2 Table). We next inhibited these factors with 3–5 shRNAs targeting each TF, with two biological replicates per shRNA, resulting in a total of 240 genetically perturbed lines. For each genetically perturbed line, we then profiled the expression of all 39 factors and housekeeping genes using high-throughput qRT-PCR. These experiments generated a large data matrix with rows corresponding to gene expression values, and columns corresponding to shRNA perturbations.
From this data matrix, we eliminated genes that were not inhibited by at least 3 distinct shRNAs. Application of the PEACS algorithm to this filtered data matrix produced a score that quantified the extent to which TF inhibition affected cell-state proportions. Based on this PEACS score, most genetic perturbations had small effects on cell state proportions, which were comparable to the effects of hairpins that did not successfully knockdown their targeted genes (Fig 5A, S3 Table). When inhibited, several genes caused large, reproducible changes in cell state proportions, which could be seen when the perturbations were plotted in 3D SVD component space or as PEACS scores (Fig 5A and 5B). We used the first three SVD components for this analysis because the elbow of the Scree plot occurred at three dimensions (S1C Fig).
The top three factors identified by this analysis were NR3C1, RUNX1 and TCF3 (Fig 5C, S3 Table). Identification of the glucocorticoid receptor (NR3C1), the highest-scoring factor, was significant because of its established role in regulating mammary ductal differentiation and lactation [7]. TCF3, the third-highest scoring factor, was recently reported to be a mammary stem cell regulator [8]. RUNX1, which was the second-highest scoring factor, is mutated in a subset of breast cancers but has not been previously implicated as a regulator of mammary stem cell biology [9–11]. Since the other hits identified by PEACS were established regulators of mammary stem cells or differentiation, we suspected that RUNX1 might also play a role in one or both of these processes, and therefore decided to further explore its function.
In this dataset, RUNX1 primarily affected the expression of SVD component 1. We therefore investigated the loadings of SVD component 1 to identify the genes that have the highest contribution to this component (Fig 5D). The highest loadings of SVD component 1 were ETS1, HIF1A, HOXA5, NFYA, RUNX1, YY1, and RB1. As expected, these genes were significantly decreased in the RUNX1 knockdown condition compared to perturbation conditions that did not change SVD component 1 (Fig 5D). While we do not know what the state corresponding to SVD component 1 is, these markers may be useful for future studies investigating mammary lineages.
To evaluate the functional role of RUNX1 we inhibited its expression with shRNAs (Fig 6B) and assessed the ability of MCF10A cells to form tissue rudiments in polymerized collagen. RUNX1-inhibited cells formed spheres that did not hollow (Fig 6D), indicating that they were not mature lobules, and rarely formed ducts or ductal-lobular rudiments (71% reduction relative to control); the rare ducts that did form were shorter in length (25% reduction) and did not exhibit the branched morphology seen in wild type structures (Fig 6A, 6C). As a control, cells that were either mock-infected or expressed a control shRNA were not affected in their ability to form tissue rudiments. These results indicated that RUNX1 is required for mammary cells to differentiate into ducts and mature lobules.
To assess if the phenotype caused by RUNX1 inhibition was reversible, we generated an MCF10A line in which RUNX1 could be reversibly inhibited by a doxycyline (dox)-inducible shRNA (Fig 7B). When cultured in collagen in the presence of dox, these MCF10A cells formed solid spheres and few ducts, recapitulating the phenotype observed above when RUNX1 was constitutively inhibited by shRNAs. When RUNX1 was re-expressed by withdrawing dox, the spheres rapidly sprouted ducts and began to hollow—often within 12–24 hours (Fig 7A). This finding indicated that the RUNX1-inhibited spheres were still capable of forming both ducts and lobules upon RUNX1 re-expression, raising the possibility that these spheres might consist of bipotent cells reversibly arrested in their differentiation.
To directly examine this possibility we assessed whether single cells from RUNX1-inhibited spheres could form tissue rudiments when seeded into collagen. Parental MCF10A cells largely lose this ability upon differentiating in collagen (Fig 7C). We seeded cells with dox to form RUNX1-inhibited spheres, harvested and dissociated the spheres by treatment with collagenase and trypsin, and then reseeded single cells into collagen with or without dox. Cells reseeded in dox again gave rise to solid spheres. However, those reseeded without dox formed lobules and ducts that matured into complex ductal-lobular structures (Fig 7C), doing so with efficiency comparable to that of parental MCF10A cells maintained in 2D culture.
These observations strongly suggested that parental MCF10A cells dissociated from tissue rudiments lost the ability to reseed tissue rudiments because they had differentiated and lost stem and progenitor activity; in contrast, cells within RUNX1-inhibited MCF10A spheres maintained their ability to reseed tissue rudiments because they did not differentiate in collagen and remained bipotent.
We next examined if RUNX1 also affected the differentiation of primary human breast stem and progenitor cells. To this end we isolated primary human breast epithelial cells from reduction mammoplasty tissue samples, modulated RUNX1 expression, and assessed stem and progenitor cells using colony forming assays (Fig 8A) [12–14]. In these assays the majority of stem and progenitor cells form colonies containing differentiated luminal or basal cells. However a fraction of bipotent stem cells proliferate but do not differentiate; these form micro-colonies of 2–16 cells that remain uncommitted and co-express both luminal and basal markers.
Inhibiting RUNX1 expression caused a 2-fold increase in the number of stem cell micro-colonies, suggesting that this transcription factor was required for primary human breast stem cells to differentiate in culture (Fig 8B). Consistent with this interpretation, inhibiting RUNX1 expression reduced the number of differentiated colonies by nearly 90%, while its over-expression led to a 300% increase in differentiated colonies.
We next examined whether transiently inhibiting RUNX1 would expand the population of functional stem cells in culture. For this experiment we first infected primary cells with the dox-inducible shRUNX1 lentivirus, and plated cells with dox to assay for colony-forming ability. After micro-colonies of stem cells had formed (7 days after plating), we removed the dox so that RUNX1 would be re-expressed. We found that re-expressing RUNX1 caused the stem cell micro-colonies to differentiate within 48–96 hours, and resulted in the formation of heterovalent colonies that included both bipotent stem cells and lineage-committed basal and luminal cells (Fig 8C). These heterovalent colonies were never observed in colony-forming assays with control primary cells, or in assays with primary cells in which RUNX1 had been stably inhibited.
Collectively, these findings indicate that RUNX1 inhibition enables primary breast stem cells to expand in an uncommitted state while retaining the functional ability to differentiate in culture.
We have shown that PEACS identifies perturbations that affect cell-state transitions, taking as input the gene-expression profiles of perturbed cellular populations. We validated PEACS by applying it to a mammary stem cell model with shRNAs as a source of perturbations. In this context, the method identified several established regulators (e.g., NR3C1 and TCF3) of mammary stem cell biology, as well as a novel gene, RUNX1, which had not previously been implicated as a mammary stem cell regulator. Follow-up studies revealed that inhibiting RUNX1 prevented mammary stem cells from differentiating, indicating that this gene is required for stem cells to exit a bipotent state. Although our study focused on shRNA perturbations, there is every reason to believe that PEACS would be equally effective for gene over-expression or chemical perturbations.
Several computational methods for analyzing gene-expression profiles have been previously reported [15–19]. PEACS differs from these in three important ways. First, the goal of PEACS is to specifically identify perturbations that influence how cells transition between differentiation states; we are not aware of other methods that do this. Second, the method does not require any markers of stem, progenitor or differentiated states. Third, our method analyzes bulk populations of cells to identify changes in cell state ratios, rather than analyzing large numbers of single cells.
We anticipate that this marker-free approach will be particularly useful in the many contexts where stem, progenitor, and differentiated cells have been identified functionally, but where markers that distinguish these states are not yet available. It is worth emphasizing that, although markers that enrich for stem and progenitor states have been identified in many systems, few systems offer markers that sort stem or progenitor cells to purity; this latter ability is essential if these markers are to be used to identify genes that regulate state transitions. In cases where such markers are in fact available—or when they are used to define states de facto without consideration of the underlying biology—we have previously shown that a Markov model can be used to quantify the rates of transition between states, and predict the equilibrium proportions of cell states [20].
Targeting RUNX1 may offer unique possibilities for therapeutic applications. Stem cells have a strong tendency to differentiate when propagated in culture, even under conditions that are intended to maintain them in an undifferentiated state. This problem has been observed with human ES cells, HSCs, and many other stem cell types. We have shown that primary human mammary stem cells can be expanded in a bipotent state by transiently inhibiting RUNX1; moreover these cells spontaneously differentiate once RUNX1 expression is re-established. A chemical compound that inhibits RUNX1 could therefore be used to propagate mammary stem cells in culture. It will be of interest to examine if inhibiting RUNX1 can also prevent other types of stem cells from exiting a bi- or multipotent state. In support of this possibility, a dominant-negative RUNX1 translocation has been found in a subset of leukemias, and expression of this protein blocks the differentiation of leukemic cells and promotes the self-renewal of hematopoietic stem cells [21,22]. Additionally, a RUNX1 ortholog, Runt, has been shown in planaria to be required for neoblast stem cells to differentiate at wound sites [23]. Taken together, these observations suggest the intriguing possibility that this function of RUNX1/Runt is conserved across species and cell types.
Primary tissues were obtained with consent in compliance with laws and institutional guidelines, as approved by the Institutional Review Board of Maine Medical Center. Exemption status for human research was obtained from the Committee on the use of Humans as Experimental Subjects (COUHES) at MIT, based on de-identification of the samples. All patient samples are de-identified prior to distribution for research use. The data collected and stored is limited to basic demographic data, specimen handling information (ex: related to chain of custody), specimen quality data, and histopathologic data. At no time is any patient identifier provided to any researcher.
MCF10A cells were obtained from ATCC and cultured in MEGM with 100 ng/ml cholera toxin, GlutaMax, Penicillin and Streptomycin (Lonza CC-3150). HEK293T cells were maintained in DMEM supplemented with 10% FBS, GlutaMax, Penicillin and Streptomycin. SUM159 (Asterland), MDA-MB-231 (ATCC), and T47D (ATCC) cells were cultured in DMEM with 10% FBS, GlutaMax, Penicillin and Streptomycin.
Human organoids were isolated from breast tissues from patients undergoing elective reduction mammoplasty. Primary tissues were obtained with consent in compliance with laws and institutional guidelines, as approved by the Institutional Review Board of Maine Medical Center. Organoids were aliquoted in 1:1 DMEM/Hams-F12 media supplemented with 5% calf serum, 10 ng/mL insulin, 10 μg/mL epidermal growth factor, 10 μg/mL hydrocortisone, and 10% DMSO and stored in liquid nitrogen. Doxycycline (dox), where applicable, was used at a concentration of 4μg/ml.
Lentivirus production, target cell infection, and selection were performed as previously described [24]. Constitutive shRNA plasmids in a pLKO.1 vector were obtained from the Broad Institute RNAi consortium (https://www.broadinstitute.org/rnai/trc3), and inducible hairpins (dox “ON”; pTRIPZ vector) were obtained from Thermoscientific. Overexpression constructs were obtained through gateway cloning of the appropriate ORF into the pLenti6.2-ccdB-3xFLAG-V5 construct.
MCF10A cells were seeded onto a 96 well plate at a density of 7500 cells per well and infected the next day with hairpin lentivirus targeting an expressed developmental transcription factor. One day after infection, cells were selected with 5ug/ml puromycin containing media. Two days later RNA was collected with the Qiagen RNeasy 96 Biorobot 8000 kit and cDNA synthesized with the iScript cDNA synthesis kit (BioRad 170–8890).
Microfluidic qPCR was carried out according to the manufacturer’s Protocol (Protocol 37: Fast Gene Expression Analysis Using EvaGreen on the BioMark or BioMark HD System). The 39 TFs profiled were selected by profiling gene-expression in MFC10As and selecting all TFs implicated in differentiation that were confirmed to be expressed by qPCR. The cDNA was preamplified for 14 cycles with a mix of 41 primer sets (39 TFs, BTub, and GAPDH) and mastermix, then treated with ExoI. Prior to analysis with PEACS, the data matrix with the Fluidigm CT values was normalized to GAPDH and median normalized by gene such that the median CT value for each gene was 0. For the idealized experiment, gene expression was profiled using standard qPCR and the 17 genes profiled were randomly selected transcription factors expressed by MCF10A cells and implicated in differentiation.
Let M be a data matrix of perturbation-expression values with rows corresponding to perturbations and columns corresponding to the genes whose expression was profiled. We used the reduced singular value decomposition to transform M, so that
M=[ ↑u1↓…↑un↓ ]︷mxn[ σ1⋱σn ]︷nxn[ ←v1t→⋮←vnt→ ]︷nxn=∑i=1nσiuivit
(1)
where u1, …, un and v1,…, vn are respectively the left and right eigenvectors corresponding to the singular values σi of the reduced singular value decomposition of M. We have assumed here that n < m, i.e., the number of perturbations exceeds the number of genes whose expression is profiled. Because the dimensions of ui and vi are (mx1) and (nx1), respectively, and the σi are scalars, M can be viewed as a weighted sum of the rank-one matrices uivit.
We then use the first k singular values and vectors to reconstruct a low-rank approximation of M:
M~[ ↑u1↓…↑uk↓ ]︷mxk[ σ1⋱σk ]︷kxk[ ←v1t→⋮←vkt→ ]︷kxn
(2)
The value k is chosen using a Scree plot, as described in the main text. In the case of k = 3:
The gene-expression vector for the perturbation p can therefore be approximated by the following weighted sum of the first 3 SVD eigenvectors:
Thus the gene expression data for each perturbation p is mapped into the space spanned by linear combinations of the first k gene-expression SVD eigenvectors v1,…, vk. Again for k = 3:
These coordinates in SVD-space are plotted as ‘Component scores’ in Figs 2B, 3, and 5A.
Finally, to determine the PEACS score, we first calculate the Euclidean distance between the u1p, u2p, u3p for a given perturbation p (averaged across all its replicates), and the median centroid vector (u¯1, u¯2,u¯3) taken across all m perturbations:
Finally, the PEACS score is calculated by dividing the distance in (Eq 6) by the standard error across replicates for a given perturbation.
To calculate a p-value, a Monte Carlo sampling algorithm was implemented. For each set of n perturbations, a null distribution of PEACS scores was obtained by sampling n random perturbations 10,000 times without regards for perturbation labels. The p-value was defined as the rank of the real PEACS score in the null distribution divided by 10,000.
The PEACS code for MATLAB is available as a supplemental file (S2 Text) and on our lab website at: http://guptalab.wi.mit.edu/.
7.5x103 MCF10A cells were resuspended in 0.2ml of collagen solution (1.25mg/ml rat tail collagen I in PBS, brought to pH 7.3 with 0.1N NaOH) and plated on a single chamber of a 4-chamber slide. Collagen was polymerized for 2 hours at 37°C, after which they were detached and cultured in 1ml of MCF10A medium.
MCF10A cells were grown in collagen matrix through day 7, at which time the collagen pads were collected and incubated in 100 ug/ml collagenase in PBS at 37°C for 10 minutes. The structures were collected by centrifugation (500 RPM, 5 min), resuspended in 0.25% trypsin, and incubated for 20–25 minutes at 37°C. Cells were counted in trypan blue, spun down (500RPM, 5 min), and resuspended in MCF10A media; 7500 living cells were reseeded into a new collagen pad.
Samples were fixed with 4% paraformaldehyde for 15 minutes at room temperature. Pads were permeabilized using 0.1% TritonX-100 and incubated with blocking solution (PBST with 10% goat serum and 3% BSA) for 1 hr at room temperature and stained with the appropriate primary antibody in blocking buffer for 1–2 hours at room temperature or overnight at 4°C. The samples were washed with PBS, and incubated with an Alexa Fluor-labeled secondary antibody. Samples were washed, stained with 1ug/ml DAPI. Images of phalloidin-AF594 and DAPI- stained collagen structures were analyzed by image segmentation software (CellProfiler; [25]), with an analysis pipeline that differentially detected lobules and ducts based on size, area and form factor adjustments.
Primary human organoids were thawed and plated on a 10cm dish in 10ml of RMFC (DMEM + 10% Calf Serum) media for 1–2 hours. The non-adherent fraction, fibroblast reduced organoids, was collected, spun 10 minutes at 233 gravity, resuspended in cold PBS and passed 10 times through an 18-gauge needle. The organoids were once again pelleted 5 minutes at 335 gravity, resuspended in 2ml of 0.05% trypsin, and incubated 10 minutes at 37°C. We then added 8ml of RMFC media and 0.5mg of DNaseI (Roche 10104159001). The cell suspension was passed through a 40 um filter and the cells counted. Thirty thousand cells were plated per well of a 6-well plate in MEGM, and assayed for cytokeratin expression after 7–11 days, using CK8/18 antibody (Vector VP-C407) and CK14 antibody (Thermo 9020-P). Some plates were visualized using IHC, while others were visualized using IF with AF488/AF555 conjugated secondary antibodies.
Colonies grown on 6-well plates were fixed in 100% methanol for 5 minutes, washed with PBS, permeabilized with 0.1% triton X-100 followed by serial blocking in 3% hydrogen peroxide and 1% BSA + 2% horse serum. The plates were incubated overnight at 4°C with 1:750 CK8/18 antibody. The plates were incubated with 1:200 αMouse-IgG-HRP (Vector BA-2000) for 30 minutes, and stained with DAB according to the manufacturer’s protocol (Vector ABC elite PK-6100; Vector ImmPACT DAB SK-4105). Excess avidin/biotin was blocked with the Vector Avidin/Biotin blocking kit SP-2001. Plates were re-blocked for 1 hour in PBS + 1% BSA and 2% goat serum, then incubated for 1 hour at room temperature with 1:750 CK14 antibody in PBS + 1% BSA then incubated at room temperature with αRabbit-IgG-HRP (Vector BA-1000) for one hour. The plates were then stained with VIP according to the manufacturer’s protocol (Vector ABC elite PK-6100; Vector ImmPACT VIP SK-4605), washed with water and stored dry. Western blots were performed with standard procedures. RUNX1 was blotted with 1:1000 Ab23980 (AbCAM).
|
10.1371/journal.pntd.0005540 | The phosphatidylinositol-3-phosphate 5-kinase inhibitor apilimod blocks filoviral entry and infection | Phosphatidylinositol-3-phosphate 5-kinase (PIKfyve) is a lipid kinase involved in endosome maturation that emerged from a haploid genetic screen as being required for Ebola virus (EBOV) infection. Here we analyzed the effects of apilimod, a PIKfyve inhibitor that was reported to be well tolerated in humans in phase 2 clinical trials, for its effects on entry and infection of EBOV and Marburg virus (MARV). We first found that apilimod blocks infections by EBOV and MARV in Huh 7, Vero E6 and primary human macrophage cells, with notable potency in the macrophages (IC50, 10 nM). We next observed that similar doses of apilimod block EBOV-glycoprotein-virus like particle (VLP) entry and transcription-replication competent VLP infection, suggesting that the primary mode of action of apilimod is as an entry inhibitor, preventing release of the viral genome into the cytoplasm to initiate replication. After providing evidence that the anti-EBOV action of apilimod is via PIKfyve, we showed that it blocks trafficking of EBOV VLPs to endolysosomes containing Niemann-Pick C1 (NPC1), the intracellular receptor for EBOV. Concurrently apilimod caused VLPs to accumulate in early endosome antigen 1-positive endosomes. We did not detect any effects of apilimod on bulk endosome acidification, on the activity of cathepsins B and L, or on cholesterol export from endolysosomes. Hence by antagonizing PIKfyve, apilimod appears to block EBOV trafficking to its site of fusion and entry into the cytoplasm. Given the drug’s observed anti-filoviral activity, relatively unexplored mechanism of entry inhibition, and reported tolerability in humans, we propose that apilimod be further explored as part of a therapeutic regimen to treat filoviral infections.
| The recent outbreak of Ebola virus (EBOV) disease in Western Africa highlights the urgent need to develop therapeutics to help quell this devastating hemorrhagic fever virus, especially in resource-limited areas around the globe. Here we show that apilimod, an investigational drug that was well-tolerated in phase 2 clinical trials for rheumatoid arthritis, Crohn’s disease, and psoriasis, is a strong inhibitor of both EBOV and Marburgvirus infections in multiple cell types. Further work shows that apilimod blocks the entry of EBOV particles into the host cell cytoplasm and that it does so by blocking the particles from reaching their normal portal of entry, in Niemann-Pick C1-positive endolysosomes. Our findings are consistent with the identity of phosphatidylinositol-3-phosphate 5-kinase as the molecular target of apilimod, as the kinase and its product phosphatidylinositol 3,5-bisphosphate are required for the proper maturation of late endocytic organelles. Hence we propose that apilimod be further explored for repositioning as part of a therapeutic regimen to help ameliorate the sequelae of filoviral infections.
| The epidemic of Ebola virus disease (EVD) that raged through Western Africa between 2013 and 2016 was the most severe filovirus disease epidemic in recorded history [1,2]. While several promising therapeutic antibodies [3–11] and novel small molecules [12–19] remain in development, no therapeutic is yet approved to treat patients with EVD. In the continuing pursuit of an anti- Ebola virus (EBOV) therapeutic, one strategy is to identify approved drugs that show anti-EBOV activity [20–28], with the goal of repurposing them for an anti-EBOV therapeutic, either alone or as part of a multi-component regimen [29–34].
Most of the approved drugs that have been identified as blocking EBOV infection inhibit the entry phase of the viral lifecycle [19–25,27,28]. Cell entry by EBOV is a complex process [35,36] entailing virus binding to cell surface attachment factors, internalization by macropinocytosis, processing by endosomal proteases, and transport to endolysosomes containing Niemann-Pick C1 (NPC1) [14,37], the intracellular receptor for EBOV [38]. Finally, EBOV fuses with the limiting membrane of NPC1+ endolysosomes [39–41], liberating its genome and associated proteins into the cytoplasm to begin replication.
The essential role of NPC1 in EBOV entry and infection was powerfully illuminated in a haploid genetic screen [37]. The same screen revealed other gene products critical for EBOV entry [42,43] including many involved in endosome and lysosome biogenesis and maturation. One of the latter proteins was phosphatidylinositol-3-phosphate 5-kinase (PIKfyve) [37], a lipid kinase that phosphorylates phosphatidylinositol-3-phosphate (PI3P) to generate phosphatidylinositol-3,5-bisphosphate (PI(3,5)P2). PIKfyve and PI(3,5)P2 are known to be critical for endosome maturation [44–53].
Apilimod is a small molecule that binds to and inhibits the phosphotransferase activity of PIKfyve [54]. The drug was developed as a suppressor of interleukin 12 and 23 production [55], and was tested in phase 2 clinical trials for treatment of Crohn’s disease [56,57], psoriasis [58], and rheumatoid arthritis [59]. Although no clinical benefit has yet been reported, apilimod is deemed to be well tolerated in humans. We chose to test whether apilimod could inhibit infections by EBOV and Marburg virus (MARV) for three reasons. The first was that apilimod binds [54] to the EBOV entry factor PIKfyve [37]. The second was because apilimod emerged from a blinded screen of 35 drugs (S1 Fig; S1 Table), which were selected as potential inhibitors based upon hypotheses of drugable targets and from theoretical considerations of pathways possibly involved in the EBOV life cycle. The third reason was because apilimod is well tolerated in humans. We find that apilimod inhibits infection by both EBOV and MARV, being notably effective in primary human macrophages, which are initial targets of filoviral infection [60,61]. Mechanistic studies revealed that apilimod blocks EBOV entry into the cell cytoplasm by working through PIKfyve and that its effect is to block viral particle trafficking to NPC1+ endolysosomes, the site of EBOV fusion [39–41]. Hence we propose that apilimod be further explored as part of a cocktail of small molecules to combat EVD.
Vero E6 (African green monkey kidney; ATCC 1586) cells were obtained from the American Type Culture Collection (Manassas, VA). Huh 7 (human hepatocellular carcinoma) cells were obtained from Dr. Hideki Ebihara (National Institute of Allergy and Infectious Diseases (NIAID), Rocky Mountain Laboratories, Hamilton, MT). Peripheral blood mononuclear cells (PBMCs) were prepared from human whole blood (Biological Specialty Corporation; Colmar, PA; Cat # 3100-03-04) and human monocyte-derived macrophages (hMDM) were generated from peripheral blood mononuclear cells at the Integrated Research Facility (IRF) immunology core laboratories as described previously [62,63]. hMDM were characterized by flow cytometric analysis for expression of major macrophage markers, including human leukocyte antigen-D related, CD11b, CD14, CD163, and CD86, to confirm that the hMDM population was mature and highly purified [63]. HEK 293T/17 (Human embryonic kidney; ATCC CRL-11268 via University of Virginia Tissue Culture Facility) and BSC-1 (Grivet monkey kidney; gift from Dr. Xiaowei Zhuang, Harvard University, Cambridge, MA) cells were maintained in growth medium: high glucose Dulbecco's Modified Eagle Medium (DMEM) supplemented with 1% L-glutamine, 1% sodium pyruvate, and 1% antibiotic/antimycotic, all from Gibco Life Technologies (Carlsbad, CA), and either 10% supplemented calf serum (SCS; Hyclone, GE Healthcare Bio-Sciences, Pittsburgh, PA) for HEK 293T/17 cells or 10% fetal bovine serum (FBS, Seradigm, VWR International, Radnor, PA) or 10% cosmic calf serum (CCS, Hyclone) for BSC-1 cells.
Toremifene citrate (CAS 89778-27-8) was purchased from Sigma-Aldrich (St. Louis, MO; Cat# T7204-25MG) and apilimod (CAS 541550-19-0) was purchased from Axon MedChem (Groningen, NL; Cat# 1369). Nocodazole (CAS 31430-18-9) was purchased from Sigma-Aldrich (St. Louis, MO; Cat # M1404-2MG).
All procedures using infectious EBOV/Mak or MARV were performed under biosafety level 4 (BSL-4) conditions at the IRF. The C05 isolate of the Makona variant of EBOV (EBOV/Mak; full designation: Ebola virus/H.sapiens-tc/GIN/2014/Makona-C05; GenBank: KX000398) and Marburg Angola virus (MARV; full name: Marburg virus/H.sapiens-tc/AGO/2005/Ang-1379v; GenBank: N/A) were propagated in BEI NR-596 Vero E6 cells and used after one or two passages.
The drug screen method was performed as described previously [63]. Briefly, Vero E6 and Huh 7 cells were seeded in 96-well plates at a density of 3 x 104 cells/well, and hMDMs were plated at a density of 1 x 105 cells/well 24 h prior to the addition of drugs. For each cell type, cells were plated in 1 black opaque 96-well plate, for the evaluation of drug cytotoxicity, and 2 clear bottom, 96-well Operetta plates, for the evaluation of drug efficacy. Drugs dissolved in dimethyl sulfoxide (DMSO; Sigma-Aldrich, St. Louis, MO) were diluted in DMEM with 10% FBS with the final DMSO concentration not exceeding 0.05%. The drug solutions were diluted two-fold in an 8-point dilution series and transferred to cell plates 1 h prior to virus infection. Efficacy plates for each cell type were infected with EBOV or MARV at a multiplicity of infection (MOI) of 0.5. After 48 h, cells were fixed with 10% neutral-buffered formalin.
Chemiluminescent enzyme-linked immunosorbent assay was used to determine virus activity. EBOV was detected with a mouse antibody against the EBOV VP40 matrix protein (B-MD04-BD07-AE11, made by US Army Medical Research Institute of Infectious Diseases, Frederick MD under Centers for Disease Control and Prevention contract) [3] and MARV was detected with a mouse antibody against the MARV VP40 protein (Cat# IBT 0203–012, IBT Bioservices, Rockville, MD) for 1–2 h at 37°C. Cells were stained with a secondary antibody, anti-mouse IgG, peroxidase labeled antibody (Cat# 074–1802, KPL Inc., Gaithersburg, MD). Luminescence was detected using Pico chemiluminescent Substrate (Thermo Fisher Scientific Inc., Rockford, IL) and an Infinite M1000 Pro plate reader (Tecan, Morrisville NC).
For quantitation of drug toxicity, 1 black opaque cell plate for each cell type was mock infected (no virus) and treated with drug dilutions under the same conditions as the infected cells. After 48 h, cell viability was measured using the CellTiter Glo Luminescent Cell Viability Assay kit according to the manufacturer’s protocol (Promega, Madison, WI). Luminescence was read on an Infinite M1000 Pro plate reader.
Following background subtraction, inhibition was measured as percent relative to untreated infected cells. Non-linear regression analysis was performed, and IC50s were calculated from fitted curves (log [agonist] vs response [variable slope] with constraint to remain above 0; GraphPad Software, La Jolla, CA). Error bars of dose-response curves represent the standard deviation of three replicates.
Entry reporter viral-like particles (VLPs) bearing GP from the Yambuku-Mayinga isolate of EBOV were prepared as described previously [24,25,41]. In brief, HEK 293T/17 cells (~80% confluent) were transfected with cDNAs encoding EBOV GP, VP40, mCherry-VP40, and β-lactamase-VP40 (βlam-VP40). The cell medium was collected 24 and 48 h post-transfection and cleared of debris. VLPs in the cleared medium were then pelleted through a 20% sucrose cushion by centrifugation, resuspended in HM buffer (20 mM HEPES, 20 mM MES, 130 mM NaCl, pH 7.4), and repelleted. The final VLP pellet was resuspended (1:100 starting volume of medium) in 10% sucrose-HM. The total protein concentration of the VLPs was determined by bicinchoninic acid (BCA) assay. All entry-reporter VLP preparations were assessed by western blot analyses (for the presence of GP as well as EBOV VP40) and titered on HEK 293T/17 cells to confirm entry competency.
The VLP entry assay scores the ability of βlam-VP40 (from incoming entry reporter VLPs) to cleave a βlam substrate preloaded into the target cell cytoplasm; this only occurs if the VLP fuses with an endosome. The assay was performed as described previously [24,25,41]. In brief, 30,000 HEK 293T/17 cells or BSC-1 cells were seeded per well in a clear 96-well plate. 18–24 h post seeding, the cells (~80%–90% confluent) were treated with the indicated concentration of apilimod (Axon MedChem; DMSO for mock) diluted in Opti-MEM I (OMEM, Gibco Life Technologies, Thermo Fisher Scientific) for 1 h at 37°C in a 5% CO2 incubator. VLPs diluted in OMEM (with DMSO or the same concentration of apilimod) were bound to the cells by spinfection (250× g) for 1 h at 4°C. After 3 h in a 37°C, 5% CO2 incubator, the βlam substrate CCF2-AM (Life Technologies, via ThermoFisher Scientific, Waltham, MA, USA) was loaded into the cells using 20 or 5 mM Probenecid (MP Biomedicals via ThermoFisher Scientific, Waltham, MA, USA), for BSC-1 or HEK 293T/17 cells, respectively. The cells were incubated overnight at RT and then fixed and analyzed by flow cytometry.
To measure corresponding cell viability, 3 x 104 HEK 293T/17 cells, seeded and grown as above but in 96-well opaque white plates were treated as above for VLP entry, but without addition of VLPs or CCF2-AM. Following overnight incubation at RT (as above), the medium was removed and replaced with 50 μL of fresh medium per well. Fifty microliters (per well) of CellTiter-Glo 2.0 (Promega, Madison WI, USA) was then added. After shaking for 2 min at RT at 575 rpm on a Jitterbug orbital shaker (Boekel Scientific, via ThermoFisher Scientific, Waltham, MA, USA), the plate was incubated at RT for 10 min, after which the luminescent signal was detected using a BioTek Synergy HT plate reader (BioTek, Winooski, VT, USA).
Transcription/replication-competent virus-like particles (trVLPs) were prepared as described [25,64,65]. In brief, HEK 293T/17 cells were seeded in six well plates and transfected 24 h later (when ~50% confluent) using TransIT-LT1 (Mirus, Madison, WI, USA) with pCAGGS-NP, pCAGGS-VP35, pCAGGS-VP30, pCAGGS-L, a tetracistronic minigenome plasmid, and pCAGGS-T7 polymerase. The minigenome plasmid encodes Renilla luciferase, as well as the matrix protein VP40, the nucleocapsid associated protein VP24, and the GP from EBOV. 24 h post transfection, the medium in each well was replaced with 4 mL fresh growth medium containing 5% FBS. 96 h after transfection, the medium (containing trVLPs harboring the Renilla luciferase-containing minigenome) was harvested, pooled, and cleared of cellular debris by centrifugation for 5 min at 800× g and used for trVLP assays as described below.
The trVLP assay measures the ability of trVLPs containing a Renilla luciferase-encoding tetracistronic EBOV minigenome to infect target cells pretransfected with plasmids encoding proteins to enhance trVLP entry (the adhesion factor Tim-1) and (other plasmids) to support replication of the minigenome. If trVLPs enter target cells, the minigenome is replicated and transcribed, leading to Renilla luciferase reporter activity [64,65]. In brief: Cells were pretreated with apilimod (Axon MedChem; DMSO for mock) as described above. The pretreatment solution was then removed and replaced with 100 μL trVLPs diluted to 200 μL in growth medium containing 10% SCS and the indicated concentration(s) of apilimod (DMSO for mock). The cells were then incubated for 48 h at 37°C in a 5% CO2 incubator, after which the medium was replaced with 40 μL of fresh growth medium containing 10% SCS. 40 μL of RenillaGlo substrate (Promega, Madison, WI, USA) was then added to each well and the plate immediately analyzed on a GloMax plate reader (Promega, Madison, WI, USA).
To assess cell viability in corresponding samples without trVLPs, the pretreatment solution was removed and replaced with 200 μL fresh growth medium containing 10% SCS and the indicated concentrations of apilimod (DMSO for mock). The cells were then incubated for 48 h at 37°C in a 5% CO2 incubator, after which the medium was replaced with 40 μL of fresh growth medium containing 10% SCS. 40μL of CellTiter-Glo 2.0 (Promega) was then added to each well and the plate placed on a Jitterbug orbital shaker (575 rpm) for 2 min at RT. The plate was then incubated at RT for 10 min, after which the luminescent signal was detected using a Synergy HT (BioTek, Winooski, VT, USA) plate reader.
BSC-1 cells were seeded in 35mm glass bottom dishes (MatTek, Ashland, MA) that were coated with 20 μg/mL fibronectin (Sigma-Aldrich, St. Louis, MO, USA). The next day, when the cells were 90–100% confluent, the cells were treated with the indicated drug at the indicated concentration, diluted in growth medium containing 10% cosmic calf serum, for 3 h at 37°C in a 5% CO2 incubator. Acridine Orange (Life Technologies, Thermofisher Scientific, Waltham, MA, USA) was added directly to each dish to reach a final concentration of 6.6 μg/mL. The cells were incubated at 37°C in a 5% CO2 incubator for 20 min and then were washed 3 times with phosphate buffered saline (PBS), 5 min per wash. Cell imaging medium [Live cell imaging solution (Molecular Probes, Cat# A14291DJ, Thermo Fisher Scientific, Waltham, MA) containing 10% FBS and 4.5 g/L glucose] was added to the dishes and images were taken using a Nikon C1 laser scanning confocal unit attached to a Nikon Eclipse TE2000-E microscope with a 100X, 1.45-numerical-aperature (NA) Plan Apochromat objective (Nikon, Melville, NY).
BSC-1 cells were seeded in 35mm glass bottom dishes (MatTek, Ashland, MD, USA) that had been coated with 20 μg/mL fibronectin (Sigma-Aldrich). The next day, when the cells were 90–100% confluent, the cells were treated with the indicated drug at the indicated concentration plus 0.05 μM TopFluor Cholesterol (Avanti Polar Lipids, Alabaster, AL), diluted in serum-free growth medium, for 18 h at 37°C in a 5% CO2 incubator. Following incubation, the cells were gently rinsed once with PBS and cell imaging medium (Live cell imaging solution (Molecular Probes) containing 10% FBS and 4.5 g/L glucose) was added to the dishes. The cells were incubated at 37°C in a 5% CO2 incubator for 30 min. Images were then taken using a 60X /1.45 numerical aperture (NA) Nikon Plan Apo total internal reflection fluorescence oil immersion objective attached to a Nikon Eclipse TE2000-E microscope equipped with a Yokogawa CSU 10 spinning-disk confocal unit, a 512-by-512 Hamamatsu 9100c-13 EM-BT camera, a motorized stage maintained at 37°C, and a Nikon Perfect Focus system.
VLP trafficking experiments were performed in BSC-1 cells essentially as described previously [41] with the following minor modifications. Cells were pretreated with apilimod or nocodazole (indicated concentrations) diluted in OMEM for 1h at 37°C prior to VLP addition. VLPs at 0.5 μg/well were bound to the cells by spinfection (250 x g) for 1 h at 4°C. After incubation at 37°C (CO2 incubator) for the indicated times, the cells were fixed and washed. Next, primary antibodies (1:1000 rabbit α-NPC1, (Abcam) or 1:1000 mouse α-early endosome antigen 1 (EEA1), BD Biosciences, San Jose, CA) were added for 45 min at room temperature (RT) and, following washing, secondary antibodies (1:1500 α-mouse or α-rabbit AlexaFluor 488, Life Technologies, Thermo Fisher Scientific) were added for 30 min at RT. The cells were washed and the coverslips were mounted overnight on glass slides using ProLong Gold Antifade reagent (Life Technologies, Thermo Fisher Scientific). The coverslips were then sealed and images were taken using a Nikon C1 laser scanning confocal unit attached to a Nikon Eclipse TE2000-E microscope with a 100X, 1.45-NA Plan Apochromat objective. Colocalization of VLPs (red, mCherry-VP40) and endosomal markers (green, NPC1 or EEA1) was assessed as Manders coefficients. Statistics were analyzed using GraphPad Prism 7. Normality of the data was assessed using the D’Agostino & Pearson normality test. Significance of normally distributed data was determined by T-test, and significance of non-normally distributed data was determined by Mann-Whitney test.
Cathepsin B+L activity was assayed as described previously [20,21,24]. (2S,3S)-trans-epoxysuccinyl-L-leucylamido-3-methylbutane ethyl ester (EST, Cat # 330005, Calbiochem, EMD Millipore, Billerica, MA), an inhibitor of cathepsin B,H, and L, was used as a positive control for inhibition at the indicated concentration. Data are displayed as fluorescence units (Ex 360/Em 460).
Thirty-five drugs obtained from the National Center for Advancing Translational Sciences (NCATS) were dissolved in DMSO at 500 μM. Drugs were diluted in DMEM (Life Technologies, Thermo Fisher Scientific) supplemented with 2 mM L-Glutamine (Q; Life Technologies, Thermo Fisher Scientific) and 100 U/ml penicillin and 100 μg/ml streptomycin (PS; Life Technologies, Thermo Fisher Scientific). Drugs were added to confluent Vero E6 cells. Drugs and cells were then incubated at 37°C and 5% CO2 in a humidified incubator in 96-well plates for final concentrations of 10, 1, or 0.1 μM in a final volume of 100 μl DMEM/PS/Q with 2% FBS (Life Technologies, Thermo Fisher Scientific). Cells were returned to the incubator for 2 h.
For efficacy studies, 50 μl DMEM/PS/Q containing 1x103 TCID50 of recombinant EBOV expressing firefly luciferase from an additional transcriptional unit (rgEBOV-luc2, Genbank Accession number KF990214.1) [66] was added to the cells. At 48 h post-inoculation the supernatant was removed and 100 μl GloLysis buffer (Promega) was added to the cells and incubated for 10 min at RT. Afterwards, 40 μl lysate was added to 40 μl BrightGlo reagent (Promega) in white opaque 96 well plates, and reporter activity was measured using a GloMax luminometer.
For cytotoxicity studies, 50 μl of DMEM/PS/Q without virus was added to the cells following the 2 h pre-incubation with drugs, and cells were returned to the incubator. At 48 h, 100 μl of supernatant was removed, and 50 μl of CellTiterGlo reagent (Promega) was added to the cells. Cells were incubated for 2 min on an orbital shaker at 60 RPM, and then for an additional 10 min without shaking at RT. Supernatants were transferred to white opaque 96-well plates, and reporter activity was measured using a GloMax luminometer (Promega). Ribavirin at final concentrations of 1 mg/ml, 100 μg/ml, and 10 μg/ml, as well as DMSO at concentrations corresponding to the DMSO concentrations found in the drug dilutions served as controls. All experiments involving infectious rgEBOV-luc2 were performed in the maximum containment laboratory of the Rocky Mountain Laboratories, National Institutes of Health, Hamilton, MT, following approved protocols.
HEK 293T/17 cells were seeded at a density of 3 x 106 cells per 10 cm plate. The next day, when the cells were approximately 60% confluent, the media above the cells was replaced with 6mL OMEM and the cells were transfected with 2.4 μg pTG-luc, 1.2 μg pCMV-MLVgag-pol, 1.2 μg pGPΔmucin (encoding Ebola GP deleted for its mucin domain), and 1.2 μg of MLV-gag-βlam diluted to 300 μL in OMEM (per plate), using 18 μL Lipofectamine 2000 (Invitrogen, ThermoFisher Scientific, Walthan, MA) diluted to 300 μL in OMEM (per plate). 4h post transfection, 6 mL of antibiotic-free growth medium containing 10% SCS was added to each plate, and the cells were incubated for 48 h at 37°C in a CO2 incubator. Cell medium containing pseudovirus was then collected, pooled, and cleared of cellular debris by centrifugation at 250 x g for 7 min. The clarified supernatant containing pseudovirus was then passed through an 0.45 μm filter and the pseudoviruses were concentrated 100-fold by high-speed centrifugation through a 25% sucrose cushion in HM buffer (20mM HEPES, 20mM MES, 130mM NaCl, pH7.4) for 75 min at 103,745 x g. The final pseudovirus pellet was resuspended in growth medium (100-fold concentrated from harvest supernatant).
HEK 293T/17 cells were seeded at a density of 5 x 105 cells per well in 6-well plates. When the cells were ~50% confluent (~18–24 h post seeding), they were transfected with plasmids encoding GFP-PIKfyve or pEGFP-Cl using TransIT LT1 transfection reagent (Mirus, Madison, WI) following the manufacturer’s instructions. 18 h post transfection, the cells were re-seeded in 96 well opaque white plates (BD Falcon, ThermoFisher Scientific, Waltham, MA) at a density of 3 x 104 cells per well. Transfection was confirmed by fluorescence microscopy. 18 h post re-seeding, the cells were pretreated for 1 h at 37°C with apilimod. MLV-luciferase particles pseudotyped with EBOV GPΔmucin were added to the cells in the presence of apilimod, and infection was allowed to proceed for 48 h at 37°C. The cells were then washed once with PBS and overlaid with 50μL PBS. Luciferase activity was then immediately assayed by adding 50μL of Britelite plus (Perkin Elmer, Waltham, MA) and reading on a Glomax plate reader (Promega, Madison, WI, USA) following the manufacturer’s instructions.
We first tested whether apilimod blocks EBOV infection in cell cultures. Apilimod blocked EBOV infection of Huh 7 (liver) cells, Vero E6 (kidney) cells, and primary human monocyte-derived macrophages (hMDMs) (Fig 1) with 6- to 247-fold higher activity than the positive control, toremifene citrate [20,22]. Apilimod also blocked MARV infection of the same cell types (Fig 2) with 38- to 1160-fold higher activity than the positive control. Apilimod was notably potent (IC50, 10 nM) against both filoviruses in hMDMs (Figs 1 and 2, Table 1). While similar potency (IC50, 15–25 nM) was seen in Vero E6 cells, apilimod was ~10-fold less potent (IC50, 140 nM) in Huh 7 cells (Fig 1, Fig 2 and Table 1).
To begin to probe the mechanism by which apilimod blocks EBOV infection, we directly compared dose-response profiles for blocking EBOV entry and replication using entry reporter VLPs [24] and trVLPs [64], respectively. Both sets of VLPs bore the GP from the Mayinga isolate of EBOV. Apilimod blocked EBOV particle entry (Fig 3A) and replication (Fig 3B) with similar dose-response profiles (Fig 3C). This finding suggested that apilimod blocks the entry phase of the filoviral lifecycle.
Since apilimod targets PIKfyve [54], since PIKfyve is required for EBOV entry and infection [37], and since apilimod blocks EBOV entry and infection (Figs 1–3), we reasoned that apilimod blocks EBOV entry and infection by targeting PIKfyve. To test this hypothesis we over-expressed PIKfyve (GFP-PIKfyve) and compared the dose response needed for apilimod to block EBOV-GP mediated pseudovirus infection. As predicted, and as seen in Fig 4, higher doses of apilimod were needed to achieve similar levels of inhibition of EBOV GP-mediated infection in GFP-PIKfyve vs. GFP expressing cells. This supports our proposal that apilimod blocks EBOV entry and infection through a PIKfyve-dependent pathway.
Recent work has shown that EBOV traffics deep in the endocytic pathway, to NPC1+ endolysosomes, for fusion and entry [39–41]. Therefore we asked whether apilimod prevents EBOV VLPs from reaching NPC1+ endolysosomes. We used BSC-1 cells for these experiments as they are more suitable (flatter and more adherent) for immunofluorescence analysis than the HEK 293T/17 cells used in previous experiments (Figs 3 and 4). We first demonstrated that apilimod blocks EBOV VLP entry into BSC-1 cells with the same approximate dose-dependency as its effects in HEK 293T/17 cells (Fig 3D). Given that, we next asked if apilimod blocks trafficking of EBOV GP VLPs to NPC1+ endolysosomes in BSC-1 cells. As seen in Fig 5, this was, indeed, the case. Apilimod blocked EBOV VLP trafficking to NPC1+ endolysosomes to a similar extent as nocodazole, a microtubule destabilizer that is known to block traffic between early and late endosomes [67].
The findings presented in Fig 5A–5D were obtained after allowing VLPs pre-bound to the cell surface to enter cells for 90 min at 37°C. This time point was chosen based on our extensive prior analysis of the time courses of EBOV VLP co-localization with NPC1+ endolysosomes and entry into BSC-1 cells [41]. To assure that apilimod did not accelerate VLP trafficking to NPC1+ endolysosomes, we analyzed co-localization of VLPs at various times up to 90 min in cells treated or not treated with apilimod. As seen in Fig 5E, at no point during this time course were VLPs seen to associate with NPC1+ endolysosomes in apilimod-treated cells, supporting our conclusion that apilimod blocks trafficking of EBOV particles to NPC1+ endolysosomes.
Concomitant with decreased trafficking of EBOV GP VLPs to NPC1+ endolysosomes, apilimod caused EBOV VLPs to accumulate in EEA1+ endosomes (Fig 6). In apilimod-treated cells, EEA1+ endosomes appeared larger than those in mock-treated cells, consistent with previous reports showing enlarged endosomes in cells genetically deficient for PIKfyve or treated with PIKfyve inhibitors [44,45,50,53].
The findings presented in Figs 5 and 6 indicate that the primary mechanism by which apilimod blocks EBOV entry and infection (Figs 1–4) is by blocking virus transport from early (EEA1+) endosomes to their site of fusion in NPC1+ endolysosomes. To further test this model, we asked whether apilimod affects other attributes of the endosomal pathway needed for EBOV entry, either endosome acidification or the activity of cathepsin B and L [42,43]. At a concentration that strongly blocked EBOV entry and infection, apilimod had no detectable effect on endosome acidification (Fig 7). Bafilomycin, an inhibitor of EBOV infection that blocks endosome acidification, was used as a positive control. Apilimod also had no apparent direct effect on the activity of cathepsin B and L (Fig 8), in contrast to EST, a known inhibitor of cathepsin B, H, and L. Several cationic amphiphilic drugs such as U18666A that block EBOV entry and infection [20–22,24,25,68] induce cholesterol accumulation in endolysosomes [20,24]. In contrast, apilimod did not cause a detectable increase in cholesterol levels in endolysosomes (Fig 9). Hence apilimod appears to block filoviral entry and infection by inhibiting virus particle trafficking to NPC1+ endolysosomes, the portal for entry of the EBOV genome into the host cell cytoplasm [39–41].
As tragically demonstrated by the recent epidemic of EVD in Western Africa (2013–2016), a pressing need remains to develop therapeutics to treat patients infected with filoviruses [1]. While novel monoclonal antibody and small molecule therapeutics are in the pipeline [3–19], a parallel approach is to consider repositioning an approved drug or a drug that has proven safe in phase 2 clinical trials to treat EVD. In this study we tested the potential utility of apilimod as an anti-filoviral agent. We evaluated this drug for three reasons: (a) apilimod directly targets PIKfyve [54], a known EBOV entry factor [37]; (b) apilimod emerged from a blinded screen of 35 drugs targeting cell signaling pathways (S1 Fig and S1 Table); and (c) apilimod was reported to be well tolerated in humans in several phase 2 clinical trials [56–59]. After demonstrating that apilimod has activity against both EBOV and MARV in several cell types, notably in human macrophages (Figs 1 and 2, Table 1), we demonstrated that, by working through PIKfyve, its primary mode of action is to block trafficking of EBOV particles to NPC1+ endolysosomes (Figs 3–9), the site of EBOV fusion and entry [39–41].
EBOV journeys deep into the cellular endosomal system, entering the cytoplasm through endolysosomes that are positive for NPC1 and two-pore channel 2 (TPC2) [39]. In addition to NPC1, its intracellular receptor [14,37,38], EBOV requires multiple factors involved in endosome and lysosome biogenesis and maturation for entry [37]. One of the latter factors is PIKfyve [37], which phosphorylates PI3P to generate PI(3,5)P2. Here we have shown that apilimod, which binds to PIKfyve [54], blocks EBOV entry and infection in a PIKfyve-dependent manner.
The inhibitory effect of apilimod on EBOV entry is likely due to a defect in the maturation of endolysosomes, as extensive evidence indicates the importance of PIKfyve and PI(3,5)P2 in this process [44–53]. Although the exact mechanism by which PIKfyve and PI(3,5)P2 orchestrate endosome maturation is not known, several mechanisms have been postulated. Considered in these mechanisms are the observations that two Ca++ channels found in (endo)lysosomes—transient receptor potential cation channel, mucolipin 1 (TRPML1) [50,69] and TPC2 [70–72]—are downstream effectors of PIKfyve and PI(3,5)P2. Through its action on TRPML1, PIKfyve has been shown to regulate the fission and consequent remodeling and maturation that reduces the size of macropinosomes containing endocytosed material from the cell surface and exterior [50]. In addition, the TPC2 channel has been reported to be activated by PI(3,5)P2 [71,72]. Intriguingly, both macropinocytosis [35,73,74] and TPC2 [75] are involved in EBOV entry and infection. Although we do not yet know all of the endolysosomal factors needed to trigger EBOV GP for fusion [36,39,43,76], it appears clear that proper endosomal maturation is required. These findings are consistent with the mounting evidence for a role of PIKfyve in EBOV entry and our observation that the PIKfyve inhibitor, apilimod, blocks transport of EBOV particles to NPC1+ endolysosomes. It is therefore likely that interconnected effects of apilimod on PIKfyve [37,54], TPC2 [75], and endolysosome maturation culminate in its blockade of EBOV entry and infection.
Several approved drugs that function as EBOV entry inhibitors (e.g., clomiphene, toremifene, and sertraline) block EBOV entry into the cytoplasm after EBOV particles have been delivered to NPC1+ endolysosomes [20,21,24,25]. Hence they likely interfere with some aspect of the virus-endolysosome membrane fusion process, per se. Other approved drugs, including chloroquine, niclosamide, atovaquone, amodiaquine and quinacrine [21–23], block endosomal acidification. Hence these drugs likely interfere with the processing of EBOV GP by acid-optimal endosomal cathepsins [42,43] and/or low pH-induced conformational changes required for fusion activity of cleaved GP [77,78].
In contrast to these mechanisms, our findings indicate that apilimod blocks EBOV entry by blocking particle delivery into NPC1+ endolysosomes. The only other approved drugs that we know of with anti-EBOV activity that are expected to have this mode of action are microtubule-disrupting agents, including colchicine, nocodazole, vinblastine, and vinorelbine [21–23]. Hence the mode of action of apilimod as an anti-filoviral agent is novel. Rather than blocking EBOV trafficking to NPC1+ endosomes by interfering with microtubules, apilimod blocks EBOV trafficking by inhibiting PIKfyve.
Our findings indicate that apilimod has similar anti-viral activity against EBOV and MARV, consistent with the need for NPC1 in endolysosomes for the entry of these and other filoviruses [37,79,80]. We therefore consider it likely that apilimod, a host-directed small molecule, will have broad or even pan-filoviral activity. Furthermore, since many other viruses, so-called late penetrating viruses [36,81], traffic beyond early endosomes for entry, it is possible that apilimod will block entry and infection by members of other virus families.
Apilimod is an investigational drug. Although it has been tested in phase 2 clinical trials for the treatment of Crohn’s disease, psoriasis, and rheumatoid arthritis, the drug has not yet been approved for any indication. Nonetheless apilimod was well-tolerated in humans in the reported phase 2 trials [56–59]. We found that intraperitoneal delivery of 10 mg/kg of apilimod to mice resulted in a Cmax of 2.53 μM. This is well above the IC50 for apilimod inhibition of EBOV infection in the three cell lines tested, ~250 times greater than the IC50 in hMDMs (10 nM), initial major targets of filoviral infections [60,61]. We therefore consider it plausible that apilimod be used in the treatment of EVD. And, while apilimod may not function as a single agent, it may perform well as a component in an anti-filoviral small molecule cocktail. In summary, we introduce apilimod, a small molecule PIKfyve inhibitor that has proven safe in phase 2 clinical trials, as a potential anti-filoviral agent.
|
10.1371/journal.pgen.1005462 | Histone H2AFX Links Meiotic Chromosome Asynapsis to Prophase I Oocyte Loss in Mammals | Chromosome abnormalities are common in the human population, causing germ cell loss at meiotic prophase I and infertility. The mechanisms driving this loss are unknown, but persistent meiotic DNA damage and asynapsis may be triggers. Here we investigate the contribution of these lesions to oocyte elimination in mice with chromosome abnormalities, e.g. Turner syndrome (XO) and translocations. We show that asynapsed chromosomes trigger oocyte elimination at diplonema, which is linked to the presence of phosphorylated H2AFX (γH2AFX). We find that DNA double-strand break (DSB) foci disappear on asynapsed chromosomes during pachynema, excluding persistent DNA damage as a likely cause, and demonstrating the existence in mammalian oocytes of a repair pathway for asynapsis-associated DNA DSBs. Importantly, deletion or point mutation of H2afx restores oocyte numbers in XO females to wild type (XX) levels. Unexpectedly, we find that asynapsed supernumerary chromosomes do not elicit prophase I loss, despite being enriched for γH2AFX and other checkpoint proteins. These results suggest that oocyte loss cannot be explained simply by asynapsis checkpoint models, but is related to the gene content of asynapsed chromosomes. A similar mechanistic basis for oocyte loss may operate in humans with chromosome abnormalities.
| Chromosome abnormalities, such as aneuploidies and structural variants (i.e. translocations, inversions), are strikingly common in the human population, causing disorders such as Down syndrome and Turner syndrome. One important consequence of chromosome abnormalities in mammals is errors during meiosis, the specialized cell division that generates sperm and eggs for reproduction. As a result of these meiotic errors, patients with chromosome abnormalities oftentimes suffer from infertility due to loss of developing germ cells. The precise molecular mechanism for germ cell losses and infertility due to chromosome abnormalities is not well understood, but is hypothesized to result from a surveillance mechanism, which has evolved to prevent aneuploidies from developing from abnormal germ cells. In mammals, meiotic surveillance mechanisms have been hypothesized to monitor for unrepaired DNA double-strand breaks (DSB) and/or chromosome pairing/synapsis errors. Here we test these hypotheses using a variety of chromosomally variant mouse models. We find that germ cell loss in female mice with chromosome abnormalities is dependent on phosphorylation of the histone variant H2AFX, an epigenetic mark involved in the transcriptional silencing of asynapsed chromosomes during meiosis. These data inform a silencing-based mechanism of germ cell loss in patients with chromosome abnormalities and for the prophase I surveillance system which safeguards against aneuploidy.
| Prophase I of mammalian meiosis entails alignment, synapsis and reciprocal recombination between homologues, which together enable crossover formation. Without crossovers, homologues mis-segregate, giving rise to aneuploidy [1]. To protect against this, germ cells exhibiting defects in the key prophase I events are eliminated by quality control mechanisms [2]. Understanding the molecular basis of these surveillance mechanisms represents an important challenge.
The pathways that drive prophase I oocyte loss in chromosomally abnormal female mice are unclear. Two models have so far predominated: persistent meiotic DNA damage and chromosome asynapsis. Evidence that persistent DNA damage can cause oocyte loss is derived from studies of mice carrying targeted mutations in meiotic recombination genes, e.g. Dmc1 [3]. In these models, DNA DSB markers, e.g. RAD51 and DMC1, persist at chromosome axes [4] and oocyte loss is partially rescued by ablating Spo11 [3], the enzyme responsible for meiotic DNA DSB formation. Persistent DNA damage triggers oocyte elimination via the CHK2/p53/p63 pathway [5]. It is important to note that in these targeted mutants DNA repair pathways are genetically disabled, while in chromosomally abnormal mice they are intact. Whether meiotic DNA DSBs persist in chromosomally abnormal mice therefore remains an open question.
Mouse mutants in which chromosome asynapsis is present but programmed meiotic DNA DSB formation is either reduced or abolished e.g. Mei1-/- [6], Mei4-/- [7] and Spo11 -/- [3,8,9] also exhibit prophase I elimination. This suggests that asynapsis per se can also cause oocyte loss. Asynapsis has been observed in chromosomally abnormal female mice [10,11], and leads to accumulation of HORMAD1/2 at chromosome axes [12–14] and serine-139 phosphorylated histone H2AFX (γH2AFX) in the adjacent chromatin [15,16]. How these changes precipitate oocyte elimination is not known. One hypothesis invokes the existence of a synapsis checkpoint that eliminates oocytes in response to asynapsis per se [17]. An alternative model suggests a role for γH2AFX-associated transcriptional inactivation [2,18]. This process, known as meiotic silencing, could cause elimination through silencing of essential germ cell-expressed genes, or through alterations in transcription factor binding on asynapsed chromosomes, or expression of non-coding RNAs or transposable elements. The synapsis checkpoint and meiotic silencing models are difficult to dissect experimentally, however, because genes with putative synapsis checkpoint functions, e.g. Hormad1 and Hormad2, are also essential for silencing [13,19–21]. Discriminating between these models requires analysis of mice carrying extra/supernumerary chromosomes, in which asynapsis is present but silencing affects genes that are not essential for germ cell development.
Further complexity in understanding the prophase I surveillance mechanisms comes from studies in males. During normal male meiosis, the X and Y chromosomes are asynapsed and thus undergo meiotic silencing. This process, called Meiotic Sex Chromosome Inactivation (MSCI) [22], is highly conserved in eutherian mammals. The fact that the X and Y chromosomes are enriched in HORMAD1, HORMAD 2, γH2AFX and other checkpoint factors, yet do not trigger prophase I elimination, is puzzling.
Here we examine the pathways that drive oocyte loss in an extensive array of mice carrying chromosome abnormalities similar to those found in humans. We present a model for prophase I elimination that unifies existing data and explains why in males the asynapsed X and Y chromosomes do not trigger loss while asynapsed autosomes do.
Initially, we sought to establish the timing of oocyte loss in mice carrying chromosome abnormalities. We first studied females with sex chromosome abnormalities, using X chromosome monosomy (XO; Turner syndrome) as our model. XO mice have a shortened reproductive lifespan owing to perinatal oocyte losses that have been hypothesized to result from X chromosome asynapsis during prophase I [23,24]. To determine the developmental timing of XO oocyte losses, we examined the percentage of oocytes with X chromosome asynapsis at successive meiotic prophase I substages. We utilized an immunostaining approach to categorize XO oocytes as being in pachynema, when autosomal synapsis is complete, or early and late diplonema, when chromosomes progressively desynapse. Our substaging criteria were based on the immunostaining patterns of SYCP3, an axial element marker, and HORMAD1, a marker of asynapsed and desynapsed axes (S1A–S1C Fig). To identify the asynapsed X chromosome, we simultaneously immunostained for serine-139 phosphorylated H2AFX (γH2AFX), which marks the chromatin of asynapsed chromosomes, but not desynapsed chromosomes, from pachynema to diplonema [16,25].
At pachynema, we found that 55% of XO oocytes carried an asynapsed, γH2AFX-enriched X chromosome (Fig 1A and 1C). In the remaining oocytes, the X chromosome was self-synapsed and, as previously reported [16], was devoid of γH2AFX (Fig 1B and 1C). Importantly, oocytes with an asynapsed X chromosome decreased in abundance by more than five-fold during diplotene progression (Fig 1C). We confirmed this result using a second approach, in which we quantified the percentage of all XO oocytes with a γH2AFX-positive X chromosome at 17.5, 18.5 and 19.5 days post-coitum (dpc; S2A and S2B Fig), the developmental time period when oocytes progress semi-synchronously from pachynema to diplonema.
Next, we examined whether the drop in the percentage of XO oocytes with an asynapsed X chromosome was also evident earlier, specifically during pachynema. To test this, we compared the percentage of oocytes with γH2AFX domains at early pachynema and late pachynema. We categorized oocytes into pachytene substages using a modification of a previously described method based on immunostaining for the DNA DSB marker RPA [26]. Foci of RPA are present on synapsed autosomes at early pachynema but disappear thereafter, with few left by late pachynema. Pachytene oocytes were subdivided into those with >30 autosomal RPA foci (early pachynema) and those with ≤30 autosomal RPA foci (late pachynema). No difference in the percentage of oocytes with a γH2AFX-enriched X chromosome was observed between these two populations (S2C Fig). Thus the decrease does not begin until diplonema. The drop in the percentage of XO oocytes with an asynapsed X chromosome during diplonema cannot be explained by conversion of asynapsed X chromosomes to a synapsed configuration, because chromosomes desynapse during this stage of prophase I. We therefore conclude that oocytes with an asynapsed X chromosome are eliminated during diplonema (see later for additional support). In addition, self-synapsis of the X chromosome is associated with protection against elimination (Fig 1C).
To further confirm our findings that sex chromosome asynapsis drives oocyte elimination, we studied a second sex chromosome variant mouse model, the In(X)1H female. The In(X)1H mouse carries a large X chromosome inversion that disrupts X-X synapsis [27] and has previously been reported to experience prophase I oocyte loss [23]. While the majority of In(X)IH oocytes displayed normal X-X synapsis, defects in X-X synapsis were found in 15% of pachytene oocytes, and again we observed selection of these oocytes during diplonema (Fig 1D and 1F, S2D Fig).
Next, we investigated whether asynapsed autosomes also elicit oocyte elimination. We studied T(16;17)43H/+ (T43H/+) female mice [28], which carry an autosomal translocation involving chromosomes 16 and 17 and in which prophase I oocyte losses have not previously been reported. In T43H/+ females, 40% of pachytene oocytes had autosomal asynapsis (Fig 1G–1I). As with the XO and In(X)1H females, we observed selection against these oocytes as they progress through diplonema (Fig 1I and S2E Fig). Therefore, both sex chromosomal and autosomal asynapsis at pachynema lead to oocyte elimination during diplonema of prophase I.
Synaptic errors have previously been reported in wild type females [10,11]. We therefore examined whether a proportion of chromosomally wild type oocytes experience diplotene elimination. This was indeed the case: synaptic errors were found in 10% of XX oocytes at pachynema but in less than one percent at late diplonema (Fig 1J–1L, S2A and S2F Fig). We conclude that the meiotic surveillance mechanism operates in both chromosomally abnormal and wild type ovaries to eliminate oocytes with chromosomal asynapsis during diplonema (Fig 1M)
Persistent meiotic DNA DSBs cause oocyte loss in targeted recombination mutants, e.g. Dmc1 nulls [3]. Furthermore, a recent study has detected DNA damage foci in Spo11 -/- prophase oocytes [29], indicating that DNA DSBs may also contribute to elimination in this model. To understand if unrepaired DNA DSBs also have a role in the elimination of oocytes with chromosomal abnormalities, we determined whether DSBs persist on asynapsed chromosomes in chromosomally variant mice, using the XO female model. We first analyzed DSB repair on asynapsed chromosomes during pachynema and diplonema using triple immunostaining for SYCP3, HORMAD2 and the DNA damage marker RPA (Fig 2). At early pachynema, RPA foci were present on the asynapsed X chromosome in XO oocytes (Fig 2A and 2B). Importantly however, at late pachynema RPA counts were lower, with many oocytes exhibiting no RPA foci (Fig 2B and 2D). Thereafter, RPA counts remained low during early and late diplonema (Fig 2C and 2D). Analysis of DNA DSB turnover on the asynapsed X chromosome using other DNA damage markers, RAD51 and DMC1, gave similar results: no foci visible on the asynapsed X at late pachynema (Fig 2E–2H). This drop in DNA DSB repair protein counts cannot be explained by oocyte elimination, because no elimination occurs during pachynema (S2C Fig). We conclude that DNA DSBs are repaired on asynapsed chromosomes by the end of pachynema (Fig 2I). These findings mirror those in males, in which disappearance of RPA/RAD51/DMC1 foci on the asynapsed X chromosome occurs during pachynema [30] and is thought to reflect sister chromatid repair [31].
Notably, the behaviour of DNA DSB markers in chromosomally abnormal mice differed to that observed in Dmc1 -/- mutants. In this mutants, RPA foci on asynapsed chromosomes persisted during late prophase I (S3 Fig). These findings show that asynapsis-associated DNA DSBs behave differently in mice with and without targeted mutations of key meiotic genes.
Curiously, while unrepaired DNA markers disappear from unsynapsed chromosomes during pachynema (Fig 2), γH2AFX is retained until the loss of asynaptic oocytes. Thus, our results indicate that oocyte elimination in chromosomally abnormal mice is linked to the presence of H2AFX serine-139 phosphorylation on the chromatin of asynapsed chromosomes. To verify this, we deleted H2afx in our XO mouse model system. We recently showed that H2afx is essential for meiotic silencing in females (Cloutier et al, submitted), as it is in males [32]. H2afx nullizygosity did not influence HORMAD1 and HORMAD2 localization to the asynapsed X chromosome (Fig 3A–3C) or the frequency of pachytene X self-synapsis (Fig 3D–3F) in XO females. Furthermore, the efficiency of autosomal synapsis (S4A Fig) and the number and timing of disappearance of RPA foci on the asynapsed X chromosome (S4B Fig) were unchanged in XO H2afx-/- mice relative to controls.
In our H2afx deletion experiments, we quantified the percentage of oocytes with an asynapsed X chromosome during successive prophase I substages using HORMAD1 and HORMAD2 double-immunostaining, which allowed us to distinguish the asynaptic X chromosome from desynapsing autosomes (Fig 3A–3C). HORMAD2 marks only asynaptic axis, while HORMAD1 marks both asynaptic and desynapsing axis [12]. In control XO H2afx+/- females, oocytes with an asynapsed X disappeared during diplonema (Fig 3D), as had been observed in XO H2afx+/+ females (Fig 1C). Notably, however, in XO H2afx-/- females, there was no significant decrease in the percentage of oocytes with an asynapsed X chromosome during this period (Fig 3E). As a result, XO H2afx-/- females had 3.5 times as many late diplotene oocytes with an asynapsed X chromosome as control XO H2afx+/- siblings (Fig 3F). H2afx deletion had no effect on the percentage of oocytes with X chromosome asynapsis at pachynema (Fig 3D), further confirming that the prophase I elimination mechanism in chromosomally abnormal mice (Fig 1) operates during diplonema. Since serine-139 phosphorylation of H2AFX is the critical epigenetic event in silencing [33], we repeated this analysis using XO females carrying a non-phosphorylatable form of histone H2AFX mutated at serine-139 [34]. Importantly, these mice also exhibited retention of oocytes with an asynapsed X chromosome during diplonema, in contrast to control XO females (S4C Fig). We conclude that H2AFX phosphorylation is required for elimination of XO oocytes with asynapsis.
Next we confirmed the rescue in oocyte loss observed in XO H2afx-/- females by quantifying oocyte numbers in newborn mice. We compared the relative numbers of oocytes, as determined histologically, in XO H2afx+/- with XO H2afx-/- ovaries at 20.5 dpc (i.e. 1 day postpartum), when most oocytes have progressed to diplonema and when losses have previously been shown to occur in XO females [23]. Consistent with our previously observed rescue, we found that XO H2afx-/- females had significantly more oocytes than age-matched XO H2afx+/- females (Fig 3G). Furthermore, oocyte counts in XO H2afx-/- females were similar to those of XX H2afx-/- females (Fig 3G). Importantly, oocyte numbers were not significantly different between XX H2afx-/- and XX H2afx+/+ females (Fig 3G), demonstrating that H2afx deletion has no effect on oocyte viability at this stage of oogenesis. In summary, H2AFX is required for the perinatal oocyte loss in XO females (Fig 3H).
In mammals, asynapsis has been proposed to cause oocyte loss through a checkpoint mechanism responding to asynapsis or through H2AFX-associated silencing, but distinguishing between these models has proved challenging because putative synapsis checkpoint proteins are necessary for silencing [13,19–21]. The checkpoint model predicts that asynapsed chromosomes will cause elimination irrespective of their gene content. By contrast, under the silencing model, the outcome of asynapsis may be linked to the gene content of the asynapsed chromosome, e.g. whether it contains essential oogenesis genes. Cot1 RNA FISH analysis revealed high global gene expression levels in XX prophase I oocytes, especially during diplonema, indicating that all mouse chromosomes harbour oogenesis-expressed genes (S5A–S5D Fig). Therefore, to separate the effects of asynapsis and silencing, we used mice carrying additional chromosomes, so-called “accessory” chromosomes, which by definition carry non-essential genes. These accessory chromosomes are hemizygous and thus are asynapsed during meiosis. If asynapsis per se were the proximal trigger of oocyte elimination, then an accessory chromosome would cause oocyte losses. However, if meiotic silencing were responsible, then an accessory chromosome would not trigger elimination, because silencing would affect genes not normally present in the oocyte genome, and would therefore not be detrimental in this scenario.
We first studied a Down syndrome mouse model, the Tc1 female, which carries a single copy of human chromosome 21 (h21). Asynapsed h21 chromosomes were present in 40% of pachytene oocytes (Fig 4A and 4C) and exhibited γH2AFX domains of size comparable to those in XO females (Fig 4D). In the remaining oocytes, the Tc1 chromosome had self-synapsed (Fig 4B and 4C). Remarkably, oocytes with h21 asynapsis were not eliminated during diplonema (Fig 4C and 4E, S5G Fig). This was in spite of the presence of HORMAD1 (Fig 3A) and other asynapsis-associated proteins on the h21 chromosome (S5E and S5F Fig). As observed in all previous chromosome variant models (Figs 1 and 2), RPA foci disappeared on the asynapsed TC1 chromosome during prophase I progression (S5H Fig).
To confirm our findings, we then studied a second supernumerary chromosome model, the sex-reversed XXYd1 female [35]. XXYd1 females harbor an accessory mouse Y chromosome, which by definition contains no oogenesis-essential genes. Notably, we also observed no selection against oocytes with an asynapsed Yd1 chromosome (S5I–S5K Fig). These data show that asynapsis per se, and the presence of asynapsis-associated proteins, are not sufficient to trigger oocyte elimination. In addition, they suggest that the oocyte loss in XO, In(X)1H, T43H and XX females (Fig 1) is instead linked to the gene content of the asynapsed chromosomes.
Here we have examined the mechanisms that cause prophase I oocyte loss in an extensive array of chromosomally variant mice, shedding new light on fundamental aspects of meiosis and on meiotic surveillance pathways (Fig 5). We show that oocyte loss in these models occurs during diplonema, later than predicted by the classical pachytene checkpoint model [17]. Furthermore, we find that DNA DSB foci disappear from asynapsed chromosomes during pachynema in chromosomally abnormal mice. DNA DSB repair was observed in all models studied, whether or not they exhibit prophase I oocyte elimination. Similar findings have been made in other models, e.g. Sycp3 null mice, in which surviving oocytes were found to resolve axial element-associated DNA DSBs [36]. We attribute the disappearance of DNA DSB foci to the existence in the mammalian oocyte of a repair mechanism for asynapsis-associated DNA DSBs that occurs independent of H2afx (S4B Fig). This most likely proceeds via inter-sister repair, which occurs more commonly during meiosis than previously thought [37]. Importantly, the behaviour of asynapsis-associated DNA DSBs in chromosomally abnormal mice differs from that seen in recombination defective mice, e.g. Dmc1-/- mutants. This is presumably because the targeted genes are themselves necessary for DNA repair, and thus their absence causes persistent DNA damage [38]. This difference emphasises the importance of studying both targeted and non-targeted mouse models to gain a full understanding of the pathways causing prophase I oocyte elimination.
We also show that deletion of H2afx rescues perinatal oocyte loss in XO females. How H2AFX phosphorylation drives oocyte elimination is unclear. However, an important finding comes from our analysis of accessory chromosome mice: this shows that the presence of asynapsed chromosomes per se and of asynapsis-associated factors, e.g. HORMAD1, HORMAD2, BRCA1, ATR and γH2AFX, is insufficient to cause diplotene oocyte elimination. This observation cannot be readily explained by simple synapsis checkpoint models. While it is conceivable that accessory chromosomes somehow do not efficiently activate such a checkpoint, we find it unlikely, since our experiments reveal no qualitative or quantitative differences in the asynapsis response between these models and those that exhibit diplotene oocyte loss. We therefore favour a scenario in which the differential effect of γH2AFX on oocyte fate in chromosomally abnormal versus accessory chromosome mice is due to the gene content of the asynapsed chromosome.
H2AFX phosphorylation could cause elimination through silencing of germ cell-specific or housekeeping protein-coding, or non-coding genes, or indirectly through changes in the balance of transcription factor binding profiles on asynapsed versus synapsed chromosomes. Our data do not allow us to discriminate between these possibilities. However, when combining our observations in the female with those in the male germ line, the most parsimonious model currently invokes a role for silencing of germ-cell expressed genes in prophase I loss. During male meiosis, accumulation of HORMAD1, HORMAD2, BRCA1, ATR, MDC1, and γH2AFX at asynapsed autosomes causes prophase I elimination, but localization of the same proteins to the asynapsed X chromosome does not. The X chromosome does not possess unique properties preventing it from triggering loss. This is demonstrated by the fact that asynapsed accessory chromosomes, such as in Tc1 males, also fail to trigger prophase I elimination (S6A–S6C Fig). Interestingly, however, in contrast to the autosomes, the X chromosome is dramatically depleted in genes required for male meiosis [39,40]. Furthermore, silencing of X-linked housekeeping genes, e.g. Pgk1, Gdpdx is compensated for by a unique system of autosomally-located, X-derived retrogenes that are expressed only in males and are essential for spermatogenesis [39,41,42]. The fact that both the X chromosome and accessory chromosomes are deficient in male meiotic genes could readily explain why H2AFX-induced silencing of these chromosomes does not induce prophase I loss.
Our analysis shows that oocyte elimination occurs during diplonema, i.e. perinatally, in all models studied, irrespective of the identity of the asynapsed chromosome. Perinatal oocyte loss is also observed in other models exhibiting asynapsis and associated meiotic silencing, e.g. Spo11 null females. Under the meiotic silencing model, the prediction would be that all chromosomes house genes required either specifically for late prophase I, or for general housekeeping functions. We note that the asynapsed chromosomes studied in our model systems carry such genes: aside from aforementioned housekeeping genes, the X chromosome is enriched in genes required for oogenesis [43], including Zfx [44], Bmp15 [45] and Fmr1 [46]. The region of maximum chromosome 17 asynapsis in T43H/+ females [47] also contains critical genes, including the splicing factor Srpk1, ribosome Rpl10a and mRNA capping factor Cmtr1. Given that meiotic silencing covers megabase-scale chromosome regions, inactivating hundreds of genes, oocyte loss could result from the contemporaneous disturbance of multiple biological pathways. How such disturbances trigger oocyte demise is unclear. Although we observe TUNEL-positive oocytes in XO females at 19.5dpc, when asynapsis-associated oocyte elimination is taking place, the numbers are not significantly elevated relative to XX females (S7 Fig).
Finally, a previous study examined the effect of ablating meiotic silencing in Smc1b null oocytes, and concluded that silencing had a limited role in oocyte loss [10]. In those experiments, meiotic silencing was prevented by removing the axial element component Sycp3. While some degree of oocyte rescue was observed perinatally, the effect was only transient, and was not sustained at 4dpp. However, it should be noted that Smc1b null oocytes harbour additional lesions, e.g. defective recombination, that are not present in the chromosomally abnormal mouse models studied herein, and could independently trigger oocyte elimination [48]. Thus, in targeted mutants, multiple surveillance pathways could conspire to drive oocyte elimination.
All animal procedures were in accordance with the United Kingdom Animal Scientific Procedures Act 1986 and were subject to local ethical review.
Females were set up in matings and checked daily for copulation plugs. The day of plugging was considered 0.5 days post coitum (dpc). Embryos were sacrificed at 17.5, 18.5, 19.5 and 20.5 dpc using UK Home Office Schedule I methods. Ovaries were dissected from embryos and flash frozen in liquid nitrogen. Material was stored at -80°C until later use. All mice were maintained according to UK Home Office regulations. XO mice were generated on a random bred MF1 background (NIMR stock) by mating XX females to fertile XY*O males, which harbor an X chromosome fused with a Y chromosome and give rise to ‘O’ gametes [49]. T(16;17)43H mice were maintained on a C57BL/10ScSnPh (B10) background, as previously described [28,47]. H2afx-/- mice [50] were generated on the MF1 background. XO H2afx-/- mice were generated by crossing XYO H2afx+/- males with XX H2afx+/- females. Dmc1-/- mice [38] were produced on a C57BL/6 background. Tc1 mice [51] were maintained on the MF1 background. The C57BL/6 males used in this cross were maintained on-site at NIMR. XXY d1 females were produced on an MF1 background by mating XY males to sex-reversed XYd1 females [35]. Mice carrying the H2afxS136/139A transgene were described previously [52].
Surface spreads and chromosome painting were performed as previously described [16,53]. The following primary antibodies were used for immunofluorescent experiments: rabbit anti-SYCP3 (1:100, Abcam: ab15093), mouse anti-γH2AFX (1:100, Upstate: 16–193; 1:100), guinea pig and rabbit anti-HORMAD1 and anti-HORMAD2 (ref. [12], 1:200), rabbit anti-RPA32/RPA2 (Abcam, ab-10359; 1:100), rabbit anti-BRCA1 (1:100, gift from Chu-Xia Deng), goat anti-ATR (1:50, Santa Cruz: sc-1887). Primary antibody incubations were carried out overnight at 37°C and secondary antibody incubations for 1hr at 37°C. For chromosome spread analyses, oocytes were first categorized into meiotic substages based upon SYCP3 and HORMAD1 staining, as described in S1 Fig. After the meiotic substage was determined, the oocytes were then assessed for γH2AFX domains or HORMAD2 axial staining (for XO H2afx-/- experiments). For quantification of the γH2AFX signal in spread oocytes (Fig 4D), images were taken with matched exposure times. We calculated the integrated intensity of γH2AFX domain area using Fiji software [54]. For each image we estimated background intensity in a region outside of the nucleus. We then calculated the background-adjusted intensity value of the γH2AFX domains by subtracting the background equal to the area of the γH2AFX domain.
Ovaries were harvested from females at 20.5 days post-coitum (dpc), fixed in 4% paraformaldehyde overnight at 4°C and then transferred to 70% ethanol. Fixed ovaries were dehydrated by three successive 5min incubations with 95% ethanol, 100% ethanol, 100% xylene and were then embedded in paraffin wax. Ovaries were serially sectioned at 5–7μm thickness. Sections were dewaxed using histoclear (2x5min) and 1:1 histoclear:ethanol (1x5min), and then rehydrated using the following ethanol series: 100% ethanol (2x5min), 95% ethanol (1x5min), 80% (1x5min), 70% (1x5min), 50% (1x3min), and PBS (1x5min). Sections were stained with DAPI and oocytes were identified based upon their distinct size and nuclear cytology, as described previously [23]. To quantify the relative number of oocytes in each ovary, we summed the oocyte counts from every tenth section, as described previously [19].
Tunel analysis was performed using the ApopTag Plus Peroxidase In Situ, Apoptosis Detection Kit, S710, Millipore, according to manufacturer’s instructions.
Statistical analysis was performed using GraphPad Prism 6.0. For statistical comparison of means, a Tukey-Kramer multiple comparison test (for >1 comparisons) or an unpaired t test (for one comparison) was performed. Calculated P values are reported in figures or legends.
Imaging was performed using an Olympus IX70 inverted microscope with a 100-W mercury arc lamp. For chromosome spread and RNA FISH imaging, an Olympus UPlanApo 100x/1.35 NA oil immersion objective was used. For ovary section imaging, an Olympus UPlanApo 20x/0.75 NA objective was used. A Deltavision RT computer-assisted Photometrics CoolsnapHQ CCD camera with an ICX285 Progressive scan CCD image sensor was utilized for image capture. 16-bit (1024x1024 pixels) raw images of each channel were captured and later processed using Fiji. Quantitation of Cot1 and γH2AFX intensities was performed as previously described [55].
|
10.1371/journal.pmed.1002306 | Association between expansion of primary healthcare and racial inequalities in mortality amenable to primary care in Brazil: A national longitudinal analysis | Universal health coverage (UHC) can play an important role in achieving Sustainable Development Goal (SDG) 10, which addresses reducing inequalities, but little supporting evidence is available from low- and middle-income countries. Brazil’s Estratégia de Saúde da Família (ESF) (family health strategy) is a community-based primary healthcare (PHC) programme that has been expanding since the 1990s and is the main platform for delivering UHC in the country. We evaluated whether expansion of the ESF was associated with differential reductions in mortality amenable to PHC between racial groups.
Municipality-level longitudinal fixed-effects panel regressions were used to examine associations between ESF coverage and mortality from ambulatory-care-sensitive conditions (ACSCs) in black/pardo (mixed race) and white individuals over the period 2000–2013. Models were adjusted for socio-economic development and wider health system variables. Over the period 2000–2013, there were 281,877 and 318,030 ACSC deaths (after age standardisation) in the black/pardo and white groups, respectively, in the 1,622 municipalities studied. Age-standardised ACSC mortality fell from 93.3 to 57.9 per 100,000 population in the black/pardo group and from 75.7 to 49.2 per 100,000 population in the white group. ESF expansion (from 0% to 100%) was associated with a 15.4% (rate ratio [RR]: 0.846; 95% CI: 0.796–0.899) reduction in ACSC mortality in the black/pardo group compared with a 6.8% (RR: 0.932; 95% CI: 0.892–0.974) reduction in the white group (coefficients significantly different, p = 0.012). These differential benefits were driven by greater reductions in mortality from infectious diseases, nutritional deficiencies and anaemia, diabetes, and cardiovascular disease in the black/pardo group. Although the analysis is ecological, sensitivity analyses suggest that over 30% of black/pardo deaths would have to be incorrectly coded for the results to be invalid. This study is limited by the use of municipal-aggregate data, which precludes individual-level inference. Omitted variable bias, where factors associated with ESF expansion are also associated with changes in mortality rates, may have influenced our findings, although sensitivity analyses show the robustness of the findings to pre-ESF trends and the inclusion of other municipal-level factors that could be associated with coverage.
PHC expansion is associated with reductions in racial group inequalities in mortality in Brazil. These findings highlight the importance of investment in PHC to achieve the SDGs aimed at improving health and reducing inequalities.
| The Sustainable Development Goals include reducing inequalities and making commitments to universal health coverage (UHC).
There is little evidence about the relationship between expanding primary healthcare (PHC)—as part of the commitment to UHC—and health inequalities, including racial inequalities. This is particularly true for low- and middle-income countries.
Racial health inequalities are important inequalities to study given the sharp disparities that exist in Brazil and globally.
We examined trends in mortality from ambulatory-care-sensitive conditions for black/pardo (mixed race) and white Brazilians from 2000 to 2013, and evaluated whether there were changes in mortality associated with expansion of PHC in municipalities.
PHC expansion was associated with reductions in mortality for both racial groups, but black/pardo Brazilians experienced a 2-fold greater reduction in mortality than white Brazilians.
The targeted rollout of PHC in Brazil to poorer and smaller municipalities and the greater unmet needs of black/pardo Brazilians at the start of the rollout are likely to explain these findings.
There is evidence of an association between expanded PHC and reductions in health inequalities in Brazil.
PHC that is preferentially expanded as part of UHC to more disadvantaged populations has the potential to reduce health inequalities.
| Reducing inequalities within and among countries is the tenth goal of the Sustainable Development Goals (SDGs). This goal includes the target to “adopt policies, especially fiscal, wage and social protection policies” that “progressively achieve greater equality” (http://www.un.org/sustainabledevelopment/inequality/). Health systems are essential for social protection and, in addition to their contributions to other SDGs for health, may play a vital role in reducing inequalities [1]. Additionally, promoting equality in access to healthcare is a core principle of universal health coverage (UHC) [2]. Investment in primary healthcare (PHC), as part of efforts to achieve UHC, may be especially important in reducing health inequalities [3–5], but evidence is largely derived from North America and Europe.
Brazil is an important setting for evaluating the relationship of PHC with health inequalities. It is a middle-income country with one of the highest levels of income inequality globally (a Gini coefficient of 52.9 in 2013 [6]) and stark health inequalities across income, education, racial, and socio-economic groups [7–13]. Brazil’s considerable investments in social protection policies over the last two decades include the rollout of conditional cash transfers under the Bolsa Família programme and a commitment to UHC with the expansion of PHC through the Estratégia de Saúde da Família (ESF) (family health strategy) [14,15]. The ESF has rapidly expanded since the mid-1990s to become the largest community-based PHC program in the world [16]. In 2014, it covered ~121.2 million individuals (~62.5% of the population) [17]. Family health teams composed of a family doctor, nurses, and community health workers deliver a broad range of comprehensive and preventive healthcare services to defined local populations (approximately 3,400 individuals) [15]. Municipal governments are responsible for the provision of local ESF services, and financial incentives provided by the federal government encourage municipalities to adopt the ESF [18]. In general, municipalities with smaller populations, higher levels of poverty, and a higher proportion of residents from black/pardo (mixed race) racial groups exhibited greater uptake of the ESF (S1 Appendix, Figs. A–C) [19]. Expansion of the ESF has been associated with reductions in infant mortality [20–22], deaths from cardiovascular disease [4], and hospitalisations from ambulatory-care-sensitive conditions (ACSCs) [5], but there is little understanding of the associations between ESF expansion and changes in health inequalities. Recent financial and political crises in Brazil are threatening funding for social protection policies, including UHC [23]. Evidence of an association between the ESF and a reduction of inequalities in health outcomes would provide a strong argument for continued investment and political support.
Assessing racial inequalities is important for evaluating the ESF, given the complex historical, sociological, and political dimensions of race in Brazil [24,25]. In contrast to ancestral and ethnic classifications of race in the US and the UK [13], institutions in Brazil use skin colour. Official classifications are branco (white), preto (black), pardo (brown/mixed), amarello (Asian), and indigenous, with white, black, and pardo accounting for over 98% of the population. Self-reported classification, whilst reflecting ancestral and cultural roots, also reflects an individual’s perceived social identity [11,13,25]. Three main ancestral roots established the Brazilian population today—indigenous individuals, European colonisers, and African slaves [25]. Today, there is considerable admixture (evidenced by a sizeable pardo population), but sharp inequalities between racial groups persist [9–13]. Black and pardo populations have higher illiteracy, have lower average incomes, and use healthcare services less [9]. In health outcomes, they have lower life expectancy, are affected more by infectious diseases (including tuberculosis, leprosy, leishmaniasis, and schistosomiasis), and have higher mortality rates from external causes, drug overdoses, and homicides [9].
Few studies have examined the potential role of PHC in reducing health inequalities in low- and middle-income countries. This study seeks to address this important gap by examining associations between ESF coverage and mortality from ACSCs in white and black/pardo populations in Brazil. We test the hypothesis that expansion of PHC coverage through the ESF in Brazil is associated with reduced inequalities in mortality between racial groups [26].
Longitudinal (panel data) regression models were employed using routinely collected municipal-level data, which have been widely applied to evaluate the ESF previously [4,20,22,27–30]. These models estimated associations between ESF coverage and mortality from ACSCs among black/pardo and white populations over time, whilst controlling for other confounding factors. The main analysis was restricted to 1,622 municipalities based on previously assessed quality of vital statistics reporting to reduce bias from under-reporting of deaths [31]. Differences in our analytic approach from previous ESF evaluations were necessary to examine associations of ESF expansion and inequalities in mortality between racial groups. These were agreed before compilation and analysis of the data (which commenced in February 2016), and are set out in detail below. In response to reviewers’ suggestions after initial submission, we explored factors associated with ESF uptake, tested for pre-existing trends, tested for biases from ill-defined death adjustments, explored interactions with Bolsa Família, and conducted sensitivity analyses with alternative model specifications and, for comparison with ACSC mortality, on mortality from accidents.
Data from individual death certificates for the years 2000–2013 were obtained from the Brazilian Ministry of Health DATASUS website [32]. Annual municipal population estimates by race and age group based on census data were obtained from the Instituto Brasileiro de Geografia e Estatística (IBGE) website [33]. Municipal-level covariate data, including illiteracy rate, poverty rate, urbanisation rate, and municipal gross domestic product (GDP), were obtained from the IBGE website [33]. Municipal ESF coverage, Bolsa Família coverage, public healthcare spending, the number of public hospital beds, the number of private hospital beds, and private health insurance coverage were obtained from the DATASUS website [32].
The mortality rate from ACSCs was the main outcome variable. ACSC deaths were encoded based on a list published by the Brazilian Ministry of Health (and restricted to those aged under 70 y) and ICD-10 codes reported on death certificates (Table 1) [34]. ACSCs were grouped by cause of death into infectious diseases, nutritional deficiencies and anaemia, chronic obstructive pulmonary disease (COPD) and asthma, cardiovascular disease, diabetes, epilepsy, and gastric ulcers. Redistribution of ill-defined deaths was performed using a published and previously utilised methodology to control for confounding trends from reductions in ill-defined deaths over time (S1 Text) [35].
Race is recorded on death certificates and as part of the decennial census in Brazil. Census recording of race is self-reported. Individuals select branco (white), preto (black), pardo (brown/mixed), amarello (Asian), or indigenous. Recording of race on death certificates (using the same categories) is usually completed by the physician certifying the death and should be based on input from the family [13]. Amarello and indigenous deaths were very few and not examined. Black and pardo deaths were merged into one group, despite issues regarding differences between these populations [13]. This was to overcome potential differences in racial classification of individuals occurring either between censuses and death certificates, or over time as individuals and/or society changed reporting behaviour. Whilst evidence indicates overlap between black and pardo classifications in reporting of race, there are significantly clearer divisions between white and pardo classifications [36].
Reporting of race is near complete in censuses (99.29% in 2000 and 99.98% in 2010) and high on death certificates (total missing for 2000–2013 was 5.8%). For completeness, values were imputed for certificates with race missing using other death certificate variables (sex, age, education level, marital status, and location of death) and municipal population estimates of racial groups (S2 Text). For the period 2000–2013, race was imputed for 39,198 of the total 588,872 ACSC deaths (of those white or black/pardo and aged under 70 y) in the municipalities included in the analysis.
Using municipal census population data, population distributions by race (white and black/pardo) and age group (0–4, 5–9, 10–14, 15–19, 20–24, 25–29, 30–39, 40–49, 50–59, and 60–69 y) were calculated for each municipality for the census years (2000 and 2010), and were linearly interpolated and extrapolated for non-census years (2001–2009 and 2011–2013). Annual total municipal population estimates were used to calculate annual age and race group population estimates for each municipality. Direct age standardisation of cause of death by race was performed, producing annual age-adjusted mortality rates for total ACSCs and ACSC groups by race. The dependent variables (for each municipality and for each year) in the regression models were the expected (from age standardisation) number of deaths from ACSCs (in total and by ACSC group) for the black/pardo and white populations and the standardised rate ratio (SRR) between total black/pardo and white ACSC mortality rates. Rate ratios (RRs) are commonly used metrics for comparing rates between groups (e.g., between males and females) [37]. In this study, the ACSC mortality rate for the black/pardo population was divided by the ACSC mortality rate for white population.
The main variable of interest was municipal ESF coverage (percent) of the population, with official calculations based on one ESF team per 3,450 individuals [17]. A 2-y average (within the year and the year prior) of ESF coverage was employed, even though comparable results were obtained with just within-year coverage or including 2- and 3-y lags. This approach was used to account for varying lagged and duration effects of the ESF that may differ between conditions and populations, to account for the time for ESF services to become fully operational and effective, and to permit simple comparison between the two racial groups.
Annual municipality-level covariate data were selected to include variables relating to socio-economic development, income, and the health system, which have been shown to affect mortality [38,39]. The covariates were scaled as percentages, in hundreds of Brazilian reais (R$100s) per person (adjusted for inflation), or per 1,000 inhabitants. Variables expressed as percentages were scaled between 0 and 1 so a one-unit increase would represent a 100% increase. Where necessary, logarithms were used to improve model fit. Covariates employed in all models were: Bolsa Família coverage (percent), illiteracy rate in those over 25 y (percent) (log-transformed), poverty rate (percent), population living in urban areas (percent), public healthcare spending (R$100s per person), public hospital beds per 1,000 population, private hospital beds per 1,000 population, private healthcare insurance (percent) (log-transformed), and GDP per person (R$100s per person) (log-transformed). An interaction between private healthcare insurance (percent) (log-transformed) and GDP per person (R$100s per person) (log-transformed) was included for model fit.
Descriptive analyses were undertaken, including national trends of ACSC mortality rates for black/pardo and white populations and the national SRR of the two rates.
Fixed-effects longitudinal regression was employed as an appropriate method for analysing annual observations of municipalities [40]. Fixed-effects models control for time-invariant unobserved factors that may affect mortality and could bias the results [40]. Consequently, only changes within municipalities over time are estimated rather than differences between municipalities. We tested for pre-intervention trends (i.e., mortality rates prior to ESF adoption and expansion) to determine whether time-varying unobserved factors could bias the results. Examining trends in the years 2000–2003 (when many municipalities still had relatively low coverage) and employing dummy variables for the years prior to ESF adoption revealed no evidence of pre-intervention trends.
In the models with dependent count variables (ACSC deaths), a Poisson model with a population offset term was employed, allowing the dependent variable (ACSC deaths) to be modelled as a rate (deaths per population). To aid interpretability, the coefficients were exponentiated and reported as RRs. These are interpreted as a ratio of the mortality rates for a one-unit increase in the independent variable (e.g., a 100% increase in ESF coverage or an additional year during the study period) (see S3 Text for more details). In other words, the difference between 1 and the RR can be interpreted as the percentage change in the rate given a one-unit increase in the independent variable. For the SRR, linear longitudinal regression was employed, and β coefficients reported. These are interpreted as the change in the SRR given a one-unit increase (i.e., from 0% to 100% coverage).
Two multiple regression models were undertaken examining the association between ESF expansion and ACSC mortality in the black/pardo and white populations separately. Differences in the effect sizes were tested for statistical significance (S4 Text). The p-values for the differences between the coefficients from the two models are reported in the text. The association between ESF coverage and the SRR was examined with a multiple regression model. Several regression models for the groups of ACSCs (infectious diseases, nutritional deficiencies and anaemia, COPD and asthma, cardiovascular disease, diabetes, epilepsy, and gastric ulcers) were employed in the black/pardo and white populations separately. Small numbers prohibited the use of SRR for groups of ACSCs. In all models, municipality-clustered robust standard errors were employed to account for possible auto-correlation and heteroscedasticity [40]. Stata 12 was used for statistical analysis.
Multiple sensitivity analyses were undertaken to check the robustness of the findings. First, alternative model specifications with sequential addition of covariates, random-effect models, and negative-binomial models were employed (S2 Appendix, Tables A and B). Second, varying classifications of ESF coverage were tested (S2 Table). Third, mortality from accidents (ICD-10 V01–X59) was tested, as an outcome that should have no association with ESF expansion (S3 Table). Fourth, the validity of imputing race on death certificates with race missing was assessed by excluding deaths where race was not recorded (S3 Appendix, Tables A and B). Fifth, the validity of redistributing ill-defined causes of death was tested (S4 Appendix, Tables A and B). Sixth, the analyses were repeated using data from all 5,565 municipalities in Brazil, not just those with adequate recording of vital statistics (S5 Appendix, Tables A and B). Seventh, because the potential for misclassification of race on death certificates exists (between the white and black/pardo populations), the effect of reclassifying black/pardo deaths (which are higher) as white was examined (S6 Appendix, Tables A–H). Eighth, an interaction between Bolsa Família and ESF coverage was examined (S7 Appendix).
Between 2000 and 2013, there were 281,877 and 318,030 deaths from ACSC causes in the black/pardo and white populations, respectively (after age standardisation). Age-standardised ACSC mortality rates fell 37.9%, from 93.3 to 57.9 per 100,000, in the black/pardo population and by 34.9%, from 75.7 to 49.2 per 100,000, in the white population (Fig 1; S7 Appendix). Mortality from ACSC causes in the black/pardo population was between 17% and 23% higher than in the white population during the study period. There was a sizeable expansion of the ESF over the period, both in terms of the number of municipalities adopting the ESF and the average municipal ESF coverage (Fig 2).
In longitudinal Poisson regression models, ACSC mortality decreased annually by 3.4% (RR: 0.966; 95% CI: 0.954–0.976) in the black/pardo population and by 2.9% (RR: 0.971; 95% CI: 0.963–0.979) in the white population in adjusted models (Table 2). ESF expansion (from 0% to 100% coverage) was associated with a 15.4% (RR: 0.846; 95% CI: 0.796–0.899) reduction in ACSC mortality in the black/pardo population and a 6.8% (RR: 0.932; 95% CI: 0.892–0.974) reduction in the white population. These coefficients were significantly different (p = 0.012).
ESF expansion (from 0% to 100% coverage) was associated with a 0.179 reduction (95% CI: 0.022–0.336) in the SRR (Table 3). Predicted SRRs from the model demonstrate that if ESF coverage were 0% in all municipalities, mortality amenable to PHC in the black/pardo population would be 29.6% higher than that in the white population (an estimated SRR of 1.296). With 100% ESF coverage in all municipalities, mortality amenable to PHC in the black/pardo population would be 11.7% higher than that in the white population (an estimated SRR of 1.117). Thus, expansion of the ESF (from 0% to 100%) yields a 60.5% reduction in the excess mortality that the black/pardo population experiences over the white population.
Over the study period, mortality from COPD and asthma decreased annually by 4.1% (RR: 0.959; 95% CI: 0.933–0.985) in the black/pardo population and by 4.5% (RR: 0.955; 95% CI: 0.939–0.971) in the white population (Table 4). Mortality from cardiovascular disease decreased annually by 3.7% (RR: 0.963; 95% CI: 0.948–0.979) in the black/pardo population and by 2.7% (RR: 0.973; 95% CI: 0.962–0.984) in the white population. For the black/pardo population, mortality from diabetes decreased 2.7% per year (RR: 0.973; 95% CI: 0.952–0.994), whilst there were non-significant trends in infectious diseases, nutritional deficiencies and anaemia, epilepsy, and gastric ulcers. For the white population, mortality from infectious diseases decreased 2.8% annually (RR: 0.972; 95% CI: 0.948–0.997), mortality from nutritional deficiencies and anaemia decreased 4.9% annually (RR: 0.951; 95% CI: 0.909–0.994), and mortality from gastric ulcers decreased 4.9% annually (RR: 0.951; 95% CI: 0.922–0.981), but there were no significant trends in diabetes and epilepsy mortality.
ESF expansion (from 0% to 100%) was associated with a decrease in mortality from cardiovascular disease of 12.9% (RR: 0.871; 95% CI: 0.801–0.947) and 7.1% (RR: 0.929; 95% CI: 0.876–0.985) in the black/pardo and white populations, respectively. In the black/pardo population, ESF expansion was associated with 27.5% lower mortality from infectious diseases (RR: 0.725; 95% CI: 0.620–0.848) and 19.3% lower mortality from diabetes (RR: 0.807; 95% CI: 0.713–0.912), but there was no significant association with mortality for these ACSC groups in the white population. ESF expansion was associated with 17.9% lower mortality from nutritional deficiencies and anaemia (RR: 0.721; 95% CI: 0.478–0.899) in the black/pardo population, but in the white population, it was associated with 25.1% higher mortality (RR: 1.251; 95% CI: 1.011–1.548). For both the black/pardo and white populations, there was no significant association between ESF and mortality from COPD and asthma, epilepsy, or gastric ulcers.
Sensitivity analyses demonstrate the robustness of our findings. Alternative model specifications (S2 Appendix, Tables A and B) demonstrate the stability and robustness of the findings. We found that controlling for additional factors (fixed effects, covariates, and state-year fixed effects) did not change our findings; in fact, the differential associations of the ESF with black/pardo and white mortality became more apparent when these factors were taken into account. Alternative classifications of ESF coverage did not change the overall differences in the associations between ESF expansion and black/pardo and white mortality, although the results of the sensitivity analysis did suggest that greater reductions in mortality in the black/pardo population accrued over a longer period (S2 Table).
Examining mortality from accidents as a control outcome revealed no significant association of accident deaths with ESF coverage in either racial group, adding to the robustness of our findings (S3 Table). Excluding deaths with race not recorded yielded highly comparable results, demonstrating that imputation of missing race data was not a source of bias (S3 Appendix, Tables A and B). Repeating the analysis without adjustment for ill-defined deaths produced similar results (S4 Appendix, Tables A and B). An analysis with all 5,565 municipalities in Brazil (not just those with adequate reporting of vital statistics) found lower ACSC mortality associated with ESF expansion only in the white population yet a highly comparable association between ESF expansion and changes in the SRR (S5 Appendix, Tables A and B). The non-significance of the association of ACSC mortality with ESF expansion in the black/pardo population (when including municipalities with inadequate reporting of vital statistics) is expected given the likelihood of black/pardo deaths being under-reported and the role of the ESF in reducing under-reporting [41]. To examine the extent to which misclassification bias (i.e., deaths that were encoded as black/pardo when individuals self-identified as white in the census) could affect the results, we randomly reclassified 10%, 20%, and 30% of black/pardo deaths as white deaths (S6 Appendix, Tables A–H). A similar association between ESF expansion and race-specific mortality was found even when 30% of black/pardo deaths were reclassified, although associated reductions in inequality were lower. Lastly, an interaction between Bolsa Família and ESF coverage was non-significant (S7 Appendix, Tables A–C).
Expansion of the ESF between 2000 and 2013 in Brazil was associated with a 2-fold greater reduction in ACSC mortality in the black/pardo compared to the white population. This differential benefit reduced racial inequalities in mortality and was driven by greater reductions in deaths from infectious diseases, nutritional deficiencies and anaemia, diabetes, and cardiovascular disease in the black/pardo population. This paper provides further evidence of the importance of expanding UHC in low- and middle-income countries.
Previous literature indicates that ESF expansion is associated with reduced child mortality, mortality from cardiovascular disease, and ACSC hospitalisations [4,5,20–22]. These changes are likely due to improved access to healthcare and a focus on prevention, health promotion, proactive outreach, and early management of conditions within the ESF [3]. Whilst there is local variation in how the ESF is implemented, federal guidelines specify minimum mandatory strategic areas ESF teams must cover, including the management of hypertension, diabetes, tuberculosis, and women’s and children’s health [42]. In this study, ESF expansion was associated with reductions in mortality in the ACSC groups that mirror these mandatory strategic areas. We found that ESF expansion was associated with reductions in cardiovascular mortality of 12.9% and 7.1% in the black/pardo and white groups, respectively. We found a 27.5% reduction in mortality from infectious diseases with ESF expansion in the black/pardo population. ESF expansion was also associated with a 17.9% reduction in mortality from nutritional deficiencies and anaemia in the black/pardo population, with children under 5 y accounting for over 25% of these deaths (compared to roughly 3% of all deaths from ACSCs). Additionally, ESF-associated reductions in mortality from respiratory diseases (COPD and asthma), epilepsy, and gastric ulcers are consistent with their inclusion within ACSC definitions and the fact that these conditions are considered amenable to PHC. We found no association between ESF expansion and mortality from accidents, which is not considered sensitive to primary care, providing reassurance that the associations of ESF expansion with ACSCs reported are not due to confounding.
The differential associations between ESF expansion and mortality in black/pardo and white populations may be explained by numerous factors, with socio-economic differences a key explanatory factor. Black/pardo populations are disproportionately affected by diseases of poverty, including infectious diseases, malnutrition, and anaemia [9,43], but these conditions may be more responsive to ESF services as they are generally easier to treat in PHC settings than complex non-communicable diseases. Additionally, black/pardo populations in Brazil have lower utilisation of healthcare and higher rates of forgone healthcare [9], suggesting ESF expansion may have facilitated access to healthcare and reduced unmet need. Lastly, the finding that ESF has benefitted black/pardo populations more than white populations may not be surprising given that ESF expansion had been prioritised within poorer areas and municipalities. Surveys indicate that black/pardo populations now have greater ESF coverage (at 57.3% in 2008) than white populations (44.6% in 2008), but lower coverage of private health insurance, suggesting they are more reliant on publicly funded and provided services, including the ESF [9].
Our findings are consistent with evidence derived largely from studies conducted in North America and Europe that show “equity-enhancing” associations from PHC expansion [3]. However, these studies mostly examine associations of PHC with health inequalities across income groups. There are fewer studies examining the association of PHC with health inequalities between racial groups. In a study in the US, increasing the supply of primary care physicians was associated with larger reductions in African-American mortality than white mortality [38]. Inequalities in low birth weight between African-American and white infants are lower among those using PHC [44]. No evidence exists on the association between PHC and race in Brazil, although a few studies have examined inequalities between municipalities. Previous Brazilian studies have shown that ESF expansion was associated with greater reductions in infant mortality in municipalities with higher infant mortality at baseline [20,45]. Another study demonstrated greater reductions in infant mortality in municipalities with lower human development, also implying improvements in equity [20].
There are important limitations to this study pertinent to the interpretation. First, these analyses were conducted on municipal-level aggregated data, and more complete, individual data with ESF enrolment, consultation rates, and associated health outcomes are required to elucidate the mechanisms determining the greater benefits experienced by the black/pardo population. Second, there could be biases from the methods employed and data manipulation. However, we conducted extensive sensitivity analyses that showed that our findings are robust to ill-defined death reclassification, varying classifications of ESF coverage, and alternative model specifications. We also found no evidence of pre-intervention trends that would bias the findings. Third, there are important caveats regarding the use of race in this study. There is potential for misclassification bias of race (with race in censuses self-reported and race in death certificates reported by either the family or physician), although sensitivity analyses indicate the robustness of the findings. Black/pardo would have to be incorrectly recorded for over 30% of black/pardo deaths for the differences found to be non-significant. Additionally, by grouping together black and pardo deaths, we do not account for the large amount of heterogeneity in health outcomes between these groups [13]. Fourth, lack of statistical power due to small numbers is apparent in our analysis of associations between ESF expansion and cause-specific deaths. This precluded any potential analysis with SRRs for ACSC groups. Fifth, we used mortality from ACSCs as our outcome measure rather than the more broadly defined concept of healthcare-amenable mortality [30,46]). This was principally to focus on conditions that have been defined as amenable to PHC within the Brazilian context and to exclude those that may be strongly influenced by hospital-based care. While previous research has generally examined hospital admissions for ACSCs, this was not feasible here due to low recording of race in hospital admission data in Brazil. We present a comparison of conditions included in the Brazilian Ministry of Health’s definition of ACSCs and healthcare-amenable mortality as defined by Nolte and McKee [46] in S4 Table.
Policy-makers should note that in Brazil, where sharp inequalities persist and an ambition to achieve UHC has been boldly pursued over the last 20 years, the equity-promoting associations of PHC are evident [30]. The strong positive relationship between PHC and reduced racial inequalities in mortality provides impetus for a renewed government commitment to the ESF. Current proposals that could limit public spending in Brazil and cause disinvestment from social protection programmes, including the ESF [23], may reverse the valuable progress made towards reducing health inequalities in the country. The health inequality impacts of policy changes influencing the ESF, which is the primary vehicle for UHC in Brazil, should be carefully monitored and evaluated.
Beyond the equity-enhancing nature of PHC itself, the impressive reductions in inequality in ACSC deaths between racial groups seen in Brazil may have been facilitated by numerous factors. These include the more rapid expansion of the ESF in poorer and more deprived areas, and the proactive outreach healthcare delivered by community health workers. Whilst challenges exist, including retaining health professionals in rural areas [15] and a lack of coverage for the urban poor [47], there are valuable lessons for other countries from Brazil’s efforts to achieve UHC. The pro-equity health gains demonstrated here reflect the country’s adoption of a pro-poor pathway to UHC. Universal access was embraced from the start, services are publicly financed, there is a focus on expanding access through community-based models of care, and strong political commitment has enabled rapid and sizeable expansion [48]. Valuable lessons may be derived from other settings including Costa Rica, which similarly expanded PHC in poorer areas preferentially [49], and countries such as Tanzania, Uganda, and Chile, which have accelerated coverage in underserved areas through flexible budget allocations [50].
In conclusion, expansion of the ESF in Brazil was associated with improved health outcomes and reductions in health inequalities between racial groups. As countries aim to “progressively achieve greater equality” as part of the SDGs, these findings reinforce the importance of strong PHC-focused health systems for improving health and reducing health inequities.
|
10.1371/journal.pntd.0001256 | Profiles of Small Non-Coding RNAs in Schistosoma japonicum during Development | The gene regulation mechanism along the life cycle of the genus Schistosoma is complex. Small non-coding RNAs (sncRNAs) are essential post transcriptional gene regulation elements that affect gene expression and mRNA stability. Preliminary studies indicated that sncRNAs in schistosomal parasites are generated through different pathways, which are developmentally regulated. However, the data of sncRNAs of schistosomal parasites are still fragmental and a complete expression profile of sncRNAs during the parasite development requires a deep investigation.
We employed high-throughput genome-wide transcriptome analytic techniques to explore the dynamic expression of microRNAs (miRNAs) and endogenous siRNAs (endo-siRNAs) of Schistosoma japonicum covering the free-living cercarial stage and all stages in the definitive host. This led us to analyze over 70 million clean reads represented both high and low abundance of the small RNA population. Patterns of differential expression of miRNAs and endo-siRNAs were observed. MiRNAs was twice more than endo-siRNAs in cercariae, but gradually decreased along with the development of the parasite. Both small RNA types were presented in equal aboudance in lung-stage schistosomula, while endo-siRNAs accumulated to 6 times more than miRNAs in adult female worms and hepatic eggs. Further, miRNAs were found mainly derived from genes located in the intergenic regions, while endo-siRNAs were mainly generated from transposable elements (TEs). The expression pattern of TE-siRNAs, as well as the pseudogene-derived siRNAs clustered in mRNAs of cytoskeletal proteins, stress proteins, enzymes related to energy metabolism also revealed distinction throughout different developmental stages. Natural antisense transcripts (NATs)-related siRNAs accounted for minor proportion of the endo-siRNAs which were dominantly expressed in cercariae.
Our results represented a comprehensive expression profile of sncRNAs in various developmental stages of S. japonicum with high accuracy and coverage. The data would facilitate a deep understanding of the parasite biology and potential discovery of novel targets for the design of anti-parasite drugs.
| Schistosomiasis, a debilitating disease, caused by agents of the genus Schistosoma afflicts more than 200 million people worldwide. Schistosomes could serve as an interesting model to explore gene regulation due to its evolutional position, complex life cycle and sexual dimorphism. We previously indicated that sncRNA profile in the parasite S. japonicum was developmentally regulated in hepatic and adult stages. In this study, we systematically investigated mircoRNA (miRNA) and endogenous siRNA (endo-siRNA) profile in this parasite in more detailed developmental stages (cercariae, lung-stage schistosomula, separated adult worms, and liver tissue-trapped eggs) using high-throughput RNA sequencing technology. We observed that the ratio of miRNAs to endo-siRNAs was dynamically changed throughout different developmental stages of the parasite. MiRNAs were expressed dominantly in cercariae, while endo-siRNAs accumulated in adult female worms and hepatic eggs. We demonstrated that miRNAs were mostly derived from intergenic regions whereas siRNAs were mostly derived from transposable elements. We also annotated miRNAs and siRNAs with stage- and gender- biased expression. Our findings would facilitate to understand the gene regulation mechanism of this parasite and discover novel targets for anti-parasite drugs.
| Schistosomiasis is a chronic debilitating disease that afflicts more than 200 million individuals in the tropics and sub-tropics regions [1]. The agents of this disease, parasitic flatworms of the genus Schistosoma, have a complex developmental life cycle characterized by a distinct parasitic phase in mammalian and molluscan hosts and a free-living phase in freshwater. There are at least seven discrete developmental stages of the parasite within the definitive (lung-stage schistosomula, juvenile, adult male and female worms, and eggs) and intermediate (sporocysts) hosts as well as the aquatic, free-swimming miracidia and cercariae, with dramatically morphological changes [2]. And they are among the few platyhelminth parasites to adopt a dioecious lifestyle and possess heteromorphic sex chromosomes, which are arrayed in 7 pairs of autosomal chromosomes and one pair of sexual chromosomes (Z, W), homozygous (ZZ) for male and heterozygous (ZW) for female [3], [4]. Previous investigations on Schistosoma japonicum had revealed that a complex gene regulation pattern was deployed by this genus of parasites [5] and its haploid genome, which is about 270 Mb in size, has been recently decoded as a valuable entity for identification of small regulatory RNAs of this parasite [6].
Small non-coding RNA transcripts, approximately 18–30 nucleotides in length, are critical regulators in silencing of target genes in fungi, plants, and metazoans [7]–[9]. Three major categories of sncRNAs, siRNAs, miRNAs, and Piwi-interacting RNAs (piRNAs) have been well established and extensively studied [10]. SncRNAs exert their regulatory functions in chromatin architecture modelling, post-transcriptional repression and mRNA destabilization, mobile genetic elements suppression, and virus defence, usually through guiding the RNA-induced silencing complex (RISC) to their target genes [7], [11]–[13]. In Drosophila, sncRNAs are generated through Dicer-dependent or independent pathways [14]. Dicer-1 generates miRNAs whereas Dicer-2 creates endo-siRNAs. Recently, it was found that the Argonaute protein family, which include the ubiquitous AGO (AGO1 and AGO2) and the germline-specific Piwi (AGO3) were devoted to different small RNA-mediated regulatory pathways [15]. AGO1 functions primarily in the miRNA-dependent pathway that silences messenger RNA, whereas AGO2 has been involved in RNAi-mediated silencing directed by exogenous and endogenous siRNAs. Further study in Drosophila somatic cells revealed that there were two classes of endo-siRNAs, one was generated from TEs and involved in retrotransposon repression; the other was produced in a Dicer-2-dependent manner from distinctive genomic loci, through refolding of RNA transcripts. The function of the second class of endo-siRNAs was likely to regulate mRNA stability in somatic cells [14].
Recently, several groups have endeavored to identify and characterize sncRNAs of schistosome with conventional cloning method and the deep-sequencing technique, mainly focused on juvenile and mixed adult worms, the two relatively closed developmental stages of the parasite [16]–[21]. A repertoire of miRNA transcripts unique to S. japonicum or those conserved to other metazoan lineages was identified. Differential expression of certain miRNAs was observed between the two developmental stages of S. japonicum (hepatic schistosomula and adult worms) and S. mansoni (7d schistosomula and adult worms), suggesting that miRNAs play a distinct role in modulating development, maturation, and reproduction of the parasite [17]–[19], [21]. Importantly, miRNA genes within one cluster could be differentially expressed, which emphasized that individual transcript might be developmentally regulated by distinct mechanisms [17], [19]. Meanwhile, a set of endo-siRNAs derived mainly from transposable elements (TEs) and the natural antisense transcripts (NATs) of S. japonicum has also been defined [17], [19]. Interestingly, the distinct length and 3′ end heterogeneity of endo-siRNAs derived from both TEs and NATs were also associated with the developmental differentiation of the parasite [17].
Though the knowledge regarding sncRNA biology within the juvenile and mixed adult worms of S. japonicum is expanding, it is indispensable to systematically profile the repertoire of sncRNAs in other stages, especially the cercariae, which is the only infectious stage to penetrate its mammalian hosts; the lung-stage schistosomula, that is viewed as the most susceptible stage for intervention [22], [23]; the tissue trapped eggs, which is the critical mediator for severe pathology in schistosomiasis, and the difference between the two sexes of adult worms. In this study, the expression profile of sncRNAs in the four critical developmental stages of S. japonicum was systematically investigated. The data, for the first time, provide a broader view of small non-coding RNAs in the parasite.
The freshly released cercariae of S. japonicum were harvested from parasite-infected Oncomelania hupensis purchased from Jiangxi Institute of Parasitic Diseases, Nanchang, China. The lung-stage schistosomula (3 days post infection) were isolated from lung tissues of infected Kunming strain mice as previously described [24]. Adult worms were obtained by hepatic-portal perfusion of New Zealand White rabbits or BALB/c mice 7-weeks post infection. Male and female worms were manually separated with the aid of a light microscope. Liver tissues deposited with schistosome eggs were obtained from BALB/c mice at day 30 and 45 post infection, respectively. All procedures performed on animals within this study were conducted following animal husbandry guidelines of the Chinese Academy of Medical Sciences and with permission from the Experimental Animal Committee. All animal work have been conducted according to Chinese and international guidelines.
Total RNAs of S. japonicum at different developmental stages (cercariae, lung-stage schistosomula, adult male and female worms perfused from infected rabbits) and liver total RNAs of BALB/c mice 30d and 45d post infection were extracted using Trizol reagent (Invitrogen, CA, USA). RNA quantification and quality were evaluated by Nanodrop ND-1000 spectrophotometer (Nanodrop Technologies, Wilmington, DE) and Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA).
Construction of small RNA libraries was carried out as described previously. Briefly, RNAs between 15–30 nucleotides (nt) were excised from a 15% TBE urea polyacrylamide gel electrophoresis (PAGE). RNA samples were purified and ligated to Illumina's proprietary 5′ and 3′ adaptors, and further converted into single-stranded cDNA with Superscript II reverse transcriptase (Invitrogen, CA, USA) and Illumina's small RNA RT-Primer. The cDNA was amplified with high fidelity Phusion DNA polymerase (Finnzymes Oy, Finland) in 18 PCR cycles using Illumina's small RNA primer set. The purified PCR products were sequenced by an Illumina Genome Analyzer at the BGI (Beijing Genomics Institute, Shenzhen, China).
Raw datasets produced by deep sequencing from the libraries (cercariae, lung-stage schistosomula, adult male and female worms, and infected liver tissues) were pooled. Clean reads were obtained after removing of the low quality reads, adaptor null reads, insert null reads, 5′ adaptor contaminants, and reads with ployA tail. Adapter sequences were then trimmed from both ends of clean reads. All identical sequences were counted and merged as unique sequences, herein referred to as sequence tags. The unique reads along with associated read counts were mapped to the S. japonicum genome sequences (http://lifecenter.sgst.cn/schistosoma/cn/schdownload.do) using the program SOAP [25]. As for the liver libraries, the unique reads were mapped to the genome of mouse (http://hgdownload.cse.ucsc.edu/downloads.html#mouse) with SOAP, and those perfectly matched ones were removed prior to mapping to the S. japonicum genome.
Briefly, the perfectly matched reads were first BLAST-searched against the 78 known mature miRNAs of S. japonicum deposited in Sanger miRBase [26], [27] (Release 15) using the program Patscan [28]. The remains were then BLAST-searched against Metazoa other than S. japonicum miRNAs allowing two mismatches to identify homologs of known Metazoa miRNAs. These homologs, as well as non-conserved reads (with rRNA, tRNA, snoRNA and high repetitive reads being filtered out [29]) were considered as potential miRNAs. To avoid repeated prediction and reduce the calculation redundancy, we then searched against the genome of S. japonicum and combined candidate unique reads located in close proximity in the reference genome with less than 150 bp and we called the joint genomic fragment as a block. For each block, 150 nt upstream and 150 nt downstream sequence were extracted for secondary structure analysis. We used software Einverted of Emboss [30] to find the inverted repeats (step loops or hairpin structure), with the parameters threshold = 30, match score = 3, mismatch score = 3, gap penalty = 6, and maximum repeat length = 240 as described [31]. Each inverted repeat was extended 10 nt on each side, the secondary structure of the inverted repeat was folded using RNAfold [32] and evaluated by mirCheck using default parameters [31]. MiRNA candidates passed mirCheck were Blast-searched against Metazoa miRNAs except those of S. japonicum using the program Patscan again and labeled with conserved and non-conserved (novel) miRNAs, respectively. The novel unique reads that sequenced less than 2 times were removed. Finally, miRNA precursors that passed MirCheck were inspected manually in order to remove the false prediction. We employed IDEG6 to identify miRNAs showing statistically significant difference in relative abundance (as reflected by TPM value) between any two small RNA libraries [33]. The general Chi2×2 test was applied to determine whether one particular miRNA was significantly differentially expressed between any two samples (P value < = 0.01). Hierarchical clustering of the known S. japonicum miRNAs was constructed based on the transformed data of log10 of TPM value.
The transposable elements in the S. japonicum genome were predicted by using REPET (http://urgi.versailles.inra.fr/index.php/urgi/Tools/REPET). TE-derived siRNAs were identified as previously described [17]. Figures were constructed to reflect the relative abundance of sense and antisense of TE-derived siRNAs during the parasite development. Briefly, the expression value of each base on TE was the sum of the expression of siRNAs that mapped to the position. After a proper bin (20–50 bases) was selected based on the length of TE sequences, the average expression value was calculated for each bin, and the expression level for four stages was marked by different colors. The natural antisense transcripts of S. japonicum were annotated and NAT-derived siRNAs were confirmed as described [17]. The small RNAs that failed to pass mirCheck were aligned to S. japonicum predicted mRNA sequences of SGST (http://lifecenter.sgst.cn/schistosoma/cn/schdownload.do) using the program SOAP, and perfectly matched reads were retained. Then a Perl script was wrote to scan the predicted mRNAs, if the region continuous covered by small RNAs is longer than 100 bp, the region was deemed as a “siRNA-Cluster”.
Stem-loop qRT-PCR was performed to quantitate the sex-biased expressed miRNAs [20], [34]. Stem-loop RT primers were designed to reverse-transcribe target miRNAs into cDNAs using total RNAs isolated from male and female adult worms, respectively (from the same smaples used for Solexa sequecing). The 20 µl reaction system contained 1 µg of total RNA, 50 nM of each individual stem-loop RT primer, 1×RT buffer, 0.5 mM dNTPs (Takara), 0.01 M DTT (Invitrogen), 0.25 µl Superscript III reverse transcriptase (200 U/µl, Invitrogen, CA, USA), and 0.1 µl RNaseOUT inhibitor (40 U/µl, Invitrogen). cDNA was synthesized by incubation for 30 min at 16°C, 30 min at 42°C, 15 min at 70°C. Real-time quantification was carried out using an Applied Biosystems StepOne Plus system. PCR reactions were set up by combining 0.5 µM miRNA-specific forward primer, 0.5 µM common reverse primer, 2 µl of RT product (1∶1 dilution), 10 µl of Power SYBR Green PCR Master Mix (ABI, CA, USA), and adjusted to a final volume of 20 µl with DEPC-treated water in triplicates. For endogenous control, 1 µg of male or female total RNA was converted to cDNA with oligo(dT). The forward primer: 5′-CCTTCATGGTAGACAACGAAGCT-3′ and reverse primer: 5′-TGTAGGTTGGACGCTCTATGTCC-3′, were used to amplify the α-tubulin gene as an endogenous control. The reaction conditions were as follows: 95°C for 5 min, followed by 40 cycles of 95°C for 5 sec and 60°C for 30 sec. The quantification of each miRNA relative to α-tubulin mRNA was calculated using the equation: N = 2−ΔCt, ΔCt = CtmiRNA - Ctα-tubulin [35]. All primers used are listed in Table S1.
5′-DIG-labeled miRCURY LNA probes were ordered from Exiqon (Vedbaek, Denmark) (Http://www.exiqon.com). Northern blot analysis was performed mianly by a method described in a previous study [36]. Total RNAs were isolated from male adult female adult worms perfused from BALB/c mice 7-weeks post infection. 10 µg total RNA of each smaple was resolved by 15% denaturing (7 M urea) PAGE and were blotted by capillary transfer to neutral Nylon Membranes (Hybond-NX, GE) with 20×SSC. RNAs were further cross-linked to the membrane by EDC (1-ethyl-3-(3-dimethylaminopropyl) carbodiimide) method [37]. Blots were pre-hybridized by incubation with DIG Easy Granule (Roche) at 37°C for 3 h. And hybridization were carried out in the same buffer containing 1 nM DIG-labeled LNA probe at temperature recommended by manufacturer (RNA Tm - 30°C) overnight. Blots were washed twice in a low stringently buffer (2×SSC, 0.1% w/v SDS), and four times in a high stringently buffer (0.1×SSC, 0.1% w/v SDS), for 30 min each, both at hybridization temperature. The membrane was rinsed in washing buffer, and incubated in blocking solution at room temperature for at least 2 h (DIG washing buffer and blocking solution Set, Roche). Subsequently, blots were incubated with a 10,000-fold dilution of anti-DIG-AP (Roche) in blocking solution at room temperature for 30 min, washed 5 times for 15 min each in washing buffer. After rinsing in detection buffer for 5 min, the blots were detected using CDP-star chemiluminescent substrate for alkaline phosphatase (Roche). Blots were stripped by boiling for 1 min at 100°C in 10 mM Tris-HCl, pH 8.0, 5 mM EDTA, and 0.1% SDS and probed up to three times.
Six small RNA libraries were generated by high-throughput RNA sequencing (see Materials and Methods and Table S2). Four libraries, SjC, SjL, SjM, and SjF, were constructed from sequences that were directly derived from the cercariae, lung schistosomula, adult male, and female worms, respectively. The two remaining libraries, SjE30 and SjE45, were egg libraries derived from the hepatic tissues of BALB/c mice 30 and 45 days post-infection, respectively. The reads that aligned to the mouse genome were filtered before they were mapped to the genome of S. japonicum. In total, 65,630,916 clean reads were obtained from libraries SjC, SjL, SjM, and SjF, which were merged into 6,989,949 unique tags, thus resulted in an average redundancy as high as 89.3 (Redundancy = 100-(Total Unique Tags/Total Clean Reads ×100)). Among these unique tags, 1,593,604 (22.80%) could be aligned to the genome of S. japonicum (Table S3). The match rate was varied among different libraries, from the lowest of 20.46% (SjM) to the highest of 31.95% (SjF), this phenomenon may related with the change of ratio of different small RNAs during development and/or between sexes. The low match ratio to the genome may be caused by either genome variation of different parasite isolates or due to less sequence information of the intergenic regions where most of the miRNAs were generated. The phenomenon was also observed in similar studies by others, and several explanations have been offered [18]. Regarding the egg libraries, 15,774 and 20,800 unique tags from libraries SjE30 and SjE45, respectively, mapped to the S. japonicum genome (Table S4). These datasets contain roughly an order of magnitude more sequence than previous similar studies.
The short ncRNA transcripts were categorized according to features related to primary and secondary structure (Figure 1 and Table S5). The majority of the ncRNA transcripts were miRNAs and TE-derived endo-siRNAs, accounting for 26.75% and 44.77%, respectively, of the total sncRNA pool (Figure 1A). Only 2.21% of the miRNAs identified in our libraries were novel, indicating that most miRNAs have been recovered from the genome. Long terminal retrotransposons (LTR) and un-annotated transposons were predominant in the set of endo-siRNAs. Interestingly, the sets of miRNAs and endo-siRNAs displayed stage- or sex-related variation in expression (Figure 1B and C). The percentage of miRNA was approximately double than that of the TE-derived endo-siRNA set in the cercariae library; the amount of miRNAs and endo-siRNAs was almost equal in lung-stage schistosomula, while endo-siRNAs were dominant in the adult worms and eggs, especially in female worms and early deposited eggs (6 times more than that of miRNAs). A class of endo-siRNAs derived from unclassified transposons was dominantly expressed in the male and female parasite compared to other stages (Figure 1B). The clear pattern of preferential expression of the genes encoding the two classes of small RNAs suggests that they play stage-specific regulatory functions. Before invasion into a mammalian host, the parasite is likely to mainly utilize miRNA pathways to regulate gene expression, while endo-siRNA mediated regulation is suppressed. The high percentage of TE-derived endo-siRNAs in females and early deposited eggs suggests that siRNAs are more functional at these developmental stages. Earlier studies in D. melanogaster and mouse oocytes demonstrated that endo-siRNAs were critical elements for maintaining genomic stability through suppression of TE activity [38]–[40]. S. japonicum possesses a faster reproductive rate than flies or mice, and thousands of eggs are released by one female adult worm each day. Efficient suppression of TE activity is likely a prerequisite for continuity of parasite development and transmission, a possible explanation for why TE-derived endo-siRNAs were dominantly found in late-stage parasites.
When the sequences of the small RNAs containing classical miRNA structure were aligned to the Sanger miRBase (Release 15), 77 sequences homologous to known or well-characterized miRNAs were identified. We found 74, 71, 69, 70, 18, and 25 such sequences in libraries SjC, SjL, SjM, SjF, SjE30, and SjE45, respectively. Only one miRNA, the previously reported sja-miR-8-5p [19], was not detected in this study (Table S6). Among the set of 77 known miRNAs, approximately 20 miRNAs were conserved homologues of sequences from the planarian Schmidtea mediterranea, the genus most closely related to Schistosoma, in previous investigations [17], [19], [20], [41]–[43], indicated that phylum Platyhelminthes contains common miRNAs that carry out similar biological function. The maximum read number of a single miRNA was 1,044,358 (library SjC, sja-miR-1; Table S7), illustrating the sequencing depth of our investigation. The range of read numbers was from the single-digits to millions, highlighting the sequencing capacity of next-generation sequencing technology and suggesting that expression variation of these miRNAs does indeed exist. This observation most likely reflects functional differentiation among the miRNAs.
Apart from the known miRNAs, 193 hairpins containing 45 conserved mature miRNAs derived from 19 families were predicted in our sequence libraries. These miRNAs along with their expression level (reflected by transcripts per million, TPM) during development were shown in Table S8. Additionally, we identified 199 novel miRNAs with various expression levels and stage specificities (Table S9). In contrast with the common or evolutionarily conserved miRNAs, most novel miRNAs identified in this study possessed low read numbers, with the exceptions of sequences sja-novel-23-5p and sja-novel-48-3p, which was mainly expressed in female adult worms and cercariae, respectively.
Previous investigations of miRNA biogenesis revealed that miRNA genes are located either in intergenic regions [44] that are controlled by their own miRNA promoters and regulatory units [45], or in introns, non-protein coding genes, or exons, and thus they are likely to be regulated in concert with host genes [46]. In an earlier study, we found that many S. japonicum miRNA genes were clustered together, and that genes within the same cluster may be regulated independently [17]. In the present study, we mapped all identified miRNA sequences to the S. japonicum genome and found that miRNAs were generated from 5′ or 3′ UTRs, intragenic, and intergenic regions in the genome; however, a majority of sequences (87.2% of the total miRNAs identified) were transcripts derived from genes located in intergenic loci (Figure 2). Thus, compared to Caenorhabditis elegans, S. japonicum has evolved more sophisticated control mechanisms for regulation of miRNA expression, possibly explaining the complicated nature of the transcription profiles of individual miRNAs in various developmental stages of the parasite.
Although the relative expression level of a particular miRNA has been proposed to be represented by the number of sequence reads, other investigations have argued that neither read counts nor northern blot signal accurately reflect actual abundance or expression level [19], [47], [48]. Here, the expression levels of each unique tag in cercariae, lung-stage schistosomula, separated adult worms and eggs libraries were normalized to TPM as previously described [18], [49], [50]. Thus, the read abundance should basically reflect the expression level of the tags in the parasites. The scale of the relative miRNA abundance during the various developmental stages appears in Figure 3. Of 77 known miRNAs, 28 miRNAs exhibited high expression levels in one or more developmental stages. The expression levels of the novel miRNAs identified in this study were generally low (Table S9). However, four miRNAs were with relatively higher expression level in one particular stage, as sja-Novel-23-3p and sja-Novel-23-5p were dominantly expressed in the female parasite, while sja-Novel-48-3p and sja-Novel-74-3p were substantially expressed in cercarial stage.
Like C. elegans, schistosomal parasites need to complete a series of biological and physiological activities, including protease secretion, tail detachment, glycocalyx shedding, and tegument transformation before developing to the schistosomula stage [51], [52]. A particular gene repertoire of S. mansoni parasites was previously shown to be up-regulated during the transition from schistosomula to adult worms [22]. Here, we observed that the expression of a set of miRNAs including sja-bantam, sja-miR-1, sja-miR-124-3p, sja-miR-2a-3p, sja-miR-3492, and sja-miR-36-3p was substantially down-regulated in lung-stage schistosomula compared to cercariae (Table S6), suggesting that the target mRNAs of these miRNAs may encode proteins fulfilling important functions at this stage.
We further explored the differential expression of miRNA genes between male and female adult worms. The expression of a set of miRNAs, sja-miR-7-5p, sja-miR-61, sja-miR-219-5p, sja-miR-125a, sja-miR-125b, sja-miR-124-3p and sja-miR-1 were dominant in male worms, while sja-bantam, sja-miR-71b-5p, sja-miR-3479-5p, and sja-Novel-23-5p were predominantly found in the female parasites (Table S6 and S9). The expression of these sex-biased miRNAs was validated by stem-loop RT-PCR (Figure 4A). The expression level of sja-miR-1 was relatively high in male adult worms (1.098±0.228) and female adult worms (0.358±0.021) when compared to other miRNAs, and was not shown in Figure 4A. The correlation between the TPM values and qPCR was investigated by a method described in a previous study (R = 0.882, Spearman's Rho, p<0.0001, n = 11) [53]. However, among individual miRNAs, the qPCR results did not exactly reflect the TPM values of the maximally expressed miRNAs, probably due to the existence of extensive miRNA isomiRs, or asymmetrical amplification during library construction. We further validated the expression differences of ten sex-biased miRNAs by northern blot analysis using the total RNA extracted from adult male and female worms isolated from BALB/c mice 7-weeks post infection (Figure 4B). The differential expression pattern of these miRNAs except sja-miR-71b-5p between male and female worms was quite consistent with the TPM values of high-throughput sequencing and the qRT-PCR results. The phenomenon was also observed in a recent study which noted that several miRNAs were expressed at similar levels in protoscoleces of G1 and G7 genotypes Echinococcus granulosus, which parasitized in different hosts [54]. Thus, these data indicated that host factors may have little impact on the expression profile or level of sncRNAs.
Although the function of these miRNAs remains to be elucidated, the significant differential expression between male and female adult worms indicated that they may participate in regulation of sexual differentiation and maintenance, pairing and reproduction of the parasite. Moreover, miRNAs may cooperate with other small RNAs (such as endo-siRNAs) and transcription factors to form a comprehensive network to regulate growth, development, differentiation, and reproduction for adaptation to a variety of environments [19]. Further studies on these miRNAs may contribute to better understanding of the developmental mechanism of sexual dimorphism in this parasite.
Recent observations of endo-siRNAs in D. melanogaster, mice, and schistosome have added more complexity to our knowledge of small RNA-mediated regulatory pathways [14], [17], [38]–[40], [55]–[58]. So far, endo-siRNAs have been found to be mainly derived from TEs, complementary annealed NATs, and the long “fold-back” transcripts known as hairpin RNAs [59]. We previously found that the TE-derived siRNAs in S. japonicum were more predominant than other endo-siRNAs, including NAT-derived siRNAs [17]. Here, we systematically analyzed the expression levels of sense and antisense endo-siRNAs that derived from various TEs in cercariae, lung-stages schistosomula, male and female adult worms (Table S10–14). The read numbers of endo-siRNAs in egg libraries were much lower than the other libraries, leading us to exclude the egg libraries from further analysis.
We observed that LINE, TIR, and LTR transposon classes were the main sources of endo-siRNAs, while the endo-siRNAs derived from other TEs were much less abundant (Figure 5). Further, endo-siRNAs mapped to the top (sense siRNA) and bottom (antisense siRNA) strands of LTR and non-LTR TEs. The expression patterns of LTR-derived sense and antisense siRNAs were relatively symmetrical, though there were obvious stage and sex specificities in expression loci (Figure 5A, B, and C). Reads mapped to the S. japonicum LTR retrotransposon SjCHGCS11 [6] were annotated as SACI-7_2p in our analysis (Figure 5A). Both sense and antisense siRNAs were concentrated in the coding region of reverse transcriptase in a manner similar to that observed in D. melanogaster somatic cells [38].
Sjpido, SjR1, and SjR2 are three classes of non-LTR retrotransposons that make up 5% of the S. japonicum genome; siRNAs generated from these elements mainly mapped to certain sequence regions (Figure 5B), contrary to our observations of LTR retrotransposons. The expression levels of siRNAs derived from SjR1 were much higher in cercariae than in male and female adult worms, indicating that these siRNAs are more functional in the earlier developmental stage (Figure 5B). Sj-alpha-1 derived siRNAs were predominantly generated from the antisense strand, while Sj-alpha-2 derived siRNAs were generated from the sense strand; however, both types of siRNAs had low expression levels (Figure 5C). In cercariae and lung-stage schistosomula, the TIR (Sj_Blaster_Recon_7337_MAP_14 annotated as SmTRC1_1p in the genome) derived siRNAs were highly expressed, while the MITE (Sj_Blaster_Grouper_1934_MAP_4) derived siRNAs were mainly expressed in the adult worms, and predominantly corresponded to the antisense strand (Figure 5D). Thus, the TE-derived endo-siRNAs of S. japonicum were more diverse than those found in D. melanogaster. Although the origin of the antisense siRNAs is not known (cis- or trans-transcription), their abundance suggests that they are stable and likely participate in regulatory pathways.
Previous studies of mouse oocytes revealed that antisense transcripts from pseudogenes formed double-strand RNAs with their functional counterparts, the sources of the endo-siRNAs, and the sense siRNAs were predominant in the endo-siRNA [55]. It has been proposed that the “passenger strand” of an siRNA is unstable due to the thermodynamic asymmetry of the two strands [60]. However, this hypothesis cannot explain our identification of many reads corresponding to the antisense siRNAs; in some cases, only the antisense strands were identified. Further dissection of the function of the two endo-siRNA classes would be an essential step toward understanding the network of gene regulation during the parasite development.
NAT-derived siRNAs are a second source of endo-siRNAs; these endo-siRNAs are further classified as cis-NAT- or trans-NAT-derived endo-RNAs [56], [61], [62]. Cis-NAT-derived endo-siRNAs are generated from transcripts from the same gene locu, while trans-NAT-derived endo-siRNAs come from NAT transcripts of distinct loci. We detected potential NAT pairs by aligning the predicted mRNA sequences to each other. Only one cis-NAT pair and 1772 trans-NAT pairs were identified in silico using data from SGST. Our sequencing results were remarkably similar to the in silico prediction; one cis-NAT pair and 225 trans-NAT pairs were detected (Table S15), indicating that bi-directional transcription was much less prevalent in schistosomal parasites and transcripts from duplicated genes are more common. Thus, trans-NAT-derived endo-siRNAs are likely the main sources of NAT-derived siRNA in S. japonicum, a scenario that differs from other organisms [63]. However, we cannot rule out the possibility that most of the cis-NAT pairs may be undetectable given the lack of information about the non-protein-coding regions of the S. japonicum genome. The identification of long non-coding RNAs in S. japonicum is still underway, and may provide an important source of NAT-derived siRNAs [64].
A previous study of D. melanogaster somatic cells demonstrated that endo-siRNAs mapped to protein-coding mRNAs rather than to transcripts of transposons that regulate mRNA expression [38]. Here, we also mapped the endo-siRNAs to the predicted mRNA sequences of S. japonicum, and found that nearly half of the siRNA-related transcripts clustered within predicted mRNAs. These mRNAs mainly encoded proteins from four categories: 1) proteins similar to pol polyprotein and endonuclease-reverse transcriptase; 2) cytoskeletal proteins such as myosin, actin, and tropomyosin; 3) enzymes or transporters such as COX1, COX2B, superoxide dismutase 1, glyceraldehyde 3-phosphate dehydrogenase, lactate dehydrogenase A, ATP-dependent RNA helicase, cation-transporting ATPase, H+-transporting ATPase, and cathepsin B and L; 4) stress proteins including heat shock protein, cold shock protein, and stress-induced phosphoprotein 1 (Table S16). We were unable to distinguish whether siRNAs clustered in pol polyprotein and endonuclease-reverse transcriptase transcripts were derived from retrotransposons or NATs. We speculated that some of the siRNAs in the other three categories were derived from trans-NATs formed by transcripts of pseudogenes and their parental genes, as suggested recently [65]; for example, the pseudogenes of hsp70 and cathepsin B exist in schistosome genomes [66], [67]. Furthermore, the pseudogenes of actin, COX, GAPDH, FABP, and histone are common in mammalian genomes. Pseudogene-derived endo-siRNAs were previously detected in mouse oocytes, with two transcripts, Hsp90ab1 (heat shock protein 90 kDa alpha, class B member 1) and Dynll1 (dynein, light chain), possessing features similar to our findings [55]. Thus, unlike the silencing of selfish genetic elements by TE-related siRNAs, trans-NAT-derived endo-siRNAs mainly regulate the expression of mRNAs coding for a diverse set of proteins.
Our current study generated comprehensive profiles of endogenous small RNAs (miRNAs and endo-siRNAs) during the four crucial developmental stages of S. japonicum. Reverse expression patterns of miRNAs and endo-siRNAs during the parasite development and differentiation process were observed. Two classes of endo-siRNAs derived from TEs and trans-NATs were identified, and the LTR retrotransposon derived siRNAs were more abundant than siRNAs from non-LTR TEs. There are likely two layers of regulatory function employed by the parasite; the antisense siRNAs directly affect the stability of mRNA transcripts, while the sense siRNAs may function indirectly by affecting the amount of antisense transcripts. Thus, the small RNA-mediated network in schistosomal parasites is more complex than networks reported in other organisms.
|
10.1371/journal.pcbi.1006624 | Reservoir computing model of prefrontal cortex creates novel combinations of previous navigation sequences from hippocampal place-cell replay with spatial reward propagation | As rats learn to search for multiple sources of food or water in a complex environment, they generate increasingly efficient trajectories between reward sites. Such spatial navigation capacity involves the replay of hippocampal place-cells during awake states, generating small sequences of spatially related place-cell activity that we call “snippets”. These snippets occur primarily during sharp-wave-ripples (SWRs). Here we focus on the role of such replay events, as the animal is learning a traveling salesperson task (TSP) across multiple trials. We hypothesize that snippet replay generates synthetic data that can substantially expand and restructure the experience available and make learning more optimal. We developed a model of snippet generation that is modulated by reward, propagated in the forward and reverse directions. This implements a form of spatial credit assignment for reinforcement learning. We use a biologically motivated computational framework known as ‘reservoir computing’ to model prefrontal cortex (PFC) in sequence learning, in which large pools of prewired neural elements process information dynamically through reverberations. This PFC model consolidates snippets into larger spatial sequences that may be later recalled by subsets of the original sequences. Our simulation experiments provide neurophysiological explanations for two pertinent observations related to navigation. Reward modulation allows the system to reject non-optimal segments of experienced trajectories, and reverse replay allows the system to “learn” trajectories that it has not physically experienced, both of which significantly contribute to the TSP behavior.
| As rats search for multiple sources of food in a complex environment, they generate increasingly efficient trajectories between reward sites, across multiple trials. This spatial navigation optimization behavior can be measured in the laboratory using a traveling salesperson task (TSP). This likely involves the coordinated replay of place-cell “snippets” between successive trials. We hypothesize that “snippets” can be used by the prefrontal cortex (PFC) to implement a form of reward-modulated reinforcement learning. Our simulation experiments provide neurophysiological explanations for two pertinent observations related to navigation. Reward modulation allows the system to reject non-optimal segments of experienced trajectories, and reverse replay allows the system to “learn” trajectories that it has not physically experienced, both of which significantly contribute to the TSP behavior.
| Spatial navigation in the rat involves the replay of place-cell sub-sequences, that we refer to as snippets, during awake and sleep states in the hippocampus during sharp-wave-ripples (SWR) [1–4]. In the awake state, replay has been observed to take place in forward and reverse direction [2, 5–8], with respect to the physical order of the initial displacement of the animal. Both forward and reverse replay are influenced by task contingencies and reward [6, 9, 10]. Reverse replay is observed to originate from rewarded locations [11], with a greater frequency of replay for locations with greater reward, which could allow a propagation of value backwards from the rewarded location [9]. An interesting example of the impact of reward on forward replay is seen in the experiments of Gupta & van der Meer [6] where rats ran the left or right (or both) sides of a dual maze. Replay occurred with equal proportions for the same and opposite side of the rat’s current location. Same side forward replay tended to be prospective, and project forward from the current location, as observed by Pfeiffer and Foster [12]. Interestingly, opposite-side forward replay preferentially occurred retrospectively, as forward sweeps to reward locations starting from remote locations. This suggests that more diverse forward replay including forward sweeps from remote locations (as observed by Gupta & van der Meer [6]) will be observed as a function of specific task characteristics and requirements. Liu and Sibille [8] have recently shown the predictive nature of such forward sweeps, using essentially statistical data analysis techniques. Our work extends this recent work in proposing an actual mechanism and neuronal model that could support it.
Thus, while it has been observed that in both 1D [5] and 2D [12], the bulk of awake replay events are prospective, and depict future paths to upcoming goals, in more complex tasks, forward retrospective sweeps from remote locations can be observed. In the current research with multiple goal locations to be remembered (no foraging), and optimization of the paths (rather than just their memorization), we argue (and the model predicts), that forward sweeps from remote locations will predominantly carry information and be used to accomplish the task.
We focus on the role of replay during the awake state, as the animal generates increasingly efficient trajectories between reward sites, across multiple trials. This trend toward near-optimal solutions is reminiscent of the classic Traveling Salesperson Problem (TSP) [13]. The TSP problem involves finding the shortest path that visits a set of “cities” on a 2D map. It is a computationally complex problem, and is one of the most intensively studied problems in optimization [14, 15]. While it is clear that rats do not solve the TSP in the mathematical sense, they remarkably display a robust tendency towards such optimization [13]. It appears likely that such spatial navigation optimization involves planning and hence awake replay but the underlying neurophysiological mechanisms remain to be understood.
One obvious advantage of replay would be to provide additional internally generated training examples to otherwise slow reinforcement learning (RL) systems. Traditional RL methods are usually inefficient, as they use each data sample once, to incrementally improve the solution, and then discard the sample. In a real-time learning situation, where the rat is optimizing in less than ten trials, this approach is unlikely to succeed. Our model proposes to add replay to reinforcement learning to overcome this problem and improve efficiency [16]. This approach has been previously exploited with good results [17]. We will go beyond this by prioritizing replay based on a spatial gradient of reward proximity that is built up during replay. Our first hypothesis is that reward-modulated replay in hippocampus implements a simple and efficient form of reinforcement learning [18], which allows recurrent dynamics in prefrontal cortex (PFC) to consolidate snippet representations into novel efficient sequences, by rejecting sequences that are less robustly coded in the input.
An example of the behavior in question is illustrated in Fig 1. Panel A illustrates the optimal path linking the five feeders (ABCDE) in red. Panels B-D illustrate navigation trajectories that contain sub-sequences of the optimal path (in red), as well as non-optimal sub-sequences (in blue). In the framework of reward modulated replay, snippets from the efficient sub-sequences in panels B-D will be replayed more frequently, and will lead the system to autonomously generate the optimal sequence as illustrated in panel A. We thus require a sequence learning system that can re-assemble the target sequences from these replayed snippets. For this, we chose a biologically inspired recurrent network model of prefrontal cortex [19, 20] that we predict will be able to integrate snippets from examples of non-optimal trajectories and to synthesize an optimal path.
Recurrent networks have excellent inherent sensitivity to serial and temporal structure, which make them well adapted for sequence learning [21, 22]. Interestingly, primate cortex is characterized by a vast majority of cortico-cortical connections being local and recurrent [23], and thus cortex is a highly recurrent network [24]. We thus model frontal cortex as a recurrent network. Interestingly, the computational complexity of credit assignment to recurrent connections is high, because it involves keeping track of the role of each connection over successive time steps as the network evolves through its temporal dynamics [21]. One solution is to unwind the recurrent network into a series of feedforward layers where each layer represents the network activation at the next time step. This is efficient [22], but introduces an arbitrary cut off of the recurrent dynamics. In order to allow the recurrent network to maintain complete dynamics, Dominey et al. [25] chose to keep the recurrent connections fixed, with a random distribution of positive and negative connections that ensured a rich network dynamics that represented the influence of new inputs, and the fading effects of previous inputs. The resulting representation of the spatiotemporal structure of the input can then be associated with the desired output function by the modification of simple feedforward connections from the recurrent network to the output neurons. Dominey and colleagues initially proposed that the prefrontal cortex corresponds to the recurrent network, and the striatum, with its dopamine-modifiable corticostriatal connections as the output layer [25]. This was in fact the first characterization of reservoir computing, which was subsequently co-discovered by Jaeger [26, 27], and Maass & Natschlager [28]. It is now well established that frontal cortex can be characterized as a recurrent reservoir model, via demonstrations that the high dimensional representations inherent to these recurrent networks is required for higher cognitive function, and is omnipresent in frontal cortex [19, 29, 30]. The use of the reservoir structure is indeed an originality of our model, and an alternative to classical plastic recurrent networks such as those used to model attractor network dynamics in hippocampus CA3 [31].
We test the hypothesis that the structure of snippet replay from the hippocampus will provide the PFC with constraints that can be integrated in order to contribute to solving the TSP problem. Two principal physical and neurophysiological properties of navigation and replay are exploited by the model and contribute to the system’s ability to converge onto an acceptable solution to the TSP. First, during navigation between baited food wells in the TSP task, non-optimal trajectories by definition cover more distance between rewards than near-optimal ones. Second, during the replay of recently activated places cells, the trajectories are encoded in forward and reverse directions [5, 11]. Exploiting these observations, we test the hypotheses that:
In testing these hypotheses, we will illustrate how the system can meet the following challenges:
The model developed in this research provides a possible explanation of mechanisms that allow PFC and hippocampus to interact to perform path optimization. This implies functional connectivity between these two structures. In a recent review of hippocampal–prefrontal interactions in memory-guided behavior Shin and Jadhav [32] outlined a diverse set of direct and indirect connections that allow bi-directional interaction between these structures. Principal direct connections to PFC originate in the ventral and intermediate CA1 regions of the hippocampus [33, 34]. Connections between hippocampus and PFC pass via the medial temporal lobe (subiculum, entorhinal cortex, peri- and post-rhinal cortex) [35], and the nucleus reuniens [36]. Indeed memory replay is observed to be coordinated across hippocampus and multiple cortical areas including V1 [37]. There are direct connections from ventral CA1/subiculum to the rodent medial frontal cortex[38] and from the mFC to dorsal CA1 [39], The connections through the RE nucleus though may be of primary importance for HC-mFC communication [40]. These studies allow us to consider that there are direct and indirect anatomical pathways that justify the modeling of bi-directional interaction between PFC and hippocampus [41].
It is important to note that the model we describe should not be considered to be fully autonomous in driving the behavior of the animal, because it relies on prior experience from which to construct new behavior. This experience is assumed to be generated by visual and olfactive processes that contribute to locally guided behavior.
Experiments are performed on navigation trajectories (observed from rat behavior, or generated automatically) that represent the recent experience from the simulated rat. Snippets are extracted from this experience, and used to train the output connections of the PFC reservoir. This requires the specification of a model of place-cell activation in order to generate snippets. Based on this training, the sequence generation performance is evaluated to test the hypotheses specified. The evaluation requires a method for comparing sequences generated with expected sequences that is based on the Fréchet distance.
A trajectory is a sequence of N contiguous two-dimensional coordinates sampled from time t1 to time tN noted L(t1→tN) that corresponds to the rat’s traversal of the baited feeders. The spatial resolution of trajectories are depicted at 20 points/m along the trajectory. Experiments were performed using navigation trajectories, including those displayed in Fig 1, based on data recorded from rats as they ran the TSP task [13] in a circular arena having a radius of 151cm. Twenty-one fixed feeders are distributed according to a spiral shape. In a typical configuration, five feeders are baited with a food pellet. For a given configuration, the rat runs several trials which are initially random and inefficient, and become increasingly efficient over successive trials, characterizing the TSP behavior [13]. Rat data that characterizes the TSP behavior is detailed in S1 Text, section Rat navigation data. The principle concept is that TSP behavior can be characterized as illustrated in Fig 1, where a system that is exposed to trajectories that contain elements of the efficient path can extract and concatenate these sub-sequences in order to generate the efficient trajectory.
The modeled rat navigates in a closed space of 2x2 meters where it can move freely in all direction within a limited range (± 110° left and right of straight ahead), and encodes locations using hippocampus place-cell activity. A given location s = (x,y) is associated with a place-cell activation pattern by a set of 2D Gaussian place-fields:
fk(s)=e−‖s−ck‖2wk
(1)
Where:
Parameter wk is a manner of defining the variance of the 2D Gaussian surface with a distance to center related parameter rk. We model a uniform grid of 16x16 Gaussian place-fields of equal size (mimicking dorsal hippocampus). In Fig 2 the spatial position and extent of the place fields of several place-cells is represented in panel A by red circles. The degree of red transparency represents the mean firing rate.
A mean firing rate close to one will result in a bright circle if the location s is close to the place-field center ck of the place-cell k. For a more distant place-field center cl of place-cell l, the mean firing rate will be less important and the red circle representing this mean firing rate will be dimmer.
Thus, at each time step the place-cell coding that corresponds to a particular point in a trajectory is defined as the projection of this L(tn) point through K radial basis functions (i.e. Gaussian place-fields spatial response)
Xin(tn)={fk(L(tn))}k∈1…K
(2)
Each coordinate of the input vector Xin(tn) represents the mean firing rate of hippocampus place-cells and its value lies between 0 and 1. Fig 2 represents in panel B the ABCED trajectory L(t1→tN) and the corresponding place-cell mean firing rate raster Xin(t1→tN) is depicted in panel C.
The hippocampus replay observed during SWR complexes in the active rest phase (between two trials in a given configuration of baited food wells) is modeled by generating condensed (time compressed) sub-sequences of place-cell activation patterns (snippets) that are then replayed at random so as to constitute a training set. The sampling distribution for drawing a random place-cell activation pattern might be uniform or modulated by new or rewarding experience as described in [1]. Ambrose and Pfeiffer [9] demonstrated that during SWR sequences place-cell activation occur in reverse order at the end of a run. We define a random replay generative model that learns to preferentially generate snippets associated to a reward by using reverse replay. For a trajectory encountered in forward direction, reverse replay allows the model to anticipate the reward by propagating the reward information in the backwards direction. Once learnt the model is able to generate snippets in forward and reverse order, hence representing parts of a trajectory in reverse or forward direction. This innovative method for spatial propagation of reward during replay yields a computationally simple form of reinforcement learning.
We define a snippet as the concatenation of a pattern of successive place-cell activations from a previously traversed trajectory:
S(n;s)=Xin(tn→tn+s)
(3)
Where: s is the number of place-cell activations (or the snippet length), and n is the offset in the trajectory. Replay occurs during SWRs at ~150-200Hz with a duration ~50–120 ms [2], so snippet length s in our experiments is typically 10 and varies from 5–20. We define a replay episode as the period between trials in the TSP experiment (on the order of 2–5 minutes) during which replay occurs. The duration of a replay episode is constrained by a time budget T, defined in simulation update cycles. Place cell activations in the simulated replay occur on each time step, with each time step corresponding to 5ms, or a 200Hz update rate. A replay episode E is a set of snippets of length s:
E(s)={S(n;s)}
(4)
such that sum of the durations of snippets replayed in E is constrained to be ≤ T. In a typical experiment described below T = 10000 and s = 10, which corresponds to 1000 snippets of length 10. In order to respect ecological orders of magnitude, we consider that during 2–5 minutes of intertrial delay in the TSP task, SWRs occur at ~1Hz, corresponding to ~120–300 SWRs. In a given SWR it is likely that across the dorsal hippocampus, 100s - 1000s of places cells will fire, corresponding to an order of 101–102 snippets per SWR. This parallel replay of multiple independent snippets within an SWR is hypothesized but has not yet been experimentally observed. Over 2–5 minutes, this corresponds to a lower bound of 1.2x103—and an upper bound of 3x104 snippets during the intertrial period. We conservatively model this at 1x103. In a given episode snippet length s is fixed. Individual snippets are spatially coherent, while successive snippets are not, and can start from random locations along previously experienced trajectories.
In Fig 2, Panel B represents a particular trajectory through feeders A, B, C, E and D. The depicted snippet is a sub-sequence of 5 contiguous locations belonging to the ABCED sequence. The B and E feeders are baited and marked as rewarding (R1 and R2). Panel B shows the spatial extent of a given snippet chosen in sequence ABCED and panel C shows the place-cell activation pattern of the ABCED trajectory and the corresponding snippet location in the raster.
We model the prefrontal cortex as a recurrent reservoir network. Reservoir computing refers to a class of recurrent network models with fixed recurrent connections. The reservoir units are driven by external inputs and the network dynamics provides a high dimensional representation of the inputs from which the desired outputs can then be read out by a trained linear combination of the reservoir unit activities. The principle has been co-developed in distinct contexts as the temporal recurrent network [20], the liquid state machine [28], and the echo state network [26]. The version that we use to model the frontal cortex employs leaky integrator neurons in the recurrent network. This model of PFC is particularly appropriate because the recurrent network generates dynamic state trajectories that will allow overlapping snippets to have overlapping state trajectories. This property will favor consolidation of a whole sequence from its snippet parts. At each time-step, the network is updated according to the following schema:
The hippocampus place-cells project into the reservoir through feed-forward synaptic connections noted Wffwd. The projection operation is a simple matrix-vector product. Hence, the input projection through feed-forward synaptic connections is defined by:
Uffwd(tn)=Wffwd*Xin(tn)
(5)
Where:
Synaptic weights are randomly selected at the beginning of the simulation. Practically speaking [42], sampling U[−1,1] a uniform distribution is sufficient. A positive synaptic weight in a connectivity matrix models an excitatory connection and a negative weight models an inhibitory connection between two neurons (that could be implemented via an intervening inhibitory interneuron). Let N be the number of neurons in the reservoir. Reservoir's neurons are driven by both sensory position inputs Xin(tn) and, importantly by the recurrent connections that project an image of the previous reservoir state back into the reservoir. The recurrent projection is defined as:
Urec(tn)=Wrec*Xres(tn−1)
(6)
Where:
Synaptic weights are drawn from a U[−1,1] uniform distribution, scaled by a S(N;K)=K1N factor. The same sign convention as in Eq (5) applies for the recurrent connectivity matrix.
Self-connections (i.e. wreci,i with i∈1…N) are forced to zero. Wrec is also fixed and its values do not depend on time. The contributions of afferent neurons to the reservoir’s neurons is summarized by
Ures(tn)=Uffwd(tn)+Urec(tn)
(7)
The membrane potential of the reservoir’s neurons Pres then is computed by solving the following ordinary derivative equation (ODE):
τ∂Pres∂t=−Pres(tn−1)+Ures(tn)
(8)
Where:
In this article, we will consider a contiguous assembly of neurons that share the same time constant. The inverse of the time constant is called the leak rate and is noted h. By choosing the Euler’s forward method for solving Eq (8), the membrane potential is computed recursively by the equation:
Pres(tn)=h*Ures(tn)+(1−h)*Pres(tn−1)
(9)
This is a convex combination between instantaneous contributions of afferents neurons Ures(tn) and the previous value Pres(tn−1) of the membrane potential. The current membrane potential state carries information about the previous activation values of the reservoir, provided by the recurrent weights. The influence of the history is partially controlled by the leak rate. A high leak rate will result in a responsive reservoir with a very limited temporal memory. A low leak rate will result in a slowly varying network whose activation values depend more on the global temporal structure of the input sequence.
Finally, the mean firing rate of a reservoir’s neuron is given by:
Xres(tn)=σres(Pres(tn);Θres)
(10)
Where:
We chose a σres≡tanh hyperbolic tangent activation function with a zero bias for Θres. Negative firing rate values represent the inhibitory/excitatory connection type in conjunction with the sign of the synaptic weight. Only the product of the mean firing rate of the afferent neuron by its associated synaptic weight is viewed by the leaky integrator neuron. See S1 Text High dimensional processing in the reservoir for more details on interpreting activity in the reservoir.
Once the model is trained, we need to evaluate its performance and the trajectories it can generate. The model is primed with the first p steps of the place-cell activation sequence the model is supposed to produce. This sequence is called the target sequence. Then the model’s ability to generate a place-cell activation sequence is evaluated by injecting the output prediction of the next place-cell activation pattern as the input at the next step. In this iterative procedure, the system should autonomously reproduce the trained sequence pattern of place-cell activations.
Predicted place-cell activation values might be noisy, and the reinjection of even small amounts of noise in this autonomous generation procedure can lead to divergence. We thus employ a procedure that determines the location coded by the place-cell activation vector output, and reconstructs a proper place-cell activation vector coding this location. We call this denoising procedure the spatial filter as referred to in Fig 3.
We model the rat action as ‘reaching the most probable nearby location’. Since only the prediction of the next place-cell activation pattern η is available, we need to estimate the most probable point s*(tn+1) = (x*(tn+1),y*(tn+1)). From a Bayesian point of view, we need to determine the most probable next location s(tn+1), given the current location s(tn) and the predicted place-cell activation pattern η(tn). We can state our problem as:
s*(tn+1)=argmaxs(tn+1)P(s(tn+1)|η(tn),s(tn))+u
(14)
Where:
u is useful at least in degenerate cases when a zero place-cell activation prediction generates an invalid location coding. It is also used for biasing the generation procedure and to explore other branches of the possible trajectories the model can generate as described in section Evaluating Behavior with Random walk.
The system is then moved to this new location s* and a new noise/interference free place-cell activation pattern is generated by the place-field model. We refer to this place-cell prediction/de-noising method as the spatial filter, which emulates a sensory-motor loop for the navigating rat in this study. Fig 3 depicts this sensory motor loop.
Once the model has been trained, it is then primed with place-cell activation inputs corresponding to the first few steps of the trajectory to be generated. The readout from the PFC reservoir generates the next place-cell activation pattern in the trajectory, which is then reinjected into the reservoir via the spatial filter, in a closed loop process. This loop evaluation procedure is called autonomous generation. In order to evaluate the model in a particular experimental condition, several instances of the same model are evaluated multiple times in a random walk procedure. The batch of generated trajectories (typically 1000) are accumulated in a stencil buffer which acts as a two dimensional histogram showing the most frequently generated trajectories. The arena is drawn with its feeders and a vector field is computed from trajectories in order to show the main direction of these trajectories. Trajectories are superimposed and summed, resulting in a two-dimensional histogram representing the space occupied by trajectories. Fig 4 shows an example of random walk trajectories, illustrating the model’s ability to autonomously generate a long and complex sequence when learning without snippets. Fig 4B illustrates a 2D histogram formed by superimposing trajectories autonomously generated by 1000 reservoirs evaluated ten times each with noise, in order to validate the robustness of the behavior.
In cases where small errors in the readout are reinjected as input, they can be amplified, causing the trajectory to diverge. It is possible to overcome this difficulty by providing as input the expected position at each time step instead of the predicted position. The error/distance measurement can still be made, and will quantify the diverging prediction, while allowing the trajectory generation to continue. This method is called non-autonomous generation and it evaluates only the ability of a model to predict the next place-cell activation pattern, given an input sequence of place-cell activations.
The joint PFC-HIPP model can be evaluated by comparing an expected place-cell firing pattern with its prediction by the readout layer. At each time step, an error metric is computed and then averaged over the duration of the expected neurons firing rate sequence. The simplest measure is the mean square error. This is the error that the learning rule described in Eq (12) minimizes.
Although the model output is place-cell coding, what is of interest is the corresponding spatial trajectory. A useful measurement in the context of comparing spatial trajectories is the discrete Fréchet distance. It is a measure of similarity between two curves that takes into account the location and ordering of the points along the curve. We use the discrete Fréchet distance applied to polygonal curves as initially described in Eiter and Mannila [45]. In [46] the Discrete Fréchet distance F between two curves A and B is defined by:
F(A,B)=minα,βmaxt∈[1,m+n][d(A(α(t)),B(β(t)))]
(15)
Where d(.,.) is the Euclidean distance, m is the number of steps of the curve A, n is the number of steps of the curve B, and α,β are reparametrizations of the curves A and B. Parameterization of this measure is described in more detail in S1 Text Frechet distance parameters.
For robustness purposes, results are based on a population of neural networks rather than a single instance. The population size is usually 1000 for evaluating a condition and the metrics described above are aggregated by computing their mean μ(.) and standard-deviation σ(.). For convenience, we define a custom score function associated to a batch of coherent measurements as:
score(X)=μ(X)+σ(X)
(16)
Results having a low mean and standard deviation will be reported as low score whilst other possible configurations will result in a higher score. We chose this method rather than Z-score, which penalizes low standard deviations. We first established that the model displays standard sequence learning capabilities (e.g. illustrated in Fig 4) and studied parameter sensitivity (see S1 Text Basic Sequence learning and parameter search), and then addressed consolidation from replay.
The model is able to learn and generate navigation sequences from place-cell activation patterns. The important questions is whether a sequence can be learned by the same model when it is trained on randomly presented snippets, instead of the continuous sequence.
In this experiment, no reward is used, and thus each snippet has equal chance of being replayed. The only free parameter is the snippet size. In order to analyze the reservoir response, we collect the state-trajectories of reservoir neurons when exposed to snippets. Recall that the internal state of the reservoir is driven by the external inputs, and by recurrent internal dynamics, thus the reservoir adopts a dynamical state-trajectory when presented with an input sequence. Such a trajectory is visualized in Fig 5D. This is a 2D (low dimensional) visualization, via PCA, of the high dimensional state transitions realized by the 1000 neurons reservoir as the input sequence corresponding to ABCDE is presented. Panels A-C illustrate the trajectories that the reservoir state traverses as it is exposed to an increasing number of randomly selected snippets generated for the same ABCDE sequence. We observe that as snippets are presented, the corresponding reservoir state-trajectories start roughly from the same point because of the random initial state of the reservoir before each snippet is replayed. Then the trajectories evolve and partially overlap with the state-trajectory produced by the complete sequence. In other words, snippets quickly drive the reservoir state from an initial random activation (corresponding to the grey area at the center of each panel) onto their corresponding locations in the reservoir activation state-trajectory of the complete sequence. Replaying snippets at random allows the reservoir to reconstruct the original intact reservoir state trajectory because the reservoir states overlap when snippet trajectories overlap.
Thus, we see that the state trajectories traversed by driving the reservoir with snippets overlaps with those from the original intact sequence. As illustrated in 5A and 5B, 100 to 1000 snippets are required for allowing the consolidation to occur in the readout layer with the learning rule described in Eq (12). A smaller learning error is achieved with 10000 snippets because the reservoir states that correspond to the whole sequence depicted in panel D are observed more often and the error gradient corrected more often by modifying the readout synaptic weights. See further details of sequence learning by snippet replay in S1 Text Sequence complexity effects on consolidation.
Here we examine how using reward proximity to modulate snippet replay probability distributions (as described in the hippocampal replay description) allows the rejection of longer, inefficient paths between rewarded targets. In this experiment, 1000 copies of the model are run 10 times. Each is exposed to the reward modulated replay of two sequences ABC and ABD having a common prefix AB as illustrated in Fig 6. The model is trained on snippets replayed from trajectories ABC and ABD. The random replay is not uniform and takes into account the reward associated with a baited feeder when food was consumed, as describe in the Hippocampal Replay section above. Effectively, snippets close to a reward have more chance to be replayed and thus to be consolidated into a trajectory.
Panel A in Fig 6 illustrates the distribution of snippets selected from the two sequences, ABC in pink and ABD in blue. At the crucial point of choice at location B, the distribution of snippets for sequence ABC largely outnumbers those for sequence ABD. This is due to the propagation of rewards respectively from points C and D. Per design, rewards propagated from a more proximal location will have a greater influence on snippet generation. Panel C shows the 2D histogram of autonomously generated sequences when the model is primed with the initial sequence prefix starting at point A. We observe a complete preference for the shorter sequence ABC illustrated in panel E.
The snippet generation model described above takes into account the location of rewards, and the magnitude of rewards. Panel B illustrates the distribution of snippets allocated to paths ABC and ABD when a 10x stronger reward is presented at location D. This strong reward dominates the snippet generation and produces a distribution that strongly favors the trajectory towards location D, despite its farther distance. Panel F illustrates the error measures for model reconstruction of the two sequences and confirms this observation. This suggests an interesting interaction between distance and reward magnitude. For both conditions, distances to the expected sequence have been measured for every trajectory generated (10 000 for ABC and 10 000 for ABD). Then a Kruskall Wallis test confirms (p-value ~ = 0) for both cases that trajectories generated autonomously are significantly more accurate for the expected trajectories (i.e. ABC when rewards are equal and ABD when reward at D is x10).
Based on the previously demonstrated dynamic properties, we determined that when rewards of equal magnitudes are used, the model would favor shorter trajectories between rewards. We now test the model’s ability to exploit this capability, in order to generate a novel and efficient trajectory from trajectories that contain sub-paths of the efficient trajectory. That is, we determine whether the model can assemble the efficient sub-sequences together, and reject the longer inefficient sub-sequences in order to generate a globally efficient trajectory. Fig 1 (Panel A) illustrates the desired trajectory that should be created without direct experience, after experience with the three trajectories in panels B-D that each contain part of the optimal trajectory (red), which will be used to train the model.
The reward-biased replay is based on the following trajectories: (1) ABCED that contains the ABC part of the ABCDE target sequence, (2) EBCDA that contains the BCD part of the ABCDE target sequence, and (3) BACDE that contains the CDE part of the ABCDE target sequence. Fig 7A illustrates how the hippocampal replay model generates distributions of snippets that significantly favor the representation of the efficient sub-sequences of each of the three training sequences. This is revealed as the three successive peaks of snippet distributions on the time histogram for the blue (ABCED) sequence, favoring its initial part ABC, the yellow (EBCDA) sequence, favoring its middle part BCD, and the pink (BACDE) sequence, favoring its final part CDE. When observing each of the three color-coded snippet distributions corresponding to each of the three sequences we see that each sequence is favored (with high replay density) precisely where it is most efficient. Thus, based on this distribution of snippets that is biased towards the efficient sub-sequences, the reservoir should be able to extract the efficient sequence.
Reservoir learning is illustrated in Fig 7B, which displays the autonomously generated sequences for 1000 instances of the model executed 10 times each. Training is based on 1000 snippets of length 10 selected from the distribution illustrated in Fig 7A. The spatial histogram reveals that the model is able to extract and concatenate the efficient sub-sequences to create the optimal path, though it was never seen in its entirety in the input. Panel C illustrate the significant differences in performance between the favored efficient sequence vs. the three that contain non-efficient sub-sequences. A Kruskal-Wallis test confirms these significant differences in reconstruction error for the efficient vs non-efficient sequences (maximum p = 5.9605e-08). These robust results demonstrate that our hypothesis for efficient sequence discovery based on reward-modulated replay is validated.
In [1], hippocampus replay during SWR is characterized by the activation order of the place-cells which occurs in backward and forward direction. We hypothesize that reverse replay allows the rat to explore a trajectory in one direction but consolidate it in both directions. This means that an actual trajectory, and its unexplored reverse version, can equally contribute to new behavior. Thus fewer actual trajectories are required for gathering information for solving the TSP problem. A systematic treatment of this effect on learning can be seen in S1 Text Analysis of different degrees of reverse replay.
We now investigate how reverse replay can be exploited in a recombination task where some sequences are experienced in the forward direction, and others in the reverse direction, with respect to the order of the sequence to be generated. We use the same setup as described above for novel sequence generation, but we invert the direction of sequence EBCDA in the training set. Without EBCDA, the model is not exposed to sub trajectories linking feeders B to C and C to D and the recombination cannot occur. We then introduce a partial reverse replay, which allows snippets to be played in forward and reverse order. This allows the reservoir to access segments BC and CD (even though they are not present in the forward version of the experienced trajectory.
Fig 8 illustrates the histogram of sequence performance for 10000 runs of the model (1000 models run 10 times each) on this novel sequence generation task with and without 50% reverse replay. We observe a significant shift towards reduced errors (i.e. towards the left) in the presence of reverse replay.
We then examine a more realistic situation based on the observation of spontaneous creation of “shortcuts” described in [47]. The model is exposed to a random replay of snippets extracted from two trajectories having different direction (clockwise CW and counter clockwise CCW). The system thus experiences different parts of the maze in different directions. We examine whether the use of reverse replay can allow the system to generate novel shortcuts.
The left and right trajectories used for training are illustrated in Fig 9A and 9B. In A, the system starts at MS, head up and to the left at T2 (counter clockwise) and terminates back at MS. In B, up and to the right (clockwise) again terminating at MS. Possible shortcuts can take place at the end of a trajectory at MS as the system continues on to complete the whole outer circuit rather than stopping at MS. We can also test for shortcuts that traverse the top part of the maze by starting at MS and heading left or right and following the outer circuit in the CW or CCW direction, thus yielding 4 possible shortcuts. The model is trained with snippets from the sequences in A and B using different random replay rates, and evaluated in non-autonomous mode with sequences representing the 4 possible types of shortcut. Fig 9C shows with no reverse replay, when attempting the CCW path, there is low error until the system enters the zone that has only been learned in the CW direction. There, the system displays clear deviations from the desired path. In the non-autonomous evaluation mode used in this experiment, after each response, the system is provided with the desired next location, which in this case creates a zigzag effect, corresponding to the spatial error. In panel D, with 50% reverse replay, this error is reduced and the system can perform the shortcut without having experienced the right hand part in the correct direction. Thus, in the right hand part of the maze, it is as if the system had experienced this already in the CCW direction, though in reality this has never occurred, but is simulated by the reverse replay. This illustrates the utility of mixed forward and reverse replay. Panel E illustrates the difficulty when 100% reverse replay is used. Fig 9F illustrates the reconstruction errors for a shortcut path as a function of degree of reverse. The trajectory is evaluated in non-autonomous mode and the position of the agent necessarily follows the target trajectory. In this case, the expected trajectory describes a CCW path. Results are not significantly different with a replay rate 25% and 75% (p = 0.02), where the best performance is observed, and all the other conditions are significantly different (p ≤ 1.1921e-07). This phenomenon was obtained for the 4 possible shortcuts.
The model demonstrates the ability to accumulate and consolidate paths over multiple trials, and to exploit reverse replay. Here we examine these effects on the more extensive and variable dataset extracted from rat behavior [13]. We show the positive effects of replay on trajectories from rats trying to optimize spatial navigation in the TSP task. In the prototypical TSP behavior, in a given configuration of baited wells, on successive trials the rat traverses different efficient sub-sequences of the overall efficient sequence, and then finally puts it all together and generates the efficient sequence. This suggests that as partial data about the efficient sequence are successively accumulated, the system performance will successively improve. To explore this, the model is trained on navigation trajectories that were generated by rats in the TSP task. We selected data from configurations where the rats found the optimal path after first traversing sub-sequences of that path in previous trials. Interestingly, these data contain examples where the previous informative trials include traversal of part of the optimal sequence in either the forward or reverse directions, and sometimes both (see S Rat navigation data). We trained the model with random replay of combinations of informative trials where informative trials are successively added, in order to evaluate the ability of the model to successively accumulate information. For each combination of informative trials, the random replay is evaluated with 0%, 25%, 50%, 75% and 100% of reverse replay rate in order to assess the joint effect of random replay and combination of informative trials. The model is then evaluated in non-autonomous mode with the target sequences that consist in a set of trajectories linking the baited feeders in the correct order. An idealized sequence is added to the target sequence set because trajectories generated by the rat might contain edges that do not relate the shortest distance between two vertices. Agent’s moves are restricted to a circle having a 10 cm radius.
Fig 10 illustrates the combined effects of successive integration of experience and its contribution to reducing error, and of the presence of different mixtures of forward and reverse replay. The ANOVA revealed that there is a significant effect for combination (F(2, 585) = 32.84, p < 0.01), as performance increases with exposure to more previous experience (Panel A). There is also a significant effect for reverse replay rate (F(4, 585) = 3.71, p ≤ 0.01), illustrated in Panel B. There was no significant interaction between consolidation and replay direction (F(8, 585) = 0.03, p = 1). This indicates that when trained on trajectories produced by behaving rats, the model displays the expected behavior of improving with more experience, and of benefitting from a mixture of forward and reverse replay.
We tested the hypothesis that hippocampus replay observed during sharp wave ripple events in the awake animal can play a role in learning by exposing the prefrontal cortex between successive trials to short sub-sequences of place-cell activation patterns. This replay can potentially play a crucial role in learning, essentially by generating synthetic data (based on experience) for training the system. The behavior of interest is a form of spatial navigation trajectory optimization in a task, mimicking the well-known traveling salesperson problem. It is a NP-Hard (non-polynomial) problem and finding an exact solution would require significant time and computing resources. Nevertheless, it has been observed that a rat was able to quickly find good solutions of simplified versions of this problem [13, 48]. The idea of exploiting replay in navigation sequence learning has been demonstrated to have a positive influence on learning [17], and here we go beyond this by further exploiting reward structure in the replay.
In the behavior of interest, rats are observed to converge quickly to a near-optimal path linking 5 baited food wells in a 151cm radius open arena. During their successive approximation to the optimal path, the rats often traversed segments of the optimal trajectory, as well as non-optimal segments. Observing this behavior, we conjectured the existence of neural mechanisms that would allow the optimal segments to be reinforced and the non-optical segments to be rejected, thus leading to the production of the overall near-optimal trajectory. We propose that the overall mechanism can be decomposed into two distinct neural systems. The first is a replay mechanism that favors the representation of snippets that occurred on these optimal segments, and that in contrast will give reduced representation to snippets that correspond to non-optimal trajectory segments. Here we demonstrate a simple but powerful method based on spatial reward propagation that implements this mechanism. Interestingly, this characterization of replay is broadly consistent with the effects of reward on replay observed in behaving animals [9].
The second neural system required to achieve this integrative performance is a sequence learning system that can integrate multiple sub-sequences (i.e. snippets) into a consolidated representation, taking into consideration the probability distributions of replay so as to favor more frequently replayed snippets. Here we considered a well-characterized model of sequence learning based on recurrent connections in prefrontal cortex that is perfectly suited to meet the sequence learning requirements.
Replay is modeled using a procedure that randomly selects a subset of place-cells coding part of a sequence, and outputs this snippet while taking into account the proximity of this snippet to a future reward. Each time a reward is encountered, it is taken into consideration in generating the snippet, and reward value is propagated backwards along the sequence, thus implementing a form of spatio-temporal credit assignment. This can be viewed in the Figs 2, 6 and 7 illustrating the snippet probability densities. The replay mechanism also implements a second feature observed in animal data, which is a tendency to replay snippets in reverse order. These two features of the replay model correspond to what is observed in the rat neurophysiology, and they also make fundamental contributions to the model’s ability to converge on an efficient navigation path. This extends previous demonstrations of the value of replay to include reward-modulated optimization [17].
Reservoir computing exploits the spatio-temporal dynamics of recurrently connected neurons that are sensitive to the spatiotemporal structure of input sequences [20, 27, 28]. The frontal cortex has been demonstrated to operate on these reservoir properties [19]. Here we demonstrated how a reservoir model of PFC meets two requirements for sequence learning: First, it can concatenate randomly replayed sub-sequences (snippets) in order to generate the complete original sequence. Second, it is sensitive to the statistics of replay, and thus can learn to ignore rare snippets (which correspond to snippets on inefficient sub-sequences, far from rewards) thus learning to optimize.
The instantaneous reward information acquired during a past experience is used for recursively updating the snippet replay likelihood in the hippocampus model. This creates a reward gradient and allows the optimal sequence to be assembled by the prefrontal cortex and striatum model. This is a novel combination of prioritized replay and reservoir computing in the context of reinforcement learning. The reward gradient is propagated along the spatial trajectory, and used to create a bias in the probability of replay. This biased replay is then provided as input to the reservoir PFC model. This is complementary to [49] who used replay to train a Dyna-Q reinforcement learning model. Both models benefit from replay, and can adapt to changes in reward contingencies. In our system, when the distant feeder is given a higher reward, this large reward produces a shift in replay probabilities (illustrated in Fig 6B), and the model learns this new distribution and favors the longer path to target D (illustrated in Fig 6D). The distinction is that we modulate the replay by reward probability, thus biasing the input to the sequence learning model towards the optimal solution. A secondary effect of rewards could be observed when rewards are sufficiently close for allowing a mutual contribution to the snippet replay likelihood surrounding the locations associated with reward delivery. Thus, we predict that a cluster of reward sites will have the effect of propagating the reward information farther than a single reward.
While we were principally motivated to study reward-prioritized replay combined with reservoir sequence learning in the TSP task, one can ask if the model generalizes to other tasks, particularly those that directly involve manipulation of reward. We thus observed that by changing the reward magnitudes, the system adapts and chooses a longer trajectory that leads to a larger reward. However, in more difficult problems that include the discovery of a long route to a single reward, the model could participate in the consolidation of partial solutions as illustrated in the current research, but would not be able to solve such problems autonomously.
The reverse replay mechanism has a dual effect. First, it provides the mechanism for the backwards propagation of reward along a trajectory. Based on this reward propagation, place-cell activation sequences leading to a nearby reward are represented more frequently and earlier than other less efficient sub-paths, which are thus rejected. This results in a form of spatio-temporal credit assignment that allows to take advantage of the reservoir network ability to combine multiple snippets into a whole sequence. We showed that it is possible to consolidate multiple sequences featuring parts of the same underlying optimal sequence into one efficient sequence and to generate it autonomously. Second, when the snippet replay likelihood is learned, a non-zero reverse replay rate allows the prefrontal cortex to be exposed to sequences of place-cell activations in both forward and reverse direction. This results in sequence learning in both directions while having experienced a place-cell activation sequence in one direction only. These results can be tested experimentally by recording place cells activities in SWR during the task.
During the intertrial period, the model predicts a co-occurrence of reverse replay from remote rewarded sites backwards to propagate the reward, and forward replay from remote locations towards rewarded sites to generate snippets from the optimal sub-sequences so as to generate the optimal path. Importantly, it also predicts that there will be a low probability of replay for subsequences that were on non-optimal trajectories. Future research should test these predictions.
A model of replay should predict which experiences should be replayed at each time to enable the most rewarding future decisions. Mattar and Daw [50] developed an elegant model of replay based on utility, characterized by a gain term that prioritizes states behind the agent when an unexpected outcome is encountered and a need term that prioritizes states ahead of the agent that are imminently relevant. This model predicts predominantly forward sequences prior to a run, and reverse sequences after a run. It accounts for a wide variety of behavioral and neurophysiological data, often in protocols where replay is observed during a run. We address a problem where the system is in a neutral area between trials in the TSP task. Thus, the current position of the animal is of low relevance. In this context, our model replays snippets that lie on the shortest route through the five baited paths. It would be interesting to observe how the Mattar and Daw model would respond during intertrial intervals in resolving the TSP problem.
The model we studied here is able to mimic the rat’s ability to find good approximations to the traveling salesperson problem by taking advantage of recent rewarding experiences for updating a trajectory generative model using hippocampus awake replay. We showed that reverse replay allows the agent to reduce the TSP task complexity by considering an undirected graph where feeders are vertices and trajectories are the edges instead of a directed graph. In this case, autonomous sequence generation is no longer possible but the information available in each prediction of the prefrontal cortex contains the expected locations. This allows the building of a navigation policy taking into account the salient actions suggested by the prefrontal cortex predictions, which are learned from hippocampus replay.
|
10.1371/journal.pntd.0007173 | The safety and efficacy of miltefosine in the long-term treatment of post-kala-azar dermal leishmaniasis in South Asia – A review and meta-analysis | Miltefosine (MF) is the only oral drug available for treatment of visceral leishmaniasis (VL) and post-kala-azar dermal leishmaniasis (PKDL). Although the drug is effective and well tolerated in treatment of VL, the efficacy and safety of MF for longer treatment durations (>28 days) in PKDL remains unclear. This study provides an overview of the current knowledge about safety and efficacy of long treatment courses with MF in PKDL, as a strategy in the VL elimination in South Asia.
Literature was searched systematically for articles investigating MF treatment in PKDL. A meta-analysis included eight studies (total 324 PKDL patients) to estimate the efficacy of MF in longer treatment regimens (range 6–16 weeks). We found a per-protocol (PP) initial cure rate of 95.2% (95%CI 89.6–100.8) and a PP definite cure rate of 90% (95%CI 81.6–96.3). Descriptive analysis showed that 20% of patients experienced adverse events, which mostly had an onset in the first week of treatment and were likely to get more severe after four weeks of treatment. Gastrointestinal (GI) side effects such as vomiting, nausea, diarrhoea, and abdominal pain were most common.
Longer treatment regimens with MF are effective in PKDL patients in India, however with the caveat that the efficacy has recently been observed to decline. GI side effects are frequent, although mostly mild or moderate. However, on the basis of limited data, we cannot conclude that longer MF treatment regimens are safe. Moreover, VL and PKDL pharmacovigilance studies indicate a risk for serious adverse events, questioning the safety of MF. The provision of safer treatment regimens for PKDL patients are therefore recommended. Until these regimens are identified, it should be considered to halt the use of MF monotherapy for PKDL in order to preserve the drug’s efficacy.
| In this study, we reviewed the available literature on the subject of safety and efficacy of the oral drug miltefosine in the treatment of post-kala-azar dermal leishmaniasis (PKDL). Literature was searched systematically in the PubMed database and eight articles, with a total of 324 PKDL patients, were included. A meta-analysis was performed to estimate the percentage of patients cured after longer (>4 weeks) miltefosine treatment. An estimated 90% of patients was found to be cured one year after treatment with miltefosine. In addition, descriptive analysis showed that nearly 20% of the PKDL patients suffered from side-effects. The majority of these side-effects, such as vomiting, nausea, diarrhea and abdominal pain, were mild and related to the gastro-intestinal tract. The findings of this study show that miltefosine is effective, although the efficacy has been observed to decline. The gastro-intestinal side effects were frequent but mostly mild. However, based on the limited data in this study we cannot conclude that longer treatment regimens with miltefosine are safe. In order to preserve the drug’s efficacy, we suggest it may be put under consideration to halt the use of miltefosine monotherapy for PKDL until alternative treatment regiments (e.g. short combination therapies including miltefosine) are identified.
| Post-Kala-Azar Dermal Leishmaniasis (PKDL) is a dermal complication of visceral leishmaniasis (VL) caused by the Leishmania donovani parasite, which is transmitted by phlebotomine sandflies. The PKDL disease can appear weeks to years after the successful cure of VL and is characterised by skin lesions, mainly present on places that are easily exposed to sunlight, such as the face [1]. The prevalence and severity of the disease vary between geographical regions. In East Africa, up to 60% of the former VL patients develop PKDL with mainly maculo-papular skin lesions, which are typically self-healing within three months. In South Asia, only 5–10% of the former VL patients develop PKDL. Most patients have hypopigmented macular lesions, however, up to 20% present with more severe papular or nodular skin lesions. Because spontaneous healing is probably limited [2,3], and may take years, treatment of more severe lesions is indicated. Considering PKDL cases are an important reservoir for transmission, potentially infecting new patients with VL [4], treatment is also required for public health reasons to achieve control of VL [1]. Because of the high endemicity limited to one geographical region and the availability of good diagnostic and treatment tools, in 2005 The Kala Azar Elimination Program was established as a regional initiative by the governments of Bangladesh, India and Nepal with the goal to eliminate VL in South Asia. Eliminating the PKDL reservoir is an important strategy in VL elimination.
The only oral drug available for the treatment of leishmaniasis is miltefosine (MF, hexadecylphosphocholine). This phospholipid derivative was originally developed as an anti-cancer drug but it was found to be unsafe after several studies indicated unacceptable renal- and gastrointestinal toxicity [5,6]. Scientists from Germany and the UK discovered the anti-leishmanial effect of the drug in the early 1990s. In 2003, MF was licensed for the treatment of VL [5]. The drug became the leading compound in the treatment of VL because it was effective, with limited side effects, and oral, so easy to administer [7]. In 2011, MF was added to the list of Essential Medicines by the WHO.
A substantial number of studies evaluated the safety and efficacy of MF in standard VL treatment of 28 days. Clinical trials have mainly been conducted in India, specifically in the state of Bihar, where VL is endemic [8]. Cure rates in VL patients range between 90–100% in a regular dose of 2.5 mg/kg per day for children aged 2–11 years; for people aged >12 years and < 25 kg body weight, 50 mg/day; 25–50 kg body weight, 100 mg/day; > 50 kg body weight, 150 mg/day; orally for 28 days. The safety concerns regarding MF mainly relate to its effect on the gastrointestinal tract [8]. Frequently observed adverse events in MF treatment regarding gastrointestinal toxicity that led to treatment interruption are nausea, vomiting, loss of appetite and diarrhoea. Other commonly observed toxicities are related to liver- and renal functions (e.g. elevated creatinine and ALT and AST levels). However, these are often not clinically relevant and normally stabilize during treatment [8]. In addition, animal studies have showed teratogenicity and impaired fertility in men and women, meaning that the use of MF could negatively influence the fetal congenital development. Impaired male fertility in humans as a consequence of miltefosine is currently under assessment by the FDA.
Miltefosine was first used in treatment of PKDL in 2006 [9]. In comparison to VL, PKDL requires longer treatment durations with MF. The drug is currently used as first-line treatment for at least twelve weeks in PKDL infected patients in India, Nepal, and Bangladesh [10]. PKDL requires longer treatment durations because of the limited skin penetration of antileishmanial drugs, and the fact that there is no other clinical marker for cure than disappearance of lesions, which may take more than one year in case of macular lesions [1]. Only few studies have investigated the safety and efficacy of the long-term MF treatment for PKDL and sample sizes in those studies are relatively small. Due to the slow clearance of MF in the body concerns are raised regarding the safety and efficacy of the drug in long-term treatment for PKDL. Therefore, this study aims to provide an overview of the current knowledge about safety and efficacy of longer treatment regimens (>28 days) with MF in PKDL patients, in order to contribute to the control of leishmaniasis.
This was a systematic review including a quantitative meta-analysis of data from different studies, in order to provide more accurate estimates of the effects of MF treatment in PKDL patients. This study was carried out in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [11].
The databases PubMed and Cochrane library were searched systematically using the following search terms: Miltefosine or hexadecylphosphocholine, Post-kala-azar dermal leishmaniasis, visceral leishmaniasis, kala-azar, safety, efficacy, tolerability, toxicity, clinical effectiveness, adverse events and South-Asia, India, Nepal, Bangladesh (Table 1). The total number of hits was 146. Fig 1 shows the flow diagram of the literature search. In addition to the computer search, reference search of all reviewed articles was performed to identify articles missed through the database search. One article was found manually.
Inclusion criteria were miltefosine monotherapy, VL or PKDL, human study population, and articles had to be written in English. Articles were excluded based on geographical location (America, Europe and Africa were excluded), in case the study used MF for treatment courses of 28 days or less and in case the study was conducted on animals. There were no further restrictions on age, sex or publication date.
All included articles were assessed on basic characteristics such as aim, methodological approach, sample size, treatment dose, treatment duration, conclusions and scientific quality. Primary outcomes of the current review were efficacy and safety. Efficacy was expressed in per-protocol (PP) cure rates and Intention-to-treat (ITT) cure rates at the end of treatment (i.e. initial cure rate) and at the end of follow up (i.e. definite cure rate).
Safety was displayed in adverse events and abnormal haematological parameters during or after treatment with MF. The seriousness of these toxicities was rated according to the Common Terminology Criteria for Adverse Events (CTC) of the National Cancer Institute [12]. Grades ranged from 1 to 5 (mild, moderate, severe, life-threatening and death). In case of mild and moderate severity (CTC grade 1 and 2, respectively), patients had to be treated with additional medication. In case of severe and life-threatening severity (CTC grade 3 and 4, respectively), treatment with MF had to be discontinued.
Data management and analyses were performed using SPSS version 25.0 [13]. Pooled estimates of initial and definite PP cure rates were calculated by random-effects regression analysis, using Wilson’s Macros for meta-analysis (Wilson, version 2005.05.23), after applying sample weights according to sample size. Moderator (subgroup) analysis was performed to indicate estimated cure rates for different duration treatment groups (a dummy variable was created for 6, 8, 12 and 16 weeks of treatment). Heterogeneity between studies was assessed using Cochran’s Q included in the meta-analysis function. A p-value of <0.05 indicated significant heterogeneity.
Eight experimental articles were included for analysis in the current study [14–21]. Table 2 provides an overview of the characteristics and main findings regarding efficacy and safety of MF in the included studies. All studies investigated longer treatment regimens of >28 days with MF in PKDL patients with the WHO-recommended standard dosing of (approximately) 2.5mg/kg/day, and were all originated from India. A total number of 324 patients were treated with MF, divided over a total of eleven treatment arms. Treatment duration ranged from six to sixteen weeks. One study investigated patients treated with MF for six weeks [21], in four study arms patients were treated with MF for eight weeks [15,18–20], in five study arms patients were treated for 12 weeks [14–18], and in one study patients were treated for 16 weeks [21].
There was some variation in methodological approaches between the included studies. First, all studies had an experimental design of which three were randomized controlled trials [14,18,21]. Furthermore, three of the included studies were single-arm trials [17,19,21], and five studies had two or more study arms [14–16,18,21]. Of those five studies, one study compared MF with another treatment (i.e. liposomal amphotericin B) [14] and the remaining four studies investigated different MF treatment durations [15,16,18,21]. Secondly, patients were treated as outpatients in five studies [14,15,17–19] while the rest of the included studies treated patients as inpatients (i.e. in hospitals) [16,20,21]. Thirdly, cure rates were assessed in two different ways. Three studies used parasite load measures by quantitative PCR (qPCR) at the end of treatment and at the end of follow up to indicate cure [14,15,17]. The remaining five studies assessed cure rate based on clinical features at the end of treatment and at the end of follow up [16, 18–21]. In those studies, patients were labelled cured if lesions had disappeared after treatment with MF. There was some variation in the length of follow up period between the studies. Two studies used a follow up period of six months [14,21], five studies used a follow up period of twelve months [15,16,18–20] and in one study a follow up period of eighteen months was used [17].
The cure rates at the end of follow up (definite cure) per study and the results of the meta-analysis are displayed in Fig 2. Meta-analysis showed an estimated PP definite cure rate of 90.0% (95%CI 81.6–96.3) and the average ITT cure rate was 74.9%. The lowest PP and ITT definite cure rates, 57% and 55%, respectively, were found in the study of Ghosh et al [17]. These numbers are substantially lower than the definite cure rates found in the other studies, which can be explained by the high number of treatment discontinuations due to severe side-effects in this study. In several study-arms, all patients, in at least one trial arm, were cured at 12-month follow up [14,17,19,20]. As can be seen in Table 2, the ITT definite cure rates ranged from 43–100%. The low ITT cure rates in the studies of Moulik et al [14], Ramesh et al [15] and Ghosh et al [17] (45%, 43% and 55%, respectively), are strongly influenced by high lost-to-follow-up in those studies. In the study of Moulik and colleagues [14], the drop-out-rate was no less than 57%. In the 8-week study arm of Ramesh et al [15] and in the study of Ghosh et al [17] the drop-out-rates were 33% and 35%, respectively.
The cure rates at the end of treatment (initial cure) per study and the results of the meta-analysis are presented in Fig 3. Five of the eight included studies reported an initial cure rate [15,18–21]. Meta-analysis showed an estimated per protocol initial cure rate of 95.2 (95%CI 89.6–100.8).
As can be seen in Figs 2 and 3, there seem to be outliers regarding both the initial and definite cure rates (i.e. numbers that lay outside of the 95%CI of the pooled estimates), which indicates heterogeneity. Analysis indicated the degree of variance in and between studies. In the analysis for initial cure rate, significant heterogeneity was indicated (Q = 15.6, I2 = 61.6% and P<0,05). 61.6% of the variance can be contributed to true heterogeneity. In the analysis for definite cure rate, no significant heterogeneity was indicated (Q = 13.4, I2 = 25.1 and P>0,05). 25.1% of the variance can be contributed to true heterogeneity.
In addition to the estimated overall cure rates, subgroup meta-analysis was performed to indicate the estimated cure rates per treatment group related to treatment duration. Table 3 shows the outcomes of this analysis with treatment duration as moderator variable. No significant differences were found in initial and definite cure rates between the different treatment durations.
Studies that were conducted in the past five years show a lower average cure rate (92.6% and 85.7% for initial and definite cure, respectively) than studies that were conducted more than five years ago (98.7% and 97.7% for initial and definite cure, respectively). However, these differences were not statistically significant (p = 0.142 and p = 0.081 for initial and definite cure, respectively).
Nearly 20% (n = 64) of all patients experienced adverse events. The most common side effects reported in the included studies are related to gastrointestinal (GI) adverse events. GI side-effects reported were nausea, vomiting, abdominal pain, diarrhoea or a combination of these events. All included studies reported that vomiting occurred in the majority of their patients. Vomiting was graded CTC 1 or 2 in nearly 10% of all patients (n = 20), however data was lacking in most studies regarding those mild and moderate adverse events. Vomiting with CTC grade 3–4 was experienced by three patients. In addition to vomiting, abdominal pain was reported in three studies (n = 10 patients) and graded CTC 1–3. In patients that experienced events graded CTC 3 or 4, treatment was discontinued. Events graded CTC 1 or 2 were treated symptomatically. In one study, six patients were treated with additional medication (Ondansetron) prior to taking MF in order to reduce repeated vomiting (CTC grade 2) [15]. In one study [17], treatment was reduced to twelve weeks because of unacceptable side effects.
Besides observable side effects, six studies provided data on haematological and laboratory tests performed before, during and after treatment. Laboratory abnormalities were seen in liver function (elevated bilirubin, SGOT and SGPT) and kidney function (elevated creatinine and serum asparate aminotransferase). However, in all but one patient, all of these laboratory abnormalities were non-severe and stabilised during treatment without interventions (e.g. additional medication, or treatment interruption). In one patient, an elevated bilirubin sample was graded CTC 2 [18].
In addition to the above-mentioned adverse events, one patient suffered from a cerebrovascular accident (CVA) [17]. This serious neurological condition (CTC grade 4) had most likely occurred as a result of the treatment with MF [17]. Ghosh et al [17] investigated the causality association between MF and the CVA with the ‘Naranjo adverse drug reaction probability scale’ [17]. However, an explanation for this association was not provided in the article.
The data provided about the time of onset of MF side-effects was lacking in the included articles. The studies of Ramesh et al [15] and Sundar et al [16] did not mention at what time during or after treatment the reported adverse events had occurred. In two studies was mentioned that the GI side-effects occurred during the first weeks of treatment. A few studies provided more concrete data on the days, or weeks, of onset of adverse events. In one study, unacceptable GI side-effects started after four weeks of treatment [15]. In addition, one study provided information on the day of onset for all gastrointestinal side effects [18]. The days of onset for vomiting graded CTC1 were: 32, 33, 38, 39, 48, 52 and 69, and vomiting graded CTC2 were: 33 and 77. Overall, adverse events were likely to occur in the first week of treatment, but became more severe after six weeks.
This study aimed to review the efficacy and safety of longer MF treatment regimens in PKDL patients. Meta-analysis showed an estimated cure rate of 95.2% and 90% for PP initial and definite cure rates, respectively. The average ITT cure rate was 74.9%. These findings are similar to literature investigating the efficacy of MF in treatment of VL with a duration of 28 days or less. Dorlo et al [8] found definite cure rates for VL ranging from 80–100% in their review. Furthermore, 97.3% of the 1100 VL patients in a large phase IV trial were cured after 28-day treatment with MF (93.2% by ITT analysis) [22]. In addition, 95% of these patients were cured at 12-month follow up (82% by ITT analysis) [22]. With regard to different treatment durations, subgroup analysis in this review showed no significant difference in initial or definite cure rates. However, the sample size of this study was small, and therefore identifying the most effective duration of MF treatment in PKDL patients requires further research.
Concerns were raised about potential toxicities as a result of the slow clearance of MF in the body, drug accumulation, and the lack of studies investigating long-term treatment. This review found that severe GI side-effects such as vomiting, nausea, abdominal pain and diarrhoea were experienced by nearly 20% of the PKDL patients. Dorlo et al [8] found similar side-effects in their review of 28-day treatment with MF for VL and explain that the GI side-effects can be attributed to MF’s working on the mucosa of the gastrointestinal tract. The current review found that adverse events in PKDL patients became more severe later in treatment (i.e. after six weeks). This can be explained by the long half-life of MF (approximately seven days) and the increasing drug levels in the patients over time. Contrary, in the trial of Bhattacharya et al [22], VL patients experienced more adverse events in the first week of treatment and those events diminished towards the end of the 28 days treatment. Bhattacharya et al [22] explained that the decrease of events over time might be a result of the rapid resolution of the VL disease features.
In the current review, one patient experienced a CVA (CTC4), which was assessed to be related to MF. To our best knowledge, this has not been seen in previous MF toxicity studies. There are, however, other severe incidental side-effects reported in VL studies, that were most likely related to MF treatment. In a VL study in India, a twelve-year-old boy was diagnosed with Steven-Johnson Syndrome (CTC4) [23]. Furthermore, one study reported the case of a male VL patient that developed fatal acute pancreatitis (CTC5) on the 13th day of treatment with MF [24]. Two recent studies conducted in Bangladesh [25,26] described five cases of ophthalmic issues (annular corneal ulcer, Mooren’s ulcer, and marginal keratitis) as a complication of the 12 weeks MF regimen in PKDL patients. In four cases the problems were reversible after discontinuation of MF. In the fifth case, MF treatment was continued as the issues were not reported. As a result, the patient has now permanent disability and blindness in the affected eye [25,26].
Phase I and II trials in the field of cancer research have indicated frequent toxicities and a lack of therapeutic efficacy in cancer patients treated with MF [27–32]. Similar to the findings in this review, the majority of side-effects were gastrointestinal. In the study by Berdel et al [28,29], 70% of the lung cancer patients treated with MF for nine weeks experienced episodes of nausea and vomiting. In the study of Unger et al [30], nearly 90% of the breast cancer patients experienced gastrointestinal side effects when treated with 100–150 mg MF daily for nine weeks. Similar results were found in a phase II trial where 90% of the cancer patients experienced episodes of nausea and vomiting when treated with MF for six weeks [31]. In addition to the gastrointestinal issues, another study indicated renal toxicities in 30% of their patients during MF treatment with doses up to 200mg per day (median treatment duration was six weeks) [32].
A challenge with MF is the reproductive toxicity. Embryo-fetal toxicity, including death and teratogenicity, was observed in embryo-fetal studies in rats and rabbits administered oral miltefosine during organogenesis at doses that were respectively 0.06 and 0.2 times the maximum recommended human dose (MRHD), based on body surface area (BSA) comparison. Numerous visceral and skeletal fetal malformations were observed in a fertility study in female rats administered miltefosine prior to mating through day 7 of pregnancy at doses 0.3 times the MRHD [33]. Therefore, female PDKL patients of child-bearing age are required to take contraceptives during and for five months after treatment with MF in order to prevent potential fetal congenital abnormalities. In addition to the teratogenicity, reduced fertility is seen in male VL patients treated with MF. Van Thiel et al [34] showed that 62% (n = 21) of the male military patients diagnosed with cutaneous leishmaniasis (CL) and treated with 150mg MF for 28 days experienced reduced ejaculation volume.
Despite the convenience of an oral treatment, patients are likely to poorly adhere to a twelve-week treatment that involves taking medication two times a day, when given non-directly observed. Because PKDL patients are typically not sick, the experience of frequent GI side-effects due to MF can easily result in missed doses and/or early discontinuation of treatment [35]. The reviewed articles showed relatively high dropout rates in groups with longer treatment durations, as a result of GI-side effects. For this reason, it was suggested that MF should be administered under clinical observation [6]. However, the practical feasibility of directly observed treatment administration can be questioned.
With regard to the non-adherence to MF treatment, Dorlo et al [8] emphasize the issue of loss of drug sensitivity and resistance that could lead to a decrease in the life-span of MF. While Dorlo et al [8] describe the drug non-susceptibility in vitro, while it is not yet demonstrated in vivo, more recent (case) studies indicate the increasing drug unresponsiveness and relapse rate in both VL and PKDL patients after MF monotherapy [8, 36–39]. The availability of expensive MF in the private sector in India ten years ago contributed to the persistence of sub-therapeutic dosage, resulting in drug-unresponsiveness [8,35]. In order to respond to the risk of resistance, the use of short combination therapies with MF is recommended. As an oral compound, MF has great potential to be used in multiple drug therapy for short durations (10 to 14 days). However, pharmacokinetic data show that it takes at least two weeks before MF reaches therapeutic blood levels [8]. Further research is necessary to identify safe and effective short combination therapies including MF in the treatment of PKDL patients.
The strength of this study is the meta-analytic design. Literature on the safety and efficacy of long-term treatment with MF is scarce and sample sizes are small. Therefore, combining the existing studies in a meta-analysis provides a more accurate estimate of the cure rates in endemic populations in South Asia. However, the results of this review need to be seen in the light of some limitations. First, all included studies were conducted in India, mainly in the state of Bihar. Although the majority of patients treated with long-term MF are Indian patients, one should be careful to generalise the results of this study to other endemic countries in South Asia. Secondly, the meta-analysis showed significant heterogeneity between studies, indicating that the variation in and between the studies was not based on standard error alone but can be contributed to methodological variations between studies (e.g. different assessments of cure, inpatient versus outpatient, and different research designs). Thirdly, the results of later studies may be affected by a decreased susceptibility to miltefosine and the overall efficacy we found may no longer reflect the reality on the ground.
In order to eliminate kala-azar in South Asia, PKDL patients need to be treated effectively. This review showed that treatment regimens with MF of six weeks or longer are effective (up to 90%) in PKDL patients in India, however with the caveat that the efficacy has recently been observed to decline. There is no straightforward answer to whether MF is an appropriate choice for the treatment of PKDL. This review showed that GI side effects are frequent in longer MF treatments, although mostly limited to mild or moderate side effects. However, on the basis of limited data included in this review, we cannot conclude that longer MF treatment regimens are safe. Moreover, information from previous VL studies and PKDL pharmacovigilance indicate a risk for serious, irreversible or even fatale adverse events, questioning the safety of longer treatment regimens with MF.
The highly common GI side effects can lead to non-compliance and form a risk for drug resistance. For this reason, directly observed treatment where possible, adequate surveillance of MF susceptibility in both PKDL and VL patients, as well as drug sensitivity monitoring in parasite isolates is required.
The provision of other treatment regimen for PKDL patients are highly recommended. It may be put under consideration to halt the use of miltefosine monotherapy for PKDL and proceed with safer alternative regimen. This will also help preserve the drug’s efficacy. In parallel, research into new treatment regimens should be encouraged.
|
10.1371/journal.pntd.0001501 | Using Molecular Data for Epidemiological Inference: Assessing the Prevalence of Trypanosoma brucei rhodesiense in Tsetse in Serengeti, Tanzania | Measuring the prevalence of transmissible Trypanosoma brucei rhodesiense in tsetse populations is essential for understanding transmission dynamics, assessing human disease risk and monitoring spatio-temporal trends and the impact of control interventions. Although an important epidemiological variable, identifying flies which carry transmissible infections is difficult, with challenges including low prevalence, presence of other trypanosome species in the same fly, and concurrent detection of immature non-transmissible infections. Diagnostic tests to measure the prevalence of T. b. rhodesiense in tsetse are applied and interpreted inconsistently, and discrepancies between studies suggest this value is not consistently estimated even to within an order of magnitude.
Three approaches were used to estimate the prevalence of transmissible Trypanosoma brucei s.l. and T. b. rhodesiense in Glossina swynnertoni and G. pallidipes in Serengeti National Park, Tanzania: (i) dissection/microscopy; (ii) PCR on infected tsetse midguts; and (iii) inference from a mathematical model. Using dissection/microscopy the prevalence of transmissible T. brucei s.l. was 0% (95% CI 0–0.085) for G. swynnertoni and 0% (0–0.18) G. pallidipes; using PCR the prevalence of transmissible T. b. rhodesiense was 0.010% (0–0.054) and 0.0089% (0–0.059) respectively, and by model inference 0.0064% and 0.00085% respectively.
The zero prevalence result by dissection/microscopy (likely really greater than zero given the results of other approaches) is not unusual by this technique, often ascribed to poor sensitivity. The application of additional techniques confirmed the very low prevalence of T. brucei suggesting the zero prevalence result was attributable to insufficient sample size (despite examination of 6000 tsetse). Given the prohibitively high sample sizes required to obtain meaningful results by dissection/microscopy, PCR-based approaches offer the current best option for assessing trypanosome prevalence in tsetse but inconsistencies in relating PCR results to transmissibility highlight the need for a consensus approach to generate meaningful and comparable data.
| Human African trypanosomiasis is a fatal disease that is carried by a tsetse vector. Assessing the proportion of tsetse which carries human-infective trypanosomes is important in assessing human disease risk and understanding disease transmission dynamics. However, identifying flies which carry transmissible infections is difficult, due to potential presence of other trypanosome species in the same fly, and concurrent detection of immature infections which are not transmissible. We used three methods to estimate the proportion of flies carrying human-infective trypanosomes: dissection and microscopic examination of flies to visualise trypanosomes directly in the fly; PCR of fly midguts in which trypanosomes were observed by microscopy; and theoretical analysis using a mathematical model of disease transmission. All three methods found the prevalence to be extremely low. Given the low prevalence, dissection/microscopy requires prohibitively large sample sizes and therefore PCR-based approaches are likely to be of most value. However, interpretation of PCR data is not straightforward; whilst PCR identifies flies carrying pathogen genetic material it does not directly identify flies with transmissible infections. This study highlights the need for a consensus approach on the analysis and interpretation of PCR data to generate reliable and comparable measures of the proportion of flies which carry transmissible human-infective trypanosomes.
| For the vector-borne diseases, pathogen prevalence in a vector population is an indicator of disease risk, and accurate measures of the proportion of vectors carrying infections are needed for (i) guiding allocation of resources or targeting intervention programs [1]; (ii) monitoring the success of control interventions [2]; and (iii) as parameters in models of disease transmission which are increasingly used to predict disease distribution and persistence, and plan control interventions [3]. Approaches for detecting parasite prevalence in vector populations, known as xenomonitoring, have until recently usually relied on dissection of insect vectors and visualisation of parasites by microscopy, which is time consuming and reliant on operator skill. PCR has presented an alternative technique for several parasite-vector systems, e.g. Plasmodium spp [4], Oncocerca volvulus [5], [6], Leishmania spp. [7], [8], and the nematodes which cause lymphatic filariasis, Wuchereria bancrofti, Brugia malaya and Brugia timori [9], [10], generally having better ability to differentiate between species of similar morphology, increased sensitivity, and hence requiring smaller sample sizes [4], [6], [8].
Human African trypanosomiasis (HAT) is caused in East Africa by Trypanosoma brucei rhodesiense and transmitted by tsetse flies (Glossina spp). Measuring the prevalence of T. b. rhodesiense in the tsetse vector is of particular importance as HAT occurs in developing countries where resources for surveillance and disease control are limited [11] and knowledge of human disease risk is important for effective targeting of available resources. In addition, HAT is characterised by its focal nature, with human cases continuing over long periods of time in specific geographical areas, but the reasons for this persistence are not clear [12]. The prevalence of infection in tsetse is an important component in understanding transmission dynamics and detecting spatiotemporal trends, which have important implications for disease control.
Assessment of the prevalence of trypanosomes within tsetse populations has traditionally comprised dissection and microscopic examination of the mouthparts, midguts and salivary glands of the fly, relying on the differing development and maturation sites of the trypanosome subgenera to identify trypanosome species [13]. Trypanosomes found only in the mouthparts are classified as Duttonella or vivax-like, trypanosomes located in the mouthparts and midguts are classified as Nannamonas or congolense-type, and trypanosomes found in the midgut and salivary glands are Trypanozoon or brucei-like. When trypanosomes are found only in the midgut, the infection is assumed to be immature. This dissection/microscopy technique has several disadvantages for use in field studies: it is not possible to differentiate below the level of subgenus (for example T. simiae cannot be differentiated from T. congolense, since they share development sites in the fly); mature and immature infections cannot always be differentiated; and mixed infections cannot be identified or discriminated. Dissection and trypanosome identification are highly dependent on operator skill, and there exist variations in protocols, with some authors only examining the midgut and salivary glands if trypanosomes are found within the mouthparts [14], [15], whilst others examine all the organs [16], [17].
A suite of molecular tools has been developed for the trypanosomatids [18], [19]. PCR and sequence analysis techniques have served to overcome some of the disadvantages of dissection/microscopy and highlighted new information about tsetse-trypanosome interactions. PCR primers with high sensitivity and specificity now permit trypanosomes to be reliably identified to species or subspecies level, for example new strains or potentially even species of trypanosome have been identified [20], [21], [22], and human-infective T. b. rhodesiense and its morphologically-identical subspecies Trypanosoma brucei brucei (not pathogenic to man) can be accurately differentiated [23]. Mixed infections are common, with approximately one third of PCR positive flies carrying more than one trypanosome species [20], [24], [25] and up to four trypanosome species identified in individual flies [24], [25].
However, when it comes to assessing the prevalence of trypanosome infections in tsetse it is clear that the results generated by dissection/microscopy do not correlate well with data generated by PCR (for example only 38% [25] to 51% [24] of Nannomonas or T. congolense-like and Duttonella or T. vivax-like infections are classified as the same species by both techniques). For T. brucei sensu lato, with its potential for human infection, this presents a particular problem. In areas where T. b. rhodesiense is known to occur in wildlife and livestock hosts, and human cases are reported, the majority of studies of T. brucei s.l. in tsetse by dissection/microscopy show prevalence of zero, even when thousands of flies are examined [16], [26]. However when whole tsetse flies have been analysed by PCR surprising amounts of T. brucei s.l. DNA has been found, with 2% of G. palpalis and 18% of G. pallidipes testing positive [27], [28]. The discrepancy between dissection/microscopy and PCR highlights the issues of assessing the true prevalence of human infective trypanosomes in tsetse populations, particularly as it is not clear how these measures relate to transmissibility. Furthermore, it would be useful if a consensus could be reached as to how best to use molecular data, either alone or in combination with results of dissection/microscopy, to generate prevalence measures.
This study presents data from a persistent focus of Rhodesian HAT in the Serengeti National Park (SNP), Tanzania. Whilst cases of HAT have been reported in this area for over one hundred years [29], recent cases in both the local population and tourists have renewed public health concerns about the disease [30], [31]. With abundant populations of G. swynnertoni and G. pallidipes, and almost 100 000 tourists visiting the SNP each year in addition to resident staff and local populations [32], understanding and mitigation of human disease risk is a priority.
Previous studies carried out in SNP have relied on dissection/microscopy to determine tsetse prevalence (Table 1). Large scale studies in 1970 and 1971 failed to identify any salivary gland infections [16], [26] but a subsequent pooled rodent inoculation study detected nine out of 11000 G. swynnertoni flies (0.08%) infected with T. brucei s.l. [33]. These findings contrast with results of a more recent study that reported a prevalence of 3.0% for T. brucei s.l. in G. swynnertoni [34] and raise questions as to whether the wide variation in detected prevalence reflects real changes in tsetse infection levels and human exposure risk, or reflect methodological differences.
This study assessed the prevalence of T. brucei s.l. and T. b. rhodesiense in the two main tsetse species in SNP, G. swynnertoni and G. pallidipes, using (i) dissection/microscopy and (ii) PCR analysis of infected midguts and salivary glands. A third approach was applied to infer the prevalence of T. b. rhodesiense in tsetse from a mathematical model of disease transmission, to examine whether previously reported low prevalences were consistent with other parameters that have been estimated for this system.
All field work was conducted in SNP, Tanzania, between October and November 2005 and August and October 2006. Tsetse sampling was carried out with the Tsetse and Trypanosomiasis Research Institute, Tanga, Tanzania. Seven sites were randomly selected for tsetse trapping in savannah and open woodland areas, within 1 km of roads and within a 40 km radius of park headquarters at Seronera, where tsetse dissection was conducted (coordinates UTM 36M (i) 711676, 9731432; (ii) 706816, 9733868; (iii) 710747, 9733536; (iv) 695691, 9727934; (v) 700825, 9746320; (vi) 693961, 9733122; (vii) 695278, 9741360). In each study site, three Epsilon traps [35] were installed for between five and eleven days, depending on trap catches. Each trap was situated at least 200 m from the next, and erected in mottled shade to reduce fly mortality. When placing traps, areas with fallen trees were avoided and traps were placed so that the entrances were directed towards gaps in vegetation, measures known to maximise fly catches by following the natural patterns of tsetse flight [36]. The location of each trap was recorded using a handheld global positioning system (Garmin Ltd, Kansas, USA). Traps were baited with 4-methylphenol (1 mg/h), 3-n-propylphenol (0.1 mg/), 1-octen-3-ol (0.5 mg/h) and acetone (100 mg/h) [37] and emptied twice daily.
All live non-teneral flies were dissected and labrum, hypopharynx, salivary glands and midgut examined for trypanosomes under 400× magnification [38]. For each fly, species, sex and the presence or absence of trypanosomes in each organ were recorded. To prevent contamination between flies and between different parts of each fly, dissection instruments were cleaned in 5% sodium hypochlorite, followed by rinsing in distilled water then phosphate buffered saline between each organ. Flies carrying trypanosome infections was categorised according to Lloyd and Johnson [13]. Confidence intervals were calculated using binomial exact 95% limits.
All trypanosome-positive midguts and salivary glands were macerated in phosphate buffered saline and applied to FTA Classic cards (Whatman, Maidstone, UK) for further analysis. A subset of trypanosome-negative midguts was also preserved on FTA cards. FTA cards were allowed to dry for two hours and stored in foil envelopes with dessicant at ambient temperature prior to processing. For each sample, one disc of diameter 2 mm was cut out from the FTA card using a Harris Micro Punch™ tool. Between cutting of the sample discs, 10 punches were taken from clean FTA paper, to prevent contamination between samples. Discs were washed for two washes of 15 minutes each with FTA purification reagent (Whatman Biosciences, Cambridge, UK), followed by two washes of 15 minutes each with 1X TE buffer (Sigma Aldrich, Dorset, UK). Each disc was dried at room temperature for 90 minutes, and then used to seed a PCR reaction. After every seven sample discs, a negative disc was included and the punch tool and mat cleaned, to reduce the risk of contamination between discs, and ensure that any potential contamination would be detected. No evidence of contamination was seen in the sequence of dissection or PCR results.
TBR primers were used to detect a 177 bp satellite repeat sequence common to T. b. brucei, T. b. rhodesiense and T. b. gambiense [39]. PCR was carried out in 25 µl reaction volumes containing 16.0 mM (NH4)2SO4, 67 mM Tris-HCl, 0.01% Tween 20 (NH4 buffer, Bioline Ltd, London, UK) 1.5 mM MgCl2, 800 µM total dNTP's, 0.4 µM of each primer TBR1 and TBR2, 0.7 Units of BioTaq Red DNA polymerase (Bioline Ltd, London, UK) and one washed disc. For samples testing positive for T. brucei s.l., T. b. rhodesiense was differentiated from T. b. brucei by detection of the serum-resistance associated (SRA) gene. Simultaneous amplification of another single copy gene, a phospholipase C (PLC) sequence found in T. brucei s.l., confirmed that there was sufficient T. brucei s.l. material present in the sample to detect the presence of T. b. rhodesiense [40]. SRA PLC PCR was carried out in duplicate in a 25 µl reaction volume containing 3 mM MgCl, 1.25 µl of Rediload dye (Invitrogen, Karlsbad, California), 1.5 Units Hot StarTaq (Qiagen, Crawley, UK), 0.2 µM of each primer and one washed disc. The SRA gives a 669 bp product, with a PLC band at 324 bp.
For all PCRs, one negative control (water) and one positive control (genomic DNA) were run for every 16 samples, in addition to negative control blank discs. PCR products were run on a 1.5% (w/v) agarose gel at 100 V, stained with ethidium bromide and visualised under an ultraviolet transilluminator.
Detection of T. b. rhodesiense in a tsetse midgut does not indicate a mature infection as only a small proportion of midgut infections will develop to mature infections in the salivary glands. The following calculation was used to predict the prevalence of mature transmissible T. b. rhodesiense infections, where Dispos is the proportion of flies with midguts which were positive by dissection/microscopy, PCRpos is the proportion of these which tested positive by PCR, PTbr/Tbb is the proportion of T. brucei s.l. positive flies with sufficient genetic material present (ie give positive results with PLC PCR) which test positive for T. b. rhodesiense (as determined by SRA PCR) and Pmat is the proportion of immature T. b. rhodesiense infections which develop to maturity in the salivary glands, estimated to be 0.12 (CI 0.10–0.14), [41], [42]:(1)This calculation relies on three assumptions: (i) that dissection/microscopy is 100% sensitive for detecting trypanosome infections in tsetse midguts, and that all flies carrying T. brucei s.l. will have midgut infectons; (ii) that TBR PCR has 100% sensitivity and specificity for detection of T. brucei s.l. in tsetse midguts; (iii) that SRA PCR has 100% sensitivity and specificity for detection of T. b. rhodesiense, if the sample is positive on PLC PCR. The implications of potential assumption violations on the prevalence estimate are addressed in the discussion.
Confidence intervals were calculated by repeat sampling from nested distributions of the data. Since the value for Pmat was taken from Milligan et al. (1995) the distribution of the original data was used, where Y is the number of flies with midgut infections and Pmat is the proportion of these which developed mature salivary gland infections (Y = 1133, Pmat = 0.12). Potential values were generated by sampling from the following nested distributions with 10 000 iterations, and ninety five percent confidence intervals calculated by taking the 2.5% and 97.5% quantiles of the values obtained: n1∼binom(N, Dispos), n2∼binom(n1, PCRpos), n3∼binom(n2, PTbr/Tbb), p1∼binom(Y, Pmat), n4∼binom(n3, p1/1133).
Rogers' [43] model of vector-borne trypanosome transmission was adapted for one host population (wildlife, x) and two vector populations (G. swynnertoni, y1 and G. pallidipes, y2). Although occasional cases of human African trypanosomiasis do occur, the rate of human feeding by tsetse is very low [0.1% of feeds on blood meal analysis, 16], so the human population was not included in the model. The model is described by the following equations:(2)(3)(4)that were simultaneously solved using the lsoda function in the package odesolve in R (http://www.r-project.org/) to give equilibrium conditions for the prevalence of T. b. rhodesiense in wildlife hosts, G. swynnertoni and G. pallidipes and which could be compared to empirically derived estimates of prevalence.
Parameters were based on those described by Rogers [43] but adjusted to reflect infection in wildlife (Table 2). Parameters specific to T. b. rhodesiense, and to G. swynnertoni and G. pallidipes, were used where possible. The proportion of tsetse developing salivary gland infection after feeding on an infected cow is 16% for G. morsitans (closely related to G. swynnertoni) and 2.1% for G. pallidipes [44]; however wildlife exhibit a degree of trypanotolerance and generally show low parasitaemia [45], which reduces the probability that a feeding tsetse will develop infection, also indicated by very low infection rates in tsetse fed on wildlife experimentally [46], [47]. A number of wildlife species do not appear to develop infection with T. brucei s.l., either proving uninfectible in experimental infections eg baboons [46] or rarely observed with natural infection despite being popular hosts for tsetse, eg elephant [16], [48], [49], so the probability that an infected tsetse feeding on a host results in an infection is also lower compared to cattle. The incubation period of 18 days follows that of Dale et al. [50] for laboratory infections of T. b. rhodesiense in G. morsitans flies; no specific data were available for G. pallidipes so the same value was used. Wildlife host parameters have been chosen to represent all wildlife species. Duration of incubation period and duration of infection are therefore estimated mean values from experimental infections of wildlife [46], [51], [52]. Although age prevalence patterns suggest the development of some immunity to T. brucei s.l. in lions [53], experimental infections do not indicate a clear immune period in other species [46]. SNP has high densities of both wildlife [54] and tsetse [34].
All statistical analyses and model solving were carried out using R 2.12.1 (The R Foundation for Statistical Computing, http://www.r-project.org).
In total, 6455 tsetse were dissected and examined, comprising 4356 G. swynnertoni (2759 females, 1597 males) and 2099 G. pallidipes (1472 females, 627 males). Overall, trypanosomes were observed (in mouthparts, midgut, or both) in 9.2% of G. swynnertoni (females 10.2%, males 7.5%), and 3.7% of G. pallidipes (females 3.9%, males 3.2%) examined. No salivary gland infections were observed. Using the classical trypanosome species identification based on the location of parasites within the fly, the prevalence of T. vivax-like, T. congolense-like and T. brucei-like trypanosomes is shown in Table 3.
For 5428 flies (all those sampled in 2006), all midguts where trypanosomes were observed (n = 133) were analysed by PCR (Table 4). No flies were found with salivary gland infections. The prevalence of flies with trypanosomes in the midgut on dissection/microscopy, which were also midgut PCR positive (Dispos×PCRpos, assumed to represent T. brucei s.l. immature infections) was 0.83% in G. swynnertoni and 0.71% in G. pallidipes. All midguts that tested positive for T. brucei s.l. were further analysed with SRA PCR, with 10 out of 43 PLC positive and 1 of these SRA positive, therefore the proportion of T. brucei s.l. testing positive for T. b. rhodesiense was 0.1. Using the expression in Eq. 1, this gives a predicted prevalence of transmissible T. b. rhodesiense infections of 0.010% for G. swynnertoni and 0.0085% for G. pallidipes (Table 4). The prevalence was also calculated separately by sex and using sex-specific maturation ratios of 0.21 for males and 0.044 for females [41]. The predicted prevalence of T. b. rhodesiense mature infections in G. swynnertoni was 0.016% for males (the number of flies testing positive on dissection/microscopy and PCR out of the total number examined was 11/1448) and 0.0035% for females (20/2289), and in G. pallidipes was 0.019% for males (5/541) and 0.0024% for females (7/1151).
Midguts from 78 flies with no trypanosomes observed on microscopy were also analysed by PCR. Of these, 3.8% (n = 3) tested positive for T. brucei s.l.. None of these tested positive with PLC or SRA.
Assuming equilibrium, the model yielded prevalences of T. b. rhodesiense of 0.0064% in G. swynnertoni and 0.00085% for G. pallidipes. The model predicted the prevalence of T. b. rhodesiense in wildlife hosts to be 2.5%, which is within the range of reported prevalences in wildlife in SNP of 1.8% and 4.3% [55], [56].
The results of all three approaches are presented in Table 5.
In this study we present data obtained from three different approaches to measuring the prevalence of transmissible T. b. rhodesiense infections in tsetse populations in Serengeti National Park. Fundamental difficulties have been identified associated with the detection of trypanosome infections in tsetse, requiring new approaches to move beyond generation of infection prevalence data to make inferences about transmissibility. The three approaches used in this study confirmed the prevalence of T. b. rhodesiense in SNP to be very low. The prevalence of T. brucei s.l. measured by dissection/microscopy was zero, despite confirmation by the other techniques that T. brucei s.l. was circulating in the area, and evidence of infection in wildlife and human hosts, highlighting a common problem with this technique. The results from PCR analysis of tsetse midguts were used to generate a measure of transmissible infections. In addition, a mathematical model of disease transmission used to predict the prevalence of transmissible infections based on other parameters for this system, confirmed the low prevalence gained by other approaches was compatible with the prevalence of T. b. rhodesiense in wildlife hosts reported in SNP. This study highlights specific challenges in measuring transmissible T. b. rhodesiense infections in tsetse, which have important implications for assessing this variable, and interpreting temporal and spatial patterns of infection in affected areas of Africa.
These results illustrate the difficulties of dissection/microscopy techniques, which in this study estimated the prevalence of T. brucei s.l. in tsetse populations as zero, despite strong evidence to indicate the presence of infection in tsetse using other techniques, and evidence for circulation of T. b. rhodesiense in vertebrate hosts in the same area [30], [31], [55]. The low prevalence commonly obtained through dissection/microscopy is often attributed to low diagnostic sensitivity of this technique, and there is evidence that some infections which would be classed as immature by microscopy may actually be transmissible. For example, inoculation of trypanosomes found in the mouthparts from flies with trypanosomes present in the mouthparts and midgut by dissection did give rise to T. brucei s.l. infections in mice, both in laboratory and field studies [57], [58], and PCR of dissection-negative salivary glands revealed additional T. brucei s.l. infected flies in Glossina palpalis palpalis in Cote d'Ivoire [59]. Whilst this may play a part in the low prevalence observed, the use of other techniques in this study confirmed the prevalence to be extremely low, and the prevalence of zero by dissection/microscopy in this study is more likely attributed to insufficient sample size than low sensitivity. With a prevalence of 0.01% (the highest of the estimates in this study) it would be necessary to examine around 30 000 flies to detect a difference from zero with 95% confidence.
Dissection/microscopy has a number of other disadvantages: it is time consuming and requires skilled technicians, and whilst it does not require substantial investment in technology, this may be outweighed by high staff costs. Identification of species, mixed infections and immature infections is unreliable, particularly if other trypanosome species are also of interest. Furthermore dissection/microscopy alone cannot differentiate between T. b. brucei and T. b. rhodesiense. The dissection/microscopy technique was first discussed in detail by Lloyd and Johnson in 1924 as an alternative to cumbersome rodent inoculation studies. However, Lloyd and Johnson relied principally on morphology of the developmental and infective forms, using the location within the fly only as an additional aid. It is clear that in areas where the prevalence is very low, dissection is less than ideal. However, since the majority of historical studies have relied on dissection/microscopy it is important to understand how these data compare to those generated by other techniques if we want to be able to detect temporal trends.
PCR-based techniques have the potential to provide a sensitive and specific tool to identify flies carrying T. b. rhodesiense. We found that 30% of microscopy-positive midguts tested positive for T. brucei s.l. by PCR in G. swynnertoni and 41% in G. pallidipes. It is difficult to compare these directly with other studies as protocols vary widely, but between 7.9% and 19% of microscopy-positive midguts have been reported testing positive for T. brucei s.l. in these tsetse species [20], [21], [25]. However, a PCR positive fly does not indicate a transmissible infection, but only indicates the presence of trypanosomal DNA. Here we have combined PCR data with information on the proportion of immature T. b. rhodesiense infections which mature to the salivary glands to estimate the prevalence of mature transmissible infections. The prevalence was within the confidence limits of dissection/microscopy and similar to the predictions of the model. Prevalence was higher in males than females, reflecting the increased probability of maturation in males [41]. Although in this study, dissection/microscopy were carried out prior to PCR, the increased likelihood of detecting immature T. brucei s.l. in midguts by PCR means the sample size can be lower for the equivalent precision, reducing field costs and time compared to the substantial sample sizes needed for dissection/microscopy only.
The calculation used to predict the prevalence of mature T. b. rhodesiense infections by incorporating dissection/microscopy and PCR data relied on assumptions regarding the sensitivity of dissection/microscopy for detecting midgut trypanosome infections, and the diagnostic sensitivity and specificity of TBR and SRA PCRs when used on tsetse midgut samples. Whilst identification of trypanosomes in the midgut is widely used in the laboratory there is little data available on the sensitivity of this technique in the field. There is however no evidence to suggest that flies can carry T. brucei s.l. without trypanosomes being present in the midgut. TBR and SRA PCRs have high specificity [40], [60]. Whilst the analytical sensitivity of TBR and SRA PCRs is known (they are both able to detect 0.1 pg of trypanosome genetic material or less, equivalent to one trypanosome [39], [40]), there is no quantitative data on the diagnostic sensitivity when used on tsetse samples. The diagnostic sensitivity of TBR on blood samples from livestock is 76% [60]; however the number of parasites in tsetse midgut samples is several fold higher than the parasitaemia in livestock (which is often <10 trypanosomes/ml [40]) hence diagnostic sensitivity is likely to be considerably higher for tsetse samples.
Imperfect test sensitivity and specificity can significantly affect prevalence estimates, particularly when the prevalence is very low [61]. Ideally the sensitivity and specificity of each technique would have been included in the analysis to produce prevalence estimates and confidence intervals that reflect this information. The paucity of data to examine these assumptions illustrates the importance of more critical assessment of these techniques, but likely reflects the difficulty of assessing sensitivity and specificity in the absence of a gold standard technique. In the absence of quantitative data, the most likely violation of the assumptions is that the sensitivity of each technique is not 100% hence the prevalence may have been underestimated.
In this study, 10% T. brucei s.l. infections were identified as T. b. rhodesiense. Whilst this is not outside the range of values found in previous studies [62], a proportion of one third has been more commonly reported [63]. SRA PCR targets a single copy gene, and therefore requires the presence of a large amount of parasite DNA. Despite an initial sample size of over 6000 flies, only ten infected midguts had sufficient genetic material present to check for T. b. rhodesiense, so our estimate of the proportion of T. brucei s.l. which are T. b. rhodesiense is not very precise (10%, CI 0.2–44%). Using the value of 33% resulted in a prevalence of T. b. rhodesiense in G. swynnertoni of 0.03% and in G. pallidipes of 0.028%.
It is interesting that 3.8% of microscopy-negative flies tested positive for T. brucei s.l. by PCR. Previous authors have found high prevalences of T. brucei s.l. by PCR (for example 18% [27]), and there are potential explanations for this high detection rate. Flies that test positive on PCR but were microscopy-negative may result from the presence of trypanosomal DNA (known to be detectable for over 10 days in the absence of live trypanosomes [64]) or a very small number of trypanosomes for example in a recent blood meal where trypanosomes are not able to establish an infection. Experimentally it has been established that only around 12–43% of susceptible flies feeding on an infected host will develop an immature infection even in teneral flies [44], [65]. In older flies, the majority of trypanosomes ingested will not develop further. Simple calculations illustrate that if trypanosomal DNA is detectable for 10 days, flies feed every 3 days and 5% of hosts carry T. brucei s.l., at any one time, up to 17% of flies may have detectable T. brucei s.l. DNA, in the absence of an immature or mature infection.
Given the drawbacks of using other techniques, it is reassuring that a model incorporating independently estimated parameters for this system predicted similar values for the prevalence of T. b. rhodesiense in tsetse. Whilst it might seem questionable whether the very low prevalence found by the other techniques is consistent with the reported prevalence of T. b. rhodesiense in wildlife hosts of 1.8–4.3% [55], [56], a simple equilibrium-based model analysis showed that with T. b. rhodesiense prevalence in wildlife of 2.5%, the prevalence in tsetse remains below 0.01%, and consistent with field measures. For diseases such as HAT where low prevalence raises diagnostic challenges, broad agreement of prevalence estimates using quite different approaches permits a measure of confidence in each.
A constraint to going forwards with making assessments of prevalence is the absence of a gold standard technique for identifying transmissible T. b. rhodesiense infections in tsetse. Dissection/microscopy requires prohibitive samples sizes and potentially may not detect all transmissible infections; PCR techniques based on amplification of DNA from midguts rely on assumptions of factors which are known to vary and tests for which the diagnostic performance is poorly defined; models require accurate knowledge of all other parameters in a system and assumptions regarding equilibrium dynamics. Even rodent inoculation may miss infections as rodents often fail to become infected due to their innate resistance to infection. However, approaches for the future are likely to rely on PCR based techniques so it is important that reliable and comparable protocols are developed. Currently, there are many different approaches reported for using PCR data to look at T. brucei s.l. in tsetse populations, including PCR of any organs found infected [25] (similar to this study although we did not include mouthparts), PCR of all organs in the fly if any organ is found infected on dissection/microscopy [59], [66] and PCR of whole tsetse flies [for example 27], [28]. This variety of protocols raises two important issues:
To interpret data from PCR analysis it is important to be clear what PCR results do or do not represent. For example, identification of T. brucei s.l. DNA by PCR in whole flies does not indicate a mature and therefore transmissible infection, but only the prevalence of T. brucei s.l. DNA. Is it possible to use this measure as a direct indicator of risk? This approach has been taken for other pathogens. For example in assessing prevalence of West Nile virus in mosquitoes, most screening programs test the whole mosquito, detecting mosquitoes with any trace of WNV present, rather than testing the salivary glands, which would give the rate of transmissible infections [1]. PCR studies to identify the nematodes which cause lymphatic filariasis in mosquito populations give a prevalence of infected mosquitoes, but cannot differentiate between pre-infective L1 and L2 larvae, and infective L3 larvae [10]. However this approach is more common where detecting pathogen presence or absence is the main aim, so the exact nature of the relationship between presence of pathogen DNA and transmissible infections is less critical. Approaches measuring the prevalence of T. brucei s.l. or T. b. rhodesiense DNA, either in infected midguts, in all midguts or in whole flies, are assuming a constant relationship between this measure, and the prevalence of transmissible infections (in turn assumed to represent human risk). In this study we relied on experimental measures of the proportion of midgut T. b. rhodesiense infections which mature to the salivary glands to estimate the prevalence of transmissible infections. However there are two areas for concern with this assumption: (i) laboratory studies may not accurately reflect the situation in the field; and (ii) this proportion is known to vary with factors such as sex, levels of certain antioxidants, mating in female flies, and environmental factors such as temperature [41], [67]. While this approach may be suitable for obtaining an approximate measure of prevalence, the validity of the assumptions would be challenged by comparative studies over different spatial or temporal situations where these factors are likely to vary. Interpretation of PCR results from analysis of whole flies or from midguts without prior dissection/microscopy is more problematic. This study illustrates the high proportion of microscopy-negative midguts which test positive by PCR and similar findings are reported from PCR of whole flies [27], [28]. It is not known how the proportion of flies testing positive by this technique relates to the prevalence of transmissible infections. Approaches involving PCR of salivary glands may hold most promise. PCR of microscopy-negative salivary glands or salivary drops has been shown to increase the prevalence compared to dissection/microscopy alone both in the field [59] and the laboratory [68]. It is not clear what these discrepancies between microscopic and PCR analysis of salivary glands means with regard to transmission and this is an area where further research is required.
The second concern is with respect to comparative data analysis, in that the variety of techniques used means it is difficult to assess trends in prevalence. This is a significant problem – prevalences measured in different ways cannot be compared between different areas or times, making it impossible to detect changing disease dynamics and human disease risk, and hindering our understanding of the complex relationships between trypanosomes, hosts and vectors. Agreement on an optimal protocol for the collection and interpretation of data on trypanosome prevalence in tsetse populations would be helpful in generating more comparable data.
This study shows that the prevalence of T. b. rhodesiense in G. swynnertoni and G. pallidipes in SNP can be sustained at very low levels. Both the PCR data and the model suggest that G. pallidipes may play a role, albeit a lesser one, in T. b. rhodesiense transmission as well as G. swynnertoni, which has always been regarded as the important vector species in Serengeti. The two species differ in both feeding preferences and vector competence; while both species include suids and bovids in their diet, G. swynnertoni feeds predominantly on warthog while G. pallidipes feeds predominantly on buffalo [69], [70]. Although both G. swynnertoni and G. pallidipes are known to avoid feeding on man, this effect is particularly evident for G. pallidipes [71], which likely decreases the importance of this species in human disease transmission. The prevalence found in this study is consistent with that of previous studies by dissection/microscopy [16], [26], [72] so we did not find any evidence of long term trends in disease transmission. However, the prevalence in this study does differ significantly from that reported in 2007 of 3% [34]. Whilst this may reflect temporal or spatial variation in prevalence within SNP, our model suggests that a sustained prevalence this high is very unlikely.
The low prevalence of T. b. rhodesiense in tsetse found in this study suggests that the risk of HAT to tourists is low. Odour-baited tsetse traps are known to target older flies [73]; flies which bite people are usually younger and less likely to be carrying a transmissible infection since the prevalence of mature infections increases with age [74]. This is consistent with the low number of cases (<5 per year) reported in Serengeti, in comparison to the large number of visitors (almost 100,000 per year [32]). However, the risk of encountering an infected fly is higher in those who spend extended periods exposed to tsetse in SNP, so it should be ensured that adequate screening and treatment provision is in place to detect cases in park and lodge staff.
In conclusion the prevalence of transmissible human infective trypanosomes in tsetse populations is an important parameter but there is no ideal diagnostic test to measure it. While new molecular diagnostic tools offer great potential for epidemiological studies, many challenges remain in the interpretation of field data generated from these tools, and these need to be recognised and addressed. Development of protocols that directly measure the prevalence of transmissible infections, and the consistent application of such protocols, would aid our knowledge of human disease risk, allow detection of spatial and temporal trends in disease transmission and add to our understanding of complex disease systems.
|
10.1371/journal.pgen.1005771 | Arabidopsis Flower and Embryo Developmental Genes are Repressed in Seedlings by Different Combinations of Polycomb Group Proteins in Association with Distinct Sets of Cis-regulatory Elements | Polycomb repressive complexes (PRCs) play crucial roles in transcriptional repression and developmental regulation in both plants and animals. In plants, depletion of different members of PRCs causes both overlapping and unique phenotypic defects. However, the underlying molecular mechanism determining the target specificity and functional diversity is not sufficiently characterized. Here, we quantitatively compared changes of tri-methylation at H3K27 in Arabidopsis mutants deprived of various key PRC components. We show that CURLY LEAF (CLF), a major catalytic subunit of PRC2, coordinates with different members of PRC1 in suppression of distinct plant developmental programs. We found that expression of flower development genes is repressed in seedlings preferentially via non-redundant role of CLF, which specifically associated with LIKE HETEROCHROMATIN PROTEIN1 (LHP1). In contrast, expression of embryo development genes is repressed by PRC1-catalytic core subunits AtBMI1 and AtRING1 in common with PRC2-catalytic enzymes CLF or SWINGER (SWN). This context-dependent role of CLF corresponds well with the change in H3K27me3 profiles, and is remarkably associated with differential co-occupancy of binding motifs of transcription factors (TFs), including MADS box and ABA-related factors. We propose that different combinations of PRC members distinctively regulate different developmental programs, and their target specificity is modulated by specific TFs.
| Polycomb group proteins (PcGs) are essential for development in both animals and plants. Studies in plants are advantageous for elucidation of specific effects of PcGs during development, since most PcG mutants are viable in plants but not in animals. Previous efforts in genetic study of plant PcGs revealed that different PcGs have both common and unique effects on plant development, but the mechanisms underlying the specific regulation of different developmental programs by PcGs are still far from clear. In this study, we quantitatively compared the change in H3K27me3 and gene expression profiles between mutants of key PcG members on a genome-wide scale in Arabidopsis seedlings, and successfully unraveled different developmental programs that are specifically regulated by different combinations of PcGs. This context specific effect of PcGs is closely associated with different sets of transcription factor binding motifs. Together, we revealed on a genome-wide scale that different combinations of PcGs, as well as their association with the binding sites of different TFs, serve to explain the specific regulation of different developmental programs by PcGs.
| The evolutionarily conserved Polycomb group proteins (PcGs) are the major epigenetic machinery regulating differentiation and development [1–4]. PcGs mediated repression is achieved by establishment and maintenance of epigenetic modifications surrounding target genes. In both plants and animals, PcGs are classified into two major multi-protein complexes PRC1 and PRC2, which participate in transcriptional repression by catalyzing H3K27 tri-methylation and H2A ubiquitination, respectively [1–4]. Depletion of various PcG components in plants lead to varied developmental defects [5–13]raising a major question about how the functional specificity of PcGs is established.
Study of the functional specificity of PcGs in both plant and animals is non-trivial. Firstly, the majority of PcG components are ubiquitously expressed, and do not have sequence-specific DNA recognition properties. Secondly, members of PcGs generally have functional redundancy and diversity, and it is difficult to distinguish the specific effect of individual members. Plants are advantageous for studying the effect of PcGs in development since most plant PcG mutants are viable, while animal development is generally vulnerable to PcG mutations. Previous efforts in genetic dissection of PcGs’ functions provide important clues as to the specialized functions of PcG members. For example, CURLY LEAF (CLF) and SWINGER (SWN) are two highly similar enzymatic subunits of the PRC2 complex [11], and they play redundant roles in plant development as double mutants clf swn show a much more severe phenotype than each of the single mutants [11]. However, this redundancy is partial since SWN cannot rescue the phenotypic defects upon loss of CLF, including early flowering and curly leaf [11]. Similarly, lack of either AtRING1 or AtBMI1, the core catalytic factors of PRC1, leads to depression of embryonic traits in seedlings, while only the atring1a atring1b double mutant displays severely fused flower phenotype [12,13]. In addition, LIKE HETEROCHROMATIN PROTEIN1 (LHP1, also known as Terminal Flower-2, TFL2) is a PRC1 component [10,14,15] capable of interacting with both PRC1 components AtBMI1 and AtRING1 in vitro [12,13,16]. However, lhp1 displays some similar phenotypic defects to those of the PRC2 mutant clf[10], and more recent evidence showed that LHP1 co-purifies with PRC2 complex [17]. Moreover, studies on different loci led to different conclusions on the interplay between PRC1 and PRC2, including their orders of recruitment into the corresponding complexes [7,18,19]. It seems that different PcGs tend to repress specific gene sets with distinct functions. However, target genes and the mechanisms of specificity for different PcGs are not sufficiently characterized, particularly from a genome-wide point of view.
Multiple mechanisms had been proposed in Arabidopsis to explain the target specificity of PcGs, e.g. specific recruitment of PcGs by diverse strategies for transcription repression [7,20–27] and selective displacement of PcGs by some transcription factors for transcription de-repression during developmental transitions [28,29]. However, most of these proposed mechanisms are based on studies of some specific genes without an overview at genome-wide scale. It is also worth to note that some conclusions drawn from genome-wide studies may not be always consistent with results obtained at specific gene loci. An example case is regarding the H3K27me3 modification change in the lhp1 mutant. The global H3K27me3 pattern in the mutant was similar to that in wild-type Col-0 plant, leading to a conclusion that LHP1 is responsible for recognizing H3K27me3 and facilitating PRC1 binding but not for depositing H3K27me3 [14]. However, some more recent ChIP-qPCR results revealed that several PRC2 targets show an obvious reduction of H3K27me3 levels in lhp1[17,23]. This discrepancy between genome-wide and ChIP-qPCR results could be due to H3K27me3 differences in lhp1 being localized to some specific genomic regions, which had been missed in detection by genome-wide profiling with relatively low resolution. Therefore, for unraveling the locus selectivity and distinguishing the specific effects of different PcGs, combining high-resolution genome-scale data with quantitative analyses methods are indispensable.
The next-generation sequencing technology has enabled the detection of epi-genome profiles with high sensitivity and specificity [30–32]. We have recently developed a package for quantitative comparison of epi-genomic data, showing a high quality in dissecting specific epigenomic modifications in animal development [33–36]. Using these newly developed methods, here we quantitatively compared genome-wide changes of H3K27me3 and gene expression profiles in loss-of-function mutants in PRC1 (AtBMI1, AtRING1 and LHP1) and PRC2 (CLF) components in Arabidopsis seedlings. We revealed that CLF collaborates with different PRC1 subunits to repress flower and embryo development. We further demonstrated that the target specificity of these different combinations of PcGs is closely associated with different sets of TF binding motifs, pointing to an active interplay between particular TFs and the specific activity of different PcGs.
To dissect the composition of PRC2 complex in Arabidopsis, we used the FIE-3XFLAG fusion protein in immunoprecipitation experiments to identify associated proteins from leaf explants cultured in callus-induction medium. Following mass spectrometry analysis (see Methods), we identified the well-documented PRC2 components, including SWN, CLF, EMF2 and VRN2, as well as the previously considered PRC1 component LHP1 (S1 Table), consistent with recent report in inflorescence [17], indicating that the interaction between LHP1 and PRC2 complex is relatively stable across different tissues. It is worth noting that although LHP1 has the ability to bind the core catalytic subunits of PRC1 including AtBMI1 and AtRING1 in vitro[12,13,16], neither component was identified here, suggesting a special role for LHP1 in association with PRC2 complex.
To compare roles of different PcGs, we used chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) to characterize the genome-wide profiles of H3K27me3 in Col-0, clf-29, tfl2-2, atbmi1a,b, and atring1a,b (S2A Table). In Col-0, 5,055 read enriched H3K27me3 regions (peaks) were identified (S2B Table), 84% of which localized in promoter and genic regions (S1 Fig). The top enriched functions for those peak targets include transcription regulation, carpel and petal development (S2C Table). Notably, almost half of the annotated MADS box proteins are marked by H3K27me3 in seedlings (S2C Table). To compare H3K27me3 marks across samples, we first ranked the read intensities of Col-0 H3K27me3 peaks from high to low (Fig 1A panel I), which is positively correlated with the binding of PcG components including FIE, EMF1 and LHP1 (Fig 1A panel VII), and is inversely associated with the expression level of surrounding genes (Fig 1B). Next, the level of H3K27me3 mark in the corresponding regions in mutants were plotted side-by-side with that of Col-0 (Fig 1A panel III-VI), which showed highly similar patterns, with the global level slightly lower in some mutants as compared to Col-0. Similarly, slight reduction of H3K27me3 has also been reported in Arabidopsis mutant lacking EMBRYONIC FLOWER1 (EMF1)[32], a plant specific PcG member [37]. These results indicate that lack of any of these PcGs does not lead to complete loss of H3K27me3.
However, we observed different sets of loci showing apparent reduction of H3K27me3 in Arabidopsis deprived of different PcG members, suggesting PcG subunits have non-redundant roles in H3K27me3 deposition in local regions (Fig 1C). Traditional peak overlap method showed poor performance in characterizing H3K27me3 changes in PcG mutants (S2 Fig). Thus, we used the pipeline of MAnorm, a software package specifically designed for quantitative comparison of ChIP-seq datasets [33]. MAnorm derives its power from definition of M value, a statistic characterizing the strength of differential binding, the higher the absolute M value, the larger the difference, with the sign (+/-) representing higher intensity in PcG mutants or Col-0. In clf-29, we identified substantially more regions with reduced H3K27me3 modification than with increased H3K27me3 marks (Fig 2A), consistent with a major role for CLF in catalyzing H3K27me3. Intriguingly, similar pattern was also observed in tfl2-2 and atbmi1a,b (S3 Fig), thus, PRC1 factors possibly contribute to H3K27me3 establishment. We also found hundreds of regions with increased H3K27me3 in clf-29, majority (68%) of which overlapped with H3K27me3 peaks in Col-0, while those non-overlapping regions also tend to be marked by H3K27me3 in Col-0, which are below peak detection cutoff, as shown by the read intensity distribution plot (S4 Fig). This increase of H3K27me3 in clf-29 could be possibly due to the effect of other PcGs. Previous studies in both human and Drosophila also observed hyper-methylation of H3K27 in Ezh2 mutants [38,39], but no consensus has been made about the underlying mechanism.
To test if the quantitative difference of H3K27me3 has an effect on differential expression of target genes, H3K27me3 regions were partitioned to consecutive groups ranked by M value, and the gene targets for each group were identified (see Methods). The percentage of genes showing differential expression, up or down regulation separately, are depicted for each group (Fig 2B). In general, target genes associated with negative M value—that is, peaks with reduced H3K27me3 in mutants—were enriched in genes more highly expressed in mutants, and vice versa, which is consistent with the repressive role of H3K27me3. This indicates that the M value determined by MAnorm reflects authentic H3K27me3 changes.
Since loss of LHP1 affects the H3K27me3 level for thousands of genes, we wondered whether the effect of LHP1 on H3K27me3 is direct or indirect. Previous evidence suggested that LHP1 co-localizes with H3K27me3 [14,15]. We compared the overlap between LHP1 binding peaks and genomic regions with change of H3K27me3 in tfl2-2 characterized by M value (Fig 2C). Notably, nearly 1/4 regions with reduced H3K27me3 level (M<0) overlap with LHP1 binding sites as determined by previous ChIP-chip study [15]. The overlap is significantly higher (P value < 1e-3) than expected by chance based on permutation test (see Methods). In addition, we observed that the higher the effect of LHP1 on deposition of H3K27me3—as indicated by low M values—the higher the percentage of corresponding H3K27me3 regions overlapping with LHP1 binding, indicating a direct relationship between LHP1 binding and H3K27me3 change.
To dissect the cooperation of these PcGs on H3K27me3 modification, we collected 3,289 H3K27me3 regions regulated by at least one PcG component, and clustered the M values of H3K27me3 change in these regions to 3 clusters (Fig 3A and S3A Table). Although LHP1 physically interacts with AtRING1 and AtBMI1 in vitro[12,13,16], the H3K27me3 change profile in tfl2-2 is closely correlated with that of clf-29, with 1,982 regions (peak set I) showing concerted reduction of H3K27me3 in both mutants. On the other hand, loss of AtBMI1A and AtBMI1B or AtRING1A and AtRING1B specifically reduced H3K27me3 level in 566 regions (peak set II), which represents PRC2 target sites affected by AtBMI1 and AtRING1 directly or indirectly, but is independent of LHP1. Thus, the dependence of H3K27me3 on LHP1 seems tightly correlated with the specific effect of CLF. In addition, 741 regions belonging to class III show increased H3K27me3 levels.
It is remarkable to find that CLF and LHP1 have a coordinated effect on H3K27me3 modifications at many genome regions. We wondered how these two factors control H3K27me3 around similar loci. A closer look at regions in peak set I indicates that loss of CLF or LHP1 lead to apparent reduction of H3K27me3 in surrounding regions, but for most cases the signal at the summit is only slightly reduced (Fig 3B). A similar finding was reported supporting CLF-dependent disperse of H3K27me3 around transgenes carrying AG regulatory sequences [42]. Heatmap in Fig 3C showed the average read intensity around summits of peak set I (Fig 3D). Statistical analysis detected this phenomenon for 51% and 53% H3K27me3 reduction regions in clf-29 and tfl2-2, respectively (S3B and S3C Table), indicating that CLF and LHP1 participate in H3K27me3 spreading.
To investigate the functional consequence of the distinct H3K27me3 profile controlled by different combinations of PcGs, we first characterized the transcriptome change in each PcG mutant, including atring1a,b, atbmi1a,b, clf-29, lhp1-6,and tfl2-2. Next, 2,438 genes with differential expression in at least one of the five mutants were collected, and partitioned to 3 groups according to the expression change pattern across the 5 samples via k-means clustering (Fig 4A and S4A Table). Genes in group I are specifically up-regulated in lhp1-6, tfl2-2 and clf-29, some of which are also induced in atring1a,b to some extent, but have no obvious change in atbmi1a,b. Comparison with expression change in clf-29 and swn-21 indicated that genes specifically increased in clf-29 as compared to swn-21 showing significant enrichment in group I (S5 Fig). Group II represents genes specifically higher expressed in atring1a,b and atbmi1a,b. Group III are genes repressed in all samples, which is perhaps not a direct effect of PcG components. It is interesting that genes from both group I and group II are upregulated in clf-29swn-21 double mutants, whose transcriptome change is closely correlated with that in atring1a,b (Fig 4B), indicating RING1 and BMI1 regulated genes tend to be concertedly controlled by PRC2.
Next, we wondered whether genes from different groups participate in distinct functions or pathways. GO enrichment analysis showed that gene set of group I is closely related with flowering, floral development and transcription (Fig 4C), whereas genes in group II are involved in nutrient reservoir, seed storage, and lipid localization (Fig 4D). Enriched terms and genes are listed in S4B and S4C Table. These functional transcriptome analyses could serve to explain the specific phenotypic defects observed in corresponding PcG mutants.
To dissect the relationship between genes with distinct change profiles of H3K27me3 and differential expression, the targets for both peak set I and peak set II shown in Fig 3A were identified followed by statistical testing of their enrichment in each of the three expression groups identified in Fig 4A. As expected, genes in expression group I are significantly over-represented in targets of peak set I, representing genes showing decreased H3K27me3 marks and increased expression level in mutants of CLF and LHP1, and thus most likely to be the direct targets of CLF and LHP1 (Fig 5A). Similarly, genes in expression group II are preferentially enriched in targets of peak set II, representing genes affected by AtBMI1 and AtRING1 via H3K27me3 (Fig 5B).
To dissect the repressive function of PcGs via H3K27me3, 108 genes with increased expression and reduced H3K27me3 marks in mutants of CLF and LHP1 are extracted (S5A Table). 164 targets regulated by AtBMI1 and AtRING1 were identified in the same way (S5B Table). The distribution of both expression and H3K27me3 change in different PcG mutants were plotted for the 108 and 164 genes, respectively (Fig 5C and 5D). The induced expression of 164 genes in clf-29swn-21 double mutant but in neither clf-29 nor swn-7 confirms that CLF and SWN could complement each other’s function for these genes. It’s also worth noting that mutation of CLF but not SWN is responsible for the induction of 108 genes.
To investigate the function of the 108 and 164 genes during development, we first classified all tissues with gene expression information into 8 groups based on their gene expression profiles (S6 Fig). Next, we calculated the enrichment of the 108 and 164 genes in the different tissue-biased genes using Gene Set Enrichment Analysis (GSEA)[43]. The 108 genes are significantly more highly expressed in flowers, while the 164 genes are preferentially expressed in embryo (S7 Fig). GSEA using RNA-seq data showed consistent profiles with much lower P values (Fig 5E and 5F and S6 Table). Notably, neither gene sets was significantly enriched in other tissues (Fig 5G and S4 Fig), indicating a prominent relationship between PcGs and specific regulation of reproductive and embryo development.
We asked whether this different combination of PcGs is associated with different TFs. We started by searching for the enriched motifs surrounding all Col-0 H3K27me3 peak summits. The top enriched motifs include the binding motifs of ABI3 type transcription factor B3 (ABI3/FUS3/LEC1), ABI4, ABF1, SPL and MYB (S8 Fig). The first three are binding motifs of ABA related TFs, which mainly participate in embryo development [44–47], and is consistent with the well-documented function of PcGs in regulation of embryo development [12,16,31]. Despite the fact that some MADS-box TFs are reported to modulate specific H3K27me3 deposition for regulation of meristem identity, flowering and floral development in individual loci [20,22,23], CArG box, the binding motifs of MADS box transcription factors [48,49], show no enrichment when all H3K27me3 regions are considered (S2D Table).
Next, we identified motifs over-represented in peak sets I and II (Fig 6A and 6B). Of note, binding motifs for MADS box and Homeobox were specifically enriched in peak set I, indicating a close relationship between transcription factors from these families and H3K27me3 levels synergically regulated by CLF and LHP1. On the other hand, motifs enriched in peak set II were similar to the result from all H3K27me3 peaks in Col-0, including ABI4 and ABF1 binding sites. The binding motif for B3 domain TFs are enriched in all H3K27me3 regions, to a higher extent in peak set II, consistent with the major role of their targets in embryogenesis, seed maturation and dormancy [50].
Since multiple TFs could bind to the same cis-regulatory sites, to identify TFs closely associated with different sets of peaks, we first collected and processed published ChIP-seq data characterizing TF binding profiles from Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/). By querying against these processed binding sites, we found that the binding regions of some MADS box TFs are significantly enriched in peak set I but not in other H3K27me3 regions (Fig 6C and S7 Table). The top enriched TFs are floral organ identity genes, including AP1 [51], AG [52], AP3 [53] and SEP3 [51]. IGV screenshots in Fig 6D and S9A Fig illustrate some examples of co-occupancy between H3K27me3 in seedlings and these MADS-box TFs. The source of the TF ChIP-seq data, the enrichment statistics and co-occupied regions are listed in S7 Table. It should be noted that these TFs mainly expressed in inflorescence where these ChIP-seq data were generated from, while our ChIP-seq data were generated in seedlings. Thus, it is likely that specific binding of these TFs in some H3K27me3 regions from peak set I is responsible for selective de-repression of common target genes. In support of this, target genes of peak set I co-occupied by these TFs in inflorescence show apparent flower biased expression, while the other target genes of peak set I have no such expression bias (Fig 6E and S9B Fig), indicating the displacement of H3K27me3 by these TFs participating in activation of their common target genes. Consistently, it has been reported that some floral organ identity genes could interact with REF6 [54], the H3K27me3 demethylase in Arabidopsis [41], which possibly work together with MADS-box TFs to remove H3K27me3 marks. Taken together, our genome-wide analyses based on both motif and ChIP-seq data revealed that the bindings of MADS box TFs are closely associated with H3K27me3 peak set I regulated by CLF and LHP1, suggesting that it might be a widespread mechanism by which the specific activities of PcG family proteins is modulated by tissue specific TFs, resulting in distinct transcriptional outputs in different tissues.
In this study, based on quantitative comparison of epigenomic and transcriptomic data in mutants of core PcG components, we revealed that CLF collaborates with different PcGs partners to achieve transcriptional repression in distinct developmental programs. Importantly, target specificity of different combination of PcGs are closely associated with different sets of TF binding motifs, suggesting a widespread mechanism for modulation of PcGs specificity by particular TFs. We propose a context-dependent model for PcGs in selective repression of flower or embryo development (Fig 7).
As an ancient machinery for developmental regulation, PcGs employed multiple protein families which underwent duplication and diversification, and thus they share common targets but also have specialized functions. The common targets may represent those that are involved in more ancient processes in development, while the unique features may have evolved more recently. In support of this, we observed that embryo development, presenting in an overwhelming majority of land plant (embryophyte)[55], is regulated by majority core components of PRC1 and PRC2, while flower development, a relatively recently evolved process in plants [56], is specifically regulated by only a subset of PcGs, with LHP1 and CLF as the major players. Then does the emergence of LHP1 or CLF parallel the appearance of flowering plant (angiosperms)? LHP1 is present in ancient plant species including Selaginella moellendorffii and Physcomitrella patens[57]. Interestingly, our phylogenetic analysis revealed that the divergence of CLF and SWN likely occurred accompanying the emergence of angiosperms (S10 Fig), with the duplication existing in all angiosperm species collected, including Amborella trichopoda, the oldest known angiosperm, but not in more ancient species, including moss, Selaginella or plant from Gymnospermae. Thus, it is possible that after the duplication event, CLF preferentially acquired the ability of regulating floral development. Similarly, AtRING1 and AtBMI1, the catalytic subunits of PRC1, have both joint and individual functions. It is likely that in addition to the common targets with AtBMI1, AtRING1 also participate in regulating H3K27me3 modifications and expression for several target genes of CLF and LHP1 (Figs 3 and 4), which possibly serves to explain aberrant flower development only observed in atring1a,b but not atbmi1a,b[12,13].
Our finding that PRC2 composition regulates differential transcriptional programs is analogous to a recent genome-wide study about human blood cell development [34]. It was demonstrated that histone methyltransferase Enhancer of zeste1 (EZH1) and EZH2, the counterparts of CLF and SWN in human, form alternative PRC2 complexes with distinct subsets of PcGs, occupy different chromatin domains and regulate distinct transcriptional activities [34]. Given that epigenetic machineries generally have multiple family members, and thus could form a suite of different combinations, different compositions of epigenetic complexes could at least partially explain their selectivity in transcriptional regulation.
Target selection by PcGs in different developmental stages is critical for correct developmental regulation. Specific recruitment of PcGs by particular factors [7,20–23], Polycomb response elements (PREs)[26,27], or non-coding RNAs [24,25] have been a major explanation for both PRC1 and PRC2 binding in plants [58]. Recent reports proposed that release of PcG by a particular MADS protein AG is the prerequisite for activation of KNU [28,29]. Whereas all these conclusions are drawn from analyses based on limited loci, we provide clues from the genome-wide scale that MADS box TFs are possibly associated with specific release of PcGs. Specific recruitment of PcGs by MADS box TF has recently been reported. SHORT VEGETATIVE PHASE (SVP), the major flowering repressor in seedlings [59], was shown to be able to interact with LHP1, and contribute to SEP3 repression via H3K27me3 [23]. Consistently, we found SVP binding sites [59] show good correlation with K27me3 modification in peak set I (S11 Fig). However, it’s binding sites are also enriched in peak set II, while the common binding sequences have no obvious relationship with CArG-box motifs. We observed that SVP ChIP-seq data have relatively high noise, and the binding regions tend to be broad, thus some coincident bindings could possibly be due to noise. Alternatively, SVP has diverse functions in addition to recruit PcGs for flowering repression.
MADS box TFs play a central role in flower development, and their family size increased explosively with the origin of angiosperm [60]. We found almost half of the Arabidopsis MADS box TFs have the H3K27me3 modification in seedlings, indicating a strict control of MADS gene expression by PcGs (S2C Table). On the other hand, we demonstrated that binding sites of MADS box TFs specifically enriched in genes controlled by CLF and LHP1, and possibly contribute to modulate the target selectivity. These findings close the loop of the transcription network regulating development processes, in which expression of tissue specific TFs is controlled by epigenetic marks, and specific TFs cooperate with epigenetic complexes in determining target selection and developmental regulation. How the epigenetic machinery co-evolved with specific TFs to cooperatively fine-tune the regulation of plant development is an intriguing phenomenon deserving further in-depth exploration.
It is intriguing that telo-box, a widespread short motif identical to the repeat (AAACCCT)n of plant telomeres [61], is specifically enriched in peak set I (Fig 6A and 6B). Telo-box is reported to be involved in regulating gene expression in cycling cells [62], and whether the association between telo-box and the function of CLF or LHP1 involved in maintaining H3K27me3 marks during cell cycle is an interesting issue for further study. Genome-wide Hi-C analyses revealed telomeric regions and H3K27me3 modifications form local interactive hot spots [64]. And a recent study showed complementary activities of TELOMERE REPEAT BINDING proteins and PcGs in transcriptional regulation of target genes [63]. It is possible that telo-box also mediates local interaction among H3K27me3 marked regions. Collectively, our study provides rich resource as well as insightful clues for further exploration of the relationship between cis-elements, TFs and PcGs in specific regulation of developmental processes.
The chromo-domain protein Heterochromatin protein 1 (HP1) and Polycomb (Pc) in animals bind H3K9me3 and H3K27me3, respectively [65–68]. Despite LHP1 sharing relatively higher level of sequence similarity with HP1, it can bind H3K27me3 in vitro and co-localize with H3K27me3 in vivo [14,15], and was originally proposed to be the counterpart of Pc, responsible for recruitment of PRC1 to H3K27me3 catalyzed by PRC2. If this is the case, LHP1 should function downstream of PRC2. However, recent studies showed that H3K27me3 modifications also require LHP1 [17,23]. Here, our genome-wide results revealed that H3K27me3 levels of thousands of loci are controlled by LHP1. Further quantitative comparison of H3K27me3 change profiles across PcG mutants revealed that the effects of LHP1 on H3K27me3 modification and target gene repression are coordinated with the non-redundant role of CLF (Figs 3 and 4). In addition, both CLF and LHP1 are involved in spread of H3K27me3 marks since loss of either component lead to localized H3K27me3 signals (Fig 3B, 3C and 3D). These findings not only confirmed previous report from studying AG transgene that CLF is indispensable for H3K27me3 spreading [42], but also identified LHP1 as an important cofactor with CLF in H3K27me3 elongation, which could finally contribute to inheritance and stability of epigenetic silencing.
Notably, LHP1 has no effect on genes jointly controlled by both PRC1 and PRC2 (targets of peak set II) despite the coincident binding of LHP1 and H3K27me3 in these regions. If the requirement of chromo-domain protein for H3K27me3 maintenance is a widespread mechanism, then there may be other chromo-domain proteins functional for spreading of H3K27me3 in these H3K27me3 regions. Alternatively, PcGs employ multiple strategies for H3K27me3 maintenance at different loci, either using LHP1 or cooperating with core subunits of PRC1 to create compacted chromatin structures [69]. There are 13 Arabidopsis proteins that have chromo-domain, and further epigenomic studies are required to have a deeper understanding about the role of chromo-domain proteins on epigenetic modifications. Due to the functional redundancy and localized effects of epigenetic machineries, the quantitative comparison pipeline applied in this study will be of great help for further exploration based on high throughput data.
The ChIP-seq and RNA-seq data were deposited in Gene Expression Omnibus (GEO http://www.ncbi.nlm.nih.gov/geo/) under the accession number GSE67322. Tracks for all sequencing data and related public data can be visualized through our local genome browser: http://bioinfo.sibs.ac.cn/gb2/gbrowse/tair10/
|
10.1371/journal.ppat.1001015 | Oseltamivir-Resistant Pandemic A/H1N1 Virus Is as Virulent as Its Wild-Type Counterpart in Mice and Ferrets | The neuraminidase inhibitor oseltamivir is currently used for treatment of patients infected with the pandemic A/H1N1 (pH1N1) influenza virus, although drug-resistant mutants can emerge rapidly and possibly be transmitted. We describe the characteristics of a pair of oseltamivir-resistant and oseltamivir-susceptible pH1N1 clinical isolates that differed by a single change (H274Y) in the neuraminidase protein. Viral fitness of pH1N1 isolates was assessed in vitro by determining replication kinetics in MDCK α2,6 cells and in vivo by performing experimental infections of BALB/c mice and ferrets. Despite slightly reduced propagation of the mutant isolate in vitro during the first 24 h, the wild-type (WT) and mutant resistant viruses induced similar maximum weight loss in mice and ferrets with an identical pyrexic response in ferrets (AUC of 233.9 and 233.2, P = 0.5156). Similarly, comparable titers were obtained for the WT and the mutant strains on days 1, 3, 6 and 9 post-infection in mouse lungs and on days 1–7 in ferret nasal washes. A more important perivascular (day 6) and pleural (days 6 and 12) inflammation was noted in the lungs of mice infected with the H274Y mutant, which correlated with increased pulmonary levels of IL-6 and KC. Such increased levels of IL-6 were also observed in lymph nodes of ferrets infected with the mutant strain. Furthermore, the H274Y mutant strain was transmitted to ferrets. In conclusion, viral fitness of the H274Y pH1N1 isolate is not substantially altered and has the potential to induce severe disease and to disseminate.
| During the 2009 pandemic of the novel A/H1N1 (pH1N1) virus, the World Health Organization recommended oseltamivir as first-line agent for treatment of patients with severe infections leading to hospitalization and for those with underlying diseases predisposing to pulmonary complications. Oseltamivir-resistant isolates started to emerge at the end of June 2009 with now more than 100 strains reported worldwide including a few outbreaks where transmission of resistant viruses may have occurred. We characterized the fitness of a pair of oseltamivir-susceptible and oseltamivir-resistant strains emerging from the same familial cluster and that differed by only a single change (H274Y) in the neuraminidase protein. We found that the drug-resistant (mutant) virus was at least as virulent as the drug-susceptible (wild-type) virus in mice and ferrets. Based on these data, we believe that the H274Y pH1N1 mutant strain has the potential to disseminate in the population and to eventually replace the susceptible strain, a phenomenon that has been already observed with seasonal A/Brisbane/59/2007-like (H1N1) viruses.
| The novel influenza A (H1N1) virus was initially detected in Mexico and California in April 2009 and then officially became the first pandemic influenza virus of the 21st century on June 11, 2009 [1], [2]. Most confirmed cases of pandemic A/H1N1 (pH1N1) infection have been characterized so far by self-limited flu-like symptoms and signs although a significant proportion of infected patients also presented with vomiting and diarrhea [2]. A minority of cases, notably those involving pregnant women, have been associated with a more severe clinical outcome leading to intensive care admission and death [3], [4], [5]. Mouse, ferret and non-human primate studies have indicated that pH1N1 isolates replicate more efficiently and produce more severe pathological lesions in the lungs than recent human A/H1N1 viruses [6], [7], [8]. Seroprevalence studies have indicated that children were initially serologically naïve to the novel pH1N1 strain whereas some degree of pre-existing immunity to this virus existed in the elderly population [6], [9], [10].
Antivirals are the cornerstone of treatment for severe influenza cases requiring hospitalization and can also be used as prophylactic agents in high-risk individuals. Early reports demonstrated that pH1N1 strains were resistant to the adamantanes due to a S31N mutation in the M2 gene but remained susceptible to neuraminidase inhibitors (NAIs) such as oseltamivir and zanamivir [6], [11]. However, oseltamivir resistance has been on the rise in recent seasonal influenza A/H1N1 viruses. Indeed, during the 2008–09 influenza season, almost all characterized influenza A/Brisbane/59/2007-like (H1N1) strains from North America and Europe were resistant to oseltamivir due to a H274Y (N2 numbering) mutation in the neuraminidase (NA) gene [12], [13], [14]. The sudden and large dissemination of this mutant A/H1N1 virus occurred in the apparent absence of antiviral pressure suggesting that it had no impairment in viral fitness. This drug resistance mutation has also been reported in some A/H5N1 viruses [15], [16] and, more recently, in several pH1N1 strains recovered from both immunocompromised and immunocompetent subject [17], [18], [19], [20]. We recently reported the emergence of such an oseltamivir-resistant H274Y mutant in a familial cluster of pH1N1 infections [21]. In this outbreak, we identified a drug-susceptible virus recovered before therapy from a 13-year old boy and a drug-resistant virus collected a few days later from his father who was receiving oseltamivir prophylaxis. We now describe the in vitro and in vivo replicative characteristics of the drug-resistant and wild-type (WT) viruses isolated from this outbreak.
As shown in Table 1, the pH1N1 isolate from the index case collected before oseltamivir therapy (A/Québec/147023/2009-WT) was susceptible to all NAIs whereas the pH1N1 isolate from the contact case recovered during post-exposure oseltamivir prophylaxis (A/Québec/147365/2009-H274Y) was resistant to oseltamivir and peramivir. Both isolates were susceptible to zanamivir and A-315675 similarly to 20 other pH1N1 isolates collected from untreated subjects in the same period. The pattern of NAI resistance of the pH1N1 H274Y mutant was similar to that of another H274Y mutant from a seasonal A/H1N1 strain (A/Brisbane/59/2007-H274Y). A pH1N1 H274Y recombinant mutant virus generated from an unrelated pH1N1 strain also exhibited high levels of resistance to oseltamivir and peramivir but remained susceptible to zanamivir and A-315675 (Table 1).
Sequence analysis of the original clinical isolates revealed the presence of only one substitution (H274Y; N2 numbering) in the NA gene of the contact case (GenBank accession number FN434454) compared to that of the index case (accession number FN434445). There was no change in the remaining 7 segments between these two strains (accession number FN434440 to FN434447 for the index case virus and FN434448 to FN4456 for the contact case virus). Phylogenetic analysis of the NA and HA genes showed that the two pH1N1 isolates described in this study were closely related to pH1N1 strains identified in North America, Europe and Asia (data not shown). The viral populations in the two clinical isolates were homogenous as 100% (16/16) of clones from the index case had the H274 sequence whereas 100% (16/16) of clones from the contact case harboured the 274Y sequence.
In vitro experiments performed in MDCK cells expressing the α2,6 sialic acid receptor indicated that the oseltamivir-resistant pH1N1 isolate replicated less efficiently than the WT pH1N1 during the first 24 h. However, there was no significant difference in viral titers subsequently i.e. from 36 to 72 h (Figure 1). The two pH1N1 isolates produced lower viral titers than seasonal A/H1N1 viruses (A/Brisbane/59/2007) including both a WT and a H274Y mutant at 36 and 48 h. Thus, the H274Y mutation resulted in either no impairment or only initial reduction in replicative capacities when inserted in seasonal and pandemic A/H1N1 backgrounds, respectively. Of note, the two pH1N1 viruses produced less well defined viral plaques on α2,6-transfected MDCK cells compared to seasonal strains (data not shown).
Two separate mouse experiments were conducted to assess weight loss, clinical signs, viral titers (on days 3 and 6 in the first experiment and on days 1, 6 and 9 in the second experiment) and histopathological changes. In the first experiment, the WT and oseltamivir-resistant pH1N1 isolates induced similar maximum weight loss, which peaked on day 8 at 16.3% for both groups (P = 0.81) although there was a more pronounced weight loss from days 3 to 7 with the mutant strain (Figure 2). In the second experiment, more weight loss was induced after infection with the H274Y mutant from days 3 to 8. By day 12, all mice from the two experiments had returned to their initial weight with no mortality. Lung viral titers, which were determined on days 1, 3, 6 and 9 post-infection, did not significantly differ between the WT and H274Y mutant viruses when assessed by quantitative viral culture (Figure 3) and real-time RT-PCR (Figure S1). Importantly, there was no unexpected change in the NA sequence of viruses recovered from lungs of euthanized mice.
Transcript levels for various cytokines/chemokines (KC [CXCL1], MCP-1 [CCL2], MIP-1α, IFN-γ, IL-4, IL-5, IL-6 and IL-10) were determined in lungs of infected mice on days 1, 6 and 9 post-infection. All cytokines/chemokines were equally expressed following infection with either of the two pH1N1 isolates (Figure S2), with the exception of increased expression of IL-6 and KC levels on day 1, following infection with the H274Y mutant virus (Figure 4).
Both pH1N1 isolates induced significant pulmonary inflammation including peribronchial, interstitial, perivascular, alveolar and pleural inflammation that peaked on day 6 post-infection (Figure S3). There was significantly more perivascular (day 6) and pleural (days 6 and 12) inflammation visualized in the lungs of mice infected with the H274Y mutant compared to the WT virus (Figure 5). A mild to moderate vascular congestion was observed in both groups of mice although pulmonary oedema was not noted in any mice. Inflammatory cellular infiltration was characterized by both acute (neutrophilic) and chronic (lymphohistiocytic) infiltrates in all mice. Thus, mouse experiments indicated that the mutant pH1N1 isolate induced more pronounced weight loss than the WT virus which correlated with increased expression of IL-6 and KC and more significant lung inflammation despite similar lung viral titers.
Intranasal inoculation of ferrets with the WT and H274Y mutant pH1N1 isolates resulted in a strong anti-A/California/07/2009 serum antibody response on day 14 (hemagglutination inhibition reciprocal geometric mean titers went from <20 to 4208 and from <20 to 3135, respectively). Notably, all ferrets had preexisting HI antibodies against seasonal A/H1N1 (A/Brisbane/59/07) but titers were similar in the two groups of ferrets pre- and post-infection (Table S1). A pyrexic response was seen between days 2 and 8 post-inoculation (Figure 6). Interestingly, temperature curves were biphasic with a major peak on days 2–3 and another lower peak on days 5–6 in both groups of ferrets. The area under the curve (AUC) of temperatures over the course of the 14-day experiment was similar for both groups of ferrets i.e. 233.9±0.5787 for the WT and 233.2±0.8669 for the H274Y mutant (P = 0.5156). Also, the mean percentage of body weight loss over time was not significantly different in animals infected with the WT or the H274Y mutant virus (Figure S4). The maximum weight loss (day 7 and day 3) was 7.54% and 4.15% for the WT and H274Y mutant viruses, respectively (P = 0.0515). By the end of the 14-day observation period, the ferrets had returned to their initial weight with no mortality.
Viruses could be recovered from nasal wash of ferrets up to 7 days post-infection with a peak on day 2 post-infection (Figure 7). Viral titers did not significantly vary at any time points when comparing the two groups of ferrets. Increased levels of IL-6, IL-12 and IFN-γ mRNA were observed in retropharyngeal lymph nodes of ferrets infected with the H274Y mutant compared to the WT on day 14 with ratios of 1.174, 1.38 and 1.183, respectively (not shown). Expression of IL-2 was decreased in ferrets infected with the mutant virus compared to the WT with a ratio of 0.8. Thus, ferret experiments showed no significant differences in clinical parameters (temperature and weight) and viral titers in the upper respiratory tract for the WT and mutant pH1N1 isolates. However, some cytokines (IL-6, IL-12 and IFN-γ) were specifically upregulated in lymph nodes of ferrets infected with the H274Y mutant.
A limited transmission study was conducted in ferrets and it demonstrated that the H274Y mutant strain was transmitted from ferrets experimentally infected intranasally to ferrets placed in the same cage 24 h after infection. All contact ferrets seroconverted for A/California/07/2009 when tested 14 days after contact (hemagglutination inhibition reciprocal geometric mean titers went from <20 to 993). All contact ferrets also shed virus and had a mean peak viral titer of 8.32×104 PFU/ml in their nasal washes.
Oseltamivir, the most frequently used NAI, is recommended for treatment of patients with severe pH1N1 infections leading to hospitalization or those with underlying diseases which place them at risk of complications. We have recently described the rapid emergence of oseltamivir resistance in a family cluster of pH1N1 infection due to the H274Y NA mutation [21], and now report on the viral fitness of this mutant in vitro and in vivo. Despite slightly reduced propagation of the mutant isolate in vitro during the first 24 h of infection compared to the original WT virus, both isolates replicated as efficiently in the lower respiratory tract of mice and in the upper respiratory tract of ferrets inducing similar maximum weight loss and pyrexic response between days 2 and 8 post-infection. Interestingly, the H274Y NA mutant induced a slightly more pronounced weight loss than the WT virus in mice, and this observation correlated with increased production of IL-6 and KC and a more important pulmonary inflammation involving the perivascular (day 6) and pleural (days 6 and 12) compartments.
Previous studies have demonstrated that oseltamivir resistance results from subtype-specific NA mutations [15]. In influenza viruses of the N1 subtype, including seasonal A/H1N1 viruses and avian A/H5N1 strains, oseltamivir resistance is mainly conferred by the H274Y (N2 numbering) mutation [16], [22], [23]. This mutation has also been recently detected in pH1N1 viruses [17], [18], [19], [20], [21]. The larger tyrosine residue at codon 274 prevents the re-orientation of glutamic acid at position 276 within the catalytic site, which is required to accommodate the bulky side chain of oseltamivir but not zanamivir [24]. In agreement with previously-described seasonal A/H1N1 viruses containing the H274Y NA mutation, our oseltamivir-resistant pH1N1 mutant showed cross-resistance to peramivir, a parenteral NAI that is in phase 3 clinical trials [15], [25] and also readily available through an emergency access program [26]. On the other hand, our phenotypic findings demonstrate an interesting potential for inhaled zanamivir and the investigational orally-available A-322278 NAI compound (the prodrug of A-315675) [27], as alternative agents for treatment of oseltamivir-resistant pH1N1 infections.
The impact of mutations conferring NAI resistance on viral fitness and transmissibility may vary depending on the genetic background of influenza viruses. In the N1 subtype, the H274Y mutation has been initially reported as impairing the viral fitness of older seasonal strains such as A/New Caledonia/20/99-like and A/Texas/36/91 when evaluated in the ferret model [28], [29]. However, transmission of the H274Y mutant strain was documented in ferrets [28]. More recently, our group showed that the same mutation had a different effect, i.e. it was associated with conserved viral fitness in the seasonal A/Brisbane/59/2007 (H1N1) background when assessed both in vitro and in ferrets [30]. In the present study, the replication of the pH1N1 H274Y mutant was initially impaired in vitro compared to the WT virus but viral titers were virtually identical on days 2 and 3 post-infection. The two pH1N1 isolates replicated less efficiently (lower titers and reduced plaque formation) than the recent WT and H274Y mutant A/Brisbane/59/2007 strains in ST6Gal I-expressing MDCK cells, which may indicate a greater affinity of the seasonal strain for α2,6 sialic acid receptors.
Different groups have reported the use of exprimental animal models for studying WT pH1N1 infection. For instance, Itoh et al. [6] and Maines et al. [7] observed significant weight loss in BALB/c mice associated with efficient viral replication in lungs when an intranasal inoculum ≥104 plaque forming units (PFUs) was used. Some mice even died from a viral challenge consisting of 106 PFUs [6]. Those results indicate that pH1N1 strains can replicate efficiently in mice in contrast to most seasonal A/H1N1 strains. The selective replication of pH1N1 virus in the BALB/c mouse model without prior animal adaptation is reminiscent of features described for highly pathogenic A/H5N1 viruses and the 1918 Spanish flu virus although they are generally less severe with pH1N1 [31], [32]. In agreement with these reports, we observed that pH1N1 can efficiently infect BALB/c mice inducing weight loss, high viral lung titers and also significant pulmonary histopathological changes. Moreover, we now report that the H274Y pH1N1 mutant virus is clearly as fit as the WT virus in this animal model. In fact, more weight loss was induced by the H274Y mutant compared to the WT virus during the first 7–8 days post-infection although all mice eventually returned to their initial weight. On the other hand, we found no significant differences in lung viral titers between the two groups of mice when assessed on days 1, 3, 6 and 9. The more prounounced weight loss observed with the mutant strain compared to the WT virus was confirmed in two separate experiments using similar viral inoculum (as confirmed by back titration) and unaltered viral genomes (as confirmed by sequencing the virus from lungs of euthanized mice). These important clinical signs correlated with slightly more severe histopathological changes observed in lungs of mutant-infected mice in particular on days 6 and 12 post-infection and more specifically in the perivascular and pleural compartments. Altogether, those results suggest that the H274Y pH1N1 mutant isolate stimulated a more important inflammatory response in mice compared to WT virus, which could be due to rapid induction of IL-6 and KC in the former. It has been reported for different influenza viruses that the early secretion of pro-inflammatory cytokines was associated with the development of pulmonary inflammation at a later stage [33] as shown here for pH1N1. Interestingly, the H274Y mutant also induced preferential expression of IL-6, IL-12 and IFN-γ in the retropharyngeal lymph nodes of ferrets compared to the WT virus. In humans, IL-6 has been shown to be the first cytokine to appear in nasal wash of infected individuals and the one likely responsible for much of the clinical symptoms [34], [35], [36]. Additional studies are required to assess the mechanism leading to increased IL-6 levels in animals infected with the H274Y NA mutant.
As shown in some but not all studies [6], [7], [8], our two pH1N1 isolates induced a strong pyrexic response (increased in temperature of 2°C) and slight weight loss (3–7%) in ferrets. Interestingly, we noticed a biphasic temperature curve for both groups of ferrets. The first and major febrile peak correlated with maximum viral titers and the second minor increase in temperature seen on day 6 might be due to cytokine release. Similar biphasic temperature curves were also seen in ferrets infected with WT A/H5N1 viruses (data not shown). Previous investigators have shown that both pH1N1 and seasonal A/H1N1 strains replicate to similar levels in the upper respiratory tract of ferrets but only the former could replicate to high levels in the lungs [6], [7], [8]. As for the comparison of the WT and mutant pH1N1 isolates, we found no significant difference in nasal wash viral titers of ferrets as determined at several time points but did not assess lung viral titers.
Among the strengths of our study is the use of two clinical isolates from the same familial cluster that differed by a single a.a. and that were only passaged twice before animal studies. Also, the use of two different animal models to characterize the virulence of these strains and the relatively similar results observed in both of them reinforce our conclusions. A limitation of our study is the incomplete assessment of the transmissibility of our pH1N1 strains. Some groups have shown that the WT pH1N1 strain can be transmitted efficiently via aerosol or respiratory droplets [6], [8]. We report here that the H274Y mutant could be transmitted by contact to uninfected ferrets but did not compare the efficiency of transmission with the WT strain and neither evaluated aerosol transmission. A confounder is the seropositive status of our ferrets for the seasonal A/Brisbane/59/2007 (H1N1) strain before challenge with pH1N1 viruses. In guinea pigs, preexisting immunity to recent seasonal A/H1N1 viruses reduced viral load and transmission of pH1N1 [37] whereas a ferret study with the seasonal A/H1N1 vaccine showed little protection from challenge with pH1N1 [38]. Although preexisting antibody levels against an heterologous strain could have an effect on pathogenicity and transmission of pH1N1, the geometrical mean antibody titers were similar for our two groups of ferrets, which should not change our conclusion about the relative pathogenicity of the H274Y mutant compared to the WT strain. Furthermore, we found similar results with the two strains in non-immune mice. Finally, our animal results may not be completely relevant to humans due to differences in distribution of HA cell receptors.
In summary, although some slight differences were observed in the two animal models, we can conclude that the H274Y pH1N1 mutant seems as virulent as the WT isolate with no obvious impairment in viral fitness. Although reports of limited person-to-person transmission in several epidemiological settings have been observed [39], currently no evidence of widespread dissemination of oseltamivir-resistant pH1N1 has been reported indicating the continued value of this drug for treatment of severe cases. Other H274Y pH1N1 mutants (with different genetic backgrounds) should be studied in terms of virulence and efficiency of transmission to confirm our conclusions. In the meantime, careful monitoring of the H274Y mutation during pH1N1 outbreaks is mandatory to rapidly identify transmission events that could lead to large-scale dissemination of an oseltamivir-resistant pH1N1 strain.
The two pH1N1 viruses that differed by only a single mutation in the NA gene were recovered from a family cluster of infection and passaged twice in ST6Gal I-expressing MDCK cells before testing. The two seasonal A/H1N1 viruses were collected in 2007 and passaged 3 times before testing.
Susceptibility profiles of pandemic and seasonal influenza A/H1N1 strains as well as recombinant pH1N1 viruses against oseltamivir carboxylate (Hoffmann La Roche, Basel, Switzerland), zanamivir (GlaxoSmithKline, Stevenage, UK), peramivir (BioCryst, Birmingham, AL) and A-315675 (Abbott Laboratories, North Chicago, IL) were determined by NA inhibition assays using methylumbelliferyl-N-acetyl neuraminic acid (MUNANA, Sigma, St. Louis, MO) as a fluorescent substrate [40].
The eight segments of pandemic influenza A/H1N1 virus (A/Quebec/141447/09) were amplified by RT-PCR and inserted into the recently-described bidirectional pLLB-A/G expression/translation plasmids by recombination in E. coli [41]. The pLLBA plasmid containing the NA segment was used in PCR-mediated site-directed mutagenesis kit (Stratagene, La Jolla, CA) for the introduction of the H274Y mutation. Plasmids were then used to cotransfect 293T cells for the rescue of recombinant viruses as previously described [42].
RNA was isolated directly from human nasopharyngeal aspirates or mouse lungs by using the QIAamp Viral RNA kit (Qiagen, Mississauga, ON, Canada). Complementary DNA (cDNA) was synthesized by using 500 ng of the influenza specific Uni12 primer [43] and the SuperScript II reverse transcriptase (GIBCO-BRL, Burlington, ON, Canada). Viral cDNA was used to amplify the eight influenza viral segments by PCR using the Pfu Turbo Polymerase (Stratagene, La Jolla, CA) and primers specific for each influenza gene [43]. PCR products were gel-purified and sequenced using an automated DNA sequencer (ABI Prism 377 DNA sequencer, Applied Biosystems, Foster City, CA). For evaluation of viral quasispecies, cDNAs of the NA gene from the original clinical samples were cloned into the pJET vector (Fermentas, Burlington, ON). Sixteen clones were sequenced to establish the ratios of WT and mutant populations within each clinical strain.
ST6Gal I-expressing MDCK cells (kindly provided by Dr. Y. Kawaoka, University of Wisconsin, WI) [44] were infected at a multiplicity of infection (MOI) of 0.001 with pandemic or seasonal A/H1N1 viruses (A/Brisbane/59/2007) containing or not the H274Y NA mutation. Supernatants were serially collected post-infection and assayed for numbers of PFUs using standard plaque assays.
Six to eight week old female BALB/c mice (Charles River, ON, Canada) were used to evaluate weight loss as well as lung viral titers, cytokines transcript levels and histopathological changes following pH1N1 infection. Anesthetized mice were challenged by intranasal inoculation of 5×105 PFUs in 50 µl virus diluent (Minimal Essential Media (MEM), 0.3% Bovine Serum Albumin (BSA), penicillin/streptomycin) of the WT or H274Y mutant influenza virus isolate. After challenge, animals were weighed daily for 12 days and monitored for clinical signs. On days 1, 3, 6 and 9 post-infection (days varied according to the experiment), 3 mice per group were sacrificed and lung tissue was placed into RNAlater (Qiagen) for RNA preservation and subsequent RNA extraction. Additional samples of fresh tissues were immediately frozen for viral isolation. All procedures were approved by the Institutional Animal Care Committee at the National Microbiology Laboratory (NML) of the Public Health Agency of Canada (PHAC) according to the guidelines of the Canadian Council on Animal Care. All infectious work was performed in biocontainment level 4 at the NML.
Lung tissues were harvested during necropsies and homogenized in MEM/BSA using a bead mill homogenizer (Tissue Lyser, Qiagen). Debris was pelleted by centrifugation (2,000 g, 5 min) and 10-fold serial dilutions of supernatant were plated on MDCK cells with six replicates per dilution. At 72–96 h post-infection, the plates were scored for cytopathic effects (CPE) and the TCID50 virus titers were calculated using the method of Reed and Muench [45]. Tissues preserved in RNAlater were homogenized using a bead mill homogenizer and RNA was isolated using the RNeasy Mini Kit (Qiagen). Pandemic H1N1 copy numbers were determined by Q-RT-PCR using the LightCycler 480 RNA Master Hydrolysis Probes (Roche Diagnostics, Laval, QC) assay targeting the hemagglutinin gene (nt position 714–815, GenBank number GQ160606). Reaction conditions were the following: 63°C – 3 min, 95°C – 30 s and cycling of 95°C – 15 s, 60°C – 30 s for 45 cycles using a Lightcycler 480 (Roche). The lower detection limit for this pH1N1 assay is 0.1 PFU/ml. The primer/probe sequences are as follow: HAForward– GGATCAAGAAGGGAGAATGAACTATT; HAReverse – AATGCATATCTCGGTACCACTAGATTT and HAProbe (TET) – CCGGGAGACAAAA-TAACATTCGAAGCAAC.
For measuring cytokines expression on days 1, 6 and 9, extracted RNA was first analyzed with the RNA 6000 Nano LabChip and Bioanalyser (Agilent, Switzerland). cDNA was prepared using RNA of standardized quality (RIN>8) and quantity (3.68 µg of total RNA), the Superscript II RNase H (Invitrogen, Burlington, ON, Canada) and 250 ng of random primer hexamers (Invitrogen). Equal amounts of cDNA were run in triplicate and amplified/detected using the Amplifluor UniPrimer system (Applied Biosystem, Foster City, CA) in which the forward or the reverse primers are tailed with the Z sequence 5′-ACTGAACCTGACCGTACA. The results were normalized based on amplification of an internal gene (18S ribosome) and amounts of target gene were calculated according to a standard curve. The primer sequences for the cytokines/chemokines IL-4, IL-5, IL-6, IL-10, KC, MCP-1, MIP-1α and IFN-γ are available upon request [46].
For lung histopathological studies, one pulmonary lobe was removed at serial times and fixed with 10% buffered formalin. Tissues were embedded in paraffin, sectioned in slices of 4 µm and stained with hematoxylin-eosin. The histopathological scores (HPS) were determined by two independent pathologists with experience in pulmonary pathology who were unaware of the infection status of the animals. A semi-quantitative scale was used to score bronchial/endobronchial, peribronchial, perivascular, interstitial, pleural and intra-alveolar inflammation [47]. Capillary vascular congestion and pulmonary edema were also evaluated using a semi-quantitative scale and the inflammatory cellular infiltrate was characterized to determine if the inflammation was acute (neutrophilic) or chronic (lymphohistiocytic).
Groups of five male ferrets (900–1500 g) (Triple F Farms, Sayre, PA) were lightly anesthetized by isoflurane and received by intranasal instillation 250 µl (125 µl/nare) of PBS containing 4.5log TCID50/ml of pH1N1 viruses with or without the H274Y NA mutation. Telemetric transmitters (DST micro-T, Star-Oddi, Iceland) were subcutaneously implanted and temperature profiles of ferrets were recorded every 15 min starting 2 days prior and up to 14 days post-inoculation. Ferrets were weighed daily and nasal wash samples were collected from animals on a daily basis during 14 days post-inoculation by instillation of 5 ml of PBS into the external nares of the animals. The work was performed in biocontainment level 2+ according to procedures approved by the Institutional Animal Care Committee of Armand Frapier Institute.
Virus titers were determined by standard plaque assays using ST6Gal I-expressing MDCK cells. In addition, serum was also collected from each ferret before intranasal infection and on day 14 post-infection to evaluate specific antibody levels against the pH1N1 A/California/07/2009 and the seasonal A/Brisbane/59/2007 strains using standard hemagglutination inhibition assays.
For transcripts analysis of IL-2, IL-6, IL-12 and IFN-γ on day 14 post-infection, RNA was isolated from retropharyngeal lymph nodes using the RNAqueous Micro (Ambion, Streetsville, Ontario) and cDNA was generated with the Transcriptor First Strand cDNA Synthesis Kit (Roche). Amplification of cytokine cDNA was performed as previously described [48]. The results were normalized based on amplification of an internal gene (GAPDH).
Five male ferrets (900–1500 g) (Triple F Farms, Sayre, PA) were lightly anesthetized by isoflurane and received by intranasal instillation 250 µl (125 µl/nare) of PBS containing 4.5log TCID50/ml of pH1N1 virus with the H274Y NA mutation. Each ferret was placed individually in a cage. Approximately 24 h following viral infection, inoculated-contact animal pairs were established by placing a naïve ferret into each cage allowing the exchange of respiratory droplets by direct contact. Inoculated and contact animals were monitored for clinical signs and nasal wash samples were collected for viral titers every day over a 14-day period. Serum was also collected from each ferret before intranasal infection and on day 14 post-infection to evaluate specific antibody levels against the pH1N1 A/California/07/2009 virus using standard hemagglutination inhibition assays. All ferrets were seronegative for pH1N1 A/California/07/2009 before intranasal infection.
Paired and unpaired t test analyses were done to compare the mutant and WT virus characteristics during in vitro and in vivo studies, respectively.
|
10.1371/journal.ppat.1002726 | Hsp90 Interacts Specifically with Viral RNA and Differentially Regulates Replication Initiation of Bamboo mosaic virus and Associated Satellite RNA | Host factors play crucial roles in the replication of plus-strand RNA viruses. In this report, a heat shock protein 90 homologue of Nicotiana benthamiana, NbHsp90, was identified in association with partially purified replicase complexes from BaMV-infected tissue, and shown to specifically interact with the 3′ untranslated region (3′ UTR) of BaMV genomic RNA, but not with the 3′ UTR of BaMV-associated satellite RNA (satBaMV RNA) or that of genomic RNA of other viruses, such as Potato virus X (PVX) or Cucumber mosaic virus (CMV). Mutational analyses revealed that the interaction occurs between the middle domain of NbHsp90 and domain E of the BaMV 3′ UTR. The knockdown or inhibition of NbHsp90 suppressed BaMV infectivity, but not that of satBaMV RNA, PVX, or CMV in N. benthamiana. Time-course analysis further revealed that the inhibitory effect of 17-AAG is significant only during the immediate early stages of BaMV replication. Moreover, yeast two-hybrid and GST pull-down assays demonstrated the existence of an interaction between NbHsp90 and the BaMV RNA-dependent RNA polymerase. These results reveal a novel role for NbHsp90 in the selective enhancement of BaMV replication, most likely through direct interaction with the 3′ UTR of BaMV RNA during the initiation of BaMV RNA replication.
| Heat shock protein 90 (Hsp90) is a highly conserved molecular chaperone in prokaryotes and eukaryotes, and regulates diverse cellular processes through ensuring the correct folding of numerous client proteins. However, there are no reports of direct interactions between Hsp90 with viral RNA. Here, we report that a new member of the Hsp90 proteins of Nicotiana benthamiana, NbHsp90, specifically interacts with the 3′ UTR of Bamboo mosaic virus (BaMV) genomic RNA, but not with the 3′ UTR of BaMV-associated satellite RNA or that of other viruses. We further demonstrate that NbHsp90 specifically involves in the immediately early stage of BaMV RNA replication. NbHsp90 directly interacts with the BaMV 3′ UTR through the domain E, the key structural differences that distinguishes the BaMV 3′ UTR from the satBaMV 3′ UTR, which might contribute to NbHsp90's differential requirement for BaMV and satBaMV replication. Our work revealed a new role of Hsp90 in the interaction with RNA molecules, and demonstrated the differential requirement of Hsp90 in the replication of BaMV and satBaMV RNAs, which provide additional leverage for understanding the complex interactions between host, virus and its associated satellite RNA.
| Viruses have limited coding capacity and require a multitude of host factors to support their biological functions during the infection cycle [1]–[4]. Researchers have used various experimental approaches in their search for host factors specifically required for viral replication. These approaches have included genome-wild screening of host factors that affect viral replication using yeast mutants [5]–[8], the direct identification of host proteins co-purified with viral replicase complexes [9]–[11], and the capturing of host proteins that bind specifically to viral RNA [12], [13]. The 3′-untranslated regions (3′ UTRs) of viral genomic RNAs contain cis-acting sequences required for the initiation of minus-strand RNA synthesis during replication [14]–[16], and hence 3′ UTRs have been used in several studies as a bait to identify host factors involved in viral replication processes. A number of host factors have been shown to physically interact with the cis-acting elements of viral RNA from Brome mosaic virus, Hepatitis C virus (HCV), tombusvirus and Tobacco mosaic virus (TMV), and are required for the efficient replication of these viruses [13], [17]–[20]. Efforts toward identifying and characterizing the various host factors required for viral RNA replication will help shed light on the molecular biology of viruses, and thus provide a valuable resource for the development of antiviral strategies.
Bamboo mosaic virus (BaMV), a member of the genus Potexvirus, has a single-stranded, positive-sense RNA genome of approximately 6,400 nt [excluding the 3′ poly(A) tail] with a 5′ cap-structure and a 3′ poly(A) tail. The five open reading frames (ORFs 1 to 5) of the BaMV genome encode ORF1 protein (155 kDa), TGBp1 (28 kDa), TGBp2 (13 kDa), TGBp3 (6 kDa), and the capsid protein (25 kDa), respectively [21], [22]. The BaMV ORF1 protein consists of three functional domains: the capping enzyme domain [23], [24], the helicase-like domain [25], and the RNA-dependent RNA polymerase (RdRp) core domain [26]. The 3′ UTR from BaMV genomic RNA can fold into four independent stem-loops (domains A to D) and a tertiary pseudoknot structure (domain E), both of which are important for BaMV replication [27]. In addition to genomic and subgenomic RNAs, the satellite RNA associated with BaMV (satBaMV RNA) is dependent on BaMV for its replication, encapsidation, and systemic movement [28]. The satBaMV RNA is 836 nt in length [excluding the 3′ poly(A) tail] and has an ORF for a 20-kDa protein [28]. The satBaMV RNA 3′ UTR structure comprises three stem-loops, SLA, SLB, and SLC, and two cis-acting elements responsible for efficient replication [29]. Although the satBaMV 3′ UTR sequence and structural elements mimic those of the BaMV 3′ UTR for recognition by the BaMV RdRp complexes, discrepancies between these two 3′ UTRs are significant, most notably the lack of a pseudoknot structure in the satBaMV RNA 3′ UTR [29], [30]. Thus, whether satBaMV RNA shares the same set of replicase complexes with BaMV or uses another set of replicase complexes (possibly with different host factors) for replication remains an open question.
Heat shock protein 90 (Hsp90) is a highly conserved molecular chaperone in prokaryotes and eukaryotes, and regulates diverse cellular processes through ensuring the correct folding of numerous client proteins [31], [32]. Hsp90 client proteins are involved in signal transduction, steroid signaling, protein trafficking, and stress response [32]–[34]. There is increasing evidence that many viruses recruit the Hsp90 chaperone function for assistance with viral protein synthesis, maturation, and stabilization [35]. For example, Hsp90 is essential for Flock house virus RNA polymerase synthesis and for efficient infection in Drosophila cells [36], [37]. Hsp90 is also essential for maturation of HCV non-structural protein NS2/3, stability of NS3, and the assembly of replicase complexes [38], [39]. Hsp90 plays a role in nuclear import and assembly of the Influenza A virus polymerase complex by binding to PB1 and PB2 polymerase subunits [40], [41]. Activated Hsp90 binds to hepatitis B virus core protein and facilitates capsid formation [42]. The dengue virus receptor-complex, comprising Hsp90 and Hsp70, is important for virus entry [43]. Another molecular chaperone family, the Hsp70 proteins, plays a key role in replication of plant viruses such as tombusvirus [11], [44], [45]. To date, however, there has been no evidence of a direct interaction between Hsp90 and viral RNA or for involvement of such an interaction in viral RNA replication.
In this article, we report a direct and specific interaction between Hsp90 and the 3′ UTR of BaMV RNA, and illustrate a novel and required function for Hsp90 during the early stages of BaMV replication by using virus-induced gene silencing and Hsp90-specific inhibitors. Our findings reveal a unique and specific role for Hsp90 in viral RNA recognition and replication, and suggest that replicase complexes recruit different host factors for replication of BaMV and satBaMV. We propose a model to illustrate the stages at which Hsp90 is most likely involved in BaMV RNA replication, and to further describe the differential requirement of Hsp90 in the replication of BaMV and satBaMV RNAs.
Results from previous studies suggested that satBaMV evolved a distinct 3′ UTR RNA structure compared to that of BaMV [29]. BaMV 3′ UTR comprises four stem-loops and one pseudoknot, while satBaMV 3′ UTR comprises three stem-loops (Figures S1A and S1B) [27], [29]. This distinction would suggest more efficient replication and the requirement of different host factors for replication. To test these hypotheses, we performed UV cross-linking assays to examine the binding of components in BaMV RdRp preparations with BaMV, satBaMV, and CMV 3′ UTRs (Figure 1A). The CMV 3′ UTR, containing a tRNA-like structure absent from both the BaMV and satBaMV 3′ UTR, was used as a control RNA for checking binding specificity (Figure 1A, lane 4). Figure 1A shows differential binding of proteins in BaMV RdRp preparations with BaMV and satBaMV 3′ UTRs. Three proteins with estimated relative molecular weights of about 70, 87, and 160 kDa (termed p70, p87, and p160, respectively) were found to have high affinity for the BaMV 3′ UTR (Figure 1A, lane 2, indicated by arrows). To test the specificity of the interaction between these proteins and the BaMV 3′ UTR, UV cross-linking assays were performed in the presence of various unlabeled competitor RNAs (Figure 1B). The intensities of p87 bound with the BaMV 3′ UTR were quantified and the percentages of the binding signals relative to that in the absence of competitor RNA (Figure 1B, lane 1) were plotted. The unlabeled BaMV 3′ UTR was an effective competitor; it was able to compete out about 70% of the labeled BaMV 3′ UTR probe at 10-fold molar excess (Figure 1B, lane 3). By contrast, the satBaMV 3′ UTR, the CMV 3′ UTR, and poly(IC) were unable to compete with the BaMV 3′ UTR probe efficiently, even at 50-fold molar excess (Figure 1B, lanes 7, 10 and 16). These results indicated that p70, p87, and p160 specifically interact with the BaMV 3′ UTR, but not with the 3′ UTRs of satBaMV and CMV, and not with double-stranded RNA. One of the major differences between the BaMV and satBaMV RNA 3′ UTR structures is the presence of a pseudoknot structure (domain E) upstream of the poly(A) tail in BaMV RNA (Figure S1A), and thus domain E of the BaMV 3′ UTR might be a specifically recognized target by these proteins. To test this possibility, poly(A) and the satM10 3′ UTR, in which the BaMV pseudoknot structure is incorporated into the satBaMV 3′ UTR [29], were also used as competitors in binding assays. Findings showed that the satM10 3′ UTR had greater competition efficiency than poly(A) did (Figure 1B, lanes 11–13 and 17–19), suggesting that the three host proteins interact with the BaMV pseudoknot more efficiently than with the poly(A) tail.
To identify p87, we sliced a protein band corresponding to the binding signal from Coomassie blue-stained gel, and analyzed the sample by MALDI-TOF MS spectrometry. Mis-Fit and Mascot software yielded 10 major protein hits, 9 of which matched to Hsp90. Thereafter, p87 was identified as one of the Hsp90 family proteins of N. benthamiana and thus designated as NbHsp90. There are three publicly available full-length coding sequences of known N. benthamiana Hsp90 family proteins, namely NbHsp90-1, -2, and -3, which are reported as AY368904, AY368905, and GQ845021, respectively [46], [47]. Among the three N. benthamiana Hsp90 proteins, the NbHsp90 identified in this study shared the highest amino acid sequence similarity with NbHsp90-2, with an E-value of 1.28×10−6 and a total of 23 NbHsp90 derived peptides as identified by MALDI-TOF MS, and covering 28% of the full-length NbHsp90-2 (Figure S2, indicated by black lines). Subsequently, we designed primers based on the NbHsp90-2 nucleotide sequence and then amplified and cloned the full-length NbHsp90 coding region into pGEX-4T-1 to express NbHsp90 in E. coli as a GST-fusion protein, designated GST-Hsp90. The full-length amino acid sequence of NbHsp90 was determined and aligned with those of NbHsp90-1, -2, and -3 using CLUSTALW [48] (Figure S2). The sequence identities between NbHsp90 and NbHsp90-1, -2, and -3 were 96%, 99%, and 99%, respectively, indicating that there is a high conservation of amino acid sequences among the NbHsp90 isoforms. Thus, the NbHsp90 protein cloned in this study appears to be a new member of the N. benthamiana family of Hsp90 proteins with features shared among other previously known members.
To test whether NbHsp90 interacts directly with the BaMV 3′ UTR, UV cross-linking assays were performed with purified GST-Hsp90 (Figure 2A). A binding signal with relative molecular weight of 107 kDa corresponding to GST-Hsp90 was observed only in samples containing both GST-Hsp90 and the BaMV 3′ UTR (Figure 2A, lanes 3–7). The GST tag alone did not interact with the BaMV 3′ UTR (Figure 2A, lane 2), confirming that Hsp90 is responsible for the binding of GST-Hsp90 to the BaMV 3′ UTR. By contrast, GST-Hsp90 did not interact with the CMV or satBaMV RNA 3′ UTRs, further demonstrating that there is a binding specificity between NbHsp90 and BaMV 3′ UTR (Figure 2A, lanes 8 and 9). The satBaMV and CMV 3′ UTRs did not outcompete the BaMV 3′ UTR for binding with GST-Hsp90 even at 50-fold molar excesses (Figure 2B, lanes 7 and 10). BaMV and satM10, whose 3′ UTRs contain a pseudoknot structure, competed effectively with BaMV 3′ UTR probe (Figure 2A, lanes 4–7; Figure 2B, lanes 2–4 and 14–16). The deletion of domain E in the BaMV 3′ UTR (Ba 3′ UTRdE) reduced the competitive activity to about 25% of that of the intact BaMV 3′ UTR (Figure 2B, lanes 11–13), further demonstrating that domain E in the BaMV 3′ UTR is responsible for NbHsp90 binding.
To further investigate the various molecular interactions here, we used Northwestern hybridization to determine which NbHsp90 domain is responsible for binding to BaMV 3′ UTR. NbHsp90 was divided into N-terminal (N), middle (M), and C-terminal (C) domains based on alignment results with yeast Hsp90, whose crystal structure was determined from a complex of an ATP-analogue with co-chaperone p23/Sba1 [49]. The three domains were analyzed for their RNA-binding activities with each domain alone or in combination (Figure 2C, left panel). As shown in Figure 2C, specific binding signals at positions consistent with those corresponding to the full-length GST-Hsp90, GST-Hsp90MC and GST-Hsp90M, all of which comprise the NbHsp90 M-domain (Figure 2C, lanes 9, 10, and 12), were detected. By contrast, Hsp90 mutants GST-Hsp90C and GST-Hsp90N, in which the M-domain is deleted, were unable to interact with the 32P-labelled BaMV 3′ UTR probe (Figure 2C, lanes 11 and 13), and neither were the control proteins GST and BSA (Figure 2C, lanes 8 and 14). These results demonstrate the existence of a binding specificity between NbHsp90 M-domain and BaMV 3′ UTR.
To investigate the requirement of NbHsp90 for BaMV RNA accumulation in planta, the TRV-based VIGS system was used to generate NbHsp90-knockdown plants. Phytoene desaturase (PDS)-knockdown plants, which exhibit the photo-bleaching phenotype 7 days post agro-infiltration (dpai), were used as indicators for the VIGS process. To examine the RNA silencing of NbHsp90, the relative level of Hsp90 mRNAs was examined at 7 dpai by real-time PCR and semi qRT-PCR (Figures 3A and S3A). It was clear that the Hsp90 mRNA level of silenced plants, TRV1/TRV2-Hsp90, was reduced to about 10% of that of the control treatment, TRV1/TRV2 (Figure 3A). To investigate the impact of Hsp90 knockdown on BaMV RNA accumulation in plants, at 7 dpai the third and fourth leaves above the infiltrated leaves were individually inoculated with BaMV, PVX, and CMV. BaMV RNA accumulation in Hsp90-knockdown plants (TRV1/TRV2-Hsp90) was reduced to 21% and 31% of levels in the control plants at 2 and 6 dpi, respectively (Figure 3B, compare lane 4 to lane 3 and lane 9 to lane 8; Figure S3B). There was no significant alteration observed in the control plants (TRV1/TRV2) or PDS-knockdown plants (TRV1/TRV2-PDS) (Figure 3B, compare lanes 3 and 5 to lane 2, and lanes 8 and 10 to lane 7) indicating that reduction in BaMV infectivity was not due to agrobacterium or TRV infection. By contrast, the accumulation of PVX and CMV RNA was not significantly affected in Hsp90-knockdown plants (Figures 3C, 3D, and S3B), revealing that the environment for virus infection in these Hsp90-knockdown plants remained unaltered. These results suggest that NbHsp90 is specifically required for the efficient accumulation of BaMV but not for PVX or an unrelated virus such as CMV.
The involvement of Hsp90 in the replication of animal viruses has been reported in several studies [40], [50], [51], albeit with no evidence of a direct interaction between Hsp90 and viral RNAs. To corroborate the possibility that NbHsp90 may facilitate BaMV RNA replication, we examined the effect of Hsp90 inhibitors on the accumulation of BaMV RNA. Two Hsp90-specific inhibitors, GA and the GA analogue 17-AAG, were used in this study. GA and 17-AAG bind to the Hsp90 N-terminal ATP-binding site where they block the Hsp90 chaperone function [52]. To test whether the blocked Hsp90 chaperone function interfered with the accumulation of viral RNA in protoplasts, GA or 17-AAG was added to the isolated N. benthamiana protoplasts for 30 min before inoculation. After incubation, 1 µg BaMV RNA was inoculated into the protoplasts and cultured in media containing 0.2 to 2 µM of GA or 17-AAG. Treatment of the protoplasts with the solvent (DMSO) alone served as a negative control and was used for normalization. The results revealed that BaMV RNA accumulation was significantly reduced for both GA and 17-AAG treatments in a dose-dependent manner, which suggests that accumulation of BaMV RNA depends on the Hsp90 function (Figures 4A and S4). DMSO alone had no effect on BaMV RNA accumulation (Figure 4A, compare lane 3 to lane 2). The inhibitory effect of 17-AAG on BaMV RNA replication was more pronounced than the inhibitory effect of GA. At a low concentration (0.2 µM), 17-AAG reduced BaMV RNA accumulation to about 30% of that in the DMSO-treated cells, whereas GA did not efficiently interfere with BaMV replication (Figure 4A, lanes 4 and 8; Figure S4). Thus, 17-AAG was used alone in subsequent experiments. To investigate the effect of NbHsp90 inhibition on the replication of other potexvirus and unrelated virus, 1 µg each of PVX and CMV RNAs was used to inoculate protoplasts individually using a similar treatment as that used for BaMV infection. The accumulation of PVX and CMV RNA was not significantly affected in GA-treated or 17-AAG-treated protoplasts (Figures 4B, 4C, and S4), which also revealed that these concentrations of Hsp90 inhibitors were not toxic to the cells. These findings are consistent with previous observations involving Hsp90-knockdown plants, again suggesting that the NbHsp90 function is specific for BaMV RNA replication.
We found that NbHsp90 interacts with the BaMV 3′ UTR but not with the satBaMV 3′ UTR, implying that NbHsp90 may be required for BaMV replication but not for satBaMV RNA replication. To test this hypothesis, BaMV and satBaMV RNA infectious clones, pCB and pCBSF4 respectively, were co-inoculated into N. benthamiana protoplasts pretreated with 2 µM 17-AAG. We found that NbHsp90 inhibition reduces both BaMV and satBaMV RNA accumulation in protoplasts (Figure 5, lanes 3 and 4). Since satBaMV RNA depends on BaMV for replication, we thought that the reduction of satBaMV RNA accumulation might have resulted from repression of BaMV replication. To examine the requirement of Hsp90 in the accumulation of satBaMV RNA, we introduced the plasmid pBaORF1 into N. benthamiana protoplasts for transient expression of BaMV ORF1 protein which served as the helper for satBaMV RNA replication. As shown in lane 5 of Figure 5, the expression of ORF1 alone was able to support the replication of satBaMV RNA. By contrast, mutant pBaORF1dGDD, which encodes a nonfunctional form of ORF1 with the deletion of the GDD motif in the RdRp domain, failed to support satBaMV RNA replication in trans (Figure 5, lane10). These results confirmed that the native BaMV ORF1 is responsible for satBaMV RNA replication in protoplasts. The BaMV free satBaMV replication system was further adopted to analyze whether Hsp90 is required for satBaMV RNA replication. As shown in Figure 5 and Figure S5, the inhibition of Hsp90 resulting from increased concentrations of 17-AAG did not interfere with satBaMV RNA accumulation (Figure 5, lanes 5–9; Figure S5), suggesting that Hsp90 is not required for satBaMV RNA replication.
Hsp90 has been suggested to play an important role in the early stages of virus infection, either for efficient polymerase synthesis or for assembly of replicase complexes [36], [40]. To study the role of Hsp90 in BaMV replication in terms of different stages in the infection cycle, N. benthamiana protoplasts were treated with 17-AAG at various time points before or after BaMV infection. As illustrated in Figure 6A, the protoplasts were treated with 2 µM 17-AAG at 0.5 hour ante inoculation (hai), or at 0, 2, 4, and 8 hour post inoculation (hpi) of BaMV RNA and harvested at 24 hpi. For treatment at 0.5 hai, 17-AAG was washed with inoculation buffer prior to BaMV inoculation. Northern blot analysis revealed that BaMV RNA accumulation was significant reduced when 17-AAG was added before BaMV infection (0.5 hai) and during the early stages (within 4 h) after BaMV infection (Figure 6A, lanes 3–6). Protoplasts treated with 17-AAG at 8 h post BaMV infection showed no significant inhibitory effect (Figure 6A, lane 7). These observations suggest that the Hsp90 function is only required during the initial stages of BaMV replication and thus may have no impact on RNA synthesis in the later stages of replication.
17-AAG's significant inhibitory effect on BaMV RNA accumulation during the early stages of BaMV infection suggests that this inhibitory effect occurs before the maturation of BaMV replicase complexes. To test this hypothesis, a preparation of BaMV RdRp complexes purified from BaMV-infected leaves [53], [54], was used in an in vitro RdRp assay using both endogenous and exogenous RNA as templates. GA or 17-AAG was added to the BaMV RdRp preparations 30 minutes before the addition of [α-32p] UTP to initiate the endogenous RdRp assay. Newly synthesized RNA emitted radioactive signals, indicating BaMV RdRp activity. The assay showed that neither GA nor 17-AAG inhibited the synthesis of BaMV RNA in vitro (Figure 6B), suggesting that Hsp90 inhibition does not interfere with the RNA synthesis activity of well-assembled replicase. In the exogenous RdRp assay, endogenous RNA templates associated with the replicase complexes were removed prior to the addition of (+) or (−) BaMV 3′ terminal RNA, BaMV 3′ UTR or BaMV (−)77, as templates to assay RdRp activities on the RNA template recognition and initiation of negative- or positive-sense RNA synthesis. The results of exogenous RdRp assay revealed that 17-AAG did not interfere with both either negative- or positive-sense RNA synthesis in vitro (Figure 6C), even at a concentration of 2 µM. This finding supports the hypothesis that Hsp90 is involved in the initial stages of BaMV infection, and that this presumably occurs before or during the assembly process of replicase complexes.
We found that Hsp90 inhibitors GA and 17-AAG, which bind to the Hsp90 N-terminal ATP-binding site and block the chaperone function of Hsp90, interfere with BaMV replication, but do not interfere with PVX or CMV replication (Figure 4). Thus, it is conceivable that NbHsp90 might play a chaperoning role in BaMV replication. To investigate whether or not the BaMV ORF1 protein is a client protein of NbHsp90, we used a yeast two-hybrid system to analyze the interaction between BaMV ORF1 protein and NbHsp90. The three functional BaMV ORF1 protein domains—the capping enzyme domain, helicase-like domain, and RdRp core domain—were individually fused to the Gal4 DNA-binding domain (BD) to probe the interaction of NbHsp90 fusion to the Gal4 activation domain (AD). Positive interactions activate His3 expression in yeast, resulting in colony formation on plates in the absence of histidine (-Trp-His/Zeo300), and also activate LacZ, which causes an increase of β-galactosidase activity during filter assay (β-Gal assay). Plates containing zeocin in the absence of tryptophan (-Trp/Zeo300) were used to screen for successful transformation of plasmid combinations as indicated (Figures 7A and 7B). Yeast cells expressing BD-Fos2 and AD-Jun were used as positive controls. The results showed that the RdRp domain of BaMV ORF1 protein (BD-RdRp) interacts with NbHsp90 (AD-Hsp90), and that the capping enzyme domain and the helicase-like domain do not. The empty vectors (BD plus AD), BD plus AD-Hsp90, and BD-RdRp plus AD served as negative controls to exclude non-specific transactivation in reporter expression (Figure 7A). Reciprocal constructs with exchanges of BD and AD showed similar results (Figure 7B). By contrast, the RdRp domain of PVX ORF1 protein (BD-PVX RdRp) did not interact with AD-Hsp90 (Figure 7B), which is consistent with previous data showing a specific NbHsp90 inhibitory effect on the down-regulation of BaMV replication but not on PVX replication (Figure 4). The results of these yeast two-hybrid assays demonstrate that the BaMV ORF1 protein interacts with NbHsp90 through the RdRp domain. To confirm the interaction between the BaMV RdRp domain and NbHsp90, a GST pull-down assay was performed (Figure 7C). The BaMV RdRp domain and NbHsp90 were expressed and purified in E. coli as a GST fusion protein (GST-RdRp) and a His(6) fusion protein (Hsp90-His(6)), respectively. GST-RdRp was bound to resin and mixed with Hsp90-His(6). After incubation, GST-RdRp was eluted from the resin and Hsp90-His(6) in the eluted sample was analyzed by immunoblotting. Specific signal of Hsp90-His(6) was detected only when both GST-RdRp and Hsp90-His(6) were present (Figure 7C, lane 9). The result further corroborated the observation that BaMV RdRp interacts with NbHsp90 in vitro.
By using the BaMV RNA 3′ UTR as a bait, we identified NbHsp90, a member of the N. benthamiana Hsp90 family proteins, to be a novel host factor associated with the BaMV replicase complexes. NbHsp90 is able to interact with the 3′ UTR of BaMV RNA specifically, and is required for efficient accumulation of BaMV RNA during the early stages of the infection cycle as demonstrated by the use of TRV-based VIGS and Hsp90-specific inhibitors. In contrast, NbHsp90 does not bind to satBaMV RNA or other viral RNAs, nor is it required for their replication. These findings revealed a new role of at least one member of the Hsp90 family proteins as a specific interaction partner of BaMV RNA involved in the efficient accumulation.
In terms of RNA-related functions, members of the Hsp90 family proteins have been shown to be involved in mRNA trafficking. For example, mutational studies showed that Hsp90 is important for the subcellular localization of specific mRNAs to the vicinity of mitochondria, and that the “control elements” for localization reside in the 3′ UTR of these mRNAs [55], [56]. A study on Drosophila melanogaster embryo development also revealed that Hsp90 is required for mRNA localization [57]. Still, there are no previous reports of a direct interaction between Hsp90 and specific mRNAs. The results of the present study represent, to our knowledge, the first direct evidence that NbHsp90 specifically interacts with a viral RNA, thus revealing a novel role for Hsp90 in viral RNA replication. The proper subcellular localization of mRNAs is important to post-transcriptional regulation of gene expression that facilitates the accurate spatial and temporal synthesis of specific proteins in cells [56]. The ability of NbHsp90 to bind specifically to the BaMV 3′ UTR raises the possibility that BaMV may have evolved to exploit the RNA-recognition function of Hsp90, and adopted this mRNA localization mechanism for the specific recruitment of BaMV genomic RNA in order to the properly assemble replicase complexes and thus initiate the replication process.
We show here that NbHsp90 directly interacts with the BaMV 3′ UTR through a pseudoknot domain (domain E) (Figure 2B). This domain is one of the key structural differences that distinguishes the BaMV 3′ UTR from the satBaMV 3′ UTR, which might contribute to NbHsp90's differential requirement for BaMV and satBaMV replication. Similarly, the PVX 3′ UTR does not fold into the pseudoknot structure (Figure S1C) [58] and does not interact with NbHsp90. This finding strongly suggests that the specific enhancement of BaMV replication by NbHsp90 is dependent on the pseudoknot structure present in the BaMV 3′ UTR. Previously, domain E of the BaMV 3′ UTR was also found to interact with viral RdRp and two host proteins, p51 and p43 [59], [60]. The addition of p51 into an RdRp reaction led to the specific down-regulation of BaMV minus-strand RNA synthesis. P43 was identified as a chloroplast phosphoglycerate kinase and as being involved in BaMV replication [60]. Thus, the BaMV 3′ UTR appears to interact with RdRp and a cohort of different host factors for replication and regulation at different stages of the infection cycle. Accordingly, like other viruses [61], the BaMV 3′ UTR may form various distinct structures that serve as ribo-switches to manage numerous events in viral replication, such as template recognition, replicase complexes assembly, and RNA synthesis. Here, we also show that the BaMV 3′ UTR interacts directly with the middle domain of NbHsp90 (Figure 2C), a conserved domain for client protein binding [32]. This suggests an alternative role for NbHsp90 in RNA chaperoning, such as the folding of the BaMV pseudoknot domain. In addition to NbHsp90's ability to bind RNA as identified in this study, other Hsps or Hsp homologues from prokaryotes to eukaryotes have also been shown to interact directly with RNA molecules [62]. For instance, the interaction between an Hsp60 homologue of the thermophilic archaeon Sulfolobus solfataricus and 16S rRNA is required for the cleavage of the rRNA precursor [63]. A conserved plant Hsp101 homologue directly binds to the 5′ end of TMV RNA to act as a translational enhancer [64], and an Hsp40 homologue from yeast interacts directly with nuclear tRNA or ribosomal RNA for individual regulation [65], [66]. Citing these specific interactions, the Hsps were proposed to either function directly as chaperones for mRNA molecules or to support the assembly of RNA-protein complexes in order to stabilize the RNA transcripts, or both [62]. By analogy, NbHsp90 might serve as a chaperone for the folding of the BaMV 3′ UTR structure to modulate its function in BaMV replication. This suggested role for NbHsp90 as a RNA chaperone remains to be explored.
Hsp90 has been proposed to assist in the assembly of viral replication complexes [38], [40], [41], [67] and in their proper localization [40], [41] via interaction with viral replication proteins. In this study, NbHsp90 was shown to interact with the RdRp domain of BaMV ORF1 protein in yeast cells but not with that of PVX (Figures 7A and 7B), implying that NbHsp90 specifically assists in the correct folding of BaMV ORF1 protein and in the assembly of replicase complexes through its chaperone function. This is consistent with the observation that NbHsp90 inhibition interferes with the replication of BaMV but not with that of PVX (Figure 4). Moreover, we propose that the BaMV 3′ UTR may attract NbHsp90 through a specific interaction to facilitate the proper folding and assembly of the newly translated components into functional replicase complexes.
Based on results presented in this study, we propose a model (Figure 8) to illustrate the stages at which Hsp90 might participate in BaMV RNA replication. This model also highlights the differential requirement of Hsp90 for the replication of BaMV and satBaMV RNA, as well as PVX concerning the differences in structure (Figure S1). We suggest that NbHsp90 has at least two distinct functional domains involved in two distinct steps in the early stages of BaMV replication. One domain is for the protein chaperone function and is sensitive to Hsp90 inhibitors GA and 17-AAG (Figure 4). This domain is required for correct folding of the BaMV ORF1 protein and for proper assembly of replicase complexes with other host factors. Once the active replicase complexes are assembled, the GA and 17-AAG inhibitory effects are alleviated (Figure 6B). This is supported by our observation that GA and 17-AAG interference in BaMV replication decreases with increased delays in treatment after BaMV infection (Figure 6A). The other domain is required for specific recognition of domain E of the BaMV 3′ UTR, which recruits the templates into the active replicase complexes and is insensitive to GA and 17-AAG (Figure 6C). In contrast, for satBaMV RNA (Figure 8, on the right), the 3′ terminus of satBaMV RNA may have evolved effective structures that serve as a scaffold for proper folding of replicase, and to facilitate the assembly of active replicase complexes with other host factors. A similar phenomenon occurred when the BaMV ORF1 protein was expressed from the plasmid pBaORF1 (Figure 5; Figure 8, on the far right). Alternatively, satBaMV RNA may hijack the preformed replicase complexes from BaMV genomic RNA for replication. Thus, satBaMV RNA replication is independent of NbHsp90 and is therefore insensitive to the inhibitory properties of GA and 17-AAG (Figure 5).
SatBaMV RNA may employ replicase complexes consisting of host factor(s) distinct from those used by BaMV. The involvement of different replicases or host factors for in cis and in trans replication processes have been previously reported. For example, the genome of Red clover necrotic mosaic virus (RCNMV) comprises bipartite RNAs, RNA1 and RNA2. RCNMV RNA2 does not encode replicase and thus must exploit viral replicase proteins encoded by RNA1 to replicate [9]. Similarly, satRNA depends on the replicase complexes provided in trans by their cognate helper viruses and host plants for replication [68]. In RCNMV, the replicase component p27 is provided in trans and directly interacts with the RNA2 3′ UTR but not with RNA1 [69]. By contrast, only ribosome-bound RNA1 (translating template RNA) interacts with the replicase proteins, suggesting that RCNMV RNA1 is recognized in a coupling between translation and replication. This cis-preferential function of the virus-encoded proteins has been reported for several positive-strand RNA viruses [70]–[73]. Likewise, NbHsp90 may regulate the cis-preferential function of BaMV replicase by specifically binding to the BaMV pseudoknot domain, which is absent in the satBaMV 3′ UTR. Thus, BaMV may retain the replication competency in the presence of satBaMV RNA. It has been demonstrated that ribosomal RNA (rRNAs) serves as a scaffold in ribosomes for the proper positioning and folding of ribosomal proteins [74], [75]. Similarly, satBaMV RNA may have evolved distinct structural elements in the 3′ UTR that could serve as scaffolds for the proper positioning and folding of the BaMV replicase complexes in the absence of NbHsp90. Thus, the differential requirement for NbHsp90 by BaMV and satBaMV RNA might reduce the competition for replication complexes and contribute to their co-existence.
BaMV strain S and the associated satBaMV RNA isolate F4 [29] were used in this study. PVX, a Taiwan strain from infected potato (Liao et al, GenBank acc. AF272736), and CMV strain NT9 [76] were included as the controls. The purifications of BaMV, PVX, and CMV virions were carried out as described previously [77].
Methods for preparing BaMV RdRp and in vitro RdRp assays were described previously [54]. The preparation of exogenous templates for RdRp assay, BaMV 3′ UTR and BaMV (−)77, were described previously [54], [78]. RdRp assay products were analyzed using RNase protection assay, and quantified using a PhosphorImager (FUJIFILM, Multi Gauge).
UV cross-linking and competition assays were used to identify host factors directly interacting with viral RNA templates [60]. Various 32P-labeled RNA probes (25 fmol) in binding buffer [Tris (20 mM, pH 8.0), MgCl2 (3 mM), KCl (10 mM), DTT (2 mM), RNase inhibitor (5 units), yeast total RNA (1 µg), BSA (1 µg), and glycerol (4%)] were added to the BaMV RdRp preparation and incubated at room temperature for 10 min. The mixture was placed in an ice bath and illuminated under a UV lamp at 254 nm wavelength (Stratagene, UV stratalinker TM 1800) for 20 minutes. The samples were then treated with RNase A (10 µg) and RNaseT1 (0.5 unit) for 30 min at 37°C, boiled in protein sample buffer for 5 min and analyzed by electrophoresis with a 10% polyacrylamide gel containing 1% SDS. Radioactive images were scanned and quantified using the BAS-1500 bioimaging analyzer (Fujifilm, Multi Gauge). BamHI-linearized pT7r138/Bam, XbaI-linearized psatBaMV/3′ UTR, and BstNI-linearized pT7CMV/tRNA were transcribed with T7 RNA polymerase in the presence of [α-32P]UTP [54] for the preparation of 32P-labeled RNA probes corresponding to BaMV 3′ UTR, satBaMV 3′ UTR, and CMV 3′ UTR, respectively. In the competition reactions, various amounts of unlabeled competitor RNA were pre-incubated with the proteins for 10 min prior to the addition of 32P-labeled RNA probe. Description of competitor RNA preparations can be found in the Supporting Information (Text S1).
Details for protein expression and purification are described in the Supporting Information (Text S1). Purified proteins (about 1 µg) were separated by 10% SDS-PAGE and transferred to nitrocellulose membranes. The membranes were incubated overnight at room temperature in renaturation buffer comprising Tris-HCl (10 mM), EDTA (1 mM), NaCl (50 mM), yeast total RNA (25 µg/ml), and 1× Denhardt's reagent at pH 7.5. The membranes were probed with 32P-labelled BaMV 3′ UTR (5×105 cpm/ml) for 2 h and washed with renaturation buffer three times (10 min per wash). The membrane was then air dried and analyzed by PhosphorImager (FUJIFILM, Multi Gauge).
The Tobacco rattle virus (TRV)-based virus induced gene silencing (VIGS) system [79] was used for knocking down the expression of specific host genes. The pTRV1, pTRV2, and pTRV2/PDS plasmids were kindly provided by Dr. David C. Baulcombe (Department of Plant Sciences, University of Cambridge, UK). Plasmid pTRV2/NbHsp90 harboring a 300-bp fragment corresponding to nt 1801–2100 of the NbHsp90 coding sequence was constructed by PCR using pGEXHsp90 as the template with the forward primer NbHsp90S-F (5′- GCTCTAGAAGCAAGAAGACCATG-3′) and reverse primer NbHsp90S-R (5′- TCCCCCGGGTTAGTCAACTTCC-3′), followed by cloning of the amplified fragment into the pTRV2 vector. The pTRV1, pTRV2, and pTRV2/NbHsp90 plasmids were individually introduced into Agrobacterium tumefaciens strain C58C1 by electroporation. For VIGS assays, A. tumefaciens cultures (OD600 = 1) containing pTRV1, pTRV2, or pTRV2/NbHsp90 were mixed in a 1∶1 ratio as indicated, and co-infiltrated by syringe onto three leaves of each test plant (N. benthamiana). At 7 days post agro-infiltration (dpai) the third and fourth leaves above the infiltrated sites were mechanically inoculated with 0.5 µg virions of BaMV, PVX, or CMV. Two and six days later, total RNA was extracted from the virus-inoculated leaves of three independent plants. BaMV, PVX, and CMV RNA accumulations were determined by Northern blot assay.
At seven dpai, total RNA of the third and fourth leaves above the infiltrated leaves were extracted for detection of Hsp90 mRNA expression by real-time PCR. First-strand cDNA was synthesized by reverse transcription using 1 µg of total RNA as the template and oligo d(T)39 primer. For real-time PCR, the Hsp90 mRNA levels were monitored by forward primer Hsp90-908F (5′-AGGGTCAGCTGGAGTTCAA-3′), and reverse primer Hsp90-1098R (5′-GGGAAGATCCTCGGAATCCAC-3′). For a negative control, we used PCR without prior reverse transcription. The actin mRNA level was determined by real-time PCR with forward primer Actin-F (5′-GATGAAGATACTCACAGAAAGA-3′) and reverse primer Actin-R (5′-GTGGTTTCATGAATGCCAGCA-3′) for normalization of the specifically silenced gene.
Total RNA was extracted from inoculated leaves at 2 and 6 dpi and from protoplasts at 24 hpi [77]. RNA samples were separated by electrophoresis after denaturation in the presence of glyoxal and transferred onto nylon membrane (Amersham, UK) for Northern blot analysis [77]. Blots were hybridized with specific riboprobes to detect BaMV, PVX, CMV, or satBaMV RNA. The 32P-labeled probes, specific for the detection of (+)-strand BaMV and satBaMV, were transcribed from HindIII-linearized pBaHB and EcoRI-linearized pBSHE using SP6 and T7 RNA polymerase, respectively, as described previously [80]. The PVX- and CMV-specific probes were prepared by linearization of pPVXHE and pCMVHE harboring PVX and CMV (−) 3′ terminal sequences downstream of the T7 promoter, respectively (Hsu et al, unpublished), with HindIII, followed by transcription with T7 RNA polymerase in the presence of [α-32P] UTP. Hybridization signals were detected and quantified by using a PhosphorImager (Multi Gauge, FUJIFILM).
Protoplasts were isolated from N. benthamiana as previously described [27]. Hsp90 inhibitors, geldanamycin (GA, Stressgen) and 17-allylamino-demethoxygeldanamycin (17-AAG, Sigma-Aldrich) were used at 0.2 to 2 µM concentration for the inhibition of ATP-Hsp90 interaction in the protoplast cells. Varying concentrations of GA or 17-AAG were added to the isolated protoplasts before or after virus infection at the indicated times. For each inoculation, 1 µg of viral RNA or 10 µg of plasmid DNA of infectious cDNA clone was used to inoculate 2×105 protoplasts. BaMV, PVX, or CMV RNA was purified as described previously [77]. The infectious clones of BaMV-S and satBaMV RNA, pCB and pCBSF4, respectively, were described previously [29], [80]. To express BaMV ORF1 and its derived mutant in protoplast cells, plasmids pBaORF1 and pBaORF1dGDD, a defective ORF1 with the GDD motif deleted [26], were generated from the full-length pCB and pCBdGDD clones respectively, by removing other ORFs using DraIII and SacI.
Yeast strain L40, pHybLex/Zeo bait and pYESTrp2 prey vectors, and positive control plasmids pHybLex/Zeo-Fos2 and pYESTrp-Jun were from the Hybrid Hunter Kit purchased from Invitrogen. In a two-hybrid system, we divided the transcription factor into two domains, a DNA binding domain (BD) and an activation domain (AD). These were fused as two separated hybrid proteins referred to as bait and prey, respectively. A positive interaction causes HIS3 and lacZ reporter gene expression in Saccharomyces cerevisiae. Preparation of yeast competent cell, yeast transformation, selection of interaction, and β-galactosidase filter assay were carried out according to the user manual provided by Invitrogen. The bait and prey designate plasmids combined as indicated were transformed simultaneously into an L40 cell and plated on -Trp/Zeo300 for selection of successfully co-transformed cells. The colonies grown on the selected plate were transferred to the filter to analyze β-galactosidase activity. Additionally, three colonies were selected at random, dissolved in water and plated on -Trp-His/Zeo300 for interaction selection. Descriptions of plasmids and their construction are detailed in the Supporting Information (Text S1).
Details for protein expression and purification are described in the Supporting Information (Text S1). GST or GST fusion proteins (1 µg each) were mixed with GST affinity resin (Novagen) and shaken gently at 4°C for 1 h, followed by the addition of His(6) fusion proteins (1 µg each) and further incubation with gentle shaking for 2 h. The proteins bound to beads were washed with 1× PBS and eluted with elution buffer (50 mM Tris-HCl, pH 8.0, 10 mM glutathione). The eluted proteins were resolved by 10% SDS-PAGE and analyzed by immunoblotting using an anti-His antibody (LTK BioLaboratories).
|
10.1371/journal.pgen.1000302 | The Chromosomal High-Affinity Binding Sites for the Drosophila Dosage Compensation Complex | Dosage compensation in male Drosophila relies on the X chromosome–specific recruitment of a chromatin-modifying machinery, the dosage compensation complex (DCC). The principles that assure selective targeting of the DCC are unknown. According to a prevalent model, X chromosome targeting is initiated by recruitment of the DCC core components, MSL1 and MSL2, to a limited number of so-called “high-affinity sites” (HAS). Only very few such sites are known at the DNA sequence level, which has precluded the definition of DCC targeting principles. Combining RNA interference against DCC subunits, limited crosslinking, and chromatin immunoprecipitation coupled to probing high-resolution DNA microarrays, we identified a set of 131 HAS for MSL1 and MSL2 and confirmed their properties by various means. The HAS sites are distributed all over the X chromosome and are functionally important, since the extent of dosage compensation of a given gene and its proximity to a HAS are positively correlated. The sites are mainly located on non-coding parts of genes and predominantly map to regions that are devoid of nucleosomes. In contrast, the bulk of DCC binding is in coding regions and is marked by histone H3K36 methylation. Within the HAS, repetitive DNA sequences mainly based on GA and CA dinucleotides are enriched. Interestingly, DCC subcomplexes bind a small number of autosomal locations with similar features.
| In sexually dimorphic species, unequal distribution of sex chromosomes requires adjustment of gene expression levels between the sexes. Male flies enhance transcription from the single X chromosome to meet the levels in females (XX). The specific recognition of sex chromosomes is a crucial step in this dosage compensation process. Intuitively, one might assume that sex chromosomes harbor distinct DNA sequence motifs for recruitment of the modulating machinery; however, no clearly defined motifs capable of fulfilling this role have yet been found. One explanation for this shortcoming could be our failure to date to identify a sufficiently large set of sites that serve as specific docking stations. In the following study, we have systematically mapped the strongest recruitment sites of the Drosophila dosage compensation complex (DCC) and identified shared sequence elements. The closer a gene resides to one of these sites the more robust is regulation by the DCC, which documents the function of our inventory of high-affinity binding sites.
| Genes residing on the single X chromosome in male Drosophila flies are transcribed at elevated rates to match the expression levels of the two X chromosomes in female cells. Transcriptional tuning in male cells depends on the activity of a ribonucleoprotein complex, the dosage compensation complex (DCC, also referred to as MSL [male-specific lethal] complex, reviewed in [1],[2]). Formation of DCC is male-specific due to the expression of the key subunit MSL2, which in turn drives the expression of the non-coding RNA components of the DCC, the roX (RNA on the X) RNAs [3],[4]. The complex associates almost exclusively with the X chromosome, which explains the selective activation of X chromosomal genes. This is at least in part due to the acetylation of lysine 16 of histone H4 (H4K16) by the histone acetyltransferase (HAT) MOF, an integral subunit of the DCC [5]. This modification may directly lead to unfolding of the chromatin fiber [6] or indirectly counteract factors that promote the formation of repressive chromatin [7],[8] rendering chromatin more permissive to the progress of transcription.
The phenomenon of dosage compensation allows the study of general principles of transcriptional fine-tuning and chromosome-wide regulation. A key question is how the DCC is recruited specifically to the X chromosome. High-resolution mapping demonstrated that the complex targets transcriptionally active regions on the X chromosome with a preference for coding sequences [9],[10]. The DCC distribution pattern cannot be easily explained by a single targeting principle, but presumably results from the successive application of two or more distinct principles. Early genetic experiments led to a concept that assumes the existence of a relatively small number of X chromosome-specific primary recruitment or chromosomal ‘entry’ sites (CES) for the DCC, from which the complex would ‘spread’ to the bulk of chromosomal binding sites that differ qualitatively from the entry sites [11],[12]. Entry sites could, for example, be defined by a particular DNA sequence element, whereas features of active chromatin combined with proximity to entry sites would be a hallmark of secondary sites. Subsequent studies disputed whether DCC binding sites should be sorted into categories defined by different recruitment principles, or whether all targeting could be explained by a single principle (e.g. DNA sequence) that was applied to define sites of higher or lower affinity [13],[14],[15],[16]. Independent of whether primary recruitment sites differ from the bulk of DCC binding sites in quality or by a quantitative feature, they attract the DCC under stringent conditions. For example, DCC is recruited to high-affinity sites (HAS) even if they are removed from the X chromosomal context and inserted on an autosome, or at low levels of DCC (genetically achieved through expression of low amounts of MSL2) [16],[17], or if the integral DCC subunits MSL3, MLE, MOF or the roX RNAs are absent [18],[19],[20]. Under the latter circumstances binding sites are demarcated by binding of a sub-complex consisting of only MSL1 and MSL2 [19]. Evidently, distribution of DCC to sites of supposedly lower affinity depends on MOF, MSL3, MLE and the roX RNAs.
More recently, Kuroda and colleagues obtained additional support for the concept that primary and secondary DCC binding sites are defined by different principles by showing that the binding to active chromatin in the vicinity of primary targeting sites is not X-specific [21]. Insertion of a roX gene in an autosome leads to extended ‘spreading’ of the DCC over the neighboring chromatin (both roX genes contain a HAS [11],[22]). Under these circumstances, the DCC associated with transcribed sequences on autosomes like it normally does on the X chromosome. Recruitment of DCC was suggested to involve binding to methylated histone H3 at lysine 36 (H3K36me3), a modification that is placed by histone methyltransferases associated with elongating RNA polymerase II (pol II) and hence marks sites of active transcription [21].
X chromosome-specific targeting may, therefore, be encoded by primary targeting sites. So far, just a few DNA elements that robustly fulfill the criteria for a primary targeting site have been characterized at the DNA sequence level. These include sites within the roX genes [22],[23], the Smr and Tao-1 genes [13] as well as a site that maps to cytological position 18D [15]. Due to this limited number, a defining feature with predictive value could not be extracted, although the presence of multiple distinct DNA sequence elements has been correlated with HAS [22],[24]. Strikingly, low complexity sequence elements such as GA- and CA-based dinucleotide repeats as well as runs of adenines have repeatedly been noted in these analyses [10],[13],[24]. Dissection of HAS DNA has yielded sub-fragments that retain limited binding activity. We therefore suggested that primary targeting is based on the local clustering of distinct sequence motifs [13],[24].
Progress on HAS definition requires moving the analysis from the anecdotal to the systematic level. We therefore mapped all DCC binding sites with highest affinities on a chromosome-wide scale by combining chromatin immunoprecipitation (ChIP) with probing of high-resolution DNA tiling arrays (ChIP-on-chip) under conditions where sites of higher affinity are preferentially visualized. This strategy not only allowed the generation of a sufficiently large training set for sequence analysis, but at the same time provided a means to directly compare the high affinity binding pattern with several chromatin features that have already been mapped along the Drosophila male X chromosome.
We followed two complementary strategies for filtering the DCC binding sites with highest affinity from the chromosome-wide binding profile. First, we attempted to reproduce in male tissue culture cells the conditions that lead to selective visualization of HAS on polytene chromosomes in mutant larvae, where MSL1 and MSL2 interact selectively with HAS in the absence of MSL3, MLE or MOF [13],[18],[19]. Towards this goal we reduced the levels of these factors in the male Drosophila cell line SL2 by RNA interference (RNAi) and monitored the residual interaction of MSL1 or MSL2 (genetic studies have established the mutual interdependence of these two subunits for their interactions with the bulk of X chromosomal sites [19]). The second strategy followed the idea that HAS should, on average, show a higher occupancy by DCC and hence should be selectively obtained by ChIP if the extent of formaldehyde crosslinking was reduced. Lower levels of crosslinking should also reveal sites of more intimate contact of MSL proteins with DNA. Reassuringly, both strategies led to a similar alteration of the MSL binding pattern with enhanced peaks along all previously known HAS. The combined data should therefore help to define an inventory of sites with similar properties.
We lowered the levels of MSL3, MLE and MOF in SL2 cells by RNAi and mapped the residual binding pattern of the DCC core components MSL1 and MSL2 by ChIP-on-chip. All knock-down experiments were controlled for non-specific effects by a parallel RNAi treatment with an irrelevant dsRNA (which corresponds to glutathione-S-transferase (GST) sequences: ‘GST RNAi’). After 7 days of treatment with double-stranded (ds) RNA we achieved approx. 90% depletion of the target proteins as compared to the RNAi GST control (Figures 1A & B). Removal of MLE also led to reduction of MSL3 levels. Furthermore, RNAi-mediated depletion of MLE, MSL3 and MOF resulted in a substantial reduction of MSL1 to as little as 20%, indicating a global destabilization of the complex (Figure 1C). A similar drop in protein levels was observed for MSL2 (not shown). These circumstances should further facilitate selecting binding sites of only the highest affinity.
In flies a genetic knockout of MLE or MSL3 leads to the most pronounced reduction of MSL1-MSL2 binding [13]. We therefore first investigated the residual MSL1 profile after RNAi against MLE or MSL3 (Figure S1). The chromosomal interaction profile showed surprisingly mild effects: only 6 and 9% of all significant MSL1 binding events were lost upon MSL3 or MLE depletion, respectively. On all MSL1 target probes we observed a moderate reduction of MSL1 signals (Figure S1C). Conceivably, the remaining DCC subunits after incomplete knockdown may suffice to sustain MSL1 binding. However, in light of results from the analysis of mutant fly strains we consider more likely that the ChIP-on-chip methodology underestimates homogeneous inter-array differences and therefore might obscure a global reduction of MSL1 binding under knockdown conditions. This would be due to disproportional procedures such as array hybridization and scanning as well as signal normalization across arrays. Visual inspection of the binding pattern, however, allowed for the identification of loci where MSL1 association was substantially reduced (such as the small gene cluster in the right half of Figure S1A). This indicates, that MSL3 and MLE RNAi cause a local redistribution of MSL1, which should contribute to the identification of high affinity regions that are supposed to be more resistant to these perturbations. In the case of MSL3, examination of the loss of MSL1 binding within distinct functional regions revealed a significantly stronger reduction in coding sequences as compared to other binding regions (p-value<2.2e-16; two-sided t-test; Figure S1C, right green box).
We next explored the usefulness of our second strategy and established the MSL2 binding profile at lower levels of formaldehyde crosslinking. We fixed cells with only 0.1% instead of 1% of formaldehyde (see Materials and Methods for details). Crosslinking with low concentrations of formaldehyde, we lost about 50% of significant MSL2 binding (Figure 2B), which preferentially affected coding sequences (p-value<2.2e-16; two-sided t-test; Figures 2C and D). We were encouraged by the fact that all genetically identified HAS were retained among the residual MSL2 peaks. For example, a previously identified HAS within the first intron of the Tao-1 gene coincides with a pronounced MSL2 peak (Figure 2A; [13],[24]).
We then applied the same crosslinking conditions to the analysis of MSL1 binding after MOF RNAi (Figure 3). Comparing the pattern to the control pattern obtained after RNAi with GST sequences, the global reduction of MSL1 interaction across all genomic regions was much more pronounced than in the case of RNAi against MSL3 or MLE (compare Figure S1C to Figure 3C). The GST RNAi control sample that was also fixed with low formaldehyde exhibited similar alterations of the MSL1 binding pattern as observed for MSL2 (Figure 2A). Depletion of MOF resulted in loss of 23% of all significant binding events, again with a strong preference for coding sequences (p-value<2.2e-16; two-sided t-test; Figures 3B and C).
In summary, we found that two unrelated strategies aimed at selecting DCC binding sites of higher affinities led to an overall reduction of chromosomal association of the core DCC components MSL1 and MSL2 with a preferential loss of binding from coding regions. We tentatively conclude that coding regions are less likely to contain HAS.
We then attempted to identify particular genomic regions that are similarly enriched under both experimental regimes: upon RNAi against the spreading factors and at low levels of crosslinking. We transformed the enrichment ratios of all residual profiles (i.e. MSL1 binding after RNAi against MOF, MLE, or MSL3; MSL2 binding after low crosslinking) to z-scores and calculated an unweighted cumulative z-score. Region thresholding on the smoothed z-score profile allowing for a maximum of 1% autosomal site detection identified 130 HAS spread all along the X chromosome (Table S1). In addition, this approach picked up one autosomal site. The median length of the sites is 800 bases and the distribution of their distances peaks between 130 and 260 kb (Figures 4C and D). Even though some of the RNAi experiments had a limited effect on the global MSL1 distribution, it turned out that all caused local redistributions of MSL1 signals with a significant retention on HAS (Figure S2).
Our set contains all genetically defined robust HAS (roX1, roX2, 18D, Smr and Tao-1 [11],[13],[15]). For example, Figure 4 documents the correspondence of the previously mapped HAS within the roX2 gene (Figure 4A) and the HAS at 18D (Figure 4B). To test this correspondence more generally we picked six new HAS and mapped them with respect to the ‘entry sites’ visualized on polytene chromosomes by Immuno-FISH. We chose female larvae of the Sxb1-2C line, which express low levels of MSL2. The reduced amount of DCC in these flies binds to a small number of chromosomal loci that coincide with the ones observed in males mutant for mle, msl3 or mof [13]. Larvae of this fly strain have previously been used to define ‘entry’ sites and to characterize HAS [13],[16]. Five out of six sites robustly colocalized with MSL2 binding sites (Figure 5). One of our strongest sites upstream of the Nej gene did, however, not show an overlap with the MSL2 pattern (see discussion). In contrast, control FISH probes located 60 kb (Or2a) or 400 kb (dpr8) away from the next HAS did not colocalize with MSL2.
We next explored whether HAS location could be attributed to a particular functional context (intergenic, UTR, intron, coding sequences, etc). Many of the 131 sites are too large for precise functional assignment, since they contain coding as well as non-coding sequences. However, we found 51 regions that are unambiguously located within a defined functional genomic context. Almost all of these sites are found in regulatory or non-coding regions within or close to genes (Figure 6B), in support of the earlier notion that DCC interactions with coding sequences are of lower affinity.
As the majority of DCC binding sites map to transcriptionally active parts of the X-chromosomal chromatin, we tested the transcription status of the HAS. Only 60% of these sites overlap with regions of elongating RNA polymerase II (data not shown). Taken together with the fact that there is also considerable binding to regions proximal to the transcription units (Figure 6B), we conclude that a substantial fraction of the HAS are not transcribed.
An important question regarding the strongest DCC binding sites is whether they contain common chromatin features that may help to explain X-chromosomal specificity. Previously, several of the few known HAS were shown to reside within regions of nucleosome depletion (DNase I hypersensitive sites) [15],[23],[24]. In order to explore whether this was a more general feature of HAS, we compared the location of the HAS with the published histone H3 profile [21]. Visual inspection of the data reveals that a large fraction of the sites colocalize with regions of low histone H3 content (see example in Figure 6A). In general, microarray probes located within HAS had significantly lower H3 signals than the ones outside (p-value<2.2e-16; two-sided t-test). Calculating the cumulative H3 distribution across all 131 HAS (Figure 6C) we found that the H3 profile clearly dropped towards the centers of the sites in a window of about 1 kb. The low resolution of the analysis based on the rather large chromatin fragments generated in our ChIP procedure (500 bp) does not allow for a robust determination of the number of nucleosomes or the length of DNA that might be affected. Clearly, however, nucleosome depletion is not alone sufficient to initiate DCC binding, as the number of X-chromosomal nucleosome depleted regions (1148) greatly exceeds the number of HA sites.
The low nucleosome occupancy of the HAS suggests that primary recruitment of the DCC to the X chromosome may involve recognition of exposed DNA sequences rather than histone modifications. We therefore performed extensive sequence analysis in order to identify motifs that are significantly enriched in the strongest binding sites. A motif that was identified most robustly in several MEME [25] analyses with varying parameters and training sets is shown in Figure 7d. An example MEME run on the strongest of our HAS is provided in the supplement (Dataset S1). Results varied depending on the size of the training set and the analysis parameters. However, dinucleotide repeats based on either GA or CA as well as runs of adenines were frequently identified. Such sequences have already been postulated to be involved in DCC recruitment [10],[13],[22],[24]. The GA-repeat based motif shown in Figure 6D is the only one that was present in a large fraction of HAS (68 of 131 by MAST analysis) and at the same time enriched on the X chromosome (1.5 fold by genome-wide search with the MEME-derived position-specific scoring matrix). Only 33% of the motifs identified on the X chromosome are, however, bound by the complex in SL2 cells, suggesting we may be missing the context within which the identified sequence motif might contribute to DCC recruitment. For example, the site may operate in conjunction with a variable cohort of secondary motifs, as suggested by previous fine-mapping of known HAS elements [24].
Interestingly, the single autosomal site that was picked up at the chosen threshold by our approach binds MSL1 robustly in the absence of significant amounts of MSL2 (Figure 7A). This is remarkable considering the widely held belief that MSL1 and MSL2 mutually depend on each other for chromosome interaction [19],[26]. To address the stoichiometry of DCC subunits at autosomal sites more systematically we identified all binding sites for MSL1, MSL2 and MOF statistically and found that these three proteins do not colocalize on most sites, with MSL2 showing the lowest occupancy (Figure 7C). Intriguingly, on autosomes MSL1 mainly binds to promoters of active genes (as seen e.g. in Figure 7B). These are depleted of nucleosomes, in analogy to X chromosomal HAS: probes within the autosomal binding sites show significantly reduced histone H3 content (p-value = 3.4e-14; two-sided t-test). Sequence analysis of the autosomal sites with strongest MSL1 binding revealed dinucleotide repeats (mainly GA and CA) similar to the ones that characterize the X chromosomal HAS (Datasets S1 and S2). One notable difference between MSL1 binding sites on the X and on autosomes is that whereas X chromosomal HAS usually display a rather broad MSL1 distribution, MSL1 binding at autosomal sites is spatially restricted and does not spread substantially onto the adjacent active chromatin, even at sites where all tested complex components colocalize (Figure 7B). This is confirmed globally by the autocorrelation of MSL1 binding on a smoothed profile with a spacing of 500 bases, which shows that binding domains on the X are much broader than those on autosomes [compare Figures 7E and F, autocorrelation (ACF) of >0.5 within 10 kb (X) or of 0.5 at a maximum of 2 kb (3R)], the latter most likely reflecting rather restricted or singular binding events blurred by the ChIP resolution of about 500 bases. Furthermore, MSL1 binding to autosomal sites not only remains upon reduction of MOF, MSL3 or MLE through RNAi, but is even increased under those conditions (Figure S3). This might reflect a re-distribution of MSL sub-complexes from the X to autosomes after elimination of the spreading component.
The 130 HAS defined by our analysis are distributed all along the X chromosome with a preferred distance between 60–300 kb. If the HAS were the primary organizers of larger dosage compensation domains, we would expect a relationship between the robustness of transcriptional compensation of genes and their distance from the nearest sites. Figure 8 shows that this is actually the case. Dosage compensation reflected by the drop of transcript upon ablation of MSL2 by RNAi decreases with growing distance (Figure 8A). On the other hand, the further away a given gene is from a HAS, the more dependent is its compensation on the spreading factors MOF and MSL3 (Figures 8B & C). This finding demonstrates that the HAS we have identified play a role as organizers of compensated domains.
Combining differential crosslinking and RNAi interference against the DCC subunits previously shown to be required for the ‘spreading’ of the complex from high affinity or ‘entry’ sites we identified 131 high-affinity sites (HAS) of the Drosophila dosage compensation complex in male SL2 cells. This set of sites contains all previously identified HAS (or chromosomal entry sites, CES) and a representative selection colocalizes with interbands on polytene chromosomes that had been described as harboring primary binding sites for the DCC in previous genetic analyses. The sites we now identified thus have similar properties to the ones identified by genetic means. Our study not only provides a much large number of such sites, but also resolves their positions and widths much more precisely than enabled by the polytene chromosome analyses. Most importantly, our study suggests that the HAS have a function in dosage compensation since we observe a positive correlation between the proximity of genes to a HAS and the extent of dosage compensation. Conversely, the further away genes reside from the nearest HAS the more they depend on the spreading factors such as MOF or MSL3 for enhancement of transcription. The 130 X-chromosomal HAS are distributed all along the chromosome with a predominant spacing between 60 and 300 kb. The realm within which loci profit from the presence of a high affinity ‘DCC attraction center’ may be of the same order of magnitude. However, we generated the inventory of HAS by applying fairly stringent thresholding criteria. Less stringent selection criteria will undoubtedly reveal a large number of sites with degenerate features and lower affinities that may serve as ‘relay stations’ for DCC spreading and may contribute cumulatively to concentration of the DCC on the X chromosome[27]. Finally, the linear display of DCC–chromosome interactions in a browser obviously does not reflect the three-dimensional path and packaging of the chromosomal fiber, which might facilitate transfer of a chromatin-bound complex between distant loci.
Under normal circumstances the DCC binds with high preference to transcribed and, indeed, coding sequences [9],[10]. Our observation that a transcribed region upstream of the Nej gene harbors a strong site in our set of binding sites but is not occupied in polytene chromosomes may, therefore, be due to differences in the transcription status between salivary glands and SL2 cells. Selection for sites of higher affinity leads to preferential loss of DCC from coding sequences, and under low-crosslinking conditions the majority of DCC binds at non-coding sequences in UTRs, introns, and also outside of the transcribed sequences in presumed regulatory and intergenic regions. Apparently, coding sequences have a lower affinity than non-coding sequences. At least part of the attraction of the DCC to transcribed sequences is due to the histone H3K36me3 mark, which is co-transcriptionally placed by Set2 and may provide a docking site for MSL3 [21]. However, this modification marks all transcribed sequences on autosomes as well and cannot be responsible for primary targeting. If, as suggested by this and previous work [10],[13],[24],[28], DNA sequence motifs contribute to DCC targeting, the observed preference for HAS outside of coding regions makes sense: assuming that binding affinity increases as sites conform with an idealized ‘consensus’ sequence, evolution of HAS with better defined sequences will be limited at coding regions where the main selective pressure is on preserving protein coding. If coding regions contain sequence elements that bind DCC they may, therefore, be of lower affinity and hence be preferentially lost as the stringency of the selection increases.
Sequence analysis of the HAS did not lead to the identification of a single motif that could explain the HAS interaction pattern. Rather, we found low complexity sequences, in particular GA and CA dinucleotide repeats, generally enriched in HAS, but in no instance present in more than 50% of the sites. The results of the sequence analysis fluctuate considerably depending on the selected training set, the analysis parameters and algorithms used. The only motif that was found consistently within the set of HAS that is also enriched on the X chromosome is an almost perfect 11mer of GA. We previously identified similar repeats employing very different strategies [10],[13]. Blocks of GA are also important for targeting the DCC to a nucleosome-free region within the roX2 gene [22].
Recently, Kuroda and colleagues published a similar study including high resolution mapping of HAS of the Drosophila DCC [29]. Even though they used Drosophila embryos and different experimental approaches (e.g. genetic knockouts instead of RNAi and Solexa sequencing in addition to tiling array analysis) the results of the two studies match surprisingly well. In fact, 90 of our 130 X-chromosomal HAS perfectly overlap with the chromosomal entry sites (CES) from the Kuroda lab (the differences in sites may well be explained by the different transcriptional status of the cells/embryos employed in the two studies). The GA-based sequence motif that we found enriched in the HAS perfectly covers the consensus MSL response element of the Kuroda lab and they also observe a comparable histone depletion among their HAS. Using a reporter gene assay a role for the GA-rich sequence element in transcription activation was documented [29]. This not only confirms the suitability of our experimental approach but also reveals that a large fraction of HAS overlap in different specimens.
How GA repeat motifs contribute to DCC loading is not known, but several scenarios may be considered. So far, a direct interaction of DCC subunits with specific DNA elements cannot be excluded. Further, DCC targeting may rely on interaction with an accessory protein with appropriate sequence preference, such as Pipsqueak or the GAGA factor (GAF) encoded by the Trithorax-like (Trl) gene. These two GAG-binding proteins colocalize at numerous sites on polytene chromosomes [30]. Hypomorph trl mutants show a male-specific lethality if the levels of MSL1 and MSL2 are reduced [31]. However, GAF only colocalizes with MSL2 at one out of 33 HAS and mutant larvae with strong Trl alleles show no obvious alteration of the DCC binding pattern on polytene X chromosomes. However, they display an increased number of autosomal binding sites, which may indicate a certain perturbation of targeting [31]. GA-rich elements may synergize with other DNA sequences (and hence other interacting factors) to form HAS, as previously suggested [13],[24]. Local clustering of two unrelated DNA sequence motifs, neither of which is particularly enriched on the X chromosome, appears to be crucial for targeting the DCC in C. elegans [32].
The affinity of a given DNA sequence for an interacting factor is strongly lowered by its nucleosomal organization [27]. Chromatin serves as a general thresholding system to present only those binding sites that reside in an appropriate non-nucleosomal context or benefit from nucleosome remodeling [33]. Interestingly, we find that the HAS, independent of whether they are located in regulatory regions, introns or outside of transcribed sequences, tend to be depleted of nucleosomes. Nucleosome depletion alone is not a stringent determinant of DCC association since many sites of low nucleosome density do not contain HAS or are not bound by the complex. Conversely, not all HAS are entirely nucleosome-free. Nevertheless, an improved definition of HAS may require considering the degree of nucleosome occupancy of sites in addition to the actual sequence itself. Nucleosome disruption may be brought about by ATP-dependent nucleosome remodeling or by competition of DCC binding with nucleosome assembly at the replication fork [34]. In the latter scenario the absence of nucleosomes would be a consequence of DCC binding rather than a requirement for interaction. Nucleosomes are also disrupted by the progression of the elongating RNA polymerase, a fact that may explain the recent observation that DCC binding to a sequence element within the MOF gene benefited from transcription [28].
Dinucleotide repeats and nucleosome depletion are also characteristic of autosomal MSL binding sites, however, these sites differ from HAS by two interesting features. First, we observed an altered stoichiometry of MSL proteins at autosomal sites, which often appear to lack MSL2. At these sites the colocalization of MSL1, MSL2 and MOF is the exception rather than the rule, suggesting that the known interdependence of MSL1 and MSL2 for chromosome association [19] is not absolute, but context-dependent. Second, binding of MSL proteins to autosomal sites appears unusually confined and does not spread onto the adjacent active chromatin as is commonly observed for X-chromosomal HAS. Lack of spreading is also found in the presence of MSL2. Because the distribution of MSL proteins from initial targeting sites is strongly facilitated by transcription of roX RNA from the same chromosome [11],[21],[35], we speculate that autosomal sites may be bound by MSL proteins in the absence of roX RNAs.
Our data are consistent with a multi-step model of X chromosomal targeting by the DCC, which involves assembly of the complex with nascent roX RNA within the X chromosomal territory, followed by its diffusion to and concentration by the set of HAS, which we have identified in this study. Distribution to all target genes may then be brought about by large numbers of low affinity sites and the transcription-associated H3K36 methyl mark.
Cultivation of the male Drosophila cell line SL2 and RNA interference of target genes were carried out as described previously [36]. In brief, 1×106 SL2 cells were incubated with 10 µg dsRNA for 1 hour in serum-free medium. After addition of serum-containing medium, cells were incubated for 7 days at 26°C before chromatin preparation. Preparation of whole cells extracts and western blot confirmation of target gene knockdown has been described previously [36]. Depletion efficiency was quantified using a Li-Cor Odyssey system using α-tubulin as a reference. Sequences of primers used for dsRNA production are listed in Table S2.
SL2 cells were crosslinked in growth medium using 1% formaldehyde for 60 minutes in icewater. Alternatively we used 0.1% formaldehyde for 10 minutes at RT (low formaldehyde crosslinking). Fixation was quenched by addition of glycine to a final concentration of 125 mM. After washing, cells were resuspended in RIPA buffer and sonicated using a Bioruptor (Diagenode, Belgium) 8 times 30 seconds using the ‘high’ setting. Fragment size of the obtained chromatin was checked to be between 300 bp and 700 bp. Chromatin was precleared using a protein A/protein G-sepharose mixture for 1 hr at 4°C. 200 µl chromatin was incubated with appropriate amounts of antibodies in a total volume of 500 µl RIPA buffer at 4°C overnight. After washing and crosslink reversal, immunprecipitated nucleic acids were purified on GFX columns (GE Healthcare). Input chromatin serving as reference sample was treated accordingly. Overall, we performed immunoprecipitations for MSL1 (4 biological replicates) and MSL2 (2 replicates) on chromatin from untreated SL2 cells. In addition, we precipitated MSL1-containg chromatin after GST, MSL3, or MLE RNAi (2 replicates each). After low formaldehyde crosslinking, we performed ChIP for MSL2 from untreated cultures (2 replicates) and MSL1 IP after GST or MOF RNAi (3 replicates each). The rabbit polyclonal MSL1 and MSL2 antibodies used in this study were described elsewhere [10],[24].
Input and IP DNA were amplified using the WGA kit (Sigma) according to an online protocol (http://www.epigenome-noe.net/researchtools/protocol.php?protid30). Labeling and hybridization to NimbleGen arrays was carried out at ImaGenes (Berlin, Germany). We used a custom array layout (approx. 1 probe/100 bases) comprising the euchromatic part of the entire X chromosome, 5 MB of 2L, 2R and 3L, respectively, as well as 10 MB of 3R. Data analysis was performed using R/Bioconductor (www.R-project.org; www.bioconductor.org). Raw signals of corresponding experimental replicates were normalized using the ‘vsn’ package [37]. Enrichment statistics (IP versus input signals) were computed using the ‘sam’ algorithm within Bioconductor [38]. Fdr values of the sam statistic were determined using ‘locfdr’ [39]. Region summarization was performed using the HMM algorithm of TileMap [40]. Probes were considered to be bound significantly if the posterior probability of the HMM was greater than 0.5. Statistical tests and presentations were performed using R defaults if not indicated otherwise. Details about high-level computations are available upon request. Visualization was carried out by loading the mean enrichment ratios as GFF files into GBrowse (www.gmod.org). All data correspond to Drosophila genome version dm2 and annotation version gadfly 4.3. Raw data was deposited at the NCBI gene expression omnibus, GEO (data series GSE12292). Wild type profiles and locations of high-affinity sites are available for browsing at http://genome1.bio.med.uni-muenchen.de.
The histone H3 profile and regions of histone depletion in SL2 cells were calculated from the GEO data series GSE8557 [21]. Gene expression changes upon RNAi of MSL2 in SL2 cells were derived from [41]. MOF and MSL3 knockout data were downloaded from ArrayExpress, accession E-MEXP-1505 [42].
FISH probes spanning the selected high-affinity sites were PCR amplified from genomic DNA. Primer sequences for the individual probes are listed in the supplement (Table S2). Immuno-FISH was performed exactly as described online (http://www.epigenome-noe.net/researchtools/protocol.php?protid4).
|
10.1371/journal.ppat.1000226 | A Structural Model of the Staphylococcus aureus ClfA–Fibrinogen Interaction Opens New Avenues for the Design of Anti-Staphylococcal Therapeutics | The fibrinogen (Fg) binding MSCRAMM Clumping factor A (ClfA) from Staphylococcus aureus interacts with the C-terminal region of the fibrinogen (Fg) γ-chain. ClfA is the major virulence factor responsible for the observed clumping of S. aureus in blood plasma and has been implicated as a virulence factor in a mouse model of septic arthritis and in rabbit and rat models of infective endocarditis. We report here a high-resolution crystal structure of the ClfA ligand binding segment in complex with a synthetic peptide mimicking the binding site in Fg. The residues in Fg required for binding to ClfA are identified from this structure and from complementing biochemical studies. Furthermore, the platelet integrin αIIbβ3 and ClfA bind to the same segment in the Fg γ-chain but the two cellular binding proteins recognize different residues in the common targeted Fg segment. Based on these differences, we have identified peptides that selectively antagonize the ClfA-Fg interaction. The ClfA-Fg binding mechanism is a variant of the “Dock, Lock and Latch” mechanism previously described for the Staphylococcus epidermidis SdrG–Fg interaction. The structural insights gained from analyzing the ClfANFg peptide complex and identifications of peptides that selectively recognize ClfA but not αIIbβ3 may allow the design of novel anti-staphylococcal agents. Our results also suggest that different MSCRAMMs with similar structural organization may have originated from a common ancestor but have evolved to accommodate specific ligand structures.
| Staphylococcus aureus (S. aureus) is a common pathogen that can cause a range of diseases from mild skin infections to life-threatening sepsis in humans. Some surface proteins on S. aureus play important roles in the S. aureus disease process. One of these bacterial surface proteins is clumping factor A (ClfA) that binds to the C-terminal region of one of the three chains of fibrinogen (Fg), a blood protein that plays a key role in coagulation. We carried out biochemical and structural studies to understand the binding mechanism of ClfA to Fg and to define the residues in Fg that interact with ClfA. Interestingly, the platelet integrin, which is important for platelet aggregation and thrombi formation, also binds to the same region of Fg as ClfA. Despite the fact that the two proteins bind at the same region, the mode of recognition is significantly different. Exploiting this difference in recognition, we have demonstrated that agents could be designed that inhibit the ClfA–Fg interaction but do not interfere with the interaction of Fg with the platelet integrin. This opens the field for the design of a novel class of anti-staph therapeutics.
| Staphylococcus aureus is a Gram-positive commensal organism that permanently colonizes 20% of healthy adults and transiently colonizes up to 50% of the general population [1]. For many years, S. aureus has been a major nosocomial pathogen causing a range of diseases from superficial skin infections to life-threatening conditions, including septicemia, endocarditis and pneumonia [1],[2]. Within the last decade a dramatic increase in the number of invasive infections caused by community-acquired S. aureus have been recorded in otherwise healthy children and young adults [3],[4]. This outbreak together with the continued increase in antibiotic resistance among clinical strains underscores the need for new prevention and treatment strategies [1].
A detailed characterization of the molecular pathogenesis of S. aureus infections may expose new targets for the development of novel therapeutics. Several staphylococcal virulence factors have been identified including capsule, surface adhesins, proteases, and toxins (reviewed in [5],[6],[7],[8]). One of these virulence factors is the MSCRAMM (microbial surface components recognizing adhesive matrix molecules) clumping factor A (ClfA). ClfA is the major staphylococcal fibrinogen (Fg) binding protein and is responsible for the observed clumping of S. aureus in blood plasma [9],[10]. Essentially all S. aureus clinical strains carry the clfA gene [11]; ClfA is a virulence factor in a mouse model of septic arthritis [12] and in rabbit and rat models of infective endocarditis [13],[14],[15].
ClfA generates strong immune responses and has shown potential as a vaccine component in active and passive immunization studies. In one study, mice vaccinated with a recombinant ClfA segment containing the Fg-binding domain and subsequently challenged with S. aureus showed significantly lower levels of arthritis compared to mice vaccinated with a control protein [12]. In another study, mice passively immunized with polyclonal or monoclonal antibodies against the ClfA Fg-binding domain were protected in a model of septic death [16]. The humanized monoclonal antibody, Aurexis®, has a high affinity for ClfA and inhibits ClfA binding to Fg [17]. Aurexis is currently in clinical trials in combination with antibiotic therapy for the treatment of S. aureus bacteremia [18]. Thus ClfA is a viable target for both vaccine and therapeutic strategies.
ClfA belongs to a class of cell wall-localized proteins that are covalently anchored to the peptidoglycan [5],[19],[20]. Starting from the N-terminus, ClfA contains a signal sequence followed by the ligand-binding A region composed of three domains (N1, N2, and N3), the serine-aspartate repeat domain (R region), and C-terminal features required for cell wall anchoring such as the LPXTG motif, a transmembrane segment and a short cytoplasmic domain [21],[22],[23]. A crystal structure of a Fg-binding ClfA segment (residues 221–559) which includes two of the domains (N2N3) demonstrates that each domain adopts an IgG-like fold [24]. This domain architecture was also determined from the crystal structure of the ligand binding segment of SdrG from Staphylococcus epidermidis, an MSCRAMM that binds to the N-terminal region of the Fg β-chain [25].
A dynamic mechanism of Fg binding termed “Dock, Lock and Latch” (DLL) has been proposed for SdrG based on a comparison of the crystal structures of SdrG N2N3 as an apo-protein and in complex with a synthetic peptide mimicking the targeted site in Fg [25]. In the SdrG DLL model, the apo-form of the protein adopts an open conformation that allows the Fg ligand access to a binding trench between the N2 and N3 domains. As the ligand peptide docks into the trench, a flexible C-terminal extension of the N3 domain is redirected to cover the ligand peptide and “lock” it in place. Subsequently the C-terminal part of this extension interacts with the N2 domain and forms a β-strand complementing a β-sheet in the N2 domain. This inserted β-strand serves as a latch to form a stable MSCRAMM ligand complex.
ClfA binds to the C-terminus of the Fg γ-chain [10],[23] and a synthetic 17 amino acid peptide corresponding to this region was shown to bind to ClfA. Interestingly, the A-region of the staphylocccal MSCRAMM FnbpA protein also binds to the same region in Fg [23]. Moreover residues in this Fg segment are also targeted by the platelet αIIbβ3 integrin [26],[27],[28] and a recombinant form of ClfA has been shown to inhibit platelet aggregation and the binding of platelets to immobilized Fg [10],[29],[30].
The current study was undertaken to characterize the interaction of ClfA and Fg to define in detail the binding of the C-terminus of Fg's γ-chain and to explore if compounds can be constructed that antagonize the ClfA-Fg interaction but does not affect the Fg interaction with the platelet-integrin αIIbβ3.
In previous studies, a segment of ClfA composed of residues 221–559 was shown to bind to the C-terminal end of the human Fg γ-chain [10]. We designed, based on structural similarities with SdrG, a smaller ClfA construct (229–545) predicted to be composed only of the N2 N3 domains and showed that ClfA229–545 retained the Fg-binding activity. To identify specific residues in Fg that are important for binding to ClfA229–545, a panel of peptides (Fig. 1A) based on the Fg γ-chain sequence 395–411 (referred to as γ1–17) were synthesized in which each position was sequentially substituted with an alanine residue (alanines 11 and 14 were changed to serines). These peptides were tested as inhibitors in solid-phase binding assays, using a peptide concentration giving about 50% inhibition by the wild-type peptide. Peptides γ1–17H6A, γ1–17H7A, γ1–17G10A, γ1–17Q13A, γ1–17A14S and γ1–17G15A were significantly less potent inhibitors than the native sequence suggesting that the Fg residues H6, H7, G10, Q13, A14 and G15 interact with ClfA (Fig. 1B). Remarkably, peptides γ1–17A11S, γ1–17D16A and γ1–17V17A showed enhanced inhibition of ClfA binding to a recombinant form of residues 395–411 of the Fg γ-chain fused to a GST protein (GST-Fg γ1–17) compared to a peptide with the wild-type sequence, indicating a higher affinity of the peptide variants for ClfA.
The ability of ClfA229–545 to bind to the peptide containing the γ1–17D16A mutation was further characterized. In solid-phase assays, ClfA binds to immobilized GST-Fg γ1–17 fusion protein with a lower affinity (Kd = 657 nM) compared to the mutated GST-Fg γ1–17D16A (Kd = 35 nM) (Fig. 1C). In solution, using isothermal titration calorimetry (ITC) assays, (Fig. 1D), ClfA also binds with a lower affinity to the native γ1–17 peptide (Kd of 5.8 µM) compared to the mutant Fg γ1–17D16A (Kd of 3 µM). Thus, although the apparent dissociation constants differ according to the assays used to estimate them, similar trends in affinity between the wild-type and the D16A mutation were observed.
Our results showed that alanine substitution at the C-terminal but not in the N-terminal region of the peptide affected MSCRAMM binding suggesting that the ClfA binding site is located at the very C-terminus of the Fg γ-chain (Fig. 1). Results also show that certain amino acid changes in the γ1–17 sequence enhance ClfA binding compared to the wild-type Fg sequence indicating that the human Fg γ C-terminal 17 residues may not be the optimum ligand for ClfA.
Analysis of the previously solved SdrG-Fg peptide complex crystal structure showed that only 11 out of the 18 peptide residues interacted with the MSCRAMM. Similarly, only a part of the 17-residue γ-chain segment may be required for binding to ClfA. In order to establish the minimum Fg peptide required for binding to ClfA229–545, a series of N- and C-terminal truncations of the γ1–17D16A peptide were synthesized (Fig. 2A). Truncations of 2, 4, 6 or 8 amino acids at the N-terminus of the Fg γ-peptide resulted in a reduced but detectable binding affinity when tested using ITC. There was a direct relationship between the length of the peptide and its affinity for ClfA. The smaller the peptide, the lower was the observed affinity for the MSCRAMM (Fig. 2B). Thus, the N-terminal residues of the Fg peptide (residues 1–8) are not critical for the interaction but may either contribute to or stabilize the binding of the peptide to ClfA. On the other hand, deletions of 2 or 4 residues from the C-terminal end of the γ1–17D16A peptide abolished binding. These results indicate that the C-terminal amino acids of Fg are critical for binding to ClfA and are in agreement with a previous report that showed that Fg lacking the C-terminal residues AGDV in the γ chain (corresponding to residues 14–17 in the peptide) or a Fg-variant that replaces the last four γ-chain residues with 20 amino acids lacks the ability to bind recombinant ClfA221–550 and induce S. aureus clumping [10].
The Fg binding mechanism of SdrG276–596 involves a transition from an open conformation, where the peptide binding trench between the N2 and N3 domains is exposed for ligand docking, to a closed conformation of the SdrG276–596 seen for the MSCRAMM in complex with the ligand peptide. The insertion of the N3 extension into the latching trench on N2, which represents the last step in the dynamic DLL binding mechanism, stabilizes the closed conformation of SdrG237–596 [31]. A closed conformation of apo SdrG N2N3, stabilized by introducing a disulfide bond between the end of the N3 latch and the “bottom” of N2, no longer binds Fg [31] demonstrating that for SdrG an open conformation is required for the initial docking of the ligand peptide. To explore if the binding of ClfA to Fg is also dependent on a movement of the latch we constructed a ClfA protein containing two cysteine substitutions. The locations of the cysteine mutations were determined using computer modeling and by sequence alignment to corresponding mutations in SdrG [31]. The mutant ClfAD327C/K541C generated a stable, closed conformation form. This recombinant His-tag fusion protein was purified by Ni+ chelating chromatography; ion-exchange and gel permeation chromatography. The ClfAD327C/K541C open and closed conformation forms were examined by SDS-PAGE analysis (Fig. 2C). Under non reducing conditions, the disulfide bonded closed form of ClfAD327C/K541C migrated faster on SDS-PAGE than its non-disulfide bonded open form. Presumably, under non-reducing conditions, closed conformation mutants are more compact and migrate faster on SDS-PAGE than open conformation constructs. Under reducing conditions, the disulfide mutant and the wild-type protein migrate at the same rate. Surprisingly, the closed conformation of the disulfide mutant ClfAD327C/K541C was able to bind Fg (Fig. 2C). Elisa-type binding assays where Fg or GST Fg γ1–17 peptide were coated in microtiter wells and incubated with ClfA showed that the closed conformation ClfAD327C/K541C bound the ligand with a much lower apparent Kd (34 nM Fg; 20 nM GST-Fg γ1–17) compared to the wild-type ClfA229–545 (apparent Kd 305 nM Fg; 222 nM GST-Fg γ1–17) (Fig. 2D). These results demonstrate that an open conformation may not be required for Fg binding to ClfA and that Fg binding by ClfA involves a mechanism that is different from the DLL mechanism employed by SdrG.
Crystallization screens were carried out with ClfAD327C/K541C in complex with several N-terminal truncations of the γ1–17D16A peptide that were shown to bind ClfA. Crystals of the stable closed conformation of ClfA229–545 in complex with several peptides were obtained, but structure determination was attempted for only the ClfA(229–545)D327C/K541C-γ5–17D16A peptide. The crystals of the ClfA-peptide complex diffracted to a 1.95 Å resolution. Two copies of the ClfA-peptide complex were found in the asymmetric part of the unit cell and are referred to as A∶C and B∶D. Although the 13 residue Fg γ5–17 chain synthetic peptide was used for crystallization, only 11 residues were identified completely in both copies of the complex. The two molecules of ClfAD327C/K541C (A and B) are nearly identical with rms deviation of 0.3 Å for 312 Cα atoms and 0.55 Å for backbone atoms. As observed in the apo-ClfA221–559 structure [24], the ClfA(229–545)D327C/K541C N2 and N3 domains adopt the DE-variant IgG fold. The overall structure of the ClfAD327C/K541C peptide complex (A∶C) and the two different orientations of the complex are shown in Figure 3A and 3B respectively. The C-terminal extension of the N3 domain makes a β-sheet complementation with strand E of the N2 domain. This conformation is locked by the engineered disulfide bond as predicted by SDS-PAGE analysis (Fig. 2C) and confirmed by the crystal structure (Fig. S1). The two copies of the Fg γ-peptide molecules are nearly identical with rms deviation of 0.5 Å for 11 Cα atoms and 0.89 Å for backbone atoms. The interaction between the ClfAD327C/K541C and the peptide buries a total surface area of 1849 Å2 and 1826 Å2 in the A∶C and B∶D complex, respectively. The interaction of the peptide with the N2 domain is predominantly hydrophobic in nature, in addition to a few main-chain hydrogen bonds (Fig. 3C). Interactions between the Fg peptide and the N3 domain are both hydrophobic and electrostatic with the electrostatic contribution coming almost entirely from the main chain-main chain hydrogen bonds due to the parallel β-sheet formation of the peptide with strand G of the N3 domain (Fig. 3C). The side-chain interactions between the peptide and ClfA are predominantly hydrophobic. The 11 C-terminal residues of the Fg γ-chain peptide sequence that interact with ClfA are composed of only two polar residues, Lys12 and Gln13. Side chain atoms of Lys12 point away and do not interact with the ClfA protein whereas Gln13 makes two hydrogen bonds with the main chain atoms of Ile384 in ClfA (Fig. 3D). A water-mediated interaction is also observed between Gln13 of the peptide and Asn525 of ClfA. Tyr338 in the N2 domain and Trp523 in the N3 domain play an important role in anchoring the peptide molecule. Tyr338 and Trp523 are stacked with residues Gly15 and Gly10, respectively. In addition, Met521 and Phe529 make hydrophobic interactions with Ala7 and Val17, respectively. The C-terminal residues of the peptide Ala14, Gly15, Ala16, and Val17 are buried between the N2–N3 domain interface with the terminal Val residue, presumably threaded through a preformed ligand binding tunnel after ClfAD327C/K541C adopted its closed conformation. A hydrogen bond is observed between Lys389 of ClfA and the C-terminal carboxyl group of the peptide (Fig. 3D). Mutational studies showed that Tyr338Ala and Lys389Ala mutant ClfA showed significantly reduced binding to Fg [24] which corroborates with the structural results. Also an earlier study showed that E526A and V527S affected the binding [32]. The structure shows that these residues make main-chain interactions with the peptide (Fig. 3C). These residues are critical for the anchoring the peptide (Lock) and redirection of the latch.
The individual N2 and N3 domains in the apo-ClfA221–559 and the closed form of ClfAD327C/K541C are almost identical with rms deviations of 0.33 and 0.42 Å for molecule A and 0.35 and 0.42 Å for molecule B, but the relative orientation of the N2 and N3 domains are significantly different (Fig. 4A). This difference affects the association of the N2 and N3 domains. In the apo conformation, the buried surface area between the N2 and N3 domains is 87 Å2 compared to 367 Å2 in the closed form of the ClfA(221–559)D327C/K541C-peptide complex. In the apo-ClfA221–559, the C-terminal residues (Ala528-Glu559) of the N3 domain fold back and do not interact with the N2 domain. Moreover the folded-back segment completely occupies the binding site (Fig. 4B). Therefore, in the folded-back conformation, the ligand binding site appears not to be accessible to the peptide and thus this conformation appears to be inactive. It is presently unclear what the spatial arrangements of the N2N3 domains are in intact ClfA expressed on the surface of a staphylococcal cell. The two structures of these domains solved so far where one is active and the other inactive form suggests a possible regulation of ClfA's Fg binding activity by external factors. One such factor may be Ca2+ which has been shown to inhibit ClfA-Fg binding [32]. Alternatively, it is possible that the folded-back conformation (which is a larger protein construct) is only one of the many possible conformations adopted by the unbound protein. Molecular modeling shows that the two domains in the folded-back conformation could adopt an orientation similar to their orientation in the ClfA-peptide complex (Fig. S2). Most likely, the structural rearrangements responsible for the transition of ClfA from an open unbound to the closed bound form are complex and involve different intermediate forms.
The major difference between Fg-binding to ClfA and SdrG is that the directionality of the bound ligand peptide is reversed (Fig. 4C). The C-terminal residues of the ligand is docked between the N2 and N3 in ClfA and makes a parallel β-sheet complementation with strand G of the N3 domain, whereas in SdrG, the N-terminal residues of the ligand are docked between the N2 and N3 domains and form an anti-parallel β-sheet with the G strand. In both cases there are 11 ligand residues that make extensive contact with the MSCRAMM but with one residue shifted towards the N3 domain in ClfA. Of these 11 residues, 7 and 11 residues participate in the β-strand complementation of SdrG and ClfA, respectively. Although the peptide binding model of ClfA is different to that of SdrG, the inter-domain orientations of the two MSCRAMMS are very similar [25]. Superposition of 302 corresponding atoms in the N2 and N3 domains of ClfA and SdrG showed a small rms deviation of 0.65 Å indicating the high structural similarity between the two MSCRAMMS. Another striking difference is that ClfA does not require an open-conformation for ligand binding, whereas Fg can not bind to a stabilized closed conformation of SdrG. ClfA binds the C-terminal end of Fg and the last few residues of the γ-chain presumably can be threaded in to the binding pocket. In the SdrG-Fg interaction, the binding segment in Fg does not involve the seven N-terminal residues of the ligand and therefore an open conformation may be required for ligand binding.
The C-terminus of Fg γ-chain, which is targeted by ClfA, is also recognized by the αIIbβ3 integrin in Fg induced platelet aggregation, a vital step in thrombosis [10],[33]. The Fg γ-chain complex with αIIbβ3 structure is not available but structures of related complexes provide clues on how αIIbβ3 likely interact with Fg [34]. In addition, the crystal structure of the αvβ3 integrin in complex with an RGD ligand provided a structural model of a similar ligand-integrin interaction [35]. In this structure, the Asp (D) residue of the RGD sequence coordinates with the metal ion in the Metal Ion Dependent Adhesion Site (MIDAS) of the integrin and thus plays a key role in the interaction. The platelet specific integrin αIIbβ3 recognizes ligands with an RGD sequence or the sequence Lys-Gln-Ala-Gly-Asp-Val found in Fg [34]. Structural studies with drug molecules that antagonize the integrin-RGD or -Fg interaction showed that each of the drug molecules contains a carboxyl group moiety that mimics the aspartic acid and a basic group that mimics the Arg (or Lys in the case of Fg) in the ligand [34]. These results suggest that the Lys and Asp residues in the C-terminal γ-chain sequence are critical for the interaction with integrin. Interestingly, our studies have shown that these Lys and Asp residues in Fg are not critical for ClfA binding (Fig. 1B). In fact, substitution of Asp with Ala (γ1–17D16A) results in a higher binding affinity. Absence of a strong interaction with Lys12 in the ClfA-peptide complex structure also correlates with the biochemical data, suggesting that Arg is not a key player in the ClfA-Fg interaction. In general, our studies show that K406 and D410, which are essential for the platelet integrin αIIbβ3-Fg interaction, are dispensable for the ClfA-Fg interaction. To experimentally examine this proposed difference, the ability of the synthesized Fg WT γ1–17 and mutated peptides (γ1–17D16A and γ1–17K12A) to inhibit full length Fg binding to αIIbβ3 was analyzed by an inhibitory ELISA type assay (Fig. 5). The WT γ1–17 peptide completely inhibited the binding of full-length fibrinogen to αIIbβ3 whereas γ1–17D16A and γ1–17K12A weakly inhibited Fg binding to αIIbβ3. These results clearly demonstrated that the γ1–17D16A and γ1–17K12A peptides bind weakly to platelet integrin and therefore could serve as specific antagonists of Fg-ClfA interaction.
Based on the results presented here, we postulate that the mechanism of interaction between ClfA and Fg is a variation of the “Dock, Lock and Latch (DLL)” model of SdrG binding to Fg. In the DLL model of binding, the apo-form of the SdrG is in an open conformation to allow the ligand access to the binding cleft. A closed conformation of SdrG is unable to bind Fg. In the ClfA model, we believe that the peptide may thread into the cavity formed in a stabilized closed configuration and therefore the ClfA-Fg binding mechanism could be called “Latch and Dock”.
In the case of CNA, a collagen binding MSCRAMM from S. aureus, the collagen molecule binds to CNA through a “collagen hug” model [36] which represents yet another variant of the DLL binding mechanism. All three MSCRAMM-ligand structures determined so far, SdrG, CNA and the ClfA have different ligand binding characteristics and mechanisms, although the overall structures of the ligand binding regions of these MSCRAMMs are very similar. These observations suggest that an ancestral MSCRAMM has evolved along different paths to accommodate different ligands without greatly altering the overall organization of the proteins.
The co-crystal structure of ClfA in complex with the C-terminal region of the γ-chain of Fg will allow the design of potent antagonist of the ClfA-Fg interaction. The Fg based peptide analogs that antagonize the ClfA-Fg interaction but not affect the αIIbβ3 integrin interaction could serve as a starting point to develop novel anti-staphylococcal therapeutic agents that do not affect the αIIbβ3.
Escherichia coli XL-1 Blue (Stratagene) was used as the host for plasmid cloning and protein expression. Chromosomal DNA from S. aureus strain Newman was used to amplify the ClfA DNA sequence. All E. coli strains containing plasmids were grown on LB media with ampicillin (100 µg/ml).
DNA restriction enzymes were used according to the manufacturer's protocols (New England Biolabs) and DNA manipulations were performed using standard procedures [37]. Plasmid DNA used for cloning and sequencing was purified using the Qiagen Miniprep kit (Qiagen). DNA was sequenced by the dideoxy chain termination method with an ABI 373A DNA Sequencer (Perkin Elmer, Applied Biosystems Division). DNA containing the N-terminal ClfA sequences were amplified by PCR (Applied Biosystems) using Newman strain chromosomal DNA as previously described [38]. The synthetic oligonucleotides (IDT) used for amplifying clfA gene products are listed in Table S1.
Cysteine mutations were predicted by comparing ClfA221–559 to SdrG(273–597) disulfide mutant with stable closed conformations [31] and by computer modeling. A model of ClfA in closed conformation was built based on the closed conformation of the SdrG-peptide complex [25]. The Cβ-Cβ distances were calculated for a few residues at the C-terminal end of the latch and strand E in the N2 domain. Residue pairs with Cβ-Cβ distance less than 3 Å were changed to cysteines to identify residues that could form optimum disulfide bond geometry. The D327C/K541C mutant was found to form a disulfide bond at the end of the latch. The cysteine mutations in ClfAD327C/K541C were generated by overlap PCR [39],[40]. The forward primer for PCR extension contained a BamHI restriction site and the reverse primer contained a KpnI restriction site. The mutagenesis primers contained complementary overlapping sequences. The final PCR product was digested with BamHI and KpnI and was ligated into same site in the expression vector pQE-30 (Qiagen). All mutations were confirmed by sequencing. The primers used are listed in Table S1.
E. coli lysates containing recombinant ClfA and GST-Fg γ-chain fusion proteins were purified as previously described [32]. PCR products were subcloned into expression vector pQE-30 (Qiagen) to generate recombinant proteins containing an N-terminal histidine (His) tag as previously described [10]. The recombinant ClfA His-tag fusion proteins were purified by metal chelation chromatography and anion exchange chromatography as previously described [23]. To generate recombinant ClfA229–545 and ClfA221–559 proteins, PCR-amplified fragments were digested with BamHI and KpnI and cloned into BamHI/KpnI digested pQE-30. The primers used to generate the recombinant constructs are listed in Table S1. The reactions contained 50 ng of strain Newman DNA, 100 pmol of each forward and reverse primers, 250 nM of each dNTP, 2 units of Pfu DNA polymerase (Stratagene) and 5 µl Pfu buffer in a total volume of 50 µl. The DNA was amplified at 94°C for 1 min, 48°C for 45 sec; 72°C for 2 min for 30 cycles, followed by 72°C for 10 min. The PCR products were analyzed by agarose gel electrophoresis using standard methods [37] and purified as described above.
The ability of the wild-type ClfA229–545 and disulfide ClfA mutants to bind Fg was analyzed by ELISA-type binding assays. Immulon 4HBX Microtiter plates (Thermo) were coated with human Fg (1 µg/well) in HBS (10 mM HEPES, 100 mM NaCl, 3 mM EDTA, pH 7.4) over-night at 4°C. The wells were washed with HBS containing 0.05% (w/v) Tween-20 (HBST) and blocked with 5% (w/v) BSA in HBS for 1 h at 25°C. The wells were washed 3 times with HBST and recombinant ClfA proteins in HBS were added and the plates were incubated at 25°C for 1 h. After incubation, the plates were washed 3 times with HBST. Anti-His antibodies (GE Healthcare) were added (1∶3000 in HBS) and the plates were incubated at 25°C for 1 h. The wells were subsequently washed 3 times with HBST and incubated with goat anti-mouse-AP secondary antibodies (diluted 1∶3000 in HBS; Bio-Rad) at 25°C for 1 h. The wells were washed 3 times with HBST and AP-conjugated polyclonal antibodies were detected by addition of p-nitrophenyl phosphate (Sigma) in 1 M diethanolamine (0.5 mM MgCl2, pH 9.8) and incubated at 25°C for 30–60 min. The plates were read at 405 nm in a ELISA plate reader (Thermomax, Molecular Devices). For the inhibition assays, recombinant ClfA229–545 was pre-incubated with Fg γ peptides in HBS for 1 h at 37°C. The recombinant protein-peptide solutions were then added to plates coated with 1 µg/well GST fusion protein containing the native human Fg γ 395–411 sequence (called GST-Fg γ1–17) and bound protein was detected as described above. If the peptide binds ClfA it would inhibit binding of the GST-Fg γ1–17 to the MSCRAMM.
For αIIbβ3 inhibition assay, Immulon 4HBX Microtiter 96-well plates (Thermo) were coated with αIIbβ3 (0.25 µg/well) in TBS (25 mM Tris, 3 mM KCl, 140 mM NaCl, pH 7.4) over night at 4°C. The wells were washed with TBS containing 0.05% (w/v) Tween-20 (TBST). After blocking with 3% (w/v) BSA dissolved in TBS for 1 h at RT, 10 nM of full length Fg was applied in the presence of either WT γ1–17, γ1–17D16A or γ1–17K12A peptides and plates were incubated at RT for another hour. The bound full length Fg was then detected by goat anti human Fg (1∶1000 dilution, Sigma) antibody followed by horseradish peroxidase-conjugated rabbit anti-goat IgG antibody (1∶1000 dilution, Cappel). After incubation with 0.4 mg/ml of substrate, o-phenylenediamine dihydrochloride (OPD, Sigma) dissolved in phosphate-citrate buffer, pH 5.0, bound antibodies were determined in an ELISA reader at 450 nm. The proteins, antibodies and peptides were diluted in TBST containing 1% (w/v) BSA, 2 mM MgCl2, 1 mM of CaCl2 and MnCl2.
The wild-type and mutated peptides corresponding to the 17 C-terminal residues of the fibrinogen γ-chain (395–411) and truncated versions of this peptide (listed in Figure 2A) were synthesized as previously described and purified using HPLC [10].
The interaction between ClfA proteins and soluble Fg peptides was analyzed by Isothermal titration calorimetry (ITC) using a VP-ITC microcalorimeter (MicroCal). The cell contained 30 µM ClfA and the syringe contained 500–600 µM peptide in HBS buffer (10 mM HEPES, 150 mM NaCl, pH 7.4). All samples were degassed for 5 min. The titration was performed at 30°C using a preliminary injection of 5 µl followed by 30 injections of 10 µl with an injection speed of 0.5 µl/sec. The stirring speed was 300 rpm. Data were fitted to a single binding site model and analyzed using Origin version 5 (MicroCal) software.
The ClfAD327C/K541C protein was purified as described earlier and concentrated to 30 mg/ml. The synthetic γ-chain peptide analogs, P16 and N-terminal truncations of P16 (P16 -2Nt, P16 -4Nt and P16 -6Nt) were mixed with the protein at 1∶20 molar ratio and left for 30 min at 5° C. This mixture was screened for crystallization conditions. Small needles of the ClfA/P16 -2Nt, -4Nt and -6Nt were obtained during initial search of the crystallization condition, but we could only successfully optimize ClfA/P16 -4Nt and ClfA/P16 -6Nt. Diffraction quality crystals were obtained by mixing 2 µl of protein solution with 2 µl of reservoir solution containing 16–20% PEG 8K, 100 mM succinic acid pH 6.0.
Crystals of ClfA/ P16 -4Nt were flash frozen with a stabilizing solution containing 20% glycerol. Diffraction data were measured on Rigaku R-Axis IV++ detector. A total of 180 frames were collected at a detector distance of 120 mm with 1° oscillation. Data were indexed, integrated and scaled using d*terk [41]. The crystals diffracted to 1.95 Å and the data statistics were listed in Table 1. Calculation of the Matthews coefficient suggested the presence of 2 copies of the molecule in the unit cell of the triclinic cell. The structure was solved by molecular replacement (MR) with the program PHASER [42] using individual N2 and N3 domains of ClfA as search model. Solutions for the N3 domain were obtained for the two copies followed by the solutions of N2 domains. Data covering 2.5–15 Å were used for the molecular replacement solution. Electron density maps calculated during the initial rounds of refinement showed interpretable density for 11 out of 13 peptide residues in both the copies of the complex. Modeling building of the peptide and rebuilding of a few loop regions were performed using the program COOT [43]. A few cycles of ARP/WARP [44] were performed to improve the map and for the building of water model. After a few cycles of refinement using Refmac5.2 [45], electron density was clear for only the backbone atoms for two remaining N-terminal residues of the peptide molecule D and one residue for peptide C. The final model of ClfA included residues 230–299, 303–452, 456–476 and 479–545 in molecule A and 230–438, 440–476 and 479–542 in molecule B. The structure was refined to a final R-factor of 21.1% and R-free of 27.9%. Stereochemical quality of the model was validated using PROCHECK [46]. Molecular modeling studies were performed using InsightII software (Accrelys Inc). Figures were made using RIBBONS [47]. The atomic coordinates and structure factors of the complex structure have been deposited in Protein data bank with accession number; 2vr3.
|
10.1371/journal.pntd.0005301 | Review of 21 cases of mycetoma from 1991 to 2014 in Rio de Janeiro, Brazil | Mycetoma is caused by the subcutaneous inoculation of filamentous fungi or aerobic filamentous bacteria that form grains in the tissue. The purpose of this study is to describe the epidemiologic, clinic, laboratory, and therapeutic characteristics of patients with mycetoma at the Oswaldo Cruz Foundation in Rio de Janeiro, Brazil, between 1991 and 2014. Twenty-one cases of mycetoma were included in the study. There was a predominance of male patients (1.3:1) and the average patient age was 46 years. The majority of the cases were from the Southeast region of Brazil and the feet were the most affected anatomical region (80.95%). Eumycetoma prevailed over actinomycetoma (61.9% and 38.1% respectively). Eumycetoma patients had positive cultures in 8 of 13 cases, with isolation of Scedosporium apiospermum species complex (n = 3), Madurella mycetomatis (n = 2) and Acremonium spp. (n = 1). Two cases presented sterile mycelium and five were negative. Six of 8 actinomycetoma cases had cultures that were identified as Nocardia spp. (n = 3), Nocardia brasiliensis (n = 2), and Nocardia asteroides (n = 1). Imaging tests were performed on all but one patients, and bone destruction was identified in 9 cases (42.68%). All eumycetoma cases were treated with itraconazole monotherapy or combined with fluconazole, terbinafine, or amphotericin B. Actinomycetoma cases were treated with sulfamethoxazole plus trimethoprim or combined with cycles of amikacin sulphate. Surgical procedures were performed in 9 (69.2%) eumycetoma and in 3 (37.5%) actinomycetoma cases, with one amputation case in each group. Clinical cure occurred in 11 cases (7 for eumycetoma and 4 for actinomycetoma), and recurrence was documented in 4 of 21 cases. No deaths were recorded during the study. Despite of the scarcity of mycetoma in our institution the cases presented reflect the wide clinical spectrum and difficulties to take care of this neglected disease.
| Mycetoma is a major health problem in tropical areas and is prevalent among people of low socio-economic status. As in many other regions of the world, the incidence and prevalence of mycetoma in Brazil is unknown. This study describes some aspects of mycetoma patients in 24 years of experience at the National Institute of Infectious Diseases at the Oswaldo Cruz Foundation, Rio de Janeiro, Brazil and contribute to the knowledge on mycetoma epidemiology globally.
| Mycetoma is a chronic subcutaneous infections caused by the inoculation of filamentous fungi (eumycetoma) or aerobic filamentous bacteria (actinomycetoma) that form grains in the affected tissues [1]. It´s considered a neglected disease by the World Health Organization (WHO) since 2016 and remains without any control program for prevention or surveillance [1, 2].
Mycetoma occurs worldwide and prevails in tropical and subtropical regions, especially in sub-Saharan areas of Africa, India, and Mexico [3,4]. In South America, cases have been reported in Venezuela, Colombia, Brazil, and Argentina [1,3,5]. The incidence and prevalence of mycetoma in Brazil are unknown, since it is not considered a public health problem, as its frequency is smaller than other diseases such as sporotrichosis, tuberculosis, leprosy, and dengue (the latter two are classified as neglected diseases by the WHO) [6]. Mycetoma evolves slowly in its clinical manifestation. Laboratory diagnosis and treatment are difficult, presenting significant medical, occupational and socioeconomic impacts [2,7].
In this study, we describe the epidemiological, clinical, laboratory, and therapeutic aspects of patients treated at a reference hospital in Rio de Janeiro, Brazil, between 1991 and 2014.
The study was approved by the Research Ethics Committee of the INI / Fiocruz, on November 25, 2013 under the number 477.037. All participants gave their written consent, with the exception of those who died before the study. In all cases the identity and information of each patient were preserved.
Histological examination was performed using haematoxylin-eosin, Grocott’s methenamine silver, Periodic acid–Schiff, and Gram-Brown-Brenn stains. Biopsy specimens were submitted for direct microscopic examination with 10% potassium hydroxide where grains were classified according to their size, shape, colour, consistency and presence of hyphae or filamentous bacteria. Culture on Sabouraud's Dextrose Agar 2% and Mycobiotic Agar was performed for eumycotic grains and in/on Lowenstein-Jensen medium, defibrinated sheep blood agar chocolate agar and thioglycolate medium with resazurina for actinomycotic grains.
Bacterial and fungal etiologic agents were identified by examination of the colonies in culture.
Ultrasound, radiography, computerized tomography (CT), and magnetic resonance imaging (MRI) were performed to determine deep tissue and bone involvement and presence of grains.
Actinomycetoma patients were treated with oral sulfamethoxazole-trimethoprim (SMX-TMP) 800/160 mg BD, alone or in combination with alternate cycles of 15 mg/kg/day intravenous amikacin for three weeks in cases with bone destruction. Other antimicrobials were given in case of secondary infections. Eumycetoma patients were treated with itraconazole (200 mg, BD) alone or, in cases without consistent clinical response after six months, in combination with fluconazole 200 mg/day, terbinafine 250 mg/day, or amphotericin B 1mg/kg. Surgical treatment was indicated for small and delimited lesions and in cases of bone destruction. Amputation was indicated in cases lacking a satisfactory antimicrobial response associated to severe bone destruction of the affected segment.
The patients were followed-up bimonthly at the outpatient clinic to assess clinical responses to treatment and drug side effects. A complete cure was defined with the healing of lesions, bone remodelling, and absence of grains upon imaging examination. After the determination of the clinical cure, outpatient follow-up turned annual, to assess the possibility of recurrence.
Data retrieved from patients records were analysed using descriptive statistics with the Statistical Package for the Social Sciences, version 20.0. Data were summarized as percentages for categorical variables and mean, median, and range for continuous variables.
A total of 21 mycetoma cases were included in the present study: 13 eumycetoma and 8 actinomycetoma patients.
The main sociodemographic aspects of the mycetoma patients are summarised in Table 1. In brief, the male to female ratio was 1.3:1, and the mean age was 46 years old (range 28–93 years). However, the mean age for eumycetoma was 51.3 years old and 38.6 years old for actinomycetoma. The non-white ethnicity/race predominated with 66,66%. Most patients (71.43%) came from the southeast region of Brazil, and 28,57% came from the northeast region. These regions correspond to the possible original infection sites.
Comorbidities occurred in 10 patients. Eight of them presented a single comorbidity and the others had two comorbidities. In general, high blood pressure, diabetes mellitus, HIV positive (Figs 1 and 2), and asthma were found.
The time from onset of signs and symptoms to medical care ranged from 2 to 420 months (mean = 77.68 months). The average time was higher for eumycetoma (mean = 105.76 months) than actinomycetoma (mean = 36.75 months).
The foot (Figs 3, 4 and 5) was affected in 17 cases (80.9%), the thigh was affected in two cases, and the hand and ankle were affected in one case each. A history of trauma was reported in 17 (80.9%) cases.
The grains were mainly identified through histopathological examination with 90.4% positivity in these methods and 9.6% through direct microscopy. We retrieved the etiological agents in 61.5% eumycetoma cases (Table 2) and in 75% of actinomycetoma cases (Table 3).
In the eumycetoma group, the Scedosporium apiospermum species complex was identified in three cases, Madurella mycetomatis was isolated from two cases and Acremonium sp. was isolated from one case. From the remaining two patients, the isolated filamentous fungi could not be identified, as they only produced hyphae without any conidia or spores, and therefore they were named Mycelia sterilia (cases 7 and 8, Table 2). It is important to note that these two organisms were consistently isolated as pure cultures in at least three consecutive mycological examinations.
In the actinomycetoma group we isolated Nocardia spp. from three cases, Nocardia brasiliensis from two cases and Nocardia asteroides from one case (Table 3). In two cases the culture were negative.
All patients underwent radiography of the affected site with exception of patient 7 of Table 3, who underwent complete excision with security margin of the lesion during the diagnostic procedure (Fig 6).
Ultrasonography was performed in 18 cases, with the observation of subcutaneous nodules in all of them. Ten patients underwent CT scans and seven patients underwent MRI. Bone involvement was present in 9 cases, five from eumycetoma and four from actinomycetoma.
Secondary bacterial infection was diagnosed in four cases, two of them had Staphylococcus aureus associated infection treated with systemic antibiotics guided by susceptibility tests. The other two cases were treated empirically.
Patients with eumycetoma received 200 mg BD itraconazole alone (8/13) or in combination with 200 mg/day fluconazole (3/13), or 250 mg/day terbinafine (2/13). In case 1 (Table 2), when the patient became pregnant during itraconazole treatment, this drug was suspended and we tried to use liposomal amphotericin B due to clinical worsening, without success.
Actinomycetoma patients received 800/160 mg sulfamethoxazole-trimethoprim BD in most of cases (75%). As monotherapy in five cases, one case with cycles of 15 mg/kg/day amikacin sulphate and another case received 500 mg cephalexin four times a day. The used of cephalexin occurred because of the repeated secondary bacterial infection. The case 7 (Table 3) with a small and well delimited lesion in lower limb underwent complete excision with security margin and therefore was not treated with antimicrobials. The case 8 (Table 3), who presented with multiple foci of bone destruction, was submitted to amikacin cycles, which had to be stopped after the fifth cycle due to changes in audiometry and increased creatinine levels, without lifelong clinical consequences. Itraconazole was used in all cases and combined with another antifungal agents (38%) in refractory cases.
The average treatment time was 35.04 months (range 6–144 months). The mean treatment time was 42.53 months for eumycetoma and 22.87 months for actinomycetoma. The average treatment time for patients with bone destruction was 70 months (median 55 months) for eumycetoma cases and 33 months (median 36 months) for actinomycetoma cases.
Amputation was recommended for three patients with eumycetoma, one of them accepted the procedure and the other two remain receiving drug treatment until now. In the actinomycetoma group, one patient accepted amputation. Surgical excision of small lesions were performed in nine eumycetoma patients and three actinomycetoma patients.
Clinical cure occurred in 11 (52.38%) of all cases. Of the 13 eumycetoma patients, seven were cured, two abandoned follow up and another two patients are still under antifungal treatment. Of the five eumycetoma patients with bone involvement, one underwent amputation, two remained in treatment, one remain under observation and one abandoned treatment.
Of the eight actinomycetoma patients, four were cured, one abandoned the treatment and three are under treatment. Of the four actinomycetoma cases with bone involvement, one patient was underwent amputation, two remained in treatment and one abandoned treatment. If we consider the cure without sequelae (amputation), the rate falls to 42.8%.
Recurrence of infection was observed in four patients: one with actinomycetoma and three with eumycetoma. The time to recurrence was 24 months for the actinomycetoma case and ranged from 8 to 96 months (mean = 36.6 months) for eumycetoma cases.
Treatment dropouts was high (23%) and recurrence was also frequent (19%) and prevailed in patients that had undergone surgery, especially in the eumycetoma group.
The broad range of treatment duration until clinical cure (6–114 months) was a striking observation of this study.
The 21 mycetoma cases diagnosed in the 24-year period of this study demonstrate the low frequency of mycetoma in our institution at Rio de Janeiro, Brazil. Most reports of mycetoma in Brazil describe one or a few cases, reinforcing the scarcity of the disease in this country. To achieve a better comprehension on this subject we performed a search of articles on PubMed (from 1980 to 2014) using the MESHterms “Mycetoma”, “Actinomycetoma”, and “Eumycetoma” alone or in combination with “Brazil”. During this period, 272 mycetoma cases were reported (Table 4). This number is smaller than that observed in Sudan and Mexico [8, 9,10]. For instance, in Mexico, where 483 mycetoma cases were diagnosed at a single hospital during the same period [11]. In 2013, van de Sande et al. [1] estimated the prevalence of mycetoma cases in Mexico and the Sudan as 0.15 and 1.81 cases per 100,000 inhabitants, respectively, compared to the prevalence of less than 0.001 per 100,000 inhabitants in Brazil.
The predominance of eumycetoma in our study might not represent the real scenery of mycetoma in Brazil, as the Brazilian literature reveals a higher frequency of actinomycetoma (Table 4) [12,13,14,15]. The involvement of male individuals above 30 years old with an acral location likely due to increased risk exposure during labour activity without safety equipment is in accordance with mycetoma characteristics [5,16,17].
Although eumycetoma and actinomycetoma share similar clinical aspects, we noted that eumycetoma cases usually tend to be more silent and chronic, while actinomycetoma cases were more inflammatory and painful. This fact may explain why patients with eumycetoma take longer to seek medical care.
We noted that six of our patients moved from the Northeast region of Brazil to the Rio de Janeiro state, in the Southeast region, probably attracted for job possibilities in a state with higher socio-economic index, higher urbanization of population and better health infrastructure. For this reason, we assume that, for these patients, the place where infection occurred was not in Rio de Janeiro.
Comorbidities are not associated to more severe or atypical forms of mycetoma and there are no changes in the course of mycetoma in the HIV infected patient [18,19,20]. Although it requires further investigation, pregnancy may be linked to more severe clinical course of mycetoma [21,22,23,24] as in case 1 (Table 2) that developed severe bone destruction during pregnancy, resulting in amputation of the affected limb [25].
The mycetoma agents identified in our study are consistent with previous reports. In the actinomycetoma group, Nocardia spp., particularly N. brasiliensis, predominated and in the eumycetoma group, Scedosporium apiospermum. From 1980 to 2014, the main bacterial agents identified in Brazil were Nocardia brasiliensis [15,26,27–32], Nocardia asteroides [15,33], Nocardia caviae [34], Actinomadura madurae [13,35,36], Actinomadura pelletieri [14], and Streptomyces somaliensis [15]. For eumycetoma were Madurella mycetomatis [15,25,37,38], Madurella grisea [39–45], Acremonium falciforme [46], Acremonium kiliense [47], Scedosporium apiospermum [12,18,48,49,50], Fusarium solani [51], Exophiala jeanselmei [44,52,53] and Aspergillus sp. [12].
In our series of cases the diagnosis of mycetoma was made mainly by histopathological examination of affected tissues with visualization of the grains (approximately 91% of cases), while the isolation of the etiologic agent by culture was obtained in 66.6% of cases [15]. Implementation of molecular tools have recently demonstrated an improvement in the sensitivity and specificity in diagnosing mycetoma [16].
Radiography and ultrasonography were the most often used imaging because of their low cost and accessibility. Ultrasonography was crucial in identifying the presence of grains before diagnosis, during and after the therapeutic follow-up. Magnetic resonance imaging is the gold standard imaging method for mycetoma diagnosis and was important to delineate the involvement of internal structures and surgical planning [54]. CT scan was used if no bone involvement was detected by radiography.
Mycetoma treatment is challenging and usually requires long periods of drug therapy with or without surgical procedures (complete excision of the lesion, bone curettage, amputation) [1,5,8,10]. Itraconazole is the most common antifungal agent used for eumycetoma treatment [2]. Voriconazole and posaconazole have been indicated for refractory cases of mycetoma [58] primarily caused by S. apiospermum and Acremonium sp. [48,59–62]. They are expensive in underdeveloped countries and are not available in our institution. Isavuconazole and ravuconazole seem to be satisfactory against M. mycetomatis [63,64] but their effectiveness against other eumycetoma agents need to be investigated.
The first patient in this series of cases was evaluated in 1991, and because of this, the combination of drugs used was based on the available drugs at that time in our institution. The combined itraconazole/fluconazole, and itraconazole/terbinafina treatment in this study was chosen because of our good experience in treating extensive cutaneous lesions of chromoblastomycosis caused by Fonsecaea pedrosoi [55]. However, currently the itraconazole/fluconazole combination for mycetoma is not effective. Although liposomal amphotericin B are no longer recommended for first-line eumycetoma treatment, due to the high minimum inhibitory concentrations required for most eumycetoma agents [16,17,21,56,57], we tried to use only in one case due to clinical worsening during pregnancy, without success [25].
The recommended treatment for actinomycetoma is SMX/TMP as monotherapy or in combination with amikacin sulphate [10]. The association usually gives a cure rate above 90% [2, 65, 66]. Laboratory tests are required to assess possible adverse effects, as ototoxicity (cochlear lesions) and nephrotoxicity, which are permanent injuries, but are not progressive when treatment is suspended. In case 8 of Table 3 a combination with amikacin sulphate was used due to bone destruction. Amoxicillin and clavulanate are alternative drugs during pregnancy, for resistant cases or for patients with adverse effects from aminoglycoside [3]. Rifampicin can be used, but in Brazil it is reserved for tuberculosis and leprosy treatment, diseases with a high burden in our country. Minocycline and moxifloxacin are also treatment options for actinomycetoma [2, 67].
Surgery is indicated for small well localised lesions or in patients who are not responding to medical therapy or to reduce disease burden in massive lesions to allow a better response to medical therapy. [68]. Usually, actinomycetoma require less surgery management then eumycetoma [10]. Amputation are indicated for those patients with massive disease with no response to medical treatment or with massive bone destruction or in case with severe secondary bacterial infection not responding to medical treatment or with severe drug side-effects. [3]
Although our institution has provided all antimicrobials necessary for the treatment free of cost to all patients, the cure rate in this study was low, which reflects the difficulties in treating this disease. Besides the inconvenience to take pills every day for a long period, the total cost of mycetoma treatment is unaffordable for people living in poor regions where the disease commonly occurs. We suggest that the low rate of cure in our study is multifactorial, including the delay to obtain a correct diagnosis, and the scarcity of specialized surgical services with knowledge about this disease that allow the management of the most advanced cases. The postponement of diagnosis favours the occurrence of severe cases that are refractory to the treatment due to the low bioavailability and efficiency of some drugs in advanced lesions. Some patients of our study took more than a year to obtain a correct diagnosis and initiate adequate treatment.
In our cases, treatment dropouts was high and they were likely related to delayed clinical responses and the prolonged treatment times. Recurrence was also frequent [56] and prevailed in patients that had undergone surgery, especially in the eumycetoma group [38]. The reasons are unknown, but may be likely due to the existence of undiagnosed subclinical lesions fungal defence mechanisms against antifungal drugs or incomplete surgical procedures. It is interesting to note that in case 2 (Table 2), the patient was considered clinically cured, but presented recurrence at the eighth year of follow-up [38]. In this case, however, exogenous reinfection cannot be ruled out. We did not observe a relationship between recurrence and a specific etiologic agent.
In rarely cases, mycetoma can spread along the lymphatics to the regional lymph node [6,68]. Few blood-spread mycetoma cases [7,16,69,70,71,72] and deaths related to the infection were reported [4,9,70], but they were not observed in our study.
Although with few cases, this study, highlights the wide spectrum of clinical manifestations of mycetoma, such as localized lesions, bone disease, worsening with pregnancy, recurrence and amputation cases. We also emphasize the challenges to treat and control this neglected disease. The accurate management of each case requires multiple experts including clinicians, surgeons, microbiologists, radiologists working together to assess the best therapeutic approach, which includes a prolonged treatment followed by a long follow up after achieving clinical cure. Rehabilitation is necessary in cases of deformity and amputation, unacceptable sequelae in the 21th century.
|
10.1371/journal.pntd.0002532 | Combinatory Microarray and SuperSAGE Analyses Identify Pairing-Dependently Transcribed Genes in Schistosoma mansoni Males, Including Follistatin | Schistosomiasis is a disease of world-wide importance and is caused by parasitic flatworms of the genus Schistosoma. These parasites exhibit a unique reproduction biology as the female's sexual maturation depends on a constant pairing-contact to the male. Pairing leads to gonad differentiation in the female, and even gene expression of some gonad-associated genes is controlled by pairing. In contrast, no morphological changes have been observed in males, although first data indicated an effect of pairing also on gene transcription in males.
To investigate the influence of pairing on males, we performed a combinatory approach applying SuperSAGE and microarray hybridization, generating the most comprehensive data-set on differential transcription available to date. Of 6,326 sense transcripts detected by both analyses, 29 were significantly differentially transcribed. Besides mutual confirmation, the two methods complemented each other as shown by data comparison and real-time PCR, which revealed a number of genes with consistent regulation across all methods. One of the candidate genes, follistatin of S. mansoni (SmFst) was characterized in more detail by in situ hybridization and yeast two-hybrid (Y2H) interaction analyses with potential binding partners.
Beyond confirming previously hypothesized differences in metabolic processes between pairing-experienced (EM) and pairing-unexperienced males (UM), our data indicate that neuronal processes are involved in male-female interaction but also TGFβ-signaling. One candidate revealing significant down-regulation in EM was the TGFβ-pathway controlling molecule follistatin (SmFst). First functional analyses demonstrated SmFst interaction with the S. mansoni TGFβ-receptor agonists inhibin/activin (SmInAct) and bone morphogenic protein (SmBMP), and all molecules colocalized in the testes. This indicates a yet unknown role of the TGFβ-pathway for schistosome biology leading to male competence and a possible influence of pairing on the male gonad.
| Schistosomiasis is an important infectious disease caused by worm parasites of the genus Schistosoma and directly affects more than 240 million people in 78 tropical and sub-tropical countries but also animals. Pathogenesis is triggered by eggs that are produced by paired females and get trapped in liver and gut causing severe inflammation. While studies have concentrated on the reproductive biology of schistosome females in the past, not much is known about males even though they are indispensable for female sexual development and egg production. Therefore, we studied pairing-dependent processes in S. mansoni males using two independent transcriptomics approaches providing a congruent and most comprehensive data-set on genes being differentially transcribed between pairing-experienced, competent males and pairing-unexperienced, naive males. Besides confirming former studies concerning changes in metabolic processes, our results give new insights into processes leading to male competence indicating among others a potential role of neurotransmitters and TGFβ signal-transduction processes. We especially highlight the follistatin gene SmFst, which codes for an inhibitor of the TGFβ-pathway. SmFst transcription was localized in the testes and found to be down-regulated in pairing-experienced males. This indicates a yet unknown function of pairing on the male gonad and a further role of TGFβ-signaling for schistosome biology.
| Schistosoma mansoni is a species of parasitic flatworms causing schistosomiasis, an infectious disease of worldwide importance for man and animals. Besides vertebrates as final hosts, the parasites' life cycle includes a snail intermediate host, and both are infected by aquatic larval stages. Schistosomiasis occurs in 78, mainly tropical and sub-tropical countries with about 600 million people at risk, of which 243 million required regular treatment in 2011. Thus it is one of the most prevalent parasitemias in the world, second only to malaria [1]–[4]. Pathology is induced by eggs deposited in the bloodstream by paired females, each producing up to 300 eggs per day [5]. A necessity for egg production is the completion and maintenance of the full development of female gonads. For this, the female depends on a constant pairing-contact to a male partner, an exceptional phenomenon in nature. Thus male worms have a key-role in the reproduction biology of schistosomes. Besides causing morphologic alterations comprising a significant increase of the body size of a paired female, which originates from mitogenic processes and differentiation of the reproductive organs ovary and vitellarium, the male even controls gene expression in its partner [6], [7]. While these effects on females have been a strong focus for research throughout the last decades [8]–[11], only few studies concentrated on pairing-dependent processes in the male. Authors of early studies presumed that different factors are transmitted from male to female controlling female body length as well as sexual maturation [12], [13]. Stimulation of the latter was reported to act locally [14], [15], also a tactile impulse was proposed, while sperm or seminal fluid were excluded [12], [14]. Vague evidence was found for a male-secreted hormone or protein to act on the female, however, none of these leads resulted in the identification of a concrete male stimulus [16]–[20]. Since glucose and other substances like cholesterol were shown to be transferred from the male to the female, also the supply of nutrients by the male was suggested to be the basis for female development [21]–[24].
So far only one glycoprotein (GCP = gynecophoral canal protein) was identified [25] and even described to be essential for pairing in S. japonicum [26]. Localization to the male gynecophoral canal as well as on the female surface further indicated the putative importance of GCP for the male-female interaction [27]. Although its function has not been clearly identified yet, evidence was obtained for the regulation of GCP by a TGFβ-dependent pathway in S. mansoni [28]. Other studies indicated that pairing may have an effect on the male as well, by influencing its capacity to stimulate mitosis in the female [29]. This led to the hypothesis that males have to reach a kind of competence before being able to induce developmental processes in the female.
Along with the progress of the genome [30]–[33] and large-scale transcriptome sequencing [34], [35] projects, additional analysis methods became available, which were used among others to compare pairing-experienced (EM) and pairing-unexperienced (UM) males. While the majority of these studies applied microarrays, several groups used serial analysis of gene expression (SAGE) alternatively. One of these studies [36] compared paired adult males and females and their pairing-unexperienced counterparts. For EM and UM the authors found differential regulation for transcripts contributing to developmental processes, metabolism and the redox-system. Already before the genome project was finished, an early microarray-based study compared EM and UM identifying 30 highly expressed genes to be exclusively transcribed in EM and 66 in UM [37]. Their identities indicated RNA metabolic processes to be differentially regulated between EM and UM, which was supported by a subsequent study [38].
Addressing the still unsolved question of male competence, here we investigated the influence of pairing on gene transcription in males. To this end we used two well established transcriptome analysis methods, SuperSAGE and microarray. The combination of both methods aimed at the production of corresponding data sets confirming, but also complementing each other to generate a comprehensive set of differentially transcribed genes in EM and UM that provides new insights into the male-female interaction. Among the most interesting genes identified here was a S. mansoni follistatin homolog (SmFst), a potential inhibitor of TGFβ pathways [39], [40]. Besides its pairing-dependent transcriptional regulation in males, our first functional analyses demonstrated not only gonad-preferential transcription of SmFst, but also its potential to interact with the TGFβ-receptor agonists SmInAct and SmBMP, which colocalized in the gonads. Thus, first evidence was obtained that TGFβ signaling plays an additional role for schistosome biology being one of probably several elements guiding male competence.
The parasite life cycle was maintained using a Liberian isolate of Schistosoma mansoni [41], Biomphalaria glabrata as intermediate snail host, and Syrian hamsters (Mesocricetus auratus) as final host. To produce EM/EF (pairing-experienced males/females) or UM/UF (pairing-unexperienced males/females) snails were infected with either several miracidia (poly-miracidial infection), or only one miracidium (mono-miracidial infection). Poly-miracidial snail infections led to populations of male and female cercariae, which were used for bisex hamster infections resulting in EM and EF. Mono-miracidial snail infections led to unisexual populations of cercariae, which upon final-host infection developed into UM or UF. After 42 days (EM) or 67 days (UM) post infection adult worms were obtained by hepatoportal perfusion. This difference is due to our experience with experimental hamster infections that revealed a positive effect on perfusion efficiency and quality of unisexual worms, when the infection period is elongated to 67 days. This elongation had no influence on further experimental procedures, which concentrated on the comparison of the pairing status of male worms. For EM enrichment, paired males from bisex hamster infections were carefully separated from their partners by feather-weight tweezers, immediately frozen, and stored at −80°C until further use.
All experiments with hamsters have been done in accordance with the European Convention for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes (ETS No 123; revised Appendix A) and have been approved by the Regional Council (Regierungspraesidium) Giessen (V54-19 c 20/15 c GI 18/10).
Total RNA from adult worms was extracted using TriFast (PeqLab) following the manufacturer's instructions. Subsequently extracted RNAs were quality-checked on denaturing formaldehyde gels.
Following total RNA extraction, cDNA synthesis was performed with the Quantitect Reverse Transcription Kit (Qiagen) following the manufacturer's protocol with 1 µg total RNA from EM or UM as template. Standard PCR reactions were performed in a final volume of 25 µl using primer end concentrations of 800 nM, an annealing temperature of 60°C, elongation at 72°C, and FirePol-Taq (Solis biodyne).
Following hamster perfusion, 50 EM or UM were collected, incubated over-night in RNAlater (Ambion) and stored at −80°C. For total-RNA isolation approximately 25 worms from each batch were washed twice in 500 µl H2ODEPC, followed by addition of 1 ml TRIzol reagent (Invitrogen). Subsequently, the worms were homogenized mechanically and incubated for 5 min at room temperature (RT) before 200 µl chloroform was added, mixed for 15 s and incubated for 2–3 min. Following centrifugation for 15 min at 12,000 g and 4°C the upper phase was transferred to a new tube and mixed with 500 µl Isopropanol. After incubation for 10 min at RT the RNA was centrifuged for 10 min at 12,000 g and 4°C. The pellet was washed with 1 ml ethanol (75%) and centrifugation repeated for 5 min at 4°C and 7,500 g. The supernatant was discarded, the pellet dried and resuspended in 25 µl H2ODEPC. Before determination of the concentration on a spectrophotometer (Nanodrop) the RNA was shortly heated to 65°C. RNA purification was done following the animal tissue protocol (Qiagen RNeasy Mini kit) with the following modifications: samples were supplemented to a volume of 100 µl, and 350 µl RLT buffer and 250 µl ethanol (70%) were added before the suspension was transferred onto a column (Qiagen RNeasy Mini kit). RNA was eluted with 30 µl H2ODEPC, and the flow-through put on the column for a second elution step. The concentration of RNA was determined again (see above), and its quality checked on a Bioanalyzer (Agilent 2100 Bioanalyzer, Agilent Technologies).
RNA was reverse transcribed, in vitro amplified, labeled with Cy3 or Cy5, and hybridized according to the Agilent technology protocol for “two color microarray based gene expression analysis”. The samples were hybridized on a 4×44 k oligoarray containing 60 mer oligonucleotides that was custom-designed by us [42], and manufactured by Agilent Technologies; the platform probe sequences are available on Gene Expression Omnibus (GEO) under the accession number GPL8606. This platform was recently re-annotated [43] according to the first draft of the genome project [30]. Three independent biological replicas for EM and UM populations each were used for microarray analyses, each with four technical replicas including dye swaps. Data were extracted using Agilent feature extraction software and raw data are available in NCBI's Gene Expression Omnibus (GEO) [44] under the accession number GSE45696 (subseries number GSE44193). Log2ratios were calculated using LOWESS normalized intensity values of UM and EM (log2 UM/EM) with R [45]. Subsequently, a manual filtering process was applied before statistical analysis, keeping only those oligonucleotides that were defined as representative (unique) for each gene (criteria “to be used in analysis”) according to the information provided together with the re-annotation of the array [43]. Remaining transcripts were submitted to a manual filtering process keeping only those transcripts that were present in all biological replicas, in at least three technical replicas of one biological replica in at least one condition (EM or UM). With these pre-selected data a statistical analysis for microarrays (SAM) [46] was performed using a one-class analysis. Significance cutoff was chosen at a FDR (false discovery rate) of 0.01, and only average log2ratios <−0.585 or >0.585 (which corresponds to a 1.5- fold difference in transcript levels) were defined as relevant. Genes fulfilling these requirements are described in the text as significantly differentially transcribed. The final analysis focused on oligonucleotides representing sense transcripts (although the array also contained oligonucleotides representing putative antisense RNAs for each corresponding gene locus).
Sample collection for SuperSAGE equaled that for the microarray approach. Total RNA was extracted from whole worm batches of 50 males (EM or UM). RNA was quality-controlled on a denaturing formaldehyde gel as well as with a bioanalyzer (Agilent 2100 Bioanalyzer, Agilent Technologies). The following experimental procedure to perform SuperSAGE was done as described previously [47] with minor changes [48]. Raw data were deposited at GEO [44] under the accession number GSE45696 (subseries number GSE45628).
Tags were annotated applying the same procedure as for the re-annotation of the microarray [43] and separated or added up according to their annotation to the predicted exon or intron parts of a CDS in either sense or antisense orientation. Counts were normalized to a library size of 1,000,000, and a filtering process was applied keeping only those transcripts that were detected in two out of three biological replicas in EM or UM. A program implementing the statistical method of Audic and Claverie [49] using a Bio_Sage script (pearl) (http://search.cpan.org/~scottzed/Bio-SAGE-Comparison1.00/lib/Bio/SAGE/Comparison.pm) was used for significance analysis of the data. The statistical cutoff was p<1−10. As with the array, only transcripts with average log2ratios <−0.585 or >0.585 were selected for further analysis. For data comparison only annotated sense transcripts were used, while 7,124 transcripts annotated as antisense as well as 19,610 tags without annotation were excluded from analysis.
Data for sense transcripts from SuperSAGE and microarray were comparatively analyzed using Spotfire [50] and Microsoft-Excel. Smp_numbers without a match in both analyses were manually checked again. The same approach was used for comparative-analysis to other data. An intersection data-set was created, which contained only transcripts detected by both analyses.
These were performed on a Rotor Gene Q (Qiagen) using SYBR-Green MasterMixes (PerfeCTa SYBR Green SuperMix (Quanta) or RotorGene SYBR Green PCR Kit (Qiagen)). PCRs were performed in a total volume of 20 µl, with a three-step thermo-profile and a final melting-curve analysis. Primers, their concentrations, annealing temperatures and efficiencies are listed in Supplementary Table S1. All primers were synthesized by Biolegio (Netherlands). Primer-efficiencies were determined with a standard-curve on diluted gel-eluate with 1∶10 dilution steps [51]. Efficiencies were considered to be optimal between 85–100%. RNAs from EM and UM were evaluated for similar quality on denaturing formaldehyde-agarose (1.2%) gels. cDNAs were diluted 80-fold for usage and added 1∶4 to the final reaction. Absolute quantification was achieved by including the standard curve in each run [52]. Fold changes (EM/UM) were calculated using UM as calibrators [52]. To facilitate comparisons to microarray and SuperSAGE data log2-values of the fold changes were determined as previously described [53], [54]. Standard-curves were performed in duplicate after initial tests for primer concentrations, and reactions on cDNA-samples were performed in triplicates. The significance of individual experiments was checked applying the “Exact Wilcoxon rank sum test” using the exactRankTests package for R [45]; [53]–[57]. Correlation of real-time PCR and transcriptome data was checked with the Spearman's correlation coefficient [53].
For stage-specific detection of SmFst transcripts, cDNAs from EM, UM, EF, UF, miracidia, and cercariae were generated and tested in standard PCR-reactions with primers for SmFst (fwd-5′- TGTTGTAAACGTGGTGGATTC-3′ and rev-5′-CGACATTfTGCATTTTGGTTC-3′) and primers for actin (fwd-5′-GGAAGTTCAAGCCCTTGTTG-3′ and rev-5′-TCATCACCGACGTAGCTGTC-3′) as positive control. PCR-products were separated on a 2% agarose gel.
To obtain organ-specific RNA a recently established protocol was used [58]. In short, adult schistosomes (about 50 individuals) maintained in M199-medium at RT were transferred to reaction vessels containing 500 µl of tegument solubilisation (TS)-buffer (0.5 g Brij35 (Roth), 0.5 g Nonidet P40-Substrate (Fluka), 0.5 g Tween80 (Sigma), and 0.5 g TritonX-405 (Sigma) per 100 ml PBS (137 mM NaCl, 2.6 mM KCl, 10 mM Na2HPO4, 1.5 mM KH2PO4 in DEPC-H2O, pH 7.2–7.4)). Following incubation at 37°C in a thermal shaker (TS-100, Biosan) at 1,200 rpm for 5 min to solubilise the tegument, the musculature was digested by protease treatment. To this end 500 µl elastase-containing medium (Sigma, #E0258; freshly dissolved in non-supplemented M199-medium, 5 units/ml) were used and the worms slightly agitated (600 rpm) in the thermal shaker at 37°C for 30–40 min. Progress of digestion was monitored by microscopic inspections using 20 µl aliquots. Upon the start of tissue fragmentation, reproductive organs such as ovary and testes were liberated. 1 ml non-supplemented M199-medium was added and the content of the vessel decanted to Petri dishes for manual collection of testes and ovaries, which were identified by their characteristic morphologies. If necessary, purification of the gonad tissue from residual parenchyma tissue was achieved by repeatedly collecting and transferring the organs to further Petri dishes containing 2 ml of non-supplemented M199-medium. Finally, the organs were collected using a 10 µl-pipette, transferred to 1.5 ml-tubes, and concentrated by centrifugation for 5 min at 1,000 g, and 1 min at 8,000 g. Following removal of the supernatant, the gonads were immediately frozen in liquid nitrogen and stored at −80°C for further use.
Total RNA was extracted from the organs as described before [58] using the PeqGOLD TriFast reagent (Peqlab; 500 µl TriFast-solution per extraction of 50 testes or 50 ovaries), and the resulting RNA pellet was resuspended in 10 µl DEPC-H2O each. RNA quality and quantity were checked by electropherogram analysis (Bioanalyzer 2100; Agilent Technologies). RT-PCRs were basically performed as described above (standard PCRs) using the following primer combinations to amplify gene transcripts of SmFST (fwd-5′-GAACCAAAATGCAAATGTCG-3′; rev-5′-GCCATGATTGTTCATTCCA-3′), SmBMP (q51- fwd-5′- GTCAAAATGAACAAAATCA-3′; q51- rev-5′- GTTACGTCGAACACTTTG-3′), and SmInAct (q1b-fwd-5′- CACAATTTGGTAATGTTCAACG-3′; q1b-rev-5′- AACTACAAGCACATCCTAAAACAA-3′).
Localization experiments were performed as previously described [59] with the following modifications: hybridization temperature was 42°C, and slides were washed up to 1× SSC. Two different probes were used for detection of SmFst transcripts: probe 1 was 571 bp long (position 208–778), while probe 2 was 306 bp long (position 914–1219). The probe for SmInAct was 440 bp long (position 206–646). Two probes were designed for SmBMP detection: probe 1 had a length of 455 bp (position 2240–2694) and probe 2 was 565 bp long (position 1041–1605).
For SmFst–SmInAct/SmBMP interaction studies, yeast two-hybrid (Y2H) assays were performed. To this end full-length SmFst was cloned into the Gal4-BD vector pBridge using the following primers, which were designed according to the sequence information available at SchistoDB 2.0 [60] for Smp_123300: fwd-5′-GAATTCATGGAAGAGAGTATATCACAATTAG-3′ (italics: cleavage site for EcoRI), rev - 5′-GTCGACTTAGAATAAATTTGAATATTTTCC-3′ (italics: cleavage site for SalI). Full-length SmInAct was cloned into the Gal4-AD vector pACT2, using the following primers: fwd-5′-CCCGGGGATGAATAGAATGTTTAAATTAATAAAAC-3′ (italics: cleavage site for SmaI), rev-5′-CTCGAGTTAACTACAAGCACATCCTAAA-3′ (italics: cleavage site for XhoI). Due to its large size, the sequence for SmBMP was split into four sub-fragments, which were separately cloned into pACT2. The following primers were used: SmBMP-Y2H-Cterm-5′- 5′-CCCGGGGAAACCAAGATCAATTAATTATCCTAAC-3′, SmBMP-Y2H-ncbi-5′– 5′-CCCGGGGATGAACTCAAATATTTTAACAAAATCAG-3′, SmBMP-Y2H-ncbi-Nterm-5′ – 5′-CCCGGGGATGGAAACAGAAAAGACAAAAC-3′, SmBMP-Y2H-overlap-5′ – 5′-CCCGGGTGAAATAAATAGTACATCATTCTACTGG-3′ (italics: cleavage site for SmaI); SmBMP-Y2H-Cterm-3′ – 5′-CTCGAGTTAACGACAAGCACAACTTTC-3′, SmBMP-Y2H-db-Nterm-3′ – 5′-CTCGAGAATTGCTTACATTATTATTATTCAGAGG-3′, SmBMP-Y2H-ncbi-Nterm-3′ – 5′-CTCGAGGTTCTTTAGATGGTTTTCGTATATTATC-3′, SmBMP-Y2H-overlap-3′ – 5′-CTCGAGGATGATTATTTGTTTGTAATACATTTG-3′ (italics: cleavage site for XhoI). PCR products were separated on 1.0% agarose gels. The amplicons were cut out from the gel, and the DNA extracted using the PeqGold Gel Extraction Kit (Peqlab) following the manufacturer's protocol. Extracted fragments were cloned into pDrive (Qiagen) and later regained by restriction-digestion to be again checked for correct size on a 1.0% agarose gel and extracted. Finally, fragments were ligated into pBridge and pACT2, respectively, using T4 Ligase (Promega). Sequences were checked for integrity and a correct ORF by commercial sequencing (LGC Genomics, Berlin).
The SmFst-containing plasmid was transformed in to yeast cells (AH109) together with either one of the other plasmids prepared for the interaction studies. To control successful transformation yeast clones were grown on selection plates (SD-Trp/-Leu/-His/-Ade). β-galactosidase (β-gal) liquid- and filter- assays were performed to confirm interactions (Yeast protocols handbook, Clontech).
The following public domain tools were used: SchistoDB (http://www.schistodb.net/schisto/; [60]), BLASTx (http://www.ncbi.nlm.nih.gov/BLAST), restriction mapper version 3 (http://www.restrictionmapper.org/). Data were analyzed for enriched genes of the ontology categories with Ontologizer [61] using contig annotations, which were the basis for the microarray design [41], [42]. Furthermore, only sense-orientated genes/transcripts were used for this analysis. Network enrichment analyses were done with the Ingenuity Pathway Analysis (IPA) tool (http://www.ingenuity.com; [62]). Only those transcripts with homology to a human molecule >60% and an e-value<10−10 were used, as defined during the re-annotation of the microarray [43]. For SuperSAGE-detected transcripts the according information was obtained from the same source as far as available. For allocation of transcripts to predicted S. mansoni metabolic pathways the function “omics viewer” of the software tool SchistoCyc was used (available at SchistoDB 2.0; [59]. The online-tool SMART (http://smart.embl-heidelberg.de/; [63], [64]) was used to predict protein domains.
Smp_135230 - dopa decarboxylase; Smp_145140 - wnt5A; Smp_155340 – frizzled; Smp_036470 - oxalate-formate antiporter; Smp_135020 - oxalate-formate antiporter; Smp_169190 - tegument protein; Smp_161500 - rhodopsin-like orphan GPCR; Smp_131110 - p14; Smp_000270 - fs800-like; Smp_000430 - ‘eggshell precursor protein’; Smp_00280 - fs800-like transcript; Smp_123300 (KC165687) – S. mansoni follistatin; Smp_090140.2 - Ftz-F1 interacting protein; Smp_090520 - purin nucleoside phosphorylase; Smp_123010 – cationic amino acid transporter; Smp_095360.x – fatty acid binding protein; Smp_065580.x – heterogeneous nuclear ribonucleoprotein k; Smp_033950 - Smad4; Smp_144390 - S. mansoni activin receptor; Smp_049760 – TGFβRI; Smp_093540.3 – ActRI; Smp_124450 - ActRI/BMPRIa; Smp_080120.2 – ActRIIa; Smp_144390 - ActRIIb.
To produce a comprehensive data set of genes differentially transcribed between EM and UM, two methods were chosen. Because both methods have successfully demonstrated their capacities in the past, microarray and SuperSAGE analyses were applied in parallel to generate complementary data sets of genes differentially transcribed between EM and UM. For microarray analyses a S. mansoni-specific 60-mer oligonucleotide microarray platform was used, which represents nearly the complete genome of S. mansoni [42], [43]. The platform-independent SuperSAGE represents a technically improved modification of ‘serial analysis of gene expression’ (SAGE) by generating 26 bp sequence tags of all sample mRNAs containing a NlaIII restriction site by a high-throughput sequencing approach [47]. Combining both methods, we expected to produce data sets complementing each other and providing independent indications for the importance of particular transcripts. To confirm differential transcription of genes from an expected overlap, or from individual data sets outside this overlap, we additionally performed real-time PCR experiments for selected candidates.
In order to facilitate the interpretation of the large scale transcriptome data and selection of molecules for first characterization studies, gene ontology (GO) analysis as well as two further analyses tools, Ingenuity Pathway Analysis (IPA) [62] and the metabolomics tool SchistoCyc [60] were applied. IPA operates with a curated biological knowledge base from the literature to generate molecular networks enriched for proteins encoded by significantly regulated genes from large-scale data sets. Similarly it searches for canonical pathways and predicts the activation of transcription factors. The SchistoCyc function ‘omics viewer’ allocates an uploaded set of genes to predicted S. mansoni metabolic pathways.
The results of the microarray analysis were extracted and evaluated with Agilent feature extraction software. Subsequent data-processing included calculation of log2ratios [53], [65], data-filtering for consistency of transcript detection, and annotation of detected transcripts. After re-annotation of the 44 k oligo array in 2011 [43], 19,197 oligonucleotides were selected as gene representatives: 11,132 detecting RNA in orientation of the predicted transcript, 8,065 detecting RNA complementary to the predicted transcript. Following hybridization and data processing, 10,115 of these representatives were detected as transcribed, 1,966 representing putative antisense and 7,494 representing sense messages. Significance analysis of microarrays (SAM) [46] was performed for all transcripts, and only sense transcripts were analyzed further. Of these, 526 transcripts were significantly differentially transcribed; 229 were up-regulated in EM and 297 up-regulated in UM. Gene ontology (GO) analyses were performed in order to get an overview about the categories these genes could be assigned to. Interestingly, enriched categories were only found for genes up-regulated in UM (Fig. 1).
Again, three samples of EM and UM were collected each, independently from each other as well as from the microarray samples. RNA was extracted from 50 worms per sample and quality-checked. The further procedure was executed by GenXPro according to an internal protocol [47], [48]. SuperSAGE-tags were annotated in the same way as the revised version of the 44 k oligo array [43], and counts for different tags were summed up if they had shown the same annotation. Similar to the microarray data-analysis, transcripts were classified into sense- and antisense orientation relative to the protein-coding gene in a given locus. Additionally, SuperSAGE-transcripts were further distinguished between predicted intron- or exon-sequences. Thus four categories of SuperSAGE-detected tags were defined, and accordingly up to four transcripts could represent one gene. This explains the large number of 25,597 gene-specific transcripts detected with SuperSAGE, which exceeds the assumed number of S. mansoni genes more than twice [33]. Statistics were performed according to the method of Audic & Claverie [49]. Subsequent to this significance analysis, 5,987 transcripts without annotation and 7,124 transcripts classified as antisense were excluded from further analyses. Of the remaining 12,486 sense transcripts, 8,969 were classified as exon and 3,517 as intron sequences. Taking out redundancy for genes represented by an exon as well as an intron sequence, sense transcripts were found for a number of 9,344 unique genes. Of these, 5,601 genes had representatives in both groups exons as well as introns. For 2,581 genes only exon-representing sense transcripts were detected, and for 219 genes only intron-representing sense transcripts were detected.
Selected from these were candidates showing a normalized detection value of at least 10 tags in at least one library and significantly differential regulation between EM and UM. These criteria were met by 253 transcripts. Of these 218 were up-regulated in EM and 35 in UM. GO analyses showed enriched categories neither for the transcripts up-regulated in EM nor for those up-regulated in UM.
Comparing sense transcripts detected by microarray and SuperSAGE, 6,326 transcripts were found by both methods, while additional 3,018 and 1,168 were exclusively detected by SuperSAGE or microarray, respectively (Fig. 2). The number of counts representing genes being differentially transcribed between EM and UM varied between the methods used. This was influenced by the underlying, method-specific statistics leading to subsets of counts that were significant for only one of the two data sets although the same transcripts were also present in the other data set (Fig. 2). The stringent analysis criteria for differential regulation were met according to both approaches by 29 transcripts. Among these were genes with various biological functions such as metabolism (NAD-dependent epimerase/dehydratase, sodium dicarboxylate cotransporter, fatty acid acyl transferase-related, oxalate-formate antiporter), neurotransmitter synthesis (aromatic amino acid decarboxylase), enzyme activity (kunitz-type protease inhibitor), microfilament organization (villin, nebulin), membrane dynamics/vesicle formation (endophilin b1, snf7-related), molecular interaction/communication (surface protein PspC, cadherin), calcium metabolism (sarcoplasmic calcium-binding protein), chromatin organization (histone h1/h5), signal transduction (pinch, dock, follistatin), and others less well defined (cancer-associated protein gene, loss heterozygosity 11 chromosomal region 2 gene) (Table 1).
Differentially transcribed genes from either microarray or SuperSAGE were assigned to predicted S. mansoni metabolic pathways using the “omics viewer” function of SchistoCyc [60] (Supplementary Table S2). Comparing the results of these analyses, transcripts coding for enzymes involved in carbohydrate metabolic processes, citrate cycle, aerobic respiration and amino acids metabolic processes were mostly down-regulated in EM compared to UM. For base metabolic processes, transcripts for enzymes involved in synthesis were rather down-regulated in EM, while transcripts for enzymes involved in degradation and salvage pathways were up-regulated in EM. Several transcripts with different directions of regulation were found for enzymes participating in lipid and fatty acid metabolic processes. Interestingly, data sets for differentially transcribed genes contained none coding for enzymes related to pentose-phosphate cycle or glycolysis. Only one enzyme within the SuperSAGE/microarray-intersection of differentially transcribed genes was allocated to metabolic pathways, aromatic amino acid decarboxylase (dopa decarboxylase, DDC, Smp_135230/NP_001076440.1). DDC was up-regulated in EM and allocated to phenylethanol and catecholamine biosynthesis. This indicates a pairing-dependent adaptation for usage of neurotransmitters like dopamine can be assumed.
IPA was performed for the data-overlap of transcripts differentially regulated according to both transcriptome analyses. While no significantly enriched canonical pathways were detected, one significantly enriched network was found, namely ‘embryonic development, hair and skin development and function, organ development’. It contained nine differentially transcribed genes, seven up-regulated in EM and two up-regulated in UM, including SmFst (Supplementary Table S2, Supplementary Figure S1). Thus IPA additionally highlighted SmFst and a putative differential regulation of neurotransmitter synthesis through DDC.
Looking for explanations why transcripts were detected by one method only without complementary counterpart, these specific transcript groups were selectively analyzed in more detail. Out of the 47 transcripts differentially regulated in the SuperSAGE-only group, oligonucleotides existed on the microarray for 25 out of 47, thus 22 transcripts were newly detected by SuperSAGE (Table 2). Next, transcripts within the microarray-only group (110) were analyzed for the presence of NlaIII restriction-enzyme recognition sites. cDNA synthesis and NlaIII restriction are the first two steps during the SuperSAGE procedure. Transcripts lacking the restriction site or with an NlaIII restriction too close to the polyA-tail will not be detected. As a snap sample of 110 candidates of the microarray-only group, we selected 10 which not only had a Smp_number but also a functional annotation other than ‘hypothetical protein’. Of these 8 had NlaIII restriction sites within their predicted CDS (Table 3). Thus indeed, the existence of NlaIII restriction may have influenced the detection of transcripts by SuperSAGE. Also other methodological differences may have contributed to the resulting number of transcripts detected by both methods, although not as differentially transcribed in both analyses. While in microarray experiments light intensities are measured, during SuperSAGE the number of transcripts is counted, resulting in different log2ratios (some below the selected threshold of log2ratio = 0.585). Also the number of technical replicas differed. While three biological replicas without technical replicas were used for SuperSAGE, four technical replicas for each biological one were done in case of the microarrays, which is part of the methodological procedure. Furthermore, different statistics had to be used for the analyses of the results from both methods, which influenced the outcome with regard to the significance of the differences detected between EM and UM.
For a sub-set of 21 transcripts we performed real-time PCR experiments (i) to confirm the results obtained in the combinatory analysis, and (ii) to confirm those microarray or SuperSAGE data, for which no complementary results existed. With respect to their putative biological relevance, transcripts for signal transduction molecules, surface molecules, metabolism-associated proteins and transcription factors were chosen for verification (Supplementary Table S1). Additionally, the regulation of three transcripts for egg-shell precursor proteins was analyzed. A Wilcoxon rank sum test for real-time PCR data detected significant differences between EM and UM transcript levels for most of the differentially transcribed genes present within the overlap of microarray and SuperSAGE data (Supplementary Table S1). The test also confirmed significant differences between the two male stages for two genes detected by only one of the transcriptome studies. Using the Spearman's rank correlation coefficient, overall results from real-time PCR were compared to those from either one of the two transcriptome approaches. Data correlated significantly (p<0.01) between real-time PCR and microarray (r = 0.676) or SuperSAGE (r = 0.621) each.
Representatives for signal transduction-associated transcripts were SmFst, dock, and pinch (Figure 3; Supplementary Table S1). For all three genes transcription regulation was confirmed by real-time PCR. Comparing the results of all approaches, dock and pinch showed a slight biological variation within only one of the three analyses. In contrast, the transcriptional activity of SmFst was found to be consistently down-regulated without exception in EM (Figure 3), which substantiated the results of the transcriptome approaches. Since follistatins (FSTs) have regulatory function for TGFβ-signaling [39], [40], we additionally tested the transcriptional activities of further genes coding for members of TGFβ-pathways (Supplementary Table S1). In those cases where microarray or SuperSAGE indicated significant regulation for a TGFβ-pathways member, this finding was not confirmed by the complementary method or real time PCR.
Furthermore, we included wnt5A (Smp_145140) and one of its potential receptors, the seven transmembrane receptor frizzled (Smp_155340) into this analysis, since wnt-pathways are linked to developmental processes [66]. According to the microarray data Wnt5A and frizzled transcripts were significantly down-regulated in EM. The direction of regulation of both transcripts detected by SuperSAGE varied strongly between the biological replicates, and differences between EM and UM were not significant. This was confirmed by real-time PCR experiments (Figure 3, Supplementary Table S1), indicating that the wnt-pathway and/or associated members may not be essential with respect to EM/UM differences and are presumably influenced by additional biological parameters beyond pairing.
The AMP-dependent ligase attracted attention due to its uniform differential transcriptional regulation in both approaches and its putative function in DNA synthesis processes. Besides SmFst, it was one of the transcripts for which real-time PCR supported the previous findings, but in contrast to SmFst, transcription of the AMP-dependent ligase was strongly up-regulated in EM (Figure 3). A similar consistent result was obtained for one of the oxalate-formate antiporters (OxlT) (Smp_036470), whose transcript amount was lower in EM. In contrast, a second OxlT (Smp_135020) was detected as significantly up-regulated in EM by the microarrays only, which was confirmed by real-time PCRs. Further transcripts for membrane proteins analyzed in real-time PCR experiments were a tegument protein (Smp_169190), significantly up-regulated in EM according to the microarray analysis (Supplementary Table S1) and a rhodopsin-like orphan GPCR (Smp_161500) up-regulate in EM according to the SuperSAGE analysis (Figure 3). Both transcripts were not detected by the respective other transcriptome analysis. Real-time PCRs confirmed the presence and direction of regulation for both transcripts substantiating that microarray and SuperSAGE experiments can complement each other. Enhanced transcript levels of the rhodopsin-like orphan GPCR in EM were also confirmed by semi-quantitative real-time PCR [data not shown]. With respect to the fact that the physical contact between the genders stimulates differentiation processes in the female, tegumental proteins as such are potentially important for male-female interaction. Since GPCRs represent the biggest receptor class in schistosomes [66], and since they are known to be involved among others in regulating differentiation processes in diverse organisms [67]–[70], there may be candidates with roles in male-female interaction as well.
Surprisingly, transcript amounts for the egg-shell precursor protein p14 (Smp_131110) were significantly elevated in EM according to the microarray analysis. Our real-time PCR experiments, however, demonstrated strong variations between biological replicas, previously also indicated by SuperSAGE results. Analogous results were obtained for two other egg-shell precursor transcripts, fs800-like (Smp_000270), and ‘eggshell precursor protein’ (Smp_000430) (data not shown). Thus it seems obvious that egg-shell precursor protein transcripts can be strongly up-regulated in EM compared to UM, but this depends on the worm batch. Indirect support for this interpretation comes from a previous study, which provided evidence for another fs800-like transcript (Smp_00280) to be up-regulated in testicular lobes of paired males [71]. Other transcriptome studies, comparing EM and UM, also found transcripts of supposedly female-specifically expressed genes in males [37]. These findings can be explained by leaky expression control of such gender-associated genes, a phenomenon that can even lead to the development of female reproductive tissue in males as previously observed [11], [72]–[74].
The follistatin homolog SmFst appeared as one of the most interesting candidates for first functional characterization because of its consistent down-regulation in EM in all analyses, and its putative participation in schistosome TGFβ signaling processes. Based on the gene prediction (Smp_123300) obtained from the Schistosoma genome project, primers were designed to amplify its full-length cDNA. Following cloning and sequencing, minor differences were detected between SmFst (KC165687) and Smp_123300, revealing parts of predicted intron sequences as exon stretches (Supplementary Figure S2).
While other typical Fsts contain three follistatin-domains (FstDs) [75], SmFst encodes an open reading frame containing two, according to a SMART-domain analysis. FstDs are further subdivided into an EGF (epidermal growth factor)- and a Kazal (a protease inhibitor)-domain, which according to SMART do not follow each other directly in SmFst (amino acid positions: EGF-domains: 52–74, 312–337; Kazal-domains: 264–307, 398–434), as is the case for other Fsts [75]. The first EGF-domain of SmFst is located upstream of the first EGF-domain of other Fsts, and the second EGF-domain of SmFst is homologous to the third one of other Fsts. The first and second Kazal-domains of SmFst are homologous to the second and third Kazal-domains of other Fsts.
To confirm the presence of SmFst in different life cycle stages of S. mansoni RT-PCR reactions were performed with RNAs from EF (pairing-experienced females), UF (pairing-unexperienced females), EM, UM and the free-living larval stages. SmFst transcripts were detected in all adult stages and miracidia, but not in cercariae. As positive control actin was amplified from all tested life cycle stages (Supplementary Figure S3). Localization studies by in situ-hybridization with couples as well as UM detected SmFst transcripts within the testicular lobes of both male groups and also in the vitellarium and ovary of female worms (Figure 4). Results concerning the detection of sense or antisense transcripts varied with probe sequence and probe batch. Organ-specific RT-PCR [58] confirmed the presence of transcripts for SmFst within ovary and testes (Supplementary Figure S4).
Fst was shown before to bind activin, which itself is an agonist of TGFβ-pathways [76], [77]. Although with lower affinity it also binds to a bone morphogenic protein (BMP), another agonist of TGFβ-pathways [78]–[81]. SmInAct, a Schistosoma activin-inhibin, was previously characterized [82] as well as SmBMP [83]. In our study, SmInAct transcripts were detected only by SuperSAGE but without differential regulation between EM and UM, which was also confirmed by real-time PCR results. SmBMP was down-regulated according to the microarray data, however, this was neither confirmed by SuperSAGE nor by real-time PCR. To obtain evidence on possible interactions between SmFst and SmInAct and/or SmBMP, co-localization and Y2H interaction studies were performed.
Previous studies had localized SmInAct transcripts in the ovary and vitellarium of EF, without providing information on their presence in testicular lobes [82]. The results obtained in our study using two replicas with a hybridization probe based on the same sequence as used previously [82] showed sense and also antisense transcripts exclusively in the ovary of EF. No signal was detected within the vitellarium or testes of EM or UM (Figure 5). However, organ-specific RT-PCR on ovary and EM testes detected transcripts not only in the ovary but also in testes of EM (Supplementary Figure S4). For the localization of SmBMP two different probes were used. Transcripts were detected in all reproductive organs of couples as well as UM, but especially for probe 2 in the area around the ootype (Figure 6). Again, sense and antisense transcripts were detected, and results for EM-testes and ovary were confirmed by organ-specific RT-PCR (Supplementary Figure S4).
Even though the major interaction partner of Fsts is activin, it can also bind to BMP [79]. For males our in situ-hybridization experiments rather indicated physical proximity of SmFst transcripts to those for SmBMP, as both transcripts were detected within testicular lobes. In females, transcripts for all three molecules were detected within the ovary by in situ-hybridization, though organ-specific RT-PCRs also indicated the presence of SmInAct in male testes.
To provide evidence for protein interaction between SmFst and SmBMP or SmInAct, Y2H interaction studies were performed. To this end full-length SmFst was cloned into the Gal4-BD vector pBridge, while full-length SmInAct and four different sequence stretches of SmBMP were cloned into the Gal4-AD vector pACT2. The SmFst-containing plasmid was transformed into yeast cells (AH109) together with SmInAct or one of the BMP variants each. Following growth on selection plates (SD-Trp/-Leu/-His), β-gal liquid- and filter-assays were performed to relatively quantify interactions. These were shown between SmFst and SmInAct as well as between SmFst and SmBMP. The result of the β-gal liquid assay indicated a stronger binding of SmFst to SmInAct, while no evidence was obtained for an interaction with SmBMP-C-term or negative controls (Figure 7). This is surprising since this part of SmBMP is most conserved in comparison to human BMPs and contains several residues essential for receptor binding, which are supposedly blocked by FST [76], [84]–[87]. However, it is possible that within its quaternary structure FST binds to other residues, besides those at the C-terminus, thereby prohibiting the receptor binding of BMP. Together with the results from the localization experiments, evidence is provided that SmFst interacts with SmInAct and/or SmBMP in the testes of EM and UM but also in the female ovary, and in case of SmBMP in the vitellarium.
The results of our study provided conclusive evidence for pairing-influenced transcriptional processes in males. Together with previous findings about pairing-dependent gene transcription in females and first transcriptome analyses for males, all data clearly demonstrate a bidirectional transcriptional influence during male-female interaction [6], [7], [36], [37], [88], [89]. Our expectation to find a comprehensive data set of genes differentially transcribed in males upon pairing was met, just as obtaining congruent and complementing results using two independent techniques. Since both are basically different, each generated additional data that were not obtained by the other method. From our point of view, both methods have their advantages and disadvantages, and none seemed superior over the other. Both methods required the application of different statistics since on the one hand signal intensities were determined (microarray) including technical replicas and on the other hand transcript counts, both resulting in different log2-ratios. Finally, also biological variability, a well-known phenomenon even within schistosome strains used for such kind of analyses [54], may have had an influence on the data obtained. Although different methodological and analytical approaches were applied, an overlap of interesting genes was obtained on the basis of stringent analysis criteria. Taking into account that we made use of a second-generation microarray, which contained the majority of genes known from S. mansoni [42], [43], as well as SuperSAGE, which can theoretically detect all genes transcribed [47] the most complete data set of genes differentially transcribed between EM and UM was obtained. The credibility of the data was confirmed by the intersection of differentially transcribed genes identified by both techniques, but also by additional real-time PCR experiments. Furthermore, a comparison to data previously generated in other studies [36], [37], [90], [91] supported the reliability of our results. Waisberg et al. [90] found a number of genes for which transcription in males was influenced by final-host sex. Analyzing the transcription of the top 30 of these genes with our data sets revealed no significant differences between EM and UM. Thus an influence of the host sex within our experimental setup can be excluded. Of the transcripts differentially regulated between male stages or strongly up-regulated in at least one male stage found by Fitzpatrick et al. [37], we found 7 to be differentially regulated by SuperSAGE, including a Ftz-F1 interacting protein (Smp_090140.2), and 14 were differentially regulated according to microarray analysis, including several genes encoding metabolism-associated proteins. From the overlap of differentially transcribed genes between SuperSAGE and microarray only two genes, the AMP-dependent ligase and dock, were found within the data-set obtained by Fitzpatrick et al. [37]. Here, transcriptional regulation showed the same direction as in our study. Our data also confirmed results of one study previously applying SAGE to reveal differences between EM and UM [36]. As far as comparison was possible, a purin nucleoside phosphorylase (Smp_090520) was up-regulated in our microarray analysis as well as in the dataset of Williams et al. [36]. Also a cationic amino acid transporter (Smp_123010) (down-regulated in EM), a fatty acid binding protein (Smp_095360.x) (down-regulated in EM), and a heterogeneous nuclear ribonucleoprotein k (Smp_065580.x) (up-regulated in EM) showed the same direction of regulation in our study and were significant within our SuperSAGE data set.
Among the transcripts identified in our study to be regulated between EM and UM were also putative antisense RNAs. Some of these may have yet unknown protein-coding function, however, the majority of these RNAs are probably non-coding RNAs (ncRNAs). Their discovery in eukaryotic genomes has significantly influenced research recently, and first evidence for regulatory functions of ncRNAs has been obtained [91]. Also for S. mansoni the existence of antisense RNAs was reported [42], and it was estimated recently that ≥10% of the transcriptome may represent ncRNAs [43]. Life-stage analyses indicated alterations in the occurrence of specific ncRNAs pointing to diversified functions in biological processes [43], and we observed antisense RNAs (microarray: 211; SAGE-exon: 261; SAGE-intron: 107) that are differentially transcribed between EM and UM. Their analysis will be subject of future studies, when more knowledge about this class of molecules will be available for schistosomes. This applies also for the number of “hypothetical proteins” identified as being differentially transcribed (Supplementary Table S4), which could add up to the list of interesting candidates for further analyses of their potential function during pairing-associated processes in males.
Applying stringent analysis criteria and focusing on sense-transcripts only, a number of candidate genes were identified, which may be responsible for male competence and/or inducing female maturation. Data interpretation based on a combination of bioinformatics tools permitted first important conclusions (i–iv). According to GO-analyses (i) EM seem to loose complexity with regard to functional categories. This interpretation is supported by a previous study, in which less enriched GO categories were found in EM applying a first-generation microarray containing a lower number of gene-representing oligonucleotides [37]. In the same study the authors concluded that worms from mixed-sex populations are transcriptionally less complex than those from unisexual populations. Because in females paired to UM the induction of mitotic activity was found to be delayed compared to the situation in females paired to EM [29], it was hypothesized that EM and UM differ with respect to their mitosis-inducing capacity, what we like to define as competence that has to be reached before males have the full capacity to govern developmental processes in their pairing partners. The biological variety of transcripts belonging to distinct functional categories found to be differentially regulated indicated that (ii) gaining male competence is a process in which different systems are involved. This may also apply to male factors inducing female maturation. Although the involvement of neuronal processes [15] as well as sperm or seminal fluid [12], [14] as players during male-female interaction was dismissed in the past, (iii) we obtained first evidence that neuronal and testes-associated factors nevertheless may be involved. IPA analysis for genes of the intersection highlighted among others DDC, which was further accentuated through the metabolomic data-analysis with DDC as one of two molecules identified. Thus, together with the enhanced DDC transcript level in EM, our data suggest the possibility that neurotransmitters such as dopamine could play a role during male-female interaction. With regard to testes-associated genes, (72 genes were found in our study to be differentially transcribed according to microarray or SuperSAGE, which were previously detected as transcriptionally up-regulated in EM testes compared to whole worms [55]. These included dock and the OxlT (Smp_135020). Our in situ-hybridization experiments for two members of TGFβ-pathways as well as the OxlT (Smp_036470) (data not shown) localized transcripts for these molecules to the testicular lobes of EM and UM. The differential regulation of molecules like OxlTs, known for their participation in the indirect proton pump of Oxalobacter formigenes [92], (iv) may indicate their pairing-dependent function in metabolism processes. Previous studies already suggested that EM support females by supplementing their partner with nutrients [21]–[24], for which they might need a wider functional assembly of molecular processes than UM. In this context it is noteworthy that base metabolic processes seem to differ between the two male stages for anabolic and catabolic pathways, indicating higher nucleic acid synthesis rates in UM. Also, transcription of enzymes involved in carbohydrate metabolic processes, citrate cycle, aerobic respiration, and amino acids metabolic processes was rather up-regulated in UM. Metabolic differences between EM and UM were also found by Williams et al. [36].
The GCP protein was previously reported to be up-regulated in EM as a result of the male-female interaction [25], [27], [93] and proposed to be essential for pairing in S. japonicum [26]. However, transcriptional differences between EM and UM for GCP were neither confirmed by our data, nor by previous transcriptome analyses [36], [37], [88], [89]. This discrepancy could be explained by post transcriptional and/or post translational regulations. Interestingly, GCP was proposed to be a downstream target of TGFβ-pathways in schistosome couples [28]. This pathway is well known in schistosomes and was previously pointed out for its possible importance in the female reproductive biology being involved in regulating mitosis and egg production [9], [10]. Here we provide first evidence for an additional role of TGFβ-pathways during schistosome development as shown by the discovery of SmFst, a potential inhibitor of the TGFβ-pathway. Besides its down-regulation in EM, confirmed by all analyses, SmFst stood out in the IPA network identified for genes within the intersection of microarray and SuperSAGE data. Besides SmFst and its two potential interaction partners SmInAct and SmBMP, two other members of TGFβ-pathways were tested in real-time PCR experiments, Smad4 (Smp_033950) and one S. mansoni activin receptor (Smp_144390). Apart from SmFst none of the transcripts was differentially transcribed between EM and UM.
First functional studies demonstrated that SmFst transcripts were present in male testes and female reproductive organs. Furthermore, Y2H experiments confirmed its potential to interact with SmInAct and SmBMP. In in situ-hybridization experiments SmFst and SmInAct both localized in the female ovary, while SmFst and SmBMP each localized in male testes and the female reproductive organs. In addition, organ-specific RT-PCRs indicated the presence of all three transcripts in EM-testes and the ovary. A previously described transcriptome study [37] did not detect SmFst to be differentially regulated between EM and UM. This may be due to the absence of the corresponding oligonucleotide on the first-generation microarray used (the respective annotation was not found in the data-set [37]), which represented about 50% of the S. mansoni transcriptome. Within the S. mansoni organ-specific transcriptome data [71] SmFst was not found to be up-regulated in EM-testes compared to whole worms using a cutoff for 2-times higher transcription. Compared to our results, this seems not surprising since the EM transcript-level of SmFst was generally low, about two times weaker than in UM according to the calculated log2ratios.
With respect to transcript detection in the ovary, results of a previous report on SmInAct [82] corresponded to our localization data. While similar SmInAct transcript levels between EM and UM were detected in the earlier study and our analysis, SmInAct protein was only detected in EM and EF but not in UM or UF before, which suggested that SmInAct expression is linked to the reproductive capacity of the worm [82]. Assuming that SmInAct is the preferential binding partner of SmFst as shown in other organisms [78], [79], [94] two possible scenarios exist. First, transcript levels detected for SmFst in this study may not be representative for translation, and thus SmFst interactions would not occur in UM. Secondly, SmFst is translated and may influence TGFβ-pathways in the male testes, however, through other TGFβ-agonists such as SmBMP. Indeed, besides our Y2H results interactions between SmFst-SmBMP have been shown also in other organisms, where they are among others involved in gonad-specific developmental and physiological processes [64], [65], [95]–[100]. However, although the presence of SmBMP protein was demonstrated, it was not detected within the testes yet [83], which may have been caused by protein amounts below the detection limit.
From all results available today, we hypothesize a tissue- and stage-dependent interplay of TGFβ-family proteins in schistosomes that also affect the gonads. This view is supported by the presence of various type I and II receptors in the S. mansoni genome [30], [33]. Most type I receptors belong to the sub-group of activin-like receptors. Among these is an alternatively spliced variant of the S. mansoni TGFβRII [28], previously described as ActRII [98]. Other members are TGFβRI [100] (Smp_049760), ActRI (Smp_093540.3), ActRI/BMPRIa (Smp_124450), ActRIIa (Smp_080120.2) or ActRIIb (Smp_144390). Since multiple receptor-combinations are possible as well as their activation by promiscuously acting agonists [101], numerous interactions are imaginable with the potential to govern tissue-specific activities. In this scenario, SmFst may regulate TGFβ signaling by binding agonist such as SmBMP in testes of UM, which may impede processes not needed or not intended to occur before pairing. Whether the role of SmFst in schistosomes covers modulating agonist activities in the extracellular environment, or whether it is involved in processing SmInAct or SmBMP pre-pro-peptides to become pro-peptides, a normal part of the activation of these agonists [102], remains unclear at this stage of its analysis but will be subject of further studies. Although typical furin cleavage sites, being necessary for processing BMP or activin proteins in vertebrates [103], are present in the schistosome homologs, a role of SmFst in SmInAct or SmBMP protein processing appears unlikely since the main function of FSTs known today is its antagonist activity in the extracellular environment. Here it was shown that FSTs among others play prominent roles in the gonads controlling different testicular and ovarian functions including cell proliferation, apoptosis, folliculogenesis, luteogenesis, hormone release and fertility [104], [105].
In summary, the presented results demonstrate that processes leading to male competence may be far more complex than hypothesized before by reports suggesting that single molecules from the male and/or nutritional support could be in charge of the fundamental consequences of pairing on female development. From our results we conclude that besides metabolic processes, neuronal processes may be involved in the initial phase of male-female interaction but also TGFβ-signaling, which has been described before to be involved in differentiation processes in fully developed females [9], [10] and embryogenesis [82]. The meaning of this pathway for schistosome biology appears to go beyond that, since SmFst has leaped into view emerging as a regulatory molecule for TGFβ signal-transduction pathways that is pairing-dependently transcribed in the male gonad probably contributing to processes leading to male competence.
|
10.1371/journal.pntd.0004002 | Meteorologically Driven Simulations of Dengue Epidemics in San Juan, PR | Meteorological factors influence dengue virus ecology by modulating vector mosquito population dynamics, viral replication, and transmission. Dynamic modeling techniques can be used to examine how interactions among meteorological variables, vectors and the dengue virus influence transmission. We developed a dengue fever simulation model by coupling a dynamic simulation model for Aedes aegypti, the primary mosquito vector for dengue, with a basic epidemiological Susceptible-Exposed-Infectious-Recovered (SEIR) model. Employing a Monte Carlo approach, we simulated dengue transmission during the period of 2010–2013 in San Juan, PR, where dengue fever is endemic. The results of 9600 simulations using varied model parameters were evaluated by statistical comparison (r2) with surveillance data of dengue cases reported to the Centers for Disease Control and Prevention. To identify the most influential parameters associated with dengue virus transmission for each period the top 1% of best-fit model simulations were retained and compared. Using the top simulations, dengue cases were simulated well for 2010 (r2 = 0.90, p = 0.03), 2011 (r2 = 0.83, p = 0.05), and 2012 (r2 = 0.94, p = 0.01); however, simulations were weaker for 2013 (r2 = 0.25, p = 0.25) and the entire four-year period (r2 = 0.44, p = 0.002). Analysis of parameter values from retained simulations revealed that rain dependent container habitats were more prevalent in best-fitting simulations during the wetter 2010 and 2011 years, while human managed (i.e. manually filled) container habitats were more prevalent in best-fitting simulations during the drier 2012 and 2013 years. The simulations further indicate that rainfall strongly modulates the timing of dengue (e.g., epidemics occurred earlier during rainy years) while temperature modulates the annual number of dengue fever cases. Our results suggest that meteorological factors have a time-variable influence on dengue transmission relative to other important environmental and human factors.
| Numerous studies have investigated meteorological and climatic influences on mosquito transmitted viruses. However, dengue ecology is complex, necessitating an understanding of the interactions among components in the system. We estimate dengue fever cases in San Juan, Puerto Rico using a mathematical model informed by relationships among meteorology, land cover, and interactions among human hosts, mosquitoes, and the dengue viruses identified from the literature. Because some of these relationships are not well known or static, we performed several thousand simulations and compared model output to dengue fever cases reported to the Centers for Diseases Control and Prevention. The model replicated reported dengue cases well, but factors related to dengue transmission patterns varied between years. During wetter years, precipitation-filled containers were the primary immature mosquito habitat in the model. Conversely, during drier years, containers filled with water by humans were the most important habitat. In warmer years there was an increased number of dengue cases that peaked following higher rainfall. These results reveal that current climatic conditions modify the relative influence of human and climatic factors on dengue transmission patterns. This knowledge can be used to develop forecasting tools for dengue outbreaks and enhance mosquito control campaigns based on weather predictions.
| In the last decade dengue infections have increased dramatically in the Americas with cases now occurring in the southern U.S., Mexico and Central America, across the Caribbean, and as far south as Argentina in South America [1]. The Pan American outbreak in 2010 resulted in 1.7 million cases of dengue fever (DF) including 21,206 cases in Puerto Rico [1]. There are four serotypes of the dengue virus (DENV), and subsequent infection with a new serotype increases the risk of severe dengue which can manifest as hemorrhagic fever (DHF). While DHF case-fatality is fairly low, ranging between 0.5% and 5.0% [2], the global burden of DENV infection is extremely high with an annual estimate of 390 million infections of which 96 million result in symptomatic disease [3]. Given the high burden of disease and that transmission is directly and indirectly regulated by meteorological factors, understanding how dengue dynamics may shift under different meteorological conditions is a key public health question [4–6].
Meteorological factors influence many components of DENV ecology, most directly through the Aedes (Ae.) mosquito vector [7]. Temperature and precipitation are important drivers of mosquito population dynamics [8]. Temperature influences development rates, mortality, and reproductive behavior [9–12]. Precipitation often provides water in containers that serve as larval and pupal habitat. Container water volume influences development rates and low water volume can increase mortality through enhanced competition between larvae and pupae at higher population densities [13,14]. Ambient air temperature also influences viral replication within the adult mosquito and is a key regulator of the length of the extrinsic incubation period (EIP), the time between when a mosquito is infected and becomes infectious [15–17]. A shorter EIP reduces the length of the cycle of transmission and increases the probability of its completion during the lifespan of Ae. aegypti.
Despite the established connections between meteorological variables and many components of DENV ecology, the relative influence of weather versus human factors on dengue epidemics is still unclear. For example, although much of the southern United States is inhabited by the vector Ae. aegypti and lies in close proximity to areas where dengue is endemic, locally acquired infections are rare. Reiter at al. [18] contend that interactions between the vector and human are limited in Laredo, Texas as compared to across the border in Nuevo Laredo, Tamaulipas, Mexico due to better infrastructure which prevents vector-human contact and, thus, limits transmission. Other studies have also found human-related factors to be of greater importance than climatic suitability. Keating [19] and Brunkard at al. [20] both found meteorological variables to influence DF cases but suspect that herd immunity, circulating serotype, and strain play a larger role in transmission. Additionally, human responses to climate may be as important as climate itself. In Australia, Beebe at al. [21] and Kearney at al. [22] found that DENV vector habitat may increase due to human water storage designed to combat drought. The complexity and uncertainty of the many factors involved in the disease system make predictions and effective interventions to reduce transmission difficult.
Process-based models can be useful in determining the relative influence of human versus meteorological factors on DENV transmission. Vector dynamics are frequently tied to variations in temperature and precipitation [23]. Dynamic models are built using known biophysical relationships between the vector, virus and the environment. Mathematical relationships are derived from studies on rates of vector development, mortality, and generational progression as well as thermal limits for population survival [11,24]. Vaidya et al. [25], for example, built a model of mosquito population dynamics using temperature dependent maturation and mortality rates and precipitation dependent carrying capacity. Such models have successfully used meteorological inputs to simulate vector-borne disease dynamics [26–28] including models for Aedes mosquitoes that are the vectors of dengue [29,30]. Other uses include exploration of outbreaks based on weather scenarios, and investigations of land use/cover change, and the evaluation of intervention strategies [31]. Because many parameters involved in DENV transmission are unknown, the ability to perform simulations under a variety of conditions and scenarios is an important and powerful aspect of dynamic models that can help characterize parameter uncertainty and whether parameter values vary in time.
Process-based models that simulate vector dynamics are rare and most do not calculate the EIP nor do they simulate transmission between the vector and human populations, in large part due to the complexity of the single and combined models [7]. But vector density does not always correlate with disease incidence [6,32]. A review of dengue transmission modeling approaches by Andraud et al. [33] promotes the use of models that include a vector component for informing public health policy. One of the first models that integrates the two components (human and vector) was developed by Focks et al. [34]. It integrates a dynamic vector population model (the container inhabiting mosquito simulation model, CIMSiM) with a dengue transmission model (the dengue simulation model, DENSiM). While CIMSiM model outputs have been validated in the field, some parameters remain highly uncertain, emphasizing the importance of developing alternative process-based models of dengue transmission [35]. Skeeter Buster, for example, has expanded on CIMSiM to include population genetics, spatial heterogeneity in habitat availability, and stochastic effects [36]. Andraud et al. [33] and another review by Johansson et al. [37] both identify model parameterization as a key challenge for simulating dengue transmission. However, employing numerous models, or many slightly-differing versions of the same model, allows researchers to explore and quantify uncertainty in the parameter space, and model ensembles have been shown to be more accurate than any single model [38].
In this study we present a new modeling framework that couples a dynamic mosquito life-cycle model calibrated for Ae. aegypti with a SEIR (susceptible-exposed-infectious-recovered) model for dengue transmission to simulate dengue outbreaks in San Juan, Puerto Rico from 2010–2013 (available for download at https://sites.google.com/site/dymsimmodel/home). This work expands upon the existing Dynamic Mosquito Simulation Model (DyMSiM) [8] by incorporating a virus transmission component and parameterizing the model for the mosquito species Ae. aegypti. The model incorporates newly collected and consolidated research related to temperature effects on vector survival and development [39] and the dengue virus extrinsic incubation period [17]. Using a Monte Carlo approach, simulations are performed for numerous combinations of parameter values and the results are evaluated using dengue case data from San Juan, PR reported to the Centers for Disease Control and Prevention (CDC) Dengue Branch. The best model simulations over the entire time period and for individual years are analyzed to determine the relative influence of different meteorological and human-mediated factors on DF case numbers and to identify how and why parameter values changed between years.
San Juan is a municipality of Puerto Rico and is located in the northeastern part of the Caribbean island. The population was 395, 326 in 2010 (US Census Bureau, 2010 Census). San Juan has a humid tropical climate with minimal variation in seasonal temperature. Precipitation occurs all year, but it is notably drier during boreal winter and early spring. San Juan is an ideal study location because dengue is endemic in the municipality, and weekly dengue case data are available through CDC ArboNet (USGS, 2014) from 2010–2013. During that time, annual case numbers ranged from 500 (2011) to 919 (2012). These data are crucial for model evaluation.
The mosquito population was simulated using the general structure of the Dynamic Mosquito Simulation Model, (DyMSiM) [8] but the model was parameterized for Ae. aegypti mosquitoes with additional components added, including the epidemiological SEIR model (S2 Table). This enabled the simulation of virus transmission between the human and mosquito populations in DyMSiM. The model is deterministic and was implemented using Euler’s method of integration in Stella 10.06 software. A conceptual diagram of model processes is provided in Fig 1. Parameter value constants and equations are provided in S1 Table, governing equations for human and mosquito populations are provided in S2 Table, and further details are discussed below.
Because it is difficult to obtain precise measurements for many of the DyMSiM parameters, especially those related to immature container habitat, we used a Monte Carlo approach to assess outcomes for a range of possible parameter values. We performed 9600 simulations with DyMSiM using discrete parameter values representative of the distribution of values identified in the literature and the preliminary model simulations (Bold in Column 3 of S1 Table) to replicate DENV transmission in San Juan, Puerto Rico from 2010–2013. During each simulation, one value was changed until all parameter value combinations were used. All humans and mosquitoes were assumed to be susceptible at the beginning of the runs, however, runs were started in 2009 to provide spin-up time to build infection and immunity (human only) within the human and mosquito population. Additionally, a background infection rate was used to initiate infections and is based off of the average number of dengue cases that occur during the low transmission season according to the 2010–2013 case data for San Juan. The assumption that the population is completely susceptible is not accurate, however, given the low incidence (3192 total cases over a population of 395,326) and short time interval over which the model is run, the influence on outcomes should be minor. Weekly CDC reported case data were used to evaluate the model.
Container habitat area estimates were determined by first performing preliminary model simulations under parameter values that would produce a maximum and minimum number of dengue cases in San Juan. By comparing these runs with the reported dengue data, an estimated range of habitat area was determined for the study. Because there is great uncertainty in the container habitat area and it strongly influences mosquito population size, the number and range of values chosen greatly exceeds that of the other parameters. Fewer values and smaller ranges were chosen for parameters where more research has been performed resulting in greater confidence in the selected values.
There were 3,192 cases of DF reported during the study period with 901 cases occurring in 2010, 500 cases in 2011, 919 cases in 2012, and 872 cases in 2013. Reported cases began increasing during week 20 in 2010, 2012, and 2013 but occurred later (week 25) in 2011 which was relatively cool compared to the other years and experienced the fewest number of cases (Fig 2). During most years (2010, 2011, and 2013) the peak in reported DF cases occurred between weeks 32 and 37, however, during 2012 the increase in cases was much more gradual and the peak did not occur until week 49 and did not decline to minimum season case numbers until week 15 of 2013. Temperatures were also much warmer during the second half of 2012 compared to other years, while precipitation was notably lower, especially during summer and fall. Although 2013 experienced similar reported case numbers as 2010 and 2012, most cases occurred during the first 15 weeks of the year (Fig 2) with fewer cases occurring during its summer peak. The decline to reported minimum season case numbers also occurred earlier than the other years. Aside from 2012, increases in reported DF followed rains occurring in late spring. The earliest increase and peak in reported DF cases occurred in 2010 which also experienced the warmest spring temperatures and earliest precipitation.
Statistical measures of accuracy between the reported data and ensemble averages using the top 1% (n = 96) best-fit simulations, for both the entire time period (S2 Dataset) and individual years (S3–S6 Datasets), are reported in Table 1. Although inter-annual variability in DF cases was well-simulated in the ensemble average from the 96 model simulations over the entire time period (r2 = 0.44, p = 0.002), the annual epidemic curves for individual years often lacked precision (Fig 2). When evaluating ensemble averages from the simulations over individual years (same simulations but the statistics are only performed on reported case data for individual years), however, intra-annual variability was replicated with much higher accuracy for 2010 (r2 = 0.90, p = 0.03), 2011 (r2 = 0.83, p = 0.05), and 2012 (r2 = 0.94, p = 0.01) (Fig 3). Simulations for 2013 were markedly less accurate (r2 = 0.25, p = 0.25) compared to the other years (Fig 3). By contrast, the yearly model accuracy assuming no a-priori knowledge to guide parameter selection was lower and more variable: the r2 values for 2010, 2011, 2012, and 2013 were 0.77, 0.64, 0.05, and 0.12 respectively.
Fig 4 displays the number of times each parameter value was used in the top 1% of simulations. Notably, in 2010 and 2011 the number of simulations retained that used a high proportion of open containers (i.e., more precipitation dependent sources) vs. a low proportion of open containers (i.e., more human-managed sources) is much larger than in 2012 and 2013 (76 and 59 vs. 0 and 0). Additionally, patterns in container habitat area, background infection rate, and carrying capacity vary considerably between years. The retained simulations for most years used the higher adult daily survival rate, and the shorter length of infectious period (except 2012). With the exception of 2011, the range of values tested for container height and host infection probability are used with equal frequency between years indicating that they did not have a strong impact on temporal variation in dengue transmission dynamics.
Fig 5 displays the average rank of simulations using each parameter value among all 9600 simulations. Simulations that used a high proportion of open containers ranked slightly better for 2011, while the opposite is true for 2012 and all years. This is consistent with the trend in the top 1% of simulations, however, there is no strong difference for 2010 and 2013. Simulations using higher values for container habitat area were markedly worse, though this effect was lesser for 2012. Additionally, the lower carrying capacity value tended to produce better simulations. Average ranks among different values did not vary markedly for the other parameters.
Precipitation was highest in 2010 (227.5 cm) and 2011 (224.0 cm), with the onset of the rainy season starting a few weeks earlier in 2010 (Fig 6). Precipitation was by far the lowest in 2012 (140.3 cm), however, temperatures were warmest in 2012, especially during late fall (~1°C warmer in November). Precipitation was 216.3 cm in 2013 –only slightly drier than 2010 and 2011—but the timing of the precipitation was different, exhibiting a marked drop off during late summer/early fall.
Our results emphasize the complexity of the influence of weather on DENV transmission. While the best-fit “all years” simulations reasonably replicate the inter-annual variability of dengue cases from 2010–2013, resolving the intra-annual variability using the same parameter values over the whole time period is difficult. As the results for individual years suggest, it is likely that parameter values change over time, as do the primary influences that drive dengue transmission. For example, preferred habitat may change seasonally or with climate conditions [58]. Averaging over 96 simulations dampens some of the variability within individual years. The best-fit “all years” simulations also captured the inter-annual variability in the length of the dengue season, though the timing and peaks of transmission were sometimes asynchronous. The ability of DyMSiM to resolve the variability of the annual case load and season length may be useful for a variety of applications, such as studies focusing on potential climate change impacts on dengue incidence and seasonality (e.g., addressing the question of whether dengue incidence may become higher or lower, or whether the season will be extended under climate change scenarios).
The best-fit models for individual years accurately resolved the intra-annual variability for 2010, 2011, and 2012 (r2 = 0.90, 0.83 and 0.94 respectively). Parameters often changed in response to meteorological variability. Optimizing the DyMSiM parameters for single years may be useful for examining the causality of seasonal trends in DF case numbers. This type of simulation could also be used for short term predictions by selecting parameter values based on currently available case data and then running simulations for forthcoming weeks using weather forecast data. Alternatively, building a databank of epidemic profiles based on model results could provide a collection of possible scenarios that could occur given present conditions.
The timing of the onset of the dengue season and peak for year 2013 was not well simulated even with parameters optimized specifically for the year, illustrating the sensitivity and complexity of the disease system. The model predicted a rise and peak in dengue cases more than 50% higher and five months later. Many or all of the components of the virus’ ecology are constantly changing and their responses to external factors (such as weather) are situation dependent. Meteorological conditions may not have had a strong influence on intra-annual variability in 2013. A number of other factors not included in the model may have dominated transmission that year such as changing patterns of herd immunity to the specific circulating dengue serotype(s), introduction of a new variant of a serotype earlier in the season, implementation of intervention methods such as novel source reduction of habitats, or other human related factors such as extensive use or reduction of water storage. While it is beyond the scope of this paper to determine which of these factors may have been influencing the transmission, a variant of one of the four serotypes could have been introduced early in the season but all four serotypes had been circulating previously in Puerto Rico [59]. Shifting herd immunity could play a role in reducing the overall level of reported cases but should not greatly influence intra-annual variability of reported cases. The greatest increase in immunity likely occurred during 2010 and 2012 when Puerto Rico experienced high levels of transmission (Fig 2). Additionally, if higher levels of herd immunity played a role, it would be expected that there would be a delay in the onset of cases while we observed that reported case data peaked much earlier than the modeled cases. It is also possible given the high level of transmission in 2012, that this early peak in 2013 was propagated transmission from the previous outbreak; with initial transmission into the general population which spread to a smaller adjacent geography. Propagated transmission of dengue has been observed in other areas of the world [60]. Further exploration into the potential factors influencing the patterns observed in 2013 are warranted. Changes in intervention strategies or patterns of container habitats could change DF dynamics by reducing or enabling transmission despite climatic conditions. For example, the model over-predicts DF cases late in the year. This could be due to fewer container sources, reductions in susceptible hosts, or increased use of pesticides.
In the present study we were able to rank the best simulations based on their fit to previously collected case data considered to be reliable. However, accurate case data may not be available in many locations, preventing the selection of parameters based on best-fit simulation. Therefore, we also tested the fit of the ensemble median of all 9600 simulations to the raw case data, in order to determine whether reasonable accuracy could be attained by assuming a wide range of parameters without any a priori knowledge about a given location. The r2 values for 2010, 2011, 2012, and 2013 were 0.77, 0.64, 0.05, and 0.12 respectively. The results indicate that the model is robust during certain years (2010 and 2011), but specific parameter values may be essential for anomalous conditions such as the extremely dry period during 2012 and early 2013. Interestingly, parameter values that produced the top 1% of simulations are not always consistent with those that produced the best simulations on average across all 9600 runs. For instance, on average the worst simulations occur with high values of container habitat area, however, many of the best simulations for 2010 and 2011 use a high value. A similar phenomenon occurs for the adult daily survival rate. Further analysis in future studies may reveal parameterizations that are robust over a wider array of conditions and locations. Currently, caution should be taken when interpreting model results that do not include either parameter selection by comparison with reported case data or data about location specific parameters such as container magnitude and type. It is recommended that parameter values producing the best simulations on average be used if there is no a priori training data.
Variations in some DyMSiM parameter values had a particularly important influence on model performance. Parameter response to meteorological variability is best exemplified in 2012, during which simulations with a higher amount of human managed water sources were best able to resolve dengue transmission. This is almost certainly a consequence of the drier conditions which made rain-dependent water sources ineffective as immature habitats. A study conducted by Barrera et al. [61] monitored adult and immature mosquito populations and habitats in San Juan and discovered that vector populations were sustained even in drier conditions through human managed water sources. Although precipitation had an obvious influence on population dynamics, human managed water sources also proved to be important habitats for the Ae. aegypti vector. We hypothesize that as has been seen in other areas [21,22] human managed water sources likely became more important habitats during the drier conditions of 2012 either because there was less precipitation to fill open containers or because of a corresponding increase in man-filled containers for water storage in response to drier conditions. Additionally, particular values were shown to be of less importance for the top 1% of simulations, including the host infection probability and container height.
While precipitation patterns likely contribute to oviposition habitat selection by Ae. aegypti adults and consequently affect intra-annual transmission dynamics, variations in temperature may be largely responsible for fluctuations in annual DF case-loads, especially given the sensitivity of the length of the EIP to temperature. While average monthly temperatures generally vary between 25°C and 30°C, this translates to an almost four day difference in the length of the EIP (14.6–10.8 days) which can accelerate transmission. For example, the comparatively warm temperatures during El Nino in the winter and spring of 2010 (Fig 6) could have facilitated early and more rapid transmission of the virus (shown in the reported and simulated case data) by accelerating the mosquito life cycle and shortening the EIP. Conversely, the cooler temperatures persisting during La Nina throughout 2011 may have limited transmission, despite the wet conditions, resulting in fewer DF cases (Fig 7). In 2012, temperatures were relatively cool during the first five months of the year but were among the warmest across all years for the final seven months, especially during late fall (Fig 7). This explains the late seasonal peak in reported and simulated DF cases and the overall high annual incidence. In another Puerto Rico study, Jury [62] also observed that yearly fluctuations in DF caseloads were related to temperature while the timing was largely regulated by precipitation patterns. These results illustrate how nonlinear relationships between temperature, the mosquito lifecycle, and the EIP can generate considerable shifts in DENV transmission dynamics from small temperature differences. Interestingly, despite moderate temperatures and precipitation, 2013 experienced low levels of DENV transmission in San Juan (Fig 7). Shifting spatial patterns of herd immunity may explain the relatively low transmission in San Juan during 2013, especially since other parts of Puerto Rico experienced high case numbers and meteorological factors did not appear to strongly influence transmission.
Although there are similarities to other models, DyMSiM differs in important ways. CIMSiM, for example, produces mosquito population estimates over a 1 ha area which interact with human populations in DENSiM to simulate virus transmission between the two populations [34]. These models, though very comprehensive, require a great deal of site-specific information and training with many preset parameters. Parameter values in DyMSiM, however, can be evaluated and selected using ensemble simulations and locally reported case data with relative ease. DyMSiM has also been parameterized using newly acquired and synthesized data on vector development and virus EIP (S1 Table). These and other models may complement each other; by comparing strengths and weaknesses for specific applications and research goals. And, when used in conjunction, simulation results can be synthesized to determine the variance and confidence levels in model predictions and increase the accuracy of results [38].
While we used the best available information from the literature to specify values of the DyMSiM parameters, some values are based on the results of only a few studies or were reasonable estimates when quantitative results were unavailable. Papers that consolidated and considered data from multiple studies proved especially useful because they accounted for variations arising from conducting research on different colonies of mosquitoes and strains of virus or using different methods to answer the same question [17,39]. Often averages were used to enhance simplicity and interpretability of the model; yet, many parameters may have values across an unknown distribution surrounding their mean due to genetic and environmental variability. For example, genetic variations have been found to alter temperature regulation of West Nile virus transmission [63]. If a similar variation in temperature sensitivity exists between or within dengue serotypes it could be an important component to the severity of DF epidemics. We attempted to mitigate our limited knowledge of some parameter values by using a Monte Carlo approach when they were particularly difficult to ascertain. The model showed that some parameters, such as container habitat area and composition of containers, were particularly important for simulating transmission. The carrying capacity density was also important as it influenced the mosquito productivity within containers, and was directly related to their area. For example, simulations were degraded when using larger container areas, but this effect was lessened when using the lower maximum larval density value. The value of the background infection rate was also important when selecting the best fit models. This may indicate the importance of varying virus introduction or virulence between years. Vertical transmission of the virus was not considered in this model though it has been observed in the laboratory and field setting and could help maintain the virus during unfavorable conditions [64,65]. However, Adams et al. [66] found it did not have considerable importance in their model and the frequency of its occurrence in nature is still uncertain [67]. As more evidence is generated establishing links among meteorological factors, vectors, virus, and humans, parameter values can be better defined.
Some of the inconsistencies between modeled and reported case data may also be attributed to the nature of surveillance data. Because the surveillance system is passive, the reported case data are only an approximation of total DF incidence for the municipality. Incomplete or inaccurate reporting can arise due to sub-clinical infections, misdiagnosis, failure to report cases, differences between where virus transmission is reported and where it was acquired, and variability in the time between transmission of the virus, development of symptoms, and clinical visits. Relative increases in dengue reporting may be seen if there is increased media attention, circulation of a more severe strain, or increases in laboratory testing ability, etc. The serotype of the virus is often not reported but could be of particular importance if a high proportion of the population is immune to the circulating serotype. It is likely that not all cases are being reported to the CDC and, therefore, the reported case data is only a subset of the actual number of cases occurring in San Juan though it is difficult to estimate the amount of underreporting. The influence of under-reporting should be minimal if the patterns of reported cases represent the yearly variability even if the overall magnitude of reported cases is inaccurate. The largest bias in the calculation of the correlation between modeled output and reported case numbers would occur if there were significant changes to the reporting system or if there are interannual differences in reporting. There were no significant changes to the surveillance system protocol during this time period.
It is difficult to determine inter-annual variability in reporting, but there is the possibility for bias. For example, if individuals at the beginning of the season were more likely to seek care and be reported through the surveillance system, or if physicians were less likely to diagnose and report cases during typically lower periods of dengue, the shape of the curve might shift. Given the uncertainty of these types of scenarios, trying to determine the relative impact they may have on the correlation between model output and reported case numbers would be highly speculative. We acknowledge, however, that under-reporting of dengue is an important issue and therefore, the model output should not be interpreted as an estimate of actual dengue case numbers but as a representation of inter-annual variability of DENV transmission. If more reliable estimates of reporting bias and information on the temporal variability in asymptomatic cases or the reporting of symptomatic cases become available, it would be useful to compare numbers of modeled and reported dengue case data. Despite these limitations, the reported case data are currently the best metric available for tracking DENV transmission and have been used in other studies to represent levels of DENV transmission in Puerto Rico [4,61,68].
The results of this study provide important information for dengue control. First, epidemics can occur during both wet and dry years. Dengue cases were comparatively high in 2010, one of the wettest years on record for San Juan, and in 2012, a drier-than-normal year in which the rains came later than average. The warmer winter and spring temperatures are likely responsible for the greater number of DF cases during 2010 while warmer fall and winter conditions in 2012 likely helped propagate dengue transmission later in the year. It is common for dry years to be hotter due to a combination of factors that include: 1) fewer clouds facilitating higher daytime temperatures and 2) less evapotranspiration freeing a greater amount of solar radiation for heating. Second, the timing of the dengue cases varies substantially from year-to-year and appears to be linked with the timing and magnitude of rainfall. In the wetter 2010 year, DF case numbers rose sharply after the onset of the rains (Week 10) and peaked near Week 30, whereas in the drier 2012 year the cases began later (Week 20) after a highly sporadic beginning to the rainy season, and gradually climbed to a much later peak near Week 48. In both instances, the onset of DF cases coincides with or begins shortly after the onset of the rainy season, and rises steadily due to the time required for the mosquito population to development in newly formed habitats and completion of the EIP. Lastly, despite a tropical climate, small fluctuations in temperature can have considerable effects on DENV transmission. Often tropical environments are thought to be warm enough for year-round transmission with precipitation being the limiting factor. Our results indicate that nonlinear relationships result in greater temperature impacts on DENV transmission than would be expected. Monitoring climate as a proxy of dengue risk may provide public health workers with a simple tool to prepare and execute transmission intervention methods and arrange testing and treatment protocols before the DF season peaks. Early warning systems can be developed based on the identification of rainfall and temperature patterns that promote rapid dengue transmission [69–72].
Dominant Ae. aegypti container habitat appeared to vary with precipitation patterns. During the wetter years, 2010 and 2011, simulations using a smaller container area and more rain dependent containers were ranked better; however, during 2012 simulations using a greater amount of human managed water sources ranked better. This suggests that Ae. aegypti inhabit rain filled containers when available, in addition to human managed water sources, but drier conditions may limit immature mosquito habitats to permanent (septic tanks, large water tanks, etc.) or human managed (animal drinking pans, plant trivets) water sources. This pattern has been observed in field studies [58]. From a public health perspective, the magnitude of precipitation in a given year may inform how to refine the message conveyed to the public about how to limit mosquito habitat. During wet years, mosquito control strategies should focus on eliminating old tires, buckets, and other items that can be filled via precipitation. Drier years should shift the focus to treating permanent or human managed water sources and covering stored water to prevent mosquito infestation. Further investigation of how weather and rainfall patterns influence human behaviors such as water storage are needed to determine if this could be incorporated into the model.
Understanding weather influences on DENV transmission is important in the context of climate change especially as DENV range expands in the Americas. Outbreaks in Hawaii, along the Texas/Mexico border, and in southern Florida indicate a reemergence of dengue in the U.S. [73–75]. Additionally, sero-prevalence studies indicate that dengue cases are likely underreported in the U.S. [73,75] due to subclinical infections, insufficient resources for testing, lack of sensitization of the medical community and patients seeking medical attention across the border in Mexico [75]. Although infrastructure differences, such as the prevalence of air conditioning, were shown to reduce transmission on the US side of the Mexico border in Laredo, autochthonous cases still occur [18]. Viral introduction in response to frequent travel across the border by residents may be a large source of infections [76] and weather may still have an important influence on transmission risk [20]. Whether DENV is endemic to an area or currently only poses a threat, understanding the influence of weather and climate on DENV ecology can facilitate strategies to prevent or mitigate transmission.
Using known relationships between climate variables, Ae. aegypti dynamics, DENV replication and transmission, and a basic SEIR model within a modeling framework known as DyMSiM we were able to simulate inter-annual variability in DF cases in San Juan County, PR for the years 2010–2013. When optimizing the DyMSiM parameters for a single year, we were able to simulate intra-annual variability well for three of the four years. Meteorological factors were important influences on dengue transmission for 2010, 2011, and 2012 but not as important in 2013. Parameter values changed in response to meteorological conditions, illustrating the complexity of DENV ecology.
The results of this study quantify meteorological impacts on DENV transmission and demonstrate the variance in parameter values that result from changing environmental conditions. Rain patterns were especially important for determining the timing of epidemics and the primary habitat for immature Ae. aegypti vectors. Abnormally high temperatures have the potential to extend the transmission season (even during dry years) as exemplified in 2012. In general, temperature had the greatest influence on annual case numbers. The sensitivity of DENV ecology to meteorological variables and their interactions underscores the utility of process-based modeling when studying the impacts of climate variability and climate change on vector-borne disease. Information from this study can be used by public health officials to improve DENV transmission intervention strategies in San Juan, PR by way of advanced preparedness through model predictions and creation of more targeted vector control campaigns. As our understanding of the ecology of the virus and its thresholds improves, so will our ability to implement effective mitigation strategies against DENV transmission and reduce the disease burden on human populations.
Future research endeavors will seek to enhance model performance and utility. Validation of DyMSiM in other areas with longer term datasets would be useful in determining its ability to simulate DENV transmission in other environments and its ability to simulate DF incidence over long time periods. The deficit of long, reliable dengue surveillance records has forced training of dengue models over periods of less than 10 years and validation on one or only a few years [70,77,78]. Inclusion of serotypes would also enhance the model given that immunity to specific virus serotypes can affect transmission dynamics by limiting the susceptible population. The current SEIR model is simple but can still be used to examine the dynamics of dengue over short time periods. Incorporating surveillance of multiple virus serotypes, including short term cross-immunity, will be a priority in future research given the importance of immunological status on virus transmission dynamics. Additionally, the inclusion of social factors that influence population risk, such as migration, age structure, personal and household level prevention behaviors and other factors influencing vector-human contact, may provide further insights into the nature of DF epidemics. Finally, future work will concentrate on the possible impacts of climate change on dengue transmission. For instance, under climate change, winter temperatures may become suitable for completion of the EIP allowing DENV transmission to continue. Understanding the effects of future weather and climate conditions on DENV transmission is vital for limiting increases in incidence of disease and range expansion.
|
10.1371/journal.pgen.1006059 | The Great Migration and African-American Genomic Diversity | We present a comprehensive assessment of genomic diversity in the African-American population by studying three genotyped cohorts comprising 3,726 African-Americans from across the United States that provide a representative description of the population across all US states and socioeconomic status. An estimated 82.1% of ancestors to African-Americans lived in Africa prior to the advent of transatlantic travel, 16.7% in Europe, and 1.2% in the Americas, with increased African ancestry in the southern United States compared to the North and West. Combining demographic models of ancestry and those of relatedness suggests that admixture occurred predominantly in the South prior to the Civil War and that ancestry-biased migration is responsible for regional differences in ancestry. We find that recent migrations also caused a strong increase in genetic relatedness among geographically distant African-Americans. Long-range relatedness among African-Americans and between African-Americans and European-Americans thus track north- and west-bound migration routes followed during the Great Migration of the twentieth century. By contrast, short-range relatedness patterns suggest comparable mobility of ∼15–16km per generation for African-Americans and European-Americans, as estimated using a novel analytical model of isolation-by-distance.
| Genetic studies of African-Americans identify functional variants, elucidate historical and genealogical mysteries, and reveal basic biology. However, African-Americans have been under-represented in genetic studies, and relatively little is known about nation-wide patterns of genomic diversity in the population. Here, we study African-American genomic diversity using genotype data from nationally and regionally representative cohorts. Access to these unique cohorts allows us to clarify the role of population structure, admixture, and recent massive migrations in shaping African-American genomic diversity and sheds new light on the genetic history of this population.
| The history of African-American populations is marked by dramatic migrations within Africa, through the transatlantic slave trade, and within the United States (US). By 1808, when the transatlantic slave trade was made illegal in the US, approximately 360,000 Africans had been brought forcibly into the US in documented voyages [1]. International and domestic slave trade continued to impose long-distance migration on enslaved African-Americans until the end of the Civil War, in 1865. By 1870, the US census reported 4.88 million “colored” individuals of which 90% lived in the South [2].
Despite the ban on slavery, economic and social perspectives for most African-Americans remained bleak. Better opportunities in the North (Northeast and Midwest) and West led millions of African-Americans to leave the South between 1910 and 1970 [3]. This demographic event known as the Great Migration profoundly reshaped African-American communities across the US [4]. Today, 45 million Americans identify as Black or African-American.
A history of slavery and of systemic discrimination led to increased social, economic, and health burdens in many African-American communities. Health disparities continue to be compounded by poverty, unequal access to care, and unequal representation in medical research. To reduce health disparity in research, many cohorts are currently being assembled to encompass more of the diversity within the US [5, 6]. These cohorts create opportunities in both medical and population genetics; they also require an understanding of genetic diversity within diverse cohorts. However, the large-scale migrations and incomplete genealogical records for African-Americans present a challenge for such an understanding. Previous studies have described the proportions of African, European, and Native American ancestries across individuals [7–13], the amount of diversity in sequence data [9, 14, 15] and inferred admixture models [12, 16, 17].
However, because previous cohorts were not representative of the general African-American populations, they provided limited information about population structure among African-Americans.
Here, we use cohorts including 3,726 African-Americans and a total of 13,199 individuals geographically distributed across the contiguous US to investigate nation-wide population structure among African-Americans. We first confirm and refine previous estimates of admixture proportions and timing in the population, and find significant differences in ancestry proportions between US regions. We then investigate relatedness among African-Americans and European-Americans through identity-by-descent analysis, and identify long- and short-range patterns of isolation-by-distance. We introduce quantitative models, incorporating both census data and fine-scale migration, to describe these isolation-by-distance patterns and infer migratory patterns in the population. Integrating quantitative models for admixture, relatedness information, and historical data, we identify ancestry-biased migrations during the Great Migration as a driving force for ancestry and relatedness variation among African-Americans. The analysis of geographically distributed cohorts through detailed mathematical modeling therefore helps us understand the distribution of genetic diversity in large cohorts and provides new insights into recent human demography.
We analyzed data from three cohorts: (a) Health and Retirement Study [18] (HRS), with 1,501 African-Americans and 9,308 European-Americans sampled representatively across all US states, and including urban and rural regions; (b) Southern Community Cohort Study [19] (SCCS), including 2,128 African-Americans sampled within the southern US in rural locations; (c) 1000 Genomes Project cohort of 97 individuals of African ancestry from the southwest USA [20] (ASW). Genotypes were obtained on Illumina Human Omni 2.5M and Human 1M-Duo platforms, and joint analyses were performed on a common set of 553,795 high-quality SNPs (for detailed information, see Materials and Methods and S1 and S2 Tables).
Individual genomes carry genetic material from multiple ancestral lineages, and each diploid locus derives ancestry from two distinct lineages. We used RFMix [11] together with 1000 Genomes Project panels from Africa, Europe, and Asia to identify the most likely continental ancestry at each locus for individuals in the cohorts (Fig 1D, S2 Fig and Materials and Methods). Here, continental ancestry is defined as the inferred location of the ancestral lineage prior to the advent of transatlantic travel. The overall proportion of African ancestry is substantially higher in the SCCS and HRS than in the ASW and the recently published 23andMe cohort [12] (Table 1).
The HRS cohort can be thought of as representative of the entire African-American population, while the SCCS focuses primarily on individuals attending community health centers in rural, underserved locations in the South. By contrast, the sampling for the ASW and 23andMe did not aim for specific representativeness, and the ascertainment in the 23andMe cohort might have enriched for individuals with elevated European ancestry (see Materials and Methods and discussion in [12]). In the HRS, average African ancestry proportion is 83% in the South and lower in the North (80%, bootstrap p = 6 × 10−6) and West (79%, p = 10−4) (Fig 1). Within the SCCS, African ancestry proportion is highest in Florida (89%) and South Carolina (88%) and lowest in Louisiana (75%) with all three significantly different from the mean (Florida p = 0.006, South Carolina p = 4 × 10−4, and Louisiana p < 10−5; bootstrap). The elevated African ancestry proportion in Florida and South Carolina is also observed in the HRS and in the 23andMe study [12], but Louisiana is more variable across cohorts (Fig 1E). As expected, European ancestry proportions largely complement those of African ancestry across the US.
Because recombination breaks down ancestral haplotypes over time (Fig 1D), the length of continuous ancestry tracts is informative of the time of admixture, with shorter tracts reflecting older admixture. We inferred the timing of admixture using Tracts [16], which fits a demographic history to the observed distribution of tract lengths (see Materials and Methods for details and S4 Table for confidence intervals). Because of the small number of Native American tracts, even a small amount of spurious Native American ancestry assignments can bias the inference. Thus, we first considered a model with two source populations: African and non-African. Assuming a single admixture event, we estimated the time of admixture onset g, where g = 1 means that the parents of the individual are the founders of the admixed population and that the current individual represents the first admixed generation. For HRS, we inferred a timing of g = 5.8 generations ago (S8 Fig). The estimated year of birth of the first admixed children is T = Ts − (g − 1)τ, where Ts = 1939.8 is the average year of birth of HRS individuals and τ is the generation time. Individuals born τ years earlier should be 1 generation closer to the onset of admixture. Correlating birth year and inferred admixture time within our cohort (Fig 2D), we inferred τ = 27.4 (r2 = 0.88, p = 10−7), which leads to an admixture year of 1808 (bootstrap 95% CI: [1805.5, 1810.4]). Note that 1808 represents the admixture time that best explains the data under the assumption of a single admixture event. The narrow confidence interval is, therefore, no guarantee that something exceptional occurred between 1805 and 1810. To investigate the role of modeling assumptions in admixture time estimate, we considered more general models.
A model allowing for two phases of European admixture outperforms the single-pulse model for HRS and SCCS (see Materials and Methods). In HRS, it suggests a first admixture event in 1740 (8.3 generations ago; bootstrap 95% CI: [1711.6, 1744.2]) and a second pulse, of approximately equal size, in 1863 (3.8 generations ago; bootstrap 95% CI: [1852.9, 1865.9]) (S8 Fig and Materials and Methods). Mean birth year in SCCS is Ts = 1946.9, supporting a single admixture event in 1802 (6.3 generations ago; bootstrap 95% CI: [1799.2, 1803.6]), or two events in 1714 and 1854 (9.5 and 4.4 generations ago; bootstrap 95% CIs: [1704.6, 1739.7] and [1849.8, 1868.7]) (Fig 2A, S8 Fig and S4 Table for confidence intervals). The two-pulse model remains a coarse simplification of the historical admixture process, but the data strongly supports ongoing admixture, predominantly before or around the end of the Civil War. This is consistent with historical accounts of “a marked decline in both interracial sexual coercion and interracial intimacy” [21] at the end of the Civil War (see also Ref. [22] and references therein).
The limited role of early 20th century admixture is further supported by the similarity in the inferred single-pulse time to admixture in all HRS census regions (between 5.4 and 6.2 generations ago, S11 Fig) and all cohorts, which is easily explained if most admixture occurred in the South prior to the Great Migration. The similar levels of African ancestry for all age groups within the HRS also support limited European admixture between 1930 and 1960 (Fig 2D). Importantly, more recent admixture is not represented in the SCCS and HRS cohorts; only two participants were born after 1970.
Time estimates point to admixture occurring when most ancestors to present-day African-Americans lived in the South. The regional differences in ancestry seen in Fig 1 are therefore unlikely to be caused by differences in recent admixture rates, and the large influx of migrants from the South would have strongly attenuated any earlier differences. An alternate explanation for regional differences in ancestry proportions is that individuals with higher European ancestry were more likely to migrate to the North and West during the Great Migration, a scenario we refer to as ancestry-biased migration.
To validate the ancestry-biased migration model, we compared ancestry proportions of HRS individuals according to their region of birth, residence, and migration status. European ancestry proportions in African-Americans who left the South (16.5%) is elevated compared to individuals who remained in the South (15.3%, bootstrap p = 0.04), confirming that ancestry-biased migrations continued at least to the mid-20th century. These migrants had substantially less European ancestry than African-Americans already established in the North (20.9%) and West (25.0%) (Fig 2E). Since the latter two groups received large contributions from the first wave of the Great Migration, this suggests that the proportion of European ancestry in first-wave migrants was higher than in the second wave—i.e., that there was stronger ancestry bias during the first wave of migration.
This change over time in ancestry-biased migration is consistent with historical accounts that southern African-American migrants to northern cities during the later stages of the Great Migration had darker complexion than North-born African-Americans (see [23], p. 179). The change could be explained by better social opportunities available to individuals with higher levels of European ancestry: Individuals with wealth and education were much more likely to migrate in the first wave of the migration (see [23], p. 167). Fig 2E shows that despite the ongoing ancestry bias, the migrations of HRS participants led to more uniform ancestry proportions across regions. Interestingly, the proportion of African ancestry among African-Americans increased in all four US regions between the time of birth and the time of survey of participants: The ancestry bias caused migrants to have levels of admixture between those of the South-born and North-born individuals. Their departures and arrivals both increased the regional African ancestry proportions.
Out of 1,491 non-Hispanic African-Americans in HRS, 11 individuals have more than 5% Native American ancestry. Within SCCS, this proportion is only 8 out of 2,128 individuals. The ASW cohort, with 8 out of 97 individuals above this threshold, is a clear outlier. The other 89 individuals, however, have similar amounts of Native American ancestry to the other studies. If we filter out individuals contributing more than 5% Native American ancestry from each cohort, the proportion of Native American ancestry in the remaining individuals is close to 1.1% in the SCCS, in all HRS census regions, and in the ASW. The filtered SCCS Louisianans have significantly more Native American ancestry (1.6%, bootstrap p = 2 × 10−5), and South Carolinians have less (0.09%, p = 2 × 10−5), than the mean Native American ancestry. We did not find a global correlation between European and Native American ancestry, except within Louisiana (S4 Fig).
A three-population admixture model accounting for Native American admixture confirmed the predominantly early, multiple-phase European admixture and suggested that Native American admixture occurred even earlier, consistent with previous findings [12]. Inferred dates of admixture onset are 1494 (bootstrap 95% CI: [1478.8,1516.0]) for the HRS (S9 and S10 Figs) and 1486 (bootstrap 95% CI: [1475.4, 1499.4]) for the SCCS (Fig 2B and 2C), as described in Materials and Methods. The presence of a small amount of spurious, short segments of inferred Native American ancestry could bias the inference toward these unrealistically early dates. The lack of longer Native American segments nevertheless suggests that most Native American ancestry in African-Americans results from contact in the early days of slavery (see, e.g., [24]). The three-population model suggests more recent European admixture dates than the two-population model, but with a higher proportion of migrants in the earlier migration. Finally, a three-population model with continuous European admixture provided qualitatively similar estimates to the two-pulse model, with an early onset of Native American admixture (1482) and European migration spanning the period between 1758 to 1887. Direct admixture between African-Americans and Native Americans is further supported by the observation that the proportion of Native American ancestry in HRS African-Americans (1.2%) is comparable to that in HRS European-Americans (1.5%). This proportion is therefore much higher than would be expected if the Native American contribution occurred through European admixture. Despite substantial disagreement as to the specific dates, all models agree on European admixture occurring predominantly prior to the Civil War.
Along the X chromosome in the HRS, we estimate 84.82% African ancestry, 12.89% European ancestry, and 2.29% Native American ancestry (bootstrap 95% CI [2.14%, 2.45%]). The higher proportion of African ancestry along the X compared to autosomes is consistent with previous studies [12, 17] and the historical record of early admixture occurring predominantly through coerced sexual interaction between European-American males and African-American females [21]. A model with a single pulse of admixture (as considered in [12]) applied to the present data suggests 28.6% Europeans among male contributors, but only 5.2% among female contributors. By contrast, it suggests almost no contribution from Native American males, and 3% from Native American females.
The US Census includes a separate category for Hispanic/non-Hispanic ethnicity. In HRS, 32 African-Americans have self-identified as Hispanics (of which only 10 are within the contiguous US). Hispanics often trace ancestry to regions colonized by Spain and Portugal, and where Native American populations contributed a higher proportion of the present-day gene pool compared than in the US. Genetic ancestry within this group is indeed distinct from the bulk of the non-Hispanic African-American population in at least two ways: elevated Native American ancestry and a higher genetic similarity to southern European populations (S5 and S6 Figs). The correlation between southern European and Native American ancestries also holds in individuals who do not self-identify as Hispanic, particularly in Louisiana (see Materials and Methods). Individuals with elevated Native American and southern European ancestry would not be identified by self-reported ethnicity or by genetic estimates of African/non-African ancestry, yet they may have distinct response patterns to medical tests [25, 26].
The classical isolation-by-distance model predicts that genetic relatedness between individuals decreases as their geographic distance increases [27]. However, large-scale migrations can dramatically alter this picture [28]. To investigate the effect of recent migrations on patterns of genetic relatedness within African-Americans, we consider genetic segments that are identical-by-descent (IBD) between pairs of individuals. We focus on long IBD segments (l ≥ 18cM), which correspond to an expected common ancestor living within the last 8 generations (see Materials and Methods) and are therefore informative of recent demography.
Fig 3A, 3B, S12 and S15 Figs show the mean pairwise relatedness among seven geographic regions in the US for African-Americans and European-Americans. Here, the relatedness of two individuals is defined as the total length of the genome shared through long IBD segments. These recent relatedness patterns differ markedly between African-Americans and European-Americans (compare Fig 3A and 3B): African-Americans exhibit a distinct enrichment in South-to-North relatedness along the main historical migration routes.
To compare these relatedness patterns with recent migration data, we used the 20th century US census data and a simple coalescent model to estimate the expected relatedness between geographic regions (see Materials and Methods). Census-based predictions (Fig 3D) are correlated with IBD-based observations (Fig 3A) if we consider non-identical pairs of regions (Mantel test p = 0.019). Limiting the comparison to the South-to-North and South-to-West relatedness, to capture migration routes specific to the Great Migration, yields p = 0.063 (using the 2010 region of residence) and p = 0.015 (using place of birth) (see Materials and Methods).
Fig 3C and S16 Fig show the relatedness between African-Americans and European-Americans. African-Americans across the US are more related to European-Americans from the South than to those from the North or West (bootstrap p < 0.0002). In addition, European-Americans from the South tend to be more related to African-Americans in the North than to those in the South (bootstrap p = 0.11). This increased relatedness with increased distance is unusual in population genetics, but is easily explained: The ancestry-biased migration is also a relatedness-biased migration. The reduced relatedness between northern European-Americans and African-Americans may also be reinforced by recent European migration, because the new migrants were more likely to settle in the North but were less likely to be related to African-Americans.
Despite the unusual long-range relatedness patterns, identity-by-descent decays with distance within African-American communities in the South, reflecting isolation-by-distance (S19 Fig). To understand how migrations affect isolation-by-distance and identity-by-descent, we introduce a quantitative model taking into account a diploid population density n and spatial diffusion constant D. In short, the displacement between parental birthplace and offspring birthplace of individuals is modeled as an isotropic random walk; the distribution of the times t to the most recent common ancestor of two individuals separated by distance R is calculated under a coalescent model; and the amount of genetic material shared IBD given a common ancestor at time t is computed as in Ref. [29]. Under this model, we can calculate the expected fraction of genome shared IBD between two randomly chosen individuals separated by a distance R. If we consider only IBD segments of length in ℓ = [lmin, lmax] (in Morgans), we find
E ℓ [ f | R ] = 1 16 π n D 2 K 0 R r min - K 0 R r max + R r min K 1 R r min - R r max K 1 R r max (1)
where r min , max = D / l min , max, and Kα(x) is the modified Bessel function of the second kind [30] (see Materials and Methods).
Fig 3E shows the presence of a background level of IBD relatedness in both African-Americans and European-Americans even at long distances. This could be attributed to false positives in IBD calling, to relatedness originating prior to ancestral migrations from Europe and Africa into the Americas, or to a small amount of distance-independent migration. We account for these effects in our model by introducing an additional distance-independent (constant) term. Using IBD segments longer than 18cM, we estimate the background IBD for African-Americans and European-Americans in HRS to be bAFR = 0.048cM and bEUR = 0.011cM respectively (see Materials and Methods for details). We estimate population density nAFR = 1.9km−2 and diffusion constant DAFR = 63.5km2/generation for African-Americans across the US, and nEUR = 7.6km−2 and DEUR = 59.6km2/generation for European-Americans (Fig 3E). The ratio of European- to African-American inferred population density is therefore 3.9. According to the 2010 US Census, 13% of the total population have self-identified as “Black or African American alone” and 72% self-identified as “White alone”. The ratio of European- to African-American population size from the census is 5.5, in good agreement to our estimate above. Interestingly, the root mean squared displacement per generation, 2 D × 1 generation ∼ 15 − 16 km, shows comparable local migration rates in European-Americans and African-Americans despite the different histories and population densities.
This root mean square (RMS) displacement is much less than the contemporary RMS parent-offspring dispersal in the US, estimated at 989km, but within the range of other modern human populations (2.6–300km) [31]. RMS displacement is heavily influenced by the largest displacement, and the latter study found approximately 27% of parent-offspring displacements in the US to be over the 1000km range. Such long-range migrations did not appear to leave a strong signature of isolation-by-distance in our IBD data and were captured by the uniform background term in our model. The RMS displacement in our model therefore does not account for such long-range migrations.
The history of African-American populations combines strong ties to place with large-scale migrations [4]. This comprehensive study shows the combined effects of fine-scale population structure, large-scale migrations, and admixture in shaping genetic diversity among African-Americans. Detailed models of genomic diversity recapitulate known historical events, such as the travel routes used during the Great Migration [4, 23] and the timing, amount, and geography of admixture between African, European, and Native ancestors [22, 24, 32–34]. They also quantify demographic effects that were less well characterized, such as ancestry-biased migration and the geographic patterns of relatedness among African-Americans. The observed ancestry-biased migrations of African-Americans suggest that the differences in social opportunity afforded to individuals with different levels of European ancestry at the time of the Great Migration [23] contributed to shaping the genetic population structure of contemporary African-Americans.
The observed patterns of relatedness have consequences for genetics research. Long IBD segments are often inherited from a recent common ancestor and are likely to carry shared but recent mutations. Such variants are more likely to be deleterious than older variants and are therefore prime targets for disease-mapping studies of rare traits [35]. Considering our analysis of long-range IBD sharing across the US, we expect rare monogenetic traits to be more often shared over long distances among African-Americans than among European-Americans, particularly along the routes of the Great Migration. Yet, their spatial distributions over short ranges should be as structured as in European-Americans.
Despite the overall correlation in regional admixture proportions among the SCCS, HRS, and 23andMe cohorts, significant differences remain in nation-wide and regional ancestry proportions. Such differences likely result from sampling biases that correlate with existing population structure through geography, urban/rural status, wealth, education level, and identity. Detailed sampling and sociodemographic modeling should therefore inform the design and analysis of large genetic cohorts that include African-Americans, as well as further efforts to understand the genetic makeup of African-American communities.
The use of these samples for the present study was approved by the IRB at McGill University and Stanford University, where the analyses were performed.
We used the genotype data of 12,454 individuals from the Health and Retirement Study [18] (HRS), genotyped on the Illumina Human Omni 2.5M platform, and of 2,169 African-American individuals from the Southern Community Cohort Study [19] (SCCS), genotyped on either Illumina Human Omni 2.5M or Human 1M-Duo platforms. The HRS cohort includes 1,649 individuals who self-identified as African-Americans (non-ambiguously in both HRS Tracker and dbGaP databases) and 10,432 individuals who self-identified as European-Americans. There are also 366 individuals labeled as “Others” whom we have not used in our main analyses (except in a PCA analysis, discussed below). The remaining 7 individuals have ambiguous, non-matching race identifiers in HRS Tracker and dbGaP, and we have, thus, excluded them from our analyses.
We performed comparisons with data from 23andMe [12] and from 97 individuals of African ancestry from the southwest USA (ASW) from the 1000 Genomes Project (at ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/supporting/hd_genotype_chip/) [20]. The 23andMe cohort includes many African-American individuals and has been the subject of a detailed population genetic analysis [12], and the ASW cohort has been a reference African-American population in recent studies. However, these two cohorts were not meant to be representative of the US population. The 23andMe database has a complex ascertainment scheme, which may cause biases in ancestry and socioeconomic status. In particular, biases in regional representation and a small amount of survey response errors might lead to a lower European ancestry proportion. These possible biases are described in detail in [12]. Similarly, the ASW cohort was assembled from duos and trios with at least one Oklahoma resident, but with no attempt to reach geographic or demographic representativeness (Morris Foster, personal communication). For comparisons with the 23andMe study, we used the global ancestry proportions reported in [12], because the genotype data is not publicly available. The global ancestry proportions reported in the 23andMe study are calculated by first using their in-house local ancestry assignment pipeline and then aggregating the results across the genome, as described in detail in [12]; we employ a similar scheme, described below in detail.
The HRS genotype data that we received had been already quality controlled, filtered, and phased. The SCCS cohort comprises data from 648 individuals in a breast cancer study (genotyped on Illumina Omni 2.5M platform) and 760 individuals in a prostate cancer study, 484 individuals in a lung cancer study, and 277 individuals in a colorectal cancer study (genotyped on Illumina Human 1M-Duo). All genotyped individuals were either cases or controls in their respective nested case-control studies. We converted the lung cancer dataset from human genome assembly hg18 to hg19 using the LiftOver utility from the UCSC Genome Bioinformatics Group and merged the four separate SCCS datasets into one using PLINK 1.9 [36]. During the merge process, we removed markers to which more than one name was assigned at the same position along a chromosome; removed markers with missing genotype calls; corrected unambiguous strand misassignments and removed ambiguous strand (mis)assignments; removed multi-allelic markers; and, finally, filtered the data for missing calls [37] first based on genotypes (PLINK argument --geno 0.0125) and then based on call rates per individual and minor allele frequency (PLINK arguments --mind 0.0125 --maf 0.01). The final SCCS dataset contains 2,128 individuals and 585,527 variants after these steps. We then used the same process to merge the HRS data with those of SCCS and ASW, resulting in a single dataset in PLINK format with 14,679 individuals and 553,795 variants. Performing a PCA on the data (pruning for LD leaves 77,902 markers), we found no batch effects (see S1 Fig). We then phased the merged data with SHAPEIT2 [38] (default arguments), and converted the output to PLINK format (while preserving the phasing information) using genetic map information from the 1000 Genomes Project data (at http://mathgen.stats.ox.ac.uk/impute/data_download_1000G_phase1_integrated_SHAPEIT2_9-12-13.html).
Geographic information in HRS is usually provided in the form of US census regions and divisions. We have used these locales in the ancestry analyses. ZIP code information for HRS study participants is available, but use of this data is restricted. We used zip code data only for the fine-scale spatial analysis of identity-by-descent relatedness. For SCCS, latitude and longitude coordinates of clinics were available. In the IBD analysis, we assigned the ASW individuals to the West South Central census division (see, e.g., https://catalog.coriell.org/1/NHGRI/Collections/1000-Genomes-Collections/African-Ancestry-in-SW-USA-ASW). In terms of geographic locations, we restrict our analyses to the census divisions in the contiguous United States (i.e., Pacific, Mountain, West North Central, East North Central, Middle Atlantic, New England, West South Central, East South Central, South Atlantic).
For the individuals in HRS, we only consider the ones born in the contiguous US who, at the time of sampling in 2010, also lived in the contiguous US; this reduces our sample size in HRS to 10,974 individuals of which 1,501 are self-identified African-Americans and 9,308 are self-identified European-Americans (with the remaining individuals being classified as “Others”). There are 4 additional individuals satisfying the geographic constraints above but who have discordant race identifiers in two different data files provided with the cohort data; these were removed from any downstream analysis. Among the unambiguous self-identified African-Americans and European-Americans mentioned above, there are respectively 10 and 427 individuals also self-identifying as Hispanics. The former 10 individuals are only included in our analysis of Hispanics status. In S1 Table, we summarize a few characteristics of the HRS African-American and SCCS cohorts, namely, the number of sampled individuals, the number of males and females, the number of Hispanics (if specified), and the locale.
African-American sample sizes in the New England and Mountain census divisions are small. We therefore merged the New England and Middle Atlantic divisions, and considered the Northeast census region as a whole. Similarly, we merged the Mountain and Pacific, and considered the West census division as a whole. The total number of geographic locales under consideration was therefore 7, namely, Northeast, Midwest consisting of 2 divisions, South consisting of 3 divisions, and West. We show in S2 Table the number of non-Hispanic individuals in our analyses separated by race and region of residence in 2010. The individuals are selected to have been born and to have lived within the contiguous United States at the time of sampling. These numbers are derived by combining the HRS, SCCS, and ASW cohorts, as described above.
We used GERMLINE [39] (arguments -err_hom 1 -haploid -bits 32 -w_extend) to infer IBD tracts of length 3cM or longer shared between individuals from the HRS, SCCS, and ASW cohorts. GERMLINE is prone to false positive IBD assignment, particularly at positions overlapping assembly gaps (see, e.g., [40]). It is therefore standard practice to filter out these regions [39, 40]. We developed a filtering strategy that improves on this practice by allowing the possibility of keeping long IBD tracts that span a troublesome regions, considering that GERMLINE is known to be more accurate for longer tracts [40].
We first count, for each genomic position, the number of overlapping IBD segments across all individuals. A chromosomal region is then marked as “forbidden” if the total number of IBD segments overlapping it is larger than a threshold, as follows. We determine a single background IBD count by comparing the total count for each position across the genome to the average count across the genome. We find that each genomic position is overlapped by approximately 15,000 IBD segments and, thus, take the threshold to be 25,000 to allow for some variation in the total number of IBD segments shared. Next, two forbidden regions will be merged as one if they are less than 0.1cM apart. IBD segments that overlap these forbidden regions are excluded from the downstream analysis unless they extend outside the forbidden regions by at least 3cM. In that case, we presume that there is sufficient evidence in the non-forbidden regions, and the segments are kept. After this filtering process, we are left with 8,664,251 IBD segments out of the total of 71,633,425, and a relatively uniform coverage of IBD across the genome.
Geographic information and inferred IBD segments were used to construct a relatedness metric between individuals and geographic regions within the cohorts. We first bin the IBD segments by length. The first bin contains segments of length between 3cM to 10cM, the second bin contains segments from 10cM to 18cM, and the last bin contains segments of length 18cM or longer. The latter bin corresponds to common ancestors living about 8 generations ago and is the focus of most of our discussion. Sorting the individuals by region and by African-American status within each region, we form two sparse relatedness matrices: L which contains the total IBD length shared between each pair of individuals, and N which contains the total number of shared IBD segments between each pair of individuals. The diagonal elements of L and N, which represent self-IBD, are set to zero by definition.
We next remove the contributions of closely related individuals from these matrices as follows. The HRS study has already identified 89 pairs of individuals having kinship coefficients greater than or equal to 0.1. To be consistent with the definition from HRS, we used PLINK to calculate kinship coefficients for SCCS and ASW individuals, labelling individuals with kinship coefficient of 0.1 or higher as related individuals. We find 22 related pairs among SCCS individuals, 62 related pairs among ASW individuals, and 1 related pairs between HRS and SCCS individuals (details below).
To see how geographic regions are associated based on the genetic relatedness of their inhabitants, we consider average pairwise IBD relatedness between regions [28]. The average pairwise relatedness L between two regions R1 and R2 is defined as the mean length of IBD segments shared between pairs of individuals, where one individual is from R1 and the other from R2. In addition, we consider the relationships between individuals of specific ancestry S1 and S2, each representing either African-American or European-American. Thus, the average total shared IBD length becomes
L ( R 1 , S 1 ) , ( R 2 , S 2 ) = ∑ i , j ′ L i j N pairs (2)
where
Using the metric defined above, we can calculate the pattern of relatedness between geographic locations among African-Americans, among European-Americans, and between African-Americans and European-Americans. The first two matrices are symmetric with respect to changes in the order of regions, whereas the last one is not.
The following criteria were used for visualization of the IBD relatedness between regions. Due to the small number of sampled African-American individuals in the northern and western regions, the total number of IBD segments shared between these regions is small compared with that between other regions (see the bottom row in S14 Fig). Relatedness estimations are noisy for such pairs, and a scale that accommodates these noisy results would not allow for detailed comparison of less noisy results. Therefore, in Fig 3, S12 and S13 Figs, we did not draw the lines between any two distinct regions for which the total number of possible pairs of IBD individuals is less than 10,000 (e.g., notice the lack of connecting lines from West North Central to West). Since a significant number of the individuals in HRS are European-Americans, the number of IBD segments shared between European-Americans residing in any two regions is large enough to ensure the significance of the results, even when we restrict the analysis to the longest IBD segments (see the bottom row in S15 Fig).
We are interested in comparing the relatedness information derived from genomic data to those described in historical migration records, e.g., available from Integrated Public Use Microdata Series (IPUMS) [2]. Here, we describe a simple coalescent-based method to calculate a relatedness metric based on census data. Despite many simplifying assumptions, this metric is able to capture the dominant relatedness patterns originating from recent migration events and, therefore, provides a first-order model to understand relatedness patterns across the US.
We downloaded census data from 1900 to 1980 and extracted census year, census region, age, race, birth place, and weighted representation of each sample; the latter is the number of people in the population represented by the sampled individual. For any decade, we focus on the people in the age group of 20- to 30-year olds and consider the migrations of African-Americans and European-Americans separately. We assume a generation time of 30 years, thereby taking census years 1900, 1910, and 1920 as generation 3; 1930, 1940, and 1950 as generation 2; and 1960, 1970, and 1980 as generation 1. For each ancestry group, we construct a matrix whose elements m i j ( g ) represent the number of migrations at generation g ∈ {1, 2, 3} from region i to region j; this matrix is highly asymmetric because of asymmetric nature of the migrations between geographical regions.
We now construct a heuristic census-based measure of relatedness between regions. Let us define p i → j ( g ) as the proportion of individuals in region j at generations g − 1 whose ancestors were in region i at generations g. In other words, the (i, j) element of the matrix P(g) is
p i → j ( g ) = m i j ( g ) ∑ i ′ m i ′ j ( g ) + m out → j ( g ) (3)
where g ∈ {1, 2, 3} denotes the generation time of the ancestral population, m i j ( g ) denotes the number of migrations from region i into census region j (as constructed above), and m out → j ( g ) is the number of migrants from outside of contiguous United States into the census region j. Had we not included migrations from outside the US into the mainland US, P(g) would have been column-normalized (i.e., normalized with respect to the destination census regions).
A three-generation transition matrix can be constructed as
P ¯ = P ( 3 ) P ( 2 ) P ( 1 ) (4)
where, by matrix multiplication of migration probabilities for all generations under consideration, P ‾ i j takes into account all possible migration routes starting at region i and ending at region j that could have taken place in the span of these three generations.
To estimate genetic relatedness between different geographic regions, we further make the coarse assumption that population sizes were constant before 1910 and that populations were randomly mating. These parsimonious assumptions allow us to model the expected relatedness within regions using coalescent theory before the massive 20th century migrations. Neither assumption is expected to hold exactly, but the randomly mating, constant-population model is expected to capture the bulk of variation in the coalescence rate across regions.
Given P ‾ k , i as the probability of a sampled individual from region i having an ancestor from region k, we define the census relatedness metric between regions i and j as
I i j = ∑ k P ¯ k i P ¯ k j 1 N k (5)
where Nk is the census population size of region k. Population size matters, because in larger populations, it is less likely that a given pair of individuals share a common ancestor. The number of common ancestors at each generation is approximately inversely proportional to Nk, and therefore the expected recent shared ancestry is also approximately inversely proportional to Nk. Thus, Iij is proportional to the probability of two individuals from regions i and j having ancestors from (any) region k times the probability that these ancestors have a recent common ancestor within region k. Unlike P ‾ which is a directional metric, I is non-directional and symmetric and can be directly compared with the genetic relatedness matrix L in Eq (2), which was estimated using IBD data. The regional relatedness patterns derived using P ‾ and I are shown in S17 and S18 Figs.
To test the hypothesis regarding South-to-North migration corridors, we consider the matrix elements corresponding to relatedness between the three southern regions (South Atlantic, East South Central, West South Central) and the three northern ones (Northeast, East North Central, West North Central), forming a 3 × 3 matrix from the census data to be compared with the corresponding matrix from IBD data. To quantify the correlation between these two matrices, we use the Mantel test (which is a standard test of correlation between matrices) as follows. We perform 9! possible permutations on the elements of the matrix derived from the census data and calculate the Pearson correlation coefficient between the original IBD matrix and the permuted census matrix. We then accept or reject each permutation based on whether the calculated correlation coefficient is lower or higher than the correlation coefficient between the two original (non-permuted) matrices. The p-value is given by the ratio of the number of rejections to the total number of permutations (see main text for the numerical values). The p-value reported in the main text for the relatedness between South to North and West are derived by performing a random subset of 107 permutations out of a total of 12! ones.
In addition to the tests above, we also perform a test using the region of birth of HRS individuals as their location, which roughly translates to the migrations during the first wave of the Great Migration. Given the average year of birth (1939.8) and the birth year distribution (S3 Table) in HRS, we only take, for consistency, generation 3 from the census data (see definition above) and write P ‾ = P ( 3 ) as our overall directional relatedness matrix (compare with Eq (4) above). We then proceed as before to calculate the non-directional (symmetric) relatedness I. Given the new census-based prediction (using only g = 3 above) and the IBD relatedness pattern (using the region of birth), we perform a Mantel test, as described above, in order to find the correlation between the data and our prediction.
Even though Fig 3 only shows pairs of regions for which 10,000 possible pairs of individuals were available, the Mantel test procedure uses all pairs of regions regardless of the number of individuals they contain.
We wish to model the expected IBD relatedness between individuals in a spatially extended population. Our starting point is an idealized population living on a set of islands (or demes), with random mating within islands and migrations between the islands. We will consider a limiting example of a continuous population below.
We are interested in the probability that a genomic segment of given length, stretching across a specific locus, is shared identical-by-descent between two randomly selected individuals living on different islands. For identity-by-descent to occur, we need two events to happen: (a) lineages at that locus must have coexisted on one unknown island at some point in the past, and (b) these two geographically coexisting lineages must also have coalesced further in the past.
We measure time in generations and track lineages backwards in time. At each generation, we assume that the displacement between parental birthplace and offspring birthplace follows a random walk. Each lineage follows a random walk on the islands, with each step representing one generation back in time, connecting an individual to the ancestor from whom the locus is inherited. The lineages are then traced back until the time at which both ancestors coexist on the same island and coalesce in the most recent common ancestor in the next step back in time. We can, therefore, symbolically write the total probability of coalescence at a given generation as the probability of coexistence times the probability of coalescence, i.e.,
Pr ( coalescence ) = ∑ island Pr ( lineage 1 , 2 ∈ island ) Pr ( coalescence | lineage 1 , 2 ∈ island ) . (6)
To derive the probability of coexistence, we first want to estimate the expected position of a lineage given its position in the past. Concretely, let x0 be the current location of an individual at t = 0. We would like to find Φ(x, t|x0), the probability that an individual’s lineage is on island x at t generations ago, given that it is currently on island x0.
By construction, the probability Φ(x, t|x0) takes into account contributions from all possible space-time paths that start at x0 and end at x at time t. For instance, a possible path is to arrive at x at t/2 and stay at that position until t, whereas another path is to arrive at x at t/3, leave x at the next step for a series of random walks to finally arrive at x again at t.
Consider a region of area ΔAi that encompasses a deme with haploid population 2n(xi, t)ΔAi, where n(xi, t) is the effective diploid population density at position xi and time t in the past. The probability that two lineages in ΔAi coalesce in a given generation is
p coal ( x i , t ) = 1 2 n ( x i , t ) Δ A i . (7)
This expression does not consider the possibility of multiple coalescent events and is therefore appropriate only for a number of generations that is much less than the population size.
The discrete probability of two lineages having coexisted on the deme at xi at time t in the past, given that they are a distance R apart (at x0 and x0+R) at present (at t = 0), is
p coex ( x i , t | R ) = Φ ( x i , t | x 0 ) Φ ( x i , t | x 0 + R ) . (8)
Therefore, the total probability of having a common ancestor t generations ago in the discrete model is
p ( t | R ) = ∑ i Φ ( x i , t | x 0 ) Φ ( x i , t | x 0 + R ) 2 n ( x i , t ) Δ A i . (9)
To go from a discrete random walk to the continuous limit, we set Φ(x, t|x0)→φ(x, t|x0)ΔA, where φ(x, t) is now a continuous probability density. Thus, in this limit (with ∑i ΔAi···→∫d2 x…), we get
p ( t | R ) = ∫ d 2 x φ ( x , t | x 0 ) φ ( x , t | x 0 + R ) 2 n ( x , t ) . (10)
The continuous limit of a random walk process is the diffusion model. In this model, the probability density φ(x, t) of finding a lineage at an infinitesimal area d2 x centered around x at generation t in the past obeys the two-dimensional partial differential equation
∂ ∂ t φ ( x , t ) = ∇ x · D ( x ) ∇ x φ ( x , t ) (11)
where the diffusion coefficient D(x) encompasses the information related, in the discrete model, to probabilities of taking a step to an adjacent island or staying on the same island (for a discussion around the connection a random walk and a diffusion process, see http://ocw.mit.edu/courses/mathematics/18-366-random-walks-and-diffusion-fall-2006/). Solving for φ(x, t|x0) amounts to solving Eq (11) with initial condition φ(x, t = 0) = δ(x − x0) where δ(x) is the (two-dimensional) Dirac delta function.
For simplicity, we consider random walks with uniform probability of transitioning to any nearest-neighbor island, which translates to a constant (position-independent) D in the continuous model. We also assume that all islands have the same constant population size, leading to a population density which, on average, is constant in the continuous model.
Under these assumptions, we have
φ ( x , t | x 0 ) = 1 4 π D t exp - | x - x 0 | 2 4 D t (12)
which, in turn, leads to
p ( t | R ) = 1 16 π n D t exp - R 2 8 D t (13)
with R = |R|.
Following Palamara et al. [29], we approximate the expected fraction of the genome shared through segments in the length range ℓ = [lmin, lmax] (in units of Morgans) as
E ℓ [ f | R ] = ∫ l min l max d l ∫ 0 ∞ d t p ( l | t ) p ( t | R ) (14)
with p(l|t) = (2t)2l exp(−2tl) the probability density of an IBD segment of length l (in units of Morgans) spanning the locus shared by the two randomly chosen individuals whose lineages coalesce t generations ago. Performing the integrals above leads to the following closed form solution for the expected fraction of the genome shared as a function of spatial separation
E ℓ [ f | R ] = 1 16 π n D 2 K 0 R r min - K 0 R r max + R r min K 1 R r min - R r max K 1 R r max (15)
where Kα(x) is the modified Bessel function of the second kind [30], and r i = D / l i with i ∈ {min, max}. Expanding for small R, we find E ℓ [ f | R ] ≃ 1 16 π n D [ ln ( l max / l min ) − ( l max − l min ) R 2 / 4 D + O ( R 4 ) ]. We can use Eq (15) to approximate the amount of IBD in a finite chromosome of length Lc by setting lmax = Lc in Eq (15). This yields E [fc|R]≡E[lmin, Lc] [f|R]. We come back to this approximation at the end of this section.
The total length of shared IBD tracts across all chromosomes, L, between a random pair of individuals, therefore, becomes
E [ L | R ] = ∑ c = 1 22 L c E [ f c | R ] . (16)
This quantity can be directly compared with that calculated from the IBD data to estimate the parameters of the model. Technically, this model makes two important relatively coarse approximations. First, in Eq (14), we have integrated from t = 0, even though coalescence from time t = 0 to t = 1 is not allowed. Second, when considering finite chromosome, Access to the exact location of clinics at which the SCCS cohort was sampled allows us to investigate the relation between IBD relatedness and spatial distance. Having inferred possible IBD segments using GERMLINE, we calculate, for each pair of individuals from SCCS, the total length of shared IBD and the distance between the clinics in which they were sampled. We make the underlying assumption that each individual lives close to the clinic at which he or she was sampled. Each pair is then placed, based on the distance between the two individuals, into one of the length bins in {[0, 1), [1, 101), [101, 201), [201, 301), …} (all numbers in kilometers). The first length bin, [0, 1), contains individuals sampled at the same clinic. For each bin, we calculate the average pairwise IBD length (the sum of the IBD lengths of all pairs divided by the total number of points in the bin) and assign it to a distance equal to the midpoint of the bin (e.g., for the length bin [1, 101), the assigned distance is 51km). The result is shown in S19 Fig.
Apart from the expected decay of relatedness with distance, we also notice the presence of a constant background IBD. This background IBD is larger for shorter IBD segments. As mentioned in the main text, this could be attributed to two possible factors: (a) GERMLINE has a higher false positive detection rate for shorter IBD segments [40] which is independent of the distance between individuals, or (b) shorter IBD segments, being much older on average, reflect history prior to migrations from Europe and Africa into the Americas. Since this relatedness patterns extends over long distances with little evidence for decay, we suppose that it is either due to false positives, or that there was enough mixing in the travels into the Americas that present-day proximity is a relatively poor proxy for the proximity of ancestors prior to transatlantic travels. In either case, the background IBD can be modeled by adding a constant term to our model in Eq (16), representing the expected fraction of the genome shared IBD by individuals over long distances.
The parameters to be inferred in this model are the haploid population density n, the diffusion coefficient D, and background IBD b. By fitting the SCCS African-American IBD data for the 18cM case (corresponding to the most recent sharing events), we find the estimated values b18 = 0.0389cM, n18 = 2.8km−2, and D18 = 88.6km2/generation. The root mean squared displacement for African-Americans in the South is thus estimated, using the IBD data from SCCS, to be 18.8km. We can use the population density and diffusion coefficient derived above to predict IBD decay for IBD segments of different lengths and estimate the background IBD for the other two cases (bins with segments of length 10cM or longer and with segments of length 3cM or longer), finding b10|18 = 0.120cM and b3|18 = 0.546cM. The resulting fits show good agreement with the data, as shown in S20 Fig.
The current model of isolation-by-distance makes two approximations in addition to the assumptions of a uniform, random-mating population. First, following Ref. [29], we approximated a discrete-generation model with a continuous-time model, as shown by the time integral in Eq (14). The integral’s lower bound at t = 0 suggests that close relatives are included in the model. Second, we assumed an infinite-genome approximation for p(l|t), as derived in Ref. [29], and accounted for finite-genome effects by setting lmax = Lc, noting that a shared IBD tract on chromosome c can be at most of length Lc. However, to properly account for finite-genome effects, it would be preferable to consider the IBD segments in the infinite-genome scenario and derive their appropriate distribution using a sliding ‘window’ to represent a chromosome of finite length [16]. To verify that our results are robust to these approximations, we computed p ˜ ( l | L c , t ) using the finite-genome approach of [16] and performed the proper integrals, from t = 1 onward, numerically. This was considerably more computationally intensive, but we found that the these corrections lead to results that are qualitatively very similar to what we have derived using the simpler approach described in Eq (16), with the effective population density n18 ≃ 2.4km−2 and the diffusion coefficient D18 ≃ 88.4km2/generation for the SCCS cohort.
Our IBD-based results could be sensitive to computational phasing errors which break up IBD tracts into shorter ones. To assess the overall effect of these errors, we used RFMix to perform phase correction on a subset of the data, used this output for IBD calling with GERMLINE, and recalculated regional relatedness patterns. We then compared these new patterns with those obtained from the same subset of data without a phase correction step. We did not observe any significant difference in relatedness values between geographical regions across the US. For the isolation-by-distance scenario, we expect that the breaking of IBD tracts would lower the overall relatedness uniformly, thereby we expect to have underestimated the densities and, similarly, overestimated the displacements by a small margin.
We expect the ascertainment bias to have negligible effect on our analyses, given that our results are based on information obtained from long haplotypes as opposed to that obtained from summary statistics based on single SNPs (e.g., allele frequency) which are more likely to be sensitive to SNP ascertainment scheme [41].
For reference, we derive the expected generation time to the most recent common ancestor (MRCA), given an IBD tract of certain length. The probability density of having an IBD segment of length l (in units of Morgans) spanning a chosen marker (denoted by ζ) inherited from a MRCA living g generations ago (assumed continuous for simplicity) is [29]
p ( l | g ) = 2 g 1 M 2 l exp - 2 g 1 M l . (17)
In the continuous limit to the Wright-Fisher model, given the shared locus ζ, the probability of having a MCRA g generation ago is
p ( g ) = 1 N e - g / N (18)
where N is the (constant) effective haploid population size. Therefore, given the length l of an IBD tract (in units of Morgans), we use Eqs (17) and (18) to find the expected value for the generation time of the MRCA
E [ g | l ] = ∫ 0 ∞ g p ( g | l ) d g = ∫ 0 ∞ g p ( l | g ) p ( g ) p ( l ) d g = ∫ 0 ∞ g p ( l | g ) p ( g ) d g ∫ 0 ∞ p ( l | g ) p ( g ) d g ≃ 3 2 ( l / 1 M ) (19)
where we have assumed that the haploid population size N ≫ 1 in the last step.
After the phasing process (discussed previously), we used RFMix [11] with arguments PopPhased --skip-check-input-format for local ancestry inference along the genome. We used available parents among the trios in the Southern Han Chinese (CHS), Yoruba in Ibadan, Nigeria (YRI), and Utah Residents (CEPH) with Northern and Western European Ancestry (CEU) populations from the 1000 Genomes Project (at ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/phase1/analysis_results/supporting/omni_haplotypes/) as a reference panel, comprising 50 CHS, 97 YRI, and 91 CEU individuals. We extracted the intersecting set of SNPs between our merged dataset and the three reference populations mentioned above, which we used as the input to RFMix. RFMix assigned continental ancestry of each marker in each sample to either CHS, YRI, and CEU, which we interpret as Native American/Asian, African, and European respectively. The local ancestry calls from RFMix for the SCCS are available from the Southern Community Cohort Study cohort through the Online Request System (ORS).
We used the local ancestry estimates obtained from RFMix to calculate global ancestry proportions for the HRS, SCCS, and ASW cohorts by dividing the total length of all tracts assigned to an ancestry (African, European, and Native American/Asian) to the total length of all assigned tracts (see S2 Fig).
For the X chromosome, a supervised run of ADMIXTURE with K = 3 reference populations (YRI representing African ancestry, CEU representing European ancestry, and CHS representing Native American/Asian ancestry) provided the ancestry breakdown shown in S3 Fig.
We performed a supervised K = 4 run of ADMIXTURE [42] on African-Americans from HRS, SCCS, and ASW, with the YRI, CHS, GBR, IBS cohorts from the 1000 Genomes Project used as the reference populations representing African, Native American/Asian, northern European, and southern European ancestral populations. Pruning for LD was performed based on the recommendations of the authors of ADMIXTURE (PLINK arguments --indep-pairwise 50 10 0.1). The mean ancestry proportions for African-Americans in HRS, as estimated by ADMIXTURE, are 81.583% for African, 17.333% for European (southern and northern combined), and 1.083% for Native American, in very good agreement with those derived using local ancestry estimates of RFMix (see main text). In comparison, the ancestry proportions for the ASW cohort are 75.726% for African, 21.881% for European (southern and northern combined), and 2.394% for Native American.
S5 Fig depicts the ancestry estimates for African-Americans in the ASW, HRS, and SCCS cohorts respectively, sorted by Native American proportions (shown in yellow). The top panel shows that ASW individuals with higher proportion of southern European ancestry (shown in green) tend to also have a higher proportion of Native American ancestry, and this pattern is repeated in the other two cohorts. This is especially true for HRS African-Americans who have self-identified as Hispanics (marked by the small black arrows in the middle plot). This correlation is also apparent in S6 Fig, which shows the proportion of southern European ancestry within the total European ancestry versus the Native American ancestry for HRS African-Americans. The correlation is particularly clear for self-identified Hispanic individuals. Note the presence of individuals who have not self-identified as Hispanics but have high proportions of both southern European and Native American ancestries. Moreover, SCCS African-Americans from Louisiana exhibit a similar pattern, as depicted by the black dots in S6 Fig.
To ensure that the inferred Native American ancestry reflects the true Native American ancestry, and not mis-assignment of European or African ancestry segments, we performed simulations based on a two- and a three-population admixture model. In both cases, we generated ancestry tracts for 50 admixed diploid genomes in a forward Wright-Fisher model with a single pulse of admixture 8 generations ago.
For the two-population admixture model, the ancestry proportions in the simulated individuals were 74.96% African and 25.04% European. We copied genotypes from one YRI sample into African ancestry segments and one TSI sample into the European segments (both samples from the 1000 Genomes Project) to generate 100 haploid chromosome 1’s. Each chromosome 1 was generated using a distinct source chromosome in the YRI and TSI population. We then inferred the ancestries of the individual i (corresponding to haplotypes 2i−1 and 2i) with panels composed of samples chosen from 91 CEU, 50 CHS, and 96 YRI, ensuring that the individual from whom the genotypes were copied was not used in the reference panel. We inferred 74.96% African, 24.95% European, and 0.09% Native American ancestry.
For the three-population admixture model, we simulated a sample of 100 haploid chromosomes with 80.9% YRI, 18.2% TSI, and 0.91% JPT ancestry, using the same method described above. In this case, the inferred proportions were 80.9% African, 18.2% European, and 0.94% Native American. These results are consistent with previous estimates of false assignment using a similar pipeline [11].
We also considered whether the amount of Native American ancestry in real samples correlated with the amount of European ancestry. If European segments are more likely to be misinterpreted as Native American, we would expect a positive correlation between inferred Native American and European proportions. Conversely, if the increased diversity in African segments led to higher rates of misidentification as Native American ancestry, we’d expect the correlation to be negative. The relation between Native American ancestry and European ancestry within SCCS is shown in S4 Fig. Within the southern states, only Louisiana shows a significant correlation. The lack of global correlation between European and Native American ancestry helps support the correctness of the inference.
Finally, we also compared global ancestry proportions inferred by RFMix and by ADMIXTURE (in supervised mode) and found an extremely high correlation between the estimates from the two methods, as shown in S7 Fig.
To infer time of admixture between ancestral populations and to identify migration models that give rise to the observed genome-wide patterns of ancestry, we use Tracts [16]. We excluded for this analysis HRS African-Americans from non-mainland US (96 individuals), African-Americans with self-reported Hispanic ethnicity (32 additional individuals), and one additional African-American who was listed as “White, non-Hispanic” in HRS Tracker but as “African-American” in dbGaP. All individuals were kept in the other cohorts. Optimization was performed for 6 models in each cohort: 2 two-population models, and 4 three-population models. Confidence intervals were then calculated.
|
10.1371/journal.pgen.1006113 | Pneumococcal Competence Coordination Relies on a Cell-Contact Sensing Mechanism | Bacteria have evolved various inducible genetic programs to face many types of stress that challenge their growth and survival. Competence is one such program. It enables genetic transformation, a major horizontal gene transfer process. Competence development in liquid cultures of Streptococcus pneumoniae is synchronized within the whole cell population. This collective behavior is known to depend on an exported signaling Competence Stimulating Peptide (CSP), whose action generates a positive feedback loop. However, it is unclear how this CSP-dependent population switch is coordinated. By monitoring spontaneous competence development in real time during growth of four distinct pneumococcal lineages, we have found that competence shift in the population relies on a self-activated cell fraction that arises via a growth time-dependent mechanism. We demonstrate that CSP remains bound to cells during this event, and conclude that the rate of competence development corresponds to the propagation of competence by contact between activated and quiescent cells. We validated this two-step cell-contact sensing mechanism by measuring competence development during co-cultivation of strains with altered capacity to produce or respond to CSP. Finally, we found that the membrane protein ComD retains the CSP, limiting its free diffusion in the medium. We propose that competence initiator cells originate stochastically in response to stress, to form a distinct subpopulation that then transmits the CSP by cell-cell contact.
| Development of competence for genetic transformation by cultures of pneumococcal cells has been considered till now as a classic example of quorum sensing, whereby a culture attaining a sufficient cell density detects a diffusible signaling molecule (in this case, Competence-Stimulating Peptide (CSP)) and switches en masse to a distinct physiological state. We find that the competence shift is dictated not by cell density but by growth for a time allowing emergence of a competence-initiator sub-population, and spreads by transmission of CSP through cell contact. This behaviour reflects the survival benefits of allowing subsets of the population to respond to environmental stress by generating signalling capacity, which prepares the entire population for a rapid and appropriate response to threatening conditions.
| Under certain circumstances, single bacterial cells can sense environmental conditions and stimulate collective behavior by using exported signaling molecules that act as auto-inducers (AI). The two first processes found to be stimulated by AI sensing were luminescence in Vibrio fischeri [1] and competence for transformation in Streptococcus pneumoniae (the pneumococcus) [2]. Other examples of collective behavior have since been found [3–5]. These AI-based systems clearly differ in their mechanism. The first that has been defined is the Quorum Sensing (QS). It was proposed to take place by a rise in the concentration of a diffusible AI to a threshold level at which it induces the entire population to switch synchronously to a new gene expression program [6]. In particular, the QS mechanism implies that induction depends on the population achieving a given cell density (quorum) and on freely diffusing AI being produced at similar rates by all cells. The original QS model has been modified to take into account environmental parameters and the relative benefits for cells as individuals or as a group. The Diffusion Sensing mechanism includes the rate of loss of the AI in an open space [7], while the Efficiency Sensing mechanism takes account of cell distribution in complex environments [8]. Inclusion of these and other parameters has considerably increased the complexity of the original QS model [9]. Furthermore, the cost of each AI-based system and the specific purpose of the inducible genetic program are other important parameters that could have distinctly shaped their mechanism [10]. One such program, genetic competence in pneumococci, has long been thought to operate according to a QS model. This assumption has been challenged but without a clear alternative model emerging [11,12]. Here, we have studied in detail how the spontaneous development of competence for genetic transformation is coordinated throughout the population in planktonic pneumococcal culture.
Competence for genetic transformation is a distinct physiological state during which most of the proteins enabling cells to take-up and integrate exogenous DNA into the genome are produced. This process is widespread throughout the bacterial kingdom, wherein it acts as a central driver of evolution by promoting horizontal gene transfer [13]. Although transformation in all species proceeds through the same general mechanism, competence differs sharply among species at many levels. First, except for a small set of proteins dedicated to specific steps of the transformation mechanism, the genes that make up the competence regulon are variable. Second, the regulatory networks controlling the expression of competence genes differ considerably. Third, the time of competence induction and its duration with respect to culture growth cycle are clearly distinct among species, as are the factors that cause competence induction. These marked similarities and differences lead to the current notion that competence is tightly integrated into the life-style of each bacterial species [13]. A prominent difference so far observed only in pneumococcus and some closely related species is the synchronous development of competence of the whole cell population grown in liquid medium [14–16].
Two major properties are known to characterize Pneumococcal competence: import and integration of external DNA, called transformation, and killing of non-competent siblings or close relatives, called fratricide (for review see: [13,17]). In combination, these processes promote horizontal gene transfer and genome plasticity in pneumococci [18,19]. In support of this view, several studies have demonstrated how rapidly the pneumococcus can modify its genotype and undergo diversification [20–22]. These adaptations contribute to evasion of vaccines, antibiotics and host immune defenses [23–25].
In liquid cultures of Pneumococcus, competence develops transiently during exponential growth. Direct, continuous monitoring of competence gene expression [26] has revealed four consecutive steps in competence development, namely pre-competence, competence shift, competence development and shut-off (Fig 1A). The competence AI is a 17 amino-acid peptide called CSP (Competence Stimulating Peptide) [15]. It is at the heart of a positive feedback loop comprising 5 genes organized into two operons (Fig 1B). CSP is the product of the first gene of the comCDE operon, synthesized as a 41 amino acid precursor (pre-CSP) which is matured and exported by a dedicated membrane peptidase transporter, ComA/ComB, encoded by the comAB operon [17] (Fig 1B). Once CSP reaches a threshold concentration, it is sensed by the ComD membrane kinase of the two component system ComD/ComE. The membrane-bound ComD, activated by CSP interaction, mediates its autophosphorylation. In turn, ComD~P is assumed to transfer its phosphate to the ComE transcription regulator, switching it from a repressor to an activator which targets the comAB and comCDE operons and creates a positive feedback loop [27] (Fig 1B). Among the early competence genes activated by ComE~P is a pair of identical genes, comX1 and comX2, which encode ComX, an alternative sigma-factor that activates the late competence genes [28–31]. Of the ~100 genes induced, 22 are required for transformation and 6 for fratricide [32,33]. Competence is transient, and its shut off has been shown to involve the late competence gene dprA and to proceed through a physical interaction between DprA and ComE~P [34,35] (Fig 1B).
Like other AI based systems, the CSP-based positive loop has been proposed to define a core sensor module here called ComABCDE, through which competence induction is coordinated within the population [36]. The ComABCDE core sensor machinery appears to be in a homeostatic equilibrium that results from the balance of positive and negative inputs during the pre-competence state [17,37], with the former leading to competence shift (Fig 1B and 1C). Once the equilibrium is disrupted by positive inputs, the core sensor is activated in a positive feedback loop that switches the cells into competence. Parameters leading to spontaneous competence shift are not fully understood but appear to be numerous. They include environmental conditions such as initial pH of the medium and the presence of antibiotics, as well as cellular contributions revealed by strain-specific differences in competence development [37–40]. One such difference underlies the question we address in this study. Håvarstein and colleagues succeeded in purifying CSP from the supernatant of a culture of the CP1200 strain, a D39 derivative [15]. This suggested that CSP was able to freely diffuse in the culture medium. In a defined volume, constant CSP production during pre-competence would result in CSP reaching a threshold concentration inducing competence which correlates to a defined and fixed cell density, consistent with a bona fide QS mechanism. In contrast, our preliminary results with the R800 strain, another D39 derivative, showed that the population shifted into competence after a fixed time of growth and independently of cell density [11]. We defined the underlying mechanism as a timing device. Such a timing device has recently been observed for spontaneous competence induction in Streptococcus thermophilus, which relies on a specific exported peptide (named ComS) which targets a regulatory system different from that of S. pneumoniae [41]. In the light of these findings, and to extend our knowledge of cell-to-cell communication during pneumococcal competence development, we have investigated in closer detail the spontaneous shift to competence during planktonic growth of various strains, including representatives of the CP1250 and R800 lineages.
We present genetic evidence demonstrating that spontaneous competence development of a pneumococcal population generally occurs independently of cell density and is linked to the metabolic state of the cells. Hence, we renamed the timing device mechanism underlying pneumococcal competence development as a growth-time dependent (GTD) mechanism. We propose that competence shift is engaged when the sum of stochastic stress perceptions and responses reaches a level that activates the ComABCDE core in certain cells, and so activates the CSP-based feedback loop via an autocrine process. The time until the competence shift depends on the lineage genotype and the growth medium used. Competence throughout the population was found to be propagated from the induced fraction of cells through CSP-mediated paracrine activation of non-competent cells. Notably, we show that the neo-synthesized CSP is retained on cells during competence development, providing strong evidence that competence relies on CSP transmission by cell-to-cell contact. Mixed culture experiments performed with a wild-type strain and strains mutant for the ComABCDE module confirmed this mode of CSP transmission. Our experiments also provide evidence that ComD is required for non-competent cells to capture CSP from competent cells.
CP1250 (derived from Rx) and R800 (derived from R6) are the two main pneumococcal lineages used to study competence (S1 Fig). However, they behave differently with regard to synchrony of competence development. The CP1250 lineage is reported to develop competence as expected for a bona fide QS mechanism, dependent on CSP diffusion [15], whereas competence in the R800 lineage is triggered after a fixed period of exponential cell growth, independently of cell density (for review see: [11,17]). The basis of this difference has not been experimentally determined. As shown in S1 Fig, CP1250 and R800 are derived from the same virulent capsular serotype II parent strain D39 by serial genetic manipulation; notably, the acquisition of a mutation causing loss of mismatch repair in 1959 has resulted in higher mutability, generating the Rx strains from which CP1250 lineage strains have been derived [42,43].
We measured spontaneous competence development for the two strains representing these lineages. To get a broader view, we also analyzed a representative of the original D39 lineage and a recent clinical isolate G54 of the capsular serotype 19F [44]. Introduction into each strain of a luc transcriptional fusion to the late ssbB competence gene promoter enabled us to monitor competence development in real time during growth [26] (Fig 1A and 1B; see M&M). Competence development was followed as a function of growth time (Fig 2A) or of cell density (Fig 2B). Cells were initially grown for several generations in medium non-permissive for spontaneous competence development. Various quantities of these ‘naive cells’ were then inoculated into medium permissive for spontaneous competence development. The inoculum size used was chosen to ensure detection of competence induction above the photon threshold sensitivity of the luminometer [26] (see M&M). Each strain developed competence in all assays but competence development occurred at different cell densities (Fig 2B). Moreover, apart from CP1250, the competence shift occurred at the same moment during cell growth whatever the size of the inoculum (Fig 2A). We confirmed this behavior by calculating and plotting the competence shift time of each experiment against the OD of each inoculum (Fig 2C). If a population develops competence through QS, the results should fit the red horizontal dashed line corresponding to a fixed OD at the competence shift (arbitrarily taken as the OD reached at the competence shift with the highest density inoculum for each strain). Instead, the results obtained with the R800, D39 and G54 lineages fit a model of competence induction based on a constant growth time, represented in Fig 2C by the blue dashed line. The CP1250 strain appears to exhibit a behavior intermediate between those expected for QS and GTD (growth time-dependent) mechanisms, possibly as a result of its genetic differences from the three other strains (see discussion). Remarkably, the three lineages that trigger competence in a GTD manner do so at a constant time that is specific for each lineage (Figs 2A and 3A).
The competence profiles recorded throughout cell growth show that after the first competence cycle (defined by competence shift to shut off, Fig 1A), subsequent competence cycles correlated with the time of exponential growth in the non-depleted permissive medium (Fig 2A). The four strains exhibit a distinct pattern of these subsequent competence cycles as the growth time elapsed before the first shift to competence is distinctive (Figs 2A and 3A). We also found that this latter property, i.e. the time of competence shift of a given population, is affected by the nature of the non-permissive growth medium used (S2 Fig). Growth of the R800 lineage in two different non-permissive media before inoculation of the same permissive medium led to competence shifts separated by 42 minutes (bottom graph, S2 Fig). This modulation occurred despite growth rate being the same during the assay (top graph, S2 Fig). Thus, a metabolic memory applies to the cells during subsequent culture in permissive medium. These results show that spontaneous competence coordination in pneumococcal populations depends on growth time and is modulated by both genotype and environmental parameters.
If each individual cell switches to competence at the same time during growth, competence development within the population should be synchronous and, therefore, should proceed at the same rate independently of inoculum size. However, as is apparent from the plots of Fig 2A, the competence development rate increases with the inoculum size for all strains. Low cell density inoculation leads to a duration of the competence development phase ranging from 46 to 98 minutes much longer than the 25 to 35 minutes observed for the higher cell density inoculation (Fig 2A). This variation was quantified by calculation of the competence development rate for each inoculum size. The competence development rate can vary as much as 3-fold (Fig 3B). This observation strongly argues against autonomous synchronous development of competence by all cells of the population. Rather, this behavior implies that a subpopulation has switched to competence first, resulting in activation of the core sensor and in a high level of CSP production, which then induces the rest of the population through CSP transmission.
The two-step scenario for competence development raised the question of the responsiveness of the cells to CSP. We measured this responsiveness by adding synthetic CSP at the time of competence shift to R800 cells grown from a low density inoculum (Fig 4, 1: wt +CSP). As shown previously [15,26], addition of excess CSP to the culture medium provoked instantaneous competence development of the whole population (Fig 4). It proceeded at a higher rate (0.27 RLU.OD−1.min−1) than when occurring spontaneously (0.07 RLU.OD−1.min−1) and at an even higher rate in cultures grown from the highest cell density inoculum (Fig 4, 40: wt) (0.19 RLU.OD−1.min−1). This result clearly showed that CSP concentration is limiting in the low density inoculum culture (Fig 4, 1: wt), while all cells are responsive to CSP. Furthermore, addition of CSP even earlier led to the same result (S3 Fig), showing that all cells are able to respond to CSP at any time during exponential growth. Thus, in addition to the GTD emergence of competence initiator cells, CSP availability is a key factor in the rate of competence development at the cell population level. We next asked how the CSP signal is transmitted from initiator cells to the responders.
Purification of CSP from a CP1250 lineage culture supernatant [15] had suggested that cells communicate through free diffusion of CSP in the medium. However, in the first attempt to isolate pneumococcal CSP, its activity was obtained not from the culture supernatant but from pelleted heat-killed R6 cells [40], implying that most CSP remained bound to the cell envelope. We re-investigated CSP distribution during spontaneous competence development in the four above-mentioned lineages. Media were inoculated with naive cells at high density to maximize synchronization of competence development, and cells were collected by centrifugation at the time of maximum competence. The supernatants and heat-treated cell pellets were assayed for the presence of CSP, using the luciferase activity of a comA- strain harboring the ssbB::luc transcriptional fusion as a reporter of ability to induce competence (Fig 5A). Synthetic CSP was used as a standard for these quantifications. In all cases, CSP was found in the cell pellet, and at similar concentrations (Fig 5B). CSP was detected in the supernatant only in the case of the CP1250 strain, whose total CSP production was 2–5-fold higher than that of the other strains. Thus, apart from strain CP1250, pneumococcal cells retain CSP on their surface without releasing it efficiently into the medium during competence development. This feature strongly suggests a basis for the difference in competence development between the CP1250 lineage and the others.
To test if cells retain CSP during competence development, we set up co-culture experiments with strains separated by a porous membrane with a 50 kD cut off (see M&M). The porosity of a 50 kD membrane should allow diffusion of the 2.2 kD CSP without permitting cell contact between the two compartments. First, we validated the ability of CSP to diffuse through the membrane. To this end, we inoculated the comA− mutants reporting competence development in both compartments at the same cell density and added CSP after 1 hour of growth in the OUT compartment (Fig 5C, left panel, comA-, luc OUT). Both cell populations (comA- luc OUT; comA- luc IN) were found to develop competence concomitantly (Fig 5C, left panel), demonstrating that the membrane does not block CSP diffusion. Next, we performed the same experiment by inoculating the OUT compartment with wild-type cells and the IN compartment with comA- cells, with both strains harboring the luc reporter (Fig 5C middle panel, wt, luc OUT; comA−, luc IN). In these conditions, wild type cells develop competence naturally but comA- cells do not develop competence concomitantly, suggesting that CSP has been retained by the wild type cells. Finally, we repeated the experiment by mixing a wild type strain that does not report competence by luciferase expression with the comA− luc strain in the same compartment, and compared competence induction with the experiment performed with these two strains cultivated into the two compartments device (Fig 5C, right panel). Competence development of comA- cells induced by wild-type cells was observed only in the mixed cells experiment. Altogether, these experiments show that CSP is preferentially retained by the producing cells and that its propagation throughout the population is favored by cell contact.
Retention of CSP by the cell indicates that its transmission could occur by random collision. At high cell density, competence propagation would thus be favored leading to a high rate of competence development. We suggest that a subpopulation of cells present at the inoculum of the culture is at the origin of competence initiation and propagation through the population (Figs 2 and 3B). We tested this prediction by measuring competence development in wild type R800 carrying the ssbB::luc fusion gene (competence reporter cells) co-cultured with an excess of isogenic non-reporter cells.
The wild-type reporter strain inoculated with a 30 fold excess of the wild-type non-reporter strain (1:wt, luc + 29:wt) developed competence at a 3-fold higher rate than when grown alone at the same low density (1:wt, luc), i.e. the competence development rate rose from 0.048 to 0.148 RLU.OD−1.min−1 (Fig 6A). This higher rate was equivalent to the rate measured with the wild-type reporter strain inoculated at high density (30: wt, luc; 0.124 RLU.OD−1.min−1; Fig 6A). Such a result was expected since, the reporter gene apart, all cells in the culture are of identical genotype. Both populations contribute to appearance of the cell fraction initiating competence and have the same CSP production and sensing capability. In the mixed culture, the wild type non reporting cells act as helper cells facilitating both production and transmission of CSP between reporter cells. However, the equivalent mixed culture experiment performed with a comA− non-reporter strain (1: wt, luc + 29: comA-, Fig 6A), prevented competence development of the wild-type reporter strain. The comA− non reporter cells, which are responsive to CSP but unable to export it, act as cheater cells by blocking CSP transmission. Because the comA− cells heavily outnumber the wild type reporter cells, most of reporter cell collisions are with comA− cells. Consequently, comA− non reporter cells reduce the frequency of wild type reporter cell collision thereby diminishing competence propagation among this minority. The comA− cells can be fully induced to competence if sufficient CSP is available (S4 Fig). Therefore, comA- reporter cells present in excess with wild type non reporter cells should switch to competence, but CSP transmission will occur only upon contact with the initiating cells. As shown in Fig 6B, in the same mixed population but with only the comA- cell reporting competence (1: wt + 29: comA, luc), a 30 minute delay in competence shift and a damping of competence development were observed. Details of CSP production by the wild type cells are presented in S4 Fig.
These results strongly support the notion that competence initiation relies on a distinct cell fraction. Moreover, when CSP is transmitted to a comA- receiver cell, it appears to be no longer accessible to the quiescent and responsive neighboring wild type cells.
Previous experiments showed that comA- cells, while reactive to CSP and to wild-type competence initiator cells, captured CSP within the population and impeded the propagation of competence. ComD is the transmembrane sensor which is presumed to transmit the CSP signal by phosphorylating its ComE partner [27,45,46]. To determine whether ComD is involved in CSP capture and retention, we used mixed culture experiments to examine how non-reporter comD- cells, which are unable to react to CSP, affect competence development of wild-type reporter cells. If ComD is the major CSP captor, then in the mixed population the comD- cell would act as deaf-mute cells since the core sensor is inactivated. The competence development rate of the mixed culture (1: wt, luc + 29: comD; Fig 6C) was found to be slightly reduced (about 1, 5 fold) in comparison to that of the pure wild type reporter culture (1: wt, luc; Fig 6C). This result sharply contrasts with the inhibitory effect of the comA- non-reporter strain on competence development of the wild-type strain (Fig 6A). Thus ComD appears to be a central element in the capture and retention of CSP by receiver cells. To test whether ComD itself or another protein of the competence regulon under its control (through ComE activation) is responsible for CSP retention, we repeated the mixed culture experiment using a comED58E strain, which renders the competence regulon constitutive [27] but containing comC- comD- mutations that prevent CSP export. The competence development rate of the mixed culture was similar to that obtained with the comD− strain. Therefore, it is unlikely that a product of the competence regulon other than ComD contributes significantly to CSP retention on receiver cells. Unproductive cell collision with the large excess of comD- cells may passively contribute to reduction of the competence development rate of the wild type reporter strain, thus producing a buffering effect of about 1.5 fold.
The four pneumococcal lineages analyzed display distinct pre-competent time periods when grown under identical conditions (Fig 3A). Moreover, the pre-competent time period of CP1250 varies as a function of the density of the inoculum, in contrast to the three other lineages for which the pre-competence time period is nearly constant (Fig 3A). This variation indicates that a property of the mechanism driving competence development is modified in CP1250. This modification concerns the pre-competence period, during which a fraction of cells has switched stochastically to autocrine activation of CSP overexpression. We propose that both the number of stress-induced cells at competence shift and the cell density in the culture determine the shift and the rate of propagation of competence throughout the population. By diluting the inoculum of CP1250, the threshold value of one or both of these two parameters is not reached at the same pre-competence time period of the previous and denser inoculum, and competence propagation thus does not occur at this time. Further exponential growth will generate new stress-induced competent cells, which could attain the two threshold values needed for triggering competence propagation later during the culture, giving a larger pre-competence time period. This also means that the distinct categories of cells that will switch individually during the pre-competence period are already present in the initial inoculum. This scenario is strongly supported by the experiment presented in the S5A Fig, where the range of cell density at the inoculum was extended to higher cell density than previously used in Fig 2. The CP1250 lineage presents a constant pre-competence time period for the inoculum above OD 0.01, while after this point, progressively higher pre-competence time periods are observed as the density of inoculums decreases. Thus, the pre-competence period is defined by the amount and concentration of stress-induced cells that have switched into CSP overexpression by an autocrine mode leading to the competent shift. The CP1250 lineage is thus convenient to investigate this hypothesis. Indeed, the use of the luc reporter gene remains in the range of sensitivity of the luminometer detector since the constant pre-competent time period is lost below 0.01 OD. We repeated the assays below and above the threshold cell density several time each to collect several pre-competence time periods. For the 12 assays performed at high cell density, we observed a reproducible pre-competence time period with a mean value of 60 minutes of growth ranging from time 59 to 64 minutes (S5B Fig, top panel). But for the 36 assays conducted with a 3 fold lower cell density, the pre-competence time periods ranged from 64 minutes to 91 minutes of growth with 27 cultures developing competence in a narrower window of time from 76 minutes and 84 minutes (S5B Fig, bottom panel).
These results support the existence of a category of cell present at the inoculum determining the pre-competence time period, which depends on the number and proportion of initiating cells at the higher cell density. By multiplying the number of assays at lower cell density inoculum, we have conserved the number of initiating cells but these are randomly scattered in the different assays which in most cases results in random loss of either the number or the ratio of these cells required maintaining the pre-competence time period of the higher cell inoculum. But some assays at lower cell density have maintained both parameters by random distribution to initiate competence with a pre-competence time period comparable to the high cell density assays (S5B Fig). The large majority of the assays at low cell density inoculum have a significant increase of the pre-competence time period spanning a wide range of time (27 minutes). This wide range corresponds to the time necessary to produce the missing number and proportion of cells able to initiate competence in an autocrine mode. This hypothesis also explains the results observe in Figs 2A and 3A with a similar pre-competent time period for two different cell densities of the CP1250 lineage below the threshold cell density.
This study provides evidence that population-wide pneumococcal competence development in liquid cultures is a two-step process: a fraction of the cells initially switches to CSP over-expression; these cells then induce competence in the whole population. We propose that this second step occurs by transmission of CSP between cells via random cell-to-cell collision. The competence shift point (Fig 1A) marking the boundary between these two steps generally occurs after a constant time of growth (termed XA; Fig 7), whatever the density of the culture inoculum (Figs 2 and 3A). The switching of a fraction of cells to competence during the XA period (Figs 1A and 7) results from auto-activation of the CSP-based positive feedback loop in those cells. CSP would then decorate the producing cells, potentially rendering it accessible to neighboring quiescent cells by random collision (Fig 1C). Conversely, the propagation of competence throughout the population lasts a longer period of time (termed XB; Fig 7) that is inversely proportional to the cell density of the inoculum (Fig 3B). It relies on a paracrine CSP transmission mode mediated by direct contact between the donor and receiver cells, as CSP does not generally diffuse easily in the medium during competence development (Figs 5 and 6). This cell-contact sensing model is clearly different from known models of sensing [6–8]. The ComAB ABC transporter delivers CSP outside of the cell, but nothing is known about how CSP reaches its ComD target. In light of this present study supporting the notion that the CSP is transmitted by cell contact, it becomes important to understand how such a transmission operates. It seems that, once bound to ComD, the CSP is not released during the competence development period (Fig 6). This suggests that only CSP freshly exported by ComAB is available for a ComD unbound by CSP located either on the producer cells or on the receiver cells.
Ours results are consistent with the idea that spontaneous competence development at the population level is a suite of events that begin by the autocrine competence activation of some individuals that may propagate competence to the neighborhood. In this model, the core sensor ComABCDE module is central to both steps of pneumococcal competence development. During the XA period, this module registers stresses at individual cell level. We propose that core sensor of each cell is either in an idle or activating mode (Fig 7). According to such a classical bistability model [47], a cell sub-population first responds most acutely to stimuli, and second compels the other cells to follow this change in order to drive a new genetic program (Figs 2A–2C and 3). Applied to pneumococcal cultures, competence is an answer to stresses, acting as an alarm by physically contacting the closest cells to spread competence during the XB period (Fig 7, S6 Fig). For several organisms, the role of stress-induced “noise” in initiation of different gene expression routes in a clonal population is considered to be the promotion of varied physiological states [47–51]. Competence may provide a way for a cell population to survive environmental threats such as antibiotics or the host immune system by acquisition of new genetic traits via transformation [24,25]. However, the benefits of a switch to competence could be offset by the possible incorporation of deleterious genes or mutations by transformation and also by the physiological change of the cell that accompanies competence, so that in order to realise the benefit it has to be shut down again quickly. It is then not surprising that competence induction is controlled according to growth conditions [16,38,39] (S2 Fig) and varies also as a function of genotype (Fig 3A). The XA value of the pre-competence period corresponds to the time taken to accumulate physiological signals that eventually push the comABCDE loop and hence CSP production over a threshold value. Thus, ComABCDE is a multi-sensor that can be tuned up or down by integrating stress signals (Figs 1 and 7). Each step in ComABCDE core sensor influences the idling balance that determines the ComE/ComE~P ratio [16]. Any factor affecting the ComABCDE core by positive input, such as initial alkaline pH of the medium, antibiotics or DNA damaging agents, may stochastically favor a shift to competence initiation [11,37]. Some signals, such as misfolded proteins, stalled replication forks or proteins like the serine threonine kinase StkP, act to stimulate the core sensor loop [52–55], while the two component regulator CiaRH represses the core sensor by its control of csRNA and HtrA proteinase expression [28,55–57]. Variability of the XA period presumably reflects response to these multiple inputs. The sensitivity of the core sensor to external and internal signals can be modulated, by many direct or indirect effects (see also S1 Text). Some responses could amount to irreversible physiological changes that commit a cell to competence well before core sensor activation boosts the CSP level (S2 Fig). Competence bistability model should be tested in further experiments. Focus on the spontaneous competence development at single cell level should be an interesting perspective to explore the stochasticity of the autocrine competence development.
We have demonstrated here that neo-synthesized CSP is retained on competent cells of four distinct pneumococcal lineages. This is a general feature of pneumococci and is central to the concept of a cell-contact sensing mechanism. The steepness of the competence development curves is directly proportional to the cell density at competence shift, as expected if determined by the probability of cell collision (Figs 2 and 7).
A notable difference between the CP1250 lineage and the three others analyzed is the presence of a large amount of CSP in the CP1250 culture supernatant at the XB period. Nevertheless, competent CP1250 cells retain one fifth of the CSP they produce, amounting to a number of CSP molecules similar to those produced by the three other lineages. CP1250 is known to express the early comCDE and comX competence genes for longer than R800 [27]. This may lead to greater CSP production before competence shut-off, and explain the release of excess CSP into the medium. But even with this large excess of CSP in the medium, competence coordination in CP1250 cultures does not match that predicted by the QS model (Fig 2B and 2C). This suggests that the cell-contact sensing mechanism plays an important and dominant role even in this lineage.
CSP retention on the cell surface (Fig 5) could be explained in part by its ability to adopt an amphiphilic helical configuration [58]. Upon export via ComAB, CSP has a hydrophobic face and could be embedded in the lipid membrane. Membrane attachment of CSP is consistent with the observation that the membrane-anchored protease, HtrA, modulates its abundance by direct degradation [57]. In addition, our results suggest that CSP is not released from comA- mutant cells and is retained by ComD (Fig 6). The relatively large number of ComD molecules present before competence induction, 1 500 per cell [27], would favor interception of CSP. In addition, ComD reaches 39 000 molecules per competent cell, which would therefore contribute to avoid free CSP diffusion. Cell contact appears central for CSP transmission throughout the cell population cultivated under planktonic conditions. Nothing is known about how this event occurs. One possibility would be through the co-aggregation of cells during the growth. It has been shown under artificial acidic conditions of growth that competent cells present a clumping capacity with non competent cells [59,60], but this study described late events in competence development involving fratricide and DNA release from the lysed cells. The pneumococcus usually lives in a biofilm in the human nasopharynx, a mode of growth resulting in a higher transformation efficiency than during septic infection [61]. CSP retention on cells and transfer by cell contact in a biofilm would favor competence propagation to the close neighbors of the induced cells and thus to their clonal siblings.
Interaction between ComD and CSP is central for physical retention of CSP on producing cells, preventing free CSP diffusion in the medium during the pre-competence and competence development periods (Figs 5 and 6C). The other CSP fraction available is the one freshly released by the ComAB transporter that is not yet captured by the resident ComD and may promote competence propagation by allowing quiescent contacting cells to acquire CSP. ComD-CSP interaction mediates ComE-dependent activation of competence genes needed for genetic transformation and fratricide [17]. An analogous case is that of the CbpD fratricin effector, which remains anchored to the teichoic acid of the cell wall and whose lytic action against neighbors is mediated by cell contact [62]. This emphasizes the importance of a basal level of comD expression in non-competent cells. It would allow non-competent cells to survive fratricide by developing competence upon CSP transmission via cell-contact and by overexpressing the immunity factor [63–66].
Conversely, this will eliminate cells that no longer express ComD. In addition, pneumococci and close relatives have evolved several specific pairs of CSP-ComD alleles [67,68]. Therefore, competence development in the initiator fraction of pneumococcal cells would also promote lysis of cells with a different CSP-ComD pair during the propagation step (S6 Fig). Such killing of siblings during competence might provide a source of exogenous DNA for the transformation process [17], which could favor genome plasticity and/or repair [13]. In addition, propagation of competence from cell to cell provides an efficient way to maintain an active CSP-based positive feedback loop in all cells of the progeny.
Within the bacterial kingdom, the ability to convert an entire population to the competent state appears to be restricted to pneumococci and close relatives. This may provide particular properties to the whole population, such as the consequences of fratricide (as discussed above) or other properties provided by the many competence genes of unknown function. The cell-contact sensing mechanism driving this collective behavior is novel amongst the AI-based sensing mechanisms characterized so far. Some of its key features apply also to spontaneous competence development in Streptococcus thermophilus. Like pneumococcal competence, it relies on an exported peptide, named ComS, and is induced in a GTD manner independently of cell density [41]. ComS appears to be retained on the cell surface but, by contrast with the CSP, is re-imported into the cell where it mediates transcriptional activation of competence genes. Interestingly, ComS might also be sensed by neighboring cells. This indicates that competence of S. thermophilus could also propagate by cell-contact sensing as revealed here for pneumococcal competence. Thus, transmission of AI signals by cell-cell contact may be far more general than the population-wide propagation of pneumococcal competence revealed by our results. In their natural environment, pneumococci can be found growing in biofilms or dispersed in liquid. These two distinct lifestyles may determine how competence is propagated in different niches during colonization and infection. Understanding the differences in competence transmission in these different modes of growth, and whether the cell-contact sensing mechanism is used in both cases, should provide insight into the importance of competence to pneumococci in different niches.
Streptococcus pneumoniae strains are described in S1 Table. To observe spontaneous competence development the following procedure is conduced. Except where noted, stock cultures were grown at 37°C to 0.2 OD 550 nm in competence non-permissive medium Casamino Acid Tryptone (CAT) adjusted with HCl to pH 6.8. The cells were washed by centrifugation, suspended at 0.4 OD 550nm in C+Y medium [69] containing 15% glycerol and frozen at -80°C. Inoculation of C +Y medium pH 7.9 (permissive medium) with these stock cells allows spontaneous competence development. To monitor competence, all strains contained the transcriptional fusions with the luc firefly luciferase gene under the control of a competence regulated promoter (S1 Table). Cultures were started by diluting various volumes of stocked cells and a 300μL volume of C+Y medium with luciferin, with the inoculum used for monitoring luc expression in clear bottomed wells of a 96-well white NBS micro plate (Corning) [26]. Relative luminescence units (RLU) and OD values were recorded throughout incubation at 37°C in a LucyI (Anthos) or a Varioskan Flash (Thermo 399 Electron Corporation) luminometer. All experiments were repeated at least 3 times. To detect the first round of competence development in the population, all assays, especially for low cell density inoculate, were designated to record RLU values from the cell population above the threshold photon detection [26]. As the sensitivity of the Varioskan Flash and LucyI are limited in OD detection for very low inoculate, a mathematical calculation to obtain the approximate OD value under the threshold detection was carried out, using the OD know by the dilution of the pre-cultured cell at time zero and the values detected above the threshold sensitivity of the Varioskan Flash and LucyI to extract the exponential parameters of the growth culture. Calculation of the competence shift time and competence development rate: the X coordinate values corresponding to competence shift time have been estimated as followed. Exponential regression was calculated by extracting at least 5 consecutive measurements in the curve portion of the competence development phase for each competence development portion curve when reporting RLU.OD−1 (y ordered) expressed against time (x-axis). However exponential regression was calculated by extracting 3 consecutives measurements for two experiments in Fig 4 since RLU and OD was recorded every 6 min. The correlation coefficient R2 for each exponential regression calculation was found to be between 0.91 and 0.99. The competence development rate corresponds to the r parameter extracted from the exponential regression equation a*exp(r*t). Each exponential regression allows calculation of an X value (time value for competence shift) which corresponds to the intersection between regression function of competence development rate with the mean value before competence shift.
CSP-induced transformation was performed as described previously [70] using pre-competent cells treated at 37°C for 10 min with synthetic CSP1 (100 ng ml-1). After addition of transforming DNA, cells were incubated for 20 min at 30°C. Transformants were selected by plating on CAT-agar supplemented with 4% horse blood (10ml), incubating for 2 hours at 37°C for phenotypic expression, and overlayed with 10 ml CAT-agar containing the appropriate antibiotic as followed: chloramphenicol (4.5 mg ml−1), kanamycin (250 mg ml−1).
Two pre-cultured strains (chosen between R895, R1313 and R800) grown in non-permissive medium (see above) were inoculated at OD 0.0067 in C+Y pH 7.9 permissive medium with the same cell density in two compartment separated by a porous membrane with a molecular cut off of 50 kD (Float-A-Lyser; Spectrumlabs). The outside and the inside compartment had a final volume of 15 ml and 5 ml respectively which allows a higher sensitivity of the inner compartment to molecule diffusion from the outer compartment. As a result, the cells are at the same density in both compartments but with a final ratio of 3 to 1. Measurements of competence were conducted in the Varioskan Flash luminometer by taking 100μl aliquots from each compartment every 6 minutes and adding these to microplates containing luciferin for RLU reading. The experiment with CSP addition was realized to reach 100ng.ml−1 in the outer compartment.
|
10.1371/journal.pgen.1002230 | Genomic Analysis of the Necrotrophic Fungal Pathogens Sclerotinia sclerotiorum and Botrytis cinerea | Sclerotinia sclerotiorum and Botrytis cinerea are closely related necrotrophic plant pathogenic fungi notable for their wide host ranges and environmental persistence. These attributes have made these species models for understanding the complexity of necrotrophic, broad host-range pathogenicity. Despite their similarities, the two species differ in mating behaviour and the ability to produce asexual spores. We have sequenced the genomes of one strain of S. sclerotiorum and two strains of B. cinerea. The comparative analysis of these genomes relative to one another and to other sequenced fungal genomes is provided here. Their 38–39 Mb genomes include 11,860–14,270 predicted genes, which share 83% amino acid identity on average between the two species. We have mapped the S. sclerotiorum assembly to 16 chromosomes and found large-scale co-linearity with the B. cinerea genomes. Seven percent of the S. sclerotiorum genome comprises transposable elements compared to <1% of B. cinerea. The arsenal of genes associated with necrotrophic processes is similar between the species, including genes involved in plant cell wall degradation and oxalic acid production. Analysis of secondary metabolism gene clusters revealed an expansion in number and diversity of B. cinerea–specific secondary metabolites relative to S. sclerotiorum. The potential diversity in secondary metabolism might be involved in adaptation to specific ecological niches. Comparative genome analysis revealed the basis of differing sexual mating compatibility systems between S. sclerotiorum and B. cinerea. The organization of the mating-type loci differs, and their structures provide evidence for the evolution of heterothallism from homothallism. These data shed light on the evolutionary and mechanistic bases of the genetically complex traits of necrotrophic pathogenicity and sexual mating. This resource should facilitate the functional studies designed to better understand what makes these fungi such successful and persistent pathogens of agronomic crops.
| Sclerotinia sclerotiorum and Botrytis cinerea are notorious plant pathogenic fungi with very wide host ranges. They cause vast economic damage during crop cultivation as well as in harvested produce. These fungi are typical examples of necrotrophs: they first kill host plant cells and then colonize the dead tissue. The genome sequences of the two fungi were determined in order to examine commonalities in structure and content and in order to find unique features that may distinguish them from other pathogenic fungi and from saprotrophic fungi. The genomes show high sequence identity and a similar arrangement of genes. S. sclerotiorum and B. cinerea differ in their regulation of sexual reproduction, and the genetic basis and its evolution could be explained from the genome sequence. The genome sequence revealed a striking difference in the number and diversity of secondary metabolism gene clusters, which may be involved in the adaptation to different ecological niches. Altogether, there were no unique features in the genomes of S. sclerotiorum and B. cinerea that could be identified as “silver bullets,” which distinguish these aggressive pathogens from other pathogenic and non-pathogenic fungi. These findings reinforce the quantitative, multigenic nature of necrotrophic pathogenesis.
| Phytopathogenic fungi have evolved a wide range of strategies to infect and colonize plants through both convergent and divergent adaptations. This is reflected in the occurrence of species within common evolutionary branches with widely diverse pathogenic lifestyles, ranging from obligate biotrophs to necrotrophs, and from host-specific to broad host range pathogens. Operationally, necrotrophs have been defined as pathogens that derive nutrients from killed host cells, biotrophs as pathogens that derive nutrients from living tissues and hemibiotrophs as pathogens that derive nutrients from a combination of feeding from living and killed host cells, respectively. The mechanisms that drive these adaptations remain largely enigmatic.
Among the few pathogens considered to be exemplary necrotrophs are the white mold fungus Sclerotinia sclerotiorum (Lib.) de Bary and the taxonomically closely related grey mold fungus Botrytis cinerea Pers. Fr. [teleomorph Botryotinia fuckeliana (de Bary) Whetzel]. Both fungi have considerably broader host ranges (>400 and >200 species, respectively) than most plant pathogens and each causes multi-millions of US dollars in pre- and postharvest crop losses world wide [1], [2]. Necrotrophs secrete an array of cell wall-degrading enzymes and toxins, which led to their reputation as relatively less adapted as compared to biotrophic fungi, which manipulate host physiology to obtain their nutrients from living tissues. Biotrophs are widely accepted to intimately interact and co-evolve with their hosts. Recent studies have, however, revealed that interactions between necrotrophs and their host plants are considerably more complex and subtle than previously appreciated. Some necrotrophs secrete effector proteins which are internalised by host cells and interact with the host in a gene-for-gene relationship to initiate disease, albeit in an inverse manner compared to biotrophs [3]. In the case of S. sclerotiorum and B. cinerea, the active modulation of the host redox status and the subversion of host (programmed) cell death pathways by the pathogen appear to be crucial for disease to develop [4]–[8]. The availability of molecular tools has considerably advanced our understanding of the infection strategies and pathogenic development of S. sclerotiorum and B. cinerea, yet only very few absolutely critical virulence determinants have been identified by candidate gene approaches [9].
Their ability to infect different plant species and tissues under a wide range of environmental conditions, as well as their ability to produce sclerotia that survive in the soil for many years, contribute to the persistent and widespread nature of these pathogens (Figure 1). The melanized sclerotium plays a central role in the lifecycle of both fungi by germinating either vegetatively for local colonization or carpogenically to initiate the sexual cycle including the production of apothecia from which ascospores are released (Figure 1). Although S. sclerotiorum and B. cinerea share many developmental and physiological features, important differences exist in their regulation and potential for sporulation. Dispersal of both species is via air-borne spores. S. sclerotiorum exclusively produces ascospores and not conidia (asexual spores). In contrast, B. cinerea, although capable of producing ascospores, is dispersed predominantly via conidia. Furthermore the regulation of sexual sporulation is quite different, S. sclerotiorum being homothallic (self–fertile) [1] and B. cinerea heterothallic (requiring a sexual partner of opposite mating type) [2]. These differences in mitotic and meiotic sporulation impact not only the life histories of these fungi but also their epidemiology and the disease control methods employed towards each.
The characteristics of S. sclerotiorum and B. cinerea pathogenicity and development stand in stark contrast to their fellow Leotiomycete powdery mildew fungi (e.g. Blumeria, Erisyphe, Podosphaera) which are obligate biotrophs often restricted at the species level to a single host genus. The recent description of genome sequences of two powdery mildew species [10] and two phylogenetically distant, restricted host range necrotrophs (Phaeosphaeria nodorum [11], Pyrenophora teres f. teres [12]) provides the opportunity to assess whether genomic features can be identified that are common to broad host range necrotrophs such as B.cinerea and S. sclerotiorum, yet distinct from other plant pathogenic fungi. Here we describe and compare the genome sequence assemblies and annotations for S. sclerotiorum and for two strains of B. cinerea. The comparative genome analyses of these two phytopathogenic fungi to each other, to a closely related powdery mildew and to distantly related necrotrophs offer insight into common genes underlying development and pathogenesis in S. sclerotiorum and B. cinerea, as well as genes that condition specific features of their pathogenic success.
S. sclerotiorum and B. cinerea are now the only fully sequenced species in the order Helotiales and with the obligate biotroph, Blumeria graminis, in the class Leotiomycetes of the Pezizomycotina, the largest subphylum of Ascomycota [10], [13], [14]. Within the Pezizomycotina, Leotiomycetes are most closely related to the sister lineage Sordariomycetes, and more distantly to the Eurofotiomycetes and the Dothideomycetes [14]–[16]. A phylogeny based on 82 completed fungal genomes anchors a well-supported and highly divergent Helotiales lineage including S. sclerotiorum and B. cinerea [17]. The order is, however, far too large and heterogeneous to be characterized by S. sclerotiorum and B. cinerea alone. Additional species are needed to increase the phylogenetic resolution.
We constructed a five-locus phylogeny rooted with Blumeria graminis, that includes two loci not previously used for this taxon sample, G3PDH and HSP60. This analysis confirms that the Sclerotiniaceae, Sclerotinia and Botrytis are closely related but distinct, monophyletic evolutionary lineages (Figure 2). This analysis also confirms that “Sclerotinia” homoeocarpa, an important pathogen of turf with morphology and etiology quite distinct from that of Sclerotinia, is not a Sclerotinia and should be reassigned to a genus in the family Rutstroemiaceae pending a reassessment of related species and generic limits. The Sclerotiniaceae includes obligate and facultative biotrophs, such as Myriosclerotinia species, as well as necrotrophs, as exemplified by Sclerotinia and Botrytis. Botrytis is divided in two sub-lineages as previously described [18]; one lineage is associated with both eudicots and monocots and the other with eudicots only. The strongly supported lineage with species of Sclerotinia on one branch, also includes the asexual Sclerotium cepivorum, an important pathogen of Allium, and a representative of the genus Dumontinia associated with wild plants such as Anemone (Ranunculaceae). Wang et al. [13] suggest that the ancestors of the lineages representing the Sclerotiniaceae and Rutstroemiaceae were associated with conifers, inferring a radiation of the Sclerotiniaceae and Rutstroemiaceae in association with the emergence and diversification of angiosperms. Co-evolution of Botrytis with host species has been investigated but evidence is inconclusive [18]; evidence would be concordant phylogenies between symbiont/pathogen species and host species, as demonstrated in Monilinia [19]. Estimates of divergence times in the phylogeny would require a molecular clock model that could be violated if some lineages have undergone accelerated evolution, as in a radiation event. Such estimates are inexact, especially when not calibrated, e.g., by fossil evidence.
Comparative genomics of S. sclerotiorum and B. cinerea has revealed an expected high degree of sequence identity and synteny, however, we observed several striking differences in gene content between these plant pathogens. The first difference was in the content of transposable elements. The data suggest a recent burst of transposition in the S. sclerotiorum genome relative to B. cinerea. The two species differ in the regulation of sexual reproduction and this is fully explained from the sequence and organization of the MAT loci. Features of these loci provide evidence for an evolution of heterothallism from homothallism. The failure of S. sclerotiorum to produce conidiospores could not be explained from the gene content, as all key conidiation genes seemed to be present and potentially functional. The genome sequences further revealed a striking difference in the amount and types of potential secondary metabolites. The botrydial toxin biosynthetic cluster and the phytohormone abscisic acid biosynthetic genes were unique to B. cinerea. The precise chemical nature and biological function of most of these metabolites remains to be determined to understand whether they contribute to the adaptation of the two species to different ecological niches. The content of CAZyme-encoding genes revealed a preference of S. sclerotiorum and B. cinerea for pectin as nutrient source. This is in agreement with the observation that S. sclerotiorum and B. cinerea are pathogens of dicots and preferentially grow on aerial plant tissues that are rich in pectin. With respect to pathogenicity determinants, there are no unique features that could be identified as ‘silver bullets’, which distinguish these aggressive pathogens from other pathogenic and non-pathogenic fungi. Comparison with two other necrotrophic fungi revealed only few (if any) genes shared among necrotrophs and absent in fungi with other trophic lifestyles (i.e saprotrophs or hemi-biotrophs). These findings point to the multigenic, variable, and sophisticated nature of necrotrophic pathogenicity whose nuance may only be revealed through systematic functional analysis of candidate activities and regulators. We suggest that the specific regulation of networks of the available suite of genes is key to pathogenic success for S. sclerotiorum and B. cinerea.
The sequenced strain of S. sclerotiorum ‘1980’ is available from the Fungal Genetics Stock Center (http://www.fgsc.net/). Plasmid (4-kb and 10-kb inserts) and Fosmid (40-kb inserts) libraries were generated with randomly sheared and size-selected DNA. Plasmid and Fosmid inserts were sequenced from both ends to generate paired reads. Sequence reads (total of 507,621) were filtered for quality, vector and other contamination, and the resulting 476,001 reads were assembled with the Arachne assembler [20]. A consensus sequence was determined from an average of approximately 9.1-fold sequence depth (7.8 depth in Q20 bases). The assembly (GenBank accession AAGT01000000) totals 38.3 Mb and consists of 36 scaffolds with an N50 of 1.6 Mb. The scaffolds consist of 679 sequence contigs with a N50 of 123 kb; at least half of all bases in the assembly fall within at least the size of the N50 scaffold or contig. The contigs total 38.0 Mb, so only 0.8% of the assembly is represented by contig gaps. This assembly contains 98.7% of bases with Q40 quality or greater.
The sequence of B. cinerea strain B05.10 was generated by Syngenta AG using four size selected shotgun libraries. Sequence reads were filtered for quality, vector and other contamination, and the resulting 291,603 reads were assembled using the Arachne assembler. A consensus sequence was generated from an average of approximately 4.5 fold sequence depth (3.9-fold depth in Q20 bases). The assembly (GenBank accession AAID00000000) totals 42.3 Mb and consists of 4,534 contigs with a N50 of 16.4 kb, which are ordered and oriented by paired end clone reads within 588 scaffolds, with a N50 of 257 kb.
The sequence of B. cinerea strain T4 (Genbank accessions FQ790245 to FQ790362) was determined and assembled by Genoscope, Evry, France. The genome was sequenced using a whole genome shotgun (WGS) strategy, by generating plasmid (3 kb and 10 kb inserts) and BAC (∼50 kb inserts) libraries. Clones from these libraries were end-sequenced to generate paired reads. Sequence reads from plasmid inserts (573,705, representing 10.5× coverage) were assembled with Arachne [20], resulting in 3,054 contigs. Contigs were ordered and oriented into scaffolds using paired sequences from BACs (35,668 reads). The gap lengths were estimated using linked clones. 118 scaffolds larger than 20 kb with gap sizes lower than 10%, were used for annotation. These 118 scaffolds (39.5 Mb, N50 of 562 kb, Table 1) contain 2,281 contigs (37.9 Mb, N50 of 35 kb). This assembly contains 98.7% of bases with Q40 quality or greater.
An optical map, a type of physical map, was created for S. sclerotiorum strain 1980 by OpGen. The restriction enzyme BsiWI was used to create ordered fragments and the resulting map contained 16 optical contigs. These likely correspond to complete chromosomes, in agreement with an estimated chromosome number of 16 using pulsed-field gel electrophoresis [21]. The assembly was compared to the optical map using an in silico digest of the assembly (http://www.broadinstitute.org/annotation/genome/sclerotinia_sclerotiorum/MultiHome.html). This comparison suggested that supercontig 4 contained an unsupported join in the assembly, between contigs 132 and 133; the assembly was updated to break scaffold 4 between these contigs, and a revised version of the assembly was submitted to Genbank. No other major disagreements were found. A total map size of approximately 38.5 Mb was estimated by calculating the ratio between the assembly fragments and map units. The assembly supercontigs cover 99.4% of the optical map (contigs cover 98.6%) (Figure S1). Only two supercontigs, 35 and 36, were not anchored to the optical map; these supercontigs are 39 kb and 16 kb respectively, and do not contain any BsiWI restriction site. Supercontig 35 consists of arrayed ribosomal DNA repeats. The optical contigs appear to correspond to complete chromosomes, as they end with a flush set of DNA molecules (data not shown). Additionally, reads containing telomeric repeat arrays of (TTAGGG) are linked to the ends of 26 of the 32 chromosome ends, and are present at the end of two additional scaffolds (Figure S1). At least two scaffolds are mapped to each optical contig, suggesting that centromeres could lie in these unassembled gaps.
68 individual progeny from a cross between strains T4 (isolated from tomato) and 32 (isolated from Vitis vinifera) were kindly provided by Caroline Kunz (Université Pierre et Marie Curie, Paris). Around 400 microsatellite markers were designed from the genomic sequences of the B05.10 and T4 strains by using the “GRAMENE” software [133]. In addition, a set of 144 SNP (Single Nucleotide Polymorphism) markers were identified through the comparison of the B05.10 and T4 genome sequences and genotyped in the progeny using the SNPlex genotyping assay (Applied Biosystems). The segregation of markers among the progeny was analysed using MAPMAKER software [134] set at min LOD3 and max Distance at 37.
For S. sclerotiorum and B. cinerea B05.10, gene structures were predicted using a combination of FGENESH and GENEID. The version of GENEID used for these calls was 1.2a. FGENESH is unversioned. As FGENESH uses a statistical model of gene structure that requires training on each organism for accurate prediction, Softberry trained FGENESH on S. sclerotiorum sequences. GENEID is an ab initio gene caller and was run with the default parameters after being trained on a set of 542 S. sclerotiorum genes that were manually curated based on EST and protein alignments.These S. sclerotiorum trained gene models were also used for B. cinerea B05.10 without additional training.
The results from these two gene callers were combined in the following manner. Both FGENESH and GENEID were run on the entire genomic sequences of S. sclerotiorum and B. cinerea (B05.10) to provide an initial set of predicted genes. This resulted in a set containing 12,961 FGENESH predictions and 14,711 GENEID predictions for S. sclerotiorum, and 13,864 FGENESH predictions and 16,907 GENEID predictions for B. cinerea. Next, gene predictions less than 30aa (90 nt) were removed, and any gene prediction less than 50aa (150 nt) was removed, unless it was overlapped by a prediction from a different program, BLAST evidence, a HMMER PFAM domain, or an EST alignment. Applying these criteria removed 2,166 S. sclerotiorum and 2,373 B. cinerea genes from consideration. We manually annotated 1,141 S. sclerotiorum and 751 B. cinerea genes. FGENESH and GENEID predictions were clustered based on overlapping exons, requiring strand consistency. For each locus, a gene structure was chosen based on BLAST similarity to other proteins (requiring ≥60% average identity and ≥80% query coverage), and selecting the gene prediction from the program with best overall agreement to EST splice sites (GENEID performed better than FGENESH by this metric). The initial gene sets contained 14,522 genes for S. sclerotiorum and 16,448 genes for B. cinerea.
The accuracy of the gene set was evaluated by comparing to EST sequences. For S. sclerotiorum, a set of 75,468 ESTs that were sequenced as part of this project or available in Genbank, were aligned to the genome using BLAT. EST alignments were clustered by combining all overlapping ESTs. Each of the resulting 7,400 EST clusters was compared to any overlapping predicted genes. In cases where multiple overlapping ESTs suggest different gene structures, the EST that most closely matched the gene structure was used. Roughly one-third of genes (5,192) have some overlap with an EST cluster; of these, 75% show no splice site disagreements. A small number of predicted genes appear partial in S. sclerotiorum: 4 are missing a start codon, 6 are missing a stop codon, and 98 span contigs. Possible missed annotations include ESTs with no overlapping gene; a total of 559 clusters are at least 200 bases from an annotated gene, and of these 88 contain canonical splice signals. After the automated annotation was complete, ESTs from additional libraries were generated and aligned to the genome. In total, we sequenced ESTs from 8 libraries, generating 96,700 quality filtered sequences (Table S30). These ESTs align to 7,942 genes.
For B. cinerea (B05.10), a set of 9,207 ESTs were aligned to the genome using BLAT. EST alignments were clustered by combining all overlapping ESTs into a cluster. Each of the resulting 2,349 EST clusters was compared to overlapping predicted genes. In cases where multiple overlapping ESTs suggest different gene structures, the EST that most closely matched the gene structure was used. Roughly one-eighth of genes (2,012) have some overlap with an EST cluster; of these, 82% show no splice site disagreements. A small number of predicted genes appear partial in B. cinerea (B05.10): 67 are missing a start codon, 153 are missing a stop codon, and 444 span contigs. Possible missed annotations include ESTs with no overlapping gene; a total of 132 clusters are at least 200 bases from an annotated gene, and of these only 4 contain canonical splice signals.
S. sclerotiorum and B. cinerea (B05.10) gene calls were further evaluated to remove transposable elements and other dubious calls. Genes with at least 100 bases of overlap with repeats called by RepeatMasker or BLAST similarity against a transposon-related protein database, or containing a PFAM repeat domain but not a PFAM non-repeat domain, were flagged as dubious gene calls. Also, genes that have 5 or more BLAST hits at ≥95% identity to the same genome and have no supporting ESTs, BLAST hits or non-repeat PFAM domains, were flagged as dubious. To identify poorly supported gene calls, two additional gene predictors, GeneMark and SNAP, were run on both genomes. Genes less than 100 amino acids that were only supported by a single gene predictor (comparing GENEID, FGENESH, GeneMark, and SNAP) and without EST, BLAST, or PFAM domain support were flagged as dubious genes. Additionally, genes with two predictions that do not share the same reading frame were flagged. In total, all these methods flagged 2,662 S. sclerotiorum genes and 2,784 B. cinerea (B05.10) genes as dubious. All predicted genes were used to query the PFAM set of hidden Markov models using HMMER (http://hmmer.janelia.org) and the public protein databases using BLASTP. Transfer RNAs were identified using the tRNAScan-SE program.
For B. cinerea T4, the automated gene prediction was performed using the URGI genomic annotation platform including pipelines, databases and interfaces, developed for fungi (http://urgi.versailles.inra.fr/). Gene prediction was performed using Eugene pipeline version 3 [135]. The gene models predicted by EuGene rely on a combination of several in silico evidences (ab initio and similarity). Ab initio gene prediction softwares are Eugene_IMM [136] (probabilistic models discriminating coding from non coding sequences), SpliceMachine [137] (prediction of CDS start sites and intron splicing sites) and FGENESH (http://linux1.softberry.com/berry.phtml) (ab initio gene predictor). Several similarity methods were used to identify genes such as BLASTN and Sim4 against B. cinerea ESTs, as well as BLASTX against Uniprot and fungal protein databases. The different results were used by Eugene to predict final gene models. The three ab initio gene prediction softwares were trained using a set of manually annotated genes from B. cinerea. FGENESH was trained by Softberry (Boston, USA), while EuGene-IMM and SpliceMachine were trained at URGI using a set of 305 genomic/full-coding cDNA pairs. One third of the set was used for training ab initio softwares, one third to optimize the parameters of EuGene and the last third to calculate the accuracy of EuGene. We finally obtained for exons and genes, a sensitivity of 97.8 and 92.1 and a specificity of 97.5 and 92.1, respectively. Genes with a size smaller than 100 nucleotides were automatically filtered out by EuGene. EuGene finally predicted 16360 genes in the B. cinerea T4 genome. In addition, 434 tRNA genes were predicted by tRNAscan-SE [138].
Genome assemblies were aligned using MUMmer [139]. The B. cinerea genomes were aligned with nucmer, selecting only 1∶1 matches (−mum), and alignments were processed with delta-filter to select 1∶1 local mapping of the reference to query. A total of 35.5 Mb of each could be aligned, or 94% of the B. cinerea T4 genome and 91% of the B. cinerea B05.10 genome. A total of 98,744 insertion/deletion positions and 175,009 SNPs were identified from the nucmer alignments using the show-snps program. This suggests that the overall rate of difference between the two genomes in aligned regions is 1 SNP every 203 bases and 1 insertion/deletion every 360 bases. S. sclerotiorum and B. cinerea are too distant to align well at the nucleotide level globally; only a total of 7.4 Mb can be aligned at 85.0% identity. Therefore the B. cinerea genomes were aligned to S. sclerotiorum using promer, selecting only 1∶1 matches (−mum), and alignments were processed with delta-filter to select 1∶1 local mapping of the reference to query. The promer alignments cover 17.4 Mb sharing 74.3% identity for S. sclerotiorum-B. cinerea T4 and 17.2 Mb sharing 74.6% identity for S. sclerotiorum-B. cinerea B05.10.
To identify larger syntenic regions between genomes, we first identified orthologs using BLASTP as well as OrthoMCL. Using BLASTP (1e-5, softmasked), S. sclerotiorum shares 8,609 best bi-directional (bbd) BLAST hits with B. cinerea B05.10 or 8,601 with B. cinerea T4; 8,088 S. sclerotiorum proteins have matches in both genomes. About 1,300 additional proteins in S. sclerotiorum have BLAST matches that are not best bidirectional. Using OrthoMCL to cluster the three genomes, which uses BLASTP similarity for clustering, there are a total of 12,095 protein families, of which 8,079 are in all three genomes and nearly all of these (7,677) are single copy. Either raw BLAST data or ortholog sets were then used to find syntenic regions of at least four genes between genomes using DAGchainer [27], and parsed using accessory perl scripts. Syntenic regions between the two B. cinerea genomes include 11,422 paired genes and cover 36.9 Mb in 249 regions (average size 148 kb). Syntenic regions between S. sclerotiorum and B. cinerea T4 include 7,752 paired genes and cover 27.7 Mb in 571 regions (average size 49 kb). Syntenic regions between S. sclerotiorum and B. cinerea B05.10 include 7,702 paired genes and cover 27.9 Mb in 602 regions (average size 46 kb). Segmental duplication blocks were identified by comparing each genome against itself, using DAGchainer [27].
Transposable Elements (TEs) were detected and annotated using the REPET pipeline [30], [31] http://urgi.versailles.inra.fr/Tools/REPET) for detection (TEdenovo) and annotation (TEannot) of transposons in eukaryotic genomes. The TEdenovo pipeline detects TE copies, groups them into families and defines the consensus sequence for each family containing at least 3 copies. The TEannot pipeline annotates TEs using the consensus sequences library. Manual annotation of the consensus sequences allowed their grouping in TE super-families corresponding to a transposon with a defined full-length copy (with both LTRs or ITRs and full ORFs when available) and classified as classI or classII TE. In addition, the genome of B. cinerea T4 was searched using TBLASTN for sequences similar to either the transposase of B. cinerea mariner element Flipper [140] or the reverse transcriptase from Boty1 [141]. Sequences with similarities to TE-encoded proteins were grouped and aligned to identify or reconstruct the corresponding full-length TE. Manually identified TEs were compared to the consensus sequences obtained from REPET.
Automated functional annotation of S. sclerotiorum and B. cinerea proteins was performed using protein sequences deduced from all gene models automatically predicted. The pipeline uses protein domain identifier InterProScan [142] which runs a set of methods including pattern matching and motif recognition. In addition, we used an automated assignment against protein domains databases such as CDD [143], KEG [144] and KOG [145]. Sub-cellular targeting signals as well as transmembrane domains were predicted using SignalP, TargetP and TMHMM [146]. Orthologs in S. sclerotiorum were determined based on best bi-directional BLAST hits. Overall, 72% of B. cinerea T4 predicted genes (11767 of 16360) encode proteins having either functional information and/or an ortholog in S. sclerotiorum. This percentage increases to 82% taking into account only reliable genes.
Four types of evidence were used to support gene calls, based on bioinformatics or experimental criteria (Figure S3). The first criterium was based on identification of functional domains (e.g. InterProScan, CDD) or topology/targeting domains (e.g. SignalP, TargetP, TMHMM) (see M7). Genes with at least one positive hit (functional or topology/targeting domain) were considered as supported by protein domain evidence. The second criterium was based on identification of orthology to genes in other fungi using OrthoMCL (see M9). The third criterium relied on expression data using ESTs (existence of at least one EST for a gene) (see M10). The fourth criterium relied on expression data using hybridization signals on a whole-genome oligonucleotide Nimblegen microarray (see M11). Among the 16,360 predicted genes in B. cinerea T4, 13,555 (83%) genes are supported by at least 1 type of evidence and 6,462 (39.5%) genes are supported by all 4 types of evidence. Among the 16,448 predicted genes in B. cinerea B05.10, 13,922 (85%) genes are supported by at least 1 type of evidence and 6,319 (38.5%) genes are supported by all 4 types of evidence. Among the 14,522 predicted genes in S. sclerotiorum, 12,283 (85%) genes are supported by at least 1 type of evidence and 4,121 (28%) genes are supported by all 4 types of evidence.
Orthologs between S. sclerotiorum, B. cinerea, and other ascomycete fungi were identified using OrthoMCL version 1.4 [147]. The Ascomycete genomes included in the analysis were B. graminis (22_02_11; blugen.org), P. nodorum (20110506 from Richard Oliver; Broad Institute), P. teres f. teres (PRJNA50389; Genbank), G. zeae (FG3; Broad Institute), M. oryzae (MG8; Broad Institute), N. crassa (NC10; Broad Institute), and A. niger (AspGD). Each set of proteins was blasted against itself and other proteomes with an e value of 1e-5. An inflation parameter of 1.5 was used for Markov Clustering with MCL. 12,120 B. cinerea (T4) proteins, 12,260 B. cinerea (B05.10) proteins, and 9,930 S. sclerotiorum proteins were identified as members of gene families; of these about 1,500 were conserved only in the two species (Table S8). OrthoMCL families enriched in S. sclerotiorum and B. cinerea were identified based on a hypergeometric distribution (p-value computed by phyper function in R; q-value computed by p.adjust in R), however significantly enriched families (q-value<0.05) included only repetitive elements or families specific to only these species. To identify functions specific to these lineage specific proteins, we mapped GO terms to the protein sets of S. sclerotiorum and B. cinerea (B05.10) using Blast2GO, and computed enrichment using Fisher's Test exact test.
To identify functions enriched in S. sclerotinia and B. cinerea proteins, we identified PFAM domains in each of the above genomes, and computed enrichment or depletion in subsets of species. Protein domains from PFAM release 25 (ftp://ftp.sanger.ac.uk/pub/databases/Pfam) were assigned to proteins in each genome using hmmsearch from hmmer3 (http://hmmer.janelia.org/software), requiring an Evalue cutoff of 1e-5. For each genome, the total number of proteins with each type of domain was computed; a protein with multiple domains of the same type was counted only once. To identify domains enriched in subsets of genomes, significant differences were identified by computing the p-value for each domain based on a hypergeometric distribution (phyper function in R), and computing q-values to correct for multiple testing (p.adjust, fdr). Domains significantly enriched in the S. sclerotiorum and B. cinerea genomes (Table S10) were filtered to remove those found in transposable elements (PF03221, Tc5 transposase DNA binding domain; PF00078, Reverse transcriptase; PF00665, Integrase core domain; PF03184, DDE superfamily endonuclease; PF03732, Retrotransposon gag protein; PF00075, RNase H; PF05225, helix-turn-helix, Psq domain). Domains significantly depleted in the S. sclerotiorum and B. cinerea genomes (Table S11) were filtered to list those that are conserved in at least 5 species.
For S. sclerotiorum, a total of 96,700 filtered EST sequenced were generated from eight cDNA libraries, prepared from mRNA from developing sclerotia, developing apothecia following 55 h light exposure, mycelia at pH 7, infected Brassica or infected tomato, two infection cushion samples, and mycelium exposed to oxidative stress (Table S30). ESTs were aligned to the S. sclerotiorum genome using BLAST and compared to predicted gene structures.
Fourteen cDNA libraries of various B. cinerea isolates (T4, B05.10, SAS56×SAS405, ATCC 58025) and stages of development have been prepared and sequenced in various laboratories (Table S31). Some of them were publicly available [148], [149] (library BcA1 and AL11), others were sequenced in the framework of the B. cinerea sequencing project by Genoscope (http://www.genoscope.cns.fr/, libraries PD0A*) while others are private (Bayer Crop Sciences/P. Tudzynski). From these 14 cDNA libraries (Table S31), 78,755 bacterial clones were obtained and sequenced once or twice (5′ end; 3′ end), leading to 83,117 ESTs (1 or 2 ESTs per clone).
Eighty six percent of ESTs, (71,238 ESTs, 67,625 clones, Figure S10) were successfully clustered on the B. cinerea T4 genome and assembled using Phrap [http://www.phrap.org/] to finally obtain 9,667 EST contigs. As some genes were lying in different contigs, the 9,667 contigs were assembled in 9,004 unisequences. In order to get an estimate of the expression of genes corresponding to the unisequences, we calculated the raw_clone_nr sum (clone number in all libraries), the raw_clone_nr (clone number in the current library), the percent_clone (100* raw_clone_nr/clone_nr in the current library), and the norm_percent (percent_clone normalized by the percent_clones in all libraries = 100 * (percent_clone)/sum(percent_clone) in all libraries). Eighteen percent of unisequences without any corresponding genes automatically predicted (mostly due to gap in the genome sequence) were used to design oligos for the Nimblegen microarrays.
A Nimblegen 1-plex array was designed using 21,200 B. cinerea gene models corresponding to 19,454 ORFs either identified in T4 or B05.10 genomes (12,071 T4 ORFs shared between T4 and B05.10; 93 B05.10 ORFs shared between T4 and B05.10; 4,072 T4 ORFs specific to T4; 3,218 B05.10 ORFs specific to B05.10), 12 experimental genes and 1,734 additional EST unisequences. Nine probes per sequence were defined, leading to 169,347 probes representing 20,889 genes and covering 20,916 provided gene models (19,222 ORFs, 12 experimental genes and 1,682 ESTs). In addition, 16,871 random probes (negative controls) were designed. Two copies of each probe were placed on the array.
A second Nimblegen 1-plex array was designed using 15,026 S. sclerotiorum gene models corresponding to 14,522 ORFs and 504 additional EST unisequences. Thirteen probes per sequence were defined, leading to 190,130 probes representing 14,801 genes and covering 14,858 provided gene models (14,360 ORFs and 498 ESTs). In addition, 9,047 random probes (negative controls) were designed. Two copies of each probe were placed on the array.
Expression of fungal genes was studied during infection on sunflower cotyledon, and compared with in vitro expression. The experimental conditions were: (i) Infection of sunflower cotyledons by mycelial plugs of B. cinerea (B05.10) and S. sclerotiorum at 2 days after inoculation (100% of the surface area was infected), (ii) mycelial cultures grown in vitro (malt agar) for each fungal strain, and (iii) non-infected sunflower cotyledons. RNAs were extracted from 3 biological replicates, labelled and hybridized to arrays. 2 or 3 biological replicates per experimental condition were exploitable. Data were analysed using ANAIS methods [150]. Probe hybridization signals were normalized using the quantile function, and summed for each gene. Genes were considered as expressed when the signal was above the defined background (1.5 fold the 95th percentile of random probes hybridization signals), in all the biological replicates of at least one experimental condition. The identification of differentially expressed genes was performed using an ANOVA test. To deal with multiple testings, the ANOVA p-values were further submitted to Bonferroni correction. Transcripts with a corrected p-value<0.05 and more than 2.0 fold change in transcript level were considered as significantly differentially expressed. Details of experiments, raw values and lists of differentially expressed genes with associated normalized values are available at http://urgi.versailles.inra.fr/Data/Transcriptome.
Complete proteome sequences were downloaded from NCBI or JGI and subjected to Merops Batch BLAST analysis [151]. Sequences from S. sclerotiorum and B. cinerea identified by Merops as putative peptidases were scrutinised based on the presence of active site and ligand residues as well as e values (cut off 1e-5). Additional BLAST analysis was performed at NCBI in order to confirm or reject suspicious hits. We excluded from the analysis the peptidases of which the activity is restricted to an autocatalytic activation of the precursor protein. We also excluded putative peptidases with problematic function inference such as the S9 and S33 proteases. The peptidases of S. sclerotiorum and B. cinerea were subjected to SignalP analysis [152]. Analyses of the other fungal peptidases were restricted to the families of which at least one sequence was predicted to correspond to a secreted peptidase and performed as described.
The best protein models of the S. sclerotiorum and B. cinerea genomes were subject to expert analysis using the CAZy database (www.cazy.org) annotation pipeline [153]. Each model was compared by BLAST [154] to libraries of known catalytic and carbohydrate-binding modules derived from the CAZy database and from previously analyzed genomes. Each identified protein model was subject to modular analysis compring BLAST [154] and HMMer [155] analysis against CAZy-derived libraries and HMM profiles, respectively, followed by human curation. Later, the quality of each identified model was manually evaluated and an expert functional annotation was proposed by comparison against characterized enzymes from CAZy. The approaches are described in more detail elsewhere [153]. Finally, comparative analysis was performed against other fungal genomes using the same principles as described before [156].
Assemblies and annotations were submitted to GenBank under the following accession numbers: AAGT01000000 (S. sclerotiorum), AAID00000000 (B. cinerea B05.10), FQ790245 to FQ790362 (B. cinerea T4).
|
10.1371/journal.pgen.1003931 | Recombinogenic Conditions Influence Partner Choice in Spontaneous Mitotic Recombination | Mammalian common fragile sites are loci of frequent chromosome breakage and putative recombination hotspots. Here, we utilized Replication Slow Zones (RSZs), a budding yeast homolog of the mammalian common fragile sites, to examine recombination activities at these loci. We found that rates of URA3 inactivation of a hisG-URA3-hisG reporter at RSZ and non-RSZ loci were comparable under all conditions tested, including those that specifically promote chromosome breakage at RSZs (hydroxyurea [HU], mec1Δ sml1Δ, and high temperature), and those that suppress it (sml1Δ and rrm3Δ). These observations indicate that RSZs are not recombination hotspots and that chromosome fragility and recombination activity can be uncoupled. Results confirmed recombinogenic effects of HU, mec1Δ sml1Δ, and rrm3Δ and identified temperature as a regulator of mitotic recombination. We also found that these conditions altered the nature of recombination outcomes, leading to a significant increase in the frequency of URA3 inactivation via loss of heterozygosity (LOH), the type of genetic alteration involved in cancer development. Further analyses revealed that the increase was likely due to down regulation of intrachromatid and intersister (IC/IS) bias in mitotic recombination, and that RSZs exhibited greater sensitivity to HU dependent loss of IC/IS bias than non RSZ loci. These observations suggest that recombinogenic conditions contribute to genome rearrangements not only by increasing the overall recombination activity, but also by altering the nature of recombination outcomes by their effects on recombination partner choice. Similarly, fragile sites may contribute to cancer more frequently than non-fragile loci due their enhanced sensitivity to certain conditions that down-regulate the IC/IS bias rather than intrinsically higher rates of recombination.
| Chromosome rearrangements are frequently associated with human cancers. Such rearrangement can result from a DNA break followed by an erroneous repair. Mammalian common fragile sites are one of the most extensively studied naturally occurring breakage prone regions of the genome. It has been proposed that fragile sites are recombination hotspots and that increased recombination activity at these loci contribute to cancer. We examined this hypothesis using a model organism, budding yeast Saccharomyces cerevisiae, where a homolog of the mammalian common fragile sites has been identified. Unexpectedly, our results showed that the rate of recombination at the fragile sites was not any higher than non fragile sites, even under the conditions that promoted chromosome breakage at the fragile sites. However, we found that the frequency of loss of heterozygosity (LOH) and translocation, the type of recombination outcomes known to contribute to cancer, to be significantly elevated at fragile sites under certain conditions. These findings suggest that the fragile sites might indeed contribute to cancer more frequently than non-fragile loci, but the reason for this is likely to be due the nature of the recombination outcome(s) rather than higher rates of recombination.
| Accidental DNA double strand breaks (DSBs) arise during unperturbed proliferation. Such “endogenous” or “spontaneous” chromosome breakage does not occur randomly throughout the genome, but at specific loci, often referred to as fragile sites. Fragile sites have been observed in organisms ranging from bacteria to mammals, suggesting that they might be a ubiquitous feature of the genome [1], [2], [3], [4]. Evidence points to the existence of multiple types of fragile sites that are distinguishable from one another based on its structure, function, and/or genetic requirement(s) for its stability [5], [6], [7].
Mammalian fragile sites are one of the most extensively studied naturally occurring breakage prone regions of the genome. They are classified as either “rare” or “common”, depending on their incidence among general population [4]. Rare fragile sites are found in less than 5% of the population. In most cases, rare fragile sites are tri-nucleotide repeats, expansion of which has been linked to conditions such as Fragile X-syndrome and Huntington disease [8]. In contrast, common fragile sites are present in all individuals and can be found on every chromosome, indicating that they are a normal component of the chromosome. Common fragile sites extend over large regions of the genome, from several hundred kilobases (kb) to over 1 megabase (Mb) with breaks or gaps occurring throughout these regions. There is no sequence determinant that defines common fragile sites [4], [9].
Studies have implicated a link between common fragile sites, genome instability, and cancer [9]. For instance, some fragile sites have been shown to be loci of frequent chromosome deletions, translocations, and/or viral genome integration [10], [11] as well as oncogenic chromosomal rearrangements (e.g. [12]). Combining these with the observations that some fragile sites in model organisms exhibit elevated rates of recombination, it has been proposed that mammalian common fragile sites are recombination hotspots and that increased recombination activities at these loci contribute to cancer.
Replication Slow Zone (RSZ) is a type of fragile site in budding yeast and a putative homolog of the mammalian common fragile sites. It was identified as loci of preferred chromosome breakage following inactivation of Mec1, the budding yeast homolog of ATR, where high levels of replication dependent single stranded DNA (ssDNA), a precursor to DSBs, accumulate [3], [13]. Like its mammalian counterpart, RSZs are relatively large (∼10 kb), and appear to be a normal component of the chromosome. Other similarities between the two include; (i) timing of their replication during normal S-phase, which occurs late, (ii) their sensitivity to mild replication stress and inactivation of the ATR/ATM family kinases, and (iii) the lack of a defining sequence determinant(s) [3], [14], [15], [16]. The mammalian ATR/ATM and their budding yeast homologs Mec1/Tel1 are conserved signal transduction proteins best known for their roles in S-phase and DNA damage checkpoint responses [17]. In addition, they also play essential roles in a number of fundamental DNA and chromosomal processes including genome duplication, meiotic recombination, and DNA repair (e.g. [3], [18], [19]).
Here, we utilized RSZ as a model to test the proposal that mammalian common fragile sites are recombination hotspots. Unexpectedly, we found that recombination rates at RSZ and non RSZ loci were comparable under all conditions tested, indicating that RSZs are not recombination hotspots. Based on these and other observations, we propose a model whereby regulation of the nature of the recombination outcome(s), irrespective of the overall recombination activity, may play a key role in controlling genome rearrangements.
The observations that stalled- and collapsed- replication forks can promote recombination [6], [7], [25], [26] and that RSZs are loci of slowed replication fork progression during normal S-phase [3], suggest that RSZs might be recombination hotspots during normal proliferation. (NB: In the current context, an arrested fork that can ultimately resume replication without intervention is referred to as a “stalled fork” whereas that requires an active fork-restart process is referred to as a “collapsed fork”). To test this, we compared the rate of URA3 inactivation at the RSZ and non RSZ loci under a standard yeast growth condition (2% glucose at 30°C, hereupon referred to as “YPD”). For each locus, the rate was estimated by the method of the median from two independently derived strains, and the average of the two was used for the locus-to-locus comparison (Figure S2) [24].
In haploids, the locus specific recombination rates varied very little, from 1.8×10−5 per cell generation at ORI to 2.1×10−5 at RSZ1 (Figure 2Ai). In diploids, the variation was greater, about 3 fold, and ranged from 1.1×10−5 per cell generation at TER to 3.3×10−5 at NON (Figure 2Aii). Importantly, the rate of URA3 inactivation at the two RSZs was not higher than the three non RSZ loci in either haploids or diploids. We conclude that RSZs are not spontaneous recombination hotspots, defined as loci of intrinsically higher recombination activities.
Chromosome breakage at RSZs does not occur during normal proliferation, but is promoted by a modest level of HU, high temperature, and/or inactivation of Mec1 [3], [21]. Thus, it is possible that the putative recombination hotspot activity associated with RSZs might also require these conditions. To test this, we examined the effect of HU. The same ten haploid and ten diploid strains analyzed above were subjected to a transient (18 hour) exposure to 10 mM HU before 5FOA selection. We found that the exposure lead to a statistically significant increase in rate of URA3 inactivation at every locus (Figure 2D). Importantly, the rates at the two RSZs were not any higher than the three non RSZ loci in either haploids or diploids (Figure 2B). We conclude that RSZs are not recombination hotspots even under a condition that promotes RSZ specific chromosome breakage.
Next, we examined the nature of genetic alternations associated with URA3 inactivation in diploids. To this end, we utilized Southern Blot analysis that enabled us to monitor the presence of the following three alleles (Figure 3A, Figure S3); (i) WT, (ii) the allele containing the hisG-URA3-hisG reporter integrated, hereupon referred to as “INT” for “integrant”, and (iii) the “pop-out” or “PO” allele, indicative of an intrachromatid or intersister (IC/IS) recombination or recombination related event (Figure 1C). As expected, the parent ura3 diploid strain exhibited a single band diagnostic of the WT allele at each of the five loci examined; in contrast, each of the URA3 heterozygotes derived from the parent exhibited an additional band(s), corresponding to the INT allele (Figure 3B, Figure S3).
We found that all 5FOAR colonies examined had lost the INT fragment, suggesting that URA3 inactivation in every case involved a relatively large structural change(s) at the integration locus (Figure 3C; data not shown). Overall, ∼80% (117/146) of the samples exhibited the diagnostic PO band, indicating that on average, URA3 inactivation was four times more likely to occur via an IC/IS-mediated event than all other mechanisms combined (Figure 3E “Total”). The latter is consistent with previous reports on strong IC/IS-bias in mitotic recombination (e.g. [27]). Locus specific PO fraction ranged from 70% (14/20) at TER to 100% (26/26) at RSZ2, suggesting that the extent of IC/IS bias might be influenced by local environment (Figure 3E). The average PO fraction for the two RSZs was higher than the three non-RSZ loci (90% versus 73%; Figure 3E); however the difference was not statistically significant (p = 0.1209). We conclude that recombination activity at RSZs during standard growth condition is comparable to non RSZ loci with regard to both the rate of recombination and the extent of the IC/IS bias.
The same Southern Blot analysis was performed on 5FOAR colonies that arose in the presence of HU. Similarly to the YPD samples, all 148 HU 5FOAR samples had lost the diagnostic INT fragment (Figure 3D; data not shown). Overall, 53% (79/148) of the HU 5FOAR colonies carried the PO allele, a significant reduction from the 80% (117/148) observed in YPD (Figure 3F “Total”; Chi square test, p<0.0001).
The negative effect of HU on IC/IS bias was observed at every locus; however, the only statistically significant reduction was at the two RSZs (Fisher's exact test; Figure 3F). At RSZ1, the PO fraction was reduced from 80%in YPD to 33% in HU, indicating that the 4∶1 bias toward IC/IS mediated URA3 inactivation in YPD was completely lost in HU, where the majority of URA3 inactivation occurred via non IC/IS mediated events. RSZ2 was notable because the extent of IC/IS bias in YPD appeared to be unusually strong (26/26; Figure 3E). In HU, we found that six of 29 had undergone a non IC/IS mechanism of URA3 loss, suggesting that the negative effect of HU might be irrespective of the intrinsic robustness of the IC/IS-bias.
In the current analysis, a 5FOAR colony exhibiting only the WT band (e.g. Figure 3CD, samples denoted by an “*”) was inferred to have undergone URA3 inactivation via a non IC/IS mediated event, such as an IH gene conversion/crossover, ectopic recombination (ECT), or chromosome loss (Figure 1C). As mentioned above, it is possible to distinguish an IH event from the others by the virtue of the fact that a 5FOAR colony that arose via an IH recombination event would carry two copies of WT allele, while the rest carries just one (Figure 1C).
Indeed, quantitative analysis of the WT fragments in Southern Blot images revealed that some of the non IC/IS samples had a two-fold greater signal associated with the WT band relative to an IC/IS sample that carried a copy of the WT and the PO allele each (Figure 3G, compare WT band intensity in “IH” and “IC/IS” lanes); these samples were inferred to have undergone URA3 inactivation via IH-recombination. We were also able to confirm the occurrence of an ectopic (ECT) event by the presence of a novel chromosome sized fragment containing the hisG sequence on a Southern Blot of a pulse field gel (PFG) (e.g. Figure 3F), suggestive of a chromosome translocation event. Applying these criteria, we were able to infer that the majority of 5FOAR colonies that arose in YPD (85/86) or in HU (78/89) had undergone URA3 inactivation via one of these three (IC/IS, IH, or ECT) mechanisms (Figure 3I and J). The rest, 1/86 in YPD and 11/89 in HU, was classified as “Other”, which would include mechanisms such as chromosome loss.
Of the 86 YPD 5FOAR colonies analyzed, about 25%, or 15, had undergone URA3 loss via a non IC/IS event(s). Among them, all but two were mediated by IH recombination. The remaining two corresponded to an ECT and an Other event each (Figure 3I, “Total”). In HU, a reduction in the fraction of IC/IS event was accompanied by a significant increase in the fractions of ECT and Other, where they rose to 8% (7/89) and 12% (11/89) from about 1% in YPD (Figure 3J, “Total”). The only significant increase in IH fraction conferred by HU was at RSZ2, where it rose from 0/26 in YPD to 6/29 in HU (Figure 3I and J).
To test whether the effect of HU on the IC/IS bias might have been due to reduced dNTP levels, we examined the effect of sml1Δ. Sml1 is a negative regulator of RNR and its deletion leads to a ∼2.5 fold increase in dNTP levels [22]. We found that the mutation conferred a modest, but statistically significant increase in the overall PO fraction from 80% (117/146) in YPD to 89% (89/100) in sml1Δ (Chi-square analysis, p<0.05; Figure 4F, “Total”). The observed effects of HU and sml1Δ suggest a positive correlation between dNTP levels and the extent of IC/IS bias, and implicate dNTP availability in regulation of recombination outcomes and genome rearrangements.
The results also revealed that the effect of sml1Δ on IC/IS bias might be locus dependent; while the overall effect was an increase, the mutation actually lead to a modest decrease in the bias at RSZ2 from 100% (26/26) in WT to 90% (18/20) in sml1Δ (Figure 3E; Figure 4F). Similarly, we found that the effect of sml1Δ on recombination rate was locus specific; while the mutation lead to an increase at RSZ1, RSZ2, and TER, but it reduced the rate at ORI (Figure 4D, E). Overall, the effect of sml1Δ on ORI differed from the rest in that it was the only locus where the mutation conferred a statistically significant reduction in recombination rate and a significant increase in the IC/IS-bias. It would be require analysis of additional origin sequences to confirm whether the observed effect might be origin specific.
No significant effect of sml1Δ was observed on the rate of URA3 inactivation in haploids (Figure 4A, C).
If chromosome breakage is mechanistically linked to recombination activity, then HU, a condition that promotes RSZ specific chromosome breakage, should also confer an RSZ specific increase in the rate of URA3 inactivation. Unexpectedly however, we found that HU increased the rate at both RSZ and non RSZ loci, indicating that the breakage was not linked to recombination activity. To confirm this further, we examined the effects of additional conditions shown to regulate chromosome breakage at RSZs. Specifically, we chose high temperature and mec1Δ sml1Δ, the two conditions shown to promote the breakage, and rrm3Δ, a mutation shown to suppress it [21]. Rrm3 encodes a DNA helicase involved in replication fork progression through ∼1000 discrete fork pause sites in the budding yeast genome, whose inactivation leads to fork stalling at these loci [28], [29]. The rrm3Δ mediated fork stalling triggers Mec1/Tel1 dependent S phase checkpoint activation and Sml1 removal. The latter in turn, promotes fork progression through RSZ and prevents RSZ breakage even in the absence of Mec1 function [21], [30]. We limited our analyses to diploids only, where the effect on IC/IS bias can be assessed in addition to the rate of URA3 inactivation.
Results showed that high temperature (37°C), rrm3Δ, and mec1Δ sml1Δ increased the average rate of URA3 inactivation by ∼2.5 fold over WT at 30°C (Figure 5A, “Average”). Low temperature (23°C), on the other hand, lead to ∼30% reduction. These observations confirmed previous reports on recombinogenic effects of rrm3Δ and mec1Δ sml1Δ. Furthermore, they revealed a positive correlation between temperature and recombination rate, suggesting that temperature itself might regulate endogenous recombination activity. Importantly, the observed effect was not specific to RSZs, providing further support for the notion that chromosome fragility and recombination activity at RSZs are not mechanistically linked.
Similarly to HU, the three recombinogenic conditions, 37°C, rrm3Δ, and mec1Δ sml1Δ lead to a reduction in the overall PO fraction (Figure 5B, “Average”). In contrast, low temperature (23°C), the only condition that decreased the rate of recombination, lead to a modest increase in the fraction. Together, these observations suggest an association between increased recombination activity and loss of IC/IS bias. Finally, results also showed that the effects of different conditions on IC/IS bias can be either RSZ specific, as in the case of HU, or general, as in the case of temperature, rrm3Δ, and mec1Δ sml1Δ (Figure 5B legend).
RSZ is a homolog of the mammalian common fragile sites, noted for its sensitivity to dNTP depletion and inactivation of Mec1 and Tel1, the budding yeast ATR and ATM, respectively. Here, we utilized RSZs to address the proposal that mammalian fragile sites are recombination hotspots and that increased recombination activity at these loci contributes to cancer development. Our analyses revealed that; (i) RSZs are not recombination hotspots, (ii) recombinogenic conditions (e.g. rrm3Δ, mec1Δ, HU, and high temperature) down-regulate IC/IS-bias in mitotic recombination, and (iii) RSZs exhibit greater sensitivity to HU dependent loss of IC/IS bias. Below, we discuss each of these findings further.
We found that rates of URA3 inactivation at RSZs were comparable to non-RSZ loci under all tested conditions. Based on this, we conclude that RSZs are not recombination hotspots, defined as loci of increased recombination activity. A key assumption is that the URA3 inactivation event monitored in the current study is a readout for recombination activities at the locus, and not for an indirect effect(s) of a more distant element(s). For example, studies of LOH in diploids indicate that most LOH events occur by a crossover between the heterozygous loci, and that the frequency of these events increases as a function of distance from centromere (e.g. [31]). In addition, Ty elements are recombination hotspots that might affect recombination activities downstream of the loci (e.g. [32]). Notably however, the frequency of the LOH events monitored in the current study was independent of the distance from CENIII (Figure 3I, J). Similarly, the rates of overall URA3 inactivation at the loci downstream (the three non RSZ loci) and upstream (the two RSZ loci) of a Ty element were comparable (Figure 2B, 4B, and 5A). Furthermore, the approach employed in the current study – i.e. integration of a reporter construct at a specific locus - is a widely utilized means of assessing local recombination activities (e.g. [33]). Together, these considerations strongly suggest that the assay utilized in the current study monitors local recombination activity.
It was surprising that RSZs did not have intrinsically higher rates of recombination than non RSZs. However, the current observation is actually consistent with the fact that RSZ breakage occurs during prometaphase in the context of topoisomerase II- and condensin- dependent chromosome condensation [3], [20], while most spontaneous mitotic recombination occurs during S phase in the context of stalled- and collapsed-replication forks [5], [7], [26]. The apparent temporal separation and differential genetic requirements suggest that chromosome breakage and recombination at RSZs might each entail a process that is independently regulated. However, it is also possible that the lack of correlation is a feature specific to RSZ (and mammalian fragile sites, by extension) and not a general feature of a fragile site.
Our results confirmed an earlier observation that the locus-to-locus variation in spontaneous mitotic recombination rate is relatively modest, in contrast to the variation in meiotic recombination rates, which can be several orders of magnitude [34], [35]. This difference is likely due, at least in part, to the nature of DNA structure that leads to recombination in each case: stalled- or collapsed-replication forks for mitotic recombination and developmentally programmed DSBs for meiotic recombination [26], [36]. Formation of meiotic DSBs are regulated at multiple levels, including targeted localization of Spo11, the enzyme that catalyzes the breakage, to the hotspot regions of the genome [37]. This in turn ensures that meiotic DSBs do not occur uniformly throughout the genome, but preferentially at DSB hotspots. In contrast, the occurrence of stalled- or collapsed-forks leading to spontaneous mitotic recombination does not appear to be regulated per se; rather, it appears to be a unintended consequence of a replication fork encountering a locus that is difficult to replicate, for example, due to either unusual DNA or chromatin structures or damaged DNA [6], [7], [25], [26], [38].
Previous studies have shown that genes encoding for proteins involved in processes such as replication fork progression, stalled fork integrity, and fork restart impact endogenous mitotic recombination rates (e.g. [39], [40]). The recombinogenic effects of HU or rrm3Δ observed in current study are likely due to increased incidences (either in the frequency and/or the duration) of fork stalling stemming from depletion of dNTP pools or loss of a replisome associated helicase activity, respectively [28], [41]. The effects of mec1Δ sml1Δ, on the other hand, is unlikely due to increased fork stalling because replication forks proceed faster in an sml1Δ background compared to WT [3]. Given that Mec1 is required for stability of stalled forks and is a key regulator of homologous recombination and recombination related processes involved in replication fork restart [18], [42], the recomginogenic effects of mec1Δ sml1Δ are likely to stem from a defect(s) in processes that occur after fork stalling. The mechanism(s) by which temperature might affect endogenous recombination activity remains unknown. Notably however, there has been a precedent in meiotic recombination, where temperature appears to play an important regulatory role(s) [43].
On average, URA3 inactivation under the standard growth condition was four times more likely to occur via an IC/IS mediated event than all other mechanisms combined. We found that HU, high temperature, rrm3Δ, and mec1Δ sml1Δ abolished this bias while low temperature and sml1Δ enhanced it. These observations suggest that partner choice in mitotic recombination might be subjected to regulation, for example by factors like dNTP availability and temperature. During meiotic recombiantion, a significant fraction of meiotic DSBs is developmentally programmed to be repaired with an IH bias, using an intact homolog as a repair template, rather than a sister chromatid. Evidence indicates that such IH bias in meiotic recombination is mediated, at least in part, by expression of several meiosis specific chromosomal proteins that fundamentally alter meiotic chromatin structure; this in turn, favours physical interaction between the homologs while minimizing that between sister chromatids, overcoming the IC/IS bias intrinsic during mitotic recombination [44], [45], [46]. The latter implicates chromatin structure, notably the status of sister chromatid cohesion, in mitotic IC/IS bias.
We utilized RSZ, a model for mammalian common fragile sites, to address the proposal that the fragile sites contribute to cancer due to increased recombination activity at the loci. The evidence presented here, however, suggests an alternative mechanism, which implicates the nature of recombination outcomes, rather than the overall recombination rate. The only RSZ specific recombination activity revealed in the current study was its greater sensitivity to HU induced loss of IC/IS bias. Although the sample size is limiting (i.e. one insertion in each of two RSZs), an implication would be that, depending on the nature of stress, recombination events at RSZs (and at mammalian common fragile sites, by extension), might be more likely to lead to a LOH and translocation, the type of alterations shown to contribute to cancer. Taken together, current observations provide a fresh insight into the ways in which fragile sites and other recombinogenic conditions may contribute to genome rearrangements.
Relevant genotypes of the strains utilized in current study are summarized in Table S1. All URA3 strains were generated by standard yeast genetics procedures including transformation, mating, sporulation, and specific phenotype selections. A wild type haploid strain NKY291, was transformed with each of the five DNA fragments containing the hisG-URA3-hisG cassette flanked by ∼500 bp upstream (L) and ∼500 bp downstream (R) genomic sequences of the targeted loci. The fragments were prepared from integration plasmids (Figure S1) by NotI digestion and gel purification. Correct integration of hisG-URA3-hisG at each locus among randomly selected URA transformants was confirmed by Southern Blot analysis (Figure S3). Two independent transformants of each locus were mated with NKY292, a WT haploid strain of the opposite mating type to generate heterozygous diploid strains, from which URA3 haploids were re-derived. These haploid strains (JDCY 463, 465, 230, 233, 239, 243, 232, 233, 235, 237) were used for all subsequent strain construction.
The rate of URA3 inactivation was determined by the method of median [24]. For each measurement, 15 colonies of comparable size (1.5–2 mm in diameter) freshly grown on YPD plates, were individually suspended in 5 mls of YPD or YPD+10 mM HU and cultured for 18 hours at 30°C (Figure S2A). Samples were then diluted in water and plated on YPD or 5FOA plates to measure the number of total viable cells or those that had undergone a URA3 inactivation event, respectively. To ensure that only the URA3 inactivation events that occurred during the 18 hour of unselected growth in liquid medium were included for the analyses, only the 5FOARcolonies of comparable size (1.5–2 mm) were counted following a three day incubation at 30°C. Statistical analyses on the rates of URA3 inactivation were assessed as described [47].
For each condition examined, a total of 20 or more independent 5FOAR colonies from each locus were subjected to molecular analysis (Figure S2B). Genomic DNA from each colony was restricted using the appropriate restriction enzyme (Figure S3; Figure 3) and subjected to Southern Blot analysis. For each locus, the L and R fragments used for plasmid construction (Figure S1) were used as probes. As a control for quantifying the relative signals associated with different diagnostic fragments (e.g. Figure 3E), a PCR amplified Mec1 fragment corresponding to nucleotide numbers 5539 to 7027 in the OFR was included as a probe (Table S2). Each restriction enzyme used for Southern analysis cleaves endogenous MEC1 sequence to generate a novel sized fragment that hybridizes to the MEC1 probe. To confirm the occurrence of ectopic recombination events, the candidate 5FOAR colonies were analyzed by Pulse Field Gel electrophoresis (PFGE) and Southern Blot analyses using hisG as a probe. For PFGE, Chromosome sized genomic DNA samples were prepared in low melting point agarous plugs as previously described [48]. Electrophoresis condition optimized for resolution around ChrIII was performed as described [49].
|
10.1371/journal.pmed.1002138 | Characterization of Novel Antimalarial Compound ACT-451840: Preclinical Assessment of Activity and Dose–Efficacy Modeling | Artemisinin resistance observed in Southeast Asia threatens the continued use of artemisinin-based combination therapy in endemic countries. Additionally, the diversity of chemical mode of action in the global portfolio of marketed antimalarials is extremely limited. Addressing the urgent need for the development of new antimalarials, a chemical class of potent antimalarial compounds with a novel mode of action was recently identified. Herein, the preclinical characterization of one of these compounds, ACT-451840, conducted in partnership with academic and industrial groups is presented.
The properties of ACT-451840 are described, including its spectrum of activities against multiple life cycle stages of the human malaria parasite Plasmodium falciparum (asexual and sexual) and Plasmodium vivax (asexual) as well as oral in vivo efficacies in two murine malaria models that permit infection with the human and the rodent parasites P. falciparum and Plasmodium berghei, respectively. In vitro, ACT-451840 showed a 50% inhibition concentration of 0.4 nM (standard deviation [SD]: ± 0.0 nM) against the drug-sensitive P. falciparum NF54 strain. The 90% effective doses in the in vivo efficacy models were 3.7 mg/kg against P. falciparum (95% confidence interval: 3.3–4.9 mg/kg) and 13 mg/kg against P. berghei (95% confidence interval: 11–16 mg/kg). ACT-451840 potently prevented male gamete formation from the gametocyte stage with a 50% inhibition concentration of 5.89 nM (SD: ± 1.80 nM) and dose-dependently blocked oocyst development in the mosquito with a 50% inhibitory concentration of 30 nM (range: 23–39). The compound’s preclinical safety profile is presented and is in line with the published results of the first-in-man study in healthy male participants, in whom ACT-451840 was well tolerated. Pharmacokinetic/pharmacodynamic (PK/PD) modeling was applied using efficacy in the murine models (defined either as antimalarial activity or as survival) in relation to area under the concentration versus time curve (AUC), maximum observed plasma concentration (Cmax), and time above a threshold concentration. The determination of the dose–efficacy relationship of ACT-451840 under curative conditions in rodent malaria models allowed prediction of the human efficacious exposure.
The dual activity of ACT-451840 against asexual and sexual stages of P. falciparum and the activity on P. vivax have the potential to meet the specific profile of a target compound that could replace the fast-acting artemisinin component and harbor additional gametocytocidal activity and, thereby, transmission-blocking properties. The fast parasite reduction ratio (PRR) and gametocytocidal effect of ACT-451840 were recently also confirmed in a clinical proof-of-concept (POC) study.
| The limited diversity of chemical mode of action in the global portfolio of marketed antimalarials along with recently observed artemisinin resistance that threatens the current first-line treatment highlights the urgent need for development of new antimalarials.
In accordance with target product profiles defined by the Medicines for Malaria Venture ([MMV]; www.mmv.org), a new model of not-for-profit public–private partnership providing guidance to the development of new drugs, ACT-451840, a new chemical class of potent compounds with a novel mode of action, was selected for early development.
This manuscript integrates a number of studies performed to characterize preclinically ACT-451840 and illustrates the new antimalarial drug development paradigm.
This study used in vitro models to investigate compound activity on sexual and asexual blood stages and a mouse model to study the human parasite P. falciparum in vivo.
While being fully active against artemisinin-resistant strains, ACT-451840 shares many of the favorable properties of artemisinins like its fast onset of action, activity against all asexual blood stage forms, and a PRR of >4 log per parasite cycle.
Pharmacodynamic properties of ACT-451840 were interpreted in two mouse models of malaria with respect to the pharmacokinetic properties, with the objective to establish the portable PK/PD parameters to estimate efficacious exposure.
Modeling the dose–efficacy relationship of ACT-451840 supported prediction of the human efficacious exposure and therefore laid the groundwork for the new clinical development phases.
Confirmed in an experimental human malaria infection model, the properties of ACT-451840, including its fast action observed in in vitro and in vivo models and its transmission-blocking activity, suggest this compound may be able to replace the artemisinin component in artemisinin-based combination therapy.
| Malaria caused 438,000 deaths worldwide in 2015, of which 70% were in children under the age of 5 y [1]. Between 2000 and 2015, strategies for malaria control and eradication reduced the incidence of malaria by 48% in the WHO African Region. The upscaled interventions consisted of increased accessibility to long-lasting insecticidal bed nets, protection of the population at risk by indoor residual spraying, and increased access to rapid diagnostic tests and artemisinin-based combination therapies. However, with the detection of parasite resistance to artemisinin, the core compound of artemisinin-based combination therapies, in five countries of Southeast Asia, the availability of efficacious combination therapies is under threat [1]. Additionally, not a single new chemical class of antimalarials has been registered since 1996 [2], and the current global portfolio of antimalarial compounds in late clinical development relies largely on novel combinations of existing drugs, not novel compounds [3]. These elements highlight the critical and urgent need for new drugs to treat malaria.
In this search, ACT-213615, a compound from Actelion Pharmaceuticals Ltd. (ACT) with a mode of action distinct from that of all registered antimalarials, was recently described [4]. The compound was discovered in a phenotypic screen and showed potent and fast-acting activity in in vitro assays against all asexual erythrocytic stages of P. falciparum (i.e., rings, trophozoites, and schizonts). Further investigations regarding the mode of action established an interaction of this compound class with P. falciparum multidrug resistance protein-1 (PfMDR1) [5,6].
The present work illustrates a novel antimalarial drug development process, in that the studies were performed in partnership with research groups around the globe from industry and academia. MMV (Geneva, Switzerland), a new model of not-for-profit public–private partnership, was instrumental in the establishment of those collaborations and has transformed the field of malaria therapy by providing guidance to the development of new drugs. The strategy of MMV is the prioritization of new treatments that provide a single exposure radical cure and prophylaxis (SERCaP). Prevention of transmission, prevention of relapse following infection by P. vivax and Plasmodium ovale, and significant post-treatment prophylaxis are the key properties sought for future medicines, as described in MMV's new target compound profiles (TCP) [7]. Achievement of the SERCaP objectives might require the combination of several compounds (two to three), each with specific profiles and properties. Ideally, a new malaria treatment requires at least one molecule capable of replacing the fast-acting artemisinin component described as target compound profile 1 (TCP1). Suitability as a TCP1 candidate can be assessed in experiments (A) determining potency in both P. falciparum and P. vivax strains, (B) including strains already resistant to other antimalarials including artemisinins, (C) measuring log kill in vitro in comparison to existing antimalarials, and (D) establishing efficacy in an appropriate in vivo animal model. The TCP2-type molecules would be used as long duration combination partners that complete the clearance of the blood stage parasites not eliminated by TCP1 molecules. Understanding the PK profile of a compound will determine its suitability as a TCP2 candidate. Molecules with a potential for blocking transmission should target mature gametocytes (TCP3b) and also antagonize other nondividing parasites such as hypnozoites (TCP3a). Finally, molecules blocking the infection of the host liver by sporozoites, or killing liver schizonts, will cover the need for TCP4 drugs [8].
Leveraging these TCPs, the continuous drug discovery effort in the new class of potent antimalarials around ACT-213615 led to the selection of ACT-451840. In the present study, the characterization of ACT-451840 is used to illustrate the new drug development paradigm in which new technical expertise and model-based methodologies are being used to improve the subsequent clinical development phases and reduce attrition rates. Firstly, the experiments to evaluate the biological potential of ACT-451840 in P. falciparum in vitro assays against all asexual erythrocytic stages and against a panel of clinical isolates including P. vivax are shown. Additionally, key PD properties of the compound are described, including its precise rapidity of action and parasite log kill, efficacy in two pharmacological murine models with both human and mouse malaria parasites, and in vitro inhibition of gamete formation from the male gametocyte stage, resulting in transmission blocking in mosquitoes. Finally, the modeling of efficacy in murine models defined either as antimalarial activity, or as survival, in relation to AUC, Cmax, and time over a threshold concentration, was used to determine the dose–efficacy relationship of ACT-451840. The subsequent prediction of the human efficacious exposure was used to guide the dose selection for the initiation of the POC study. The goal of the here described preclinical assessment of activity and dose–efficacy modeling of ACT-451840 was to lay the groundwork for the next phases of clinical development.
Animal experiments performed at Actelion Pharmaceuticals and at the Swiss Tropical and Public Health Institute (Swiss TPH), or performed at GlaxoSmithKline (GSK) were approved, respectively, by the Swiss Cantonal Authorities and by the Diseases of the Developing World Ethical Committee on Animal Research. The animal studies carried out at GSK were in accordance with European Directive 2010/63/EU and the GSK Policy on the Care, Welfare and Treatment of Animals and were accredited by the Association for Assessment and Accreditation of Animal Laboratory Care for the ones performed at Diseases of the Developing World Laboratory Animal Science facilities. The results from the animal experiments are reported as per ARRIVE guidelines (S1 Checklist). The human biological samples were sourced ethically, and their research use was in accord with the terms of the informed consents. Ethical approval for the field studies were obtained from the Human Research Ethics Committee of the Northern Territory Department of Health & Families and Menzies School of Health Research, Darwin, Australia (HREC 2010–1396), and the Eijkman Institute Research Ethics Commission, Jakarta, Indonesia (EIREC 47).
The synthesis of ACT-451840 developed at Actelion Pharmaceuticals (Allschwil, Switzerland) is shown in the supporting material (S1 Fig). ACT-451840 is patented (WO 2011/083413).
The in vitro antimalarial activity of ACT-451840 was measured at the Swiss TPH (Basel, Switzerland) using the [3H]-hypoxanthine incorporation assay [9]. Results were expressed as the concentration resulting in 50% inhibition (IC50). In vitro time-, stage-, and concentration-dependent effects were assessed using pyrimethamine as a stage-specific and slow-acting control, as described elsewhere [10].
Ring-stage survival assays (RSA0-3h) were carried out at Columbia University Medical Center (New York, United States) as previously described [11,12]. Parasite cultures were synchronized using 5% sorbitol (Sigma-Aldrich), and schizonts were incubated in RPMI-1640 containing 15 units/ml sodium heparin for 15 min at 37°C to disrupt agglutinated erythrocytes. These schizonts were concentrated over a gradient of 75% Percoll (Sigma-Aldrich), washed once in RPMI-1640, and incubated for 3 h with fresh erythrocytes to allow time for merozoite invasion. Cultures were then subjected again to sorbitol treatment to eliminate remaining schizonts. The 0–3 h postinvasion rings were adjusted to 1% parasitemia and 2% hematocrit in 1 mL volumes (in 48-well plates) and exposed to a range of ACT-451840 drug concentration (700–35 nM) or 0.1% DMSO for 6 h. Duplicate wells were established for each parasite line ± drug concentration. One-mL cultures were then transferred to 15-mL conical tubes, centrifuged at 800xg for 5 min to pellet the cells, and the supernatants carefully removed. As a washing step to remove drug, 10 mL culture medium were added to each tube, and the cells were resuspended, centrifuged, and the medium was aspirated. Fresh medium without drug was then added to cultures, which were returned to standard culture conditions for 66 h. A media change was performed, and parasite viability was assessed by flow cytometry at 96 h. Stained parasites were detected using a BD Accuri 6 Flow Cytometer and analyzed using FlowJo Software. IC50 values were determined by nonlinear regression using GraphPad Prism 6.0. For statistical analyses, Student’s t tests were performed, and significance was set at p < 0.05.
Ex vivo drug susceptibility testing in P. falciparum and P. vivax (often a mix of rings, trophozoites, and schizonts) was assessed in clinical isolates at the Timika Research Facility (Papua, Indonesia), an area with confirmed high levels of multidrug resistance to chloroquine, sulfadoxine-pyrimethamine, and amodiaquine in both species [13–15]. Thick blood films made from each well were stained with 5% Giemsa solution for 30 min and examined microscopically. The number of schizonts per 200 asexual-stage parasites was determined for each drug concentration and normalized to the control well. The dose–response data were analyzed using nonlinear regression analysis (WinNonLin 4.1, Pharsight Corporation), and the concentrations resulting in 50% inhibition (IC50) were derived using an inhibitory sigmoid Emax model. Ex vivo IC50 data were only used from predicted curves for which the E0 and Emax were within 15% of 1 and 100, respectively.
In vitro PRR testing was conducted at GSK (Tres Cantos, Madrid, Spain) as previously described [16]. The assay used the limiting dilution technique to quantify the number of parasites that remained viable after drug treatment. P. falciparum strain 3D7 (MR4) was treated with drug concentration corresponding to 10 x IC50. Conditions of parasites exposed to treatment were identical to those used at GSK in the IC50 determination (2% hematocrit, 0.5% parasitemia). Parasites were treated for 120 h. Drug in culture medium was renewed daily over the entire treatment period. Parasite samples were collected from the treated culture every 24 h (24, 48, 72, 96, and 120 h time points); drug was washed out of the sample, and parasites were cultured drug-free in 96-well plates by adding fresh erythrocytes and culture medium. To quantify the number of viable parasites after treatment, 3-fold serial dilutions were used with the above mentioned samples after removing the drug. Four independent serial dilutions were performed with each sample to correct experimental variations. The number of viable parasites was determined after 21 and 28 d by counting the number of wells with growth using [3H]-hypoxanthine incorporation. The number of viable parasites was back-calculated by using the formula Xn-1 where n is the number of wells able to render growth and X the dilution factor (when n = 0, number of viable parasites is estimated as zero) [16]. The PRR, defined as the logarithm to base 10 of the number of parasites the drug can kill in one parasite life cycle and indicating the killing rate of the compound investigated, was calculated as the decrease in viable parasites over 48 h. A lag phase was considered to occur for as long as drug treatment did not produce the maximal rate of killing, and this period of time was excluded for PRR calculation. The parasite clearance time to kill 99.9% of the initial population was determined using a regression calculated on the log-linear phase of the parasite reduction, taking into account the lag phase.
In vitro (ex vivo) activity against the mouse malaria parasite P. berghei was measured at the Swiss TPH as described elsewhere [17], with the following modifications: heparinised blood of P. berghei infected and uninfected control female NMRI mice (final hematocrit of 2.5%) was used and exposed to compounds for 16 h followed by 8 h of [3H]-hypoxanthine incorporation.
In vivo efficacy against P. berghei was conducted at the Swiss TPH as previously described [18], with the modification that female NMRI mice from Harlan (Horst, The Netherlands) (n = 3) were infected with a GFP-transfected P. berghei ANKA strain (kindly donated by A. P. Waters and C. J. Janse, Leiden University Medical Center, The Netherlands) and parasitemia was determined using standard flow cytometry techniques. ACT-451840 was dissolved or suspended in corn oil before dosing and administered 24 h after infection (single-dose regimen) or 24, 48, and 72 h after infection (triple-dose regimen). With the single-dose regimen, blood was collected on Day 3 (72 h after infection). Samples for the triple-dose regimens were collected on Day 4 (96 h after infection). Antimalarial activity was calculated as the difference between the mean percent parasitemia for the control and treated groups expressed as a percent relative to the control group, consisting of five mice. Control mice were euthanized on Day 4 in order to prevent death typically occurring on Day 6. For the consistency in PK/PD modeling, Day 5 instead of Day 6 was considered the day of death, because treated animals were always euthanized one day before their natural death for animal welfare reasons. Animals were considered cured if there were no detectable parasites on Day 30 post-infection. To confirm the reliability of this assumption, blood from the surviving mice in curative experiments was inoculated into uninfected mice, and no parasitemia developed in these mice even after 30 d [19]. The antimalarial efficacy is defined either as antimalarial activity or as survival: the two readouts of the P. berghei murine model.
In vivo efficacy against P. falciparum was conducted at GSK according to the standard assay previously described [20,21]. Female NODscidIL2Rγnull mice were engrafted with human erythrocytes and infected with the isolate P. falciparum PfNF540230/N3, a strain developed at GSK for growth in engrafted mice. At Day 3 after infection, chloroquine (at 2 or 5 mg/kg) or ACT-451840 (3, 10, 30, or 100 mg/kg; n = 3 mice per dose, formulated in corn oil) were dosed via oral gavage once a day for four consecutive days (quadruple-dose regimen). Blood samples were collected on Days 3, 4, 5, 6, and 7 (n = 3 per time point) for parasitemia measurement by flow cytometry. Similarly, blood samples were collected at 0.5, 1, 2, 4, 6, 8, 12, 24, and 48 h after a single administration of 4.7 mg/kg of ACT-451840 to a parallel group of three P. falciparum-infected mice for PK parameters determination. Results are expressed exclusively as antimalarial activity in the P. falciparum murine model.
Mature P. falciparum NF54 gametocytes were cultured at Imperial College London (London, United Kingdom), as previously described [22]. The cultures, which displayed between 0.2 and 0.4% exflagellation centers as a percentage of the total blood cell population, were then taken and divided into 96-well plates containing ACT-451840 in dose response (range 0.84 nM–20 μM). Final DMSO concentration in the assay did not exceed 0.25%. Plates were incubated for 24 h in the presence of ACT-451840 before male and female gametocyte viability was assessed by stimulating gamete formation by temperature decrease from 37°C to ~22°C and the addition of the gametocyte activating factor xanthurenic acid (5 μM). Male gamete formation was quantified by computer-aided automated identification and counting of exflagellation centers. Female gamete formation was visualized by live immunofluorescence with a Cy3-conjugated antibody specific for Pfs25—a female gamete expressed surface protein. Gamete formation was expressed as a percent inhibition, taking into consideration DMSO-negative control wells and 10 μM methylene blue-positive control wells, methylene blue being a potent inhibitor of the functional viability of male and female stage V gametocytes [22]. IC50 values were calculated using GraphPad Prism version 6 using the logarithmic versus normalized response (variable slope) function (Graphpad Software, San Diego, US).
ACT-451840 was tested in the standard membrane feeding assay at TropIQ (Nijmegen, The Netherlands) in two modes, as previously described [23]. In the direct mode, the compound was directly added to a bloodmeal containing mature stage V gametocytes that was fed to Anopheles stephensi mosquitoes. In the indirect mode, mature stage V gametocytes were pre-exposed for 24 h to a serial dilution of compound (10−5 to 10−9 M final concentration) prior to mosquito feeding. Gametocytes were adapted to a hematocrit of 50% in full serum and fed to A. stephensi mosquitoes. Seven days after the bloodmeal, the number of oocysts in the midgut was determined by microscopy for both modes. ACT-451840 was tested in duplicate in a 9-point dose response series (1010–106 M in 0.5 log dilutions). Vehicle control (0.1% DMSO) and positive control (10 μM dihydroartemisinin) were included. Twenty mosquitoes were dissected per sample, and oocyst load was analyzed per mosquito. All assays passed the quality control criterion of >50% infected mosquitoes in the control cages.
Female NMRI mice were purchased from Harlan (Horst, The Netherlands) by Actelion Pharmaceuticals. ACT-451840 was formulated as a solution in corn oil and dosed via oral gavage at 10, 100, and 300 mg/kg. Serial blood samples were collected at 0.5, 2, 4, and 8 h after dosing from two mice and after 1, 3, 6, and 24 h from two other mice, creating composite PK profiles from, in total, four mice per dose. Blood samples were analyzed using liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) after protein precipitation. PK parameters were estimated using non-compartmental analysis from the composite profiles.
A single compartment PK model with saturable absorption was obtained using naïve pooling to fit simultaneously the observed blood concentration time data from the PK studies done at a dose of 10, 100, and 300 mg/kg, as described above. Using this model, blood concentration time profiles were predicted for all dosing regimens used in the mouse antimalarial efficacy studies. Cmax, AUC, and time above a threshold concentration were subsequently calculated using non-compartmental analysis. PD modeling was done using an Emax model using Cmax, AUC, or time above a threshold concentration versus survival or antimalarial activity. All modeling and PK analyses were done using Phoenix Winonlin 6.3 (Cetara, Princeton, US).
Antimalarial asexual blood-stage activity of ACT-451840 was evaluated against a panel of P. falciparum culture adapted isolates by a [3H] hypoxanthine incorporation assay (Table 1). ACT-451840 showed a low nanomolar activity against the drug-sensitive P. falciparum NF54 strain (mean IC50 = 0.4 ± 0.0 nM; IC90 = 0.6 ± 0.0 nM; IC99 = 1.2 ± 0.0 nM, n ≥ 3; ± SD) and was almost equally active against a number of drug-resistant isolates. The IC50 values of the control compounds, e.g., artesunate, chloroquine, and pyrimethamine, against the NF54 strain were 3.7 ± 0.5 nM, 11 ± 2.1 nM, and 18 ± 0.8 nM, respectively.
Antimalarial asexual blood-stage activity of ACT-451840 was also evaluated against a Cambodian clinical isolate (Cam3.II, which has a K13 R539T mutation) that was gene-edited to the wild-type K13 gene (Cam3.IIrev) and against Cam3.IIC580Y, a gene-edited line that is artemisinin resistant and encodes the C580Y mutation in the K13 gene [11,12]. The mutation C580Y, conferring resistance to artemisinin, was found to not alter parasite susceptibility to ACT-451840, mefloquine, or artesunate in conventional 72-h dose–response assays that determine IC50 growth inhibition values on both the Cam3.II and V1/S backgrounds (S1 Text). Additionally, both Cam3.II lines were subjected to the recently developed ring-stage survival assay (RSA0-3h), in which parasites were exposed for a 6-h pulse to a range of drug concentration (35–700 nM), and parasite survival was assessed 66 h later [11,12]. No difference in survival rates was observed between the parasite harboring a wild-type K13 allele or the C580Y mutation (Fig 1), confirming that ACT-451840 is fully efficacious against artemisinin-resistant parasites. In contrast, the C580Y strain showed the expected grade of resistance versus the exposure to 700 nM dihydroartemisinin [11,12].
In ex vivo assays against a range of clinical isolates collected from patients with malaria from Papua, Indonesia, a region where multidrug-resistant malaria is endemic, ACT-451840 was as potent as artesunate against both P. falciparum (median IC50 = 2.5 [Range 0.9–9.0] nM, n = 27) and P. vivax (median IC50 = 3.0 [Range 0.5–16.8] nM, n = 34) (Fig 2). The IC50 value of ACT-451840 against the murine malaria parasite P. berghei was 13.5 nM in in vitro ex vivo assays.
In an in vitro PRR assay [16], ACT-451840 showed a fast parasite killing profile against P. falciparum similar to that of chloroquine (log PRR of 4.4 and 4.5, respectively) without lag phase in the killing curve in contrast to atovaquone (48 h). The parasite clearance time to kill 99.9% of the initial population was 28 h for ACT-451840, similar to that of artemisinin and chloroquine (less than 24 h and 32 h, respectively) (Fig 3). In vitro stage specificity assays [24] with synchronous cultures of P. falciparum NF54 were consistent with this rapid killing rate. ACT-451840 rapidly reduced parasite growth and affected all asexual erythrocytic stages equally in a time- and concentration-dependent manner, similar to artemether (Fig 4) [24].
In vivo activity of ACT-451840 was assessed in the P. falciparum mouse model established at GSK (DDW, Tres Cantos, Spain). After an oral quadruple-dose regimen, ACT-451840 had a rapid onset of action and an effective dose, resulting in 90% antimalarial activity (ED90) of 3.7 mg/kg (3.3–4.9 mg/kg), which was comparable to the one of chloroquine (ED90 of 4.9 mg/kg) (Fig 5).
In a P. berghei mouse model, the ED90 after an oral triple-dose regimen of ACT-451840 was 13 mg/kg (11–16 mg/kg) (calculated from the data in S2 Text). This ED90 value was about 4-fold higher than the one observed in the P. falciparum mouse model (ED90 3.7 mg/kg). The difference in in vivo activities correlates with the difference in in vitro ex vivo activities of ACT-451840 against P. berghei (IC50 13.5 nM, see above) and P. falciparum ex vivo (IC50 2.5 nM, clinical isolates; Fig 2). Despite the lower activity of ACT-451840 against P. berghei, 20% and 100% of the infected mice were cured after 100 and 300 mg/kg triple-dose regimen, respectively (Fig 6 and S2 Text). In the same in vivo model, artesunate, the semisynthetic derivative of artemisinin, cured 50% of the mice at 100 mg/kg after a quadruple-dose regimen.
In summary, ACT-451840 had an IC50 below 10 nM in vitro against a panel of sensitive and resistant strains of P. falciparum and was fully active against recently identified artemisinin-resistant strains. Furthermore, the compound showed ex vivo activity against P. falciparum and P. vivax as well as a single digit mg/kg ED90 in a P. falciparum murine model. In addition, the compound inhibited growth of all asexual blood stages of P. falciparum in vitro, and its in vitro PRR was similar to that of chloroquine. These key properties fully meet the criteria described in TCP1 [8].
Using two bioassays, the effects of ACT-451840 on the functional viability of both male and female mature gametocytes and oocysts of P. falciparum were explored to illustrate how the question of transmission is addressed.
Within the human host blood, P. falciparum gametocytes are classified morphologically into five stages, denoted I–V, whilst P. vivax gametocytes do not show the same morphology. Although the various sexual stages do not cause clinical disease, stage V gametocytes are infective to mosquitoes. These gametocytes are not eliminated by the majority of current antimalarial agents [25] and therefore remain circulating in the human bloodstream for up to 3 wk, well after the disappearance of clinical symptoms of malaria [26]. Gamete formation is a functional readout for the viability of the mature stage V gametocyte [27]. ACT-451840 potently prevented male gamete formation from the gametocyte stage with an IC50 of 5.89 nM ± 1.80 nM (SD), as shown in Fig 7.
Interestingly, ACT-451840 had no effect on the functional viability of female gametocytes up to a concentration of 20 μM. Nevertheless, successful transmission requires viable gametocytes of both sexes to be present, as seen similarly with two male-specific molecules, MMV007116 and MMV085203, that demonstrate transmission blocking in a standard membrane feeding assay [27], supporting the potential of ACT-451840 to block malarial transmission.
In line with this, ACT-451840 did indeed block transmission when tested in the standard membrane feeding assay. In this assay, parasite cultures containing P. falciparum stage V gametocytes were exposed to compound for 24 h prior to mosquito feeding. Seven days postinfection, control feeds showed an average parasite density of 26.2 oocysts per mosquito, which is well within the range of oocyst densities observed in laboratory infections [28]. ACT-451840 dose-dependently blocked oocyst development in the mosquito, with an IC50 of 30 nM (95% confidence interval: 23–39 nM) (Fig 8A). At this parasite exposure, the prevalence of infection (number of infected mosquitoes) was inhibited, with an IC50 of 104 nM (95% confidence interval: 98–110 nM) (Fig 8B).
As the parasite exposure in the field is much lower (1–3 oocysts/mosquito), ACT-451840 potency might be underestimated and might have a greater effect on oocyst prevalence [29]. To address the specific effect of the compound against the parasite forms that develop in the mosquito midgut, a different mode of standard membrane feeding assay was performed, in which the compound was added to the parasite at the time of mosquito feeding. Here, ACT-451840 did not show any activity against P. falciparum oocysts at concentrations up to 1 μM (S2 Fig). These data strongly suggest that the transmission-blocking effect of ACT-451840 relies on its gametocytocidal activity.
In order to initiate investigation in humans, the preclinical safety profile of ACT-451840 was analyzed in dedicated toxicities studies performed in compliance with Good Laboratory Practice regulations and following the International Council for Harmonization guidelines. In safety pharmacology studies, no effects on central nervous and respiratory parameters were observed in rats dosed up to 1,000 mg/kg, and no effect on cardiovascular parameters were observed in dogs dosed up to 500 mg/kg. ACT-451840 was not genotoxic when assessed in vitro in a reverse mutation test or in a chromosome aberration study and in vivo in a micronucleus test. In the general oral toxicity studies performed with ACT-451840 for up to 4 wk, rats and dogs were dosed up to 2,000 mg/kg/day and 1,000 mg/kg/day, respectively. There was no evidence of local gastrointestinal toxicity. The no-observed-adverse-effect levels were established at 100 mg/kg/day in both species based on reduced body weight gain and clinical chemistry and histology findings, respectively. Based on these results, safety margins were calculated to cover the whole dose range to be tested in the single ascending dose Phase I study.
Towards the goal of initiation of a POC study and based on the results from the P. falciparum and P. berghei malaria mouse models, the human effective dose was predicted. The approach to model the PK/PD relationship is illustrated in Fig 9.
As a first step, PK parameters of ACT-451840 were determined in healthy mice after a single oral dose of 10, 100, and 300 mg/kg body weight. The PK parameters were calculated and are shown in Table 2.
To assess the PK parameters in infected mice, blood samples were taken at 1, 4, and 24 h after the first dose of a triple-dose regimen in the P. berghei efficacy studies. The employed doses were 15 or 30 mg/kg bis in die (b.i.d.), and 30, 50, 100, 300 mg/kg quaque die (q.d.) (n = 3/timepoint and dose). No significant difference in blood concentration was found between the time points studied in infected mice compared to those in healthy mice (S3 Fig).
In order to analyze the PK/PD relationship, drug exposure was predicted for all tested doses in P. berghei mouse model using a single compartment model with saturable absorption based on the concentration–time profile observed after 10, 100, and 300 mg/kg single-dose regimen in healthy mice. PK model parameter estimates are shown in Table 3.
The PK model was used to predict the blood PK parameters for all dosing regimens used in the P. berghei mouse model, as shown in Fig 6 and S2 Text.
The next step was to determine which of the PK parameters (Cmax, AUC, or time above a threshold concentration) was driving the curative efficacy of ACT-451840 in the P. berghei mouse model. To establish this PK/PD relationship, the predicted and measured PK properties were used to model curative efficacy in relation to AUC, Cmax, and time over a threshold concentration of 1,000 ng/mL. This threshold was defined as the minimum concentration needed to significantly increase the survival of mice treated with ACT-451840 after both single- and triple-dose regimens over that of control animals. The value of 1,000 ng/mL was estimated from the data in S2 Text, in which the triple-dose regimen of 10 mg/kg with a Cmax of 617 ng/mL showed no increase in survival compared to control animals, whereas the triple-dose regimen of 30 mg/kg with Cmax of 1,670 ng/mL showed an increase in survival up to 13 d.
The PK/PD relationships for the P. berghei in vivo efficacy experiments are given in Fig 10 and Fig 11 and the resulting PK/PD parameters are shown in Table 4.
The best fit for predicted versus observed PK/PD relationship was found for AUC and time above a threshold concentration versus survival, and time above a threshold concentration versus antimalarial activity, as illustrated by the correlation coefficients ≥0.96 (Table 5).
Survival and AUC were selected as parameters for the predictions in human. Survival is a robust readout and the relevant endpoint for clinical efficacy. AUC, rather than time above threshold, was used because the threshold value was based on PD endpoint in infected mice with P. berghei and may be different in humans infected with P. falciparum.
Having established AUC as the driver for survival in the P. berghei mouse model, this insight was used to predict an efficacious exposure against P. falciparum in humans with the results of the P. falciparum mouse model. In the P. falciparum mouse model after a quadruple-dose regimen, the ED90 of 3.7 mg/kg corresponded to a total AUC in blood of 227 ng*h/mL (linearly extrapolated from the exposure measured at a single dose of 4.7 mg/kg). The total AUC giving an ED90 in the P. berghei model (16,300 ng*h/mL after a triple-dose regimen of 13 mg/kg) was 23-fold lower than the total AUC giving cure (377,000 ng*h/mL after a triple-dose regimen of 300 mg/kg). Using this factor, the efficacious human total AUC was estimated to be 227 x 23 = 5,250 ng*h/mL, expressed in mouse blood. Correcting for the measured blood/plasma ratio of 0.6 (S3 Text), this corresponds to a predicted total efficacious human plasma AUC of 8,750 ng*h/mL. This exposure is about six times the exposure (1,408 ng*h/mL) observed after a single 500 mg dose in fed human subjects [30].
This report presents results from vitro and in vivo experiments to preclinically characterize a novel antimalarial compound (ACT-451840). The data shown here indicate that ACT-451840 shares many of the favorable MMV TCP1 properties of artemisinin and its derivatives, such as its potency against P. falciparum and P. vivax, fast onset of action, activity against all asexual erythrocytic stages, and PRR of >4 log per cycle of asexual blood stage forms. In addition, although the potency observed in the standard membrane feeding assay was 10-fold lower than the potency of the compound against asexual blood stage parasites, ACT-451840's gametocytocidal activity blocks transmission, fulfilling the properties requested for a TCP3b [8].
In the global portfolio of malaria medicines, other recently discovered novel compounds are the synthetic peroxide OZ439 [18], the spiroindolone KAE609 [31], the imidazolopiperazines KAF156 [32], the dihydroorotate dehydrogenase inhibitor DSM265 [33], and the translation elongation factor 2 inhibitor DDD107498 [23]. ACT-451840 shares the fast-acting properties [34] with OZ439 and KAE609, which are both foreseen to be part of a single-exposure radical cure product, as the compounds fit the TCP1, 3b and TCP1, 2 and 3b, respectively [7,35]. Although KAF156 and DDD107498 share the transmission-blocking characteristics with ACT-451840, both are seen to be agents with potential for chemoprophylaxis due to their in vitro activity against liver schizonts [23,36]. Finally, DSM265 exhibits properties of a TCP 2, positioning this compound with its long duration as potential partner of a single-exposure combination, in addition to its prophylaxis potential (TCP4) [33].
In contrast to the artemisinins, the here described compound class acts through a distinct mode of action by interacting with the P. falciparum multidrug resistance protein-1 [4,5]. Combined with the finding that ACT-451840 was well tolerated in a recent single ascending dose study and exhibited a much longer terminal half-life than artemisinin, comparable to OZ439 (34 h, 2 h, and 25–30 h, respectively) [30,37,38], this novel compound has the potential to be used as a substitute for the artemisinin derivatives in combination treatments. Additionally, the in vitro activity against artemisinin-resistant strains positions ACT-451840 as a strong candidate in the race for malaria eradication.
PD properties of ACT-451840 were interpreted in two mouse models of malaria with respect to the PK properties, with the objective to establish the portable PK/PD parameters to estimate efficacious exposure following similar methodology validated with KAE609 [35]. In this PK/PD approach, the robust endpoint of animal survival and cure were chosen for clinical relevance. Using survival as endpoint, both AUC and time above a threshold concentration were drivers for efficacy. The AUC resulting in cure in the mouse models was further translated to a predicted efficacious exposure in man, which was actually about six times the exposure observed after a single dose of 500 mg in fed state in healthy subjects. Two elements might give prospect to this conservative prediction. Firstly, the threshold concentration specified in the current PK/PD analysis (1,000 ng/mL in the P. berghei model) was defined as two times the 99% inhibitory concentration in a recent publication [35]. The alternative approach used to assess efficacious concentrations in vivo calculated the minimum inhibitory concentration and estimated a minimum parasiticidal concentration from modeling of the exposure–efficacy in the P. falciparum murine model. Because exposure was measured at a single dose of ACT-451840, this method cannot be applied in this case. However, given P. berghei and P. falciparum differences, it is likely that both concentrations would be much lower for ACT-451840 than the threshold concentration in the P. berghei model. Considering the 99% inhibitory concentration of ACT-451840, the value of the threshold would be 1.8 ng/mL (2.4 nM) and would be covered for 72 h, corresponding to one and a half asexual blood reproductive cycles, after a single dose of 500 mg in fed state in healthy subjects. Secondly, the results of the phase I study indicated the presence of circulating active metabolites with, on average, a 4-fold increase in antimalarial activity [30]. Providing these active metabolites were not present in the malaria murine model (not investigated) and, therefore, not taken into account for the human efficacious dose prediction, the overall antimalarial activity in humans dosed with ACT-451840 could be even higher than expected.
The aim of efficacious dose prediction is to provide guidance for the selection of optimal doses to be investigated in POC clinical trials as monotherapy. Although artemisinin monotherapy offers rapid recovery and fast parasite clearance [39], the use of combination therapy is recommended to reduce the risk of recrudescence associated with monotherapy, which is high for treatment regimens less than 7 d in length [40], and to decrease the potential development of drug resistance [41]. This approach allowed short-course combination therapies to be deployed with excellent efficacy, tolerability, and adherence, as illustrated with the lumefantrine/artemether combination therapy Coartem [42,43]. In addition, as fat intake enhances absorption, particularly for lumefantrine [44] and ACT-451840 [30], the development of a lipid-based formulation in humans might be the next major improvement towards limiting the variability and increasing exposure. Of note, work performed at Actelion Pharmaceuticals (S4 Text) has shown that the food effect observed with ACT-451840 [30] could be mimicked by the use of lipid-based formulations in fasted conditions in dog PK studies. Therefore, dose prediction might be refined during the development of the future medicine with new formulation and combination partners.
ACT-451840 has been tested as a monotherapy in an experimental human malaria infection model, a clinical trial developed by MMV in partnership with QIMR Berghofer of Medical Research Institute (Brisbane, Australia) [45]. The dosing regimen for this POC study (500 mg in fed condition) was constrained by the conditions tested in the phase I study. The results of this study confirmed the fast action of ACT-451840 observed in in vitro and in vivo models and its gametocytocidal effect. This validates that ACT-451840 might be a good candidate for TCP1 and TCP3b in humans [34]. The future steps will consist of the development of a lipid-based formulation, the selection of a combination partner, and the determination of the activity of ACT-451840 in patients.
|
10.1371/journal.ppat.1006677 | KSHV inhibits stress granule formation by viral ORF57 blocking PKR activation | TIA-1 positive stress granules (SG) represent the storage sites of stalled mRNAs and are often associated with the cellular antiviral response. In this report, we provide evidence that Kaposi’s sarcoma-associated herpesvirus (KSHV) overcomes the host antiviral response by inhibition of SG formation via a viral lytic protein ORF57. By immunofluorescence analysis, we found that B lymphocytes with KSHV lytic infection are refractory to SG induction. KSHV ORF57, an essential post-transcriptional regulator of viral gene expression and the production of new viral progeny, inhibits SG formation induced experimentally by arsenite and poly I:C, but not by heat stress. KSHV ORF37 (vSOX) bearing intrinsic endoribonuclease activity also inhibits arsenite-induced SG formation, but KSHV RTA, vIRF-2, ORF45, ORF59 and LANA exert no such function. ORF57 binds both PKR-activating protein (PACT) and protein kinase R (PKR) through their RNA-binding motifs and prevents PACT-PKR interaction in the PKR pathway which inhibits KSHV production. Consistently, knocking down PKR expression significantly promotes KSHV virion production. ORF57 interacts with PKR to inhibit PKR binding dsRNA and its autophosphorylation, leading to inhibition of eIF2α phosphorylation and SG formation. Homologous protein HSV-1 ICP27, but not EBV EB2, resembles KSHV ORF57 in the ability to block the PKR/eIF2α/SG pathway. In addition, KSHV ORF57 inhibits poly I:C-induced TLR3 phosphorylation. Altogether, our data provide the first evidence that KSHV ORF57 plays a role in modulating PKR/eIF2α/SG axis and enhances virus production during virus lytic infection.
| Mammalian RNA granules, including stress granules (SG), are important components of the host cell antiviral responses and their assembly is widely counteracted by RNA viruses. In Kaposi’s sarcoma-associated herpesvirus (KSHV) lytically infected B cells, we found that KSHV infection inhibits the assembly of SG by expression of viral lytic protein ORF57 and ORF37 (vSOX). KSHV ORF57 blocks arsenite-induced SG formation by binding to and preventing PACT from activating PKR. ORF57 also interacts with two double-stranded RNA binding motifs of PKR and prevents its binding with poly I:C and autophosphorylation, subsequent inhibition of eIF2α phosphorylation and SG formation. Consistently, knockdown of PKR increases production of KSHV virions. This function of KSHV ORF57 is conserved in homolog HSV-1 ICP27, but not in EBV EB2. We conclude that KSHV ORF57 antagonizes host antiviral defenses for virus lytic infection and production.
| Mammalian somatic cells produce two types of RNA granules, processing bodies (P-bodies, PB) and stress granules (SG) [1,2]. Both granules are physically and mechanistically distinct compartments with many unique biomarkers. While GW182 is confined to PB, RNA-binding proteins TIA-1, poly(A) binding protein (PABP) and G3BP are specific markers of SG. PB appear during normal cell growth and contain enzymes for RNA de-capping and degradation [1,3], and have been shown to store and degrade siRNA- or miRNA-guided mRNA [4,5]. SG on the other hand, lack de-capping/de-adenylating machinery and appear during cell stress to play a role in global translational arrest by storing mRNA [1]. Therefore, SG represent a central and dynamic warehouse where stored mRNA is protected and exchanged with polysomes or PB for further translation or degradation, respectively [3,6].
SG contain 40S ribosomal subunits, mRNAs, dozens of RNA-binding proteins and many translation initiation factors including eIF4G, eIF4E, eIF3, and PABP [3,7,8]. SG assembly is initiated by phosphorylation of the α subunit in eIF2 at a specific serine (Ser 51) residue [9]. eIF2 is a translation initiation factor which forms a ternary complex with GTP and the initiator methionine-tRNA (eIF2-GTP-tRNAi-Met) [10] and in turn loads the initiator tRNAmet onto the small ribosomal subunit [11–13]). Different types of stress (oxidative, heat, or nutrient deprivation) can induce eIF2α phosphorylation by activation of four different eIF2α kinases (GCN2, PKR, PERK, and HRI) [14,15]. Phosphorylation of heterotrimeric eIF2 on its regulatory α subunit increases its affinity with eIF2β (the subunit responsible for GTP binding) and thus reduces its availability for GTP exchange. This deficiency in GTP exchange inhibits the ability of eIF2 to reach its active GTP-bound state and therefore prevents ternary complex formation and arrests translation initiation [16]. Consequently, polysomes disassemble leaving translationally arrested mRNPs to be recognized by the RNA binding proteins, TIA-1 and TIAR, and sequestered through their prion-like aggregation property to nucleate SG [17,18]. Many RNA binding proteins including tristetrapolin (TTP), and fragile X mental retardation proteins (FMRP) also join the assembly in SG [19,20].
Virus infection imposes stress on multiple biosynthetic pathways in host cells, including translation [21], and captures the host translation machinery to ensure virus translation and production [22]. During infection, the presence of viral double stranded RNA (dsRNA) activates host cell PKR to induce eIF2α phosphorylation and SG formation and thereby, triggers the host cell antiviral response and shutting down host cell translation [23,24]. To bypass this response, however, many RNA viruses utilize alternative ways of translation [25]. Notably, several RNA viruses suppress the formation of SG by a viral factor [26–30]. Poliovirus C3 cleaves Ras-GAP (Ras GTPase activating protein) SH3 domain-binding protein (G3BP), a component of SG that initiates the assembly of SG and interacts with inactive PKR [28,31,32]. Semliki Forest virus nsP3 targets G3BP [33], and influenza virus NS1 inactivates dsRNA-activated PKR [26,27]. Hepatitis C Virus instead induces SG formation, but co-opts SG proteins for its replication and production [34,35]. The regulation of SG formation during infection with large DNA viruses is poorly understood although herpesviruses are proposed to produce viral proteins to regulate SG formation [33,36,37].
Kaposi’s sarcoma-associated herpesvirus (KSHV) is a γ-2 herpesvirus [38] and infects human B lymphocytes and endothelial cells. KSHV infection leads to development of Kaposi’s sarcoma, primary effusion lymphoma (PEL), and multicentric Castleman’s disease (MCD) [39,40]. Like other herpesviruses, KSHV infection undergoes two alternative life-cycle programs. Viral lytic infection is characterized with the expression of all viral genes to produce infectious virions; whereas latent viral infection features highly restricted expression of only a few viral genes. Although the underlying mechanism responsible for the switch between lytic and latent infection, or vice versa, remains an unresolved topic, it is known that both viral and host factors are involved in the shift of KSHV infection [41,42]. A viral replication and transcription activator (RTA or ORF50) is an immediately early protein and is essential for transactivation of almost all other viral genes in the lytic infection. RTA expression from the latent KSHV infection can be induced experimentally by chemicals such as sodium n-butyrate (Bu) [43] or valproic acid (VA) [44,45]. Another important KSHV protein is ORF57 (mRNA transcript accumulation or MTA) which is an early lytic RNA-binding protein responsible for posttranscriptional processing of viral transcripts and virus production [46,47]. ORF57 stabilizes viral RNAs [48–51]), promotes splicing of intron-containing viral mRNA [52,53], and inhibits miRNA function to promote viral gene expression [48,54] through its interactions with targeted RNA and numerous host factors [47]. In this report, we discovered a novel function of ORF57 to inhibit SG formation during KSHV lytic infection. Mechanistically, ORF57 directly interacts with PACT and PKR and prevents phosphorylation of PKR and eIF2α and, thereby, prevents both SG formation and the stalling of RNA translation.
To determine whether KSHV is able to modulate SG formation in infected cells, we employed a defined strategy (Fig 1A) to study SG formation in KSHV-infected BCBL-1 cells [55] and HEK293-derived Bac36 cells [45] by arsenite, a common chemical inducer which causes oxidative stress and robust formation of SG [10]. Both BCBL-1 and Bac36 cells harbor an episomal KSHV genome at the latent stage and can be reactivated to lytic KSHV infection in the presence of 1 mM Valproic acid (VA) or 3 mM Butyrate (Bu) (Fig 1A) [45,55,56].
On cells with or without virus lytic induction, we performed IF staining for SG-specific TIA-1, an RNA-binding protein that promotes the assembly of SG [18]. In the absence of arsenite treatment BCBL-1 cells either with latent or VA-induced lytic KSHV infection showed no visible SG (S1A Fig). In contrast, after arsenite treatment the majority of cells displayed ~3–6 TIA-1-positive SG per cell (Fig 1B). These arsenite-induced SG were also stained positive for PABPC1 and G3BP, two other SG-specific markers (S1B and S1C Fig). However, BCBL-1 cells expressing viral lytic genes, indicated by the presence of ORF57, exhibited a remarkable reduction of arsenite-induced SG (Fig 1B, S1B and S1C Fig), both by numbers of SG per cell and by numbers of cells with SG (Fig 1B and 1D bar graph). We also assessed the ability of Bac36 cells containing a wt KSHV genome (Bac36-wt) or ORF57-null KSHV genome (Bac36-Δ57) to prevent arsenite-induced SG formation during virus lytic infection, although viral lytic induction in Bac36 cells is less efficient than BCBL-1 cells [45]. We detected RTA in Bac36-Δ57 cells and ORF57 in Bac36-wt cells for butyrate-induced expression of viral lytic genes. While the ORF57 expressing Bac36-wt cells showed complete abrogation of arsenite-induced SG formation, the Bac36-Δ57 cells with RTA expression in lytic infection did not (Fig 1C and 1E bar graph), nor in LANA or ORF45-expressing cells (S1D Fig). The TIA-positive granules induced by arsenite in Bac36-Δ57 cells are bona fide SG and were sensitive to cycloheximide [6] by which blocks the flux of molecules between fully formed SG and polysomes (S1E Fig). These data indicate that, unlike latent infection, the presence of viral early protein ORF57, but not viral RTA, ORF45 or LANA during KSHV lytic infection is capable of preventing the arsenite-induced SG formation.
We also stained Bac36 cells (wt or Δ57) for TIA-1-positive SG in the absence of arsenite treatment, but in the presence of ectopic RTA expression to induce viral lytic infection. We did not see any visible SG in Bac36 wt or Δ57 cells (Fig 2A, control), unless the cells were treated with arsenite (Fig 2A, control). However, when cells are induced to the lytic phase by ectopic RTA expression, SG were found in approximate 20% of the RTA-expressing Bac36-Δ57 cells, but only in ~3% of the RTA-expressing Bac36-wt cells (Fig 2A for RTA activation Image and bar graph). Again, the RTA-induced SG formation in Bac36-Δ57 cells was sensitive to cyclohemixide treatment (S1F Fig) [6]. Importantly, the presence of ectopic RTA alone does not induce SG in KSHV-negative HEK293 cells, from where Bac36 cells were originated [45]. Collectively, these results indicate that KSHV lytic infection provides a stress to the infected cells and the expression of ORF57 but not RTA during virus lytic infection is required for suppression of SG formation.
The complex nature of lytic replication in KSHV makes it difficult to confirm that the observed suppression of SG formation is a result of ORF57 function. To examine whether ORF57 alone is sufficient to block SG formation in the absence of other viral lytic proteins, we transfected HeLa cells with an ORF57-expressing or an empty vector. Transfection of these plasmids did not induce SG formation as confirmed by staining for TIA-1 and PABPC1 (S2A Fig). We then treated the cells with arsenite for 30 min and performed similar immunostaining. As expected, all cells without ORF57 displayed SG positive for TIA-1, PABPC1 and G3BP1 staining, but a dramatic reduction in SG formation was found in ORF57-expressing cells (Fig 2B, S2B Fig), indicating that ORF57 is a viral lytic protein responsible for efficient suppression of SG formation.
There are three nuclear localization signals (NLS) present in the N-terminal domain of ORF57 that are important for the various functions of the viral protein. Introduction of point mutations in any two of the NLS renders ORF57 dysfunctional and unable to bind other partner proteins [57]. We next compared the ability of wild-type ORF57 (ORF57 wt) and an NLS mutant ORF57 (ORF57 mt) protein to abrogate arsenite-induced SG formation in HeLa cells. KSHV ORF59, a DNA polymerase processivity factor, served as another viral lytic protein control. While ORF57 wt showed nearly complete abrogation of SG formation, ORF57 mt and ORF59 did not (Fig 2C, panels on the left). By quantification, we measured ~20 SG per cell in all cells transfected with an empty vector, ORF57 mt, or ORF59, but ~80% of the ORF57-expressing cells did not exhibit any SG and the remaining ~20% of the cells showed a dramatically reduced number of SG (2–5 SG/cell) (Fig 2C, bar graphs on the right). Three-dimensional image reconstructions acquired by confocal microscopy verified the absence of SG in ORF57 expressing cells (S1 & S2 Videos). As an initial investigation into how ORF57 may prevent SG formation, we checked whether ORF57 affects the expression of TIA-1, PABPC1, G3BP, eIF4E and its phosphorylated form (p-eIF4E) and found no difference in the protein levels for all of these SG components, compared to control cells (S2C–S2E Fig).
During stress, TIA-1 aggregates in a similar manner to prion-like proteins to form SG [18]. The concentrated TIA-1 in SG can be detected in the insoluble pellets after high speed centrifugation of cellular extracts (S3A Fig). As expected, arsenite treatment of HeLa cells led to remarkable increase of TIA-1 in the pellet (Fig 2D and S3B Fig) in a time-dependent manner (Fig 2E and S3C Fig). Of particular interest, ORF57 wt, but not ORF57 mt, could be found in the pellets (S3B Fig) and prevent TIA-1 recruitment into the pellets (Fig 2D and S3C Fig). We calculated that ORF57 wt blocked ~75% of total TIA-1 from being aggregated in 30 min of arsenite treatment (Fig 2E) and this function of ORF57 begins even at 10 min of arsenite treatment of HeLa cells in this study (Fig 2E and S3C Fig). All together, these results indicate that a novel function of ORF57 is to establish the conditions that maintain the solubility of TIA-1 to prevent SG formation during stress.
The formation of SG is initiated as a downstream event after elF2α phosphorylation which leads to the prion-like aggregation of TIA-1. Normally, eIF2α is required to initiate mRNA translation by promoting the binding of tRNAmet to the 40S ribosome in a GTP-dependent manner. Stress induces phosphorylation of eIF2α to attenuate eIF2α activity (Fig 3A) and thereby promote TIA-1 aggregation to form the SG where mRNA translation is stalled [14].
Accordingly, we examined whether ORF57 could affect eIF2α phosphorylation. HeLa cells transfected with an ORF57-expressing or empty vector and treated with arsenite were blotted for the phosphorylation status of eIF2α (p-eIF2α). As expected, arsenite was found to induce eIF2α phosphorylation ~15 times greater than basal level (Fig 3B, lane 3 vs lane 1), whereas ORF57-expressing cells remarkably inhibited eIF2α phosphorylation upon arsenite induction (Fig 3B, lane 4 vs lane 3, and Fig 3C), with the total level of eIF2α protein remaining the same.
To correlate kinetic production of KSHV ORF57 with both total eIF2α and phosphorylated eIF2α, we induced KSHV-infected BCBL-1 cells with 1 mM VA for the indicated time and then treated the cells with arsenite for 30 min before collecting cell lysates for Western blotting analysis. Although lytic KSHV infection did not increase eIF2α phosphorylation over that of latent KSHV infection in the cells without exogenous stress (Fig 3D and 3E), we found that arsenite-induced phosphorylation of eIF2α was in reverse correlation with kinetic ORF57 expression (Fig 3F and 3G). Higher ORF57 expression resulted in less eIF2α phosphorylation, while the total eIF2α remained unchanged (Fig 3F and 3G). Altogether, these data indicate that ORF57 is inhibitory for eIF2α phosphorylation both when expressed alone and when present with other viral proteins during viral lytic infection.
During cellular stress eIF2α can be phosphorylated by four different kinases, and which kinase is activated depends on the cause of stress (Fig 4A) [14]. Of particular interest, both viral infection [24,58] and arsenite [59] commonly activate PKR although GCN2 could be activated by Sindbis virus in a report [60]. PKR is well known for its antiviral activity by induction of interferon and is both a cytoplasmic and a nuclear dsRNA-binding protein [61,62]. We confirmed that arsenite did induce phosphorylation of PKR and SG formation in both HeLa and BCBL-1 cells, which could be specifically blocked by a PKR inhibitor (S4A–S4D Fig). These observations exclude the possibilities of other three pathways being involved in the studied SG formation in this report, although arsenite was also reported to mediate eIF2α phosphoryaltion through HRI in erythroid cells [63] and mouse fibroblast cells [64]. Virus infection activates PKR through the binding of viral dsRNA to the dsRNA-binding domain (RBD) of PKR, whereas arsenite activates PKR by inducing PACT to bind with and activate PKR [65,66]. PKR contains at least 15 autophosphorylation sites, but phoshorylation at both Thr 446 and Thr 451 is critical for its activation, and subsequent phosphorylation of eIF2α [67,68].
To elucidate the mechanism of ORF57-mediated inhibition of eIF2α phosphorylation, we investigated whether ORF57 could inhibit PKR activation and phosphorylation in HeLa cells under three different stress conditions (arsenite, double-stranded poly I:C, or heat stress) (Fig 4B). An optimized poly I:C dose was used to mimic dsRNA [69] to specifically activate PKR-mediated eIF2α phosphorylation (S5A Fig). Heat treatment at 44°C for 40 min served as an alternative route of cell stress resulting in eIF2α phosphorylation. As expected, both arsenite and poly I:C induced PKR phosphorylation along with eIF2α phosphorylation (Fig 4B, compare lanes 1 vs 3 and 5 vs 7). However, heat shock induced eIF2α phosphorylation without PKR phosphorylation. Thus, the heat shock-induced phosphorylation of eIF2α is not related to PKR activity (Fig 4B, compare lanes 9 vs 11) as has been reported in reticulocytes [63]. We did not see any induction in phosphorylation of PERK (PKR-like endoplasmic reticulum kinase) by arsenite (S5B Fig). Interestingly, we observed a dramatic reduction (~75%) of both phosphorylated PKR (p-PKR) and p-eIF2α when ORF57-expressing cells are exposed to either arsenite or poly I:C (Fig 4B, compare lanes 3 vs 4 and 7 vs 8, and Fig 4C), with only minimal or no change in overall PKR and elF2α protein levels. In contrast, ORF57-expressing cells exhibited no effect on the phosphorylation of eIF2α induced by the heat shock (Fig 4B, compare lanes 11 vs 12, and Fig 4C), but a dose-dependent, increased expression of ORF57 exhibited a steady decrease in PKR phosphorylation (Fig 4D and 4E). The ORF57 prevention of arsenite- and poly I:C-induced PKR phosphorylation with no change in total PKR levels was also observed in HEK293 cells (S5C Fig), indicating that the inhibitory effect of PKR phosphorylation by ORF57 is not cell-specific and could take place as early at 15 min of arsenite treatment, the earliest time point of the sample collection in this study (S5D Fig). Moreover, ORF57 was found to block the phosphorylation of TLR3 (toll-like receptor 3) induced by dsRNA (poly I:C) (S5E Fig) to activate Interferon regulatory factor 3 (IRF3) and production of type 1 interferon [70–73].
By IF microscopy, we found that both poly I:C and heat stress induce SG formation in HeLa cells (Fig 4F and S6 Fig), however, ORF57 was found to inhibit SG formation when induced either by arsenite or by poly I:C, but not when induced by the heat stress (Fig 4F and 4G and S6 Fig). The functional relevance of how ORF57 modulation of SG-formation impacts KSHV gene expression and replication was further explored in BCBL-1 cells. Because ORF57 does not affect heat shock-induced SG formation in BCBL-1 cells with lytic KSHV infection (Fig 4H), the effect of heat-induced SG formation on the expression of viral RTA, a viral replication and transcription activator and ORF45, a viral tegument protein, was investigated in the VA-induced cells. We found that 40-min heat shock-induced SG formation showed no effect on their RNA levels by RT-qPCR, but led to reduction of RTA protein expression by 30–40% in VA-induced BCBL-1 cells. The 40-min heat-shock had little effect on the expression of ORF45, a less-sensitive viral early gene to RTA transactivation [74] and a relative stable and abundant protein [75] (Fig 4I). Together with our previous findings [47], these results reveal that ORF57 inhibits PKR activation, disrupts the PKR-mediated phosphorylation of eIF2α and, therefore, blocks SG formation to promote viral gene expression.
Given that both viral infection [24,58] and arsenite treatment [66] activate PKR and induce SG formation, and arsenite activates PKR by inducing PACT to bind PKR [65,66], we examined the mechanism by which ORF57 inhibits arsenite/poly I:C activation of PKR. By co-immunoprecipitation (co-IP) in combination with Western blot analysis we found that ORF57, but not the ORF57-mt, interacts with both PKR and PACT independent of RNA (Fig 5A–5C). Interestingly, the interaction of ORF57 individually with either PKR or PACT disrupts the interaction between PACT and PKR. As shown in Fig 5B and 5C, various co-IP experiments using an anti-PKR antibody (Fig 5B) revealed a remarkable reduction in the amount of PACT associated with PKR when ORF57 is present. Similarly, using an anti-PACT antibody (Fig 5C), we observed a significant reduction in the association of PKR with PACT in the presence of ORF57. This mechanistic function of KSHV ORF57 in blocking SG formation resembles that of the TRBP-PACT interaction and inhibition of PKR activation [76], but differs from other viruses that rely on cleavage or direct interaction with G3BP to block SG formation [28,33]. In this regard, ORF57 did not interact with G3BP (Fig 5D), nor alter G3BP expression (S2D Fig). In addition, ORF57 failed to interact with eIF2α (Fig 5E), TIA-1 or eIF4G1 (Fig 5F), although ORF57 did interact with PABPC1 [50] (Fig 5D) and eIF4E (Fig 5F), two common components of SG. Importantly, we further confirmed by co-IP and Western blotting that ORF57 interacts with both PACT and PKR in BCBL-1 cells during viral lytic infection (Fig 5G). Altogether, these studies indicate that PACT, PKR, PABPC1 and eIF4E are ORF57-interacting proteins and ORF57 binds to PACT and PKR and blocks PKR activation.
PACT contains two RNA-binding motifs (RBM) in its N-terminal half and a PKR-activation domain (PAD) in its C-terminal half. To determine the specific domain of PACT interacting with ORF57, the cell lysates containing individual Flag-PACT deletion mutants were mixed with the cell lysates containing ORF57 protein and followed by RNase A/T1 treatment to avoid any possible RNA-mediated protein-protein interaction in the subsequent anti-Flag antibody co-IP for PACT-associated ORF57 or anti-ORF57 antibody co-IP for ORF57-associated PACT. Western blot analysis of the proteins pulled down by the co-IP revealed that ORF57 interacts with PACT through its two RBM motifs and this interaction is independent of RNA. As shown in Fig 6, deletion of either RBM1 (PACT-Δ1) or RBM2 (PACT-Δ2) from PACT significantly reduced the binding of PACT to ORF57 (Fig 6B and 6C, compare lanes 7 and 12 to lanes 8–9 and 13–14), but deletion of the PAD (PACT-Δ3) did not (Fig 6B and 6C, compare lanes 7 and 12 to lanes 10 and 15). This study also demonstrated that further deletion of both RBM1 and RBM2 from PACT (PACT-Δ1,2) completely prevented the binding of PACT to ORF57 (Fig 6D, compare lane 9 to lanes 7–8 and 10). Based on these data, we concluded that ORF57 interacts with PACT via its two RBM motifs.
PKR has a N-terminal regulatory domain containing two dsRNA-binding motifs (RBM1 & RBM2) and a C-terminal kinase domain (Fig 7A). To determine the specific domain of PKR interacting with ORF57, we generated two deletion mutants of PKR either by deletion of the kinase domain (ΔPK) or by deletion of the dsRNA-binding domain (ΔRBM). Both deletion mutants and the full-length (FL) PKR have a chimeric Myc-Flag tag and were individually expressed in HEK293 cells separately from untagged ORF57. By mixing the PKR cell extract with the ORF57 cell extract, followed by RNase A/T1 treatment to avoid any possible RNA-mediated protein-protein interaction in the subsequent anti-Flag or anti-Myc co-IP for PKR-associated ORF57, we found that the FL PKR and the N-terminal RBM domain (ΔPK), but not the kinase domain (ΔRBM) interact with ORF57 (Fig 7B, compare lanes 6–7 to 8 for ORF57). Interestingly, the ΔPK exhibited ~6.5-fold greater binding to ORF57 than did FL PKR (Fig 7B, compare lanes 7 to 6). Moreover, we found that the phosphorylated FL PKR increased its binding capacity toward ORF57 ~3 times more when activated in cells treated with arsenite (Fig 7C, compare lanes 6 to 5). Overall, we find that ORF57 interacts with the N-terminal RBM-containing domain of PKR, and the conformational change that occurs in PKR during activation allows ORF57 to bind to the N-terminus of PKR with greater affinity.
PKR activation depends on its binding to dsRNA via its two RBMs. This interaction induces PKR dimerization at the C-terminal kinase domain, which in turn leads to autophosphorylation of PKR (Fig 8A). Once phosphorylated, each subunit kinase domain in the dimerized PKR can independently phosphorylate the substrate eIF2α [14,77] (Fig 8A). To investigate the mechanism of how the ORF57-PKR interaction prevents phosphorylation of PKR and eIF2α, we first performed an in vitro competitive binding assay by which Myc-Flag-tagged PKR and its ΔRBM mutant immobilized separately on the anti-Myc beads were compared for competitive binding with 32P-poly I:C and recombinant ORF57. To do this, 32P-poly I:C was first mixed with recombinant ORF57 or BSA before allowing to interact with the immobilized PKR. As shown in Fig 8B, we observed that while BSA didn’t compete with 32P-poly I:C to interact with PKR, the recombinant ORF57 protein significantly reduced this interaction to ~50%. The PKR ΔRBM mutant and Flag control both served as negative controls, and ORF57 had no effect on their basal level of binding to 32P-poly I:C. These data suggest that ORF57 interacts with the N-terminal domain of PKR to prevent PKR binding to dsRNA.
We next examined that ORF57 prevention of PKR binding to dsRNA might affect PKR autophosphorylation. Subsequently, an in vitro autophosphorylation assay was conducted by incubation of recombinant ORF57 or BSA (a negative control) with immobilized PKR beads first before adding poly I:C and [γ-32P]-ATP. By examining the 32P-labelled PKR, we demonstrated that poly I:C did stimulate PKR autophosphorylation in the absence or presence of BSA, but this poly I:C induction of PKR autophosphorylation could be reduced by ~50% in the presence of ORF57 (Fig 8C).
In a separate experiment, we also tested whether ORF57 prevents the phosphoryated-PKR (p-PKR) from in turn phosphorylating its substrate eIF2α. To do so, we performed an in vitro kinase assay using a GST-eIF2α. The Myc-Flag-tagged PKR expressed in HeLa cells treated with or without arsenite was immobilized on anti-Myc beads and used to phosphorylate GST-eIF2α in the presence of recombinant ORF57 or BSA in a kinase reaction containing [γ-32P]-ATP]. By examining the amount of 32P-labelled GST-eIF2α, we found that while the inactive PKR from the cells without arsenite treatment did not exert much kinase activity on GST-elF2α, the arsenite-activated p-PKR did actively phosphorylate GST-eIF2α equally well both in the presence of BSA or ORF57 (Fig 8D). The inability of ORF57 to inhibit phosphorylation of GST-elF2α was not due to a lack of interaction between ORF57 and the active p-PKR. In fact, ORF57 was found to associate more efficiently with the activated p-PKR (Fig 7C). From these results presented above, we conclude that ORF57 interacts with the RBM motifs of PKR and prevents PKR binding to dsRNA or PACT and PKR activation by autophosphorylation, consequently preventing eIF2α phosphorylation and SG formation. However, once PKR is phosphorylated, ORF57 is unable to prevent p-PKR from phosphorylating eIF2α by its C-terminal kinase domain.
In the course of drafting our manuscript for publication, Finnen and colleagues reported that viral vhs protein encoded by UL41 in herpes simplex virus type 2 (HSV-2) suppresses SG formation [78]. Although α-herpesvirus vhs and γ-herpesvirus vSOX are not homologs, both are RNA endonucleases and exert their host shutoff function by digesting host mRNAs [79–82] which are fundamental for SG formation [8]. In assumption of KSHV vSOX encoded by viral ORF37 in blocking SG formation via a mechanism similar to HSV-2 vhs [78], we transfected both HEK293 and HeLa cells with a Flag-tagged KSHV vSOX expression vector for 24 h and stained the cells for TIA-1-specific SG formation in the presence or absence of vSOX after induction by arsenite for 30 min. We confirmed vSOX expression in HEK293 cells by anti-Flag antibody staining of HEK293 cells (Fig 9A) and by anti-Flag antibody Western blot (Fig 9B), but had difficulty to express vSOX in HeLa cells. As expected, HEK293 cells expressing no vSOX displayed SG formation induced by arsenite, but all cells expressing vSOX did not (Fig 9A). In contrast to ORF57, vSOX in transfected HEK293 cells had no effect on arsenite-induced phosphorylation of eIF2α. These results clearly indicate that KSHV vSOX inhibits SG formation in HEK293 cells, similar to HSV-2 vhs [78] by degradation of RNA [80–82].
Considering that KSHV interferon regulatory factor 2 (vIRF-2) might play an inhibitory role in PKR activation and PKR-mediated phosphorylation of eIF2α [83], we compared ORF57 with vIRF-2 in regulation of PKR activation and phosphorylation of eIF2α in arsenite-treated HeLa cells. Because the actual vIRF-2 ORF splits into two separate exons, with exon 1 in K11.1 and exon 2 in K11 [84] and encodes a full-length vIRF-2 having 680 aa residues, the annotated vIRF-2 ORF encoding 163 aa residues in an early report [83] based on initial ORF annotation in the KSHV genome [85,86] was not an authentic vIRF-2 ORF. Thus, we cloned a full-length vIRF-2 encoding 680 aa residues and expressed as a Flag-vIRF-2 in HeLa cells with or without arsenite treatment. As shown in Fig 9C, vIRF-2 was expressed predominantly as a cytoplasmic protein, but exhibited no effect on arsenite-induced SG formation. When compared with ORF57, we found that vIRF-2 in HeLa cells did not inhibit arsenite-induced PKR phosphorylation, nor eIF2α phosphorylation (Fig 9D, compare lane 4 to lane 3). In contrast, ORF57 at the same condition displayed the expected inhibition on phosphorylation of both PKR and eIF2α (Fig 9D, compare lane 6 to lane 2). According to these results, we conclude that the full-length vIRF-2 which modulates the host antiviral response [87–89] has no inhibitory function in activation of PKR pathway.
Previously, we and others demonstrated that KSHV ORF57 is essential for KSHV replication and virus production [45,90]. Our observations in this study showed that KSHV ORF57 inhibits PKR activation and disrupts the PKR-mediated phosphorylation of eIF2α to block SG formation (Fig 4B–4G). A well-recognized outcome of SG formation is to trigger the host cell antiviral response and inhibit virus production [91]. Therefore, we postulate that PKR might be a host inhibitory protein to block KSHV production and therefore, one of the ORF57 functions in blocking PKR activation and SG formation is to promote KSHV gene expression and virus production. To confirm this hypothesis, we examined KSHV virion production in a newly established iSLK-BAC16 cell line [92] with or without siRNA-mediated PKR knockdown. By using a PKR-specific siRNA, we found that efficient knockdown of PKR expression from iSLK-BAC16 cells (Fig 10A) resulted in significantly increased production of KSHV virions and led the iSLK-BAC16 culture supernatants being highly infectious for HEK293 cells (Fig 10B). Quantitative analyses by flow cytometry indicate that siRNA knockdown of PKR expression in iSLK-BAC16 cells led to ~78-fold increase of KSHV virion production over the cells with the normal level of PKR expression (Fig 10C).
Given that HSV ICP27 and EBV EB2 (SM) proteins are homologues to ORF57, we reasoned that the ability to block the PKR/eIF2α/SG axis by KSHV ORF57 might be a conserved function in other herpesviruses. To investigate this possibility, we expressed ICP27 from HSV-1 and EB2 from EBV in HeLa cells and examined their influence on SG formation and phosphorylation of both PKR and eIF2α. The expression of ORF57, ICP27 or EB2 protein failed to induce SG in HeLa cells, but the three homologues exhibited a functional disparity in cells treated with arsenite. While both ORF57 and ICP27 abrogated the formation of SG in ~85% of arsenite-treated cells (Fig 11A and 11B), EB2 did not exert such an inhibitory function on SG formation (Fig 11A). By Western blotting of the lysates prepared from untreated or arsenite-treated cells transfected with an empty control vector or a vector expressing individual viral proteins, we found that both ICP27 and ORF57, but not EB2, inhibited the phosphorylation of ~70% PKR and ~75% eIF2α (Fig 11C, lanes 6–7 vs lanes 5–8 and Fig 11D), along with only a minimal effect on total PKR or eIF2α protein levels (Fig 11C and 11D). Altogether, these data indicate that the ability to inhibit SG formation in herpesviruses is conserved in HSV-1 through the ORF57 homologue ICP27.
Virus infection inevitably induces host cell stress. Thus, SG formation in the infected host cells has been widely appreciated as an antiviral defense mechanism [91,93]. SG formation is a downstream event of eIF2α phosphorylation which stalls translation initiation [10,94]. Of the four cellular kinases which can phosphorylate eIF2α and induce SG formation [77], PKR is a major player during viral infection. Viral dsRNA activates PKR and, therefore, induces SG formation. Although counteracted by many RNA viruses, regulation of the PKR/eIF2α/SG pathway by DNA viruses is poorly understood. Besides its activation by viral dsRNA, PKR is activated by PACT during arsenite stress [65,66]. Here, we utilized arsenite stress to explore the differential ability of KSHV infection to block SG in B cells and HEK293-derived Bac36 cells with a wt or ORF57-null (Δ57) KSHV genome. We found that KSHV, a DNA virus expressing the viral early protein ORF57, confers the infected B cells and Bac36-wt cells refractory to SG induction during lytic infection. Viral ORF57 alone blocks activation and phosphorylation of PKR and thereby SG formation. KSHV vSOX bearing intrinsic endoribonuclease activity also affects SG formation by degrading RNA. Other KSHV-encoded proteins examined in this study, such as RTA, ORF45, ORF49, LANA and vIRF-2, have no such function.
TIA-1 protein is a robust marker of SG and the N-terminal RRM domain TIA-1 binds to targeted RNA transcripts that have an AU-rich or C-rich element in the 5’ or 3’ UTRs [95–97]. TIA-1 nucleates SG formation via its C-terminal glutamine-rich prion-related domain (PRD) responsible for self-association [18]. Consistently, we found TIA-1 enrichment in the insoluble cell pellets during arsenite treatment, but this enrichment is prevented when ORF57 is present. The ORF57-mediated reduction of TIA-1 in the pellet is most likely a consequence of ORF57 inhibiting SG formation, but why a proportion of ORF57 resides in the insoluble pellets remains unknown, presumably being associated with ribosomes, microtubes, or other cellular debris. ORF57 does not interact directly with TIA-1 or G3BP1, another SG nucleator in addition to TIA-1 [98] and therefore, must indirectly influence the biochemistry of TIA-1 and G3BP1. Unlike TIA-1 and G3BP1, ORF57 does bind to PABPC1 and eIF4E, two important components of SG. ORF57 interaction with PABPC1 reduces the cytoplasmic pool of PABPC1 by promoting the redistribution of PABPC1 to the nucleus [50]. Although an important component of SG, PABPC1 is viewed as a passenger and does not function directly in SG formation [10,99], but may influence indirectly the ability of TIA-1 and G3BP1 to form SG. We also find that ORF57-mediated inhibition of SG formation is accompanied by a significant reduction in the amount of cytoplasmic PABPC1. PABPC1 binds to the 3’ poly(A) tail of eukaryotic mRNAs, and its interaction with the N-terminus of eIF4G stabilizes RNA and promotes both ribosome recruitment and translation initiation [100–102]. Whether the reduction in cytoplasmic PABPC1 mediated by ORF57 affects recruitment of the polyadenylated RNA transcripts by TIA-1 into SG needs to be investigated. Cap-binding protein eIF4E is a translation initiation factor which binds mRNA’s 5’ cap and mediates the cap structure of mRNA directly binding to the 40S ribosomes. Similar to PABPC1, eIF4E could be a passenger protein and does not function directly in SG formation, although ORF57 interaction with eIF4E, but not with eIF4G1, might be involved in regulation of protein translation initiation. The inability of ORF57 to block heat stress-induced SG formation indicates that the ORF57 interaction with PABPC1 and eIF4E is insignificant with regards to SG formation, at least under conditions of heat shock-induced cell stress.
KSHV ORF57 is an RNA binding protein and a posttranscriptional regulator of viral RNA transcripts [47]. Our finding that ORF57 functions as an inhibitor of PKR/eIF2α phosphorylation is intriguing and surprising. Further characterization of this inhibitory function of ORF57 led us to discover that KSHV ORF57 interacts with both PACT and PKR via their RBM motifs to prevent activation and phosphorylation of PKR and thereby, to inhibit eIF2α phosphorylation and SG formation (Fig 12). PKR is a dsRNA-binding protein important for the antiviral action of IFN and is a major cellular kinase that controls translation by phosphorylation of eIF2α [14]. We demonstrated that knocking down PKR expression in iSLK-BAC16 cells significantly promoted KSHV lytic infection and virion production. PKR is activated during viral infections [24,58] and experimentally by arsenite treatment [59]. The activation mechanisms are slightly different in that, viral infections activate PKR through the binding of viral dsRNA to the RNA-binding motif of PKR [103–105], whereas arsenite activates PKR via PACT, a PKR activating protein which heterodimerizes with PKR and activates PKR in the absence of dsRNA [65,66]. When we experimentally induce PKR phosphorylation by either arsenite treatment or incubation with poly I:C, the presence of ORF57 blocks PKR phosphorylation by binding directly to both PACT and PKR leading to a reduction in PACT-PKR and dsRNA (poly I:C)-PKR interactions. Other viral proteins, including TRS1 of cytomegalovirus [106] and UL41 of HSV-2 [36,37], exhibit a similar function in blocking PKR phosphorylation and activation to block SG formation. Even though ORF57 interacts with the N-terminal RBM domain of PKR and blocks it’s binding to dsRNA and its autophosphorylation, it has no direct effect on phosphorylation of eIF2α once PKR is activated and autophosphorylated and does not interact with eIF2α. The enhanced interaction of ORF57 with the p-PKR might be a result of conformational changes in p-PKR. This mechanistic function of ORF57 resembles that of TRBP which inhibits PKR activity through the interaction with PKR RBMs [76,107], but differs from poliovirus 3C protease which cleaves G3BP1 [28], SFV nsP3 and HSV-2 ICP8 which suppress SG formation by their FGDF motifs interacting with G3BP1 [33], and HSV-2 vhs which requires its endoribonuclease activity in disruption of SG formation [78]. KSHV ORF57 does not have a FGDF motif or interact with G3BP1 and bears no endoribonuclease activity. Although α-herpesvirus vhs and γ-herpesvirus SOX are not homologs, both are RNA endonucleases that digest host mRNAs [79,81]. Thus, it is not to our surprise that KSHV vSOX is capable of blocking SG formation by a mechanism similar to HSV-2 vhs [78]. Moreover, poxviruses induce a SG-like antiviral granule formation which does not entirely depend on PKR or eIF2α [108,109] and can be induced in the cells lacking eIF2α [109] or TIA-1 [110], although vaccinia virus E3L protein antagonizes PKR function and blocks the antiviral granule formation [110].
The discrepancy between EBV EB2 (SM) and both KSHV ORF57 and HSV-1 ICP27 in regulating the phosphorylation of PKR and eIF2α is another interesting result described in this report. KSHV ORF57, HSV-1 ICP27 and EBV EB2 are three well-known homologous proteins in regulation of viral RNA biogenesis at various lytic stages of viral infections. Although ORF57 deviates from ICP27 in protein structure and the function of RNA splicing, EBV EB2 is a more closely matched homologue to ORF57, with many functional similarities [47]. Here we show that both ICP27 and ORF57, but not EBV EB2, inhibits arsenite-induced SG formation by blocking phosphorylation of PKR and eIF2α (Fig 11). EBV EB2 was described as a dsRNA-binding protein and inhibits PKR activation via its RXP triplet repeats [111], but this domain with the RXP triplet repeats doesn’t exist in EB2 homologs and the function of EB2 in promoting protein translation appears independent of the PKR pathway [112]. ICP27 has not been characterized as a protein being able to block SG formation, but does appear in spontaneous SG during virus infection [36,37] and inhibits IFN signaling [113]. Our data show an extreme similarity of HSV-1 ICP27 with KSHV ORF57 in preventing SG formation by blocking phosphorylation of PKR and eIF2α, providing evidence for ectopic ICP27 being directly involved in regulating the PKR pathway in the absence of HSV infection.
KSHV encodes many proteins that function to evade the host immune system in multiple ways. Most of these viral proteins are homologous to cellular proteins and interfere with both innate and adaptive immune responses. In particular, immune evasion of the interferon pathway and the TLR pathway by KSHV vIFNs and other host mimics is one of the important strategies for KSHV to escape from the host innate immune response [114]. In this regard, KSHV vIRF-1, -2, or -3 blocks TLR3-mediated activation of IFN-responsive promoter activity [73] and vIRF1 decreases phosphorylation and nuclear translocation of IRF-3 in response to TLR3 activation [73]. Although a non-existing short-form vIRF-2 of KSHV was found to interact directly with PKR and inhibit PKR autophosphorylation and eIF2α phosphorylation [83], the full-length vIRF-2 in our study does not exhibit such a function or inhibition of SG formation. In addition, we found ORF57 also blocks poly I:C-induced phosphorylation of TLR3. Together with the finding that KSHV ORF57 regulates PKR pathway to suppress SG formation and PKR inhibits KSHV production, our data provide the first evidence that KSHV ORF57 plays a critical role in modulation of PKR/eIF2α/SG axis to enhance KSHV lytic infection.
Human HEK293 and HeLa cells (ATCC, Manassas, VA) were cultivated in DMEM (Thermo Fisher Scientific, Waltham, MA) supplemented with 10% fetal bovine serum (FBS, GE Healthcare, Logan, UT). Primary effusion lymphoma BCBL-1 cells (KSHV+) [55] obtained from the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH, were grown in RPMI 1640 (Thermo Fisher Scientific) containing 10% FBS. KSHV lytic infection in BCBL-1 cells was reactivated by 1 mM sodium valproate (cat. no. P4543, Sigma-Aldrich, St. Louis, MO) for 24 h. HEK293-derived Bac36 cell lines stably harboring a wt KSHV genome (Bac36-wt) or an ORF57-null KSHV genome (Bac36-Δ57) were established in our lab as described [45]. KSHV lytic infection in Bac36 cells was reactivated with 3 mM sodium butyrate (cat. no. B5887, Sigma-Aldrich) for 24 h. All plasmid transfections were performed using LipoD293 transfection reagent (SignaGen Laboratories, Gaithersburg, MD) according to the manufacturer’s instruction. Unless indicated, for IFA and Western analysis, HeLa (2.5 ×105) and HEK293 (5 × 105) cells were plated a day prior to 1 μg (2 μg for vSOX) of plasmid DNA transfection in a 6-well plate. For IP, HeLa (2 × 106) and HEK293 (5 × 106) cells were plated a day prior to 5 μg plasmid DNA transfection in a 10-cm Petri dish.
The custom made rabbit polyclonal and mouse monoclonal anti-ORF57 antibodies were described earlier [45]. Mouse monoclonal and rabbit anti-RTA and mouse monoclonal anti-ORF45 antibodies were kindly provided by Drs. Yoshi Izumya and Fanxiu Zhu, respectively. Mouse monoclonal anti-β-tubulin (cat. no. T5201), anti-Flag M2 (cat. no. F1804), anti-c-Myc (9E10, cat. no. M4439) and rabbit polyclonal anti-Flag antibody (cat. no. F7425) were obtained from Sigma-Aldrich. Rabbit polyclonal antibodies anti-eIF2α (cat. no. 9722S) and anti-phospho-eIF2α (ser51) (cat. no. 9721S) and rabbit monoclonal anti-GAPDH (cat. no. 2118), anti-phospho-PERK (Thr 980) (16F8, cat. no. 3179) and anti-PERK (C33E10, cat. no. 3192) were obtained from Cell Signaling Technology. Other antibodies used were: rabbit polyclonal anti-TLR3 antibody (phospho-Tyr759) (cat. no. LS-C19344, LifeSpan BioScience Inc. Seattle, WA), anti-PABPC1 (cat. no ab21060, Abcam, Cambridge, MA), anti-PACT/PRKRA (cat. no 10771-AP, ProteinTech group, Rosemont, IL), anti-phospho (Ser 209)-eIF4E (cat. no. ab47605, Abcam), and anti-phospho PKR (pThr451) (cat. no. 527460, EMD Millipore, Billerica, MA); mouse monoclonal anti-eIF2AK2 (PKR) (cat. no. H000005610-M01, Abnova), anti-eIF4E (cat. no. 610269, BD Biosciences, San Jose, CA), anti-PACT/PRKRA (cat no. H00008575-M01, Abnova, Taipei, Taiwan), and anti-eIF4G1 (cat. no. ab54970, Abcam); goat polyclonal anti-TIA-1 (cat. no. sc-1751, Santa Cruz Biotechnology, Dallas, Texas), and rat monoclonal anti-HHV-8 LANA (cat. no. MABE1109, EMD Millipore). The peroxidase-conjugated secondary antibodies used in Western blotting were obtained from Sigma-Aldrich and all Alexafluor-conjugated secondary antibodies used in IFA were purchased from Thermo Fisher Scientific (Waltham, MA). Sodium arsenite (cat. No 38150), valproic acid sodium salt (cat. no. P4543), sodium butyrate (cat. no. B5887) and cycloheximide (cat. No. C-7698) were obtained from Sigma-Aldrich. PKR inhibitor (PKRi, cat. no. 527451) and its negative control (PKRc, cat. no. 527455) were obtained from EMD Millipore.
The expression vectors were used to express recombinant proteins: KSHV ORF57 (pcDNA-ORF57), ORF57-FLAG (wt, pVM7; mt NLS 2+3, pVM89), ORF57-GFP (wt, pVM8; mt NLS 2+3, pVM36), KSHV ORF59-FLAG (pVM18), pKY15 (HSV1-ICP27-FLAG) and pGS113 (myc-EBV-EB2) [48,52,57]. Empty pcDNA 3.0 and pCMV-FLAG 5.1 (Sigma-Aldrich) were used as a negative control. KSHV vSOX (ORF37) ORF was amplified by PCR using a primer pair of oVM400 (5’-CCGGAATTCACC/ATGGAGGCCACCCCCACAC-3’, nt 57273–57291) and oVM401 (5’-ACTGTCTAGA/CGGGCTGTGAGGGACGTTTG-3’, nt 57291–57273) on total DNA from BCBL-1 cells. The resulting PCR product was cloned into pFLAG-CMV-5.1 vector (Sigma-Aldrich) via EcoRI and XbaI sites to create plasmid pVM116. The identity of inserted ORF was verified by Sanger sequencing. The expression of vSOX-Flag fusion protein (486 aa + Flag tag, ~52 kDa) was confirmed by Western blot using anti-Flag antibody upon transfection in HEK293.
Full-length Protein Kinase R encoding plasmids (p-CMV-Entry, PKR-Myc-Flag or pPKR-FL) (cat. no., RC210792) was obtained from Origene (Rockville, MD). PKR deletion mutants (ΔRBM, aa 1-12/153-551) was generated by overlapping PCR using following primers: oVM78 5’-CCGTTGACGCAAATGGGC-3’ and oNS1 (5’-TGCTTCCTGTTT/CTCCATGAAGAAACCTGC-3’; PKR nt 1487-1476/1064-1047) for first PCR and oNS2 (5’-TTCTTCATGGAG/AAACAGGAAGCAAAACAATT-3’; PKR nt 1053-1064/1476-1495) and oNS3 (5’-CATCACTGGTCTCAGGATC-3’; PKR nt 2043–2025) for second PCR. Obtained two PCR products were combined and re-amplified with oNS3 and oVM78. The final PCR product was used to replace full length PKR using Asp718 and BclI sites in pPKR-FL plasmid. The PKR ΔPK mutant (aa 1–270) was generated by PCR using oVM78 and oNS4 (5’-AGTATTACGCGTATCCATGCCAAACCTCTTG-3’, PKR nt 1826–1808) as forward and reverse primers, respectively. The resulting PCR fragment was cloned to replace FL-PKR at BamHI and MluI sites in PKR-FL plasmid.
Full-length PACT-Myc-Flag encoding plasmid (pPACT-FL) was obtained from Origene. Deletion of PACT-RBM1 (Δ1, aa 1-34/100-313), RBM2 (Δ2, aa 1-126/193-313) or both RBM1 and RBM2 (Δ1,2) was generated by overlapping PCR on pPACT-FL using following primers: oVM78 5’-CCGTTGACGCAAATGGGC-3’ in the CMV IE promoter region and oNS16 (5’- TGCATTGGCTTT/TGTTTTCCCTGGCTTAGCT-3’; PACT nt 1337-1326/1130-1112) for first PCR and oNS17 (5’-AGCCAGGGAAAACA/AAAGCCAATGCAAGTATTT-3’, PACT nt 1117-1130/1326-1344) and oNS18 (5’-CACTGGAGTGGCAACTTC3-3’; PACT nt 2166–2149) for second PCR were used for deletion of RBM1; oVM78 and oNS19 (5’-AGAAATATTACT/ATTAAGCTGGTTCTTTGGT3’-3’; PACT nt 1616-1605/1406-1388) for the first PCR and oNS20 (5’-AGAACCAGCTTAAT/AGTAATATTTCTCCAGAGA-3’; PACT nt 1393-1406/1605-1623) and oNS18 for the second PCR were used for deletion of RBM2. Obtained two PCR products in oligo-mediated deletion of RBM1 or RBM2 were combined and re-amplified with oVM78 and oNS18. The final PCR product (PACT-Δ1 or– Δ2) was used to replace full length PACT using BamHI and MluI sites in the plasmid pPACT-FL. To make the PACT-Δ1,2 (aa 1-34/100-126/193-313), oVM78 and oNS16 were used for the first PCR and oNS17 and oNS18 were used for the second PCR on plasmid PACT-Δ2. The overlapped PCR products from oVM78 and oNS18 were cloned as above. The PACT- ΔPAD mutant (Δ3, aa 1–239) was generated from the parent pPACT-FL plasmid by PCR using oVM78 and oNS21 (5’-AGTATT/ACGCGT/TGTATTTGGAATACTAAGGA-3’, PACT nt 1745–1726) as forward and reverse primers, respectively. The resulting PCR fragment was cloned to replace the PACT-FL into the parent plasmid pPACT-FL using BamHI and MluI sites.
Because vIRF2 is a split gene in the KSHV genome [84], the overlapping PCR was also performed by using oVM386 (5’-TACTCAGAATTCACC/ATGCCTCGCTACACGGAGT-3’, vIRF2 nt, 94127–94109) and oVM389 (5’- TCGCTCTGTGACCGTGATGAA-3’, vIRF2 nt 93435–93453) for the first PCR and oVM388 (5’-AGTCGCCACGCCCACAACAT-3’, vIRF2 nt 93496–93472) and oVM387 (5’-ATCGTGGATCC/GTCTCTGTGGTAAAATGGG-3’, vIRF2 nt 93435–93453) for the second PCR on cDNA derived from VA-activated BCBL-1 cell total RNA. Following by annealling the two PCR products and re-amplification using oVM386 and oVM387 to generate a spliced isoform of full-length vIRF2 in size of 2161 bps, the final PCR product was digested by EcoRI and BamHI and cloned into p-Flag-CMV-5.1. The resulting plasmid was subsequently named as pVM105. All plasmids were verified by restriction digestion and sequencing.
Unless indicated otherwise, protein samples for Western blot were prepared by direct lysis of the cells in 2 × SDS sample buffer (Quality Biological) containing 5% 2-mercaptoethanol (Sigma-Aldrich). Samples were resolved on a 4%-12% SDS-PAGE gel in 1 × MOPS buffer (Thermo Fisher Scientific). The signal was detected with SuperSignal West Pico or Femto Chemiluminescent Substrate (Thermo Fisher Scientific).
Sodium arsenite (cat. No. 38150, Sigma-Aldrich) solution (0.83 M) was prepared in water to serve as a 1660× stock solution. To induce oxidative stress the cells were cultivated in fresh culture medium containing 0.5 mM of sodium arsenite for 30 min. To mimic the cellular stress induced by dsRNA, double-stranded poly I:C (lyophilized polyinosinic–polycytidylic acid sodium salt) obtained from Sigma-Aldrich (cat. no. P0913) was dissolved in sterile DEPC-treated H2O containing 0.98% NaCl to make 5 mg/ml stock solution. Before use, poly I:C was denatured by incubation at 50°C water bath for 20 min followed by slowly cooling to room temperature for at least 45 minutes for proper annealing. The size of poly I:C was determined on a 1% agarose gel and was in size of ~250–300 bp. The amount of poly I:C required for induction of eIF2α phosphorylation was experimentally determined by transfection of HeLa cells with 0.1–5 μg of annealed poly I:C using Lipojet transfection reagent (SignaGen). To study inhibitory effect of ORF57 on PKR activation by poly I:C, HeLa or HEK293 cells with or without ORF57 expression were transfected with 1 μg poly I:C and incubated additional 8 h before harvesting. The heat stress was induced by incubation of cells at 44°C for 40 min. Before harvesting, the cells were washed with 1 × phosphate-buffered saline (PBS, Thermo Fisher Scientific).
The activation of PKR was inhibited by pre-incubation of cells for 1 h with increasing dose (1, 10 and 100 μM) of PKR inhibitor (PKRI) or PKR inhibitor control (PKRC) dissolved in DMSO. DMSO alone was used as a negative control. Cells were subsequently washed and incubated with fresh medium containing 0.5 mM of sodium arsenite for 30 min before harvesting.
BCBL-1 cells (2 × 106 cells) were induced with 1mM VA for 8 h followed by 40 min incubation at 44°C. The cells cultivated at 37°C were used as negative controls. Subsequently, the cells were harvested for preparation of total proteins in SDS sample buffer and total RNA extracted by TRIzol. Quantitative real-time RT-PCR (RT-qPCR) was performed using customized TaqMan probes and TaqMan Gene Expression Master Mix (Thermo Fisher Scientific) with standard 2-Δ(ΔCT) protocol.
HeLa cells with or without ORF57 expression were treated with 0.5 mM arsenite for 30 min to induce SG. After washing with ice-cold 1 × PBS, cells were lifted from plates using a cell scrapper and in RSB-200 buffer (10 mM Tris-HCl [pH 7.5], 200 mM NaCl, 2.5 mM MgCl2, 0.1% NP-40) supplemented with 1 × protease and phosphatase inhibitor cocktails (Roche) followed by sonication using 10 strokes at level 4 on a sonic dismembrator (Model 100, Fisher Scientific). The cell lysates were incubated on ice for 15 min and centrifuged at 15871 × g for 15 min at 4°C to separate soluble (supernatants) and insoluble (pellets) fractions. The cells without arsenite treatment were used as a negative control.
IP was performed as described earlier [50]. Briefly, ectopic expression of proteins was obtained by individual plasmid transfection in HeLa or HEK293 cells grown in a 10-cm plate. Cells with or without plasmid transfection were washed with 1× PBS, lysed in 500 μl of 1× RSB-200 lysis buffer (10 mM Tris-HCL [pH 7.5], 200 mM NaCl, 2.5 mM MgCl2, 0.1% NP-40 and protease inhibitor cocktail [Roche]), sonicated (10 strokes at level 4) and cleared by centrifugation at 11500 × g for 10 min at 4°C. In the assays designed to see the interaction of multiple overexpressed proteins, 100 μl total cell extract containing one overexpressed protein was mixed with 100 μl total cell extract with another overexpressed protein. The mixed cell lysates (200 μl) were then incubated with 5 μl of RNase A/T1 mixture containing 1.25 U RNase A and 50 U RNase T1 (Thermo Fisher Scientific) for 10 min at room temperature followed by pre-cleaning with pre-washed sepharose CL-4B beads (Sigma-Aldrich). The pre-cleaned cell lysates were mixed with 80 μl of antibody-coated protein A/G beads (50% slurry, Sigma-Aldrich) in 1 ml of IP buffer (50 mM HEPES [pH 7.5], 200 mM NaCl, 1 mM EDTA, 2.5 mM EGTA, 10% glycerol, and 0.1% NP-40, 1 × Roche’s protease inhibitor cocktail) and incubated overnight at 4°C followed by extensive wash with IP buffer. Immunoprecipitated complexes on the beads were dissolved in 70 μl of 2 × SDS protein sample buffer containing 50 mM DTT. Alternatively, the immunoprecipitated protein complexes on the beads were eluted by incubation with 50 μl of 100 μg/ml 3×Flag (Sigma-Aldrich) in IP buffer for 2 h at 4°C. The eluant in the supernatant after spinning was collected and mixed with 20 μl of 5 x SDS protein buffer containing 50 mM DTT. All samples were heat-denatured at 95°C for 5–10 min before SDS-PAGE. Western blotting for individual proteins was carried out using the 3–5% input lysates and 30%-50% immunoprecipitated proteins. Densitometric quantification of the individual protein band intensity was performed using Image J software (NIH).
Adherent HeLa and HEK293 cells were grown directly on glass coverslips. The non-adherent BCBL1 cells were immobilized by spotting of cells suspension on poly-D-lysine-treated glass coverslips. Immunofluorescence staining was performed as described previously [50,57,115]. Briefly, the cells were washed with PBS, fixed with 4% paraformaldehyde, permeabilized with 0.5% Triton X-100 and blocked with 2% BSA (bovine serum albumin, Promega, Madison, WI) dissolved in Tris-buffered saline containing 0.05% of Tween-20 (TTBS). Primary antibodies diluted in blocking buffer were incubated with slides overnight at 4°C. AlexaFluor-conjugated secondary antibodies (1: 500, ThermoFisher Scientific) were diluted in blocking solution and incubated with slides at 37°C in humidified chamber. The slides were washed with TTBS and before mounting the cells nuclei were visualized by 5 min counterstaining with wash buffer containing Hoechst dye 33342 (Sigma-Aldrich). Confocal fluorescence images were collected with a Zeiss LSM780 laser-scanning microscope (Carl Zeiss, Inc., Thornwood, NY) equipped with 20x Plan-Apochromat (numerical aperture, 0.8) and 63x Plan-Apochromat (numerical aperture, 1.4) objective lenses. The x-y pixel sizes of 0.4 and 0.07 μm and optical slice thicknesses of 1.5 and 0.9 μm were used to acquire confocal images with the 20× and 63× objectives lenses, respectively. Volume reconstructions were generated using the Imaris (version 8.0.2) image processing software (Bitplane, Inc., Concord, MA).
One microgram of reconstituted poly I:C was labelled using 10 units of T4 polynucleotide kinase (T4 PNK, Thermo Fisher Scientific) for 10 min at 37°C in a 25-μl reaction containing 25 μCi [γ-32P]-ATP (Perkin Elmer). The reaction was terminated by adding 5 mM EDTA to the reaction mixture and the labeled poly I:C was purified using an illustraMicrospin G-25 column (GE Healthcare, Marlborough, MA). In a pull-down assay, HeLa cells in a 10-cm culture dish were transfected with a Myc-Flag-tagged PKR expression vector or an empty vector. The Myc-Flag-tagged PKR protein in the HeLa cell lysate was immunoprecipitated using 80 μl of anti-mycEZview beads (Sigma-Aldrich) overnight at 4°C. The beads were then washed four times with IP buffer and subsequently resuspended in 100 μl of IP buffer. Approximately, 40 μl of the resuspended beads coated with PKR were incubated with a mixture of 100 ng [γ-32P]-labelled poly I:C and 500 ng of purified recombinant ORF57-Flag protein or BSA (negative control) in 750 μl of IP buffer for 2 h at room temperature. The beads were extensively washed with IP buffer, resuspended in 5 ml of liquid scintillation cocktail (CytoScint, MP Biochemicals, Santa Ana, CA) and radioactivity was counted by a liquid scintillation counter.
Recombinant full-length PKR was purified from HeLa cells in a 10-cm petri dish by by immunoprecipitation using 80 μl of anti-mycEZview beads as described above and subsequently dissolved in 100 μl of 1x kinase buffer (10 mM Tris-HCl [pH7.6], 50 mM KCl, 2 mM magnesium acetate, 20% glycerol). In a 32-μl kinase reaction, 12 μl of beads-attached PKR protein were first incubated with 200 ng of recombinant ORF57-Flag or BSA proteins for 10 min at room temperature. The resulting mixture was sequentially supplemented with 50 ng poly I:C, 1x kinase buffer supplemented with 0.83 mM MgCl2 and 20 μCi [γ-32P]-ATP. The reaction was incubated at 30°C for 10 min and terminated by addition of equal volume of 2 × SDS protein sample buffer. After SDS-PAGE the gel was mounted in an exposure cassette and analyzed by a PhosphorImager (GE Healthcare).
HeLa cells with expression of Myc-Flag-tagged PKR in a 10-cm Petri dish were treated with arsenite to phosphorylate PKR. Myc-Flag-tagged PKR purified by immunoprecipitation as described above from the cells with (for activated PKR) or without (for inactive PKR) arsenite treatment was finally resuspended in 100 μl of 1x kinase buffer. Ten ul of the resuspended beads coated with PKR were mixed with 500 ng of recombinant ORF57 protein or BSA and incubated for 10 min at room temperature. The mixture was sequentially supplemented with 1x kinase buffer, 0.83 mM MgCl2, 50 ng GST-eIF2α (cat# H00001965-P01, Abnova), and 20 μCi [γ-32P]-ATP to a final volume of 32 μl. The kinase reaction involving eIF2α phosphorylation by phosphorylated (activated) PKR was allowed to proceed for 40 min at 30°C. The reaction was terminated by addition of 2x SDS sample buffer and the samples were resolved by SDS-PAGE. The gel was exposed to a PhosphorImager and an X-ray film for signal quantification.
KSHV infected iSLK-BAC16 cells [92] growing in a six-well plate (2.5 x 105 cells/ well) were transfected twice, respectively, at an interval of 24 h with 40 nM of ON-TARGETplus SMART-pool PKR siRNAs targeting human PKR/EIF2AK2 (L-003527-00-0005, Dharmacon, GE Healthcare, Lafayette, CO) or ON-TARGETplus Non-targeting siRNA #1 negative control (D-001810-01, Dharmacon, GE Healthcare) using LipoJet transfection reagent (SL100468, SignaGen Laboratories, Gaithersburg, MD). Total cell extract was collected 24 h after the second siRNA transfection to measure the knockdown efficiency by immunoblotting.
For KSHV virus production and titration assays, iSLK-BAC16 cells in a six-well plate without or with PKR siRNA transfection at 24 h of the second round of siRNA transfections described above were induced for KSHV lytic infection with 1 mM sodium Butyrate and 1 ug/ml doxycycline in 2 ml of DMEM medium. On the third day, fresh 2 ml of DMEM medium containing the same amount of sodium butyrate and doxycycline were added to make the culture medium in total of 4 ml per well for another two more days. For virus production and titration, the iSLK-BAC16 culture supernatants were harvested on the 5th day after induction, cleared by centrifugation at 2000 rpm for 10 min, and filtered through Sterile Millex 0.45 μM filter units (cat# SLHA033SS, Millipore, Billerica, MA). 400 μl of the supernatants were used to infect HEK293 (5 x 105 cells/ well) in a six-well plate. KSHV-infected GFP+ HEK293 cells at 48 h after infection were observed by a fluorescent microscopy and analyzed by flow cytometry.
|
10.1371/journal.ppat.1001121 | Ebolavirus Is Internalized into Host Cells via Macropinocytosis in a Viral Glycoprotein-Dependent Manner | Ebolavirus (EBOV) is an enveloped, single-stranded, negative-sense RNA virus that causes severe hemorrhagic fever with mortality rates of up to 90% in humans and nonhuman primates. Previous studies suggest roles for clathrin- or caveolae-mediated endocytosis in EBOV entry; however, ebolavirus virions are long, filamentous particles that are larger than the plasma membrane invaginations that characterize clathrin- or caveolae-mediated endocytosis. The mechanism of EBOV entry remains, therefore, poorly understood. To better understand Ebolavirus entry, we carried out internalization studies with fluorescently labeled, biologically contained Ebolavirus and Ebolavirus-like particles (Ebola VLPs), both of which resemble authentic Ebolavirus in their morphology. We examined the mechanism of Ebolavirus internalization by real-time analysis of these fluorescently labeled Ebolavirus particles and found that their internalization was independent of clathrin- or caveolae-mediated endocytosis, but that they co-localized with sorting nexin (SNX) 5, a marker of macropinocytosis-specific endosomes (macropinosomes). Moreover, the internalization of Ebolavirus virions accelerated the uptake of a macropinocytosis-specific cargo, was associated with plasma membrane ruffling, and was dependent on cellular GTPases and kinases involved in macropinocytosis. A pseudotyped vesicular stomatitis virus possessing the Ebolavirus glycoprotein (GP) also co-localized with SNX5 and its internalization and infectivity were affected by macropinocytosis inhibitors. Taken together, our data suggest that Ebolavirus is internalized into cells by stimulating macropinocytosis in a GP-dependent manner. These findings provide new insights into the lifecycle of Ebolavirus and may aid in the development of therapeutics for Ebolavirus infection.
| Ebolavirus (EBOV) is an enveloped, single-stranded, negative-sense RNA virus that causes severe hemorrhagic fever with high mortality rates in humans and nonhuman primates. Previous studies suggest roles for clathrin- or caveolae-mediated endocytosis in EBOV entry; however, questions remain regarding the mechanism of EBOV entry. Here, we demonstrate that internalization of EBOV particles is independent of clathrin- or caveolae-mediated endocytosis. Specifically, we show that internalized EBOV particles co-localize with macropinocytosis-specific endosomes (macropinosomes) and that their entry is negatively affected by treatment with macropinocytosis inhibitors. Moreover, the internalization of Ebola virions accelerated the uptake of a macropinocytosis-specific cargo, was associated with plasma membrane ruffling, and was dependent on cellular GTPases and kinases involved in macropinocytosis. We further demonstrate that a pseudotyped vesicular stomatitis virus possessing the EBOV glycoprotein (GP) also co-localizes with macropinosomes and its internalization is similarly affected by macropinocytosis inhibitors. Our results indicate that EBOV uptake into cells involves the macropinocytic pathway and is GP-dependent. These findings provide new insights into the lifecycle of EBOV and may aid in the development of therapeutics for EBOV infection.
| Viruses have evolved a variety of mechanisms to enter host cells [1], [2], [3], including clathrin- and caveolae-mediated endocytosis, phagocytosis, and macropinocytosis. The main route of endocytosis, mediated by clathrin, is characterized by the formation of clathrin-coated pits (CCP) of 85–110 nm in diameter that bud into the cytoplasm to form clathrin-coated vesicles. Influenza virus, vesicular stomatitis virus (VSV) and Semliki forest virus all enter their host cells via this pathway [4], [5], [6]. Although Listeria monocytogenes is larger than a CCP in diameter, it exploits non-classical clathrin-mediated endocytosis along with actin rearrangement to facilitate its infection [7], [8]. Caveolae are small vesicles of 50–80 nm in diameter enriched in caveolin, cholesterol, and sphingolipid, and have been implicated in simian virus 40 (SV40) entry [9]. Clathrin- and caveolae-mediated endocytosis requires large guanosine tryphosphatases (GTPase) dynamin 2 for vesicle scission [3].
Phagocytosis plays a role in the uptake of microorganisms, cell debris, and apoptotic cells [10]. It is initiated by the interaction of cell surface receptors, such as mannose receptors, Fc receptors and lectin receptors, with their ligands at the surface of the internalized particles. Particles are internalized through a dynamin 2- and actin-dependent mechanism [11] that results in the formation of phagosomes, large particles of >500 nm in diameter. Human herpes simplex virus and acanthamoeba polyphaga mimivirus are internalized through this mechanism [12], [13].
Macropinocytosis is characterized by actin-dependent membrane ruffling and, unlike phagocytosis, was thought to be independent of receptors or dynamin 2 [14], [15], [16], [17]. Macropinocytosis is constitutively activated in some immune cells, such as dendritic cells and macrophages [18], [19], [20]. In the other cell types, including epithelial cells and fibloblasts, macropinocytosis is initiated by growth factor stimulation [21], [22] or expression of ruffling kinases [23], [24], [25]. Macropinocytosis is also associated with the activation of Rho GTPases, such as Rac1 and Cdc42, which are responsible for triggering membrane ruffles by actin polymerization [26], [27], [28], [29]. Macropinocytosis is dependent on a series of kinases; a serine/threonine p21-activated kinase 1 (Pak1) is activated by Rac1 or Cdc42 and is essential for the regulation of cytoskeleton dynamics [24], [30]. In addition Pak1 plays a role in macropinosome closure by activating carboxy-terminal-binding protein-1/brefeldin A-ADP ribosylated substrate (CtBP-1/BARS) [30]. Phosphatidylinositol-3-kinase (PI3K) and its effectors are responsible for ruffling and macropinocytosis [23], [31]. Protein kinase C (PKC) is activated by a receptor tyrosine kinase or PI3K and also promotes plasma membrane ruffling and macropinocytosis [23]. Membrane ruffling is associated with the formation of macropinocytosis-specific endosomes, macropinosomes, of approximately 0.5–10 µm in diameter [32]. Human adenovirus type 3 (Ad3) [33], vaccinia virus [26], Kaposi's Sarcoma Associated Herpesvirus [34], and Nipah virus [35] enter cells via macropinocytosis. Human immunodeficiency virus (HIV) [36], [37] and Ad2/5 [38] may also trigger this pathway.
Ebolavirus (EBOV) is an enveloped, single-stranded, negative-sense RNA virus that belongs to the family Filoviridae. In humans and nonhuman primates, it causes severe hemorrhagic fever with mortality rates of up to 90%. Ebolavirus virions are long, filamentous particles of varied length (typically, 1–2 µm) and a diameter of 80–100 nm. EBOV infects a wide range of host cells [39], suggesting that its entry into target cells is mediated by the binding of its surface glycoprotein (GP) to a widely expressed and highly conserved receptor, or by GP binding to different host receptors. Several cellular proteins have been reported as EBOV receptors or co-receptors, including folate receptor-α (FR-α) [40], several lectins [41], [42], [43], [44], [45], [46], and integrin ß1 [47]. In addition, EBOV entry is facilitated by members of the Tyro3 protein kinase family [48], [49].
The mechanism of EBOV cell entry is currently poorly understood. EBOV is likely internalized by an endocytic pathway, since its entry is dependent upon low pH [50], [51] and the endocytic enzymes cathepsin B and L [52], [53], [54], [55], [56]. Several studies suggest that EBOV internalization depends on cholesterol, a major component of caveolae and lipid-rafts [50], [57], [58]. Another study suggests a role for clathrin-mediated endocytosis in wild-type EBOV and retrovirus psendotyped with EBOV GP entry [59], [60]. These discrepancies may reflect differences in the experimental systems and/or conditions used. Most studies have been carried out with retroviruses or vesicular stomatitis virus (VSV) pseudotyped with EBOV GP [52], [53], [54], [56], [58], [61]. These pseudotyped systems have limitations because the morphology of the virions differs significantly from that of authentic Ebola virions (spherical for retrovirus or VSV-pseudotyped virions versus filamentous for authentic Ebola virions).
To better understand EBOV entry, we conducted internalization studies with fluorescently labeled, biologically contained EBOV [62], and Ebolavirus-like particles (Ebola VLPs), both of which resemble authentic EBOV in their morphology [62], [63], [64], [65]. Our results suggest that EBOV uptake into cells involves the macropinocytic pathway and is GP-dependent.
To assess the mechanism of EBOV entry, we established a real-time monitoring system for fluorescently labeled, biologically contained Ebola virions [62], and fluorescently labeled Ebola VLPs [63], [64], [65]. The biologically contained EBOV (EbolaΔVP30) lacks the gene for the viral transcriptional co-activator VP30 and can only replicate in VP30-expressing cells [62]. EbolaΔVP30 resembles authentic EBOV [62] and thus provides an ideal system to study EBOV entry. Likewise, co-expression of the EBOV GP glycoprotein and the VP40 matrix protein yields virus-like particles (VLPs) with filamentous architecture [63], [64], [65]. Since co-expression of the EBOV nucleoprotein (NP) increases the efficiency of VLP generation [66], we generated VLPs by co-expressing GP, VP40, and NP. To establish a real-time monitoring system for EBOV cell entry, EbolaΔVP30 virions and Ebola VLPs were generated and purified as described in the Material and Methods, and labeled with a lipophilic tracer, 1,1′-dioctadecyl-3,3,3′,3′-tetramethylindocarbocyanine perchlorate (DiI), which is incorporated into the envelope of the virions [67], [68], [69]. The infectivity of DiI-labeled EbolaΔVP30 was equivalent to that of unlabeled virions as measured by plaque assays (data not shown), demonstrating that DiI labeling did not interfere with virion binding and infectivity.
We synchronized the adsorption of DiI-labeled EbolaΔVP30 and Ebola VLPs to African green monkey kidney epithelial (Vero) cells, which support EBOV replication, for 30 min on ice. We assessed the effect of low temperature incubation on the internalization of the DiI-virions by incubation on ice, room temperature, or 37°C in parallel, followed by a temperature shift to 37°C and found that there were no appreciably differences in the total numbers of internalized virions across these conditions, suggesting that incubation of cells and virions on ice had a limited effect on the subsequent viral internalization (Figure S1).
After adsorption, we shifted the temperature to 37°C and visualized the labeled particles by using confocal laser scanning microscope at various times. DiI-labeled EbolaΔVP30 and Ebola VLPs were visualized as red particles of various sizes (red, Figure S2A), indicating that viral particles of various lengths had been produced, an observation that we confirmed by electron microscope (Figure S3). Both DiI-labeled EbolaΔVP30 and Ebola VLPs were internalized efficiently, migrated immediately after the temperature shift, and eventually trafficked to intracellular compartments (Figure S2A and B, left panels, Video S1). As a control, we tested VLPs that lacked GP [Ebola VLPs (-GP)]. These particles bound to the cells with low efficiency and remained stationary even after long-term incubation at 37°C (Figure S2A and B, right panels, Video S2), confirming the requirement of GP for binding and internalization of EBOV.
Previous studies suggested that EBOV enters cells via clathrin-mediated endocytosis [50], [59]. The typical architecture of Ebola virions (length 1–2 µm and diameter 80–100 nm) is larger than the diameter of clathrin-coated pits (85–110 nm). However, Listeria monocytogenes is internalized into cells via non-classical clathrin-mediated endocytosis [7], [8], Therefore, we visualized clathrin-coated pits (CCPs) via the expression of clathrin light chain a (CLCa) fused to enhanced green fluorescent protein (eGFP) to assess the significance of this pathway for EBOV internalization. The functional integrity of clathrin is not compromised by fusion to eGFP and the expressed fusion protein forms CCPs with endogenous CLCa [70], [71]. We did not detect co-localization of eGFP-labeled CLCa (CLCa-eGFP) with DiI-labeled EbolaΔVP30 virions (Figure 1A, left panel and Video S3) or Ebola VLPs (Figure 1A, right panel) at 15 min or 60 min after the temperature shift, whereas fluorescence-labeled Transferrin (Tf), a specific ligand of the clathrin-mediated pathway, was co-localized with eGFP-CLCa (Figure S4, left panel). These results suggest that clathrin-mediated endocytosis may not be critical for EBOV entry.
To further assess the role of clathrin-dependent endocytosis in EBOV entry, we down-regulated endogenous clathrin heavy chain (CHC) with small interfering RNAs (siRNA) and assessed the effect of CHC down-regulation on the internalization of Ebola virions. Down-regulation of CHC expression (red) was confirmed by immunofluorescent staining in Vero cells (Figure 1B, lower right panel). To remove the surface-bound uninternalized virions, we treated the cells with trypsin 2 h post-temperature shift (Figure S5). The uptake of Alexa Fluor-Tf was abrogated in CHC siRNA-treated cells, indicating that the clathrin-mediated endocytosis was blocked in these cells (Figure 1B, lower right panel). However, internalization of Ebola VLPs was not blocked by down-regulation of CHC (Figure 1B, upper right panel and Figure 1C), further suggesting that clathrin-mediated endocytosis is not critical for EBOV entry.
Previous studies also indicated a role for caveolin-mediated endocytosis in EBOV internalization [50], [59]. Using a similar strategy to that described above, we assessed the co-localization of eGFP-fused caveolin 1 (Cav1-eGFP), which does not impair the internalization of caveolae [9], [72], with DiI-labeled EbolaΔVP30 (Figure 2A, left panel, Video S4) and Ebola VLPs (Figure 2A, right panel). We did not observe efficient co-localization of labeled Ebola virions with Cav1, indicating that caveolaes may not play a critical role in EBOV entry. Alexa Fluor-Chorela toxin B subunit (CtxB), which is internalized via caveolae- and clathrin-mediated endocytosis [73], was co-localized with some of the Cav1-eGFP (Figure S4, right panel).
The role of caveolin-mediated endocytosis was further tested by inhibiting Cav1 expression with siRNA. Down-regulation of Cav1 expression was confirmed by immunofluorescent staining and western blotting in Vero cells (Figure 2B, lower right panel). Cav1 down-regulation did not prevent DiI-Ebola VLP internalization (Figure 2B and 2C), upper right panel and Figure 2D), further suggesting that caveolin-mediated endocytosis does not play a critical role in EBOV internalization. Our finding that DiI-labeled EbolaΔVP30 virions enter Cav1-deficient human hepatoblastoma Huh7 cells [74] (Figure 2E) further supports this concept.
Clathrin-, caveolae- and phagocytosis-mediated endocytosis all depend on dynamin 2, a large GTPase that plays an essential role in vesicle scission during clathrin- and caveolae-dependent endocytosis and phagocytosis [75]. Treatment with a dynamine-specific inhibitor, dynasore [76], reduced the internalization of Alexa Fluor-labeled Tf (green; Figure 2F, right panel); however, dynasore did not affect the internalization of DiI-labeled virions (Figure 2F, right panel and Figure 2G). These data indicate that EBOV internalization does not involve clathrin-, caveolin-, or phagocytosis-mediated endocytosis.
Our data argue against a role for clathrin-, caveolae-, or phagocytosis-mediated endocytosis in the internalization of EBOV. We therefore considered macropinocytosis as a potential mode of EBOV entry. Induction of macropinocytosis leads the formation of macropinocytosis-specific endosomes (macropinosomes), which are large enough (0.5–10 µm of diameter) [32] to accommodate Ebola virions.
Sorting nexin (SNX) 5 comprises a large family of peripheral membrane proteins that associate with newly formed macropinosomes and are involved in their maturation [77], [78]. To assess the role of macropinocytosis in EBOV internalization, we first generated Vero cells expressing an eGFP-SNX5 fusion protein and confirmed that a specific ligand of macropinocytosis, dextran Mw 10,000 (Dex Mw 10K) co-localized with expressed eGFP-SNX5 (Figure S6A) but not with CLCa-eGFP or Cav1-eGFP in Vero cells (Figure S6B). We then asked whether DiI-labeled EbolaΔVP30 and Ebola VLPs co-localize with eGFP-SNX5-positive vesicles. Approximately 70% of DiI-labeled EbolaΔVP30 (blue bars in Figure 3B) and 45% of DiI-labeled Ebola VLPs (yellow bars in Figure 3B) associated with eGFP-SNX5-positive vesicles within 10 min of the temperature shift to 37°C (Figure 3A, upper panels, Figure 3B, and Video S5). Co-localization of viral particles with eGFP-SNX5-positive vesicles continued for 30 min after the temperature shift and then decreased (Figure 3B). On the other hand, DiI-labeled influenza viruses, which are mainly internalized by clathrin-mediated endocytosis [5], did not appreciably co-localize with eGFP-SNX5-positive vesicles (Figure 3A, red bars in lower panels, Figure 3B, and Video S6). We further confirmed co-localization of Ebola VLPs with endogenous SNX5 (Figure S7A). These observations suggest an association of internalized Ebola virions with macropinosomes.
Once internalized, macropinosomes mature into endocytic vesicles [77], [79]. However, the endocytic pathway is also part of the clathrin- and caveolin-mediated entry processes. Several groups have shown that authentic EBOV and EBOV GP-pseudotyped virions enter cells in a low pH- and cathepsin B/L-dependent manner, consistent with endosomal entry [50], [51], [52], [53], [54], [55], [56]. Here, we sought to confirm endosomal localization of EbolaΔVP30 and Ebola VLPs, both of which more closely resemble authentic EBOV than do pseudotyped viruses.
The small GTPase Rab7 specifically associates with late endosomes [80], [81] and serves as a marker for this compartment. We, therefore, analyzed the co-localization of internalized DiI-labeled virions with eGFP-Rab7-positive vesicles after the temperature shift. About 20% of DiI-labeled EbolaΔVP30 virions (blue bars in Figure 4B) and Ebola VLPs (yellow bars in Figure 4B) co-localized with eGFP-Rab7 within 10–20 min of the temperature shift; within 2 h of the temperature shift, 70%–80% of EbolaΔVP30 particles and Ebola VLPs co-localized with eGFP-Rab7 (Figure 4A and 4B). Internalized Dex Mw 10K, a specific ligand of macropinocytosis, was also observed in Rab7-positive vesicles (Figure S8). We further confirmed co-localization of Ebola VLPs with endogenous Rab7 (Figure S7B). At 3–4 h after the temperature shift, the DiI-signals were enlarged and overlapped with eGFP-Rab7 (Figure S9, left panel), suggesting fusion of the DiI-labeled viral envelopes with endosomal membranes. Following treatment with NH4Cl, which inhibits the acidification of endosomes, the DiI-signals localized with eGFP-Rab7 but remained small (Figure S9, right panel), indicating that NH4Cl inhibited membrane fusion. Similarly, VLPs possessing a fusion-deficient GP mutant (F535R) [82] trafficked to eGFP-Rab7-positive vesicles but the signals remained small (Figure S10). Collectively, these findings indicate that internalized EBOV particles are transported to late endosomes, where low pH-dependent membrane fusion occurs.
To further test whether Ebola virions are internalized via macropinocytosis, we assessed several inhibitors for their effects on EBOV uptake. DiI-labeled influenza virus particles which are internalized via clathrin-mediated endocytosis, served as a control. An actin depolymerizing agent, cytochalasin D (CytoD) was used because macropinocytosis depends on actin bundle formation; however, an intact actin skeleton is also critical for other endocytic pathways [83]. Since macropinocytosis relies on PI3K activation [23], [31], we also tested two inhibitors of this kinase, wortmannin (Wort) and LY-294002 [84]. Finally, we used EIPA [5-(N-ethyl-N-isopropyl) amiloride], an inhibitor of the Na+/H+ exchanger that specifically inhibits macropinocytosis [26], [34], [35], [85], [86]. These inhibitors all inhibited the uptake of Dex Mw 10K (Figure S11A and B). Treatment of cells with the inhibitors appreciably blocked co-localization of EbolaΔVP30 (blue bars in Figure 5B) and VLPs (yellow bars in Figure 5B) with late endosomes, as visualized by eGFP-Rab7 expression (Figure 5A and 5B). CytoD treatment also affected co-localization of DiI-labeled influenza virus with late endosomes (red bars in Figure 5B and S12); this observation was expected because actin is also critical for the internalization of influenza viruses [69], [72], [87]. Entry of influenza virus was also moderately affected by the PI3K inhibitors (Figure 5B and S12), a result consistent with a previous report of PI3K-dependent influenza virus cell entry [88]. However, the internalization of influenza virus was not inhibited by EIPA (Figures 5B and S11), whereas the uptake of DiI-EbolaΔVP30 and Ebola VLPs was appreciably reduced in the presence of this compound (Figure 5A and 5B). These findings suggest that EBOV is internalized via macropinocytosis.
Constitutive macropinocytosis occurs in specific cell types such as dendritic cells and macrophages [18], [19], [20]; however, in epithelial cells, it is initiated in response to growth factor stimulation [21], [22] or expression of ruffling kinases [23], [24], [25]. To assess whether Ebola virions activate macropinocytosis to allow EBOV be internalized into the cells, we asked whether the virions accelerated the uptake of a macropinocytosis marker, Dex Mw 10K. In the presence of Ebola virions, the uptake of Dex Mw 10K was accelerated (Figure 6A and S13), and this event was inhibited by EIPA. Co-localization of DiI-EbolaΔVP30 and Alexa Fluor-Dex Mw 10K was also observed (Figure 6B).
The Rho GTPases (Rac1 and Cdc42), protein kinase C (PKC), and Pak1 are involved in the regulation of macropinocytosis[23], [24], [27], [29], [30]. Therefore, we examined the role of Rac1 by use of dominant-negative Rac1 (dnRac1) [89]. Expression of eGFP-fused dnRac1 inhibited the internalization of Ebola virions (red) into cells by 80% (Figure 6C, lower right panel; Figure 6D) compared with that of eGFP-fused wild-type Rac1 (wtRac1) (Figure 6C upper right panel; Figure 6D). dnRac1 expression also interfered with the uptake of Dex Mw 10K (blue) (Figure 6C, lower middle panel), indicating that expression of dnRac1 inhibited macropinocytosis. The role of PKC in the internalization of Ebola virions was tested by use of the specific PKC inhibitor staurosporine [90]. Staurosporine reduced the internalization of DiI-virions (red bars in Figure 6E and left panels in Figure S14A) and Dex Mw 10K (blue bars in Figure 6E and right panels in Figure S14A) by 80% and 70%, respectively. The effect of down-regulation of Cdc42, and Pak1 by siRNAs on Ebola VLP uptake was also tested. Down-regulation at the mRNA level was assessed by RT-PCR (Figure 6F). Knockdown of Cdc42 and Pak1 appreciably interfered with DiI-Ebola VLP internalization (red bars in Figure 6G and left panels in Figure S14B) and also reduced the uptake of Dex Mw 10K (blue bars in Figure 6G and right panels in Figure S14B). Since plasma membrane ruffling precedes macropinocytosis [14], [15], [16], we monitored ruffling formation in the internalization of DiI-virions by use of Vero cells expressing eGFP-actin [91]. Time-lapse analysis revealed that prominent plasma membrane ruffling was associated with DiI-labeled virions after the temperature-shift (Figure 6H and Video S7). Appreciable actin rearrangement was not observed in the absence of EBOV virions (Figure S15 and Video S8). Together, these results demonstrate that Ebola virions stimulate macropinocytosis along with the activation of the cellular factors involved in actin polymerization that allow the virions to be internalized.
Our data indicate that the EBOV particle internalization occurs via macropinocytosis, whereas previous studies suggest that clathrin- or caveolin-dependent endocytosis mediate the internalization of wild-type EBOV and EBOV GP-pseudotyped VSV or retroviruses [50], [57], [58], [59]. To determine if these conflicting findings result from differences in assay systems (i.e., use of pseudotype viruses) and/or experimental conditions used, we tested whether a VSV pseudotyped with EBOV GP (VSVΔ*G-GP) was internalized by macropinocytosis. Although VSV is known to be internalized via the clathrin-dependent pathway [6], DiI-labeled VSVΔ*G-GP virions did not co-localize with CLCa-eGFP or Cav1-eGFP (Figure S16), whereas DiI-labeled VSVΔ*G-G virions co-localized with CLCa-eGFP (Figure S17). By contrast, DiI-labeled VSVΔ*G-GP virions co-localized with eGFP-SNX5 (Figure 7A, left panel), indicative of macropinocytosis. No significant co-localization with eGFP-SNX5 was observed for a DiI-labeled control virion possessing the VSV G glycoprotein (DiI-VSVΔ*G-G; Figure 7A, right panel). EbolaΔVP30 particles possessing authentic morphologies (blue bars in Figure 7B) and VSV pseudotyped with EBOV GP (green bars in Figure 7B) co-localized with eGFP-Rab7-positive vesicles with similar kinetics, indicating that the smaller size of the VSV virions relative to that of the Ebola virions did not affect the kinetics of internalization. The kinetics of DiI-VSVΔ*G-G trafficking to late endosomes/lysosomes (red bars in Figure 7B) was consistent with a previous study of authentic VSV [92]. EIPA, which specifically interferes with macropinocytosis, blocked the co-localization of eGFP-Rab7 with DiI-labeled VSVΔ*G-GP (green bars in Figure 7C), but not with DiI-VSVΔ*G-G (red bars in Figure 7C). The PI3K inhibitors significantly reduced the co-localization of eGFP-Rab7 with DiI-labeled VSVΔ*G-GP (green bars in Figure 7C) but not with VSVΔ*G-G (red bars in Figure 7C), which is consistent with previous findings [93].
The effect of these inhibitors was further assessed in a viral infection system by use of a VSV pseudovirion encoding eGFP. Vero cells were pre-treated with one of the inhibitors and then infected with VSVΔG*-GP (green bars in Figure 7D) or VSVΔG*-G (red bars in Figure 7D) in the presence of the inhibitors. The infection efficiency of each pseudovirus was determined by measuring the number of GFP-positive cells. EIPA blocked the infection of VSVΔ*G-GP (green bars in Figure 7D), but not VSVΔ*G-G (red bars in Figure 7D). The PI3K inhibitors reduced the infection of VSVΔ*G-GP (green bars in Figure 7D) but not VSVΔ*G-G (red bars in Figure 7D), which is consistent with the results of the co-localization of DiI-VSV pseudovirions and SNX5 (Figure 7C). These findings demonstrated that in this viral infection system, VSV pseudotyped with EBOV GP is internalized by macropinocytosis, as are EbolaΔVP30 and Ebola VLPs. Therefore, regardless of the size of the virions, our data indicate that EBOV GP induces receptor-dependent macropinocytosis, unlike those in a previous report which showed that macropinocytosis is receptor-independent [32]. Our finding is consistent with a recent report describing receptor-dependent macropinocytosis in adenovirus type 3 [33].
Viruses accomplish cell entry by hijacking the cellular endocytic machinery. In this study, with EBOV particles that resemble authentic EBOV, the data lead us to conclude that EBOV is internalized into host cells via macropinocytosis in a viral GP-dependent manner.
Our conclusion that EBOV is internalized via macropinocytosis is based on the following observations: (i) the internalized viral particles co-localize with a marker of macropinosomes, SNX5 (Figure 3); (ii) the internalization of viral particles was blocked by inhibitors of actin polymerization and PI3K, which are known players in macropinocytosis and also by a specific inhibitor of macropinocytosis, EIPA (Figure 5); (iii) the internalization of Ebola virions accelerated the uptake of a specific cargo for macropinosomes Dex Mw 10K (Figure 6A) and the internalized virions co-localized with Dex Mw 10K (Figure 6B); (iv) the internalization of viral particles was blocked by a dominant-negative Rac1 (Figure 6C and 6D), a PKC inhibitor (Figure 6E) and the down-regulation of Cdc42 and Pak1 (Figure 6F and 6G); and (v) the internalization of viral particles was associated with membrane ruffling (Figure 6H). These findings suggest a model in which the binding of EBOV glycoprotein to cellular receptor(s) activates multiple macropinocytosis inducers (PI3K, Rac1, PKC, Cdc42, and Pak1), triggering plasma membrane ruffling and macropinocytosis (Figure 8). Internalized Ebola virions then traffic to Rab7-positive late endosomes/lysosomes (Figure 4), where membrane fusion occurs (Figure 8).
Two findings, the inability to enter cells of Ebola VLPs lacking GP (Figure S2) and the macropinocytic uptake of VSV particles pseudotyped with EBOV GP (Figure 7), support a role for GP in the macropinocytic internalization of EBOV particles. Macropinocytosis was thought to be receptor-independent [32] until a recent study showed that Ad3 entry via macropinocytosis requires receptors (CD46 and integrins) [33]. This finding, together with our observations, supports the concept of receptor-mediated macropinocytic pathways. The exact mechanism of GP-mediated macropinocytosis remains to be elucidated; however, mannose-binding lectin, a potential EBOV co-receptor [43], is known to accelerate macropinocytosis and phagocytosis for the uptake of apoptotic cells and bacteria into macrophages [94], [95]. In addition, integrins, which are also potential EBOV co-receptors [47], play an important role in Ad3 entry via macropinocytosis [33]. Thus, macropinocytosis is likely initiated through GP interaction with EBOV co-receptors on the cell surface (Figure 8).
Recently, one study demonstrated that the entry of Ebola VLPs and pseudovirions depends on the PI3K-Akt signaling pathway and Rac1 [93]. PI3K and its effectors are responsible for ruffling and macropinocytosis [23], [31]. Rac1 is also critical for the induction of actin filament accumulation at the plasma membrane, which leads to membrane ruffling and macropinocytosis [27]. Moreover, membrane-bound Rac1 localizes to macropinosomes [26], [27], [28]. Other study demonstrated that overexpression of RhoC GTPase facilitated wild-type EBOV entry and VSV pseudotyped with EBOV GP [96]. Although a role of RhoC in viral entry has not been specifically characterized, the overexpression of RhoC resulted in increased dextran uptake and in formation of increased actin organization [96], suggesting that RhoC plays a role in EBOV entry mediated via macropinocytosis.
Taken together with our findings, these observations support the model of EBOV entry through macropinocytosis.
Clathrin-mediated endocytosis was thought to contribute to EBOV entry based on findings that specific inhibitors of clathrin-mediated endocytosis blocked the expression of viral antigens in EBOV-infected cells [59]. However, some of these inhibitors caused severe cytotoxicity, which may have induced the down-regulation of viral antigen expression [59]. Recently, by using specific inhibitors of clathrin-mediated endocytosis, a dominant-negative Eps15, which abrogates CCP formation, and siRNA for CHC, a possible role for the clathrin-dependent pathway in the internalization of retrovirus pseudovirions with EBOV GP was suggested [60]. The discrepancy between this study and ours may originate from the difference in pseudotype systems (retrovirus versus VSV or Ebola virions) and specific cell types [60]. Our data demonstrate that down-regulation of cellular CHC, which specifically blocks clathrin-mediated endocytosis, does not interfere with the internalization of Ebola virions which resemble authentic EBOV in their morphology into Vero cells (Figure 1B).
Caveolae- and lipid-raft-mediated endocytosis were also thought to play a role in EBOV entry because FR-α, a potential co-receptor of filovirus entry, localizes to lipid rafts and is internalized through lipid raft-associated caveolae [40]. However, the role of FR-α in EBOV entry remains controversial [51], [97]. The internalization of EBOV GP-pseudotyped virions was sensitive to the depletion of cholesterol, a major component of caveolae and lipid rafts [50], [57], [58]; however, cholesterol is also required for membrane ruffling and macropinocytosis [98]. Moreover, the internalization of Ebola virions into cells transfected with siRNA for Cav1 (Figure 2B and 2C) or that lacked Cav1 (Figure 2D), argues against a role for caveolae-mediated endocytosis in EBOV entry.
One study [59] ruled out macropinocytic uptake of wild-type EBOV based on the use of an amiloride; however, the concentration of the drug used was one tenth of that typically used and may not have allowed the authors to detect an effect of this anti-macropinocytic drug on EBOV internalization.
After internalization, EBOV particles traffic to late endosomes, as suggested by their co-localization with Rab7-positive vesicles (Figure 4). This finding is consistent with previous studies that identified low pH- and cathepsin B/L-requirements for the internalization of EBOV and pseudovirions with EBOV GP into host cells [50], [51], [52], [53], [54], [55], [56].
Currently, no antivirals or vaccines are available for EBOV infections. Since viral entry is an attractive target for therapeutic intervention, it is imperative that we understand the mechanism of EBOV cell entry. Our finding that EBOV is likely internalized through macropinocytosis may stimulate the development of compounds that interfere with the EBOV internalization process.
Human CLCa, Cav1, and Rab7 genes were amplified by RT-PCR from total RNA derived from HeLa cells, and subcloned into pEGFP-N1 or pEGFP-C1 plasmids (Clontech, Mountain View, USA). The eGFP-SNX5 and eGFP-actin expression plasmid was a kind gift from Drs Rohan D. Teasdale (University of Queensland, Brisbane, Australia) and David Knecht (University of Connecticut), respectively. The eGFP-fused genes were cloned into a moloney murine leukemia virus-based retrovirus plasmid [99], a kind gift from Dr. Bill Sugden (University of Wisconsin-Madison, Madison, USA). Expression plasmids for eGFP-fused wild-type and dominant-negative Rac1 were purchased from Addgene (Cambridge, USA). DiI, Alexa Fluor 633-labeled Tf and Alexa Fluor 647-labeled Dex Mw 10K were purchased from Invitrogen (Carlsbad, USA). Dynasore, Cytochalasin D, Wortmannin, LY-294002 hydrochloride, EIPA, and Staurosporine were purchased from Sigma-Aldrich (St. Louis, USA). Antibodies for human clathrin heavy chain and Caveolin 1 were purchased from Abcam (Cambridge, UK).
African green monkey kidney epithelial Vero cells were grown in minimum essential medium (MEM) supplemented with 10% fetal bovine serum (FBS), L-glutamine, vitamins, nonessential amino acids, and antibiotics. A Vero cell line stably expressing the EBOV VP30 protein [62] was maintained in complete MEM containing 5 µg/ml puromycin (Sigma-Aldrich). Human embryonic kidney 293T cells and human hepatoblastoma cell line Huh7 cells were grown in high-glucose Dulbecco's modified Eagle's medium (DMEM) containing 10% FBS and antibiotics. Cells were maintained at 37°C in 5% CO2. Plasmid transfections in Vero cells were carried out with FuGENE HG (Roche, Basel, Switzerland).
Recombinant retroviruses for the expression of CLCa-eGFP, Cav1-eGFP, eGFP-SNX5, -actin and -Rab7, were produced and purified as previously described [99]. For retroviral infections, Vero cells were grown to 20%–30% confluence, at which point the culture medium was replaced with ice-cold MEM supplemented with 10% FBS and 20 mM Hepes (pH 7.4), and the cells were incubated with viral stocks (107–108 infectious units/ml) for 1 h at 4°C at a multiplicity of infection (m.o.i) of 5. After being washed twice with complete medium, the cells were cultured in complete medium for 48 h.
For the purification of EbolaΔVP30, Vero cells stably expressing VP30 were infected with EbolaΔVP30 stock [62] at a m.o.i of 0.1 in MEM containing 4% BSA and 2% FBS. EbolaΔVP30-containing culture medium was harvested 5 days post-infection and centrifuged at 3,500 rpm for 15 min to remove cell debris. The virions were precipitated through a 30% sucrose cushion by centrifugation at 11,000 rpm for 1 h at 4°C with an SW28 rotor (Beckman, Fullerton, USA). Precipitated virions were resuspended in TNE buffer [10 mM Tris-HCl (pH 7.6), 100 mM NaCl, 1 mM EDTA], and fractionated by use of a 2.5%–30% Nicodenz (Nycomed Pharma AS, Oslo, Norway) gradient in TNE buffer at 27,000 rpm for 2.5 h at 4°C with an SW40 rotor (Beckman). The purification efficiency was confirmed by Coomassie Brilliant Blue staining and western blot analysis with antibodies to VP40 and NP. The infectious titer was determined by plaque assay, as described previously [62].
For purification of Ebola VLPs, equal amounts of the expression plasmids for EBOV VP40 [100], [101], NP [100], and GP [100], [101] were transfected into 293T cells by using TransIT LT-1 (Mirus, Madison, USA). Forty-eight hours post-transfection, the culture supernatants were harvested and released VLPs were purified, as described above. Incorporation of viral proteins in the purified VLPs was confirmed by western blot analysis with antibodies to VP40, GP and NP, and the morphology of the VLPs was confirmed by negative staining (Figure S3).
Influenza virus A/PR/8/34 was prepared and purified as described previously [102]. VSV pseudotyped with EBOV GP (VSVΔG*-GP) was generated as described previously [61] and purified as described above. Protein concentrations of the individual virion fractions were measured by use of a Bradford protein assay kit (BioRad, Hercules, USA).
Viral particles were fluorescently labeled as described by Sakai et al. [67]. Briefly, 1 ml of fractionated virions (100 µg/ml) was incubated with 6 µl of 100 µM stock solution of DiI in the dark for 1 h at room temperature with gentle agitation.
For real-time imaging of the internalization of DiI-labeled viral particles, Vero cells expressing CLCa-eGFP, Cav1-eGFP, eGFP-SNX5, eGFP-actin or eGFP-Rab7 were cultured in 35 mm glass-bottom culture dishes (MatTek corporation, Ashland, USA), washed in 1 ml of phenol red-free MEM (Invitrogen) containing 2% FBS and 4% BSA, and incubated with DiI-labeled virions in 50 µl of the same medium on ice for 30 min. The cells were washed with the ice-cold medium and incubated for various times in a temperature-controlled chamber on the stage of a confocal laser scanning microscope (LSM510 META, Carl Zeiss, Oberkochen, Germany); the chamber was maintained at 37°C with a humidified atmosphere of 5% CO2. Images were collected with a 40x oil objective lens (C-Apochromat, NA = 1.2, Carl Zeiss) and acquired by using LSM510 software (Carl Zeiss). For presentation in this manuscript, all images were digitally processed with Adobe Photoshop. For co-localization analysis, the images were acquired randomly, the number of DiI-labeled virions that co-localized with eGFP-SNX5 or eGFP-Rab7-positive vesicles were measured in 10 individual cells (approximately 10–20 dots/cell), and the percentage of co-localization in the total DiI-virions was determined for each time point. Each experiment was performed in triplicate and the results are presented as the mean ± standard deviation.
Target sequences corresponding to the human CHC [103], Cav1 [104],and Cdc42 [105] -coding sequences were selected, respectively (Table S1), and synthesized (Dharmacon, Lafayette, USA or Qiagen, Hilden, Germany). siRNA for Pak1 down-regulation was purchased from Cell Signaling (Trask Lane, USA). Synthesized siRNA was transfected into Vero cells by using TransIT-TKO (Mirus, Madison, USA). For analysis of the efficiencies of internalization of Ebola virions, DiI-virions were adsorbed to the siRNA-transfected cells 48 h post-transfection, as described above, and then incubated for 2 h at 37°C. Uninternalized surface-bound virions were removed by the addition of 0.25% trypsin for 5 min at 37°C and the number of DiI-virions in 10 individual cells was counted. Each experiment was performed in triplicate and the results are presented as the mean ± SD.
The efficiency of CHC and Cav1 down-regulation was assessed by immunofluorescent staining with antibodies specific to CHC and Cav1 (Abcam, Cambridge, UK). The down-regulation of endogenous Cav1 was also examined by western blot analysis by using an antibody specific to Cav1 (Abcam). The efficiencies of Cdc42 and Pak1 [106] were assessed by RT-PCR with oligonucleotides to amplify each gene (Table S1).
Vero cells or Vero cells expressing eGFP-Rab7 cultured in 35 mm glass-bottom culture dishes were pretreated with 100 µM dynasore (Sigma-Aldrich), 2 µM cytochalasin D (Sigma-Aldrich), 50 µM LY294002 hydrochloride (Sigma-Aldrich), 50 nM wortmannin (Sigma-Aldrich), 100 µM EIPA (Sigma-Aldrich) or 100 nM staurosporine (Sigma-Aldrich) for 30 min at 37°C. DiI-labeled virions were adsorbed to the cells for 30 min on ice in the presence of these inhibitors in phenol red-free MEM (Invitrogen) containing 2% FBS and 4% BSA. Cells were then washed with the same medium and incubated for 2 h at 37°C in the presence of the inhibitors. As a control, cells were treated with dimethyl sulfoxide (DMSO, Sigma-Aldrich). Efficiencies of internalization of DiI-labeled viral particles into Vero cells or co-localization of DiI-labeled viral particles with eGFP-Rab7 were analyzed by using confocal laser scanning microscope as described above.
Vero cells treated with dynasore, transiently expressing dominant-negative Rac1, were incubated with DiI-labeled virions on ice for 30 min in MEM containing 2% FBS and 4% BSA. The cells were washed with the same medium and then incubated for 2 h at 37°C. Cells were then incubated with 2 µg/ml Alexa Fluor 633-Tf for 10 min or 0.5 mg/ml Alexa Fluor 647-Dex Mw 10K for 60 min at 37°C. To remove surface-bound labeled virions, Tf, or Dex Mw 10K, the cells were treated with trypsin as described above for 5 min at 37°C. Cells were then washed twice with the same medium and internalized DiI-labeled virions, Tf or Dex Mw 10K were analyzed by use of confocal laser scanning microscope. To assess the effect of staurosporine or the siRNA treatment on fluid phase uptake, after staurosporine pretreatment or 48 h post-transfection of individual siRNAs, Vero cells were incubated with 0.5 mg/ml AlexaFluor 647-Dex Mw 10K, harvested by treating with trypsin, washed twice with ice-cold PBS, and fixed with 4% PBS-buffered paraformaldehyde for 10 min at room temperature. The mean fluorescence intensities in the cells were analyzed by use of flow cytometry (FACSCalibur; Becton Dickinson, Franklin Lakes, USA).
VSV pseudotyped with EBOV GP (VSVΔG*-GP) or VSV G (VSVΔG*-G) expressing GFP was generated as described previously [61]. Vero cells were treated with a series of inhibitors for 30 min at 37°C and infected with each virus at a multiplicity of infection (as titrated with Vero cells) of 0.002 to 0.005 in the presence of the inhibitors. 1 h post-infection, surface-bound virions were removed by trypsin as described above and cultured for another 24 h. Infection efficiencies for VSVΔG*-GP or VSVΔG*-G were determined by measuring the number of GFP-positive cells by conventional fluorescent microscope. Each experiment was performed in triplicate and the results are presented as the mean ± SD.
Methods for supporting information files are described in Text S1.
|
10.1371/journal.pntd.0000481 | The Role of Human Movement in the Transmission of Vector-Borne Pathogens | Human movement is a key behavioral factor in many vector-borne disease systems because it influences exposure to vectors and thus the transmission of pathogens. Human movement transcends spatial and temporal scales with different influences on disease dynamics. Here we develop a conceptual model to evaluate the importance of variation in exposure due to individual human movements for pathogen transmission, focusing on mosquito-borne dengue virus.
We develop a model showing that the relevance of human movement at a particular scale depends on vector behavior. Focusing on the day-biting Aedes aegypti, we illustrate how vector biting behavior combined with fine-scale movements of individual humans engaged in their regular daily routine can influence transmission. Using a simple example, we estimate a transmission rate (R0) of 1.3 when exposure is assumed to occur only in the home versus 3.75 when exposure at multiple locations—e.g., market, friend's—due to movement is considered. Movement also influences for which sites and individuals risk is greatest. For the example considered, intriguingly, our model predicts little correspondence between vector abundance in a site and estimated R0 for that site when movement is considered. This illustrates the importance of human movement for understanding and predicting the dynamics of a disease like dengue. To encourage investigation of human movement and disease, we review methods currently available to study human movement and, based on our experience studying dengue in Peru, discuss several important questions to address when designing a study.
Human movement is a critical, understudied behavioral component underlying the transmission dynamics of many vector-borne pathogens. Understanding movement will facilitate identification of key individuals and sites in the transmission of pathogens such as dengue, which then may provide targets for surveillance, intervention, and improved disease prevention.
| Vector-borne diseases constitute a largely neglected and enormous burden on public health in many resource-challenged environments, demanding efficient control strategies that could be developed through improved understanding of pathogen transmission. Human movement—which determines exposure to vectors—is a key behavioral component of vector-borne disease epidemiology that is poorly understood. We develop a conceptual framework to organize past studies by the scale of movement and then examine movements at fine-scale—i.e., people going through their regular, daily routine—that determine exposure to insect vectors for their role in the dynamics of pathogen transmission. We develop a model to quantify risk of vector contact across locations people visit, with emphasis on mosquito-borne dengue virus in the Amazonian city of Iquitos, Peru. An example scenario illustrates how movement generates variation in exposure risk across individuals, how transmission rates within sites can be increased, and that risk within sites is not solely determined by vector density, as is commonly assumed. Our analysis illustrates the importance of human movement for pathogen transmission, yet little is known—especially for populations most at risk to vector-borne diseases (e.g., dengue, leishmaniasis, etc.). We outline several important considerations for designing epidemiological studies to encourage investigation of individual human movement, based on experience studying dengue.
| For vector-borne pathogens heterogeneity in patterns of contact between susceptible hosts and infectious agents is common [1],[2],[3]. Some hosts will be exposed to, harbor, and pass on more parasites than others. Variation in contact patterns can amplify [4],[5] or dampen [6] the rate of transmission, even as it also potentially reduces disease prevalence and epidemic stability (i.e., likelihood of an outbreak; [7]). Understanding and describing what drives heterogeneous contact patterns is thus important for designing improved disease surveillance and prevention programs [3]. If the characteristics of hosts most often infectious or important for transmission are known they could be targeted to more efficiently prevent disease [8]. To be useful for targeted control across different contexts the mechanisms underlying heterogeneous contact patterns must be elucidated. Here we examine the role of individual human movement as a critical behavioral factor underlying observed patterns of vector-borne pathogen transmission, because movement determines exposure to infectious agents; i.e., bites from infected mosquito vectors.
Little is known about individual human movement patterns and even less about their epidemiological consequences, even though such knowledge would be a valuable contribution to the understanding and control of many vector-borne diseases. We begin our investigation of this topic by reviewing studies of human movement. Next, based on an existing typology, we examine the relevance of movement patterns to the dynamics of different diseases. Using the mosquito-borne virus dengue as an example, we develop a conceptual model that illustrates how human and vector behavior can influence pathogen transmission dynamics. We end by outlining key issues important to the design of future research and explaining potential benefits to disease prevention of an improved understanding of host movement.
Historically epidemiologists have viewed human movement from the perspective of populations of susceptible hosts moving into high risk areas or infected hosts moving into susceptible populations as explanation for disease occurrence and spread. Indeed, across different scales and diseases, movements of hosts affect pathogen transmission in a variety of ways. Thirty years ago Prothero [9] provided a typology to facilitate study of the role of human movements in epidemiology based on his experience in Africa. Drawing on geography literature concerned with understanding human movement [10],[11],[12], Prothero highlighted the difference between circulatory movements, where individuals return home after some period, and migratory movements, which tend to be permanent changes of residence (see Figure 1 in [11]). He further characterized movements by their ‘spatial scale’, which he categorized in terms of a rural-urban gradient, and temporal scale based on the time and timing of displacements. He qualified these categories in terms of their relevance to public health. For instance, seasonal movements from one rural area to another for agriculture could potentially expose individuals to different ‘ecological zones’ where the risk of malaria or African trypanosomiasis is high [13]. His argument was that knowing something about the nature of such movements would help explain the incidence and prevalence of disease in a population and provide informed options for control [9]. In Figure 1 we generalize Prothero's typology in terms of the spatial and temporal scale (sensu [14]) of human movement and extend it to include most vector-borne disease contexts.
At broad spatial scales (e.g., national, international) individual movements drive pathogen introduction and reintroduction (far right, Figure 1). Global spread of dengue virus via shipping routes was characterized by periodic, large, spatial displacements [15]. Globalization and air transportation have changed the dynamic of pathogen spread by dramatically shortening the time required to travel around the globe [16],[17],[18]. The recent chikungunya epidemic in the Indian Ocean that subsequently spread to Italy is an example [19]. At finer scales (e.g., regional, urban-rural, intra-urban; far left of Figure 1), movement associated with work, recreation, transient migration, and other phenomena is important to patterns of pathogen transmission and spread [9],[20]. Movements into high-risk areas not only lead to individual infection, but can also contribute to local transmission when infected hosts return home and infect competent vectors. For example, in the Chocó region of Colombia most malaria transmission occurs in rural areas and many cases diagnosed in the city of Quibdó are due to travel to these areas [21]. Transmission also occurs locally within Quibdó [22], however, most likely because of infected travelers returning and infecting competent vectors. Understanding the origin of infections and the relative importance of human movement at different scales to both local and regional transmission dynamics would increase effectiveness of disease prevention programs by, for example, identifying individuals at greatest risk of contracting and transmitting pathogen.
Generally, a key significance of human movement for vector-borne disease at any scale lies with exposure to vectors. Exposure is local in space and time and variation in exposure due to individual host movement could strongly influence the transmission dynamics of pathogens. For instance, circulatory movements associated with working in rural areas and variation in movement patterns among cultures may explain heterogeneous patterns of onchocerciasis incidence. While men in Cameroon and Guatemala both experience similar parasite loads reflecting exposure to vectors when working in fields, women in the 2 countries show different patterns of infection partly due to differences in exposure [23]. The type of movement most relevant for exposure will depend on site specific differences, the ecology of the arthropod vector, human behavior, and the relative scale of host and vector movement. For pathogens transmitted by vectors able to move long distances in search of a host, fine scale host movements may not be important, while they are for pathogens transmitted by sessile vectors. Aedes aegypti—the principal vector of dengue virus—bites during the day [24], disperses only short distances [25] and is heterogeneously distributed within urban areas [26],[27]. Conversely, humans move frequently at local scales (bottom-left of Figure 1), allocating different amounts of time to multiple locations on a regular basis. This will influence individual risk of infection with dengue virus [28] and thus overall patterns of transmission [29],[30],[31].
The dynamics of human movement, the locations used and the paths between them, is conceptualized by the ‘activity space’ model developed in the 60's by human geographers [12],[32],[33]. This model, much like the ‘home-range’ concept in ecology, is effective because organisms exhibit habitual behavior in their use of space [34]. For our purposes of studying dengue, the ‘activity space’ refers to those few locations where humans commonly spend most of their time [32],[35] and ‘movement’ refers to the use of these locations. Thus, exposure to host-seeking female Ae. aegypti is the sum of exposure across an individual's activity space. For other vectors and pathogens, human movements per se (e.g., walking between the house and a water source) and/or visits to less common destinations could be relevant for the transmission of other pathogens (e.g., African trypanosomiasis) depending on the behavior of the vector and the relative scales of vector and host movement.
The activity space model represents movement associated with the regular activity of individuals [36]. We present a version of this model in Figure 2 for understanding how movements within an urban area might contribute to risk of exposure. Risk at locations within an individual person's activity space will vary depending on the number of infected, host seeking vectors present and their biting behavior. For instance, visits to locations during the day are of minimal risk for bites from nocturnal An. gambiae, but are relatively high for day active Ae. aegypti (Figure 2). Exposure to vector bites may also depend on how long a person stays at a given location. If vectors are stimulated by the arrival of an individual to a location (as may be the case for Ae. aegypti and Aedes albopictus [37],[38]), then a bite is most likely to occur early after arrival (i.e. the cumulative probability of a bite during a visit, e(t), accumulates rapidly). Alternatively, for vectors like triatomine bugs, which are less opportunistic than mosquitoes, long visits will be expected to pose a higher risk of host-vector contact (e(t) slowly accumulates over time). How vectors respond to hosts arriving at a site is important because it weights the risk of visits differently depending on their frequency and duration. If a vector is stimulated to host seek by the arrival of a host, then multiple short visits to that site will carry greater risk than a single long visit of equivalent total duration.
In summary, a person's risk of exposure to an infective vector can be represented with a simple exposure model for indirectly transmitted disease:(1)
Here, the risk of exposure (i.e., being bitten by a vector) for individual i, ri, over some observation period is simply the sum across sites visited, j, of vector abundance, Vj, conditioned on the time and duration of all visits to that site, k, as determined by vector behavior (where K is the total number of visits during the observation period). The biting rate, ak, is the number of bites expected per visit and is drawn from the day biting rate distribution for the times of the visit.(2)
How vectors respond to the appearance of a host at a site is captured by ek, the cumulative probability of a bite given the time spent in the site, and is bounded by the unit interval.(3)
Visits, k, are defined by an arrival time, t0, and a departure time, t1, in hours and are in reference to a single day. At the limit (where t1−t0 = 24 hours), ak becomes the day biting rate, a, and ek goes to 1 and we recover the model often assumed for vector-borne diseases where exposure occurs in the household. Note that although we imply here that a site comprises a household or other edifice because of our focus on dengue, in truth it simply demarcates a location where the abundance and activity of vectors is independent of other locations and is defined by the scale of vector movement.
Site-specific exposure risk is calculated as:(4)and has units of bites*humans for the observation period. Note that in this formulation, risk among individuals using the same site is assumed to be independent (i.e., the expected number of bites at a site is the product of humans present and vector activity). This may not be realistic if hosts occupy a site at the same time, which would be expected to dilute the number of bites individual hosts receive, and can be corrected (see below) by incorporating the actual amount of time individual humans spend in a location. The estimate of risk, rj, can be used to estimate the transmission rate, R0, which is the number of secondary infections expected from the introduction of a single infective individual into a wholly susceptible population. Woolhouse et al. (1997) use the following approximation for R0:(5)where vj is the proportion of vectors at site j, hj is the proportion of hosts living in site j, and J is the total number of sites. Risk as estimated above is incorporated by replacing vj with site associated risk, rj, discounted by the proportional use of that site within some interval by people, hj:(6)
For example, if a site is used by 2 individuals for 6 hours each over a week, hj = (2 humans * 6 hours)/(24 hours/day * 7 days) = 0.07 humans. The activity space model elaborated here illustrates that host and vector behavior are very important for determining who gets bitten and has the greatest risk of contracting or transmitting a pathogen.
The activity space model when coupled with our knowledge of vector behavior provides a tool for determining what human movements are important for transmission (e.g., Figure 1). Specifically, it allows us to identify places and individuals that contribute disproportionately to pathogen transmission dynamics. For example, consider the following scenario depicted in Figure 3 for a human population at risk for dengue virus infection like the one we are studying in Iquitos, Peru (Figure 3, Text S1 and Table S1). Briefly, individuals spend their time at a number of different sites, both commercial and residential, during their regular weekly activities (Sites, first column in Figure 3). Sites have different numbers of female mosquitoes and are visited at different rates and for different durations. We can estimate the risk of exposure to host-seeking female mosquitoes (ri) for each person (columns 1–13 in Figure 3) at each site (rows in Figure 3) and then estimate R0. In this particular example, R0 as approximated when accounting only for the home (eq. 5) is 1.3 and the site with the highest estimated risk is house 5 (in bold in column under R0). If we account for exposure at all locations in addition to the home and assume the biting rate at night is 10% of the rate during the day [39], our estimate of R0 (eq. 6) jumps nearly 3-fold and the most important site is 13, a clinic (in bold under R0e). This latter result arises because of the relatively large number of bites per person expected at that site, determined largely by the significant amount of time a single person spends there (e.g., their workplace). In this example, all individuals except individual 10 experience the greatest exposure to bites in their homes because that is where they spend the most time. Individual 10, however, experiences the highest risk at site 4, which represents their workplace. This individual is also at the greatest risk in the host population.
This example illustrates that the key sites are not necessarily those of greatest vector abundance, as is commonly assumed. For this example scenario, R0j increases monotonically with vector abundance when transmission is assumed to occur only in the home (Figure 4). When exposure rates are accounted for, however, there is no relationship between R0j and vector abundance (Figure 4). Similarly, people living where vector abundance is greatest are not necessarily at greater risk. Human movement and subsequent variation in exposure thus becomes more important than vector density per se. Because heterogeneity in contact patterns has a large influence of the rate of pathogen transmission, variation in exposure rates due to individual movement patterns could have considerable impact on disease dynamics [40],[41].
To fully understand the implications of movements, however, data should be incorporated into network, individual-based or metapopulation models [5],[42],[43]. Network models, in particular, capture heterogeneity explicitly and intuitively, allowing precise prediction of trends and patterns in human infection and disease [44]. For dengue, one imagines a dynamic network of individuals most likely to become infected or infect mosquitoes and of locations where transmission is most likely to occur [29]. These are the key nodes of pathogen transmission that, if identified and understood, would be excellent targets for intervention (e.g., [8]).
The value of estimating actual exposure rates and incorporating these into models to better understand pathogen dynamics is clear for dengue, which is mostly transmitted when people are engaged in daily activities [29]. For this reason we are currently monitoring human movements in Iquitos, Peru. The activity space model as we describe it, however, highlights that movements may be important for the transmission of many pathogens typically thought to be transmitted at night when hosts are inactive. Sand fly vectors of American visceral leishmaniasis are active at dusk [45], move short distances [46], and are heterogeneously distributed among homes [47], which, in combination with human behavior, may be key to understanding leishmaniasis incidence patterns [48],[49]. Similarly, Michael et al. [50] found that 27% of Culex quinquefasciatus resting within households had fed on hosts from outside that home despite its nocturnal habit, with implications for transmission of lymphatic filariasis. There are thus many reasons for increased examination of individual human movement patterns.
As an aid to future research, in the remainder of this article we discuss key issues and considerations for designing studies of human movement based on our experiences with dengue.
Because patterns of contact between pathogens and susceptible hosts are heterogeneous, disease interventions can be made more effective and efficient by targeting the key points or ‘nodes’ of transmission [3]. Even where heterogeneous patterns are clearly documented, not knowing the factors driving such patterns impedes one's ability to effectively target control. Is a biting preference toward young adults [60] because they are intrinsically more attractive to a host-seeking mosquito or, because of their behavior, they are more likely to be exposed to mosquitoes? Although many different causes of host-vector contact heterogeneity have been proposed (summarized by [6]), variation in exposure due to human behavior is likely to be key across disease systems. The role of other risk factors (e.g., host-preference) will always be conditioned by exposure rates. The study of human movement is thus critical to the identification of key individuals and key locations. Nevertheless, movements have largely been neglected in studies of indirectly transmitted disease even though it is becoming increasingly easy to measure.
Quantifying and describing human movements promises more than just characterization of key heterogeneities. Quantification of the collective dynamics of human populations provides information necessary for models intended to predict disease outbreak and spread and to evaluate control alternatives to halt epidemics [8],[35],[51]. Buscarino et al. [61], for instance, predict that movements within a population have an important effect on the epidemic threshold, lowering this as individuals move over larger distances more frequently. Additionally, quantifying movements and applying that information to a variety of diseases creates the opportunity to identify common places where infection occurs across diseases and, thus, the potential to leverage public health programs by allowing limited resources to be targeted to the most important locations for more than one disease.
Rigorous examination of the role of human movement across different scales will significantly improve understanding of pathogen transmission, which will be critical to increasing the effectiveness of disease prevention programs. As transmission rates are reduced through intervention efforts, we expect the importance of heterogeneity in exposure to increase and to play an even more important role in pathogen persistence. Characterization of movements will thus not only facilitate the elimination of disease, it will help to prevent its return.
|
10.1371/journal.ppat.1000780 | HIV Controller CD4+ T Cells Respond to Minimal Amounts of Gag Antigen Due to High TCR Avidity | HIV controllers are rare individuals who spontaneously control HIV replication in the absence of antiretroviral treatment. Emerging evidence indicates that HIV control is mediated through very active cellular immune responses, though how such responses can persist over time without immune exhaustion is not yet understood. To investigate the nature of memory CD4+ T cells responsible for long-term anti-HIV responses, we characterized the growth kinetics, Vβ repertoire, and avidity for antigen of patient-derived primary CD4+ T cell lines. Specific cell lines were obtained at a high rate for both HIV controllers (16/17) and efficiently treated patients (19/20) in response to the immunodominant Gag293 peptide. However, lines from controllers showed faster growth kinetics than those of treated patients. After normalizing for growth rates, IFN-γ responses directed against the immunodominant Gag293 peptide showed higher functional avidity in HIV controllers, indicating differentiation into highly efficient effector cells. In contrast, responses to Gag161, Gag263, or CMV peptides did not differ between groups. Gag293-specific CD4+ T cells were characterized by a diverse Vβ repertoire, suggesting that multiple clones contributed to the high avidity CD4+ T cell population in controllers. The high functional avidity of the Gag293-specific response could be explained by a high avidity interaction between the TCR and the peptide-MHC complex, as demonstrated by MHC class II tetramer binding. Thus, HIV controllers harbor a pool of memory CD4+ T cells with the intrinsic ability to recognize minimal amounts of Gag antigen, which may explain how they maintain an active antiviral response in the face of very low viremia.
| HIV infection, if left untreated, leads to the progressive disruption of the immune system, the destruction of the CD4+ T cell population, and the occurrence of multiple opportunistic infections. However, a small fraction of HIV-infected individuals (less than 1%) avoid these deleterious effects by spontaneously controlling HIV replication to very low levels in the absence of antiretroviral therapy. Emerging evidence indicates that these rare patients, named HIV controllers, contain HIV through a very active T cell-mediated immune response. In this study, we found that memory CD4+ T cells from HIV controllers had the capacity to respond to minimal amounts of antigen derived from the viral protein Gag. This property was intrinsic to controller CD4+ T cells, and resulted from the expression of T cell receptors (TCRs) with high avidity for a particular Gag peptide. The presence of high avidity CD4+ T cells may explain how HIV controllers maintain the antiviral immune response in constant alert, even though the amount of virus inducing this response is minimal. Based on this study, we propose that future candidate vaccines against HIV should induce high avidity memory CD4+ T cells, to mimic the rapid and persistent antiviral response characteristic of HIV controllers.
| HIV controllers are rare individuals who spontaneously control HIV replication in the absence of antiretroviral treatment [1],[2]. HIV controllers harbor plasma viral loads that remain undetectable by conventional assays and cell-associated HIV DNA loads that are in the very low range, close to one log below those detected in patients receiving efficient antiretroviral therapy [3]–[5]. HIV controllers show a very low risk of progression to AIDS [3], emphasizing the importance of limited viral dissemination in maintaining a healthy status in the long term.
Emerging evidence indicates that controllers suppress HIV replication through a very active immunological process. HIV controllers harbor effector memory CD8+ T cells capable of rapidly killing infected autologous CD4+ T cells through a cytotoxic mechanism involving the upregulation of perforin and Granzyme B [6],[7]. Signs of immune activation are more prominent in HIV controllers than in efficiently treated patients, and include increased plasma LPS [8], increased expression of T cell activation markers [9], and increased propensity to secrete IFN-γ and MIP-1β upon polyclonal stimulation [10]. Longitudinal studies of efficiently treated patients who achieve undetectable viral load have shown a waning of cellular antiviral responses, which paralleled the progressive decrease in viral burden [11]. In contrast, HIV controllers maintain polyfunctional effector memory T cells with the capacity to secrete multiple cytokines [12]–[14]. How controllers maintain an active antiviral response in the long term in spite of a very low viral burden remains poorly understood.
One element contributing to the persistence of an active immune response may be the quality of the HIV-specific central memory (CM) compartment. CM T cells are thought to be responsible for the long-term maintenance of immune memory, due to their long half-life, high proliferative potential, and capacity to replenish the pool of effector and effector memory (EM) T cells that directly control pathogens [15]–[17] The progressive depletion of the CM CD4+ T cell compartment parallels disease progression in a simian model of AIDS [18]. CM CD4+ T cell functions, such as proliferation and IL-2 secretion, are impaired as early as the primary infection stage in progressive HIV infection [19]–[21], and are only partially recovered in efficiently treated patients [22],[23]. Chronic antigenic stimulation is thought to drive an accelerated differentiation of CM into effector CD4+ T cells, and thus contribute to T cell exhaustion. Importantly, CM CD4+ T cell numbers and functions are preserved in HIV controllers, who appear protected from this accelerated differentiation process [24],[25]. A recent study suggests that inactivation of pro-apoptotic molecules may contribute to the remarkable proliferative capacity of CM CD4+ T cells of HIV controllers, which can exceed that seen in healthy controls after non-specific stimulation [26].
We have previously shown that signs of CD4+ T cell immune activation could be detected in HIV controllers who nevertheless had an intact CM CD4+ T cell compartment, with preserved IL-2 secretion capacity and efficient proliferative responses [10]. How chronic immune activation was induced in controllers, and why it did not generally lead to accelerated CD4+ T cell differentiation and exhaustion remained unclear. To explore these issues, we tested the capacity of Gag-specific memory CD4+ T cell to differentiate in vitro, comparing primary CD4+ T cell lines derived from HIV controllers and efficiently treated patients with equivalent duration of infection. We found that HIV controller harbored a pool of memory CD4+ T cells able to differentiate into effector cells with high functional avidity for an immunodominant Gag epitope. This heightened sensitivity to Gag antigen could be explained by a high avidity interaction between the TCR and the peptide/MHC complex, as measured by class II tetramer binding. The capacity to mount a CD4 recall response in the presence of minimal amounts of Gag antigen may help explain how HIV controllers maintain a continuously activated antiviral response in spite of very low viremia.
Memory CD4+ T cell responses were compared in patients who spontaneously controlled HIV replication (HIC group, n = 17) and in patients who achieved viral control following successful antiretroviral therapy (HAART group, n = 20). Patients in both groups had viral loads <40 HIV RNA copies/ml plasma. The duration of infection and the CD4+ T cell count did not differ significantly between the two groups (Table 1).
We analyzed the properties of memory CD4+ T cell precursors by determining their capacity to generate CD4+ T cell lines specific for three immunodominant HIV-1 Gag peptides (Table 2). The peptides were chosen because of their broad immunodominance and their capacity to bind multiple HLA-DRB1 alleles [20], [27]–[30]. The frequency of response was determined by measuring the percentage of patients for whom viable CD4+ T cell lines (defined by a growth ratio >0.7 at day 14) could be obtained after stimulation with a Gag 20-mer peptide. PBMC from HIV-seronegative donors did not yield viable cell lines (not shown). 89 out of 90 cell lines obtained from HIV-seropositive donors proved peptide-specific, as indicated by a positive IFN-γ response measured in ELISPOT assay. The frequency of response to the most immunodominant peptide, Gag293, was remarkably high in both the HIC and HAART groups, with 94% and 95% of responders, respectively (Table 3). Responses to the second peptide, Gag263, were also frequent, with 82% responders in the HIC group and 77% responders in the HAART group. These findings confirmed that several CD4 epitopes in Gag could achieve strong immunodominance in patients with controlled HIV-1 infection. Interestingly, responses to the third peptide, Gag161, were more frequent in the HAART group than in the HIC group, with 91% versus 53% responders, respectively (P<0.05). We did not detect an association between the lack of response to Gag161 and particular HLA-DR genotypes. Ex vivo IFN-γ ELISPOT responses to the 3 Gag peptides were low, as expected for CD4 responses directed to single HIV peptides (Fig. S1). However, it was interesting to note that 7/13 Controllers had a detectable ex vivo response to Gag293 while only 1/13 Controller responded to Gag161 (P<0.05). This finding supported the notion of a higher frequency of Gag293-specific than Gag161-specific CD4+ T cells in Controller patients. Taken together, these observations suggested that the HIV controller status may be associated with a change in the immunodominance pattern of Gag CD4 epitopes.
As controls, we generated CD4+ T cell lines specific for the CMV pp65 protein. Since the immunodominance pattern of CMV responses proved more variable than that observed for HIV-1, we optimized the generation of CMV-specific CD4+ T cell lines by choosing the pp65 peptide in function of the HLA DR genotype of the patients (Table 2) [31]–[33]. Using this strategy, close to half of the patients responded to pp65 peptides in both groups, which was consistent with CMV seroprevalence in the studied populations.
CD4+ T cell lines typically showed an initial loss of cells due to apoptosis, followed by growth due to multiplication of HIV- or CMV-specific cells. Measurement of the growth ratio at day 7 showed that viable CD4+ T cells lines from HIV controllers had a faster growth kinetics than those from treated patients, with a significant difference in response to the 3 Gag peptides but not in response to CMV peptides (Fig. 1A). Analysis of CD4+ T cell lines derived from a control group of untreated patients with HIV-1 viremia (VIR) showed limited growth capacity in all cases, consistent with the notion that active HIV-1 replication impaired the proliferative capacity of memory CD4+ T cells [19]. Measurement of growth ratios at day 14 (Fig. 1B) confirmed the rapid amplification of controller CD4+ T cell lines in response to Gag but not to CMV peptides. The clearest differences between the HIC and HAART groups were seen in response to Gag293 (P = 0.003), suggesting that HIV controllers harbored CD4+ T cell precursors with particularly good proliferative capacity in response to this immunodominant Gag peptide.
Analysis of IFN-γ production by ELISPOT at day 8 showed a prominent response in HIV controller CD4+ T cell lines (median SFC/106 cells = 7,716), while most cell lines from treated patients remained negative. However, the fact that cell lines from treated patients had not yet entered the exponential growth phase could account for these differences. To compare CD4+ T cell lines at equivalent growth stages, all following measurements were made at doubling time (mean doubling time = 8 days in the HIC group, 13 days in the HAART group, and 14 days in the VIR group). In these conditions, IFN-γ production remained higher in the HIC group compared to the HAART group in response to the immunodominant Gag293 peptide (Fig. 2, P = 0.006). In control experiments, CD4+ T cell depletion abrogated the ELISPOT signal, confirming that IFN-γ was produced by CD4+ T cells, and not by the CD8+ T cells that may have escaped CD8 depletion (Fig. S2).
IFN-γ ELISPOT responses measured at doubling time were higher in the HIC group than in the VIR group for all peptide tested (Fig. 2). However, we and others have previously shown that when responses are measured ex vivo, which gives an evaluation of ongoing effector responses, CD4+ T cells from viremic patients produce as much IFN-γ as those of HIV controllers, while CD4+ T cells from treated patients have low IFN-γ production [10],[11],[22],[24]. The hierarchy of IFN-γ responses measured after proliferation of CD4+ T cell memory precursors is different, and rather reflects the capacity of these precursors to differentiate into cells with effector functions. We conclude that HIV controllers harbor memory CD4+ T cell precursors that can differentiate into efficient cytokine-secreting cells.
To assess the sensitivity of memory CD4+ T cells to antigenic stimulation, we measured IFN-γ production in response to serial peptide dilutions (Fig. 3). We did not observe a significant difference in the dose of Gag293 peptide that induced a half-maximal ELISPOT response (median EC50 = 1.38 10−6 M in HIC vs. 1.55 10−6 M in HAART, P = 0.09). However, we observed that the shapes of the response curves differed, with a marked trailing end in the HIC group, suggesting the presence of a high avidity component within the responding CD4+ T cell population (see representative examples in Fig. 3A and Fig. S3). To extend this observation, we measured the last peptide concentration that gave a positive ELISPOT reading at least 2 fold above background. We verified that measurement of this concentration was reproducible in duplicate experiments (Table S1). Importantly, all measurements were carried out on CD4+ T cell lines at doubling time, to normalize for growth stage.
This analysis revealed that the functional avidity of CD4+ T cells recognizing the immunodominant Gag293 peptide was higher in HIV controllers (Fig. 3B). In the HIC group, 10 out of 15 patients tested had a positive ELISPOT reading at peptide concentrations ≤5 10−8 M, while in the HAART group only 1 out of 17 patients tested had a positive response at the same concentrations (P = 0.005). In contrast, functional avidities measured for Gag263, Gag161, and CMV peptides did not differ significantly between groups. These results pointed to a particular efficiency of CD4+ T cells specific for the immunodominant Gag293 in HIV controllers. Interestingly, the functional avidity of these cells correlated with the level of the IFN-γ response measured by ELISPOT assay at high peptide dose (R = −0.68, P<0.0001). Thus, CD4+ T cells responses to Gag293 appeared both sensitive and potent in the group of controller patients.
Patients were genotyped for the HLA DRB1 gene at 4 digit resolution. Table 4 reports the frequency of the most common HLA DRB1 alleles among the 34 HIV controllers and the 34 efficiently treated patients who were genotyped at initiation of the study. Interestingly, the frequency of the DRB1*0701 allele was 44% in the controller group and 18% in the treated patient group, which yielded a significant difference as measured by Fisher's exact test (P = 0.03). No other DRB1 allele showed significant differences. The frequency of the DRB1*0701 allele in the French population was reported to be 26% [34], close to that seen in the group of treated patients. In contrast, the frequency of DRB1*0701 appeared increased among HIV controllers.
To further explore the possibility that DRB7*0701 conferred an advantage in CD4+ T cell memory function, we compared functional parameters in DRB1*0701 positive versus DRB1*0701-negative individuals included in the study. We did not detect significant differences within the HIC group in terms of growth ratio, IFN-γ response, or functional avidity of CD4+ T cells following Gag293 stimulation. Within the HAART group, DRB1*0701-positive individuals showed lower IFN-γ responses (median SFC/106 cells = 2458, n = 6) than DRB1*0701-negative individuals (median SFC/106 cells = 6116, n = 9, P = 0.02). Taken together, these findings suggest that the DRB1*0701 allele may confer an increased chance to acquire a controller phenotype upon HIV infection, but that once the controller phenotype is established, the presence of the DRB1*0701 allele does not confer further benefit in terms of CD4+ T cell function.
To further characterize the nature of Gag293-specific CD4+ T cells, we identified these cells through MHC class II tetramer labeling. Analysis of Gag293-specific CD4+ T cell lines revealed the presence of tetramer-positive (Tet+) cells for all the patients tested, confirming the antigen specificity of the cell lines (representative examples in Fig. 4A and 4B), and the capacity of the Gag293 peptide to bind multiple HLA-DR alleles [27]. At doubling time, the frequency of Tet+ cells was in the order of 1%. (Fig. 4A). The population of tetramer-negative (Tet−) cells may correspond to Gag293-specific cells restricted by an HLA-DR, -DP, or DQ allele distinct from that used in the tetramer, or to cells amplified through bystander effect. At later time points, the population of Tet+ cells could reach up to half of the CD4+ T cells (Fig. 4B), suggesting an efficient amplification of Gag293-specific cells restricted through HLA-DR. Comparison of the percentage of Tet+ cells at doubling time showed no significant difference between the HIC and HAART groups (Fig. 4C), which validated our normalization strategy. Namely, CD4+ T cell lines analyzed at equivalent growth stages contained equivalent numbers of peptide-specific cells, and could thus be usefully compared.
To determine the proliferative capacity of Tet+ cells, Gag293-specific CD4+ T cell lines were labeled with CFSE at doubling time and analyzed by flow cytometry 3 days later (Text S1, supplementary methods). The percentage of CD4+ Tet + cells that had divided (CFSElo) was comparably high for the 3 HIC and 3 HAART cell lines analyzed (Fig. S4). The proliferative index, which represents the average number of divisions undergone by the population that divided, showed a trend toward higher values in the HIC cell lines. Interestingly, the difference became more apparent when the number of cells that had undergone 5 divisions or more was computed. In HAART cell lines, 6 to 7% of Tet+ cells had undergone 5 or more divisions, while in HIC cell lines these percentages were of 40%, 35%, and 18%. These data suggested that Tet+ cells from HIV controllers comprised a population endowed with an intrinsically high proliferative capacity and a short generation time.
The TCR Vβ specificities of Tet+ and Tet− cells within CD4+ T cell lines were determined at doubling time by immunostaining with a panel of anti-Vβ antibodies (Fig. 5A and B). The TCR Vβ repertoire of Gag293-specific CD4+ T cells was diverse in both the HIC and HAART groups and varied depending on the individual. Table 5 lists the Vβ specificities showing an amplification within the Tet+ population, as defined by a ratio Tet+/Tet− ≥4 in a Tet+ population ≥2% of CD4+ T cells. The frequency of amplified Vβ populations is reported in supplementary Table S2. This analysis showed that Vβ1 was amplified in 3 out of 4 cell lines in the HAART group and in 2 out of 4 cell lines in the HIC group. Vβ9 and Vβ13.2 were also amplified in half of cell lines tested. These observations suggest that some Vβ chains may be preferentially selected in response to the Gag293 peptide. However, we did not detect a Vβ signature characteristic of the HIC group. Taken together, these data suggest that multiple clones contribute to the high avidity memory CD4+ T cell population in HIV controllers.
The high functional avidity of controller CD4+ T cells could result from increased avidity of the TCR for the pMHC complex, or from multiple factors that facilitate APC/T cell interactions or effector functions, including expression tuning of costimulatory molecules or efficiency of the IFN-γ secretion system [35]. To explore this issue, we set to directly test the avidity of the TCR/pMHC interaction. The TCR avidity was evaluated by measuring the percentage of Gag293-specific Tet+ CD4+ T cells detected as a function of decreasing class II tetramer concentrations, as described in reference [36] (Fig. 6A). The concentration measured at half-binding (EC50) did not show significant differences between the HIC and HAART groups. However, the shape of the binding curves differed between groups, with a persistence of detectable binding at low tetramer concentrations in the HIC group (Fig. 6B). The TCR avidity was measured by the inverse of the last concentration that gave a Tet+ staining at least 2 fold higher than control CLIP-tetramer staining. Comparison of Gag293-specific CD4+ T cells at doubling time showed that the TCR avidity was significantly higher in the HIC group than in the HAART group (Fig. 6C), indicating a difference in the nature of CD4+ T cell clones responding to the immunodominant Gag293 epitope. Thus, the high functional avidity of HIV controller CD4+ T cells could be explained, at least in part, by an intrinsic property of their TCR.
This study provides evidence that HIV controllers harbor a pool of high avidity memory CD4+ T cell precursors directed against an immunodominant Gag peptide. Memory CD4+ T cells specific for the Gag293 peptide were endowed with rapid growth potential and, importantly, with IFN-γ secretion capacity, suggesting that they would rapidly generate a pool of CD4+ T cells with effector function upon antigenic stimulation in controller patients. The high functional avidity of Gag293 specific cells points to their capacity of initiating recall responses in the presence of minimal amounts of HIV antigen. The high functional avidity could be explained, at least in part, by a high avidity interaction between the TCR and the cognate Gag293 peptide/MHC complex. Thus, the high sensitivity of controller CD4+ T cells to antigen appeared intrinsic, rather than dependent on the antigen presentation context or the cytokine milieu. The Vβ repertoire of tetramer-positive Gag293-specific cells proved diverse, suggesting that multiple clones contributed to the high avidity CD4 response in HIV controllers. This property may favor the long-term persistence of a high avidity response, since the presence of multiple clones reduces the probability of viral escape or of immune senescence. Taken together, these findings suggest that CD4 recall responses to Gag293 are rapid and efficient in the group of controller patients. We propose that the rapid triggering of recall responses may contribute to viral control. A rapid CD4+ T cell response upon occurrence of “viral blips” will keep the immune system in alert, provide immediate help for CD8+ T cells to exert efficient cytotoxic function, and possibly provide direct antiviral effector function [37]. This rapid recall response may help keep HIV-1 replication under a low threshold, and avoid the progressive undermining of the immune system associated with repeated viral replication episodes [38].
The presence of high avidity CD4+ T cells helps explain how HIV controllers maintain an active T cell response in the face of very low viremia. We and others have previously shown that the level of the specific CD4+ T cell responses in HIV controllers exceeds that seen in efficiently treated patients, even though both groups have very low antigenemia [10],[11],[22],[24]. The triggering of recall responses at very low antigenic load in the controller group may account for this difference. Emerging evidence suggests that the CD8+ T cell response may also be of high avidity in HIV controllers. In particular, individuals harboring the protective HLA-B27 allele frequently develop a high avidity response against the immunodominant KK10 CD8 epitope, the avidity of the response correlating inversely with viral load [39]. The fact that HIV controllers maintain antiviral CD8+ T cells with high cytotoxic potential in spite of their low viral load is also suggestive of a high avidity response [6],[7],[40]. Thus, both the CD4+ and the CD8+ T cell compartments may contribute to the high sensitivity to antigen characterizing antiviral responses in the controller group. Studies in mouse models of chronic viral infections have shown that efficient CD8 responses do not persist in the long term without CD4 help [41]. Therefore, a high avidity CD4 response may be essential in maintaining the quality of the CD8 response in low viremia conditions.
A correlate of a heightened sensitivity to HIV antigens may be a chronic level of immune activation, due to the recall of cellular responses upon each viral replication episode, however limited. Indeed, we have previously reported on signs of ongoing immune activation in the effector memory CD4+ T cell compartment of HIV controllers, as measured by the expression of HLA-DR, the downregulation of the IL-7 receptor, and the secretion of MIP-1β [10]. Other signs of activation include raised levels of LPS in plasma [8] and increased expression of HLA-DR within the HIV-specific CD8+ T cells [6] as compared to efficiently treated patients. These observations confirm the notion that viral control is achieved through an active immunological process. One should note that excessive chronic activation may be deleterious in the long term, as suggested by a trend toward CD4+ T cell decrease in controller patients with the highest degree of immune activation, even in the persistence of undetectable viral load [9]. It will be important in future studies to determine if such individuals show a decrease in T cell functional avidity, which may lead to more prolonged induction of recall responses to achieve viral control, and consequently to prolonged episodes of immune activation. On the other end of the activation spectrum, a few controllers appear to have low HIV-specific CD8 responses and a generally quiescent immune system [42]. This phenotype may result from a particularly successful viral control, with an antigenemia so low that it would not activate the high avidity memory T cell population for long periods of time. It also remains possible that non-T cell based, alternate mechanisms of viral control predominate in these rare individuals.
It was intriguing that HIV controllers responded less frequently to the Gag161 peptide than efficiently treated patients, while the quality of the CD4+ T cell response appeared generally better in the former group. This observation pointed to possible changes in the immunodominance pattern associated with the controller status. Responses to Gag161 may have become subdominant in the controller group due to competition by high avidity CD4+ T cells responding to other epitopes, including that present in the Gag293 peptide. Indeed, high avidity has been shown to sharpen immunodominance in mouse models [43]. The key mechanism appears to be the increased proliferative capacity of high avidity T cells, which progressively fill the memory T cell niche, a phenomenon accounting for the apparent avidity maturation of T cell responses over time [36],[44]. Importantly, in the present study, the duration of HIV-1 infection in the group of efficiently treated patients did not differ significantly from that in the controller group, with median of 12 (7–20) vs. 15 (10–21) years, respectively. Thus, a longer infection time was unlikely to account for the presence of high avidity CD4+ T cells in the controller group.
The CFSE analysis identified a population of Gag293-specific cells with high proliferative capacity in HIV Controller cell lines, which was consistent with the presence of a pool of high avidity CD4+ T cells. The number of divisions undergone by Tet+ cells was heterogeneous, with only a fraction reaching 5 generations and above. This may reflect a range of avidities for the Gag293 antigen, with only a fraction of Tet+ cells being endowed with high avidity and thus high proliferative capacity. This notion is also supported by the shape of the functional avidity curves, which suggests the presence of both high and low avidity populations within the pool of Gag293-specific cells from HIV Controllers. However, the high avidity component was absent in the Gag293-specific CD4+ T cell population from treated patients, independent of the method of analysis (functional avidity, tetramer avidity, or proliferation of Tet+ cells). One should note that the growth ratio of CD4+ T cell lines depended on the intrinsic proliferative capacity of specific cells but also on the frequency of these specific cells at the initiation of culture. We have previously shown that the frequency of p24 Gag-specific cells measured ex vivo by intracellular cytokine staining was approximately 3 fold higher in HIV controllers than in efficiently treated patients [10]. The analysis of ex vivo ELISPOT responses to the Gag293 peptide also showed a trend for higher values in the controller group. Thus, it is likely that both an increased precursor frequency and a higher proliferative capacity contributed to the efficient growth of CD4+ T cell lines from HIV controllers. Both properties may also contribute to the long term persistence of CD4 responses in these patients.
The genetic background may play a role in conferring a better ability to mount high avidity CD4+ T cell responses. The increased frequency of HLA DRB1*0701 in the controller group could suggest a beneficial effect of this allele on the development of anti-HIV CD4 responses. However, since we did not detect an association between the presence of HLA DRB1*0701 and the level or avidity of the CD4 response within the controller group, we speculate that this allele may play a role in initially facilitating viral control, rather than in maintaining high avidity CD4+ T cells. Alternatively, HLA DRB1*0701 may be in linkage disequilibrium with a protective MHC class I allele associated with viral control. A beneficial effect of the HLA DRB1*13 alleles on CD4 responses has also been suggested [13],[20], though we did not detect a significant effect in our study. It will be important to confirm these findings in cohorts of patients powered for large scale genetic studies. An intrinsic advantage in CD4+ T cell growth capacity may also promote efficient CD4 responses in controllers. Van Grevenynghe et al. [26] have reported an increased growth capacity of controller CD4+ T cell lines in response to polyclonal stimulation, as compared to cell lines derived from efficiently treated patients or even from healthy donors. These authors demonstrated a role for the activation of the PI-3 kinase pathway, and the resultant inactivation of the downstream apoptosis inductor FOXO3a, in this particular growth phenotype. We did note a trend for higher growth ratios in controller CD4+ T cell cultures in the presence of CMV peptides, even in the absence of a positive IFN-γ ELISPOT response (not shown). However, an increase in growth propensity only partially accounts for the CD4 response characteristics observed in HIV controllers. High avidity CD4+ T cells were directed against HIV but not CMV, pointing towards a selective advantage in the induction of anti-HIV responses.
The selection of high avidity gag-specific CD4+ T cells may result from a lower exposure to HIV antigens during the acute infection stage, when the repertoire of responding T cells is initially shaped. The few reported cases of acute HIV-1 infection followed by spontaneous viral control support the notion of a lower viral peak in patients who acquire a controller status [45]. Mouse models indicate that low antigen exposure is associated to the development of a high avidity response, since only the high avidity T cells receive sufficient signals through the TCR to proliferate in the long term [44],[46]. Such a scenario may predominate in patients who spontaneously control HIV replication. The presence of high avidity T cells may in turn stabilize the controller status by limiting viral replication episodes.
On the other hand, we cannot rule out that high avidity Gag-specific CD4+ T cells are selected but subsequently lost in progressor patients. Since high avidity CD4+ T cells are the first to respond in the presence of low HIV antigen amounts, they may be the first to get activated in the presence of replicating HIV, and may represent the initial wave of target cells available to the virus. HIV is known to preferentially infect HIV-specific cells [47], and among those it may well preferentially infect the most readily activated population. A recent report suggests that responses to several CD4 epitopes can be detected during the acute infection stage but are subsequently lost in progressor patients, which supports the idea of a rapid culling of the CD4 repertoire [21]. Another reason for the loss of high avidity CD4+ T cells may be senescence due to overstimulation by high antigenic loads in progressor patients. The observation that high avidity CD8 responses can be lost after acute HIV infection supports such a model [48]. An important area of future research will be to elucidate mechanisms that protect high avidity CD4+ T cells from depletion in HIV controllers.
In conclusion, this study provides evidence for the presence of high avidity CD4+ T cells directed against Gag in HIV controllers. It is remarkable that the distinctive properties of HIV-specific T cells in Controllers, including high proliferative potential [14],[24],[49], polyfunctionality [12],[13] and high cytotoxic capacity per cell [6],[7], are all known attributes of high avidity T cells [36],[40],[50],[51]. Thus, high avidity may underlie many of the characteristics of an efficient adaptive immune response against HIV. The presence of high avidity T cells has been associated with control of chronic viral infections in mice [50],[52], monkeys [53], and humans [54]. Since high avidity also confers long-term memory and rapid reactivation in presence of antigen [36],[44],[55], it represents a desirable property to be induced by candidate T cell vaccines against HIV.
HIV controllers (HIC group; n = 17) were recruited through the French “Observatoire National des HIV Controllers” established by ANRS. HIV controllers were defined as HIV-1 infected patients who had been seropositive for >10 years, had received no antiretroviral treatment, and for whom >90% of plasma viral load measurements were <400 copies of HIV RNA/ml. All HIV controllers included in the present study had current viral loads <40 copies/ml. Control groups included: (1) HAART group (n = 20): HIV-1 infected patients successfully treated with antiretroviral therapy for more than 5 years and with a viral load <40 copies of HIV RNA/ml; (2) VIR group (n = 10): viremic patients with viral loads >10,000 copies HIV RNA/ml. Viremic patients had been infected with HIV-1 for more than 1 year and had not received antiretroviral therapy. Patients from the HAART and VIR groups were recruited through the SEROCO-HEMOCO cohort and the Bicêtre hospital.
The study was promoted by ANRS under number EP36 and approved by the Comité de Protection des Personnes IDF VII under number 05–22. All participants gave written informed consent prior to blood sampling.
PBMC from HIV infected patients were plated at 2×106 cells per well in 24-well plates in the presence of one HIV-1 Gag or pp65 CMV peptide (10 µM) in RPMI 1640 supplemented with 10% human AB serum, 2 mM L-glutamine, 10 mM HEPES, 100 ug/ml penicillin/streptomycin, 0.5 µM AZT, 5 nM Saquinavir and 5 ng/ml recombinant IL-7 (Cytheris). The peptides used to stimulate the culture were highly purified 20-mers (>99% purity; PolyPeptide Laboratories). Recombinant IL-2 was added after 2 days to a final concentration 100 U/ml. Cell lines were restimulated with IL-2 every 2 days until the end of culture. Starting from day 7, cells were counted every day by trypan blue exclusion to determine the growth ratio (GR: observed number of cells/number of input cells at day 0). The CD8+ T cell population represented a median of 5.9% (range: 0–22.7%) of the CD3+ population. CD8+ T cells were depleted with magnetic beads (IMag particles, BD Biosciences) at doubling time (GR = 2), before performing functional assays. Less than 1% CD8+ T cells remained after CD8 depletion (Fig. S2A).
IFN-γ secretion by CD4+ T cell lines was evaluated by ELISPOT assay as previously described [56]. Briefly, 96-well nitrocellulose plates were coated with 1 µg/ml anti-human IFN-γ capture monoclonal antibody (Mabtech). Cell lines starved off IL-2 for 16 h were plated in duplicate at 30,000 cells/well in coated ELISPOT plates and incubated with 4 µM peptide for 24 h at 37°C. Wells were then washed, incubated with a biotinylated anti-IFN-g detection antibody (Mabtech), followed with alkaline phosphatase-labeled extravidin (Sigma-Aldrich), and with a chromogenic alkaline phosphatase-conjugated substrate. IFN-γ spot-forming cells (SFC) were counted with a Bioreader 4000 system (Bio-Sys). The ELISPOT response was expressed as SFC/106 cells after subtracting background. Wells were counted as positive if the number of SFC was at least two times above background level. Functional avidity assays were carried out on all cell lines with ELISPOT responses >1000 SFC/106 PBMC. ELISPOT responses were measured in response to serial peptide dilutions from 4×10−6 to 10−11 M, and the last dilution that gave a number of SFC at least two times above background was determined.
At initiation of the study, 34 HIV controllers and 34 treated patients were genotyped for HLA-DRB1. Patients were included in the study if their genotype matched at least one of the 6 HLA-DRB1 alleles available for MHC class II tetramer studies. This panel allowed the analysis of ≥70% of Caucasian patients [34]. PE-labeled tetramers for the DRB1*0101, DRB1*0301, DRB1*1501 and DRB5*0101 alleles were obtained through the NIH Tetramer Facility at Emory University. HLA-DRB1*0401, DRB1*0701, and DRB1*1101 biotinylated monomers were produced in insect cell cutures as previously described [57]–[59]. Monomers were loaded with 0.2 mg/ml peptide by incubation at 37°C for 72 h in the presence of 2.5 mg/ml n-octyl-b-D-glucopyranoside and 1 mM Pefabloc SC (Sigma-Aldrich). Peptide-loaded monomers were tetramerized using APC- or PE-conjugated streptavidin (eBioscience). To each tetramer loaded with the Gag293 peptide corresponded a control tetramer loaded with the CLIP peptide.
The class II tetramer labeling protocol was adapted from [38]. CD4+ T cell lines were incubated with 4 µg/ml class II tetramer for 90 min at 4°C in PBS-1% BSA buffer. Surface marker antibodies CD4-PerCP, CD3-AF750-APC, CD14-FITC (eBioscience), CD8-FITC, CD19-FITC (BD Biosciences), and the Aqua Live/Dead viability dye (Invitrogen) were added for the last 20 min of labeling. The percentage of tetramer-positive (Tet+) cells was measured in the live, CD3+, CD4+, CD8−, CD14−, CD19-gate. Events were acquired on a CyAn flow cytometer (Beckman Coulter, Fullerton, CA) and analyzed using the Flowjo software (Tree Star). Negative controls were obtained by staining with HLA-DR matched tetramers loaded with the CLIP peptide. To determine the avidity of the TCR/pMHC interaction, CD4+ T cell lines were incubated with decreasing concentrations of class II tetramer from 1 to 0.01 µg/ml. The avidity was defined as the inverse of the last concentration that gave a percentage of Tet+ cells at least 2 fold higher than CLIP-tetramer control values.
The TCR Vβ repertoire of HIV-specific CD4+ T cell lines was determined by co-staining cells with an MHC class II tetramer and a panel of Vβ-specific antibodies (IOt-Test Beta Mark TCR Vβ repertoire kit, Beckman Coulter), according to the manufacturer's instructions. The kit covered approximately 70% of human Vβ specificities. The Vβ nomenclature is that of Wei et al. [60]. A Vβ specificity was considered amplified when the Vβ frequency was increased at least 4 fold in the Tet+ compared to the Tet- population. A ratio of 4 was above the range of Vβ variation observed in the ex vivo repertoire of HIV-infected patients [61] and within the range of Vβ expansions induced by superantigens in vitro [62].
Data are expressed as medians and range. Analyses were performed with the GraphPad Prism 5.0 software, using nonparametric statistical tests in all cases. Differences in variables between groups were analyzed with the Mann-Whitney U Test. Differences in percentages of response were analyzed with the Fisher's exact test. Correlations were analyzed with Spearman's coefficient R. All significant differences between groups (P<0.05) were reported on data plots.
|
10.1371/journal.pbio.1002491 | Paneth Cell-Rich Regions Separated by a Cluster of Lgr5+ Cells Initiate Crypt Fission in the Intestinal Stem Cell Niche | The crypts of the intestinal epithelium house the stem cells that ensure the continual renewal of the epithelial cells that line the intestinal tract. Crypt number increases by a process called crypt fission, the division of a single crypt into two daughter crypts. Fission drives normal tissue growth and maintenance. Correspondingly, it becomes less frequent in adulthood. Importantly, fission is reactivated to drive adenoma growth. The mechanisms governing fission are poorly understood. However, only by knowing how normal fission operates can cancer-associated changes be elucidated. We studied normal fission in tissue in three dimensions using high-resolution imaging and used intestinal organoids to identify underlying mechanisms. We discovered that both the number and relative position of Paneth cells and Lgr5+ cells are important for fission. Furthermore, the higher stiffness and increased adhesion of Paneth cells are involved in determining the site of fission. Formation of a cluster of Lgr5+ cells between at least two Paneth-cell-rich domains establishes the site for the upward invagination that initiates fission.
| The intestinal tract undergoes many changes during development, and after birth it has to significantly elongate and widen in order to increase the surface area for absorption. Crypt fission is a key process in intestinal tissue expansion and is also involved in adenoma growth. Despite the importance of crypt fission, the mechanisms controlling it are poorly understood. Understanding how crypt fission is regulated in normal tissue can help us to determine how the process changes in cancer. Here, we describe cellular behaviour during crypt fission. We identify a specific cellular arrangement in the intestinal stem cell niche that is associated with crypt fission and reveals insights into the mechanisms controlling crypt fission. There are two different cell types at the crypt base, Lgr5+ and Paneth cells, which play distinct roles in this process. We find that both their location and differences between them in proliferation, stiffness, and adhesion are important for fission. Based on our data, we propose a model in which stiffer and more adhesive Paneth cells are necessary to shape the crypt base and establish where fission occurs, whereas softer Lgr5+ cells allow shape changes and proliferation to expand newly formed crypts. Our model is an important step in understanding how crypt fission is initiated in normal tissue and provides a framework to understand how the process changes in tumorigenesis.
| The structures of many adult epithelia arise from branching events during development. For instance, the organisation of adult lung, kidney, and mammary epithelia arises by branching of epithelial tubes that ceases once the tissue is fully formed. A related but distinct form of branching is important in the gut, where the crypts of Lieberkühn divide in a fissioning process to elongate and widen the intestinal tract during postnatal development [1]. Crypt fission involves the division of a single crypt into two daughters (Fig 1). The incidence of crypt fission is highest in young animals and decreases with age but does not completely stop [2]. Importantly, crypt fission is reactivated in cancer and drives the clonal expansion of mutant crypts in adenoma [3–7]. For instance, polyps in ApcMin/+ mice and in familial adenomatous polyposis (FAP) patients are initiated by and expand through crypt fission [8–10]. Many reports describe the importance of crypt fission in growth of healthy and cancerous tissue; however, a detailed understanding of the underlying mechanisms is lacking.
The crypt base in the small intestine contains two major cell types: Lgr5+ cells, including stem cells; and secretory Paneth cells. Producing two crypts of normal size from one crypt requires an increase in the number of Paneth and stem cells between fission events. However, there is currently no consensus about the requirement of either of these cell types for the formation of new crypts. It has been proposed that crypt fission is driven by an expansion of the stem cell pool [11]. On the other hand, budding of new branches from intestinal organoids, a process related to fission, has been proposed to require Paneth cells [12–14]. However, the ability of intestinal tissue lacking Paneth cells to repair after injury questions the requirement of Paneth cells in this process [14,15].
To complicate matters further, recent reports have challenged the classical model of crypt fission as a bifurcation of a parental crypt, and instead propose that it occurs as “asymmetric budding,” with daughter crypts formed by budding from a larger parental crypt [16]. In intestinal organoids, new crypts can also form by budding from a spherical structure [12–14,16].
To understand the processes that govern normal fission, we utilised 3D imaging of whole mount tissue [17]. We examined crypts undergoing fission at high resolution and detected multiple types of fission during normal postnatal development. Monitoring Lgr5+ and Paneth cells, we found a cluster of Lgr5+ cells at the earliest stages of fission. This cluster marks the site of the bifurcation that initiates fission. Using whole tissue and organoids, we determined that Paneth cells have more β4 Integrin on their basal surface and are more adherent to laminin, a major component of the basement membrane. We also find that adhesion by β4 Integrin is important for normal fission. Computational models of tissue dynamics support roles of both Paneth and Lgr5+ cells in fission. We propose that Paneth cells promote deformation of the crypt base. A cluster of Lgr5+ cells forms between two Paneth-cell-rich areas, creating an area with lowered adhesion and mechanical stiffness that permits tissue buckling.
Crypt fission has been studied mostly using sectioned tissue or isolated crypts. Details of the earliest stages of fission are difficult to identify by either of these approaches. Sectioned tissue provides limited information about the 3D relationship of tissue structures. In isolated crypts stained only with DAPI [16], information about the position of daughter crypt bases relative to surrounding connective tissue is lost. Here, we use whole mount mouse tissue to identify mechanisms involved in crypt fission. High-resolution multiphoton confocal microscopy allowed visualisation of the 3D structure of crypts undergoing fission in tissue stained to visualise F-actin and nuclei (Fig 1). Previously unknown details about cell and tissue arrangements at all stages of fission were revealed.
Intestinal crypts have a straight lumen extending from the stem cell compartment at the base toward the opening facing the gut lumen (Fig 1A). Bifurcation of the crypt lumen marks fissioning crypts. To identify specific features of fission, we first documented different stages of fission based on the position of the bifurcation relative to the crypt base. The earliest stages are characterised by widening of the crypt base and the appearance of a branch point between two very short prospective daughter crypts (Fig 1B). As fission propagates, the bifurcation ascends upward from the crypt base (Fig 1C). Fission is complete when the bifurcation reaches the top of the crypt producing two new, separated daughter crypts.
The lumens of daughter crypts can be of equal or unequal length, distinguishing symmetric (Fig 1B and 1C) and asymmetric fissions (Fig 1D–1F), respectively. Asymmetric fission occurs when the initial bifurcation is not symmetrically placed (Fig 1D and 1E) or when fission is initiated above the crypt base (Fig 1F). Previously, fissions producing unequal daughter crypts have been referred to as “asymmetric budding” [16]. However, this classification relied on imaging isolated crypts released from the underlying connective tissue prior to fixation and imaging. Our in situ approach allows us to distinguish between asymmetric fission that results from a branch forming on the side of a crypt (i.e., away from the crypt base, which we define as budding) and asymmetric placement of the bifurcation at the crypt base, which we define as asymmetric fission. Using our method, we also observed examples of trifurcating crypts (Fig 1G) producing three daughter crypts. We found that such “triple fission” can be symmetric (Fig 1G), with three daughter crypts of equal length, or asymmetric (S1 Fig), with daughter crypts of different lengths.
Observation of different types of fission raises the question of their prevalence in normal development. We therefore counted and scored fission in mouse intestinal tissue. Three regions in each of the small intestine and the colon were examined from mice aged 2, 3, 4, 5, 6, and 10 wk. Asymmetric fission was defined visually as fission with daughter crypts differing by approximately 25% in length; both budding and asymmetric fissions were counted as asymmetric. One hundred to 200 crypts were scored in each region from three individual animals in each age group.
Consistent with previous reports [1,2,11,18–20], fission was most prevalent in rapidly growing young mice (Fig 2, S1 Table and S1 Data). The highest incidence of fission was at 2 wk of age, when approximately 40% of all crypts in region 2 (jejunum) and 60% of all crypts in region 5 (middle colon) were fissioning. These numbers are similar to those reported in isolated crypts [2], but are higher than reported in tissue sections [11]. Sectioned tissue permits identification of fission events only when they are orientated parallel to the sectioning plane. Since the orientation of fission is not uniform (S2 Fig), the incidence of fission is likely underestimated when using tissue sections.
Fission incidence declines with age, and by 10 wk the highest incidence was approximately 5% in regions 1 and 4. The rate of decline in fission varies between different regions but is most pronounced in the distal regions of both the small intestine and colon, where only approximately 1% of crypts undergo fission in 10-week-old animals (Fig 2, S1 Table, and S1 Data). Symmetric and asymmetric fissions occur with similar frequencies throughout development and in all regions. Triple fissions are rare, except in rapidly growing tissue of 2-week-old animals when 12.4% of all crypts in region 5 are undergoing triple fission. No triple fissions were detected in 10-week-old mice. Triple fission may only be a feature of rapidly growing tissue in young mice, explaining why it has not been identified in normal tissue previously.
Visualising fission events at all stages at high resolution facilitated observing the behaviour of cells during fission. The crypt base contains at least two cell types: secretory Paneth cells and Lgr5+ cells. The arrangement of these two cell types in early fission was determined in crypts from mice expressing Lgr5GFP [21] and stained with antibodies against Lysozyme to mark Paneth cells (Fig 3). In the base of single, non-fissioning crypts, Lgr5+ cells and Paneth cells are arranged in an alternating pattern (Fig 3A, 3E and 3G). At the earliest stage of fission, Paneth cells were absent from the middle of the crypt and were only found at either side of the position that marks the initiation of the bifurcation (Fig 3B and 3C). We examined 61 crypts at this stage and found that Paneth cells were completely absent from the initial bifurcation site in 54 cases, and only a single Paneth cell was detectable at this site in the remaining 7 cases. The same exclusion of Paneth cells from the bifurcation was observed in fissioning crypts in human tissue (S3 Fig). Correspondingly, we found that Lgr5+ cells formed a cluster in the middle of the crypt base, and this cluster marked the region where tissue appeared to buckle upward (Fig 3F and 3H). Such a cluster was observed in all 53 examples examined. These data reveal that clustering of Lgr5+ cells creates the site where bifurcation initiates fission and that Paneth cells are excluded from this region.
At later stages of fission, Paneth cells continued to reside exclusively at the base of the daughter crypts (Fig 3D). Lgr5GFP expression was lost from the cells forming the bifurcation once it had been displaced above the crypt base (Fig 3I). At this stage, both Lgr5+ and Paneth cells were restricted to the base of the two new daughter crypts. Based on these results, we divided fission into two distinct phases. The first phase, early fission, involves a change in the distribution pattern of cells in the stem cell niche that initiates a bifurcation and establishes the base of two distinct prospective daughter crypts. The second phase, late fission, involves an expansion of the bifurcation upward to complete formation of two new crypts.
The pattern of cell distribution we discovered in early fission prompted us to ask how this arrangement could promote the tissue buckling that drives fission; specifically, how differential properties of Paneth and Lgr5+ cells contribute. Buckling of the crypt wall in areas formed by the clustered Lgr5+ cells led to the hypothesis that these regions are more easily deformed than neighbouring Paneth-cell-containing areas. Paneth cells are more mechanically rigid than Lgr5+ cells [12]. However, in order to remain at the crypt base and resist the forces created by the buckling, they also have to adhere to their substrate more strongly. Attachment of epithelial cells to the underlying basement membrane is mediated by Integrins. The most common basal anchoring structures in the gut are dynamic focal adhesions formed by β1 Integrin [22], and hemi-desmosomes, which are more stable anchoring structures requiring β4 integrin [23]. We found that Paneth cells have an approximately 1.3-fold higher average signal intensity of β4 Integrin on their basal surface than crypt base columnar (CBC) cells (Fig 4A–4C, S2 Data). Together with the larger basal surface of Paneth cells, these data suggest that they have more β4 Integrin at their basal surface and therefore may attach more strongly to the basement membrane. To test this idea directly, we performed adhesion assays using isolated single cells from mouse small intestinal crypts. Specifically, we compared the adhesion of Paneth cells and other epithelial cells from intestinal crypts to laminin, the common ligand for β1 and β4 Integrin (Fig 4D and 4E and S3 Data). Cells were plated on laminin-coated substrates and allowed to adhere for 1 h. Weakly adherent cells were removed by shaking (Fig 4D). The increased proportion of Paneth cells present after shaking (Fig 4E) indicated that they were more adherent than other cells in intestinal crypts. Relatively stronger adhesion of Paneth cells may help to anchor the bases of prospective daughter crypts on either side of the Lgr5+ cell cluster and facilitate buckling. Paneth cells in intestinal polyps from ApcMin/+ mice do not have an increased signal intensity of β4 Integrin on their basal surfaces compared to neighbouring cells (S4 Fig, S2 Data). The aberrant fission in adenoma in this situation is consistent with a role for the differential adhesion of Paneth cells in normal fission.
If differences in adhesion to the basement membrane are a feature of normal fission, they are expected to also exist in the colon, where there are no Paneth cells. In colonic crypts, the stem cell niche contains secretory cells arranged in an alternating pattern with stem cells [24]. We found that, similar to Paneth cells in the small intestine, Muc2+ cells in the colonic stem cell niche had an increased signal intensity of β4 Integrin on their basal surfaces compared to neighbouring cells (S5 Fig). Therefore, crypt fission in the colon could also be supported by differences in adhesion of secretory cells in the stem cell niche to the basement membrane. Similarly, UEA-I+ Paneth/Goblet precursor cells have a higher β4 Integrin signal intensity (S6 Fig) on their basal surfaces than their neighbours. In mouse models lacking mature Paneth cells, such UEA-I+ secretory precursors may support fission.
Our tissue data suggested that differential adhesion between Paneth and Lgr5+ cells, together with a specific pattern of cell distribution, is important for fission. Testing this idea experimentally requires the ability to manipulate crypts. Intestinal organoids provide this opportunity. Organoids are epithelial structures derived from intestinal stem cells grown in matrigel that maintain many aspects of normal gut epithelial organisation, including alternating Paneth and stem cells at the crypt base [13]. Similar to the situation in tissue, crypt numbers increase with time as organoids grow.
The formation of new branches in organoids has been attributed to a budding-like process defined as formation of a new branch that initiates above the crypt base [16]. However, fission-like branching also occurs in organoids. To determine whether this branching process is comparable to fission in situ and to establish organoids as a suitable model to study crypt fission, we examined fixed organoids (Fig 5A–5C) and live organoids expressing LifeAct-GFP (Fig 5D). Fixed organoids contain crypts representing different stages of fission similar to those identified in tissue (Fig 5A–5C). However, in organoids, daughter crypts elongate at a 90°–180° angle (Fig 5C), rather than parallel as in tissue. The lack of physical constraint imposed by adjacent crypts and surrounding contractile cells is the most likely reason for this difference. Indeed, fissioning crypts removed from tissue before fixation exhibit a more splayed conformation (S7 Fig). The localisation of Lgr5+ and Paneth cells is similar to that identified in fissioning crypts in tissue. In early fission, Lgr5+ cells are located in the region forming the bifurcation, whilst Paneth cells are located on either side of the bifurcation (Fig 5B). In late fission, both Lgr5+ and Paneth cells are located primarily at the base of the two daughter crypts (Fig 5C). Furthermore, the average intensity of β4 Integrin on the basal surface of Paneth cells was approximately 1.3-fold higher than their neighbours (Fig 5E and 5F and S2 Data), almost identical to the difference measured in whole tissue (S8 Fig, S2 Data). Therefore, fission in organoids is similar to fission in tissue, and also may involve differential adhesion of Paneth cells and their neighbouring Lgr5+ cells. These similarities confirm organoids as a valid system to study the dynamics of fission. We therefore utilised organoids to probe mechanisms responsible for the initiation of fission. Specifically, we examined how the absolute numbers of Paneth and Lgr5+ cells per crypt, the position of Paneth cells within a crypt, and their increased adhesion to the basement membrane contribute to fission.
The documented expansion of the crypt base [25] prior to fission suggests that the number of Paneth and/or Lgr5+ cells in a crypt increases as a crypt ages. We found a strong correlation between the length of a crypt and the number of Paneth cells it contained (Fig 6D and S9 Fig, R2 = 0.67, S4 Data). Comparing single and fissioning crypts suggested that organoid crypts grow longer as they age until they reach approximately 100 μm and contain approximately 12 Paneth cells (S10 Fig, S4 Data) before they fission, similar to the number of Paneth cells reported in tissue [26].
Organoids permit manipulating the cellular composition of crypts so that the requirement for different cell types in fission can be examined. Lgr5+ cell numbers can be increased by supplementing standard growth media (ENR) with Chiron99032 (C) and Valproic acid (V) (ENR-CV), while Paneth cell number can be increased by ENR-CD (standard growth media supplemented with Chiron99032 and DAPT [D]). Both growth conditions caused altered crypt morphology: crypts in ENR-CV were longer, while crypts in ENR-CD were shorter and rounder (Fig 6B–6D, S9 Fig). In both conditions, the correlation between Paneth cell number and crypt length was lost (Fig 6D, S9 Fig, and S4 Data; ENR, R2 = 0.67; ENR-CV, R2 = 0.10; and ENR-CD, R2 = 0.33). The fissioning of crypts in organoids grown in ENR-CV was not altered measurably (Fig 6E, S5 Data), but crypts grew longer than 100 μm before fission (S10 Fig, S4 Data). On the other hand, fission incidence was reduced in organoids grown in ENR-CD (Fig 6D, S4 Data). These data indicate that onset of fission is not related to the absolute numbers of Paneth or Lgr5+ cells per crypt, but that a combination of their relative abundance and position is important.
New branches in organoids are reported to form at the location of Paneth cells; therefore, we examined whether mislocalising Paneth cells could affect crypt fission. Eph/Ephrin signalling is important for the normal positioning of Paneth cells [27], and inhibiting Eph/Ephrin signalling in mice causes mislocalisation of Paneth cells from the crypt base [28]. We found that including inhibitory Eph fragments in organoid culture media also caused mislocalisation of Paneth cells to positions above the crypt base (Fig 7A–7C, S6 Data). Fission events became more asymmetrical in the presence of the Eph fragment, indicating new branches formed further from the crypt base (Fig 7D, S7 Data). Time-lapse imaging revealed that the number of crypts increased at similar rates in Eph-treated and control organoids, indicating fission incidence was unaffected (Fig 7E–7G, S8 Data). Importantly, the Eph fragment had no effect on the increased β4 Integrin signal intensity on the basal surface of Paneth cells (S11 Fig, S2 Data). If the increased adhesion of Paneth cells is important for their role in positioning fission, we would expect inhibiting cell-substrate adhesion to reduce fission. Indeed, organoids grown in the presence of a β4 Integrin blocking antibody showed reduced incidence of fission compared to controls (ENR + Y27632; Fig 8, S9 Data). Together, these data suggest that Paneth cells are important in positioning daughter crypts and that the differential adhesion of Paneth cells and their neighbours is involved in normal fission. This supports the idea that differential adhesion of Paneth and Lgr5+ cells is important in positioning the buckling event that initiates fission.
Experimental data from organoids suggests a role for both Paneth and Lgr5+ cells during fission. However, our data that increasing the number of Paneth cells reduced fission contradicts our observation that new branches form where Paneth cells are located. To examine this in more detail, we utilised computational modelling of connected epithelial layers to investigate the role of Paneth cells in fission. Paneth cells were modelled to be 4.5 times stiffer than neighbouring cells [12], but they were otherwise identical to other cells in the layer. One consequence of these parameters was that Lgr5+ cells were smaller than Paneth cells (S12 Fig, S10 Data). Paneth cells were 1.4-fold or 1.14-fold larger than Lgr5+ cells for epithelial layers containing 40% or 80% Paneth cells, respectively. This effect is more marked when cells are in a constrained stable circular layer (as seen at 40% Paneth cells), as in a highly constrained system the cell stiffness has a greater influence on cell size and shape (S12A Fig). We observed more buds forming in epithelial cell layers containing a higher proportion of Paneth cells (Fig 9A). The mechanism of budding initiation was similar to that reported by Pin et al. [12], in which an Lgr5+ cell is pushed out by surrounding Paneth cells and proliferates, resulting in bud formation. Circularity of the epithelial layer was used as a measure of deformation. Epithelial layers with fewer Paneth cells had circularity values distributed around one (Fig 9B, S11 Data), indicating a circular shape. Increasing the proportion of Paneth cells led to decreased circularity with values tending toward zero (as demonstrated in Fig 9C), indicating that the epithelial layer had deformed. These data support a model in which Paneth cells are required to initiate outward buckling, mimicking a budding event. This appears to be contradictory to our observation that increasing Paneth cell numbers reduces fission (Fig 6). It is likely that this is due to the lack of Lgr5+ cell clustering under these conditions. Indeed, when we examined how the ratio of cell stiffness influenced fission, we found (for epithelial layers containing 80% Paneth cells) a similar progression from stability to fission as we increased the stiffness ratio from 3 to 4.5. Importantly, after budding is initiated, two Lgr5+ cells adjacent to the budding site are more likely to neighbour each other (Fig 9D, S12 Data). This is reminiscent of the Lgr5+ cluster that we found experimentally in early fission. Together with experimental data, these findings suggest that Paneth cells are required to deform the crypt base, while Lgr5+ cells may expand the region between Paneth cells to separate the new prospective crypt bases. In organoids, the region between Paneth cells is expanded by cell division [12]. Consistently, we observed mitotic cells in the bifurcation in fissioning crypts in tissue (S13 Fig), indicating that mitosis may play a role in expanding the Lgr5+ cell cluster in tissue.
Crypt fission is at the core of normal growth and maintenance of intestinal tissue. Importantly, the normal decrease in fission that occurs with ageing is reversed in early cancer, and aberrant fission in adult tissue is associated with adenoma growth. This study aims to determine how fission works normally so that mechanisms involved in fission in adenoma growth can be identified more readily. We found that the initiation of fission was marked by a cluster of Lgr5+ cells in the middle of the crypt base. Paneth cells appeared only lateral to this region and marked the bases of prospective daughter crypts. We discovered that differential adhesion of Paneth and neighbouring Lgr5+ cells to the basement membrane via β4 Integrin is likely involved in fission. This conclusion was supported by our finding that blocking β4 Integrin inhibits fission in intestinal organoids.
To identify how fission changes in cancer, we first had to establish what constitutes normal fission in situ. We found symmetric and asymmetric fission with similar frequency in healthy tissue, contrary to a recent report that identified “asymmetric budding” as the prevalent form of crypt fission [16]. These differences are easily accounted for by different definitions of asymmetric fission. We classified asymmetric fissions as those producing daughter crypts varying by ≥25% in length compared to >1% [16]. We also found that triple fission is a normal event and is not restricted to adenoma. However, it was only detectable during early postnatal development. It is possible that triple fission is a feature of rapidly growing tissue, when the space available due to lower crypt density allows triple fissions to occur more readily. Indeed, higher fission rates after irradiation have been reported to be a result of lower crypt densities [2]. Similarly, the high fission incidence in early postnatal development observed here may reflect lower crypt densities. We report that fission incidence reduces most rapidly in the distal regions of both the small intestine and colon, which may be explained by the different crypt morphologies [29] or different signalling pathway activities (e.g., Wnt, [30,31]) along the proximal-distal axis.
Fission incidence in different regions of the gut correlates with mitotic rate [32]. However, simply increasing mitotic rate does not stimulate crypt fission [33]. Instead, we identified changes in cell patterning, specifically the appearance of a cluster of Lgr5+ cells between at least two Paneth-cell-rich areas to establish regions of different mechanical properties at the crypt base.
One question that remains is how the Lgr5+ cluster forms. Our computer model provided some clues. First, it showed that the greater stiffness of Paneth cells deforms the crypt wall (Fig 9), consistent with our finding that Paneth cells determine where fission occurs (Fig 7 and [12]) and suggesting that Paneth cells may determine the shape of the crypt base. Second, the model showed that in the vicinity of the deformation caused by the Paneth cells, Lgr5+ cells are more likely to neighbour each other (Fig 9). The predicted increased probability of two Lgr5+ neighbours near deformations is consistent with our finding of Lgr5+ cell clusters at these sites. Together this suggests that the Lgr5+ cell cluster may be a consequence of crypt wall deformation initiated by Paneth cells. Alternatively, or in addition, the differences in Lgr5+ and Paneth cell behaviour may lead to Lgr5+ cell cluster formation. Lgr5+ cells divide [21], while Paneth cells are generated slowly [34] and reside at the crypt base for 18–23 days [35,36]. These data and our findings that Paneth cells are more adherent suggest that Paneth cells lack mobility. Together, this suggests that Lgr5+ cells produced at the extreme crypt base may be trapped by their larger, stiffer, and less mobile neighbours, the Paneth cells, causing Lgr5+ cells to cluster. This idea is further supported by the recent report that displacement of stem cells from the extreme crypt base is less likely than displacement from more lateral positions bordering the niche [37].
Once the Lgr5+ cluster has formed, it creates a softer area in the crypt base. This area deforms in response to forces exerted by neighbouring crypts and results in the upward buckling event that initiates the formation of new crypt walls. Expansion of these walls is likely achieved by proliferation of the Lgr5+ cells in the cluster, as suggested by the mitotic cells we observed in the bifurcation of fissioning crypts (S13 Fig). One consequence of Lgr5+ clustering is a loss of their direct contact with Paneth cells. This is predicted to cause Lgr5+ cells to become transit-amplifying cells [37], as indicated by the loss of Lgr5-GFP expression in bifurcations extending beyond the stem cell compartment (Fig 3). When there are too many Paneth cells, Lgr5+ cells may not be able to form a cluster, explaining the reduced incidence of fission in organoids grown in ENR-CD (Fig 6).
An important role for Paneth cells in normal fission is supported by many other reports: organoids lacking Paneth cells do not form branches [13,14,38]; high incidence of fission in tissue in young mice coincides with a rapid increase in the number of immature Paneth cells per crypt [39–41]; and Sox9flox/flox mice, which are depleted of Paneth cells, have wider and deeper crypts [42,43], which could reflect a failure to fission.
Apparently inconsistent with our conclusions that Paneth cells are important for crypt fission are studies suggesting they are not required for recovery after injury. In Paneth-cell-depleted mice (CR2-tox176, Math1flox/flox and Gfi1KI/KI [14,15,44–47]), intestinal tissue can recover after radiation-induced injury. However, 10Gy 137Cs irradiation [14,15] does not induce severe crypt loss [15,48,49], and recovery may not require crypt fission. Alternatively, other secretory cells may substitute for Paneth cells. Indeed, the UEA-I+ and Muc2+ cells still present in GfiKI/KI mice [44,46,47] also had an increased signal intensity of β4 Integrin on their basal surface compared to their neighbours (S6 Fig), a feature we found to be important for fission.
Increased adhesion of Paneth cells to the underlying basement membrane (Fig 4) is consistent with increased β4 Integrin on their basal surfaces [50–54]. Inhibiting β4 Integrin causes reduced fission incidence in organoids, consistent with a role of cell-substrate adhesion in fission (Fig 8). Although our model does not take substrate adhesion into account and only incorporates the increased stiffness of Paneth cells, it still produces results consistent with our experiments showing that Paneth cells drive the budding process required for crypt formation in organoids. The higher stiffness of Paneth cells may be sufficient to drive fission, while their increased adhesion may make the process more efficient in normal tissue by anchoring the crypt bases. Apc mutant organoids have a uniform distribution of β4 Integrin on their basal surface and fail to form branches (S4 Fig; [55]), suggesting that differential adhesion between the different cell types in the crypt is important for fission.
Understanding how the differential accumulation of β4 Integrin in secretory cells is achieved, and how it changes in polyps, requires identifying factors that regulate its expression. One potential factor is Wnt, a key regulator of intestinal crypt homeostasis [56]. For instance, the Wnt-responsive transcription factor c-Myc can regulate activity of the Itgb4 promoter [57–59].
Growth of many other tissues also involves branching structures that bear some similarities to fissioning crypts and also involve differential adhesion; for instance, milk duct formation in the breast. However, there are also distinct differences between these cases. Crypt fission in the intestine and colon initiates from a stationary crypt base that is relatively fixed in space by the surrounding tissue layer. Mammary ducts achieve branching by growth of two terminal end buds into extracellular space with a stationary, non-growing area between. Weaker substrate adhesion has been reported at the terminal end buds of breast tissue [60]. Therefore, similar to crypt fission, substrate adhesion is weaker in areas where tissue actively changes shape and moves. Another similarity is that β1 and β4 Integrin are important regulators of breast development and may play a role in breast tumour progression [61]: β4 overexpression correlates with more rapid breast cancer progression [62] and is required for ErbB2-driven tumorigenesis [63]. As a consequence, integrins have been identified as potential targets for treatment of breast cancer [61]. Any antagonists developed may provide useful tools for investigating the role of β4 Integrin in colorectal cancer and may lead to new avenues for treatment.
We report here mechanisms involved in intestinal crypt fission (summarised in Fig 10). Crypt fission involves a specific arrangement of cells within the intestinal stem cell niche. Lgr5+ cells cluster between Paneth cells, possibly as a result of mitotic daughters at the extreme crypt base becoming trapped between neighbouring Paneth cells. The Lgr5+ cell cluster creates a softer region that is more prone to changing shape in response to mechanical pressure from surrounding tissue. The Lgr5+ cell cluster expands upward from the middle of the crypt base, causing a bifurcation of the parental crypt as described by Wright [64]. The presence of mitotic cells in this region (S13 Fig) suggests that proliferation may play a role in the expanding bifurcation. The related process of crypt budding, which occurs above the crypt base, involves the proliferation of a cell positioned between two Paneth cells as reported by Pin et al. (Fig 10) [12]. The new bud grows through proliferation of cells positioned between the Paneth cells. In both fission and budding, proliferation is likely to support expansion of new crypts. While fission (either symmetric or asymmetric) is the prevalent form of crypt formation in tissue, new crypts form in organoids through both fission and budding-like branching. In both fission and budding, the position of Paneth cells dictates the site of fission. Their higher stiffness and adhesion helps to anchor and shape the new crypt bases.
Crypt fission is important for growth of the gut and plays a role in recovery from injury, but it is also responsible for expansion of mutant tissue in polyps and adenoma. Here, we describe how fission occurs normally. Our findings suggest that regulation of β4 Integrin, particularly at the crypt base, is involved. Mis-regulation of Integrins in colorectal cancer development may contribute to the formation of aberrant crypt fissions common in adenomas. Understanding regulation of β4 Integrin in normal and early cancerous tissue may reveal how crypt fission increases during adenoma formation. Further examination of the role of β4 Integrin and cellular adhesion in crypt fission requires the development of novel mouse models, allowing the conditional depletion of adhesion molecules in the intestine. Potential similarities between the role of Integrins in colorectal and breast cancer may make drugs that target Integrins useful for treatment of both of these common human cancers.
All experiments involving animals were performed in accordance with United Kingdom Home Office approved guidelines and were approved by the Home Office Licensing Committee (project license PPL60/4172).
The Tayside Tissuebank subcommittee of the Local Research Ethics Committee approved the collection of tissue samples. Normal human samples were obtained from surgical resections for hemi-colectomy.
All experiments involving animals were performed under the UK Home Office guidelines. CL57BL/6 wild-type, Lgr5-EGFP-IRES-creERT2 (Lgr5GFP/+), and ApcMin/+ were sacrificed by cervical dislocation or exposure to CO2, and the entire intestine (small and large) was removed immediately. Tissue was washed and divided into proximal, medial, and distal regions of the small and large intestines using percentage length from the gastro-duodenal and ileo-caecal junctions [8] corresponding to six regions. Regions for the small intestine were: region 1, 10% = duodenum; region 2, 50% = jejunum; region 3, 90% = ileum. Regions for the large intestine were: region 4, 10% = caecum; region 5, 50% = medial; region 6, 90% = distal. A 5 mm x 5 mm section from the middle of each of these six regions was immersed in cold fixative containing 4% paraformaldehyde in PBS (pH 7.4) overnight at 4°C before processing for staining [17].
Tissue samples were prepared for optical imaging of F-actin and nuclei as described previously [17]. For staining of Lgr5GFP+ cells and Paneth cells, an additional antibody incubation step and additional washing was added. Lgr5GFP+ cells were visualised using anti-GFP (Abcam, Cambridge, UK) and Paneth cells with Lysozyme (1:2,000, Dako, Cambridge, UK). Secondary antibodies (AlexaFluor-conjugated 1:250, Molecular Probes, Eugene, OR) were added along with fluorescently labelled Phalloidin (AlexaFluor-conjugated 0.26 μM [8 units], Molecular Probes) and Hoechst (50 μg/ml, Molecular Probes). Specimens were mounted in BABB after dehydration through an ethanol series [17].
Organoids were generated from mouse small intestinal crypts as described previously [13]. Briefly, small intestine was removed from the mouse and flushed. The intestine was opened longitudinally and villi removed by scraping the luminal surface with a coverslip. Tissue was washed with PBS, incubated in 3 mM EDTA (20 min), and crypts detached mechanically by vigorous shaking. Crypt suspension was washed twice in PBS then once in Advanced DMEM/F12. Crypts were dissociated to single cells with TrypLE Express (Life Technologies, Carlsbad, CA) at 37°C for 5 min; Advanced DMEM/F12 (Life Technologies) was added, then cells were filtered through a 40 μm cell strainer (Greiner, Frickenhausen, Germany). Single cells were then suspended in Growth Factor Reduced Phenol red free Matrigel (BD Biosciences, Oxford, UK). Organoids were grown in crypt media (Advanced DMEM/F12 supplemented with 10 mM HEPES, 2 mM Glutamax, 1 mM N-Acetylcysteine, N2 [Gemini, Sacramento, CA], B27 [Life Technologies], Pen/Strep [Sigma-Aldrich, St. Louis, MO]) containing growth factors (EGF [50 ng/ml; Invitrogen], Noggin [100 ng/ml; eBioscience], and R-Spondin conditioned media [1:4]). Additional growth factors Chiron99021 (3 μm; Invitrogen, Waltham, MA), Valproic acid (1 mM; Invitrogen), and Y27632 (10 μm; Cambridge Bioscience, Cambridge, UK) were added for the first 48 h culture. Organoids were passaged by first physically breaking up Matrigel, then washed in Advanced DMEM/F12 and dissociated into individual crypts by pipetting. Individual crypts were re-suspended in Matrigel and grown in crypt media containing growth factors.
Additional reagents were added along with the crypt media. Chiron99021, Valproic acid, and Y27632 were added at the same concentrations as described above. 1 μg/ml recombinant mouse EphB2 Fc chimera (R&D Systems, Minneapolis, MN) was added at concentrations corresponding to serum values previously reported [28]. Anti-β4 Integrin (Abcam) was added at 5 μg/ml. In the presence of a β4 Integrin antibody, organoids become moribund and disintegrate after 2 d. However, including the Rho kinase inhibitor Y27632 to inhibit anoikis [13] prevents this demise, confirming that adhesion to a basement membrane is compromised by the β4 Integrin antibody.
Organoids were grown in Matrigel in 8-chamber μ slides (Ibidi, Munich, Germany) for 3–5 d and fixed in warmed 4% paraformaldehyde in PBS (pH 7.4) for 20 min (37°C), permeabilised for 1 h in 1% Triton-X100 (this and all subsequent steps were carried out at RT), and blocked for 1 h (1% BSA, 3% normal goat serum, 0.2% Triton-X100 in PBS). Organoids were incubated in antibodies overnight in Working Buffer (0.1% BSA, 0.3% normal goat serum, 0.2% Triton-X100 in PBS): β4 Integrin (1:100 Abcam), Lysozyme (1:2,000), washed 5x in Working Buffer, then incubated in secondary antibodies (1:250, AlexaFluor-conjugated [Molecular Probes]), along with 5μg/ml Hoechst 33342 (as above) and Phalloidin (as above) overnight in Working Buffer, washed 5x in Working Buffer, and mounted in ProLong Gold antifade (Molecular Probes).
Small intestine and colon were isolated from mice and flushed with PBS followed by 4% paraformaldehyde (pH 7.4). Tissue was opened longitudinally, washed briefly in PBS, and incubated in 4% paraformaldehyde (pH 7.4) at 4°C overnight. A small, square piece of tissue was excised with a scalpel, embedded in 3% Low Melt Temperature Agarose, and sliced into 200 μm thick sections with a Vibratome (Leica, Wetzlar, Germany). Tissue sections were permeabilised for 2 h in 2% Triton-X100 (this and all subsequent steps at 4°C), blocked for 2 h (1% BSA, 3% normal goat serum, 0.2% Triton-X100 in PBS), and incubated in primary antibodies for 3 d (Lysozyme and β4 Integrin as described above, Muc2 [1:200, Santa Cruz Biotech, Santa Cruz, CA]), washed in Working Buffer (5x in 1 h), incubated in secondary antibodies along with Phalloidin and Hoechst (as described above) and Rhodamine-labelled UEA-I (1:200, Vector Labs, Peterborough, UK), washed in Working Buffer (5x in 1 h), and mounted in ProLong Gold antifade (Molecular Probes). Sections were mounted on coverslips between 2x 120 μm spacers to preserve tissue structure.
Tissue and organoids were imaged with a Zeiss LSM 710 microscope (Carl Zeiss AG, Oberkochen, Germany) using 25x or 40x Zeiss objective lenses and immersion oil with a refractive index of 1.516. Multiphoton excitation was provided by a Coherent Chameleon Titanium Sapphire laser at 820 nm to simultaneously excite Hoechst and Rhodamine-Phalloidin for whole mount tissue imaging. Organoids and sectioned tissue were imaged in conventional confocal mode using 40x LD Plan-Neofluar objective lens, and Z stacks were taken at 1 μm steps.
Image processing and analysis were performed in Volocity (PerkinElmer, Waltham, MA) or Imaris (Bitplane, Windsor, CT). Fission in tissue was measured by importing image stacks of 3D whole mount tissue into Volocity. Using the images, all crypts were marked and those undergoing fission identified to report the proportion of all crypts undergoing fission. Symmetry and fission stage was determined visually: using a cut-off of approximately 25% divergence in daughter crypt lengths and sizes to mark an asymmetric fission. The structure of fission was determined in Imaris: surfaces were drawn around the crypt lumen and outer wall based on Phalloidin staining. Using Imaris, crypt length was determined by measuring the position of the crypt base and the crypt opening, while symmetry was assessed by dividing the length of the shorter daughter crypt by the length of the longer daughter crypt. Each Paneth cell was marked and its distance from the crypt base measured to analyse number and distribution of Paneth cells. β4 Integrin levels on basal cell membranes were measured using Imaris; surfaces were drawn around the basal membrane of Paneth cells and their neighbouring cells. Average signal intensity in these surfaces was measured for each individual cell. The intensity for β4 Integrin in Paneth cells was divided by the corresponding value for its neighbours.
We extend the crypt model outlined in Dunn et al. [65] to model the cross-section of a confluent epithelial layer. We represent cells by the positions of their centres and define spatial connectivity through a Delaunay triangulation. Cells interact with each other through forces applied along the edges of the triangulations. Attraction and repulsion between cells is modelled through a linear spring force law. Furthermore, epithelial cells are subjected to a basement membrane force, where the basement membrane is defined along the boundary separating epithelial cells and the Matrigel. Detailed mathematical descriptions of these forces and simulation details are provided in Dunn et al. [65,66]. The code used to run these simulations can be found at https://chaste.cs.ox.ac.uk/trac/wiki/PaperTutorials/CryptFissionPlos2016.
We consider a 20-by-20 square array of cells, consisting of a layer of proliferative epithelial cells surrounded by terminally differentiated non-epithelial cells (representing the surrounding stromal cells or Matrigel). Each simulation begins with a circular epithelial layer mimicking the base of the crypt (i.e., two bases joined together, or the initial geometry of spherical organoids [13]). The region surrounded by the epithelial layer represents the crypt lumen. Outside the square, additional layers of non-epithelial cells are fixed around the edges, ensuring the layer may deform without compromising the overall shape of the box. Any cell that moves beyond any of the four walls of the square also becomes fixed.
The epithelial layer consists of both Lgr5+ and Paneth cells. Epithelial cells proliferate according to a stochastic cell cycle model, in which the cell cycle duration is sampled from a Uniform Distribution U(12,14) [67]. Each cell can divide symmetrically or asymmetrically according to a pre-defined division parameter controlling the proportion of stem cells in the model, called the target proportion. We model Paneth cells as stiffer than Lgr5+ cells, by assigning the stiffness of springs connecting one or more Paneth cells to be 4.5 times stiffer [12] than springs joining only Lgr5+ cells or non-epithelial cells (whose stiffness is as defined in [64]). Apart from stiffness, Paneth cells are physically identical to stem cells in simulations. Note that we are considering the cross-section of a three-dimensional epithelial layer and modelled all epithelial cells to have the same size. Normally, Paneth cells are 2–4-fold larger than Lgr5+ cells. As a consequence, the model requires a larger number of Paneth cells than normally present in tissue. To prevent over-crowding in the layer, cell death in the form of anoikis is introduced. If an epithelial cell loses contact with the surrounding non-epithelial cells and is pushed into the lumen, it is removed from the simulation.
For each target proportion value, 100 simulations were run for 100 h each. This length of time was in accordance with the length of time taken for organoids to undergo budding [13]. To determine incidence of crypt fission, we track the circularity of the epithelial layer, defined by the formula C=4π∙Area(Perimeter)2. More circular forms have circularity values closer to one, while more complex shapes will have circularity values near zero. An epithelial layer with low circularity indicates crypt fission has occurred. As a buckled epithelial layer cannot revert back to a circular shape, we need only to record the circularity at t = 100 h.
Live imaging of organoids was performed using confocal microscopy, as described above, for LifeAct-GFP organoids (kind gift from Prof. Laura Machesky, Beatson Insitute, Glasgow, UK). Brightfield movies were recorded with a Leica DMIRB (Leica) inverted microscope and analysed in Fiji [68]. The number of branches per organoid was counted every 6 h using the Cell Counter plugin. Numbers of branches shown on graphs correspond to the average number of branches per organoid.
Single cells were isolated from mouse small intestinal crypts, as described above for organoid culture. Cells were plated in Advanced DMEM/F12 on laminin-coated (Sigma-Aldrich) surfaces and allowed to adhere for 1 h. Cells were shaken at 2,000 rpm for 15 s to remove weakly adherent cells. Cells were fixed in 4% paraformaldehyde, and stained with Hoechst and Lysozyme (as described above). The total number of cells and Paneth cells were counted and compared to non-shaken controls.
All statistical analyses were performed using GraphPad Prism 6.0a (GraphPad, La Jolla, CA) for Windows. Tests performed are described in individual Figure legends, along with p-values and significance (ns = not significant, * = p < 0.05, ** = p < 0.01, *** = p < 0.001, **** = p < 0.0001).
|
10.1371/journal.pcbi.1001040 | Phase-Locked Signals Elucidate Circuit Architecture of an Oscillatory Pathway | This paper introduces the concept of phase-locking analysis of oscillatory cellular signaling systems to elucidate biochemical circuit architecture. Phase-locking is a physical phenomenon that refers to a response mode in which system output is synchronized to a periodic stimulus; in some instances, the number of responses can be fewer than the number of inputs, indicative of skipped beats. While the observation of phase-locking alone is largely independent of detailed mechanism, we find that the properties of phase-locking are useful for discriminating circuit architectures because they reflect not only the activation but also the recovery characteristics of biochemical circuits. Here, this principle is demonstrated for analysis of a G-protein coupled receptor system, the M3 muscarinic receptor-calcium signaling pathway, using microfluidic-mediated periodic chemical stimulation of the M3 receptor with carbachol and real-time imaging of resulting calcium transients. Using this approach we uncovered the potential importance of basal IP3 production, a finding that has important implications on calcium response fidelity to periodic stimulation. Based upon our analysis, we also negated the notion that the Gq-PLC interaction is switch-like, which has a strong influence upon how extracellular signals are filtered and interpreted downstream. Phase-locking analysis is a new and useful tool for model revision and mechanism elucidation; the method complements conventional genetic and chemical tools for analysis of cellular signaling circuitry and should be broadly applicable to other oscillatory pathways.
| Key to robust discernment of cell circuit architecture is to have as many distinct response features as possible for comparison and evaluation. One under-appreciated characteristic of oscillatory circuits is that under periodic stimulation, these systems will exhibit responses synchronized to this stimulatory input, a phenomenon termed phase-locking. We demonstrate that phase-locked response characteristics vary noticeably depending on circuit activation and recovery properties; these response characteristics thereby provide a unique set of criteria for oscillatory circuit architecture analysis. The concept is validated through experiments on an oscillatory calcium pathway in mammalian cells; the experimental setup allowed us to explore, for the first time, the properties of chemically induced phase-locking of intracellular signals. Observations of this phenomenon were then used to test the predictions of several existing mathematical models of calcium signaling. Most of the models we evaluated were unable to match all our experimental observations, suggesting that current models are missing mechanistic elements in the context of calcium signaling for the cell type and receptor/stimulant tested. The observations of phase-locking further led us to identify one simple mechanistic modification that would account for all the experimental observations. The techniques and methodology presented should be broadly applicable to a variety of biological oscillators.
| Determining the circuit architecture of cellular signaling pathways is challenging. Analysis using perturbative tools including siRNA [1], [2], protein over-expression [3], chemical inhibitors [4], or caged compounds [5] usually reveal multiple plausible models that require further refinements and clarification, not just one definitive conclusion. Thus, there is always a need for additional tests and readouts that shed light on signaling circuit architecture in a robustly discriminating manner.
Most perturbations applied to biochemical circuit analysis are genetic or chemical in nature and alters the circuit architecture itself. Furthermore, the analysis usually looks at how such perturbations change signaling in response to a single step change with no further time variation in stimulation parameters. While these types of analyses are very useful, the circuit-destructive and temporally non-varying nature limits information that can be obtained concerning dynamic properties of the intact signaling system [6]. We hypothesized that analysis of the frequency-dependent response characteristics of the intact biological oscillator circuit to periodic extracellular chemical stimulation would reveal critical activation and recovery properties of biological oscillators to enable elucidation of molecular mechanisms. Here we demonstrate and validate this concept for the oscillatory calcium pathway of the G-protein coupled receptor (GPCR) M3 muscarinic system.
The biochemical recovery properties of this system were evaluated by reducing the rest period between pulses of the M3 ligand, carbachol (CCh), and observing the resulting calcium responses. We noted the emergence of beat skipping upon periodic stimulation. The phenomenon whereby an oscillatory system becomes synchronized to a periodic stimulation input is referred to as phase-locking. As the rest period between stimulation pulses was decreased, the number of system responses of the signaling pathway of interest became less than the number of stimulatory inputs thereby revealing biochemical pathway recovery properties not attainable by continuous stimulation. Furthermore, the skipped beats often were not completely absent, but instead appeared as small calcium transients that we here termed “sub-threshold” spikes; these have been observed previously in electrical responses of cellular systems [7]. The sub-threshold spikes provided insight into the activation properties of the signaling system. The complete absence of a sub-threshold spike would suggest that a switch-like mechanism produced calcium spikes; their presence, however, would suggest that a graded mechanism was more plausible.
These experimental observations of phase-locking properties were compared to the activation and recovery properties of nine models of oscillatory calcium signaling; while these models exclusively deal with the temporal dynamics of calcium signaling and we note that more elaborate models that also include spatial dynamics and IP3 receptor noise are available [8], [9]. In the main text we focus upon two highly different models: the Chay et al. model [10], and the positive feedback Politi et al. model [11]. The former model is the first that theoretically analyzed calcium dynamics in chemically-induced phase-locking; the latter model was recently published, features experimental work to support its proposed mechanisms, and carries dynamic features from previous models and experiments [12], [13], [14]. In addition, both models are able to account for a wide range of calcium oscillation periods (10s of seconds to minutes) upon continuous stimulation. The activation properties of the Chay et al. model are characterized by switch-like activation of phospholipase C (PLC) by G-protein, and it also features basal inositol triphosphate (IP3) production, which represents a recovery mechanism that ensures that IP3 quickly returns to its pre-stimulus levels. The Politi et al. model does not have such a recovery mechanism, and features graded PLC activation. To produce oscillations in the Chay et al. model, the products of the switch-like activation of PLC (IP3 and diacylglycerol) negatively feedback on upstream pathway components (G-proteins). In the Politi et al. model, IP3, produced by graded activation of PLC, feeds back on downstream elements (IP3 receptor) and calcium feeds back upon upstream elements (PLC) to create oscillations. A large number of oscillatory calcium models feature the aforementioned feedback mechanisms [15], [16], [17], [18], [19].
Under continuous stimulation, both models exhibit calcium oscillations with increasing frequencies upon increasing stimulation concentration, as seen in a host of experimental data [20], [21], [22]. Both models were thus appropriate but indistinguishable by conventional stimulation methods. The discriminating features provided by phase-locking analysis, however, revealed that neither of the calcium models correctly predicted all the experimental behaviors based upon their activation and recovery dynamics. Furthermore, by analyzing the sources of discrepancy between the predictions and experiments, we were able to propose a mechanism and parameter modification to account for all the experimental observations of phase-locking.
Although phase-locking can be thought of as a general property of biological oscillators [23], it has not been previously explored experimentally in the context of chemical stimulations. While recent reports have claimed that phase-locking events are largely independent of detailed mechanism [24], we show that the properties of phase-locking can be employed for elucidation of some of the activation and recovery properties of an oscillatory calcium system. We demonstrate that phase-locking, which can only be observed using temporally patterned stimulation, complements conventional chemical and genetic tools for elucidating non-linear oscillatory pathways.
We assessed cellular responses to square-wave stimulation through use of a microfluidic platform [modification of 25], which enabled exploration of phase-locked rhythms induced by chemical input signals (Fig. 1a–c). With fixed stimulant concentration (C) and stimulation duration (D), increases in the rest period (R) resulted in increases in the phase-locking ratio (Fig. 1d); phase-locking ratios were calculated by dividing the number of system responses by the number of chemical inputs (See Fig. 1,2 in Text S1). Analysis of the phase-locking rhythms also uncovered the existence of sub-threshold calcium spikes in individual cellular calcium responses (Fig. 1b). In addition, we explored the phase-locking trends induced by varying C and D (See Fig. 3a, b in Text S1). These observations collectively provided robust discrimination markers for rigorous evaluation of mathematical models of oscillatory calcium signaling in order to elucidate molecular mechanisms.
Nine oscillatory calcium models were chosen as a test set against our experimental results, based upon the inability to discriminate their behaviors using continuous stimulation despite significant differences in their activation and recovery mechanisms. Here we show phase-locking analysis of two of these models: the Chay et al. model [10] and the Politi et al. model [11] (Fig. 2). Under continuous stimulation, both the Chay et al. and Politi et al. models exhibited oscillatory calcium responses in physiologically relevant frequency ranges (Fig. 3a–c); furthermore, both depicted the same behavior as the strength of stimulation was increased, as depicted in Fig. 3a and b. We demonstrate that phase-locking analysis is able to effectively dissect the differences in recovery and activation properties between the models (Fig. 3d–i).
We first analyzed the Chay et al. model [10] (Fig. 2a). As depicted in Fig. 3d, we found that as the rest period (R) between stimulation events was increased, the phase-locking ratio increased. Despite the agreement of the model with the effects of R on phase-locking ratio observed in our system (compare Fig. 1d with Fig. 3d), it could not account for the presence of sub-threshold calcium spikes (compare Fig. 1b with Fig. 3g), suggesting inaccuracies in its activation properties. We attributed the lack of sub-threshold spikes to the model mechanisms, and not model parameter values, as we used a sampling algorithm (Latin Hypercube Sampling (LHS)) to survey a range of parameter values and found no parameter set able to result in sub-threshold calcium spikes (Fig. 4). The Chay et al. model assumes that G-protein activation of PLC is a switch-like response with a Hill Coefficient of 4. Therefore if activated G-protein levels are not sufficiently high to surpass the threshold for PLC activation, a calcium spike will not result. However, the presence of sub-threshold calcium spikes in our experiments suggested that such a sharp activation threshold does not exist. While some experiments suggest that Gq-protein activation of PLC is graded [26], to our knowledge, there are no studies that have conclusively determined the nature of this interaction; furthermore, these activation properties may be cell type or signaling pathway dependent. When the Hill coefficient of the G-protein/PLC interaction was reduced below 3.5 in the Chay et al. model, calcium oscillations could not be obtained under continuous stimulation (See Fig. 4a in Text S1); furthermore, periodic stimulation of the model with Hill coefficients between 3.5 and 4 did not yield sub-threshold calcium spikes for a wide range of stimulation conditions (See Fig. 4b in Text S1). These results have important implications in terms of how extracellular chemical signals are filtered and interpreted by downstream elements. In particular, intracellular calcium is not only frequency encoded [27], but also amplitude encoded [28], which means that sub-threshold calcium responses might affect cellular responses compared to the non-responses that were noted in the Chay et al. model. Therefore, from a mechanistic standpoint, the ability to capture behaviors such as sub-threshold spikes may prove critical. In addition, these findings show that the reaction mechanisms and model parameters need to be re-evaluated for the Chay et al. model, which has been used for analysis in many other studies [5], [29], [30], [31].
Our experimental observations were then used to evaluate the Politi et al. model (Fig. 2b). Individual calcium graphs portrayed sub-threshold calcium spikes upon exposure to square-wave stimulation pulses (Fig. 3h). However, the model incorrectly predicted that larger R resulted in smaller phase-locking ratios (Fig. 3e), suggesting that the recovery properties of the model are not accurate. LHS analysis indicated that the choice of model parameter values alone could not explain these inaccuracies, suggesting that reaction mechanisms used to formulate the model needed revision.
Thus, neither of the calcium models tested was able to account for all of our experimental observations. We noted that the Politi et al. model showed continued IP3 decay between stimulation pulses, while in the Chay et al. model, IP3 levels exhibited recovery between stimulation pulses (See Fig. 5 in Text S1). In the latter model, IP3 recovery between stimulation pulses is due to a mechanism for basal IP3 production. Addition of basal IP3 production to the Politi et al. model was able to correct its deficiencies in recovery dynamics (Fig. 3 right column); the IP3 production value used in our study was similar to that of reference [32]. This model revision may provide crucial insight into physiological systems where cells or tissues require fidelity of its calcium signals to periodic chemical stimulation in order to carry out their function [33]. Accurate capture of the recovery properties of oscillatory pathways may also play a pivotal role in the entrainment of such systems [34]. We note that other mechanisms may be found that can account for our experimental observations, but basal IP3 production provides the simplest explanation and is supported by the literature [35], [36], [37]. Collectively, this would suggest that the activation and recovery mechanisms reflected in our revised Politi et al. model (positive feedback mechanism of calcium upon PLC activity, graded PLC activation by G-proteins, and basal IP3 production) are a good fit for the pathway studied here.
We also analyzed seven additional calcium oscillation models. We first explicitly included ligand-receptor-G protein dynamics in both the Chay et al. and Politi et al. models analyzed above, to test whether this would affect our predictions. Those modifications did not change the outcomes of the phase-locking analysis (phase-locking ratio vs. C, D, and R and presence of sub-threshold spikes) (See Fig. 6 in Text S1), suggesting that the discrepancy between model and experiment did not lie in the simplified way stimulation was represented in the original models. We also tested a precursor to the Chay et al. model, a model by Cuthbertson and Chay [38]. Like the Politi et al. model described above, it did not contain a basal level of protein activity, and it too yielded a descending staircase as rest period (R) was increased (See Fig. 7 in Text S1). We next tested the model developed by Atri et al. [16], and found that it produced the correct recovery behavior as well as sub-threshold spikes (See Fig. 8 in Text S1); these results can be attributed to a basal flux term and graded activation, respectively. However, the calcium oscillation dynamics of the Atri et al. model are significantly faster than the range of oscillation periods we observed experimentally. As a result, we then analyzed a version of the Li and Rinzel model [18] that features slower dynamics, as presented in the study by Sneyd et al. [5]. While the model did exhibit calcium oscillation periods closer to what we saw experimentally, it exhibited a decrease in phase-locking ratio as both C and D were increased (See Fig. 9 in Text S1). This behavior was perhaps due to an augmented inhibitory effect of calcium upon the activation of the IP3 receptor; in addition, the model did exhibit sub-threshold spikes and showed the correct recovery properties, which could be attributed to a basal flux term and graded activation, respectively. Finally, we performed phase-locking analysis on the oscillatory calcium models developed by Dupont et al. [17] and Kummer et al. [39]. The former model features feedback of calcium upon PLC activity and IP3 metabolism, similar to the Politi et al. model, and the latter model features G-protein and PLC dyanmics. While the Dupont et al. model did exhibit sub-threshold spikes, phase-locking analysis revealed that it exhibited a decrease in phase-locking ratio for increases in R (See Fig. 10 in Text S1); the Kummer et al. model exhibited sub-threshold spikes as well, but also did not show a change in phase-locking ratio with changes in C (See Fig. 11 in Text S1). Thus, although we have not performed an exhaustive search, the modified Politi et al. model developed here best describes the qualitative features of our data on the M3 pathway.
In sum, we employed a combination of microfluidics, real-time imaging, and mathematical modeling in order to probe the circuit architecture of an oscillatory signaling pathway in mammalian cells. Here chemical-induced phase-locking was explored and analysis of its properties was used to test mathematical models and elucidate molecular mechanisms. Previous reports have claimed that phase-locking events are mostly robust to mechanism details [24], [40]; this study reports that the properties of phase-locking, however, largely depend upon some of the recovery and activation properties of the molecular mechanisms of an oscillatory signaling system.
As microfluidic setups become more elaborate in their ability to generate temporal stimulation patterns, we can expect even more discriminating markers for signaling studies [41]; the diverse waveform stimulation patterns generated by microfluidic setups such as the “chemical waveform synthesizer” [42] and the “chemical signal generator” [43] should prove useful to this end. While a single optical readout (calcium) was employed for this study, the experimental setup is amenable to the use of multiple real-time readouts of cellular signaling, thereby further enhancing the number of discriminating markers for elucidation of signaling pathways. Finally, although this paper focused on calcium oscillations, we believe our approach would be well-suited for studies on various biological oscillators such as ERK [44], NFκB [45], and components involved in cell cycle [24], circadian [46], and ultradian [47] rhythms. For example, we have performed phase-locking analysis of two popular circadian oscillator models [48], [49] and seen dramatic differences in phase locking behavior between the two, despite similar behaviors under conventional stimulation conditions (See Fig. 12 in Text S1). Thus, these types of phase-locking analyses provide experimentally testable hypotheses for elucidating molecular mechanisms and show that the method is applicable to a broad range of oscillatory pathways.
HEK293 cells were cultured in Dulbecco's Modified Eagle's Medium (DMEM) (Invitrogen) supplemented with 10% Fetal Bovine Serum (FBS) (Gibco) and were maintained at 37°C with 5% CO2 in 24-well plates. 0.25% Trypsin/EDTA (Gibco) was used to detach cells from plates and transfer them to the microfluidic setup. These cells were stably transfected with the M3 muscarinic receptor (selected with 0.4 mg/mL Geneticin (Gibco)). Cells were transiently transfected with the calcium FRET probe YC3.60 [50]. Transfections were carried out with Lipofectamine2000 (Invitrogen) using the manufacturer's protocol.
Microfluidic device molds were fabricated based upon the ones described in Futai et al. [25]. Front-side photolithography [51] was used to construct the outlet channel where cells were cultured; the remaining channels (inlets and “Braille” channels) were constructed with backside photolithography [52]. With the resulting glass mold, PDMS (1∶10 ratio of curing agent to base) was cast upon the positive relief features and allowed to cure for at least 2 hours in a 60°C oven. The resulting device was then irreversibly sealed against a thin (∼100 µm) PDMS sheet through 30 s plasma oxidation. Once sealed, the device was filled with Phosphate Buffered Saline (PBS) and sterilized for 2 hrs in a UV oven. To ensure cell adhesion, the chip was subsequently filled with 100 µg/mL laminin (Invitrogen) and allowed to incubate at 37°C for two hours. After this, the chip was flushed and refilled with DMEM supplemented with 10% FBS. Transfected HEK293 cells were then seeded from the outlet port and were appropriately positioned in the outlet hydrodynamically. The cells were then allowed to attach overnight.
A custom program written in Visual Basic was used to control the dynamic pumping mediated by Braille-actuation [53], and thereby create the various temporal stimulation patterns used in experiments (Fig. 1a); experiments with fluorescein solution confirmed the nearly square-wave shape and reproducibility of these patterns. Carbachol (CCh) dissolved in imaging media [54] was added to one of the inlet reservoirs, and the other reservoir was filled with stimulant-free imaging media. Cells in the devices were maintained at 37°C via a transparent indium tin oxide heater [55], situated between the objective and the thin PDMS-sheet upon which the cells were cultured. Fluid flow did not elicit detectable intracellular calcium responses.
Cells were imaged with a TE2000-U Nikon inverted microscope, using a 20× objective, a standard 100W mercury lamp, and a 490 nm long pass dichroic mirror. A CoolSnap HQ2 camera (Photometrics, Tucson, AZ) was used to capture fluorescence images of YC3.60-transfected cells. Cells were excited at 450 nm and the emission signals were captured at 490 and 535 nm (filters from Chroma Technology Corp, Rockingham, VT). An ND4 neutral density filter was used to reduce photo-bleaching. The excitation and emission filter wheels were controlled by the Lambda 10-3 Shutter Controller (Sutter Instruments, Novato, CA). Images were acquired every 3 s, and an exposure time of 100 ms was used. The program MetaFluor (Molecular Devices, Downington, PA) was used for image acquisition and processing; for each emission image (at 490 nm and 535 nm) the background was subtracted, ratiometric images were constructed (intensity at 535 nm/intensity at 490 nm), and calcium FRET ratios of individual cells were generated with this software. These FRET ratios (I) were normalized by the minimum FRET ratio obtained in the experimental run (I0), and accordingly I/I0 was plotted in our figures, as has been done previously [56].The normalized ratio values of the calcium peaks fell between 1.2 and 7.5, which was in accord with previously obtained values using the same FRET indicator [50].
The resulting images were then analyzed to calculate the phase-locking ratios by dividing the number of calcium spike events by the number of CCh stimulation inputs. Since at least several cells always responded to a particular stimulation pulse, we concluded that when cells did not respond, it was due to phase-locking and not a malfunction with the microfluidic setup (Video S1).
Cells were exposed to 9–18 stimulation inputs, and the number of calcium responses for each run was recorded. For instance, for a cell that had been exposed to 12 CCh stimulation pulses and responded with 6 calcium spikes, the phase-locking ratio was computed as 0.5. Calcium spikes that were above levels of background noise (typically more than 10% maximum calcium spike height) but did not reach an amplitude greater than 33% of the maximum calcium spike height were not counted as true calcium spikes and were deemed sub-threshold calcium spikes (See Fig. 1 and Fig. 2 in Text S1). Phase-locking ratios were computed for individual cells, and averages and standard errors of the mean were computed for each experimental condition. Statistics were based upon three experiments (each of no less than 20 cells) for each experimental condition. Between 85–106 cells were examined for each experimental condition. The unpaired Student t-test was used to statistically compare pairs of experimental conditions; p<0.05 was used as a threshold of statistical significance.
Nine mathematical models of oscillatory calcium signaling were evaluated in our study: the Chay et al. model [10] (Fig. 2a), the positive feedback Politi et al. model [11] (Fig. 2b), the Cuthbertson and Chay model [38], the Li and Rinzel model [13], the Atri et al. model [16], the Chay et al. and Politi et al. models with ligand/receptor/G-protein dynamics from Ref. [57], the Dupont et al. model [17], and the Kummer et al. model [39]. For all these mathematical models, we used the equations and initial conditions defined in the original publications (except for the Li and Rinzel model, for which we used the adaption developed in Sneyd et al. [5]); model equations, parameters, initial conditions, and brief model descriptions for all models used in this study are provided in Text S1, starting on page 13. For the Chay et al. model, it was assumed that receptor-mediated G-protein activation was proportional to stimulant concentration. For the Politi et al. model, it was assumed that the maximal rate of PLC-mediated IP3 production was proportional to stimulant concentration. These assumptions are based upon those from the original publications. For the Politi et al. model, we used calcium flux strength ε = 5 to reflect the role of extracellular flux in calcium oscillations [58]. The mathematical systems were exposed to 12 square-wave stimulation pulses and the corresponding number of calcium spike responses was counted in order to compute phase-locking ratios; the criteria for assessing the phase-locking ratio were the same as those for experiments, as described earlier in the Materials and Methods Section. To assess the effect of rest period on the phase-locking ratio, this parameter was varied, while stimulant concentration and stimulation duration were fixed; we then plotted the resulting phase-locking ratio against the rest period (Fig. 3- middle row). The same procedure was applied to assess the effects of stimulant concentration and stimulation duration on the phase-locking ratio, respectively (See Fig. 3 in Text S1). Stimulation parameters for the mathematical models were chosen such that the range in behaviors under periodic stimulation matched those observed in experiments. The stimulation concentration ‘C’ is represented differently for each model, as noted in Text S1.
Original parameters were used for both circadian models [48], [49].
All models were coded in MATLAB version 7.8.0 (MathWorks Inc, Natick, MA) and the system of ODEs was solved with the stiff solver ode15s.
We used Latin Hypercube Sampling (LHS) to check if inaccuracies in model parameter values alone could account for differences between experimental results and model predictions. LHS is a highly effective method for exploring parameter spaces for mathematical models [59], [60], [61], [62]. Using LHS code from Marino et al. [60] (http://malthus.micro.med.umich.edu/lab/usadata/), we varied model parameter values by sampling from a normal distribution with a 25% standard deviation; original parameter values were used as the mean. Larger standard deviations (100%) did not yield results different from those at 25% standard deviation. We also sampled parameters from a uniform distribution; the boundaries of the distribution were set by using one tenth of the original parameter value as the minimum and ten times the original parameter value as the maximum. As was the case for sampling from a normal distribution, sampling from a uniform distribution did not yield any parameter sets that could account for the discrepancies between models and experiments. For the Chay et al. model, we varied all twelve independent parameters; for the Politi et al. model, we varied all 17 independent parameters, except for β, which represented the ratio of ER to cytoplasm volume. LHS was run for 500 iterations on each model, and each model output was analyzed to decipher whether the results matched experimental observations (either by constructing ‘phase-locking ratio vs. rest period’ graphs for the Politi et al. model or by looking at individual model runs for the Chay et al. model, as depicted in Fig. 4).
|
10.1371/journal.pbio.3000021 | If a fish can pass the mark test, what are the implications for consciousness and self-awareness testing in animals? | The ability to perceive and recognise a reflected mirror image as self (mirror self-recognition, MSR) is considered a hallmark of cognition across species. Although MSR has been reported in mammals and birds, it is not known to occur in any other major taxon. Potentially limiting our ability to test for MSR in other taxa is that the established assay, the mark test, requires that animals display contingency testing and self-directed behaviour. These behaviours may be difficult for humans to interpret in taxonomically divergent animals, especially those that lack the dexterity (or limbs) required to touch a mark. Here, we show that a fish, the cleaner wrasse Labroides dimidiatus, shows behaviour that may reasonably be interpreted as passing through all phases of the mark test: (i) social reactions towards the reflection, (ii) repeated idiosyncratic behaviours towards the mirror, and (iii) frequent observation of their reflection. When subsequently provided with a coloured tag in a modified mark test, fish attempt to remove the mark by scraping their body in the presence of a mirror but show no response towards transparent marks or to coloured marks in the absence of a mirror. This remarkable finding presents a challenge to our interpretation of the mark test—do we accept that these behavioural responses, which are taken as evidence of self-recognition in other species during the mark test, lead to the conclusion that fish are self-aware? Or do we rather decide that these behavioural patterns have a basis in a cognitive process other than self-recognition and that fish do not pass the mark test? If the former, what does this mean for our understanding of animal intelligence? If the latter, what does this mean for our application and interpretation of the mark test as a metric for animal cognitive abilities?
This Short Report received both positive and negative reviews by experts. The Academic Editor has written an accompanying Primer that we are publishing alongside this article (https://doi.org/10.1371/journal.pbio.3000112). The linked Primer presents a complementary expert perspective; it discusses how the current study should be interpreted in the context of evidence for and against self-awareness in a wide range of animals.
| The ability to perceive and recognise a reflected mirror image as self is considered a hallmark of cognition across species. Here, we show that a fish, the cleaner wrasse, shows behavioural responses that can be interpreted as passing the mark (or mirror) test, a classic test for self-awareness in animals. We ask whether these behaviours should be taken as evidence that fish are self-aware or whether the test itself needs to be revised. In particular, we interrogate whether tests such as these can be reliably employed in animals as divergent from humans as fish and how we might understand cognition in nonprimates.
| The mark test, in which a coloured mark is placed on a test subject in a location that can only be viewed in a mirror reflection, is held as the benchmark behavioural assay for assessing whether an individual has the capacity for self-recognition [1,2]. In human infants, approximately 65% of individuals pass the mark test by 18 mo of age by touching the mark with their hands while viewing their reflection [3], although some individuals pass earlier, and some never pass. Accumulating reports claim that many other animal species also pass the mark test, including chimpanzees [1], elephants [4], dolphins [5,6], and corvids [7], while many other species are apparently unable to pass the test [8] (but see [9–11]). Nevertheless, the interpretation of these results is subject to wide debate, and the certainty with which behavioural responses during the mirror test can be taken as evidence of self-awareness in animals is questioned [8,12,13]. This problem is exacerbated when the taxonomic distance increases between the test species and the primate taxa for which the test was initially designed. For instance, can the behavioural results recorded for chimpanzees during the mirror test be meaningfully compared with the responses of a bird? If so, does this mean a bird that passes the mirror test is self-aware? More generally, if we are interested in understanding and comparing cognition and problem solving across taxa, can we assume that equivalent behaviours represent equivalent underlying cognitive processes? With particular reference to the mark test, here we explore what forms of behaviour in fish could be taken as evidence of self-awareness and whether the same conclusions that have been drawn in other taxa can also be drawn for fish.
Given that the mark test as designed for primates relies on hand gestures towards the marked region and changes in facial expression, we also ask whether it is even possible to interpret the behaviour of divergent taxonomic groups during the mark test in the same way as for the taxa for which the test was initially designed. If not, the usefulness of the mark test across taxa must be questioned, as should our confidence in sharp divisions in cognitive abilities among taxa. To explore these questions, we here test whether a fish, the cleaner wrasse L. dimidiatus, displays behavioural responses that can be interpreted as passing the mark test. We then ask what this may mean for our understanding of self-awareness in animals and our interpretation of the test itself.
To date, no vertebrate outside of mammals and one bird species has passed the mark test. This is despite many species in other vertebrate classes, such as fish, showing sophisticated cognitive capacities in other tasks [14–17], including transitive inference [18,19], episodic-like memory [20], playing [21], tool use [22,23], prediction of the behaviour of others by using one’s own experience during coordinated hunting [24,25], cooperating to warn about predators [26,27], and cooperative foraging [28]. These studies reveal that the perceptual and cognitive abilities of fish often match or exceed those of other vertebrates [15,17] and suggest the possibility that the cognitive skills of fish could more closely approach those found in humans and apes [14,16,17,24,28]. Clearly, a claim such as this requires rigorous testing to be held up and an accepted framework in which the results of any test can be interpreted.
It can be challenging to employ standardised cognitive tests across species when performance in the test depends on specific behavioural responses that are not present in all taxa or, perhaps more importantly, that are difficult for human observers to objectively interpret. This may be the case for the mark test, which has been specifically designed to suit the behavioural repertoire of primates [1,2]. Animals that cannot directly touch the marks used in mirror self-recognition (MSR) tests may therefore be inherently poor test candidates [2,5,29] regardless of their cognitive abilities, making direct comparison across taxa challenging [30–33]. Manta rays (Chondrichthyes), for instance, show unusual behaviour on exposure to a mirror, and it has been suggested these are self-directed behaviours in response to seeing their own reflection [34], although no mark test was performed and this interpretation is contested [35,36]. This controversy highlights the need to ask what type of behavioural response would be taken as evidence of contingency testing, self-directed behaviour, or self-exploration in an animal with such divergent morphology and behaviour from typical test species.
To make a comparison across taxa, initially it may be useful to choose species with perceptual abilities and a behavioural repertoire that allow them to respond to coloured marks placed on the body (this is not a given when the sensory systems of animals differ so greatly) and do so in a manner that can be effectively interpreted by a human observer. The cleaner wrasse, L. dimidiatus, is potentially such a species because it forms mutualistic relationships with larger client fish by feeding on visually detected ectoparasites living on the skin of the clients [37]. Therefore, the cleaner wrasse has sensory and cognitive systems that are well equipped for visually detecting spots of unusual colour on the skin surface, as well as the motivation to behaviourally respond to marks. Importantly, the natural response to removing parasites from clients—directly biting them—would result in cleaner wrasse biting at the mirror surface rather than performing self-directed behaviour, which would constitute failing the test. The role of hard-wired behavioural responses to parasites could therefore be ruled out. Additionally, this species is highly social, interacting with the same individuals repeatedly over long periods of time, and has sophisticated cognitive abilities, including tactical deception [38–40], reconciliation [41], and the ability to predict the actions of other individuals [41,42]. These are traits requiring cognitive abilities that may be correlated with the ability for self-recognition [e.g., 16,29,43–45].
During the mark test, animals must visually locate a mark in a mirror image that cannot be viewed directly. Given their sensory biology, it is reasonable to predict the wrasse will notice the coloured marks and that marks may generate an attentional response that culminates in a removal attempt [46–47]. However, lacking hands or trunks, any attempt to remove or interact with the mark would necessarily take a different form than is seen in many other taxa. Fortunately for the question at hand, many fish display a characteristic self-directed behaviour that functions to remove irritants and/or ectoparasites from the skin surface, termed glancing or scraping [48,49]. Similarly, mammals such as dolphins that lack hands may scrape their own bodies, and this behaviour has been interpreted as self-directed behaviour during application of the mark test in those species [29,50]. We therefore consider the cleaner wrasse to possess the prerequisite sensory biology and behavioural repertoire to adequately implement the mark test and here use a modified experimental design to test for MSR in a fish. Importantly, this experiment allows us to ask a broader question of whether the criteria that are accepted as evidence for MSR in mammals and birds can be applied to other taxa, and if these fish fulfil these criteria, what it means for our interpretation of the test itself.
Prior to the provisioning of a coloured mark, transitions among three behavioural phases after initial exposure to a mirror are typically observed [1,4,5,6]. These transitions among behavioural phases are interpreted as additional evidence of self-recognition, although in themselves do not constitute passing the mark test, which specifically requires mark removal attempts [1,4]. The first phase (i) is a social reaction towards the mirror, apparently as a consequence of the reflection being perceived as an unknown conspecific. In phase (ii), animals begin to repetitively perform idiosyncratic behaviours that are rarely observed in the absence of the mirror. These behaviours are interpreted as contingency testing between the actions of the subject and the behaviour of the reflection [e.g., 1,4]. In phase (iii), the subject begins to examine their reflection and uses the mirror to explore their own body in the absence of aggression and mirror-testing behaviour [1,4,5]. Finally, a coloured mark is applied, and observations of removal attempts are recorded. Here, we first tested whether the cleaner wrasse pass through all three behavioural phases upon exposure to a mirror (Fig 1) and then provided a mark using subcutaneous injections of transparent or pigmented elastomer to test for removal attempts.
Prior to starting the experiments, the focal fish swam around the tank and showed no unusual reactions to the covered mirror. Immediately after initial exposure to the mirror, seven of 10 fish responded aggressively to their reflection, attacking it and exhibiting mouth fighting (Fig 1 and S1 Video [45,46]), suggesting that the focal fish viewed the reflection as a conspecific rival. The frequency of mouth fighting was highest on day 1 and decreased rapidly thereafter, with zero occurrences by day 7 and almost no aggression throughout the remainder of the experimental period (Fig 1A; cf. a similar decrease in aggression seen in chimpanzees and shown in Figure 2 of [1]). This initially high and subsequently decreasing aggression is consistent with phase (i) of the mark test as reported in other taxa.
As mouth fighting towards the mirror reflection decreased, the incidence of atypical behaviours (e.g., swimming upside-down, a highly unusual behaviour typically never observed in cleaner wrasse; Table 1 and S2 and S3 Videos) significantly increased and was highest on days 3 to 5 (Fig 1A). On days 3 and 4, the estimated average frequency of these atypical behaviours across the seven individuals was extremely high—36 times per hour. Each of these atypical behaviours was of short duration (≤1 s), often consisting of rapid actions with sudden onset within 5 cm of the mirror, and could be loosely grouped into five types (Table 1). While it is possible to interpret these behaviours as a different form of aggression or social communication, they have not been recorded in any previous studies of social behaviour in this species [46] and were not likely to be part of a courtship display because all of the subject fish were females. Moreover, we did not observe these behaviours in our own control experiments when presenting a conspecific across a clear divide (Fig 1C), further demonstrating they were unlikely to be forms of social communication.
These atypical behaviours were individually specific, with each fish performing one or two types of behaviour (Table 1; Fisher’s exact probability test for count data with simulated P value based on 2,000 replicates of P = 0.0005). Crucially, these behaviours occurred only upon exposure to the mirror and were not observed in the absence of the mirror (i.e., before mirror presentation) or during conspecific controls. Almost all of the behaviours ceased by day 10 (Fig 1A) and were rarely observed thereafter. These behaviours were different from the previously documented contingency-testing behaviours of great apes, elephants, and magpies [1,4,7], but given the taxonomic distance between them, this could hardly be otherwise. While primates and elephants may perform more anthropomorphic behaviours such as changing facial expression or moving the hands, legs, or trunk in front of the mirror, wrasse and other fishes cannot perform behaviours that are so easily interpreted by a human observer. Nevertheless, behaviours such as upside-down swimming are indeed unusual for a healthy fish and could represent alternative indices of contingency that are within the behavioural repertoire of the study species. Moreover, the atypical movements observed in cleaner wrasse were consistent with behaviour previously interpreted as contingency testing in other species [1,4,5,7] in that these behaviours were atypical and idiosyncratic, repetitive, displayed only in front of the mirror, absent in the absence of a mirror, shown after a phase of initial social (here aggressive) behaviour, displayed over a short period of time, and distinct from aggressive behaviour. Although we reserve judgement as to whether these behaviours should unequivocally be interpreted as evidence that these fish are examining and perceiving the reflection as a representation of self, we nevertheless argue that on an objective basis, these behaviours fulfil the criteria as presented for contingency testing and are consistent with phase (ii) of MSR as presented for other taxa [1,4,5,7].
In phase (iii), species that pass the mark test increase the amount of time spent in front of the mirror in nonaggressive postures, apparently visually exploring their own bodies [1,4,5,7]. This interpretation is again rife with pitfalls because it requires an assessment of the intentionality of nonhuman animal behaviour. An agnostic approach is to simply measure the amount of time animals spend in postures that could reflect the body in the mirror [2], giving an upper measurement of the time in which animals could observe their reflection while making no inferences about the intentionality of the act. We observed an increase in the amount of time spent in nonaggressive postures while close to the mirror (distance of <5 cm), peaking on day 5 after mirror presentation and remaining consistently higher than days 1 to 4 (Wilcoxon sign-ranked test, T = 36, P = 0.008; Fig 1A). Although we did not observe directed viewing behaviour as seen in chimpanzees and elephants, this would in any case be difficult given challenges of assessing gaze direction in animals like fish (although see [45] for a recent technological solution). We therefore consider that in terms of time spent in postures that would facilitate viewing the mirror reflection, this behaviour was consistent with phase (iii) of MSR.
Species with MSR distinguish their own reflection from real animals viewed behind glass [e.g., 29]. When we exposed naïve cleaner wrasse to conspecifics behind glass, we observed fundamentally different responses towards their mirror image (S1 Text). Aggressive behaviour frequency towards real fish was generally low yet did not diminish appreciably during the 2-wk testing period (Fig 1C). Time spent within 5 cm of the glass in the presence of conspecifics was also higher than that in the presence of the mirror. Importantly, no atypical or idiosyncratic behaviour (that might be considered contingency testing) was exhibited towards conspecifics. These behaviours were only observed upon exposure to the mirror. Similar to many previous MSR studies [1,4,5,7], not all individuals we tested passed through every phase of the test. After the initial presentation of the mirror, three fish showed low levels of aggression and rarely performed atypical behaviours during period E1 (Fig 1B). Instead, these three individuals spent relatively longer periods in front of the mirror, as is typically observed during phase (iii), and we conclude these fish failed the test (but see S1 Text for an alternative explanation).
In the second part of the experiment, we used a modified standard mark test protocol to assess reactions to visible (pigmented) or sham (transparent) marks. We used subcutaneously injected elastomer (see Materials and methods) to apply a small amount of colour below the skin surface, a widely used procedure that has been repeatedly shown not to affect fish behaviour [51–54, Northwest Marine Technology]. Moreover, the combined use of coloured and transparent sham marks provided an internal control for the effects of application, including irritation or tactile sensations around the marking site. Nevertheless, the procedure certainly resulted in higher tactile stimulation than, e.g., paint marks on elephant skin, necessarily so because of the requirements of provisioning marks in the aquatic environment and on animals covered in a protective mucus coating. We must therefore consider recent studies showing that visual–somatosensory training induced self-directed behaviour in rhesus monkeys [10,11] that could not be achieved through visual stimuli alone. Our study differs in that we do not provide direct somatic stimulation during the mark test and that we observed no response during our sham-mark phases, which also used a subcutaneous injection. However, given the nature of the mark application, we cannot rule out that a combination of visual and tactile cues produces the behavioural responses we describe, and our test might therefore be considered more similar to the modified tactile–visual mark test than the original mark test.
Fish were marked at night while under anaesthesia, and they swam normally the next morning in the no-mirror condition. After the initial settlement period ‘E1’ (i.e., the initial 2 wk of phases i–iii), we evaluated behaviour during periods ‘E2’ (no mark), ‘E3’ (injection with transparent sham mark), ‘E4’ (injection with coloured mark with no mirror present), and ‘E5’ (coloured mark with mirror present) during a subsequent 2-wk period. The sham and coloured marks were applied on the right side of the head of two fish, on the left side of the head of two other fish, and under the throat in a further four fish; these areas were only visible in the mirror. Each mark was in the form of a small brown mark with the intention of mimicking a natural ectoparasite in colour, size, and shape.
We first examined whether fish assumed postures in front of the mirror that would reflect the marked site by categorising all body postures performed within 5 cm of the mirror into three categories: postures exposing the right side of the head to the mirror, postures exposing the left side of the head, and frontal–vertical postures exposing the head, throat, and underside to the mirror. These postures would reflect the right face mark, the left face mark, and the throat mark, respectively. We predicted that if fish were attempting to observe the coloured marks on body parts reflected in the mirror, they would assume postures that facilitated this observation of the mark significantly more frequently during E5 (mirror, colour mark) than in E2 (mirror, no mark) or E3 (mirror, transparent sham mark). Two independent analyses of the videos were conducted (by MK and JA), as well as two further blind analyses by unrelated researchers of a subset (15%) of the videos; the frequencies were highly correlated between the analyses (r = 0.988).
Posturing behaviours that would reflect the marked sites during periods E2 and E3 were infrequent, and all sides were presented equally (Fig 2A) in all fish except fish #7 (Table 2), suggesting the marking procedure itself had minimal effect on posturing behaviour. In contrast, time spent posturing while viewing the marked sites was significantly higher in the colour-marked (E5) versus no- (E2) and sham-marked (E3) periods (Fig 2A). This pattern held for all individuals except fish #2, regardless of the sites marked (Table 2). Note that no comparisons to E4 can be made with respect to observations of reflections because no mirror was present during that period. Moreover, the time spent in postures reflecting the two remaining unmarked sites (e.g., right side of head and throat for a fish marked on the left side of the head) were not different among periods (Fig 2B). Taken together, these findings demonstrate that cleaner wrasse spend significantly longer in postures that would allow them to observe colour-marked sites in the mirror reflection, and in previous studies on dolphins, similar patterns of activity were considered to constitute self-directed behaviour [5]. These reactions also demonstrate that tactile stimuli alone are insufficient to elicit these responses because they were only observed in the colour mark/mirror condition. Rather, direct visual cues or a combination of visual and tactile stimuli are essential for posturing responses in the mark test.
Although they cannot touch their own bodies directly, many species of fish scrape their bodies on a substrate to remove irritants and/or ectoparasites from the skin surface [48,49]. When we marked fish with brown-pigmented elastomer on the lateral body surfaces in locations that could be viewed directly (i.e., without the mirror), we observed increased scraping behaviour on the site of the mark (S1 Fig). We therefore hypothesised that when marked in locations that could only be seen with the aid of a mirror, wrasse would similarly scrape their bodies in an attempt to remove the marks. We hypothesised this would occur in fish after viewing these marks in the mirror, and crucially, that they would not scrape transparent sham marks nor coloured marks in the absence of a mirror. Like many natural behaviours, some scraping of the body flanks was observed outside the mirror condition in our studies and was also difficult to distinguish from face scraping. Because of this, we restricted our analysis only to throat scraping and took this behaviour as the only evidence of a putative self-directed behaviour because it was never observed outside the period E5 in any of the subject fish.
After throat marking, three out of four fish scraped their throats against the substrate upon exposure to the mirror during period E5 (Fig 3 and S4 Video), but none of the four fish exhibited this behaviour during E2–E4 (control, transparent mark, and coloured mark without a mirror). This is a ratio comparable to other species tested previously; one of three Asian elephants passed the test [4], as did two of five magpies [6]. In total, we observed 37 separate instances of throat scraping during E5 (15 for fish #1, 16 for fish #4, 6 for fish #21; Friedman test, χ2 = 9.0, degrees of freedom (df) = 3, P = 0.029; binomial test within individuals, E2, E3, and E4 versus E5: 0 versus 15 scrapings, P < 0.0001 in fish #1, 0 versus 16 scrapings, P < 0.0001 in fish #4, 0 versus 6 scrapings, P = 0.031 in fish #21). The motivation for scraping the mark is potentially to remove a perceived ectoparasite, which these wild-caught fish would have experienced previously. Crucially, these scraping attempts are the opposite to what would be expected if cleaner wrasse were ‘hard-wired’ to remove anything resembling a parasite. In this case, we would expect fish to attempt to bite at the mark itself as though they were cleaning a client. To control for this possibility, we placed identical marks on the surface of the mirror itself but observed no attempts to remove these marks nor any scraping behaviour (S1 Text), demonstrating the scraping behaviour during the mark test was not a consequence of an innate response to marks that resemble parasites. Given that scraping behaviour is accepted as being self-directed in mammals during the mark test [29,50], we similarly interpret this behaviour as being self-directed in fish. Alternative interpretations risk introducing subjective taxonomic biases, setting moving goal posts, and precluding scientific comparisons among certain taxa. If scraping behaviour is therefore interpreted as self-directed, these results constitute compelling evidence that three of the four throat-marked fish passed through all prephases of the test and subsequently attempted to remove visually perceived coloured marks from their bodies after viewing them in the mirror. By extension and comparison to similar mark test studies, this leads to the crucial question of whether fish are aware that the mirror reflection is a representation of their own body.
The mark test is a controversial assessment of animal cognition [8] and perhaps even more so when applied to fish, a taxonomic group considered by some to have lesser cognitive abilities than other vertebrate taxa. Nevertheless, we provide compelling evidence that cleaner wrasse show behavioural responses that can be reasonably interpreted as passing through all stages of the mark test and attempt to remove a mark only when it is able to be viewed in the mirror (Fig 3). The results we present here will by their nature lead to controversy and dispute, and we welcome this discussion. We consider three possible interpretations of our results and their significance for understanding the mark test. The first (I) is that the behaviours we document are not self-directed and so the cleaner wrasse does not pass the mark test; the second (II) that cleaner wrasse pass the mark test and are therefore self-aware; and the third (III) that cleaner wrasse pass the mark test, but this does not mean they are self-aware.
If the reader takes position (I), rejecting the interpretation that these behaviours are self-directed, it should be necessary to justify the grounds for this rejection. As noted above, touching or scraping behaviour is taken as evidence of a self-directed behaviour in mammals, and so if these and other behaviours are not similarly considered self-directed in fish, the question must be asked why. For a test to be applicable across species, an objective standard is required. What behavioural criteria need to be fulfilled to define self-directedness in a fish? What is the definition of contingency testing in animals with vastly divergent sensory ecologies? How do we determine an animal is visually exploring its own body when its visual system is nothing like our own? Without such a standard, the behaviours shown in the mark test can be differently assessed depending on the taxon being investigated. This introduces an impossible and unscientific standard for comparison that can never be resolved by debate among differing subjective opinions and therefore undermines the value of the mark test as a comparative tool. This may be an inherent difficulty in comparative studies of animal behaviour, but we do not consider it intractable. Rather, we see great value in computational approaches to behavioural analysis [55] (S5 Video), allowing researchers to decompose behaviour into constituent elements and ask, e.g., whether some kinematic signatures of behaviour are only observed during specific periods of the mark test, or to compute the visual field and determine whether an animal is truly able to see its own reflection. This may allow an objective standard for assessing whether behaviours are unusual, idiosyncratic, or contingent based on quantitative rather than qualitative analysis. It would at the very least provide a quantitative basis for categorisation of different behaviours and thereby facilitate comparison and discussion.
Alternatively, if the behaviours reported here in cleaner wrasse are accepted as being functionally equivalent to those in other taxa during the mark test, position (II) or (III) must be taken. The original interpretation of the mark test by its inventor Gallup posits that species passing the mark test are self-aware [1,56]. A strict adherence to this interpretation would logically lead us to take position (II), that cleaner wrasse are also self-aware. This would require a seismic readjustment of our cognitive scala naturae. We are more reserved with our interpretation of these behaviours during the mark test with respect to self-awareness in animals and therefore take position (III). We do not consider, even if our observations are taken as successful behavioural responses to all phases of the mark test, that this should be taken as evidence of self-awareness in the cleaner wrasse. Rather, we consider the interpretation that makes fewest assumptions to be that these fish undergo a process of self-referencing [32,57], in which direct or indirect (e.g., in a mirror reflection) observations of the physical self are perceived as part of one’s own body by the observer but without this involving theory of mind or self-awareness [32,57].
Our conclusion is therefore that cleaner wrasse show behavioural responses that fulfil the criteria of the mark test as laid out for other animals, but that this result does not mean they are self-aware. This position raises a number of difficult questions. Can passing the mark test be taken as evidence of self-awareness in one taxon but not another? We argue not, because a position that holds the same results in a standardised test can be interpreted different ways depending on the taxon from which they are gathered is both logically untenable and taxonomically chauvinistic [58]. Are we instead mistaken in our conclusion that these behaviours even fulfil the criteria of the test? If so, this ambiguity suggests the mark test needs urgent re-evaluation in the context of comparative cognition studies. Finally, while we make no claims that our study proves fish are self-aware, we do hope our results ignite further discussion of fish as cognisant, intelligent animals.
All experiments were conducted in compliance with the animal welfare guidelines of the Japan Ethological Society and were specifically approved by the Animal Care and Use Committee of Osaka City University.
The cleaner wrasse, L. dimidiatus, is a protogynous hermaphrodite teleost that lives in coral reef habitats [46,59]. We used 10 wild fish obtained from commercial collectors in this study. Prior to our experiments, the fish were housed in separate tanks (45 cm × 30 cm × 28 cm), and each fish was kept for at least 1 mo prior to beginning the experiments to ensure acclimation to captivity and the testing conditions and that they were eating and behaving normally. Fish were between 51–68 mm in length; this is smaller than the minimum male size, thus strongly suggesting that these individuals were functionally female. Individual fish sizes were as follows: 68 mm for fish #1, 62 mm for fish #13 and #20, 61 mm for fish #21, 58 mm for fish #4, 55 mm for fish #5, 53 mm for fish #6, 52 mm for fish #2 and #7, and 51 mm for fish #3). Each tank contained a 5 cm × 5 cm × 10 cm rock in the corner and a PVC pipe that provided shelter on a coral-sand substrate 3–4 cm deep. The water was maintained at 24°C–26°C and was aerated and filtered. The fish were maintained on a 12 h:12 h light/dark cycle. Artificial flake food (Tetramin USA) and small pieces of diced fresh shrimp were given twice daily.
The mirror presentation method (e.g., duration, timing, position, and mirror size and shape) has important consequences for successful implementation of MSR studies [1,4,5]. We presented a 45 cm × 30 cm high-quality mirror on a glass wall of the same size inside the experimental tank. The mirror was positioned at night, while the fish were sheltered within the PVC pipe, 1 wk before beginning the experiments; it was then completely covered with a white plastic sheet (45 cm × 30 cm). At the start of the experiments, the white cover on the mirror was removed, and the subject fish were exposed to the mirror until the end of the series of experiments, with the exception of a half-day experiment during which the mirror was completely covered with the white sheet (see below).
We studied fish behavioural responses during five experimental periods, from experiment 1 (E1) to experiment 5 (E5; in chronological order). All behaviours were recorded by video camera (HDR-CX370; Sony, Tokyo, Japan), and the field of view encompassed the entire experimental tank. After the start of the experiment (i.e., uncovering of the mirror), the behavioural responses of fish were video recorded eight times over a 2-wk period: every day for the first 5 days and then again on days 7, 10, and 15 (period E1: initial responses to mirror). Three or four days after the end of period E1, fish behaviours were recorded for two consecutive days as a control for the mark test (E2: no mark). Two or three days after E2, a nonvisible sham mark (white or nonpigmented elastomer) was made on all 10 fish, and behavioural responses were recorded over the next two days (period E3: sham marking). Two days after period E3, the fish were marked with a coloured (brown) elastomer mark at night, and their behaviours were recorded in the morning of the following day while the mirror was covered with the 45 cm × 30 cm white plastic sheet (period E4: colour mark with no mirror present). After this observation phase, the mirror was uncovered, and behaviours were recorded on the afternoon of the same day and also on the morning of the following day (period E5; colour mark with mirror present). All experiments were necessarily run in this order because coloured tags could not be removed from fish once applied; hence, transparent (sham) marks always preceded coloured marks. The four fish that were marked on the head showed an increase in scraping of the marked area during period E5. However, three of these fish were also observed scraping facial areas prior to colour marking, indicating that face scraping cannot be taken as unequivocal evidence of mirror-induced self-directed behaviour.
Elastomer tags and visible implant elastomer (VIE) marking (Northwest Marine Technology Inc., Shaw Island, USA) via subcutaneous injection are widely used in studies of individually marked live fish and do not affect fish behaviour [51–54, NMTI]. Our fish were taken from their tanks at night together with their PVC pipe and placed in eugenol solution to achieve mild anaesthesia (using FA100; Tanabe Pharmacy, Tokyo, Japan). A nonpigmented gel mark was injected subcutaneously in an area of 1 mm × 2 mm at one of three sites during the sham mark period: on the right side of the head (two fish), on the left side of the head (two fish), or under the throat (four fish). The entire injection process took no longer than 5 min, and the fish were returned to their original tank together with the pipe after the mirror was covered with the white plastic sheet. We ensured that the fish were swimming normally the next early morning and that they showed no behavioural changes as a consequence of the tagging procedure. We initially used white pigment on the pale-coloured body areas but found that the skin in these areas had a slight blue tint and that the white tag was visible in two fish; these fish were not used in further experiments. A brown-pigmented elastomer colour mark was applied as a colour mark at night before the day of E4. After confirming that all marks were of the same size (1 mm × 2 mm), the fish were returned to the tank. Given the location of the tags relative to the field of view of cleaner wrasse, direct observation of the marks on the head was unlikely and was definitely impossible for throat marks. To standardise the testing procedure, the brown-coloured mark was injected at the throat directly adjacent to the transparent marked site. Even with both marks applied, the total volume of the tag was lower than the minimum recommended amount, even for small fish, and <13% of the size of tags used in studies with other fish (biologists who applied VIE to small fish in previous studies, i.e., 26-mm brown trout [51] and 8-mm damselfish [54], stated that the amounts used were minute, but for the former species, 2–3 mm tags were made with 29 G needles [51]. Willis and Babcock used large tags (10 mm × 1 mm × 1 mm [127/ml]) in Pagrus auratus [53]). Our own tagging method was therefore very unlikely to have caused irritation. Moreover, we saw no evidence during period E4 (colour tag, no mirror present) of any removal attempts or scratching behaviour, further confirming that the tags did not stimulate the fish to perform any of the behaviours we report.
Videos were observed for all behavioural analyses. Fish performed mouth-to-mouth fighting frequently during period E1, and the duration of this behaviour was recorded (Fig 1 and S1 Video). Unusual behaviours performed in front of the mirror, which have never been observed before in a mirror presentation task nor in the presence of a conspecific, were often observed during the first week of E1, and the type and frequency of these behaviours was recorded.
In the latter half of E1, fish occasionally swam slowly or remained stationary in front of the mirror, and the duration (in seconds) of these behaviours, when performed within 5 cm of the mirror, was recorded. The duration of postures in which the marked area was reflected in the mirror was recorded during E2 (no mark), E3 (sham mark), and E5 (coloured mark with mirror present). Posturing within 5 cm of the mirror was categorised into three types: right-sided posture (i.e., reflecting the right side of the head), left-sided posture (reflecting the left side of the head), and frontal–vertical posture (reflecting the throat). The duration (in seconds) of each of the three types of posture was recorded during six separate 5-min observation periods for a total of 30 min per fish for each of the periods when a mirror was present (E2, E3, and E5). A subset of 15% of the videos was blindly analysed by two researchers outside our team; their analysis was highly correlated with the main analysis (r = 0.887, P < 0.0001), and statistical tests showed no significant differences between the two data sets (two-way repeated-measures analysis of variance [ANOVA], blind effect: F = 0.06, P = 0.80; blind effect × observation site: F = 0.77, P = 0.45). Scraping behaviour, including the location on the body that was scraped, was recorded during periods E2–E5 when it occurred. During period E5, when the fish were colour marked and exposed to the mirror, individuals often displayed the marked site to the mirror immediately prior to and following a scraping behaviour. Therefore, we also recorded the time interval between displaying and scraping during E5 (S1 Text).
Statistical analyses were performed using SPSS (ver. 12.0; SPSS Inc., Chicago, IL, USA) and R software (ver. 2.13.2; R Development Core Team 2011). During period E1, the responses of the subject fish to the exposed mirror changed significantly over time. Changes in the duration of mouth fighting and time spent within 5 cm of the mirror over time were analysed with linear mixed models (LMMs). Similarly, changes over time in the duration of mouth fighting and time spent within 5 cm of the mirror were analysed with LMMs for the experiments using real fish across glass dividers. The frequency of unusual mirror-testing behaviours was analysed using a generalised linear mixed model (GLMM) with a log–link function and assuming a Poisson distribution. Time spent in postures reflecting the right side of the head, the left side of the head, and the throat were compared between mark types during the mark tests (E2: no mark, E3: sham mark, and E5: coloured mark with mirror present) using repeated-measures ANOVA. Note that the marked and unmarked positions were analysed separately. Individual-level statistics on postures that reflected the marked sites are shown in Table 2 (Mann–Whitney U test with duration in seconds of the six different behaviours per 5-min observation in periods E2, E3, and E5). To detect the effect of throat marking on the frequency of scraping behaviour, a Friedman test was used on the entire data set (E5 versus E2, E3, and E4) and a binomial test was used for comparison between periods (E5 versus E2, E3, and E4). No throat scraping or unusual behaviours were observed when individuals interacted with conspecifics across a glass divider, so no statistical tests were performed for that condition.
|
10.1371/journal.pcbi.1002640 | A Network-based Approach for Predicting Missing Pathway Interactions | Embedded within large-scale protein interaction networks are signaling pathways that encode response cascades in the cell. Unfortunately, even for well-studied species like S. cerevisiae, only a fraction of all true protein interactions are known, which makes it difficult to reason about the exact flow of signals and the corresponding causal relations in the network. To help address this problem, we introduce a framework for predicting new interactions that aid connectivity between upstream proteins (sources) and downstream transcription factors (targets) of a particular pathway. Our algorithms attempt to globally minimize the distance between sources and targets by finding a small set of shortcut edges to add to the network. Unlike existing algorithms for predicting general protein interactions, by focusing on proteins involved in specific responses our approach homes-in on pathway-consistent interactions. We applied our method to extend pathways in osmotic stress response in yeast and identified several missing interactions, some of which are supported by published reports. We also performed experiments that support a novel interaction not previously reported. Our framework is general and may be applicable to edge prediction problems in other domains.
| Networks of protein interactions encode a variety of molecular processes occurring in the cell. Embedded within these networks are important subnetworks called signaling pathways. Pathways are initiated by upstream proteins (called sources) that receive signals from the environment and trigger a cascade of information to downstream proteins (targets). Modeling the interactions that occur within this cascade is important because pathway disruption has been linked to several diseases. Further, the interactions help us better understand how cells respond to various conditions and environments. Unfortunately, interaction networks today are largely incomplete, which makes this analysis difficult. We provide a framework to model missing interactions in pathways by searching for interactions that putatively result in quicker and more efficient source-target cascades. We find that we can substantially shorten source-target distances with only a few additional edges and that many of our predicted edges have support in several knowledge databases and literature reports. We believe our approach will be useful to identify interesting and important pathway-centric interactions that have been missed by previous experimental assays.
| Networks of protein interactions can reveal how complex molecular processes are activated in the cell. However, even for model species, only a fraction of true physical interactions are known [1], [2] and experimental verification of all remaining potential interactions is unlikely in the near future. Furthermore, interactions are often condition- or tissue-specific [3] while current experimental methods often focus on one condition and one cell type [4]. Thus, computational techniques to predict protein interactions have flourished as a means to build more complete interaction maps [5], [6].
Signaling pathways are subnetworks of proteins that communicate via a series of interactions and are often only activated under specific conditions (e.g. stress response, development, etc.). Perturbations of proteins within such pathways have been linked to several diseases [7]. In addition, pathways are often conserved, thus studying their interactions in model organisms may help elucidate cellular response mechanisms in other organisms [8].
Signaling pathways typically contain upstream proteins (e.g. receptors on the cell's surface) that sense changes in the environment or that are directly involved in host-pathogen interactions. These proteins trigger a signaling cascade that leads to downstream transcription factors (TFs), which consequently carry forth regulatory programs. The former set of proteins can be considered sources that transmit information to a set of targets. Experimental protocols can infer source proteins based on their interactions with external stimuli (e.g. host-pathogen interactions [9]), and likewise targets can be determined via expression or knockdown assays. This motivated several techniques that have been proposed to extract pathways from global interaction networks by searching for efficient and robust paths between the given sets of sources and targets [10]–[13]. These techniques, however, do not try to infer putative interactions that are missing from the network. We model this problem computationally by searching for missing edges that increase the network's ability to explain the signaling cascade from sources to targets.
Many methods have been proposed to computationally predict protein-protein interactions. These methods leverage a variety of data sources, including physical docking models and protein structure [14], [15], evidence based on orthologous proteins in related species [16], microarray expression profiles [17]–[21], literature mining [22], sequence-level features [23]–[27], or a combination of heterogeneous features to learn a predictive model or classifier [28]–[32] (for reviews, see [5], [6]). Network-only approaches range from completing defective cliques [33] to analyses based on the shared topology or the distance between two candidate proteins [34], [35] to embeddings of the network to find non-interacting but adjacent proteins in the new space [36], [37]. None of these approaches, however, leverage known sources and targets to make pathway-aware predictions. Further, most other approaches use local cues of similarity, whereas our approach attempts to optimize a global distance function. There has also been theoretical work on predicting “shortcut edges” in graphs to minimize the average shortest-path distance amongst all nodes in the graph [38] or the diameter of the graph [39]–[42]; however, these works also do not exploit specific sources and targets when making predictions.
In this paper, we propose a combinatorial optimization framework to identify missing interactions that putatively mediate the passage of signals within pathways. Formally, we seek the edges to add to the network that maximally decrease the shortest-path distances between sources and targets (Figure 1). We consider several variants of the problem: an unrestricted setting where long paths are allowed; a restricted setting where source-target paths are bounded by a maximum number of hops; and a setting where each target is only required to be regulated by a single source. In computational experiments using a confidence-weighted protein interaction network for S. cerevisiae under the high osmolarity glycerol (HOG) osmotic stress response pathway, we find that we can drastically reduce source-target distances via the addition of only a few edges. Several new interactions predicted by our method, while missing from current databases, are supported by the literature; other interactions are novel predictions. We selected one of our novel predictions, , for condition-specific follow-up experiments. New knockout microarray experiments suggest that Sok2 is indeed functionally downstream of Tpk2 in the osmotic stress response, and previous evidence suggests that this could be due to Tpk2's direct phosphorylation of Sok2.
We first present our framework for predicting missing edges in graphs based on their ability to connect a given set of sources and targets. We show that our collection of problems are NP-hard to solve optimally and describe two efficient greedy optimization algorithms to address them. We then describe our testing setup, followed by our computational and experimental results.
We assume we are given a directed protein interaction network with nodes () corresponding to proteins and edges () to physical interactions. Protein interaction networks inferred from high-throughput experiments are often noisy [2], [43], therefore we assume each edge is weighted by a value denoting our confidence in the interaction [13]. We also assume we are given a set of sources and targets . The sources are typically upstream proteins in pathways that initiate a signaling cascade to the downstream targets (transcription factors). Our goal is to predict missing (directed) edges that lie centrally “in-between” the sources and targets. These edges putatively belong to the pathway but are not present in current databases. Formally:
Problem 1 [Shortcuts]. Given a directed and weighted graph and a set of sources and targets , add edges to to minimize , i.e. the total shortest-path distance between all source-target pairs.
We use the shortest-path distance to measure the distance between proteins and in the weighted network (as opposed to other distance measures, such as those based on random walks [44], [45]) because the shortest path represents a direct and specific series of high-likelihood signaling events.
The shortest path between two nodes in a weighted graph can be very long (either because the diameter is long or if the path uses many high confidence, and hence lowly weighted, edges). This may not be biologically reasonable since pathway targets are typically no more than 5 edges away from their closest sources [13]. Thus, we also propose a hop-restricted version of our problem. Let be the shortest-path distance between and that uses at most links ( if no such satisfying path exists). Formally:
Problem 2 [Shortcuts-X (restricted)]. Given a directed and weighted graph , a set of sources and targets , and a maximum allowable number of hops , add edges to to minimize , i.e. the total hop-restricted shortest-path distance between the pairs.
Both of these problems (general and hop-restricted) assumes that each transcription factor receives signal from each source. Another variant of these problems asks to minimize the distance between each target and any single source (biologically, the same source does not need to regulate all targets, but every target is regulated by some source). Formally:
Problem 3 [Shortcuts-SS (single source)]. Given a directed and weighted graph and a set of sources and targets , add edges to to minimize , i.e. the total shortest-path distance between each target and its single closest source.
We also consider the analogous problem in the hop-restricted setting:
Problem 4 [Shortcuts-X-SS (restricted, single source)]. Given a directed and weighted graph , a set of sources and targets , and a maximum allowable number of hops , add edges to to minimize , i.e. the total hop-restricted shortest-path distance between each target and its single closest source.
In the Supporting Text (Text S1 and Figure S1) we prove that these four edge predictions problems are NP-hard.
Given these hardness results, we consider a heuristic greedy algorithm for our suite of edge prediction problems. The Greedy algorithm selects edges to add iteratively: in each step, it predicts a single edge that maximally reduces the objective function. In the case of the Shortcuts problem, this means the algorithm will pick, from amongst all possible non-existent edges, the edge that maximally reduces the global shortest-path distance between all sources and targets.
In a network with nodes and directed edges, there are non-existent edges (excluding self-loops). In the yeast network we use, and , which means there are almost 20 million directed edges to test. Each edge can alter the shortest path from any source to any target hence, done navely, this would require recomputing the shortest-path lengths from each source to each target 20 million times just to add a single edge.
One trick to make the search more efficient is to notice that, if a candidate edge reduces the distance from source to target then the new shortest path from to consists of three components: the shortest path from to , the candidate edge , and the shortest path from to . If it does not reduce the distance, then the distance from to remains as it was without . Thus, the procedure can be made more efficient by pre-computing the shortest-path distances from every source to every other node in the network, and separately from every node in the network to every target. (This latter step can be further optimized by computing the distance from every target to every other node in the reverse graph, where edge directions are reversed.) To compute the cost reduction of candidate edge with weight we check if:(1)
The left-hand side sums the (pre-computed) distance from to , the weight of the new edge, and the distance from to ; the right-hand side is the previous distance from to without the new edge. (If we do not know the weight of the non-existent edge we set to encourage its usage; other values, e.g. based on the predicted likelihood of the interaction that is derived from other data sources may also be reasonable). The minimum of these two values is stored and is summed over each source-target pair, yielding the new objective function cost assuming exists in the graph. The edge that maximally decreases the cost function over all possible edges is added to the graph. Box 1 shows the pseudocode for the Greedy algorithm for the Shortcuts problem.
This trick reduces the algorithm's complexity in each step from in the nave case to . The first term considers all possible non-existing edges, each of which requires a constant lookup (Equation 1); the second term is the pre-computation of single-source shortest-path distances using Dijkstra's algorithm. Thus, we get a runtime reduction of a factor of , which in our case is roughly 60,000 for each iteration.
For the hop-restricted problems (Shortcuts-X and Shortcuts-X-SS), we seek short paths between sources and targets with the restriction that each path uses a maximum of hops. This bound stems from the fact that many pathways in signaling databases such as KEGG [46] depict on average 5 edges between a target and its closest source [13]. Other approaches have used similar bounds (3–4 [47]).
To constrain the shortest paths to use at most edges, we use a modified version of the Bellman-Ford algorithm [48], [49]. This algorithm computes single-source shortest paths starting from a node by relaxing every edge in each step (i.e. checking if traveling along the edge yields a shorter path to the destination node). Shortest-path distances are propagated through the graph and, as a result, after iterations, the algorithm computes the shortest-path distance from to every other node in the graph using at most hops.
Computing the updated cost for a candidate edge, however, requires a slightly different strategy than the one used before. The main challenge is that the new edge induces one hop, and hence, the two sub-cases ( and ) must be constrained to use hops in total. This leads to 6 possible cases to consider for the each candidate edge when computing the new distance from source to target , and each can be computed in constant time:(2)In the first case, the new path from to uses hop to reach , hop to reach (via the new edge), and hops to reach . The cost of this path consists of the Bellman-Ford distances shown (where e.g. is the distance from to that uses at most hops) plus the weight of the new edge (). Cases 2 and 3 follow similarly. If either endpoint of the candidate edge involves or , then a similar rule is checked (cases 4 and 5). Each case is considered and the one that yields the minimum distance is compared with the previous distance from to (without the new edge; case 6). For the Shortcuts-X problem, this is repeated for each source-target pair; for Shortcuts-X-SS this is done for each target to find the hop-restricted distance to its closest source.
After an edge is added, the Bellman-Ford distances are re-computed (from sources to all nodes in the graph and from targets to all nodes in the reversed graph) and the process is repeated greedily. This algorithm takes time per step. The first term evaluates the benefit of each possible edge (Equation 2); the second term is the pre-computation of single-source hop-restricted shortest-path distances using the Bellman-Ford algorithm.
We compare our Greedy algorithm to several other popular algorithms for predicting missing interactions.
Several strategies have previously been used to validate network-based edge predictions [34], [61]. First, we describe the notion of potential edges, and then we describe four validation techniques using these edges.
The STRING database aggregates protein-protein associations from over a dozen other pathway and protein interaction databases and combines these with computational predictions based on sequence, co-expression, literature mining, interactions between orthologous proteins, and other biological features to provide a comprehensive protein relationship resource [50]. Only a small subset of these relationships, however, represent physical binding interactions. The remainder, which we term potential edges, are composed of other types of experimentally- or computationally-derived non-physical associations. STRING assigns edge weights for both types of edges (physical and potential) based on biological and computational evidence supporting the link. One benefit of the STRING weighting scheme is that weights for both the physical and potential edges are computed in the same manner and thus are directly comparable. Edges supported by multiple types of evidence have higher weights [62]. Our predictions are based solely on the network topology and source-target connectivity — they do not rely on sequence, gene expression, or any of the other data types — and are therefore completely independent of the STRING predictions.
Starting from only the STRING physical interactions, one way to test our predicted edges is to count how many of them exist within the set of STRING potential edges. The STRING potential network contains 659,719 of the approximately 20 million possible interactions (3.5%), hence identifying the correct interactions is still very challenging.
Although identifying STRING potential edges is useful, these predictions may not bear any relevance to the HOG pathway from which the sources and targets are derived. Our second validation approach considers a prediction as correct if it exists within the STRING potential edges and it connects two proteins from the set of sources, targets, and other known HOG pathway members [46], [53]; otherwise it is incorrect. KEGG and the Science Signaling Database of Cell Signaling provide an unbiased set of pathway members that are not dependent on our own subjective curation efforts. Although these pathway databases omit some HOG members reported in recent literature (e.g. the upstream proteins in de Nadal and Posas [54]) and other uncharacterized proteins that partake in the osmotic stress response, the proteins and interactions they do contain are provided by pathway experts and are thus trustworthy. Therefore this test serves as a strong proxy for each method's ability to make high quality and pathway-relevant predictions.
Our third test measures the quality of an edge prediction based on how much its addition reduces the objective function cost. This approach directly quantifies the method's ability to reduce the distance between sources and targets.
Finally, as a fourth test, we conducted the following cross-validation experiment: We started with the unoriented STRING PPI network and identified all the edges connected to at least one HOG-relevant node (there were 1079 such edges). Because our algorithm specifically predicts edges that lie between sources and targets, these HOG-related edges were used as the cross-validation set. We performed 5-fold cross-validation for the Greedy algorithm using the Shortcuts and Shortcuts-X objective functions and counted how many of the top 10 predictions exactly recovered a left-out edge. The probability that a random prediction would recover a left-out edge from amongst all the potential edges is extremely small (0.033%), and thus this test is also very challenging. It is also challenging because it is difficult to decouple training and test sets of edges. Leaving out even a very small number of edges may result in an entirely different pathway structure in which alternative paths may emerge as more likely. This is especially prevalent on small scales: for example, if edges exist and the edge is left-out, then it is entirely reasonable to predict edge as a shortcut of the path chain. More generally, any chain can be shortcutted by directly connecting the ends (which may often be hubs through which the paths diverge), and single-use edges that play a peripheral role in the pathway may be bypassed altogether.
To summarize, we consider four approaches to validate edge predictions. The first test compares the prediction accuracy of each method in identifying STRING potential edges. The second test compares the prediction accuracy of each method when predicting STRING potential edges that are also relevant to the HOG pathway. The third compares each method's ability to reduce the objective function cost. And the fourth measures the cross-validation accuracy of the Greedy algorithm.
We started with sets of HOG pathway sources and targets and an undirected, weighted PPI network for S. cerevisiae from STRING composed of only physical binding edges (Table 1). We oriented the network [13] and used the three source-target-based algorithms (Greedy, Betweenness, Direct-ST) and two global algorithms (Jaccard, Short-Path) to predict directed edges in this network using the relevant objective functions (Shortcuts, Shortcuts-X, Shortcuts-SS, Shortcuts-X-SS). We evaluated each method with respect to its ability to: 1) reduce the objective function cost; 2) predict edges that lie within the STRING potential edges; and 3) predict edges that lie within the STRING potential edges that also connect known HOG-related nodes. For the Greedy method, we also performed cross-validation experiments.
Our Greedy algorithm achieves the greatest cost reduction compared to the other four methods over all variants of the pathway-aware edge prediction problems (Figure 2). Moreover, Greedy substantially decreased source-target distances after adding only a few edges. For example, after adding 3 edges, the Shortcuts cost (measured as the total shortest-path distance amongst source-target paths) can be reduced to approximately 60% of the original cost. In contrast, it takes 10 edges for Direct-ST to achieve the same ratio. The Betweenness algorithm does monotonically decrease the cost, however, because edges are added based on greater usage (as opposed to greater explicit cost reduction), its reduction is much slower than Greedy overall. The global methods (Jaccard and Short-Path) do not leverage the sources and targets and therefore are unable to reduce source-target distances at all; in general, there are an enormous number of possible edges that play no putative role in the pathway and it is difficult for these methods to disambiguate these edges from HOG-relevant edges. The tremendous cost reduction seen with the Greedy predictions implies that there are a few missing edges in the network whose addition may cover a large bulk of the information flow in the network.
For Shortcuts-SS and Shortcuts-X-SS, both Greedy and Direct-ST perform equally well. This is because there are only 11 paths to optimize over instead of 55 (each target to a single source). Thus, a viable strategy is to find the target that is furthest away from any source and connect a source directly to it. This can greatly reduce the cost function, even if no other path uses this edge, though this need not be the case in general.
Next, we judged the quality of the predictions based on how well they overlapped with the STRING potential edges and with HOG-relevant proteins (Figure 3). In these tests, the accuracy of the method is the percentage of predicted edges, made from amongst all possible non-existent edges, that lied in the relevant set.
When only considering support in STRING (Figure 3A), we find that the global methods (Jaccard and Short-Path) significantly outperform the source-target-based methods. In particular, every prediction made by the Jaccard algorithm is correct according to STRING as are over 60% of the Short-Path predictions. This result agrees with previous studies that showed that network distance and shared topology are strong indicators for functional or physical relatedness [33], [35], [37], [57]–[59]. The probability of predicting a STRING potential edge from amongst all possible edges is only 3.5%, and thus most approaches perform significantly better than baseline.
This test, however, does not tell us whether the predictions bear any relevance to the HOG pathway, which is the primary focus of this study. To better home-in on HOG-relevant predictions, we filtered the STRING potential edges to only include those edges that connected two known HOG-related proteins. Figure 3B shows that the global methods do not make any predictions that relate to the HOG pathway. On the other hand, the Greedy predictions remain at the same level in both tests, which implies that its predictions tend to be highly accurate and lie amongst HOG-related nodes. The difference is especially pronounced in the hop-restricted cases, where Greedy is more accurate than any other method by roughly 40% (Shortcuts-X). Two of these edges connect Hog1 to known HOG transcription factors, Msn4 and Cin5 — both previously established interactions in KEGG [46] or the literature [63] (which are missing from the STRING database and thus do not appear in the original network we used). The probability of predicting a HOG-relevant STRING potential edge from amongst all possible edges is only 0.076%, which is much lower than the accuracy of all three source-target-based algorithms.
Of the top 15 predictions made by Greedy and Betweenness for the Shortcuts-X problem, only one prediction overlaps, and a similar trend holds for the other objectives. This likely stems from the fact that Greedy takes the magnitude of the cost reduction into account, whereas Betweenness only computes the number of shortest paths that use the candidate edge. Because both algorithms perform significantly better than baseline, this implies that they may provide complementary predictions and both may be reasonable depending on the use case.
Interestingly, despite their similar performance in cost reduction for Shortcuts-SS and Shortcuts-X-SS (Figure 2), Greedy makes more accurate predictions than Direct-ST (Figure 3). This is because there are many cases where a direct source-to-target prediction can be equivalently replaced by a target-target interaction. For example, if was added in the first step, the predictions and (regulated via ) both equally reduce the cost from a single source () to the target . However, target-target interactions are more likely to exist within the STRING potential edges than direct source-target edges, and indeed Greedy makes several TF-TF predictions (e.g. ), thereby giving it an advantage.
To show that the orientation step is indeed useful in extracting HOG paths given sources and targets, we ran each algorithm on the unoriented STRING PPI network (Figure S2). We found that for both hop-restricted objective functions, the Greedy algorithm makes more HOG-relevant predictions when using the oriented network (53% vs. 46% for Shortcuts-X and 40% vs. 20% for Shortcuts-X-SS, compared to using the unoriented network). Moreover, the global methods (Short-Path and Jaccard) also benefited significantly from the orientation, which implies that defining network neighbors more precisely can help in identifying putative interactions.
Overall, these results show that the global methods perform well in identifying putative interactions, but that the Greedy algorithm can home-in on more pathway-consistent interactions while drastically reducing source-target distances.
While predicting plausible edges from amongst all possible edges serves as a strong validation technique, in practice, we would also like to leverage other data sources (such as expression, sequence, and literature evidence) when making predictions. To naturally integrate these sources into our framework, instead of predicting from amongst all possible edges, we only predict from amongst the set of STRING potential edges (Methods). Each potential edge is weighted by STRING with a confidence value in , which we explicitly set to (Equations 1 and 2; in the previous sections, was given a default weight of 0). By using these data types and weights together, we can pinpoint putative interactions that have evidence from a wide variety of biological sources as well as evidence from the network.
Table 2 presents the top 10 predictions made by the Greedy algorithm for the Shortcuts objective function, many of which are known physical interactions missing from STRING. The and predictions have direct evidence of physical interaction according to BiOGRID [64], but were not present in the STRING network. The and predictions lied within the STRING binding edges (and thus represent physical interactions), but were either oriented in the opposite direction or were left out of the oriented network. was originally oriented , but the Greedy algorithm suggests that that this edge was either oriented incorrectly or is bidirectional. was left out of the network because the orientation algorithm did not find any length-bounded paths that included this edge. Although in general biological pathways are short, this prediction exemplifies an exception where considering longer pathways through the edge improves the source-target connectivity. These correct predictions demonstrate that our approach can correct for limitations of the edge orientation.
For the following three predictions, we verified both the physical interaction between the two nodes and the directionality (which is not possible for edges validated with the undirected STRING or BioGRID databases). The prediction () involves two general stress TFs that play a substantial role in the HOG pathway [51]. Harbison et al. [65] showed that indeed Msn4 binds the MSN2 gene in the succinic acid stress condition. This study did not profile Msn4 DNA binding in osmotic stress, but it is plausible that this stress-activated TF could bind MSN2 in other conditions as well. The prediction () was recently shown by Pokholok et al. [63] to occur in osmotic stress. We discuss the prediction () at length in the next section.
Overall, 7 of the top 10 predictions have support for direct physical binding in the cell. In addition, the prediction was not directly supported in the literature but warrants further study. Both Reg1 and Msn4 have been shown to physically associate with the 14-3-3 proteins Bmh1 and Bmh2 [66] but have not yet been shown to directly interact with one another. Proteins with a common physical interaction partner may be more likely to directly interact themselves than proteins with other types of functional connections (e.g. genetic interactions) [33], [35], [57].
Table 3 presents the top 10 predictions made by the Greedy algorithm for the Shortcuts-X objective function, which attempts to model more biological constraints by imposing a hop-restriction on the source-target paths. Remarkably, the top three predictions (, , and ) represent best-case predictions: The two genes/proteins involved are known to physically interact, the directionality is correct, and the interaction is highly relevant to osmotic stress response. In particular, and are core HOG pathway interactions that are well-characterized [51] and appear in KEGG [46], but lack evidence for physical binding in STRING. The MAPK Hog1 is central to the HOG response program, and its activation of downstream TFs is a critical component of the response. The other two validated predictions involve HOG pathway members as well. Sho1 is a transmembrane osmosensor, and its branch of activation of Hog1 is known to be mediated by interaction with Cdc42 [67]. The interaction is also present as part of the related starvation subpathway of MAPK in KEGG [46]. Finally, the prediction () is between two members of the Sho1 HOG pathway input branch [53]. Overall, of the 659,719 STRING potential edges considered, only 0.0011% are in KEGG, and thus the fact that 3 of the top 10 predicted edges lie in KEGG is highly significant (, Fisher's exact test).
Other predictions whose physical interaction could not be validated also involve pairs of HOG pathway members. Some predictions occur between the two independent upstream input branches in the pathway (e.g. and ) or between upstream proteins and proteins that are very far downstream (e.g. ). From an algorithmic standpoint, these edges do indeed provide faster diffusion of signal from sources to targets; however, they may not represent direct interactions that occur in the cell. In contrast, the prediction is a shortcut within the Sho1 input branch, which contains the cascade [54]. Note that several of these predicted edges have very high weights (e.g. ) from STRING reflecting their strong functional dependencies, which makes them more likely to be selected by our algorithm. However, several predictions were made despite lower evidence (e.g. ), which suggests that their addition strongly aided source-target connectivity. Interestingly, none of the top 10 predictions directly connects a source to a target. This further necessitates an approach like ours versus Direct-ST.
To further validate our ability to extract accurate pathway-relevant predictions from within the potential set, we conducted 5-fold cross-validation experiments by leaving out HOG-relevant edges (see Methods). The probability that a random prediction would recover a left-out edge from amongst all the potential edges is extremely small (0.033%). Using the Greedy algorithm, we found that 12% (16%) of the top 10 predictions for Shortcuts (Shortcuts-X) recovered a left-out edge. Recovering one correct edge (10%) yields a P-value of and recovering two correct edges (20%) yields a P-value of (Fisher's exact test). Both values are significant (our results lie between them) further supporting the ability of our method to make accurate edge predictions.
To explore the sensitivity of our results to the hop-restriction length, we repeated our computational experiments using a hop-restriction length of . Overall, we found similar qualitative performance for the algorithms when predicting from amongst all possible edges (Figure S3). However, when predicting from amongst the potential set, we found only a few overlapping predictions with those made when the hop length was 5. Interestingly, these included the well-known HOG interactions , and , suggesting that the most confident and likely predictions are not wholly affected by the decreased hop restriction. Of course, some different predictions are also to be expected; for example, using a hop length of 4, the algorithm makes predictions for and . While these predictions make sense algorithmically, they do not make sense biologically because they attempt to shortcut the sources of the pathway directly to a core node (Hog1). This suggests that 4 hops may be too restrictive and may motivate using a hop restriction of 5 in future efforts.
We also found that our approach was able to recover missing interactions when not leveraging the STRING-derived weights (see Text S1). This implies that our approach is not entirely dependent on the potential edge weights and that our objectives are well-defined.
To demonstrate our approach's ability to make novel, biologically meaningful predictions we selected for experimental validation. This was a top prediction for two objective functions (for Shortcuts-SS it was the prediction and for Shortcuts it was the uncharacterized prediction; Table 2). As we showed, the addition of a few edges can greatly reduce the objective function cost, and therefore we place more confidence in these top edges.
Verifying a directed protein-protein interaction at the mechanistic level requires extensive experimentation and is beyond the scope of this work. However, genetic experiments such as gene deletions can establish condition-specific causal relationships between proteins in signaling pathways. For instance, loss-of-function mutations and gene over-expression were used to identify and order the genes along the apoptosis pathway in C. elegans [68]. In our case, if Tpk2 controls the TF Sok2 in osmotic stress, TPK2 deletion should affect Sok2's regulatory activity in this condition. Because many interactions along signaling pathways occur post-translationally, we would not expect the SOK2 gene to be differentially expressed in the mutant even if Tpk2 does activate or inhibit Sok2 at the protein level. Instead we determine the degree to which the deletion alters Sok2's function as a transcriptional regulator. As predicted, the knockout significantly affected genes bound by Sok2 (, Fisher's exact test; see Supporting Text S1 for microarray details and Table S1 for lists of affected genes). The knockout alone cannot confirm whether the interaction is direct or indirect, but clearly establishes that there is a functional connection between these proteins that is active in osmotic stress. Moreover, the orientation of the predicted edge is correct because if Sok2 were upstream of Tpk2 in the pathway, its bound genes would be unaffected by TPK2 deletion.
To test the significance of our knockout (KO) with other perturbation experiments, we used the Rosetta compendium [69] of 300 KO expression experiments and compared the overlap between differentially expressed (DE) genes in each experiment with the list of Sok2 targets (see Supporting Text S1). Of 301 experiments, only 31 (10.3%) had a lower P-value than the one obtained from our TPK2 KO. In the other direction, we considered 117 additional TFs for which a high confidence set of targets exists [70]. For each, we computed the significance of the intersection between their targets and genes affected by the TPK2 deletion using Fisher's exact test. Similar as the test above, of the 118 tests only 14 (11.9%) had a lower P-value than our predicted Tpk2-Sok2 pair. Combined, our predicted interaction ranked close to the top 10% in these two independent analyses further supporting our prediction.
Protein interaction networks encode a variety of signaling processes that occur in the cell, however, many interactions are still missing and experimental validation of all putative interactions is unlikely in the near future. This has led to a proliferation of computational methods to aid in identifying putative interactions. One particularly important task when mining these networks is to identify pathways. Experimental protocols have made it possible to identify upstream proteins that trigger information cascades to downstream transcription factors. Many techniques have been proposed to extract likely subnetworks from within global interaction networks, however, these approaches do not formally model interactions that are missing from the network.
We presented a new framework for predicting missing edges that lie “in-between” given sets of sources and targets within the network. Compared to four other edge prediction algorithms, our Greedy algorithm was able to home-in on more pathway-consistent interactions while substantially reducing source-target distances by only adding a few edges. We also showed how to naturally integrate other biological features into the pipeline and used this evidence to recapitulate many known but missing physical interactions, including several interactions reported in KEGG and other databases and reports.
Our ability to correctly predict context-specific directed PPIs by reducing source-target distances with the Greedy algorithm yields high-level biological insights into signaling network topology. In many cases the endpoints of a predicted edge are already connected via a longer alternate pathway. Shortcut edges between connected proteins form alternate paths for signal flow, which may lead to a greater degree of robustness in the pathway. In addition, such edges may indicate that the two proteins are participating in a feed-forward loop. The feed-forward loop motif can provide precise control of activity timing and noise filtering [71] so recognizing that a pair of proteins belong to a feed-forward loop instead of a linear chain improves our understanding of their role in the signaling pathway. Our objective functions encourage adding edges that reduce the distance between multiple source-target pairs, and indeed, we find that the first few predictions (those that improve the objective function the most) when using the Shortcuts or Shortcuts-X objective benefit many such pairs. For Shortcuts, the first 3 added edges decrease the distance of 27 of the 55 source-target pairs (49.1%). Likewise, the first 3 Shortcuts-X predictions reduce the distance for 18 pairs (32.7%). These first few predictions are also highly accurate (Tables 2 and 3), indicating that edge-reuse is an important principle in signaling networks.
In general, the predictions varied as more constraints were added to the objective function: with respect to Shortcuts, 50% of the top 10 predictions overlapped with Shortcuts-SS and only 20% with Shortcuts-X and Shortcuts-X-SS. Initially, without any hop-restriction, the average number of hops to connect a source and target is 7.8 (with total distance 12.91). When applying the 5-hop-restriction (as in Shortcuts-X), alternative edges are forcibly used that have lower confidence, and thus the total distance increases to 18.24. The hop-restricted objectives thus lead to a restructuring of the source-target paths and tend to select central nodes through which much signal flows (e.g. Hog1). The non-hop-restricted algorithms may induce alternative longer paths that circumvent these hubs. This implies that there is a trade-off between the likelihood of a series of interactions (the weights along the path) and the efficiency of the source-target cascade (the number of hops along the path). The former is characterized by the Shortcuts objective, while the latter is captured by Shortcuts-X. While evidence exists supporting predictions from both objectives, the hop-restricted versions found more predictions that were actually in the KEGG HOG pathway (3 versus 0) and that connected two known HOG pathway members (8 versus 3; compare Tables 2 and 3). This suggests that Shortcuts-X predictions may have greater fidelity with the condition-specific pathway (which is our focus here). On the other hand, Shortcuts made more predictions whose physical binding could be verified than Shortcuts-X (7 versus 5), which suggests that this objective may be capturing more general interactions that aid overall network connectivity.
Our knockout experiment examines the predicted relationships between Tpk2 and the target TF Sok2 in hyperosmotic stress conditions. Tpk1, Tpk2, and Tpk3 form the catalytic subunit of protein kinase A (PKA), the complex at the heart of the Ras/cAMP/PKA signaling pathway [72]. Through interactions with its many substrates, PKA is involved in general stress response, metabolism, growth, ribosome biogenesis, and various other biological processes [72], including osmotic stress response. PKA's involvement in the osmotic stress response is parallel to the HOG pathway [73]. Msn2, Msn4, and Sko1, which along with Hot1 are considered to be the primary HOG pathway TFs [51], are each affected by PKA in osmotic stress [73], [74]. Decreased PKA activity modulates the repressive effects of Sko1 in this condition. This behavior is complementary to Hog1's phosphorylation of Sko1, which also alleviates Sko1 repression of its target genes [73]. While Tpk2's role in osmotic stress is well-established, Sok2 is not considered to be a core HOG pathway TF, but was rather assumed to be controlled by the primary TFs [52]. However, genetic screens illustrate that its role in the osmotic stress response may be larger [75], [76].
Our TPK2 knockout establishes a functional link between Tpk2 and Sok2 in which Sok2 is downstream of Tpk2. A previous genetic interaction reported by Ward et al., who suggested that PKA may directly phophorylate Sok2, supports this directionality and relationship [77]. Subsequent experiments confirmed that active PKA phosphorylates Sok2 when glucose is the carbon source [78]. However, this link does not appear in other conditions. For example, Sok2 was found to function in a pathway parallel to PKA [79] and Tpk2 [80] in pseudohyphal growth and adhesive growth, respectively. In addition, Tpk2 does not interact with Sok2 in a mutant yeast strain that is sensitive to exogenous cAMP [81]. These findings highlight the importance of pathway-specific predictions of missing interactions as opposed to general protein interaction predictions.
Our results showing that Tpk2 functionally affects Sok2 in osmotic stress coupled with previous evidence that the Sok2 sequence contains a consensus PKA phosphorylation site at amino acids 595 to 598 [7], [78] and that PKA phosphorylates Sok2 in other conditions, suggests that the predicted interaction warrants direct experimental validation. Despite their high sequence similarity, the three Tpk's have distinct sets of substrates [82] so confirmatory future work must specifically examine Tpk2 phosphorylation. Because in vivo verification of a kinase-substrate interaction is challenging, the next step experimentally will be to show that Tpk2 phosphorylates Sok2 in osmotic stress in vitro. Peptide arrays and kinase assays have been used to validate computational phosphorylation predictions in vitro [83]. Proteome chips did not detect Sok2 as a Tpk2 substrate in vitro [82], highlighting the need for osmotic stress-specific experiments in order to validate our condition-specific prediction. Following in vitro confirmation any number of in vivo strategies could be used to decisively validate the interaction (see Morandell et al. [84] for a review). For instance, electrophoretic mobility shifts in kinase deletion strains can provide in vivo evidence of phosphorylation and validate in vitro interactions [82], [83].
Our analysis comparing the set of Sok2 targets and affected TPK2 knockout (KO) genes with other binding and KO experiments indicated that the overlap between these two sets lies close to the top 10% in both tests. It is not surprising that the deletion of other genes also leads to the differential expression of some Sok2 targets, but the fact that this occurs for only a fraction of experiments suggests that our KO holds against the statistical background. Further, of the 31 KOs with a higher overlap, none correspond to protein products that directly bind to Sok2 according to STRING. As for the overlap between the other TF targets and our TPK2 KO set, again, it is not surprising that other TFs were affected by the KO because deletions can affect both direct binding partners and proteins further downstream. The more significant Tpk2-TF associations do not correspond to direct binding in the interaction network — the average distance in the interaction network is 4.8 edges — which suggests that these are not candidates for missing interactions.
Recently, there has been a great increase in the amount of experimentally derived protein interaction data in several species [85] and in our ability to experimentally query host-environments and host-pathogen interactions [9]. Given these networks, the problem of identifying response pathways can now be tackled in multiple species. A key problem in such studies is dealing with missing interactions, as these prevent algorithms from recovering the correct information flow. The method we presented in this paper is the first to address this issue in a pathway-specific context and can be applied to any species for which such data exists. Further, our method may have use in other domains, for example, in network design where the goal is to reduce routing lags or to aid the flow of information between entities in a network.
|
10.1371/journal.pbio.0060277 | Chromatin- and Transcription-Related Factors Repress Transcription from within Coding Regions throughout the Saccharomyces cerevisiae Genome | Previous studies in Saccharomyces cerevisiae have demonstrated that cryptic promoters within coding regions activate transcription in particular mutants. We have performed a comprehensive analysis of cryptic transcription in order to identify factors that normally repress cryptic promoters, to determine the amount of cryptic transcription genome-wide, and to study the potential for expression of genetic information by cryptic transcription. Our results show that a large number of factors that control chromatin structure and transcription are required to repress cryptic transcription from at least 1,000 locations across the S. cerevisiae genome. Two results suggest that some cryptic transcripts are translated. First, as expected, many cryptic transcripts contain an ATG and an open reading frame of at least 100 codons. Second, several cryptic transcripts are translated into proteins. Furthermore, a subset of cryptic transcripts tested is transiently induced in wild-type cells following a nutritional shift, suggesting a possible physiological role in response to a change in growth conditions. Taken together, our results demonstrate that, during normal growth, the global integrity of gene expression is maintained by a wide range of factors and suggest that, under altered genetic or physiological conditions, the expression of alternative genetic information may occur.
| Recent studies have shown that much more of the eukaryotic genome is transcribed into RNA than previously thought. In Saccharomyces cerevisiae, when particular factors are defective, cryptic promoters within several coding regions become active and produce shorter transcripts corresponding to the 3′ portions of genes. (Transcription proceeds from the 5′ end of genes to the 3′ end.) A comprehensive analysis of cryptic transcription identified the factors that normally repress this event. We find that at least 50 factors, many involved in chromatin structure and transcription, are required to repress cryptic transcription. Other results suggest that the potential for cryptic transcription is widespread, initiating from at least 1,000 locations across the S. cerevisiae genome. In mutants in which cryptic transcripts are produced, some of the transcripts are translated into proteins not normally made in unmodified, wild-type cells. Finally, in wild-type cells, a subset of cryptic transcripts is transiently induced following a nutritional shift, suggesting a possible role for cryptic transcription. Taken together, our results demonstrate that the normal pattern of gene expression is maintained by a wide range of factors and suggest that, under altered genetic or physiological conditions, the expression of alternative genetic information may occur.
| Several recent studies have demonstrated that transcription occurs across large eukaryotic genomes in a much more widespread and complex pattern than previously imagined. The recent findings of the ENCODE project, which analyzed transcription of 1% of the human genome [1], demonstrated the use of multiple transcription start sites and transcription across most sequences, including intergenic regions (reviewed in [2]). Many other recent studies have also identified extensive transcription across human sequences, including antisense transcription (reviewed in [3–5]). Similarly, in Drosophila melanogaster, recent studies estimate that 85% of the genome is transcribed, with extensive intergenic transcription and multiple transcription start sites [6]. Although the function of most of this pervasive transcription is currently not understood, there is evidence that a significant amount of it is regulated, raising the possibility that it is required for previously unknown modes of regulation or that it allows the expression of previously undetected genetic information [3–5]. Strong precedents exist for regulatory roles for intergenic transcription (for example, [7,8]; see [4,9] for recent reviews).
In Saccharomyces cerevisiae, similar to larger eukaryotes, several recent genome-wide studies have demonstrated widespread transcription across coding and noncoding regions [10–15]. In a small number of cases in S. cerevisiae, intergenic transcription [16–18], antisense transcription [19,20], and initiation within coding regions [21,22] have been shown to play biological roles. In addition to transcriptional events that occur in wild-type strains, other studies have revealed that transcription initiation can be activated from within coding regions in particular mutants [23,24]. Such initiation was originally observed in strains containing mutations in SPT6 and SPT16, which encode conserved, essential transcription factors believed to be involved in nucleosome disassembly and assembly [23–27]. In an spt6 mutant, the use of a transcription start site within the FLO8 gene was shown to be dependent upon a consensus TATA element within the FLO8 coding sequence, suggesting the existence of a cryptic promoter within FLO8 that is normally repressed in a wild-type strain but becomes activated in an spt6 mutant [23]. Evidence suggested that in spt6 mutants, the failure to reassemble nucleosomes in the wake of elongating RNA polymerase II (RNAPII) allowed transcription initiation factors to bind to and activate cryptic promoters [23].
Several transcription factors are required to repress cryptic promoters in S. cerevisiae. An early study revealed that several different mutants allow cryptic initiation [23]. Subsequent analysis has suggested that the level of histone modifications in coding regions, as regulated by the Set2 histone methyltransferase and the Rpd3S histone deacetylase complex, also controls cryptic initiation [28–30] and that set2Δ mutations allow cryptic initiation in a large set of genes [31]. Additional work has identified other mutants that allow cryptic initiation, including asf1 and ctk1, [32,33], as well as particular combinations of double mutants, revealing roles for other elongation factors, including the Paf1 complex, Bur1-Bur2, the HIR complex, Spt2, and Elf1 [34–36]. These studies suggest that the repression of cryptic promoters requires a variety of factors that play roles in transcription elongation and chromatin structure. These factors appear to be entirely distinct from those that suppress cryptic intergenic transcripts [37].
In this paper, we present the results of genome-wide approaches to comprehensively study cryptic transcription from within open reading frames (ORFs) in S. cerevisiae. First, we used both spontaneous mutant selection and a synthetic genetic array (SGA) screen to identify new mutations that allow cryptic transcription. These mutations have varying effects on the expression of a set of cryptic transcripts, suggesting the existence of different classes of cryptic promoters and mechanisms for their activation. Second, we used microarray analysis to identify cryptic transcripts throughout the S. cerevisiae genome that are activated in spt6 and spt16 mutants. These experiments showed that cryptic transcription is widespread, occurring in at least 1,000 genes (17% of all genes). We have also investigated the possibility of a physiological role for cryptic transcription, as it is not understood whether it represents unwanted transcription from fortuitous promoters that are activated only in mutants in which chromatin structure has been altered, or whether it serves a biological role in some cases, possibly to express different gene products. Here, we demonstrate that a number of cryptic transcripts expressed in an spt6 mutant are translated into corresponding short proteins. In addition, we show that some cryptic transcripts are modestly activated in wild-type (SPT6+) strains upon a nutritional shift and that this activation is dependent upon Ras2. Taken together, our results show that cryptic transcription from ORFs can occur in a widespread fashion throughout the S. cerevisiae genome and suggest that some cryptic promoters may normally serve to express alternative genetic information during environmental changes.
Previous results have shown that cryptic promoters are active in several mutants that impair transcription and chromatin structure. However, no systematic isolation of cryptic initiation mutants has been performed. To comprehensively identify factors that regulate cryptic promoters, we first constructed a reporter to allow easy detection of activation of the FLO8 cryptic promoter. In this reporter, we replaced the region of FLO8 3′ of the cryptic transcription start site with the HIS3 coding sequence (Figure 1A; Materials and Methods). The HIS3 coding sequence was inserted out-of-frame with respect to the FLO8 coding sequence, using the first ATG within FLO8 that follows the cryptic start site. As this ATG is in the +2 reading frame, functional HIS3 mRNA can only be made by transcription initiation at the FLO8 cryptic start site (Figure 1A). In one version of this reporter, the normal FLO8 promoter was replaced with the GAL1 promoter to allow regulation of full-length FLO8-HIS3 transcription by growth on different carbon sources and in a second version, the wild-type FLO8 promoter was maintained. Both growth assays on plates lacking histidine and northern analysis demonstrated that the FLO8-HIS3 fusion constitutes a sensitive reporter for mutants that allow cryptic initiation (Figure 1B and 1C).
Using FLO8-HIS3, we employed two methods to identify mutants that are permissive for cryptic initiation: direct selection and a screen of the S. cerevisiae nonessential deletion set (Materials and Methods). Direct selection was valuable for identification of strong mutations that are not in the deletion set, in particular, mutations in histone genes, described below. The deletion set screen allowed systematic testing of all nonessential genes. Overall, we identified mutations in 50 genes that allow cryptic initiation at FLO8-HIS3 (Table 1). These 50 mutants are permissive for the FLO8 cryptic promoter to varying degrees and several are dependent upon expression from the upstream GAL1 promoter in the FLO8-HIS3 reporter (Figure 2A). Overall, the majority of genes identified encode histones, regulators of histone gene expression, histone chaperones, and other factors implicated in transcriptional control.
Among this large collection of mutants, histone H3 mutants are of particular interest as some identify previously unstudied changes in H3 that may play roles in transcription elongation. These H3 mutants are likely gain of function mutants, as deletion of either HHT1 or HHT2, the genes encoding histone H3, causes only a very weak His+ phenotype with FLO8-HIS3, whereas the H3 mutants isolated by our selection are dominant and confer a strong His+ phenotype (Figure 2A, unpublished data). The majority of these H3 mutants are inviable when the second, wild-type H3 gene is deleted, suggesting that the H3 mutants are incapable of forming a functional nucleosome on their own (Table S1). One class of H3 mutants of interest includes four clustered changes in one region of H3: I51N, I51S, Q55H, and S57P. These changes are of interest due to their proximity to K56 of histone H3, whose acetylation has been shown to be important for resistance to DNA damaging agents, histone gene expression, and transcriptional silencing [38–42]. However, H3 K56 acetylation does not affect cryptic initiation, as an rtt109Δ mutation, which abolishes K56 acetylation [43–46], does not activate the FLO8 cryptic promoter (unpublished data).
To test whether the mutants we identified activate cryptic transcription from multiple genes, we performed northern analysis on 14 cryptic initiation mutants, examining transcription of FLO8, SPB4, and STE11, three genes previously shown to have cryptic promoters [23]. Our results show that there are different patterns of cryptic promoter activation among the mutants (Figure 2B). Most of the mutants express the FLO8 short transcript, with the exceptions of hir1Δ and chd1Δ (Figure 2B, lanes 10 and 14; also see [35,36]), suggesting that in some cases, the FLO8-HIS3 reporter is more sensitive in detecting cryptic initiation than northern analysis. Conversely, an spt16-197 mutant appeared weakly His+ with the FLO8-HIS3 reporter, whereas northern analysis indicated high levels of expression of the FLO8 short transcript (Figure 2A and 2B, lane 3). This effect with spt16-197 may be due to the slow growth of the spt16-197 mutant. In addition, short transcripts could be detected for SPB4 and STE11 for most of the mutants, indicating that cryptic initiation was not specific to FLO8. However, there were differences in the pattern of cryptic transcription among the mutants tested. For example, spt6-1004, eaf3Δ, and rtt106Δ confer distinct patterns of activation of FLO8, SPB4, and STE11 cryptic transcripts (Figure 2B, compare lanes 2, 8, and 13). These distinct patterns suggest that there are distinct classes of cryptic promoters and different mechanisms for their repression. Other evidence suggesting differential expression of cryptic transcripts has recently been described [47].
Recent results have shown that Set2-dependent methylation of histone H3 at K36 plays a role in the repression of cryptic transcription [28,47,29]. Furthermore, both H3 K36 dimethlyation and trimethylation have recently been shown to be defective in spt6 and spt16 mutants, as well as in set2 mutants [28,48]. Therefore, we tested whether this histone H3 K36 methylation defect might be a common phenotype among cryptic transcription mutants. Our results show that, of 50 mutants tested, only five showed a significant decrease in total H3 K36 di- and trimethylation (spt6-1004, set2Δ, ctk1Δ, ctk2Δ, and ctk3Δ) (Figure 3, Table S2). The histone H3 K36 methylation defects in these five mutants have been previously reported [28,33,47–49]. We note that under our growth conditions, the spt16-197 mutant had wild-type levels of H3 K36 di- and trimethylation, in contrast to a previous report [48], yet still showed a high level of cryptic transcription. These results show that the majority of the cryptic transcription mutants regulate at a step other than H3 K36 methylation.
Previous studies of cryptic initiation in an spt6-1004 mutant identified only a few genes with cryptic promoters [23]. However, the frequency at which they were found among a small set of genes tested suggested that cryptic promoters may be widespread. To test this possibility, we assayed for cryptic transcription within ORFs on a genome-wide scale by microarray analysis. In these experiments, we compared mRNA from a wild-type strain to that from an spt6-1004 mutant, using microarrays with six probes across each coding region (Materials and Methods). Using a stringent threshold (Materials and Methods), our results suggest that out of the 5,689 ORFs represented on the microarray, at least 960 genes (17%) have active cryptic transcription in the spt6-1004 mutant (Figure S1; Table S3). As detailed in Materials and Methods, this method may unavoidably be biased towards identifying cryptic transcripts from genes with lower transcript levels, likely resulting in an underestimate of the actual number of cryptic transcripts (Materials and Methods; Figure S2). In support of the ability of the microarrays to identify genes with cryptic transcripts, we used northern analysis to test five genes predicted by the microarrays to have cryptic transcripts and found that all five indeed produce short transcripts (Figure 4A).
To test whether another mutant permissive for cryptic transcription allows production of the same large set of cryptic transcripts, microarray analysis was performed on the temperature-sensitive spt16 mutant, spt16-197. These experiments identified approximately 1,130 genes predicted to have cryptic transcripts in the spt16-197 mutant (Table S4). Between the spt6-1004 and spt16-197 results, there is a striking overlap (correlation coefficient r = 0.83, Figure 4B and 4C), indicating that these two mutants affect cryptic transcription similarly at most genes. Taken together, these results strongly suggest that approximately one sixth of all S. cerevisiae genes produce detectable cryptic transcripts in spt6 and spt16 mutants.
To determine whether the genes that produce cryptic transcripts share any particular traits, we examined several different characteristics of the genes that we identified in the spt6 and spt16 microarray experiments as having cryptic transcripts. With respect to the length of coding regions, the average length of the genes with cryptic transcription in both spt6 and spt16 mutants is 2.4 kb, significantly longer than the average length of the 5,869 genes on the microarray (1.5 kb; Wilcoxon rank-sum test, p-value < 2.2 × 10−16). The majority of genes with cryptic transcription also have lower transcriptional frequencies (for spt6-1004, average = 2.46 mRNA/hour [p-value < 2.2 × 10−16] and for spt16-197, 1.93 mRNA/hour [p-value , 2.2 × 10−16]) when compared with the whole genome (average = 7.57 mRNA/hour) [50]. The enrichment for longer genes with lower transcriptional frequencies was expected, as these two characteristics correlate, and our method for detection of cryptic transcripts enriched for genes with lower transcription levels.
In addition, we focused on TATA elements, both within coding regions and in 5′ noncoding regions. Since cryptic initiation within the FLO8 coding region depends on the presence of a TATA element [23], we first tested whether genes showing cryptic transcription are enriched for those with TATA motifs in their coding sequence. We searched for the TATA consensus sequence in S. cerevisiae, TATA(A/T)(A/T)A(A/T)(A/G) [51]. We found that genes with at least one TATA element in their coding region are three times more likely to have a cryptic transcript in the spt6-1004 mutant than genes without a TATA box (p-value < 2.2 × 10−16). We see an even stronger enrichment for the spt16-197 mutant (p-value < 2.2 × 10−16) (Table S5). Given that our set of genes was enriched for those that are longer, we also examined whether these findings were still significant when corrected for gene size (longer genes are more likely to contain TATA motifs by chance) and found that they were (Figure S3). Thus, the genes with cryptic promoters identified by the spt6-1004 and spt16-197 microarray results suggest that cryptic transcription tends to be located in coding regions that contain TATA consensus sequences. We also classified the normal promoters of genes with cryptic promoters as to whether they have a TATA element or not. Genes with TATA elements tend to display more cell-to-cell and strain-to-strain variation in expression [52–57]. We found that cryptic transcripts are two times (Fisher exact test, p-value = 2 × 10−12) and 2.4 times (Fisher exact test, p-value = 2 × 10−16) more likely to be from genes with natural TATA-less promoters than from genes with TATA-containing promoters for the spt6-1004 and spt16-197 mutants, respectively, after correction for gene expression levels (Figure S4).
Given the large number of cryptic transcripts, it seemed likely that many of them would have the potential to encode proteins. We examined the potential for cryptic transcripts to be translated by mapping all ATGs in the three reading frames downstream of the 5′-most limit of transcription initiation established in the spt6-1004 microarray analysis. For each of those ATGs, we mapped the first stop codons in the same frame to infer the peptide sequence that would result from translation from the internal ORFs. The results of this analysis (Table S6) show that most ORFs could encode proteins if the cryptic transcripts were to be translated: 820, 825, and 731 ORFs in frames +1, +2, and +3, respectively. However, the two alternative reading frames primarily encode short peptides, while, as expected, the +1 frame encodes much longer sequences.
To test directly whether genes with cryptic transcripts express proteins, we screened 146 genes that are predicted to have cryptic promoters and that have at least one internal ATG codon in the +1 frame located 3′ of the predicted cryptic start site. To screen these strains, we used the tandem affinity purification (TAP)-tagged set of S. cerevisiae strains in which each ORF is fused at its 3′ end to a sequence encoding the TAP epitope tag [58]. The TAP-tagged strains corresponding to the 146 selected genes were crossed to an spt6-1004 strain to obtain TAP-tagged versions in both SPT6+ and spt6-1004 backgrounds. These strains were then screened by western analyses using an antibody recognizing the TAP tag to determine whether any altered proteins are made in the spt6-1004 strains. We note that this method will only detect proteins produced by translation in the same reading frame as the full-length protein, because it requires that the TAP epitope tag be expressed. Our results show that 20 of the 146 genes tested produced a detectable shorter protein in the spt6-1004 mutant but not in the SPT6+ strain (Table S7, examples shown in Figure 5A). The short proteins were all in the size range predicted by the microarray results, and several of them encode domains with known activities lacking their normal amino-terminal sequences (Figure S5). Northern analysis of these genes verified that corresponding short transcripts of the appropriate sizes were indeed expressed in the spt6-1004 mutant (Figure 4A; unpublished data).
To verify that the short proteins were produced by translation initiation from their corresponding short transcripts and were not simply degradation products of the full-length proteins, we analyzed the expression of short proteins made from two genes, APM2 and PUS4. For each of these genes, we constructed and analyzed mutations that alter the initiation codon for both the normal, full-length protein and for the shorter protein and analyzed each by western analysis. Our results show that mutation of the normal ATG initiation codon eliminated expression of the full-length protein, but had no effect on expression of the short protein expression (Figure 5B, lanes 7 and 15). Furthermore, mutation of the internal ATG specifically abolished expression of the short protein (Figure 5B, lanes 6 and 14). We also observed that this mutation in APM2 resulted in apparent degradation products (Figure 5B, lane 6), perhaps due to the amino acid change in the mutant protein. This mutation in APM2 also causes increased expression of the full-length Apm2 protein specifically in the spt6-1004 mutant and may be due to changes in either mRNA or protein stability. Taken together, these results demonstrate that at least a subset of transcripts expressed from cryptic promoters are translated to produce alternative, shorter proteins. The functions of these proteins are likely to be different from the full-length proteins because they often lose predicted protein domains (Table S8).
The expression of the cryptic transcripts that we have identified is normally repressed in wild-type strains when cells are grown in rich medium. If some of the cryptic transcripts serve a biological function, however, they might be expressed in a wild-type background under particular growth conditions. To screen for such an effect, we used northern analysis to assay the transcription of 16 genes with cryptic transcripts under 20 different growth conditions. Most of these genes were selected from those shown to produce a protein from the cryptic transcript. The conditions tested included starvation for carbon, nitrogen, phosphate, or sulfate, as well as heat shock, high salt concentration, or exposure to different drugs such as 3AT or menadione. Of the 20 different growth conditions tested, one of them, a shift from rich medium (YPD) to minimal medium (SD), caused modest expression of cryptic transcripts in three of the 16 genes tested, CHS6, FLO8, and SPB4 (Figure 6, lanes 3–6). For these genes, cryptic transcripts were detectable by 30 min after the shift, and for two of the genes, CHS6 and FLO8, it was transient, no longer detectable by 2 h after the shift. In all cases, the level of the short transcript was clearly less than observed in the spt6-1004 mutant, indicating that an spt6 mutant represents an extreme condition for cryptic initiation in the genome, relative to what may be seen in a wild-type strain under different growth conditions.
Previous studies have shown that a nutritional shift from rich to minimal media causes other transient effects with very similar kinetics to what we have observed. Among these effects is the induction of translation of the transcription factor Gcn4 [59–61], which occurs in a Ras2-dependent fashion [62]. We therefore tested whether either Gcn4 or Ras2 plays a role in the expression of cryptic transcripts that we observe by assaying gcn4Δ and ras2Δ mutants during a nutritional shift. Although gcn4Δ did not affect cryptic transcript levels (unpublished data), our results showed that the expression of the CHS6 and FLO8 cryptic transcripts upon the nutritional shift was strongly Ras2-dependent, whereas the expression of the SPB4 cryptic transcript appeared to be largely Ras2 independent (Figure 6, lanes 7–10). These results also suggest that the cryptic initiation induced at CHS6 and FLO8 after the nutritional shift is not simply the result of the increased expression of the full-length transcript seen for both genes following the media shift. Even though full-length expression of CHS6 and FLO8 is still greatly increased following the shift in the ras2Δ mutant, cryptic transcripts are not expressed, indicating some form of regulation of the cryptic promoters under these conditions. Thus, our results suggest that a subset of cryptic promoters can be specifically activated upon a nutritional shift in a Ras2-dependent fashion.
In this work, we have investigated cryptic transcription and its consequences in S. cerevisiae on a genome-wide scale. Our results have established that a large number of chromatin- and transcription-related factors are required to repress widespread cryptic transcription from within coding regions throughout the S. cerevisiae genome. Most of the cryptic transcripts contain ORFs, and our results suggest that when these cryptic transcripts are expressed, such as in an spt6 mutant, many of them are translated to produce proteins that are not normally made. Thus, loss of Spt6 causes a dramatic change in the mRNAs and proteins produced genome-wide. Furthermore, a small subset of cryptic transcripts have been shown to be modestly expressed in wild-type strains during a nutritional shift. Taken together, these results demonstrate the widespread existence of cryptic transcription and the expression of alternative genetic information in S. cerevisiae.
Several results strongly suggest that multiple mechanisms control the expression of cryptic transcripts. Below, we discuss these possible mechanisms in terms of distinct classes of cryptic promoters. We note that our microarray results have established widespread cryptic transcription, but have not demonstrated that these transcripts all arise from cryptic promoters. However, based on our earlier studies of the FLO8 and SPB4 genes ([23] and unpublished data), we think it is likely that most or all of the cryptic transcripts identified are the result of activation of cryptic promoters. Testing this possibility will be the focus of future investigations. First, the mutants identified in this study vary greatly in their strength of cryptic initiation, based both on the FLO8-HIS3 reporter and on northern analysis. Second, one of the most permissive mutants for cryptic initiation, spt6-1004, is known to impair at least two features of normal transcription elongation that individually contribute to repression of cryptic promoters: histone H3 K36 methylation [28,47,48] and the recruitment of the transcription factor Spt2 [35]. Consistent with this observation, both set2Δ, which abolishes histone H3 K36 methylation, and spt2Δ are less permissive for cryptic initiation than is spt6-1004 (our results and [28,31,34,35]. In addition to these effects, spt6-1004 likely causes other effects on chromatin structure [23,26]. Third, our results also showed that most mutations that allow cryptic initiation do not impair H3 K36 di- or trimethylation; therefore, loss of this histone modification is not the sole mechanism by which cryptic promoters are derepressed. This conclusion is consistent with recent studies that showed enhanced cryptic initiation in double mutants that lack Set2 and another factor, indicating that mechanisms other than histone H3 K36 methylation play an important role in this regulation [34]. Fourth, previous analysis has identified cases in which mutations that impair distinct aspects of transcription can combine to cause strong effects on cryptic initiation [34–36]. Finally, assay of a small set of cryptic promoters showed that they were activated in distinct patterns among different cryptic initiation mutants. For example, the pattern of cryptic initiation in mutants that impair Rpd3-mediated histone deacetylation was different from cryptic initiation in mutants affecting histone assembly (Figure 2B). Thus, cryptic promoters may be similar to normal promoters in terms of the complexity of regulation by distinct sets of factors, raising the possibility that additional transcription factors may regulate specific subsets of cryptic promoters. Consistent with this idea, our analysis of the FLO8 cryptic promoter has shown that it requires a UAS-like element as well as a TATA element (V. Cheung and F. Winston, unpublished data).
The microarray experiments that we have performed suggest that there are at least 1,000 cryptic promoters in the S. cerevisiae genome that are activated in spt6 or spt16 mutants. The similarity between these two mutants suggests that they serve similar roles in normally repressing cryptic initiation, likely by helping to establish or maintain a repressive chromatin structure across coding regions [23,25]. Another recent set of microarray studies examined cryptic initiation in set2Δ mutants [31] and identified 621 genes with cryptic transcription on the sense strand. That study also identified 494 antisense transcripts, something not measured in our analysis. Similar to our results, the genes identified by Li et al. [31] were enriched for long genes transcribed at low level. Although we would expect that the cryptic promoters activated in set2Δ mutants would be a subset of those found in spt6 and spt16, only 45% of those found in set2Δ were found in spt6. This degree of overlap, while still quite significant, was likely affected, at least in part, by differences in the microarrays and analysis of the datasets. The smaller number of cryptic promoters in set2Δ mutants compared to spt6 and spt16 fits with our results that mechanisms beyond histone H3 K36 methylation control cryptic initiation. The possible role of antisense transcripts is unknown, although recent studies have demonstrated roles in transcriptional regulation [19,20].
Other evidence suggests that promoters within coding regions occur on a wider scale than indicated by our microarrays of spt6 and spt16 mutants. One study, that examined the S. cerevisiae transcriptome in a wild-type strain by serial analysis of gene expression (SAGE), identified 384 genes with transcription start sites located within the 3′ half of the coding region [12]. Only 55 of these 384 genes (14.3%) were identified in our spt6-1004 microarrays to express short transcripts. This small overlap is expected, as our experiments were designed to identify cryptic promoters activated specifically in spt6 mutants. In addition, our spt6-1004 microarrays were designed to detect short transcripts only from the sense strand, while the SAGE analysis was able to detect both sense and antisense short transcripts. More thorough microarray and transcriptome analysis of additional cryptic initiation mutants and other growth conditions will provide a more comprehensive map of cryptic promoters in the yeast genome.
The question still remains as to why so many cryptic promoters are found in the S. cerevisiae genome and what role they serve, if any. We can envision at least four possible roles for cryptic promoters, none of which are mutually exclusive, as all are possible for different subsets. First, some cryptic promoters may direct the expression of gene products that carry out specific functions, being expressed in response to particular environmental changes. In this way, use of cryptic promoters would be analogous to other mechanisms of expressing different genetic information, such as alternative splicing or use of internal ribosome entry sites. Although our results have not demonstrated a function for a product of cryptic initiation, precedent exists for using an internal promoter to express an alternative protein, sometimes under particular growth conditions [22,63–66]. In mammalian cells, the use of alternative promoters has been shown to have numerous roles in normal gene expression and in disease-associated genes [67]. Other results have also shown the potential to express shorter gene products in response to an environmental change [68]. Our results, showing that many cryptic transcripts are translated and that some cryptic promoters are activated by a nutritional shift, also fit with this possibility. We note that we did test for evidence of conservation between S. cerevisiae genes with cryptic transcription and S. bayanus orthologs, but did not detect any significant reduction in either synonymous or nonsynonymous changes in genes with cryptic transcription when compared to genes without cryptic transcription, but of similar length (unpublished data). Second, the information expressed from cryptic promoters may provide the potential for an adaptive mechanism in which, under appropriate selective conditions, expression of such products would enable improved growth or survival, thereby facilitating evolutionary genetic changes. Such an idea was previously suggested for the yeast prion [PSI+], which affects the fidelity of translational termination and thus allows for the possible production of novel protein products [69–71]. Strains containing [PSI+] can acquire complex phenotypic traits distinct from [psi−] strains, and when outcrossed to wild-type strains, these phenotypic traits can sometimes be maintained even after treatment to remove [PSI+] [70,71]. A possible role for intergenic RNAs has also been previously suggested [72]. Third, some cryptic promoters may serve to regulate transcription or control chromatin structure without producing a functional gene product. A previous study demonstrated that a promoter within PRY3 of S. cerevisiae serves to repress PRY3 expression during mating [21]. In this case, transcription from the internal promoter does not appear to play any functional role. In other cases, the act of transcription may alter chromatin structure in some beneficial way, as previously suggested [68]. Finally, some cryptic promoters may be “noise,” existing as one of many transcriptional events that serve no apparent biological role [73]. In such a scenario, a significant role of the genes we identified in our screen would be to minimize such “noise,” similar to that of Trf4, Air1, Air2, and components of the exosome in the removal of cryptic intergenic transcripts [37]. Given the very large number of cryptic promoters in S. cerevisiae, it seems reasonable to speculate that all of these reasons and others may turn out to be true. The analysis of specific cryptic promoters will likely yield additional insights into their roles and into previously unknown aspects of gene expression.
All S. cerevisiae strains are listed in Table S9. Strains with the prefix “FY” are isogenic with a GAL2+ derivative of S288C [74]. Strains were constructed by standard methods, either by crosses or by transformations [75]. The spt6-1004 [23]), spt16-197 [76], (hta1-htb1)Δ0::LEU2 [77], spt2Δ0::kanMX [35], spt21Δ0::kanMX [78], spt4Δ0::URA3 [79], and RAS2val19 [80] alleles have been previously described. The can1Δ0::STE2pr-LEU2 allele was generously provided by the Boone lab [81]. The ctk3Δ0::kanMX, set2Δ0::kanMX, eaf3Δ0::kanMX, rco1Δ0::kanMX, asf1Δ0::kanMX, rtt106Δ0::kanMX, chd1Δ0::kanMX, cdc73Δ0::kanMX, rtf1Δ0::kanMX, ras2Δ0::kanMX, and ygr117cΔ0::kanMX deletion mutations are from the S. cerevisiae haploid nonessential genome deletion library [82], and the deletions were verified to be correct by PCR. The ctk1Δ0::URA3, ctk2Δ0::URA3, chd1Δ0::HIS3, hir1Δ0::LEU2, and hir1Δ201::kanMX deletion mutations were constructed by replacing the ORFs with URA3, HIS3, LEU2, or kanMX by standard methods [83–85]. The spt10-302::URA3 allele consists of a Tn10-LUK transposon [86] inserted at the SPT10 locus. The [SPT10::URA3]dup duplication allele was constructed by integrating plasmid pFW216 (a derivative of pRS306 containing SPT10 and URA3) at the ura3-52 genomic locus [87,88]. To construct the hht2-S57P allele, a 1.9-kb SmaI-EcoRI fragment containing HHT2 and HHF2 from plasmid pDM18 was ligated into vector pRS306 (URA3) [88,89]. The TCT codon (Ser) at base pair position +172 of HHT2 (relative to the +1 ATG start codon) was mutated to a CCT codon (Pro) by site-directed mutagenesis (QuikChange kit, Stratagene). The resulting plasmid was used to replace the wild-type HHT2 allele in strain FY2716 by two-step gene replacement. The presence of the mutant hht2-S57P allele in the genome was verified by sequencing. Strain FY2724 contains a point mutation in HHT2 changing the GAA codon (Glu) at base pair +316 to an AAA codon (Lys) and was created by UV mutagenesis of strain FY2713 and verified by sequencing.
To construct the kanMX-GAL1pr-flo8-HIS3 reporter, a 2-kb cassette containing the kanMX marker and the GAL1 promoter was amplified by PCR from plasmid pFA6a-kanMX6-PGAL1 [90]. This cassette was used to transform strain FY2425 by integration at the FLO8 promoter, replacing base pairs −1,147 to −1 (relative to the FLO8 +1 ATG start codon), to create strain FY2174. The HIS3 ORF (663 bp) was amplified by PCR from plasmid pRS403 [84] and transformed into strain FY2174 at the genomic FLO8 locus, replacing the 3′ end of the FLO8 ORF and the first 105 bp of the 3′ UTR (base pairs +1,727 to +2,505 of FLO8 relative to the +1 ATG start codon). Successful transformants were selected on SC-His medium and verified by PCR. The HIS3 ORF is inserted out-of-frame with respect to the FLO8 ORF and is inserted 3′ of both the internal FLO8 TATA element (+1,626 to +1,631) and the cryptic transcription initiation sites of the FLO8 short transcript (+1,679 to +1,685) [23]. To construct the flo8-HIS3 reporter, the HIS3 ORF was transformed into strain FY2425 and inserted at the FLO8 genomic locus as described above.
The APM2-TAP::His3MX, DDC1-TAP::His3MX, OMS1-TAP::His3MX, PUS4-TAP::His3MX, and SYF1-TAP::His3MX alleles are from the S. cerevisiae TAP-tagged library [58]. The apm2–1-TAP::His3MX allele contains a point mutation in the in-frame ATG codon at base pair position +1,420 of APM2 (relative to the +1 ATG start codon), changing it to a TTG codon (Leu). The apm2–2-TAP::His3MX allele contains three point mutations at the +1 ATG start codon of APM2, changing it to a CGT codon. The apm2–3-TAP::His3MX allele contains both the +1 ATG and the +1,420 ATG mutations in APM2. The pus4-1-TAP::His3MX allele contains a point mutation in the in-frame ATG codon at base pair +478 of PUS4 (relative to the +1 ATG start codon), changing it to a GTG codon (Val). The pus4–2-TAP::His3MX allele contains three point mutations at the +1 ATG start codon of PUS4, changing it to a CGT codon. The pus4-3-TAP::His3MX allele contains both the +1 ATG and the +478 ATG mutations in PUS4. All ATG mutations were constructed by a two-step gene replacement using a previously described method [91] and verified by sequencing.
For liquid cultures, strains were grown in either YPD rich medium (1% yeast extract, 2% peptone, and 2% glucose) or SD minimal medium (0.15% yeast nitrogen base, 0.5% ammonium sulfate, and 2% glucose) as indicated. Synthetic complete media plates (SC) and synthetic complete drop-out media plates (SC-His) were made as previously described [75]. SC + Gal plates and SC-His + Gal plates were made using 2% galactose instead of glucose as the carbon source. For the spontaneous mutant selection, 3-aminotriazole (3AT) was added to SC-His plates at the concentrations described below.
Cryptic initiation mutants were isolated using the following three methods: spontaneous mutant selection, synthetic genetic array (SGA) analysis with the S. cerevisiae genome nonessential deletion set [82,92], and direct testing of candidate genes. Spontaneous mutant selection was performed using the parental wild-type strains FY2393, FY2713, FY2717, and FY2718, each containing the kanMX-GAL1pr-flo8-HIS3 reporter. Parental strains were grown overnight in 5-ml YPD cultures at 30 °C, washed twice in water, and then either 1 × 107 cells or 1 × 108 cells from each culture were plated on SC-His media plates containing 0, 1, 2, 3, 4, 5, or 10 mM 3AT. Plates were either UV-irradiated (5,000 μJ/cm2) or left untreated, and then grown at 30 °C to select for His+ mutants. Potential His+ cryptic initiation mutants were single-colony purified and retested to verify their His phenotype. Mutant genes were identified by diploid complementation, plasmid complementation, linkage analysis, and cloning by plasmid complementation with an S. cerevisiae genomic library [93]. A total of 254 different mutants were isolated, and 226 of them were identified as belonging to the following groups: SPT21, SPT10, HTA1-HTB1, HHT1, HHT2, HIR1, HIR2, HIR3, HPC2, and mutations linked to the kanMX-GAL1pr-flo8-HIS3 reporter. SGA analysis was performed as previously described [92], using the query strain L1102 and screening for deletion mutants that allowed growth on SC-His media plates. Potential positive candidates from SGA analysis were individually crossed with strain FY2506, and their His phenotype was verified by tetrad analysis.
Probe sequences corresponding to 5,869 ORFs of the S. cerevisiae genome were submitted to Agilent Technologies for microarray production. Each ORF was represented by six 60-mer probes spaced evenly along its coding sequence, with the most-5′ probe beginning at base pair position +1 (relative to the +1 ATG start codon) and the most-3′ probe ending at the final stop codon. Strains FY80 and FY2425 were used for four independent spt6 microarray experiments, and strains FY70 and FY347 were used for two independent spt16 microarray experiments. Experimental pairs were performed in dye reversal. Wild-type and mutant cells were grown in YPD medium at 30 °C to mid-log phase (1–3 × 107 cells/ml), shifted to 37 °C for 80 min, and then harvested as previously described [94] Sample preparation, labeling, and hybridization of microarrays were performed as previously described [94,95]. Microarray images were acquired and spots quantified with a GenePix 4000B microarray scanner and 3.0 software, respectively (MDS Sciex). Spatial detrending and variance stabilization normalization of raw microarray data were performed as previously described [95]. Genes were detected as expressing short transcripts in either the spt6 or the spt16 mutant using the following criteria. The mutant/wild-type ratio was calculated for each probe on the microarray using the normalized spot fluorescent intensity values. For each ORF, the 3′/5′ ratio was calculated by dividing the mutant/wild-type ratio of the most-3′ probe by the mutant/wild-type ratio of the most-5′ probe. Genes with high 3′/5′ ratios were predicted to express short transcripts, whereas genes with low 3′/5′ ratios (close to 1.0) were predicted to not express short transcripts. The location of the internal transcription start site for genes generating short transcripts was estimated by calculating the mutant/wild-type ratio of each probe in the gene relative to the corresponding ratio of the most-5′ probe. Based on the microarray results for five genes previously known to express short transcripts in an spt6 mutant (FLO8, SPB4, STE11, RAD18, and VPS72) [23], a 3′/5′ ratio threshold was set at 2.5, where only genes with a 3′/5′ ratio greater than 2.5 in either all four spt6 microarray experiments or both spt16 microarray experiments were predicted to express a short transcript. Using this criterion, 960 genes were predicted to express short transcripts in an spt6 mutant, and 1,130 genes were predicted to express short transcripts in a spt16 mutant. It is likely, though, that even more cryptic promoters exist, as the method of calculation likely and unavoidably discriminates against the identification of cryptic transcripts from highly transcribed genes. This discrimination arises from the fact that the hybridization signal from the 3′ probe is the sum of the signals for both the full-length and cryptic transcripts. Thus, for genes with a high level of the full-length transcript, the level of a cryptic transcript would need to also be high to be detectable. The 3′/5′ ratio from the microarray results are shown plotted according to expression levels in Figure S2. In support of a greater number of cryptic transcripts, when a more relaxed threshold was used (3′/5′ ratio of 2.0 rather than 2.5), 620 additional genes were predicted to express cryptic transcripts. When five genes were randomly selected from these 620 genes, four of them expressed short transcripts as detected by northern analysis (unpublished data). However, it is clear that not all genes produce cryptic transcripts, as northern analysis of ten other genes at random showed that only one produced a detectable cryptic transcript (I. Ivanovska, J. Pamment, and F. Winston, unpublished data).
mRNA preparation and northern hybridization analysis were performed as previously described [96]. Unless otherwise indicated, RNA was prepared from cells grown in YPD at 30 °C to mid-log phase (1–3 × 107 cells/ml). For temperature shift experiments, cells were grown in YPD to mid-log phase at 30 °C and then shifted to 37 °C for 80 min. For media shift experiments, cells were grown in YPD to mid-log phase at 30 °C, washed twice with SD, and then grown in an equivalent volume of SD at 30 °C for the indicated times. Double-stranded northern probes were amplified by PCR from genomic DNA and were designed to hybridize to the 3′ ends of FLO8 (+1,515 to +2,326), SPB4 (+1,605 to +1,812), STE11 (+1,868 to +2,110), APM2 (+1,449 to +1,786), DDC1 (+1,489 to +1,739), OMS1 (+1,084 to +1,351), PUS4 (+861 to +1,134), SYF1 (+2,032 to +2,525), and CHS6 (+1,917 to +2,295). A probe for SNR190 (+1 to +190) was used as a loading control for all northern analyses. Because the probes are double stranded, they could anneal to either sense or antisense transcripts. The base pair positions given for each probe is relative to the +1 ATG start codon of the respective gene.
For Western analysis of histone H3 and H3 K36 methylation, whole-cell protein extracts were prepared as previously described [79]. The protein concentration of extracts was determined by Bradford assay (Bio-Rad). Fifty micrograms of whole-cell extracts were separated on 15% acrylamide SDS-PAGE gels, transferred to immobilon-P membrane (Millipore), and analyzed by immunoblotting as previously described [48]. Antibodies were used that recognized total histone H3 (1:5,000 dilution, Abcam), dimethylated H3 K36 (1:10,000 dilution, Upstate), and trimethylated H3 K36 (1:10,000 dilution, Abcam). Antibodies were detected by chemiluminescence (PerkinElmer).
For western analysis of TAP-tagged proteins, whole-cell protein extracts were prepared as follows: 50 ml of cells were grown in YPD at 30 °C to mid-log phase (1–3 × 107 cells/ml) and then shifted to 37 °C for 80 min. Cells were washed twice with wash buffer (20 mM Tris-Hcl, 150 mM NaCl [pH 7.5]) and resuspended in 400 μl of lysis buffer (50 mM Hepes-KOH [pH 7.5], 150 mM NaCl, 10% glycerol, 0.5% NP-40, 1 mM EDTA, 1 mM PMSF, 2 μg/ml Leupeptin, 2 μg/ml Pepstatin A). One milliliter of glass beads was added, and cells were lysed by vortexing in an Eppendorf multihead shaker 5432 for 40 min at 4 °C. The cell lysate was spun out through a hole punctured in the bottom of the tube, by spinning for 2 min at 1,000 rpm. The lysate was spun for 5 min at 14,000 rpm, and the supernatant was saved and spun again for 15 min at 14,000 rpm. The supernatant was saved from this final spin and used for western analysis. Total protein concentration of extracts was determined by Bradford assay (Bio-Rad). Equal amounts of whole-cell extracts were separated on 8% acrylamide SDS-PAGE gels, transferred to immobilon-P membrane (Millipore), and analyzed by immunoblotting as previously described [79]. The TAP tag was detected by chemiluminescence (PerkinElmer Life Sciences) using the peroxidase anti-peroxidase antibody (1:5,000 dilution, Sigma). Pgk1 was used as a loading control and visualized with anti-Pgk1 antisera (Molecular Probes) that was generously provided by Angelika Amon's laboratory.
To examine which protein domains are present and lost, we obtained data on proteins from SGD (ftp://genome-ftp.stanford.edu/pub/yeast/sequence_similarity/domains, last updated on September 25, 2007) and mapped them onto the proteins encoded by genes with cryptic transcription initiation. We considered the first ATG after the most-5′ limit of the cryptic transcript as a conservative limit for the length of the short protein being produced; i.e., cases in which a minimal number of residues would be lost. A domain was called to be absent if the position of the ATG was downstream of the domain start site. Using published data on protein domains that are at the physical interface of the interacting partners [97], we also examined whether these lost domains are known to mediate physical interaction among proteins. Finally, in order to estimate how common these domains are among yeast proteins, we tabulated how many proteins in the genome have these domains.
The microarray data, accession number GSE12272, can be found at GEO (http://www.ncbi.nlm.nih.gov/geo/).
|
10.1371/journal.pntd.0000834 | Metabolomics-Based Discovery of Diagnostic Biomarkers for Onchocerciasis | Development of robust, sensitive, and reproducible diagnostic tests for understanding the epidemiology of neglected tropical diseases is an integral aspect of the success of worldwide control and elimination programs. In the treatment of onchocerciasis, clinical diagnostics that can function in an elimination scenario are non-existent and desperately needed. Due to its sensitivity and quantitative reproducibility, liquid chromatography-mass spectrometry (LC-MS) based metabolomics is a powerful approach to this problem.
Analysis of an African sample set comprised of 73 serum and plasma samples revealed a set of 14 biomarkers that showed excellent discrimination between Onchocerca volvulus–positive and negative individuals by multivariate statistical analysis. Application of this biomarker set to an additional sample set from onchocerciasis endemic areas where long-term ivermectin treatment has been successful revealed that the biomarker set may also distinguish individuals with worms of compromised viability from those with active infection. Machine learning extended the utility of the biomarker set from a complex multivariate analysis to a binary format applicable for adaptation to a field-based diagnostic, validating the use of complex data mining tools applied to infectious disease biomarker discovery and diagnostic development.
An LC-MS metabolomics-based diagnostic has the potential to monitor the progression of onchocerciasis in both endemic and non-endemic geographic areas, as well as provide an essential tool to multinational programs in the ongoing fight against this neglected tropical disease. Ultimately this technology can be expanded for the diagnosis of other filarial and/or neglected tropical diseases.
| Onchocerciasis, caused by the filarial parasite Onchocerca volvulus, afflicts millions of people, causing such debilitating symptoms as blindness and acute dermatitis. There are no accurate, sensitive means of diagnosing O. volvulus infection. Clinical diagnostics are desperately needed in order to achieve the goals of controlling and eliminating onchocerciasis and neglected tropical diseases in general. In this study, a metabolomics approach is introduced for the discovery of small molecule biomarkers that can be used to diagnose O. volvulus infection. Blood samples from O. volvulus infected and uninfected individuals from different geographic regions were compared using liquid chromatography separation and mass spectrometry identification. Thousands of chromatographic mass features were statistically compared to discover 14 mass features that were significantly different between infected and uninfected individuals. Multivariate statistical analysis and machine learning algorithms demonstrated how these biomarkers could be used to differentiate between infected and uninfected individuals and indicate that the diagnostic may even be sensitive enough to assess the viability of worms. This study suggests a future potential of these biomarkers for use in a field-based onchocerciasis diagnostic and how such an approach could be expanded for the development of diagnostics for other neglected tropical diseases.
| Onchocerciasis, commonly referred to as “river blindness” is classified by the World Health Organization (WHO) as a neglected tropical disease, afflicting approximately 37 million people in Africa, Central and South America and Yemen, with 89 million more at risk [1]. Symptoms of the disease include acute dermatitis and blindness, the result of which is the loss of 1 million disability-adjusted life years (DALYs) annually [2]. The causative agent, the filarial nematode Onchocerca volvulus, is transmitted in its larval stage between human hosts through the bite of a Simulium (sp.) black fly. Once these parasites have matured into the adult form, they can live for approximately 14 years in subcutaneous nodules within a human host [3]. The drug ivermectin (Mectizan) has served as the principal means of onchocerciasis control [4], however, after initially reducing the number of microfilariae, within a year, the microfilariae return to levels of 20% or higher than that prior to treatment [5]. The combination of the lack of effect of annual ivermectin treatment on adult worm survival and the fecundity of adult females, along with significant fly and human migration patterns has helped to perpetuate the disease.
In Africa, where onchocerciasis control programs have been in place since the founding of the Onchocerciasis Control Programme in West Africa (OCP, 1974–2002) and are currently being conducted by the African Programme for Onchocerciasis Control (APOC, 1995-present), diagnosis is an essential aspect of the determination of treatment and distribution of medication. In the Western hemisphere, accurate and robust diagnostics are essential for attaining the goal of disease elimination. Twice yearly dosage of ivermectin, through the efforts of the Onchocerciasis Elimination Program for the Americas (OEPA, 1992-present), has lead to a minimization of infection to 13 foci within six countries in Central and South America. Although mass treatment of onchocerciasis foci in the Western hemisphere is slated to be suspended in 2012 [6], achieving the goal of elimination is contingent upon continued surveillance of the disease. However, proper surveillance is directly dependent on the availability of robust diagnostic technologies used for infection assessment. This need is further underscored in studies of antibiotic treatments being investigated for targeting Wolbachia endosymbiotic bacteria [7]–[10] as well as reports of sub-optimal response to ivermectin treatment [11], [12]. In both of these cases an accurate diagnostic is critical for the analysis of drug efficacy and patient drug response.
Currently, multinational control and elimination programs primarily rely on various techniques for diagnosis including: entomological studies of Simulian flies, Ov specific antigen tests, antibody tests, analysis of microfilariae in skin snips, nodule palpation and quality of those nodules that can be excised. There are a number of technical concerns with each technique including: a lack of sensitivity and reproducibility, invasiveness, and the inability to distinguish past from present infection or between filarial diseases [13]–[15]. A small molecule/metabolite based test has the potential for reflecting a more accurate picture of infection status, as it is a comprehensive measure of the effects of posttranslational modification and regulation. Furthermore, small molecules are frequently constitutively produced (e.g., excretory-secretory products), diffuse easily and are inherently non-immunogenic in vivo, thus avoiding some of the technical challenges associated with DNA and protein-based diagnostics.
Although adult O. volvulus worms do not reside directly in the blood, the highly vascularized subcutaneous nodules of the human host allow for the potential diffusion of adult parasite-derived compounds into the blood where compounds involved in host response to infection might also be present. Since the microfilariae (mf) and third infective larval stage (L3) of the O. volvulus life cycle do come in contact with the vascular system during vector transmission, it is additionally possible that some mf or L3 produced compounds might also be localized to this biological sample. Certainly, as a starting point, the blood matrix serves as an easy to obtain, chemically complex data rich matrix for metabolite analysis [16], [17].
However, a technical challenge of analyzing a large number of metabolites stems from the shear size and complexity of the resulting data set. Initially devised and applied to the analysis of highly dimensional gene micro-array data, a number of machine learning approaches have been expanded and used for identifying patterns of biomarkers resulting from the multidimensional analysis of genes, proteins, and metabolites that can be linked to early detection [18], survival prediction [19], and disease outcomes [20]. Although identification of a single biomarker “smoking gun” is perceived as the ideal scenario, more attention is being focused on the use of multiple markers for improving overall diagnostic accuracy [18], [21], [22] and model stability [23].
Herein, we report a liquid chromatography-mass spectrometry (LC-MS) based approach to the discovery of a set of molecules that, in combination, provide a statistically relevant characteristic of onchocerciasis infection. An initial untargeted analysis was applied to the profiling of O. volvulus infected and uninfected blood plasma and serum samples representing a variety of geographic regions and disease states, including other tropical diseases. This analysis resulted in a set of statistically significant mass features identified for their potential as onchocerciasis-specific biomarkers. Using multivariate statistics and machine learning algorithms, these metabolic signatures were further evaluated for their ability to discriminate O. volvulus-infected and uninfected individuals, therefore, creating the basis of a small molecule-based diagnostic for onchocerciasis.
The use of human serum and plasma samples in the study was approved by the Scripps Health Human Subjects Committee. Samples with geographic origins outside of the United States of America (USA) consisted of pre-existing, unidentifiable diagnostic specimens collected with written informed consent and in cases of illiteracy, a literate witness signed and a thumbprint was made by the participant. These samples were determined by the Scripps Health institutional review board (IRB) to be exempt from formal review under 45 CFR 46 101. O. volvulus negative controls from the USA consisted of serum and plasma samples and were obtained with written informed consent from healthy donors through The Scripps Research Institute Normal Blood Donor Service and approved by the Scripps Health IRB. All patient codes have been removed in this publication.
Onchocerciasis positive samples were collected in characterized endemic areas and their status confirmed by either positive skin snip (mf +) or nodule palpation (nodule +). Several sample groups used in this analysis were collected during previously published studies including serum from Liberia [24], [25] and Ghana collected in 2003 [9]. The Ghana sera collected in 1986 and 1991 were obtained from the College of Public Health, University of South Florida. Cameroon samples were obtained as part of a nodulectomy campaign conducted in villages surrounding Kumba, Cameroon in 2006 and consist of plasma from O. volvulus-positive individuals (nodule + with nodules containing live females), O. volvulus-negative individuals (skin snip - volunteers with no current or prior symptoms of O. volvulus infection), and ambiguous samples (nodules contained either dead, calcified worms or lipomas with no evidence of worms, or for which there were no particular disease symptoms recorded). Guatemala sera were obtained as part of a nodulectomy campaign conducted by the Guatemala Ministry of Health and the Centro de Estudios en Salud, Universidad del Valle de Guatemala in several villages within the Guatemalan Central Endemic Zone from 2007–2008. Nodules were surgically removed from all individuals sampled, and nodule dissection was conducted to assess worm viability on nodules from five of the 21 individuals whose serum was analyzed in this study. Of those dissected, no live worms were found. Leishmaniasis positive, Chagas disease positive, and onchocerciasis negative sera were obtained from the Centro de Estudios en Salud, Universidad del Valle de Guatemala. Indian lymphatic filariasis positive plasma samples were obtained from the Laboratory of Parasitic Diseases, U.S. National Institutes of Health. A detailed summary of the samples analyzed in this study is presented in Table 1.
Solvents used were of high performance (HPLC) grade. A methanol precipitation of proteins was conducted by adding 400 µl aliquots of ice cold methanol to 100 µl aliquots of serum and plasma samples. The samples were immediately vortexed for 30 sec and allowed to rest on ice for 20 min. After centrifugation at 13,780×g for 5 min, the metabolite containing supernatent was removed from the precipitated protein pellet and transferred to fresh tubes. The supernatent samples were dried in a GeneVac EX-2 Evaporation System (GeneVac Inc., Valley Center, New York, USA) at ambient temperature and then resuspended to a 50 µl volume in water: acetonitrile (95∶5), vortexed for 30 sec and then centrifuged again at 13,780×g for 5 min. After being transferred to LC vials, samples were stored at 4°C and transferred to the LC-MS thermostated autosampler (6°C), typically within 48 h of their preparation.
In order to minimize instrumental drift, sample sequences were composed of a single injection of each sample in randomized order. To monitor any potential instrument irreproducibility and to confirm the absence of sample carry over within the chromatographic run, a mobile phase blank and an external standard were injected every 24 h throughout the duration of the analysis.
Experiments were performed with an electrospray-ionization time-of-flight (ESI-TOF) MS (Agilent 1200 LC, TOF 6210, Agilent Technologies, Santa Clara, CA, USA). Each sample analysis consisted of an 8 µl injection of extracted sample with chromatographic separation across a reverse phase C18 column (Zorbax 300SB C18 Capillary, 3.5 µm, 1 mm×150 mm; Agilent Technologies, Santa Clara, CA, USA) at a capillary pump flow rate of 75 µl/min. Mobile phase A was composed of water with 0.1% formic acid, and mobile phase B was acetonitrile with 0.1% formic acid. Each sample was analyzed over a 60 min run time with a gradient consisting of a 45 min linear gradient from 5% to 95% B and 15 min isocratic hold at 95% B. Between sample injections a wash step was used to minimize carry over. It consisted of a saw-tooth linear gradient beginning with a hold at 95% A for 10 min. Then, linear ramping between 5% and 98% B for 5 minute increments throughout the 35 min wash cycle was followed by a 20 minute final re-equilibration of the column with an isocratic hold at 95% A.
Consistent mass accuracy (<2 ppm) was maintained through the constant infusion (2 µl/min) of reference masses via a second nebulizer. Data were collected in positive electrospray ionization (ESI) mode scanning in centroid mode from 75 to 1,100 m/z with a scan rate of 1.0 spectrum per second in 2 GHz extended dynamic range. The capillary voltage was 3,500 V; the nebulizer pressure, drying gas flow and gas temperature were set to 20 psig, 12 l/min and 350°C, respectively.
All mass spectral data was collected in .d format and converted to .mzData using the Mass Hunter Qualitative Analysis software version B.03.01 (Agilent Technologies, Santa Clara, CA, USA). XCMS [26] software was used for peak matching, non-linear retention time alignment and quantitation of mass spectral ion intensities across all .mzData mass spectral files. Statistical comparison of the intensity data was conducted using the XCMS built in Welch's t-test. False discovery rate (FDR) analysis was conducted with the q-value program [27] in R version 2.9.0 [28]. Principal Components Analysis (PCA) was conducted with Statistica software version 8.0 (StatSoftInc., Tulsa, OK, USA), machine learning algorithms were implemented using Weka Explorer version 3.6.0 [29] with 10 fold cross-validation settings.
The molecular formula assignment made for the 10 selected small molecule biomarkers was conducted through a combination of LC-MS/MS fragmentation using a quadrupole- TOF MS (QTOF 6510, Agilent Technologies, Santa Clara, CA, USA) and sub-2 ppm accurate mass measurements using a Bruker Daltonics Apex II 7.0 Tesla Fourier transform ion cyclotron (FT-ICR) MS (Bruker Daltonics., Billerica, MA,USA). For the QTOF analysis chromatographic conditions were identical to those reported for the profiling experiment and serum plasma samples from either the Scripps normal blood or pooled patient samples were used for the analysis. The average m/z and retention times of each of the biomarkers obtained through XCMS analysis, were used for targeted MS/MS analysis with a starting collision-induced dissociation energy of 20eV. Fragmentation patterns were analyzed with the Agilent Mass Hunter Qualitative Analysis software version B.03.01 using the targeted MS/MS and formula generation algorithms and compared with the MS/MS fragment data in the METLIN database [30].
The FTMS system was equipped with a custom machined electrospray source with two nebulizers for dual spray ionization. The main orthogonal nebulizer was used for LC-eluent, while the second nebulizer was used to introduce a calibration mixture containing two compounds (aminoantipyrine at 204.1132 m/z and quinidine 325.1911 m/z) at 3 mM concentration mixed with 1∶10 dilution of Agilent low concentration tune mix. A linear calibration fit was used in the narrow range to internally calibrate individual mass spectra. The chromatographic conditions were identical to those reported for the profiling experiment with an additional analysis using a smaller i.d. column with the same stationary phase composition (Zorbax 300SB C18 Capillary, 3.5 µm, 0.3 mm×150 mm; Agilent Technologies, Santa Clara, CA, USA) at a capillary pump flow rate of 4µl/min. Pooled serum and plasma samples from either the Scripps normal blood or patient samples were used for the analysis.
A metabolomic approach was developed to address the need for improvement of diagnostics in onchocerciasis detection. Profiling of blood biomarkers is much less invasive than skin snipping or nodulectomy, and should have the added advantage of increased sensitivity. Antigen tests have been attempted in the past, however, the immunogenicity of the proteins has been a consistent deterrent [31], [32]. An advantage of profiling low molecular weight compounds is that they are typically not immunogenic (i.e., M.W. <1,100 amu) and therefore not subject to such a limitation. It is important to note that profiling molecules with molecular weight less than 1,100 amu will also include peptides and/or protein fragments, expanding the pool of available analytes that can be detected.
The most important aspect of any clinical analytical study resides with the quality of the samples used; here representative serum and plasma samples from a variety of subject populations were incorporated to minimize the effects of non-relevant metabolic variation (e.g., nutrition, sex, age, race) and magnify those metabolic differences that are not only statistically significant between specific populations, but relevant in identifying the changes in metabolism that can be directly attributable to infection.
One of the analytical limitations with an untargeted LC-MS metabolomics approach is that of inter-sequence reproducibility (i.e., sample preparation, instrument drift, column and mass spectral baseline variation) when comparing samples directly between analytical sequences. Such inter-sequence variability can introduce shifts in ion intensities that can interfere with the accuracy of downstream statistical analysis. Therefore, this study was conducted with single injections of each sample, analyzed in randomized order consecutively within one analytical sequence (Figure 1). Due to such analytical constraints, small groups of representative samples were selected from various sample classes (e.g., O. volvulus-infected and uninfected individuals from various geographic regions and individuals infected with other parasitic diseases). XCMS analysis of the sample mass spectral data files (n = 136) resulted in the measurement of a total of 2,350 mass features. Testing the overall reproducibility of the analysis, the coefficient of variation (CV) was found to be 15.9% as calculated from all mass feature intensity values compared across triplicate injections of a single plasma sample analyzed throughout the analytical sequence. This value is comparable to previous studies of analytical variation within plasma and serum analysis by our laboratory and consistent with a number of other LC-MS based metabolomics studies [33], [34]. Statistical comparison between all onchocerciasis positive samples (n = 76) and all onchocerciasis negative samples (n = 56), including those infected with other tropical diseases, by Welch's t-test resulted in 194 features with a p<1×10−4; with a false discovery rate FDR of 54%. To reduce the number of potentially erroneous markers and focus on those mass features with the most potential in distinguishing disease, the top 35 mass features (p<1×10−7) were chosen for more stringent analysis through assessment of the quality of the resulting extracted ion chromatograms (EICs) (Figure S1). While XCMS pre-processing software contains a robust retention time correction and peak alignment algorithm, an important aspect of this study is the statistical quantitation of biomarkers, therefore any features with questionable quantitation, observed as imperfect alignment or inconsistent peak boundaries across samples were ruled out of further analysis. Additionally, since several mass features may redundantly describe one chemical metabolite due to the presence of in-source fragments, adducts, or multiply charged species and overlapping retention time. The features were separated into unique peak groups and representative ions with the highest overall abundance were included in a subset of 14 features for further analysis (Table 2). Interestingly, the majority of these features were detected at lower levels in infected individuals relative to those without onchocerciasis. Analysis of the selected biomarkers with MS/MS and FTMS analysis has provided molecular masses and assigned molecular formulas that could be used to classify the biomarkers into distinct chemical classes; of the 14 markers identified 10 were small molecules and four were protein fragments or small peptides.
Beginning with a subset of the larger sample set, the mass spectral data for the top 14 candidate biomarkers were investigated for their ability to discriminate O. volvulus infected individuals (n = 55) from healthy controls (n = 18) from the African serum and plasma samples. PCA of the effect of these 14 biomarkers was used to visualize the variation between these samples groups (Figure 2A). A distinct clustering of the O. volvulus infected versus the healthy individuals was observed across the x-axis of the PCA score plot, implying that principal component 1 (PC1) contained the variance of the data set required to distinguish these two sample groups. The next greatest amount of variation within the data set appeared to have little effect on discriminating infection or even geographic differences, but appears to be more representative of the heterogeneity present among healthy controls.
The top 14 candidate biomarkers were also applied to a larger sample set comprised of multiple geographic regions including O. volvulus-infected individuals (n = 76) and healthy and disease controls (n = 56). PCA of these 14 biomarkers (Figure 2B) revealed the inherent complexity encountered when employing a metabolite profiling approach to diagnostic development. As with the initial African samples, there is general clustering of the onchocerciasis positive individuals with the variance contained in PC1 having good discriminatory power. The disease and healthy controls cluster separately from the onchocerciasis positive individuals, however, there is some overlap between one of the Chagas disease and two of the leishmaniasis positive individuals. Interestingly, the lymphatic filariasis samples, infected with the closely related filarial parasite Wuchereria bancrofti, cleanly cluster with the healthy controls.
Ideally serum and plasma samples would not be directly compared against each other as the two matrices have distinct chromatographic differences (Figure S2). However, given the nature of onchocerciasis sample banks that have been collected over the past 20 years, it was important to determine if the resulting biomarker results would be biased to one biological sample type over another. Importantly, our results show that the plasma samples from Cameroon as well as the Indian lymphatic filariasis plasma samples consistently align as expected with the multi-region serum sample set in distinguishing onchocerciasis infected from uninfected individuals.
As evidenced in Figure 2C, there is little clustering of the Guatemalan individuals initially classified as onchocerciasis positive; rather there appears to be a continuum of onchocerciasis disease variation within those samples. However, dissections of excised nodules at the time of nodulectomy revealed no live worms, as opposed to the results of the Cameroon samples where infection status was confirmed by the extraction of live O. volvulus worms.
Although tools such as PCA provide a graphical means of distinguishing between sample groups, they do not have the ability to provide a quantitative diagnostic assessment as would be needed nor are they intended to be used for field applications of an onchocerciasis diagnostic. Alternatively, machine learning algorithms do provide the necessary binary output, as well as calculate confidence intervals of a given classification. The mass spectral intensity values for the onchocerciasis serum and plasma data set were used as inputs in a collection of machine learning algorithms. The algorithms were chosen to provide a survey of the various types of machine learning algorithms that could be used with mass spectral data in diagnostic assessments, either alone or in combination in more sophisticated algorithms. Results of this analysis are summarized in Table 3 where sensitivity (true positive rate) and specificity (1–false positive rate) are displayed. The receiver operating characteristic (ROC) areas present a numerical value description of the relationship between sensitivity and specificity for a given diagnostic test [35], [36]. In the context of a binary classification problem as presented here, a value of 0.5 indicates there is no discrimination within the test and shows any result is essentially the same as a random guess, while a value of 1.0 indicates a perfect test prediction. Based upon the data, it is clear that the inclusion of the Guatemala samples within the sample analysis dramatically increases the number of reported false positives, compromising the accuracy of the test overall. However, it is important to note that within the context of the Africa sample set, the ROC area approaches, or is equal to, a perfect test prediction in numerous cases, and in the case of the functional trees classification tree algorithm, perfect sensitivity and specificity can be achieved.
Metabolomics, or the measurement of all the metabolites present in an organism, and metabolite profiling, in which a smaller subset of metabolites are measured, have become established as useful tools in the “real-time” measurement of organismal metabolism. For infectious disease, previous metabolomics approaches have included mice challenged with the protozoan parasites Trypanosoma brucei brucei [37] and Plasmodium berghei [38], trematode parasites Echinostoma caproni [39] and Schistosoma mansoni [40] and some viruses [41]. This study represents the first investigation of a metabolomic approach to the discovery of biomarkers and creation of a diagnostic test for identifying and classifying onchocerciasis infection. Through the use of multivariate statistics and machine learning algorithms, the potential of metabolomic analysis has been demonstrated for uncovering biomarkers for specific determination of not only onchocerciasis infection but holds promise for the diagnosis of other parasitic diseases. Specifically, this was demonstrated by the clustering of the W. bancrofti infected samples with those individuals that were not infected with O. volvulus in the multivariate PCA. This clustering showed the potential specificity of the biomarkers for the discrimination of onchocerciasis from other filarial diseases. Although this analysis consists of only four representative lymphatic filariasis samples, the distinct clustering of these samples with uninfected individuals is noteworthy and argues for future analysis that includes other filarial disease pathogens (e.g., Brugia malayi, Loa loa).
The 14 candidate biomarkers showed excellent performance in the African specific sample set with up to 99–100% sensitivity and specificity when examined with the single machine learning algorithms. With 99% of onchocerciasis disease prevalence in Africa [42] and the presence of multiple regions of ongoing transmission [43], this is the most clear test of the biomarker strategy.
When applied to a multi-region sample set, the multivariate PCA of the biomarker analysis resulted in a wide spread of results across the range of infected and uninfected individuals. This observation raises several questions regarding the unique epidemiological challenges of measuring onchocerciasis in the Americas. In the context of the PCA, the Guatemalan patients did not classify as expected if nodule presence alone is used as an indicator of infection. However, nodule presence as a diagnostic is known to have exceedingly poor sensitivity and specificity. A possible explanation of this data is that the observed heterogeneity is related to microfilarial load. Unfortunately, skin snip samples with mf counts were not collected for the Guatemala sample set. Nonetheless, if this observed spread of data were correlated with variation in the presence of the mf, then in a region such as the Guatemalan Central Endemic Zone (CEZ) where biannual dosage of ivermectin reaches high coverage levels [44], mf should be nearly absent and we would expect to see no spread of the data but rather a distinct cluster with or near the uninfected individuals. Alternatively, the observation that a quarter of the nodules from these infected individuals from the Guatemalan CEZ did not contain living worms, indicates that these biomarkers may be sensitive to not only the presence, but also the viability of the infective worms. The results of this PCA are consistent with an increasing body of evidence that biannual ivermectin treatments, as are received in the Guatemalan CEZ, have an effect on the viability of adult female worms and ultimately on the elimination of parasites [45]–[47]. Since the Guatemalan O. volvulus positive samples do not segregate along clear lines with the clinically confirmed samples from Africa, it is possible that the continuum seen in the PCA plot reflects a range of infection that could be correlated qualitatively or quantitatively to the health of the worms (e.g., live healthy, dying, and dead) in vivo. Given that an individual with dead or dying worms does not need further treatment in the context of ivermectin mass drug administration, this finding is particularly valuable in the context of onchocerciasis elimination progress. Ideally, a biomarker determination study would involve independent sample sets for training, validation, and testing. Due to sample limitations inherent to onchocerciasis and many neglected tropical diseases in general, we have chosen to use an approach that trains on the majority of the sample set, and through the 10-fold cross validation machine learning analyses, conduct tests on small subsets of the full sample set [48].
In this study, we report only those features detected in positive ion mode with the highest statistical significance and the most accurate intensity values by XCMS analysis. Consistent among these 10 small molecule features is that they are all fatty acids and related fatty acid derivatives. Further investigations into the biological roles of these fatty acids and fatty acid sterols in onchocerciasis disease progression and potential interaction with the down-regulated proteins is of distinct interest, not only in the development of a diagnostic but also to more clearly understand the biology of this disease. Almost certainly, other biomarkers could be discovered and validated simply by altering the chromatographic (e.g., HILIC) and/or ionization conditions (e.g., negative mode ESI, APCI). It is possible that additional markers can be eventually be added to the repertoire of biomarkers used for onchocerciasis detection, further increasing assay specificity.
The achievement of the goals of elimination and eradication of onchocerciasis and of the neglected tropical diseases in general, ultimately depends upon the ability to measure and track the progress of disease elimination and recrudescence. Our study highlights advantages of a metabolomics based diagnostic over onchocerciasis diagnostics currently implemented including: sensitivity, reproducibility, invasiveness, and the potential for multiplexing with biomarkers for other filarial and/or neglected tropical diseases. Fine calibration of this test in the Western Hemisphere would require characterized samples from individuals with confirmed active infection. Unfortunately, these samples are rapidly becoming a rarity due to the success that has been achieved by OEPA. Further refinement and validation of this metabolomic based diagnostic approach calls for an expansion of the mass spectral analysis with larger sample sets, while inclusion of a greater demographic representation will allow for further validation of the test in specific populations (e.g., children, adults, different genetic backgrounds). Eventually, the optimized biomarkers can be ported into field-based technologies (e.g., immuno-chromatographic or micro-fluidic-based tests) for use as a point-of-care diagnostic, a determinant for the distribution and duration of treatment, and ultimately for long-term disease surveillance.
|
10.1371/journal.pbio.0060187 | Multivariate Patterns in Object-Selective Cortex Dissociate Perceptual and Physical Shape Similarity | Prior research has identified the lateral occipital complex (LOC) as a critical cortical region for the representation of object shape in humans. However, little is known about the nature of the representations contained in the LOC and their relationship to the perceptual experience of shape. We used human functional MRI to measure the physical, behavioral, and neural similarity between pairs of novel shapes to ask whether the representations of shape contained in subregions of the LOC more closely reflect the physical stimuli themselves, or the perceptual experience of those stimuli. Perceptual similarity measures for each pair of shapes were obtained from a psychophysical same-different task; physical similarity measures were based on stimulus parameters; and neural similarity measures were obtained from multivoxel pattern analysis methods applied to anterior LOC (pFs) and posterior LOC (LO). We found that the pattern of pairwise shape similarities in LO most closely matched physical shape similarities, whereas shape similarities in pFs most closely matched perceptual shape similarities. Further, shape representations were similar across participants in LO but highly variable across participants in pFs. Together, these findings indicate that activation patterns in subregions of object-selective cortex encode objects according to a hierarchy, with stimulus-based representations in posterior regions and subjective and observer-specific representations in anterior regions.
| As early as 1031 a.d., the Arab scholar Ibn al-Haytham suggested that visual experience was not veridical, but inherently subjective. During the last few decades, this observation has given rise to one of the core questions in visual neuroscience: how does the subjective experience of visual stimuli relate to their neural representations in the brain? It is well-known that visual shape is represented in a brain region called lateral occipital complex (LOC). However, do these representations reflect physical or perceptual stimulus characteristics? We presented observers with a set of complex visual stimuli and obtained three measures of similarity for these stimuli: a physical similarity measure based on stimulus parameters; a behavioral similarity measure based on discrimination performance; and finally a neural similarity measure based on multivariate pattern analyses in LOC. We found that in anterior LOC, neural stimulus similarities correlated with subjective perceptual similarities, but not with physical stimulus similarities; the reverse was true in posterior LOC. In addition, neural similarities were consistent across participants in posterior LOC, but highly variable across participants in anterior LOC. Together these findings suggest a two-part answer to the question of how cortical object representations relate to subjective experience: anterior regions appear to contain subjective, individually variable shape representations, whereas posterior regions contain stimulus-based shape representations.
| What is the neural code for object shape? This question has been at the core of systems neuroscience for decades. In monkeys, inferotemporal (IT) cortex has been shown to contain cells selective for complex shapes [1]; in humans, functional magnetic resonance imaging (fMRI) has identified a brain region known as lateral occipital complex (LOC) as a neural center for object representation [2,3]. This region responds more to intact than scrambled images of everyday objects [2,3] and is thought to be critical for object recognition [4,5]. However, the nature of the representations in these object-selective regions remains poorly understood.
A number of previous studies suggest that the coding of objects in high-level visual cortex may reflect subjective perceptual experience of shapes. For instance, LOC adapts across changes in low-level physical stimulus properties that leave perceived shape unaltered, but not across changes that affect perceived shape [6,7]. Furthermore, the fMRI signal in LOC tracks recognition performance more accurately than activation in retinotopic cortex [5,8], and both IT neurons and the fMRI signal in LOC reflect the perceptual similarity of stimuli [8,9]. Finally, Kayaert et al. [10,11] found that IT cells are more strongly modulated by perceptually salient stimulus changes (nonaccidental properties) than by metric changes of equal physical magnitude.
FMRI studies of visual processing have traditionally focused on mean activation levels, looking for brain regions showing a difference in activation between different stimulus conditions. More recent studies, in contrast, have illustrated the importance of the distributed pattern of activation in representing information about stimulus conditions [12–14]. Haxby et al. [12] first showed that even when there is no difference in the mean activation levels of specific conditions across occipitotemporal cortex, object category can still be determined from the distributed pattern of activation using a correlation method. Recently, Williams et al. [8] demonstrated that activation patterns contain object-specific information only on trials where recognition is successful. This finding raises the question whether activation patterns contain detailed information about subjective visual experience.
We used a combination of human fMRI and psychophysics to test the hypothesis that distributed activation patterns in LOC reflect perceived shape. We created a novel artificial shape space, in which physical similarity was controlled by gradual, parametric changes in aspect ratio and skew. Perceptual similarity was measured by psychophysical discrimination performance between the shapes, and neural similarity was measured by the correlations between the fMRI activation patterns of these shapes in LOC. (Note that we use the term “neural” to refer to fMRI activation patterns because of the high correlation between the BOLD signal and neuronal activity [15]) We found significant correlations between neural and perceptual similarity measures in LOC. Interestingly, this finding was restricted to the anterior portion of LOC (pFs); in the posterior portion (LO), we found significant correlations between neural and physical similarity measures. In addition, neural similarities were consistent across participants in posterior LOC, but highly variable across participants in anterior LOC. Together, these results suggest that object representations in posterior LOC reflect the physical stimulus, while representations in anterior LOC reflect subjective shape experience.
We created a novel artificial stimulus space consisting of four complex objects (Figure 1A). Each stimulus had four radially arranged protrusions (two half-parabolas joined at the vertices), which varied parametrically in aspect ratio and skew across stimuli. This stimulus space had five important features. First, IT cortex contains cells that are tuned to aspect ratio and skew, independently of one another [16]; thus, our stimuli varied along dimensions likely to be relevant in object-selective areas. Second, the four shapes used in the experiment were equidistant in aspect ratio and skew, and we thereby controlled important aspects of the their physical similarity; we refer to these aspects of similarity as “physical similarity,” while noting that other definitions of physical similarity are possible (see below for an analysis using a V1-like measure of similarity). Third, the stimuli were novel, allowing us to investigate shape similarities without confounds from semantic or learned associations. Fourth, spatial fMRI activation patterns in LOC have recently been shown to contain information sufficient for discrimination of such novel shapes [8,14,17]. Finally, the stimuli were chosen such that perceptual similarities correlated somewhat with physical similarity, but not perfectly, leaving room for the neural similarities to correlate, e.g., with perceptual similarity without necessarily also correlating with physical similarity, and vice-versa.
To study the relationship between perceptual, physical, and neural similarities in pFs and LO, we obtained three similarity measures for each pair of stimuli as follows. First, for each of the six possible pairs of nonidentical stimuli, physical similarity was measured by the inverse pairwise distances of the four shapes in the aspect ratio/skew space (Figure 1B, top right panel). As pointed out above, aspect ratio and skew were chosen because these dimensions are thought to be of relevance in high-level visual cortex [16]. Since the four stimuli formed a continuum with equal distances between adjacent stimuli, the six possible pairs of the four stimuli had distances of 1, 1, 1, 2, 2, and 3 steps; these distances were converted to similarities by inverting their values, to yield the similarity values 3, 3, 3, 2, 2, and 1 (see below for a different measure of physical similarity).
Second, to obtain a measure of perceptual shape similarity, we conducted a separate behavioral experiment outside the scanner with the same participants. On each trial, two shapes were shown in succession, and participants responded whether the two shapes were identical or different. Each shape was shown for 17 ms, with a forward and a backward mask of 50 ms each (without gaps between stimulus and masks), followed by a 1,500 ms response period. The proportion of trials on which a particular participant responded “identical” to a pair of stimuli that were in fact different was used as a measure of the perceptual similarity of that pair of stimuli. An example of a perceptual similarity matrix is shown in Figure 1B (top left panel). Note that we use the confusion rate merely as a proxy for true, first-person perceptual similarity, and do not wish to argue strongly that two stimuli that are confused with high probability necessarily also have highly similar qualia. Our definition of perceptual similarity is therefore merely an operational one in the context and for the practical purposes of this experiment.
Finally, to obtain a measure of neural similarity, we scanned the brains of eight participants using fMRI. Since we had a specific hypothesis about neural coding in object-selective cortex, we first identified the human object-selective region LOC in an independent localizer scan, using the standard comparison of intact versus scrambled everyday objects [2] (p < 10−4). LOC can be subdivided into a posterior portion, LO, on the lateral surface of occipitotemporal cortex; and an anterior portion, pFs, on the fusiform gyrus of the temporal lobe [18]. These two anatomically distinct portions of LOC were defined as separate regions of interest (ROIs). The ROI approach is of advantage because it is not subject to multiple comparisons problems.
In separate scans, we presented participants with the four shapes, using an event-related design. Each stimulus was shown for 300 ms, followed by a blank period of 1,700 ms, during which the participants had to respond whether the current stimulus was identical to that on the previous trial (one-back task). The purpose of this task was to keep participants' attention focused on the stimuli. We then extracted the spatially distributed activation patterns of each individual stimulus from the two LOC ROIs, on a voxel-by-voxel basis. Thus, in each participant and for each ROI, we obtained four vectors, each representing the voxelwise activation pattern of one particular shape in that ROI. We then computed the correlations between each of the six pairs of activation patterns for nonidentical stimuli, separately for pFs and LO. This resulted in six correlation coefficients for each ROI, one for each possible pair of the four shapes. An example of a neural similarity matrix is shown in Figure 1B (bottom panels).
Thus, we obtained physical, perceptual, and neural similarity measures for each possible pair of stimuli. We next compared these six-element similarity matrices to one another, by computing their correlation coefficients within participants and ROIs (Figure 1B). A high correlation between, e.g., the perceptual similarity matrix and the neural similarity matrix in a given ROI would indicate that if two stimuli are similar perceptually, they are also similar neurally in that ROI, i.e., their neural activation patterns are highly correlated with one another. Our hypothesis predicted that neural and perceptual similarity should be correlated in LOC.
The results confirmed this hypothesis, with an interesting twist. Neural and perceptual similarities were positively correlated in pFs, with an average correlation of 0.35 across participants, whereas in LO the average neural-perceptual correlation was only 0.001 (Figure 2). Conversely, neural and physical similarities were strongly positively correlated in LO (average correlation 0.41), but much more weakly in pFs (average correlation 0.10; Figure 2).
To quantify these results, we initially applied the Fisher z transformation to all correlation coefficients. This method transforms the non-normally distributed correlation coefficients into normally distributed variables, which allows the use of standard analysis of variance methods [19] (for details, see Materials and Methods). Statistical analysis after Fisher z transformation confirmed that across participants, the correlation coefficients between neural and perceptual similarities were significantly greater than zero in pFs (t(5) = 5.66, p < 0.001), but not greater than zero in LO (t(5) = 0.10, p = 0.38). Conversely, the correlations between neural and physical similarities were significantly greater than zero in LO (t(5) = 2.66, p < 0.05), but not in pFs (t(5) = 0.70, p = 0.29). A two-way analysis of variance (ANOVA) with region of interest (pFs versus LO) and correlation type (neural-perceptual versus neural-physical) as factors revealed a significant interaction of ROI and correlation type (F(1,5) = 13.79, p < 0.005), confirming the dissociation between these ROIs: neural pattern similarities in pFs correspond to subjective shape similarities, while neural pattern similarities in LO correspond to physical shape similarities.
These results were not due to differential mean signal levels for any of our stimuli, for two reasons: first, the correlation analysis does not take into account mean levels of activation; second, there were no differences in mean signal between the four stimuli in either region of interest (pFs: F(3,21) = 1.86, p = 0.17; LO: F(3,21) = 0.17, p = 0.92). Moreover, these results cannot be due to task performance, since critically the perceptual similarity measure was obtained in a separate testing session, while in the scanner participants performed an easy one-back task. (A control analysis for potential effects due to this task is reported below.)
As a further test of this finding, we speculated that if neural similarities in pFs reflect subjective perceptual similarities, the correspondence of neural similarities across participants in this region might be low: if a given pair of stimuli is neurally similar in pFs in one participant, the same pair may be neurally different in another participant whose subjective percept is different. Conversely, if neural similarities in LO reflect physical similarities, they should not differ greatly across participants. In other words, in pFs we would expect low inter-participant reliability of the neural similarities, whereas in LO we would expect high inter-participant reliability. To test this hypothesis, we correlated the neural similarity matrices of all individual participants with one another, separately for each ROI. This resulted in two 8 × 8 matrices, where each cell represents the correlation between the neural similarity matrices of two individual participants in one ROI (Figure 3). A high correlation in a given cell indicates that in these two participants, the stimulus pairs that are neurally similar in one participant are also neurally similar in the other participant.
As predicted, inter-participant reliability was low in pFs (mean across-participant correlation: 0.07; not different from zero, t(25) = 0.92, p = 0.26; Figure 3), but high in LO (mean: 0.42; greater than zero, t(25) = 7.83, p < 0.00000005; Figure 3). The difference between these two ROIs was significant (t(25) = −2.80, p < 0.05).
Note that this analysis is independent of the results described above: whether the neural similarities in pFs correlate across participants (as tested here) does not depend on whether they correlate with the behavioral similarities within participants (tested above).
To test whether the results described above are specific to object-selective cortex, we defined a set of further regions of interest: a retinotopic ROI based on activation at the occipital pole during the localizer task, as described before [8]; the fusiform face area (FFA [20]; see also [7]), and the occipital face area (OFA; [21,22]), based on the standard functional contrast of faces against objects (p < 10−4); and the parahippocampal place area, PPA [23], based on the standard contrast of scenes against objects (p < 10−4). In none of these regions did the neural similarities exhibit significant correlations with either perceptual or physical similarity (Figure 4A). Moreover, we found no significant inter-participant reliability in any of these regions (Figure 4B). With the exception of PPA, each of these regions contained at least as many voxels as pFs, ruling out the possibility that this finding is due to an inability to detect a correlation in small datasets. However, this possibility remains for PPA, which contained significantly fewer voxels than pFs. Note that FFA did not overlap with pFs in any of our participants.
The psychophysical same-different task used outside the scanner to obtain a measure of perceptual similarity was made as difficult as possible by presenting stimuli for extremely short durations (17 ms each) and using both forward and backward high-energy noise masks (50 ms each). Nevertheless, the performance level was high, with an average of 93% ± 1% correct performance on the same-different task. However, this level of performance corresponds to an average of 44 ± 9 errors over the course of the psychophysical experiment; this number of errors, distributed over the four stimuli used, proved sufficient to obtain a reliable measure of perceptual similarity. In support of this claim, the inter-participant reliability of perceptual similarity was on average r = 0.28 across participants (different from zero: t(25)=2.77, p < 0.05), showing that pairs of stimuli that a particular participant confused with high probability were also perceptually similar for the other participants. However, at the same time the fact that this correlation was not extremely high leaves room for subjective, observer-specific patterns of perceptual similarities.
The average correlation between the physical and perceptual similarity measures across participants was r = 0.38; this value was significantly different from zero (t(5) = 2.75, p < 0.05). Thus, perceptual similarity correlated with physical similarity, but not perfectly, again leaving room for the neural similarity measures to correlate with either the perceptual or the physical similarities without necessarily also correlating with the other.
The one-back task participants performed in the scanner was designed purely to keep participants' attention focused on the stimuli and was very easy to perform; none of our participants made a single mistake on this task. However, to nevertheless control for any possible effects of performance in the scanner, we analyzed the reaction times during the task in the scanner as follows. Since the order of stimulus presentation was randomized using m-sequences, each stimulus was preceded equally often by each other stimulus; therefore, each of the six possible discriminations among the four stimuli entered the neural pattern an equal number of times, and differential performance on any of these conditions could therefore not influence the neural pattern. However, it is possible that the task was easier for some stimuli in general; such a difference would enter the neural pattern of those stimuli and could thus potentially bias our results. Indeed, we observed a significant difference in the reaction times of stimuli 1 and 3 (mean reaction time [RT], stimulus 1: 602 ms ± 13 ms; stimulus 3: 624 ms ± 14 ms; t(7) = 8.11, p < 0.001), and stimuli 1 and 2 (mean RT, stimulus 2: 619 ms ± 17 ms; t(7) = 2.88, p < 0.05). To assess the effect of this difference, we used the reaction time differences among the four stimuli during the task in the scanner as a new behavioral similarity measure within each participant; e.g., if a particular participant took an average of 635 ms to respond to stimulus 1, and 649 ms for stimulus 2, the new behavioral similarity of these stimuli would be the negative reaction time difference, i.e. −14 ms (negative to turn the measure from a “difference” into a “similarity” measure). We then re-computed the correlations between the neural similarities and this new behavioral similarity measure. None of the resulting correlations were significant in any of our ROIs, although there was a nonsignificant trend towards a correlation between neural similarities and the new reaction time similarities in LO (mean r = 0.33, t(5) = 1.75, p = 0.09).
As pointed out above, it is conceivable in principle that differential numbers of voxels in different ROIs affect the likelihood of detecting correlations. Indeed, LO contained more voxels on average than pFs (pFs: mean 127 ± 36 voxels, LO: mean 287 ± 107 voxels). We therefore wished to test whether the results described above depend on the number of voxels in each ROI. To this end, we conducted a control in which we randomly excluded 50% of the voxels of each ROI. This procedure was repeated 100 times, and an average correlation estimate was obtained by averaging over the 100 bootstrapping iterations. The results were the same as in the main analysis reported above (Figures 5 and 6): the average correlation between neural and perceptual similarities in pFs was 0.25 (different from zero: t(5) = 6.32, p < 0.001; Figure 5), but that between neural and physical similarities in this region was only 0.05 (not different from zero: t(5) = 0.56, p = 0.32); in contrast, LO exhibited a significant correlation between neural and physical similarities (mean r = 0.32, t(5) = 2.33, p < 0.05; Figure 6), but not between neural and perceptual similarities (mean r = 0.06, t(5) = 0.56, p = 0.32). The interaction was again significant (F(1,5) = 12.00, p < 0.005). Similarly, the inter-participant reliability was again high in LO (mean r = 0.21, t(25) = 2.82, p < 0.05), but low in pFs (mean r = 0.04, t(25) = 0.60, p = 0.33; Figure 6). The difference between LO and pFs was significant (t(25) = 2.04, p = 0.05). Note, however, that subsampling reduced the inter-participant reliability in LO by a factor of one-half (mean r = 0.42 to mean r = 0.21). In light of this change, and the fact that the average sizes of pFs and LO differed by a factor greater than two, we repeated the subsampling for the inter-participant reliability analysis using not 50% of voxels, but instead equalizing voxel numbers across the two ROIs. Specifically, we excluded random subsets of voxels from the larger ROI, until its size matched that of the smaller ROI, again with 100-fold bootstrapping. The results were comparable to that of the initial analysis: the inter-participant reliability was high in LO (mean r = 0.31, t(25) = 7.79, p < 0.0001), but low in pFs (mean r = 0.07, t(25) = 0.91, p = 0.26), with a significant difference between LO and pFs (t(25) = 2.65, p < 0.05). Thus, our results are independent of the size of our ROIs.
The physical similarity measure reported above was based on the distances of the stimuli from each other in terms of aspect ratio and skew parameters. The high correlation of these aspect ratio/skew distances with neural similarities in LO is consistent with the proposal that LO encodes stimuli in terms of aspect ratio and skew, as has been reported previously for high-level visual cortex in monkeys [16,24]. This finding suggests that LO might no longer correlate with physical similarity if it was defined in a different fashion. As a test of this hypothesis, we replaced the aspect ratio/skew distance measure with an alternative physical similarity measure designed to mimic the properties of area V1: the images were convolved with a set of Gabor filters with orientation and spatial frequency selectivities similar to those found in V1 [25] (see Materials and Methods); the resulting filtered images were then compared for pixelwise similarity. The resulting mean physical similarity matrix correlated well (r = 0.67) with the physical similarity measure reported above, i.e., closeness of the stimuli in parameter space. However, the neural similarities of pFs and LO showed no correlation with this V1-type physical similarity measure: in pFs, the mean correlation of the neural similarities with the mean Gabor similarity measure was r = 0.06 (not different from zero: t(5) = 0.42, p = 0.34); in LO, it was r = 0.07 (t(5) = 0.45, p = 0.34). In addition, we repeated this analysis for each individual Gabor filter (4 orientations × 5 spatial frequencies); none of the resulting 20 correlations between neural and physical similarities were significantly different from zero across participants in either pFs (correlations ranging from −0.09 to 0.10, none significant across participants) or LO (correlations ranging from −0.11 to 0.16, none significant across participants). Thus, the correlation of neural and physical similarities in LO appears to be specific to the case when the physical similarities are described in terms of aspect ratio and skew [16].
The neural patterns in retinotopic cortex only showed a weak correlation with the physical stimulus distances based on this V1-type similarity measure; the mean correlation between neural and Gabor similarities in retinotopic cortex was r = 0.15, which did not differ from zero across participants (t(5) = 0.76, p = 0.27). Moreover, none of the correlations were significant when the individual physical similarity matrices resulting from each of the 20 Gabor filters were correlated one-by-one with neural similarities (correlations ranging from 0.12 to 0.17, none significant across participants). This result is probably due to the fact that the images were presented with a random jitter of ∼2 degrees during scanning, which likely resulted in sufficiently nonoverlapping activations in retinotopic cortex to disrupt the neural similarity estimates in this region, and therefore also any correlation between neural and physical similarities. In support of this hypothesis, the stimuli were not distinguishable in retinotopic cortex using Haxby's pattern discrimination method [12] (mean percent correct discrimination: 44% ± 5%), while they were easily discriminable in pFs (66% ± 4% correct) and LO (65 ± 3% correct; for more details see next section).
The results reported above indicate that the neural patterns in our regions of interest contain fine-grained information about perceptual and physical stimulus similarity. These analyses were based on correlations between the neural activation patterns of pairs of stimuli; this measure controls for noise in the data because the formula for the correlation coefficient includes a division by the standard deviations of the data vectors. However, we additionally wished to confirm with a conventional analysis method that these neural patterns were indeed stable and contained information about stimulus identity. To this end, we applied the widely used technique of Haxby et al. [12]: we extracted the activation patterns separately for even and odd runs, and compared “within” and “between” correlations. The mean “within” correlations were 0.10 ± 0.05 and 0.04 ± 0.02 in pFs and LO, respectively, while the mean “between” correlations were 0.04 ± 0.02 and −0.001 ± 0.04 for pFs and LO, respectively. These correlations were low because we used an event-related design. Importantly, however, the “within” correlations were significantly higher than the “between” correlations, indicating that the patterns contained enough information to discriminate between same versus different stimuli (F(1,7) = 3.67, p < 0.05, two-way ANOVA with ROI and within/between as factors). As a further test of pattern discriminability, we computed the “Haxby Index” [8,12,13,26]. This index estimates classification performance between pairs of stimuli based on the within and between correlations, where 50% is chance performance and 100% is optimal performance. Discrimination performance was 66% ± 4% in pFs and 65% ± 3% in LO; these levels of performance were significantly above chance (pFs: t(7) = 4.20, p < 0.005; LO: t(7) = 4.55, p < 0.005). Thus, the patterns in both ROIs were stable enough across the split halves to successfully discriminate between our stimuli.
In sum, we have found that distributed activation patterns in human object-selective cortex contain information about the subjective perceptual similarities between complex visual stimuli. Specifically, we show a dissociation between neural coding of perceptual versus physical similarities within LOC: using independent measures of neural, perceptual, and physical similarity on our set of novel artificial shapes, we find that the neural similarities of shapes in anterior LOC (pFs) correlate with their perceptual similarities. Conversely, the neural similarities in posterior LOC (LO) correlate with the physical similarity of the shapes in the stimulus space. Furthermore, the agreement across participants of the neural similarities is high in LO, but low in pFs, consistent with a physically based representation in LO and a representation based on observer-specific subjective shape experience in pFs.
These results are specific to object-selective cortex, i.e., the regions LO and pFs; additional ROIs including retinotopic cortex, FFA, OFA, and PPA did not show significant correlations between either neural and perceptual or physical similarities. Moreover, the results did not depend on the number of voxels in each ROI.
Our findings confirm previous studies showing that object representations in the ventral stream reflect subjective perception [5–9,11,27], and extend them by showing that the distributed pattern of activation in LOC contains information about idiosyncratic perceptual similarities on a fine-grained scale [14].
Similarly, the finding that posterior LOC shows a correlation between neural and physical stimulus similarity confirms previous studies that have shown selectivity for physical shape features of moderate complexity in high-level visual cortex. In area V4, single-cell and fMRI studies have demonstrated tuning to contour curvature in monkeys [24,28,29] and selectivity for radial and concentric gratings [30] and intermediate-complexity object parts [31] in humans. Single cells in monkey IT cortex have been shown to be tuned to metric changes in simple geometrical shapes [10,16] and to particular combinations of simple shapes [1]. Moreover, IT responses are sensitive to low-level visual properties such as object size, position, and viewpoint [32–34].
Putting the results from pFs and LO together, our findings are consistent with previous evidence regarding an anterior-posterior functional subdivision within LOC: Grill-Spector et al. [35] showed that pFs exhibits more location- and size-invariance than LO; Lerner et al. [36,37] found that pFs was more vulnerable to object scrambling than LO; and Kourtzi et al. [38] showed that pFs does not adapt across changes that alter an object's subjective appearance (convex versus concave), while LO does. Together, these studies suggest that object representations in pFs are more high-level, abstract, and closer to subjective perception than those in LO. Our results substantiate this claim by showing directly that neural similarities correlate with perceptual similarities in pFs but not LO, while neural similarities correlate with physical similarities in LO but not pFs. In contrast to these previous studies, our experiment shows the correspondence between perceptual and neural similarities directly and within individual participants, by using participant-specific measures of perceptual and neural similarity. Furthermore, by computing these neural similarities on the activation pattern across individual voxels, we obtain a richer and more informative measure of neural similarity than can be achieved by averaging the activation across the entire ROI [12].
However, it should be noted that other recent studies have found evidence highlighting the informativeness and behavioral relevance of neural activation patterns in LO: Eger et al. [14] showed that support vector classification within categories was better in LO than pFs; Williams et al. [8] found that correct versus incorrect recognition was reflected in the activation patterns of LO but not pFs. Thus, pFs may not always be the seat of conscious shape perception; instead, the cortical regions whose representations are most closely associated to subjective shape perception may vary with the stimulus, task, and viewing conditions [8].
Three previous studies have shown correlations between pattern information and subjective perception in object-selective cortex. First, Edelman et al. [9] used multi-dimensional scaling (MDS) to uncover the perceptual and neural similarities of a set of categorized stimuli, and found that both the behavioral and the neural measures of similarity followed the stimulus category boundaries (e.g., four-legged animals versus cars). However, their stimulus set consisted of photographs of familiar, every-day objects from a restricted set of categories; thus, it is unclear whether the correspondence between the neural and behavioral measures is due to low-level visual similarity of objects in the same category, category membership itself, or even matching semantic associations of stimuli in the same category. By using novel objects from a parametrized stimulus space we can disentangle perceptual from physical similarity. Moreover, we show a functional dissociation between the two subregions of LOC, which were not analyzed independently in the Edelman et al. [9] study.
Second, Op de Beeck et al. [27] showed in monkeys that the firing rates of neurons in IT cortex reflect perceptual similarities in a set of complex stimuli. The present experiment was inspired by this study, and we replicate its findings in a different species (human) and using a different technique (fMRI). In addition, we show that human object-selective cortex contains both physically based and perceptually based similarity metrics, organized in a posterior-anterior hierarchy.
Third, Williams et al. [8] used a pattern analysis approach similar to the one used here to show that the activation pattern in LOC contains information sufficient for stimulus discrimination only if the participant successfully categorizes the stimulus. This study is similar to ours in that it establishes a link between the distributed activation patterns in LOC and behavioral performance. However, the question it addresses is substantially different from ours: Williams et al. asked whether the information contained in the LOC activation patterns correlated with successful object categorization; in contrast, we ask whether the activation patterns in LOC reflect physical stimulus similarities or subjective perceptual similarities. This difference is also the likely cause for a further one, namely that Williams et al. find correspondence between neural patterns and task performance in LO, but not in pFs, while we show correlations between neural and perceptual similarities in pFs but not in LO (see also [14]). In addition, Williams et al. used visually highly distinctive stimulus categories, whereas the differences between our stimuli were more subtle. Together these differences suggest that coarse category membership [8] and more fine-grained similarity (this study) may be neurally distinguishable at the level of LOC.
In conclusion, our results indicate that object shape may be coded in terms of physical features in posterior LOC, and in terms of subjective shape experience in anterior LOC. Of course, this claim simply replaces one puzzle with another: “What is the neural code for object shape?” is transformed into the equally difficult question, “What are the determinants of subjective shape experience?” However, the advantage of this new question is that it has been the subject of intensive study since the time of the Gestalt psychologists [39], and the accumulated evidence is a rich source of new hypotheses about the neural code for object shape. Combining behavioral measures with fine-grained fMRI pattern analysis methods [8,12,13] may prove a powerful means of solving the puzzle of visual object recognition.
We recruited eight participants from the MIT Human Subject Pool. Each participant was compensated US$60. The study was approved by the MIT Committee on the Use of Humans as Experimental Subjects (COUHES). All participants gave informed consent.
Localizer scans: The LOC was localized as the region that responded more strongly to grayscale images of intact objects than to images of scrambled objects (p < 10−4), as described previously [2,6]. The FFA [20] and the OFA [21,22] were defined as the regions responding more to faces than objects (p < 10−4). The PPA [23] was defined as the region responding more to scenes than objects (p < 10−4). The retinotopic ROI was defined based on activation at the occipital pole in a contrast between all stimulus conditions versus baseline in the localizer scans (p < 10−4; [8]).
Experimental scans and behavioral experiment: Four novel stimuli, each measuring 10 degrees across, were used for the experimental scans. The use of novel stimuli ensured that correlations were not due to semantic associations with the stimuli; this was a potential confound in previous pattern similarity studies [9]. Furthermore, we wished to use shape features that are likely to be encoded in object-selective cortex. Single-cell studies have shown that aspect ratio and skew are two such features [10,16]; we therefore created our stimuli based on parametric changes in aspect ratio and skew. Specifically, each stimulus consisted of four protrusions arranged radially around a central disk. Each protrusion was composed of two adjoining half-parabolas of the form y = a xn. The parameters a and n could be used to vary the skew and aspect ratio of each protrusion parametrically. In doing so, the total area of the stimuli was always kept constant to avoid low-level confounds. We defined aspect ratio as the ratio of the height to the base width of each protrusion, and skew as the position of the vertex with respect to the center of the base; for instance, 0% skew indicates a vertex directly above the center of the base, skew of 100% indicates a vertex directly above the right end of the base, and skew of −100% indicates a vertex directly above the left end of the base. From the left to the right end of the stimulus spectrum, the aspect ratio of the second and fourth protrusions (counting clockwise, beginning at 12 o'clock) decreased by 1.4 and 1.6 on each morph step, respectively; the aspect ratio of the first and third protrusions was fixed. For skew, the first, second, and fourth protrusions moved counterclockwise by 60%, 24%, and 24% for each step, respectively (where a cumulative skew change greater than 100% simply meant moving the vertex of the protrusion beyond its base); the skew of the third protrusion changed in the clockwise direction by 25% on each step. Thus, the four stimuli used were equidistant in terms of aspect ratio and skew, forming a straight line in the stimulus space. The magnitude of the parametric distances between the stimuli was chosen based on informal testing to be at the same time discriminable and not too obvious. The stimuli were filled with random dots, with a mean luminance of 50%, to ensure activation throughout the ventral visual stream. In addition, a chair and a face were included in the stimulus set, to prevent adaptation in ventral visual cortex due to the high similarity among the novel shapes.
fMRI experiment: Each participant was run in one session of about 2 h, consisting of eight experimental scans and four LOC localizer scans. Stimuli were presented using the Psychophysics Toolbox [40] and Matlab (Mathworks).
The localizer scans were run as described previously [6,20,23].
The experimental scans were event-related, and each scan contained 144 stimulus trials and 36 fixation trials. On each trial, one of the six possible stimuli (four novel shapes, one face, one chair) was presented at the center of the screen for 300 ms, followed by a 1,700-ms response period during which participants indicated whether the current stimulus was identical or different from the previous one. The purpose of this task was to keep participants' attention focused on the stimuli. The order of stimulus presentation was optimized using m-sequences (Optseq). Each stimulus occurred 24 times per scan, resulting in a total of 192 times for the whole experiment.
Behavioral experiment: In a separate behavioral session that followed the fMRI experiment with a delay of at least 1 wk, each of the original participants performed a same-different task on pairs of the same four shapes, plus the face and the chair, that were presented in the fMRI experiment. On each trial, two stimuli were shown sequentially, for 17 ms each, with a forward and a backward mask (consisting of a full screen noise field of random letters with high density and overlap) of 50 ms each, followed by a 1,500-ms response period. Perceptual similarities were obtained by computing the proportion of trials on which a particular pair of different stimuli was erroneously considered “identical”. Participants performed 630 trials total. Each stimulus appeared with equal probability on each trial, and with equal probability as the first and second stimulus of each pair.
fMRI scanning was performed on a 3T Siemens Trio Scanner (Siemens) at the Athinoula A. Martinos Center for Biomedical Imaging at the McGovern Institute for Brain Research at MIT. A Gradient Echo single-shot pulse sequence was used (TR = 2 s; TE = 30 ms). Twenty-five slices were collected with a 12-channel head coil. Slices were oriented roughly perpendicular to the calcarine sulcus and covered most of the occipital and posterior temporal lobes, as well as some of the inferior parietal lobes. Slices were 2 mm thick, with a 10% gap, and had an in-plane resolution of 1.6 × 1.6 mm.
Data analysis was performed using FS-FAST (http://surfer.nmr.mgh.harvard.edu), fROI (http://froi.sourceforge.net), and custom-written software. Before statistical analysis, images were motion corrected [41], and the data from the blocked localizer scans (not the event-related scans) were smoothed (3 mm full width at half maximum Gaussian kernel).
The LOC was defined as the set of contiguous voxels in the central occipitotemporal cortex that showed significantly stronger activation (p < 10−4, uncorrected) to intact objects than to scrambled versions of the same objects [2]. Two subregions of LOC were defined as ROIs, as described previously [42]: a posterior portion, LO, on the lateral surface of occipitotemporal cortex; and an anterior portion, pFs, on the fusiform gyrus of the temporal lobe [18]. Furthermore, we defined four control regions of interest: the FFA [20] and the OFA [21,22], based on the standard functional contrast of faces against objects (p < 10−4); the PPA [23], based on the standard contrast of scenes against objects (p < 10−4); and a retinotopic ROI based on activation at the occipital pole in a contrast between all stimulus conditions versus baseline in the localizer scans (p < 10−4; [8]). The FFA did not overlap with pFs in any of our participants, as is sometimes the case.
For the blocked localizer scans, statistical maps were calculated by correlating the signal time course with a gamma function (delta = 2.25, tau = 1.25) for each voxel convolved with the block timecourse. For the event-related scans, the hemodynamic response was extracted using a deconvolution analysis, without any assumptions about the shape of the response. The peaks of the fMRI responses of each of the four novel shapes were extracted from each ROI, for all voxels separately. This resulted in four patterns per ROI, each representing the distributed activation pattern to a particular stimulus in that ROI. Neural similarities were obtained by computing the Pearson correlation coefficient between these patterns, as described above.
To assess the statistical significance of the correlation matrices results, we first applied the Fisher z transformation to the data and then performed t-tests and ANOVAs. This transformation is necessary because correlation coefficients do not follow a normal distribution, and are therefore strictly not amenable to analysis of variance statistics [19]. The Fisher z transformation converts correlation coefficients into normally distributed variables and thereby makes t-tests and ANOVAs possible. Given a correlation r, the Fisher z is given by
We took care to use the standard error formula specific to the Fisher z:
This formula has fewer degrees of freedom and is therefore more conservative than the conventional standard error.
To obtain a V1-like physical similarity measure, we applied a set of Gabor filters to the images and then computed pixelwise similarities between the images. This analysis was motivated by the fact that V1 cells exhibit tuning profiles that are well-described by Gabor filters [25,43]. Each image was convolved with Gabors of four different orientations (0°, 45°, 90°, and 135°; these orientations cover the whole unit circle because of the symmetry of the Gabors) and five different spatial frequencies (2, 4, 6, 8, and 10 cycles per degree). These parameters are representative of the tuning properties found in early visual cortex [25]. We then correlated the resulting physical similarity matrices with the neural similarities from our regions of interest, as described above; this was done both for the average across the V1-like similarities, as well as the individual matrices. The average V1-like physical similarity matrix correlated well (r = 0.67) with our other physical similarity measure, i.e. distance of the stimuli in parameter space.
|
10.1371/journal.ppat.1005620 | Ecosystem Interactions Underlie the Spread of Avian Influenza A Viruses with Pandemic Potential | Despite evidence for avian influenza A virus (AIV) transmission between wild and domestic ecosystems, the roles of bird migration and poultry trade in the spread of viruses remain enigmatic. In this study, we integrate ecosystem interactions into a phylogeographic model to assess the contribution of wild and domestic hosts to AIV distribution and persistence. Analysis of globally sampled AIV datasets shows frequent two-way transmission between wild and domestic ecosystems. In general, viral flow from domestic to wild bird populations was restricted to within a geographic region. In contrast, spillover from wild to domestic populations occurred both within and between regions. Wild birds mediated long-distance dispersal at intercontinental scales whereas viral spread among poultry populations was a major driver of regional spread. Viral spread between poultry flocks frequently originated from persistent lineages circulating in regions of intensive poultry production. Our analysis of long-term surveillance data demonstrates that meaningful insights can be inferred from integrating ecosystem into phylogeographic reconstructions that may be consequential for pandemic preparedness and livestock protection.
| It is assumed that AIV outbreaks in poultry are introduced from wild birds. To test this, we incorporated ecosystem and location of isolation into a comparative genetic analysis. We show high rates of viral transmission from domestic to wild birds within a region and, that wild birds could transmit AIV to poultry between regions. However, the highest rates of viral flow between regions was among domestic populations, indicating poultry trade may play a major role in spreading viral populations. We demonstrate that interactions between migratory birds and animal productions systems contribute to the emergence of AIV. The assumption of unidirectional viral flow from wild birds to domestic poultry provides an incomplete picture of influenza ecology and may confound control efforts.
| Intensive agriculture has allowed AIV circulating in wild bird populations and multi-host poultry systems (domestic food birds including chicken, duck, goose, pigeon) to interact, shaping the diversity of subtypes with pandemic potential [1]. The recently emerged H7N9 viruses containing H9N2-origin internal genes highlight that co-circulation of subtypes concealed within poultry systems can enhance the pandemic threat of influenza [2]. Such interactions are not unique. Over the last decade, H9N2 viruses have donated gene segments to several virus subtypes infecting poultry and humans, including the highly pathogenic avian influenza (HPAI) H5N1 panzootic that emerged in 2003 and persists to the present day [3–5]. Transmissions across the wild-domestic bird interface and genomic reshuffling within poultry have contributed to the emergence, spread and persistence of novel H5, H6, H7, and H9 AIV genotypes, which have caused human infection [4, 6, 7]. Despite the importance of viral transmission between natural and domestic systems, the role of these interactions in determining viral diversity and distribution has not been adequately studied.
AIV transmission between wild reservoirs and domestic animals takes place where natural and agricultural ecosystems overlap, a scenario that occurs worldwide. For example, transmission between wild and domestic birds led to H6 outbreaks in Californian poultry [8]; low pathogenic (LPAI) H5 viruses from wild birds in Italy were linked to poultry disease in Asia [7, 9]; and wild bird-origin H9 viruses circulating in Korea later emerged in domestic flocks [10]. Most recently, H5 viruses detected in East Asia have spread to European and North American poultry, consistent with intercontinental migration of wild birds [11]. Although often detected in domestic birds, viral communication between populations is not one-way. In 2005, HPAI H5N1 jumped from domestic birds to infect bar-headed geese (Anser indicus) at Qinghai Lake, China (12). This triggered large-scale wildlife die-offs, contributed to the global spread of HPAI H5N1 virus and placed considerable burdens on government resources to mitigate global spread [12, 13].
H9 viruses are endemic in terrestrial poultry in China, the Middle East, and Europe and occur globally in a diversity of wild bird taxa [14]. Although H9 in poultry is not a notifiable disease, it is listed by the World Health Organization as a candidate for the next global pandemic along with H5 and H7 [15]. Gene segment exchange with H9 has facilitated generation of novel reassortants, including the 1997 H5N1 strain that caused human infections and fatalities in Hong Kong [3]. Furthermore, periodic human infections with H9 subtype viruses have led to intensive epidemiological surveillance of wild and domestic birds to identify the infection source [3, 6, 16, 17]. These events have prompted regular collection and sequencing of H9 subtype viruses providing a globally sampled genetic dataset where the contributions of wild and domestic animals to the global spread can be modelled using phylogeographic approaches.
Despite an expanding AIV host range, the role of ecosystem interactions (i.e. transmission between wild and domestic animals) in generating and spreading novel influenza viruses over large distances remain poorly characterized. In this study we integrate the geographical and ecological context of transmission into a statistical phylogenetic model. Using H9, H3 and H6 sequence data, including newly acquired sequences from North American wild birds, we characterize the contribution of wild and domestic birds to the global distribution of AIV and estimate the role of inter-ecosystem dynamics on viral diversity and persistence. We hypothesize that viruses circulating among wild migratory birds may connect a global network of influenza outbreaks that occur locally in domestic birds. To test this, we incorporated ecosystem and geographic location of isolation into a comparative genetic analysis of H9 hemagglutinin (HA) sequence data (see S1 Text and S1 Table). H9 subtype viruses are commonly isolated from domestic poultry and wild birds [3, 7, 18–24]. Periodic human infections have resulted in intensive surveillance of domestic birds in order to identify the sources of infection [16, 17]. As a result, a robust dataset of H9 subtype viruses is available from both domestic and wild birds throughout the world. Since poultry and wild birds infected with H9 viruses do not often manifest major disease symptoms, vaccination or active control efforts are limited for this subtype. We utilize the H9 virus surveillance and sequence reporting, enhanced with novel sequences acquired from North American wild bird surveillance collected between 1974 and 2013. We build on previous phylogeographic methods [25] to integrate wild and domestic ecosystems in order to 1) estimate an asymmetric spatial transition matrix; 2) model viral spread among wild or domestic populations; and 3) assess the relative risk of virus emergence from discrete locations and ecosystems. We extend our analysis to other avian influenza A virus subtypes (H3 and H6) in order to assess if inferences made from comparative genetic analysis of H9 gene sequences could be generalized.
Poultry production density and disease surveillance effort vary considerably between countries. We therefore incorporated this information in defining our discrete geographic units for our migration model. Mapping all available poultry and wild bird virus isolates allows for a qualitative assessment showing that global sampling sites correspond well with the intensity and distribution of poultry production systems, particularly in China and the Middle East (Fig 1A–1C). Samples from wild and domestic avian hosts spatially overlap. It is possible that surveillance in wild birds was in response to AIV detection in domestic birds. There are few long-term virological surveillance programs and most detection was based on opportunistic sampling [2,4,6–8,10,11]. As a result, there are many more isolates from domestic birds than wild birds, despite the importance of these hosts in the ecology, evolution and emergence of AIV [11]. Sampled sites similarly overlap with areas of intense duck production, although domestic duck rearing is far less common than chickens worldwide (S1 Fig). While we assume that the virus behaves the same in all poultry hosts, experimental data indicates this may be a poor assumption [18]. Unfortunately, limited information on virus prevalence or epidemiology in various domestic host species between countries makes it difficult to treat individual species separately, thereby necessitating the grouping used here.
We assume that if AIV is transmitted between domestic populations it is likely correlated with live poultry trade between countries (Fig 1B). We use this framework to interpret our phylogeographic reconstruction results where, we believe that the movement of viruses between discrete states is limited to the movement of live animals, whether through trade or wild animal migration. It is possible that transmission between regions could be linked to trade of animal product or other cryptic means, but for this study we have not considered other mechanisms of viral spread. According to official trade data available from the Food and Agriculture Organization of the United Nations (UN-FAO), intercontinental movement of live poultry primarily involves exports from Europe and North America into the eastern hemisphere: Middle East, China and Southeast Asia (Fig 1B). Regional flow of animals throughout Europe showed that few countries (i.e. Germany, Netherlands) were the main importers of live animals from other European nations and were also exporters to the Middle East and Southeast Asia. The majority of poultry production in China is for domestic consumption, although trade with Southeast and East Asia (Japan/South Korea) accounted for the bulk of live animal exports (Fig 1B). After the emergence and spread of HPAI H5N1, there was a marked decline in poultry exports with the majority of official trade restricted to supplying Hong Kong and Macau. Detailed trade flow and animal population data led us to subdivide China into 3 distinct regions (East, Central and West China) based on production and consumption, and to combine Japan and South Korea into a discrete isolated (particularly after 2004) geo-region relative to China (Fig 1B and 1C). Based on the distribution of available AIV sequences from poultry and global patterns of live poultry trade, we therefore identified 9 discrete regions for which viral migration patterns could be estimated (Fig 1C, S2 Fig and S2 Table).
Wild bird surveillance is often conducted to identify potential virus sources following outbreaks in poultry [26]. Consequently, large overlap between wild bird and poultry surveillance sites exists (Fig 1C and S2 Table). Notable exceptions include long-term influenza surveillance studies of migratory birds in North America (Delaware Bay, Alberta) and Europe (Ottenby, Sweden) [27]. With the exception of waterfowl, most other wild bird species are opportunistically sampled rather than targeted [28]. Sampling inconsistencies across different species mean that sample sizes of virus isolates are not appropriate to infer rates of interspecies transmission among wild bird taxa (i.e. mallard, Anas platyrhynchos to northern pintail, A. acuta). Nevertheless, the data from wild birds was well suited for estimating viral flow between wild and domestic systems. In our analysis we assumed that the virus behaves similarly in all wild species and classified isolates as either ‘wild’ or ‘domestic’ to estimate viral transmission rate between populations.
Phylogenetic reconstruction of the H9-HA gene showed a mixture of North American and Eurasian wild bird isolates and two recently diverged poultry lineages (G1 and Ck/Bei lineages; Fig 2A and S3 Fig), consistent with previous studies [3, 29–31]. Our analysis demonstrated that the H9 lineage was younger and less geographically structured than other HA subtypes [8, 32, 33]. The estimated tree root age from sampled H9 strains revealed an origin between 1964 and 1975, suggesting a recent selective sweep removed genetic signals of long-term geographic structuring from the population. Periodic sweeps have been implicated in lasting changes in the viral population of multiple subtypes [33, 34].
We estimated the ancestral locations of isolates collected from the 9 discrete geographic regions defined above (Fig 2A) and stratified the observations as viruses collected from either wild or domestic birds. We found frequent viral flow among regions and identified East/Central China (Ck/Bei lineage) and the Middle East (G1 lineage) as the primary sources for H9 virus spread among poultry (Fig 2B and Table 1). We determined the number of transitions emerging from each ecosystem state from 1998 through 2013, and show that the viral population was primarily emerging from domestic populations with most transitions into other domestic populations (Fig 2B). Approximately 90% of all transitions emerged from the domestic populations. The mean waiting time in each state as a proportion of the total phylogenetic tree time show that the populations was primarily circulating among poultry located in East/Central China and the Middle East (Table 1). Date estimates suggested these two regional populations diverged around 1988 (95% Bayesian Credible Interval: BCI 1985–1990), coinciding with industrialization of poultry production in Asia [35]. Our results also suggest that both CK/Bei and G1 lineages emerged from independent wild bird introductions. Over the last decade these regions have acted as two distinct and persistent gene pools (Fig 2B and Table 1), reflecting the establishment of East Asia and the Middle East as independent centers of poultry production [14].
In our analysis we used a non-reversible model such that the direction of migration could be inferred to assess whether a region acted as a source or a sink from 1998 through 2013 [25, 36, 37]. For each location, we averaged the number of state changes observed/year for trees sampled from the posterior distribution. Even though East China was a persistent source population, the transmission was primarily limited to within China. East China primarily receives viruses from Central China and vice versa (Fig 2C and S4 Fig). Prior to 2003 there was some exchange with Japan. The rapid disruption in viral flow between these regions corresponded with restrictions on live poultry exports after the emergence and regional spread of highly pathogenic avian influenza H5N1 [35]. The Middle East region was primarily a sink for G1 lineage viruses originating in South Asia and Europe (S4 Fig). Transmission from the Middle East to South Asia and Europe, although rare, was evident in other lineages (S4 Fig).
Our results showed enhanced risk of viral emergence from East China, whereas Western China and Southeast Asia represent regions with enhanced risk of receiving H9 viruses (Table 2). By evaluating the relative risk for regional source/sink transmission from domestic or wild birds, we see enhanced risk for transmission originating from wild birds in Japan/South Korea and Europe. In contrast, wild birds from the Middle East, South Asia, East and Central China were more likely to receive viruses. There was enhanced risk for domestic birds in East China, South Asia and the Middle East to be a source for transmission of H9 viruses. Domestic birds in Japan/S Korea, Western China, Southeast Asia and Europe were more likely to be sinks (Table 2).
Domestic-to-wild virus communications were restricted to within regions, whereas wild-to-domestic transmissions occurred both within and between regions (Table 3). Two-way viral flow between ecosystems occurred within China and the Middle East, emphasizing the importance of these regions for ecosystem interactions (Table 3). Domestic-to-domestic transmission was between adjacent regions with no significant long distance migrations (Fig 3A). In contrast, supported wild-to-wild transmissions occurred over long and short distances (Fig 3A). Our results suggest that poultry systems have provided a persistent source for regional H9 spread, whereas wild bird-mediated dispersal provided a mechanism for both intercontinental (Japan/S Korea-North America and Japan/S Korea-Europe) and regional spread (East Asia-Southeast Asia) along known waterbird migratory routes [38, 39] (Fig 3A). Despite strong statistical support for these migration events, the importance of Japan and South Korea for long distance wild bird carriage of viruses is difficult to assess due to sparseness of wild bird sampling in this region (Fig 2A). We also found strong support for wild-bird mediated viral migration between North America and the Middle East (Table 3). In both locations, viruses were isolated from gulls and shorebirds. In our reconstruction there was more than 10 years of under-sampled diversity between viruses in gulls from the Republic of Georgia and the putative North American ancestor. It is unlikely that this was a direct transmission between populations. The most parsimonious explanation is that this virus transmitted among gulls during this period, implying that persistence in wild bird populations may contribute to the long-distance spread of H9 viruses.
Integrating ecosystem and region into our migration model provided the most comprehensive description of global H9 distribution and diversity. Domestic-to-domestic transitions among regions, especially between East China and Central China, were significantly faster than all other estimated migration patterns, suggesting poultry trade was likely responsible for spreading H9 subtype virus (Fig 3B and 3C). This trend was also reflected at the global scale, whereby migration rates among domestic birds was significantly higher than among wild birds and between ecosystems (wild-to-domestic or domestic-to-wild) (Fig 3).
We chose to extend our model with analysis of AIV subtypes H3 and H6 HA gene sequences. Similar diffusion patterns (i.e. ecosystem transition patterns or rates) between subtypes may suggest similar underlying processes. Analysis of H3 subtype HA gene sequences showed similar results to those described above for H9 (S5 Fig, S6 Fig and S3 Table). In contrast, the role of virus ecosystem interactions in determining viral distribution of H6 subtype viruses was less clear, even though similar transmission patterns were observed (S7 Fig, S4 Table). Despite evidence for two-way transmission of H6 viruses between wild or domestic populations (S8 Fig and S4 Table), our analysis shows no support for either ecosystem playing a larger role in the distribution of these viruses (see S1 Text).
Our model for ancestral state reconstruction was restricted to locations and ecosystems where viruses were observed, which results in an inherent bias in the reconstruction of migration or ecosystem transitions. For example, in our analysis of H9 subtype virus, wild bird isolates were not observed from western China and therefore, this state was not represented in our ancestral state reconstruction. While it is possible that the virus population did spend time circulating among wild birds in western China, the lack of contemporary observations imposes limitations on the ancestral state reconstruction approach taken in this study. Similarly, North American H9 isolates from domestic birds were not present in our reconstruction of migration patterns and ecosystem interactions. To investigate if our ancestral reconstruction was sensitive to sampling bias, we randomized the location assignments at the tips throughout the MCMC procedure to determine if the posterior transition rate estimate and root state probability converged on the expected prior under the sampling scenario we used, as well as randomly generated alternatives [37]. For each subtype analyzed the posterior empirical frequency converged to the prior root state probability for all sampling scenarios (S9 Fig). In addition, the ecosystem transition rate estimates converged on the prior expectation where all rates were approximately equal (S10, S11 and S12 Figs). Despite uneven sampling of domestic and wild populations these analyses suggest that signals of ecosystem/spatial structure in the data inform our estimates and were not biased by the sampling scheme.
Our findings highlight that transmission among domestic flocks drove the majority of H9 dispersal to adjacent regions. Generally, wild bird migrations provide opportunities for the widespread movement of viruses with pandemic potential, but this did not play a greater role in viral spread than transmission among poultry populations, despite frequent transmission between wild and domestic birds. The most significant contribution from wild birds to emerging AIV ecology is the redistribution of gene segments over large distances, thereby increasing biodiversity and creating opportunities for novel variants to emerge. Failure to control H9 outbreaks in domestic populations can therefore contribute to the emergence and spread of new influenza variants [20]. Our results suggest that viral flow between wild and domestic systems contributes to the persistence, and spread of AIV.
The continued circulation of AIV in domestic populations poses a public health risk [2]. In this study, we capitalized on the robust surveillance of H9 viruses among avian populations to elucidate the role of viral transmission between wild and domestic ecosystems in the global spread and persistence of AIV. Parallels exist between the mechanisms of H9N2 global dispersal uncovered here and other subtypes. Most notably, transmissions of HPAI H5N1, and recently HPAI H5N8, have been perpetuated and spread by both wild bird migration and domestic poultry trade [13, 40, 41]. The complex interactions characterizing these systems have contributed to the genetic diversity and widespread diffusion of these viruses throughout Asia, Africa, Europe, and most recently, North America [4, 11].
Similar models may be extended to understand the mechanisms of spread and emergence of other influenza subtypes, although the paucity of surveillance data may be a limiting factor. While analysis of H3 subtype virus HA gene sequences supported our findings that transitions among domestic populations may be driving the spread of influenza viruses, our analysis of H6 viruses was less clear. H6 subtype viruses were prevalent in domestic ducks in southern China [24]. While 75% of the global duck production (including reared wild ducks) occurs in China [42] production is primarily for domestic consumption, with Hong Kong as the largest importer of duck meat [35]. Even though the large-scale transmission patterns were similar across subtypes, the binning of discrete geographic states used in our analysis of H6 could not capture the majority of domestic trade within China. It is likely that alternative sampling strategies are necessary to investigate the role of ecosystems interactions and poultry trade in maintaining H6 subtype populations.
Production systems that promote the two-way transmission of viruses between wild and domestic avian hosts facilitate the generation of potentially pandemic AIV and may lead to widespread outbreaks that are difficult to contain. Surveillance programs focused on detecting highly pathogenic subtypes in symptomatic poultry falls short of identifying the mechanisms of emergence, spread and genomic reshuffling. Active systematic surveillance for AIV in both wild and domestic populations allow for the continued development of models needed to test the role of various species or populations in viral persistence. A limitation of this study is that reconstruction of viral movement patterns was limited to transmission among the populations surveyed. It is likely that unsampled populations play a role in the spread and persistence of AIV. Systematic surveillance programs are critical to assess the risk of disease emergence and spread of AIV by wild migratory birds. Analysis of long-term surveillance data enables meaningful insights necessary to develop appropriately informed predictive models. Inferences from such models are consequential for pandemic preparedness and livestock protection.
Using available census data for chicken and ducks we produced heat maps showing poultry production intensity. Mapping all available H9-HA sequence data on top of the heat maps allows us to visually assess the distribution of sampling and production regions. This assessment was used to determine appropriate geographic regions to incorporate into our model and identify regions with few samples. Data on the international trade of live poultry from 1995–2011 (the most current year of data available) was downloaded from the United Nations, Food & Agriculture Organization (UN-FAO: faostat3.fao.org, accessed June 11, 2015). The quantity of chickens and ducks traded was chosen as the metric to assess trade relationships. All years were summed to generate long-term estimates of international trade and countries were aggregated into 7 regions: Japan/South Korea, China, South East Asia, South Asia, Middle East, Europe and North America, broadly consistent with georegions used for the phylogeographic model. The quantity of exports and imports was compared for each region and only the maximum trade quantity linking two regions was recorded due to inconsistencies in data reported.
All available H9 influenza A HA gene sequences were downloaded from the Influenza Virus Resource database (http://www.ncbi.nlm.nih.gov/genomes/FLU/FLU.html) on March 30, 2014. Accession numbers of newly sequenced viruses are presented in S1 Table. The data analyzed, including isolation dates, latitude, longitude and accession numbers are presented in the attached S1 Dataset. Sequences included in the dataset were subject to the following criteria: a) sequences had known location, host, and isolation date; b) for sequences with the same location, date of isolation and 100% similarity a single representative was retained; c) vaccine, derivative, and recombinant sequences were excluded; and d) sequences less than 1000 nucleotides in length were excluded. Locations with fewer than 10 taxa, as well as taxa collected prior to 1970 were excluded.
The remaining taxa were coded by both geographic region and ecosystem (‘wild’ or ‘domestic’). The wild classification included migratory birds (i.e. Anseriformes, Charadriiformes, etc.). The agricultural ecosystem included domestic birds raised for consumption (Galliformes including chicken, quail and pheasant; and Anseriformes including domestic duck and goose). See S1 Text for detailed descriptions of data stratification and subsampling. The final dataset consisted of 955 taxa, which were coded into 9 geographic regions: Japan/South Korea (n = 116), China–East (n = 147), China–Central (n = 179), China–West (n = 94), Southeast Asia (n = 18), South Asia (n = 93), Middle East (n = 210), Europe (n = 36), and North America (n = 62). 178 taxa were isolated from wild birds and the remaining from domesticated poultry. S2 Table presents detailed stratification of the dataset by region and ecosystem and S2 Fig shows the H9-HA sequence/location/year before and after subsampling.
Two additional datasets were assembled to investigate if model inferences from analysis of the H9 dataset could be generalized to other influenza A virus subtypes. Low pathogenic avian influenza A H3 and H6 subtype viruses have been sampled from both ecosystems and were chosen as comparison datasets. Even though AIV has a global distribution, surveillance and reporting is inconsistent and data availability for both wild and domestic birds can be limited. The AIV H3 and H6 subtype spatial distribution of wild and domestic birds sampled were similar to those of H9, but not identical. For the H3 subtype, sequence data was available from North Asia, including Russia and Mongolia, but none were available from western China, South Asia or the Middle East. For the H6 subtype no sequence data was available from South Asia. Limiting our ancestral state reconstruction to location states that overlapped between datasets would result in the exclusion of substantial data. Variation in data availability, locations sampled, and dataset design is discussed in the S1 Text. All available H3 and H6 subtype HA gene sequence data was downloaded from the Influenza Virus Resource database and screened based on the criteria described above (S2 Table).
For each of the gene segments analyzed, Bayesian phylogenetic trees were estimated using BEAST v.1.8 [43] with an uncorrelated lognormal relaxed molecular clock [44] that allows for rate variation across lineages. The general time-reversible model of nucleotide substitution (+ gamma + invariant sites) was used along with a Bayesian skyline coalescent tree prior. A minimum of three independent runs of 150 million generations were performed and combined after removal of burn-in to achieve an Effective Sample Size of >200 as diagnosed in Tracer v1.6.
Phylogenies record the history of viral exchange between ecosystems and the gene flow between spatially sampled populations [4, 6, 7, 32, 45–48]. By integrating both ecosystem and geography into the phylogenetic model, we can estimate the relative contribution of each to the global distribution and diversity of viruses in circulation. We used a non-reversible continuous-time Markov chain model to estimate the migration rates between geographical regions and the general patterns of H9 virus circulation in different avian populations [25]. Here we estimate the network linking the discrete wild and domestic populations distributed across regions. By defining the geographically and ecologically discrete characters in our model we were able to distinguish whether inter-regional migration was between domestic populations or wild animals. In addition, we estimated the rate of viral transmission between wild and domestic flocks and where the ecosystem interface was porous. A limitation of this approach is that realistic measurements of bird density and disease prevalence were not accounted for within our model.
A Bayesian stochastic search variable selection (BSSVS) was employed to reduce the number of parameters to those with significantly non-zero transition rates [25]. The BSSVS explores and efficiently reduces the state space by employing a binary indicator (I). A Bayes factor (BF) can be computed to assess the support for individual transitions between discrete states. We identify a transition as important when P(I = 1|data) >0.5. This analysis was conducted with a Poisson prior on the number of non-zero rates with a mean equal to the minimum number of rates required to connect the discrete ecological/geographical region states. We applied this to our analysis of the H9 dataset and determined the critical BF >14 (Poisson prior mean = 15). The same criterion was applied to our analysis of H3 (Poisson prior mean = 11) and H6 (Poisson prior mean = 14) datasets and determined a critical Bayes factor >10 and > 12 respectively. Strength of statistical supports were interpreted as follows; 10≤ BF <30 indicating strong support, 30≤ BF <100 indicates very strong support and BF >100 indicating decisive support [8, 25, 32].
We assessed statistical support of rate differences (wild > domestic and domestic < wild) by computing Bayes factors. The Bayes factors for differences in migration rates (r) were estimated by the ratio of posterior odds (P(r1 > r2 | Data)/P(r2 > r1 | Data)) versus prior odds P(r1 > r2)/P(r2 > r1), where the prior odds ratio was approximately 1 [8, 32].
Domestic populations were sampled much more intensively than wild populations (S1 Text). To investigate if our reconstruction was sensitive to data heterogeneity, we consider the prior expectation for the root discrete state frequencies and mean ecosystem transition rates for the sampling scheme used. If the discrete state distribution at the root is correlated with the location frequencies at the tips, we can expect that ancestral reconstruction throughout the entire phylogeny will be influenced by this tip-location sampling frequency. We randomized the location assignments at the tips throughout the MCMC procedure to investigate this possibility [37]. Similarly, if uneven sampling influences ecosystem transition rates then we can expect the mean rates to deviate from the prior. We further tested the model sensitivity to alternative sampling procedures where the number of sequences sampled was randomly pruned from a maximum of 955 sequences to a minimum of 73 sequences. Analysis of each sampling scheme was repeated three times. This sensitivity analysis was carried out for the final H3 and H6 subtype datasets. Our results quickly converged on the prior root location probability for all subtypes (equal state frequencies; S5 Fig) and ecosystem transition rate probability (equal mean rates; S6 Fig) indicating that the sampling frequencies had little impact on the model inferences a posteriori.
To assess the contribution of each region as a viral source or sink in the migration network, state jumps at the tree nodes, representing a state transition event (i.e. migration or ecosystem interaction) were counted [36]. We used a non-reversible model and therefore the direction of gene flow between states can be determined to assess which region was either source or sink. Heat maps representing the average number of jumps per year estimated from the last 1000 posterior sampled trees were generated. Due to the increase in surveillance efforts post-1997’s Hong Kong H5N1 outbreak and recent intensification in poultry production, our analysis only considered migration events between 1998 and 2013.
Even though the explicit migration events are not observed, the waiting times between state changes can be tracked on the phylogeny. The duration that a particular state is observed before transitioning to another state (Markov reward) was recorded on a branch-by-branch basis from the posterior sampling of phylogenetic trees [49]. These Markov rewards were calculated for each tree and ecosystem/location state in our model.
We further assessed the relative risk of each location as a viral source or sink population using 2x2 contingency tables [50]. The total jump counts in to or out of each discrete state were obtained for each step of the Bayesian MCMC. Contingency tables containing four cells (A, B, C, and D) were populated for each combination of regions. For example, to calculate the relative proportion of times viruses emerged from Japan and migrated to North America, the columns of the contingency table denote events in which Japan was the geographic source (left column) and events in which Japan was not the source (right column). Likewise, the rows of the contingency table can be denoted as events in which North America was the geographic sink of a transition (top row) and events in which was not the sink (bottom row). Therefore, cell A in this example includes the total number of estimated events in which viruses from Japan were introduced into North America; cell B includes the total introductions from other regions (not Japan) into North America; cell C includes the total number of introductions from Japan into other regions (not North America), and cell D includes the total number of mutually exclusive events in which Japan was not the geographic source and North America was not the geographic sink. The relative proportion of introductions from Japan into North America will be calculated as [A/(A+B)]/[C/(C+D)]. The proportion of total times where North America was the sink when Japan was the source was represented by [A/(A+B)]. Likewise, [C/(C+D)] represents the proportion of total times when viruses entered other regions (not North America) where Japan was the source. The ratio of these proportions represent the relative risk of Japan as the source when North America was the sink compared to when other regions were the geographic sinks. These ratios were calculated per MCMC step and averaged across all steps for each in order to incorporate phylogenetic uncertainty.
Data deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.601fd. [51]
All studies involving the collection of samples from wild and domestic animal species are conducted in compliance with the policies of the National Institutes of Health and the Animal Welfare Act, and with the approval of the St. Jude Children's Research Hospital Institutional Animal Care and Use Committee (Protocol Number 546-100324-10/14, approved July 20, 2015) and Massachusetts Institute for Technology (Protocol Number 0515-046-18, approved May 3, 2015).
|
10.1371/journal.ppat.1006917 | SpdC, a novel virulence factor, controls histidine kinase activity in Staphylococcus aureus | The success of Staphylococcus aureus, as both a human and animal pathogen, stems from its ability to rapidly adapt to a wide spectrum of environmental conditions. Two-component systems (TCSs) play a crucial role in this process. Here, we describe a novel staphylococcal virulence factor, SpdC, an Abi-domain protein, involved in signal sensing and/or transduction. We have uncovered a functional link between the WalKR essential TCS and the SpdC Abi membrane protein. Expression of spdC is positively regulated by the WalKR system and, in turn, SpdC negatively controls WalKR regulon genes, effectively constituting a negative feedback loop. The WalKR system is mainly involved in controlling cell wall metabolism through regulation of autolysin production. We have shown that SpdC inhibits the WalKR-dependent synthesis of four peptidoglycan hydrolases, SceD, SsaA, LytM and AtlA, as well as impacting S. aureus resistance towards lysostaphin and cell wall antibiotics such as oxacillin and tunicamycin. We have also shown that SpdC is required for S. aureus biofilm formation and virulence in a murine septicemia model. Using protein-protein interactions in E. coli as well as subcellular localization in S. aureus, we showed that SpdC and the WalK kinase are both localized at the division septum and that the two proteins interact. In addition to WalK, our results indicate that SpdC also interacts with nine other S. aureus histidine kinases, suggesting that this membrane protein may act as a global regulator of TCS activity. Indeed, using RNA-Seq analysis, we showed that SpdC controls the expression of approximately one hundred genes in S. aureus, many of which belong to TCS regulons.
| Staphylococcus aureus is a major human pathogen, and has become a significant worldwide health concern due to the rapid emergence of antibiotic resistant strains. Like most bacteria, S. aureus adapts to its environment by adjusting its genetic expression through sensing and regulatory systems. We show here that the SpdC membrane protein is a novel virulence factor of S. aureus, controlling biofilm formation and pathogenesis. We show that SpdC interacts with the WalK histidine kinase to inhibit activity of the WalKR two-component system. SpdC also interacts with nine other histidine kinases of S. aureus, suggesting it acts as a pleiotropic global regulator.
| Two-component systems (TCSs) are composed of a histidine kinase, usually membrane-bound and acting as an environmental sensor, which phosphorylates a coupled response regulator, often controlling gene transcription. Although these systems have been extensively studied and play an essential role in bacterial adaptation to the environment, the signal(s) to which they respond and additional factors positively or negatively controlling their activities remain mostly unknown. Staphylococcus aureus, a major human pathogen, causes diseases ranging from superficial cutaneous abscesses to life-threatening infections affecting all major organs [1]. S. aureus is also a commensal bacterium, colonizing approximately half the human population asymptomatically, essentially within the anterior nares [2]. In addition to its considerable arsenal of virulence factors, S. aureus must rapidly adapt to environmental conditions encountered during host colonization. Among the 16 TCSs encoded by the S. aureus genome [3], the SaeSR, AgrCA, and WalKR systems are particularly important for controlling virulence and innate immune evasion factors [4–8].
The WalKR system is the only S. aureus TCS shown to be essential for cell viability, suggesting that its activity may respond not only to environmental conditions but could also be controlled by intrinsic bacterial factors [9–11]. Indeed, it is becoming increasingly apparent that so-called two-component systems frequently involve additional proteins regulating the phosphorylation levels of the response regulator [12, 13]. These include accessory phosphatases such as CheZ, Spo0E or RapA, respectively dephosphorylating the E. coli CheY [14], B. subtilis Spo0A [15] and Spo0F [16] response regulators, or antikinases such as KipI and Sda that inhibit B. subtilis KinA [17, 18] and FixT which inhibits the Sinorhizobium meliloti FixJ kinase [19]. A sub-class of histidine kinases, known as intra-membrane sensing kinases [20], require the permease component of an associated ABC transporter for signal sensing, such as the S. aureus BraS [21] and GraS [22] kinases. Many TCS histidine kinases act as so-called bifunctional sensors, acting on their cognate response regulators both as kinases and phosphoprotein phosphatases [23]. Accordingly, several accessory proteins act by binding to the histidine kinase to inhibit its kinase activity or stimulate its phosphatase activity towards the response regulator. These include the PII protein acting on the NtrB kinase to control nitrogen assimilation by dephosphorylating NtrC [24], the Streptococcus agalactiae Abx1 Abi-domain membrane protein which interacts with the CovS histidine kinase to inhibit activity of the CovR response regulator [25], and the SaePQ protein complex, which stimulates the phosphatase activity of the SaeS histidine kinase in S. aureus [26].
In Bacillus subtilis, the WalK histidine kinase is thought to coordinate cell wall plasticity with cell division, with two membrane-bound accessory proteins, WalH and WalI, inhibiting WalK kinase activity [11, 27–29]. However, the WalH and WalI proteins of S. aureus share no significant sequence similarities with those of B. subtilis and their role is not as clear-cut. Indeed, although they are also membrane proteins and interact with the WalK kinase, WalH and WalI do not seem to play a major role in negatively controlling WalKR activity, suggesting that their functions have evolved [30].
In an effort to identify additional factors controlling the WalKR system, we showed that the S. aureus SpdC Abi-domain protein negatively affects WalK activity and expression of WalKR-regulated genes. We showed that SpdC, previously identified as playing a role in the display of surface proteins [31], forms a complex with the WalK histidine kinase and that the two membrane proteins preferentially localize at the division septum, suggesting that this interaction regulates WalK histidine kinase activity. The ΔspdC mutant displays a pleiotropic phenotype, including altered resistance towards compounds targeting the cell wall, as well as strongly diminished biofilm formation and virulence. Using RNA-Seq analysis, we showed that SpdC controls the expression of approximately one hundred genes in S. aureus. Indeed, SpdC activity appears to extend well beyond the WalKR system, since we have shown it also interacts with several other S. aureus histidine kinases suggesting it could be involved in controlling multiple regulatory/adaptive pathways.
We previously performed an extensive transcriptome analysis in order to define the scope of the S. aureus WalKR regulon [6]. Our results showed that expression of the spdC gene (SAOUHSC_02611) was increased 3.5-fold in a S. aureus strain producing a constitutively active form of the WalR response regulator (D55E) [6]. SpdC, a membrane-anchored protein with 8 predicted transmembrane segments and an Abi domain (CAAX protease self-immunity), (Fig 1A), was previously reported as playing a role in the display of surface proteins such as protein A [31]. In order to confirm that spdC is a member of the WalKR regulon, we used quantitative real-time PCR (qRT-PCR) to measure its expression in a S. aureus strain where the walRKHI operon is placed under the control of the IPTG-inducible Pspac promoter [9]. Cells were grown overnight in TSB with 0.05 mM IPTG, and cultures were inoculated at OD600nm = 0.05, with or without different IPTG concentrations (0.05 and 1 mM) to induce expression from the Pspac promoter. RNA samples were prepared from exponentially growing cells harvested at OD600nm = 0.5, before cessation of growth of the culture lacking IPTG, and walR and spdC mRNA levels were measured by qRT-PCR. As shown in Fig 1B, walR transcription is increased approximately 4-fold when cells are grown with 0.05 mM IPTG and 8-fold at 1 mM IPTG. Under the same conditions, spdC expression followed that of walR, and was increased 2-fold and 5-fold with 0.05 or 1 mM IPTG, respectively (Fig 1B), confirming positive regulation by the WalKR system.
As shown above, the WalKR TCS controls spdC expression. In Streptococcus agalactiae, another bacterial Abi-domain protein, Abx1, has been shown to inhibit activity of the CovSR two-component system [25]. In order to determine whether SpdC has a regulatory role in S. aureus, we generated a ΔspdC mutant in strain HG001 and performed a comparative RNA-Seq analysis. The ΔspdC mutant did not display any gross morphological changes or growth defects. Indeed, although it had a slight lag during the first hour post inoculation, the growth rate and final OD600nm were not significantly different from those of the parental strain (S1 Fig). Doubling times (http://www.doubling-time.com/compute.php) calculated during the exponential growth phase (S1 Fig, 90 min to 210 min) gave identical values of 32 min for both strains. The HG001 strain and ΔspdC mutant were grown in TSB until early exponential phase (OD600nm = 1) and total RNA was extracted for RNA-Seq analysis (See Materials and Methods). We verified that spdC is well expressed under these conditions using a lacZ reporter fusion with the spdC promoter region (S2 Fig).
Three biological replicates were analyzed by RNA-Seq for each strain and the data presented as the mean fold-change. Using a value cut-off greater than 2 with a P value less than 0.05, we found that the expression of 42 genes was lowered in the ΔspdC mutant strain and that of 65 increased (Table 1). In order to perform a general analysis of the transcriptomic data we generated an ontological grouping of SpdC-regulated genes (Fig 2A). Among the genes positively controlled by SpdC, 10 are known virulence factors, suggesting that SpdC may influence S. aureus pathogenicity, and 11 are involved in capsular biosynthesis. The S. aureus capsule is known to impede phagocytosis and promote host colonization, however although the HG001 strain used in this study carries the serotype 5 capsule gene cluster, a missense mutation in the cap5E gene prevents capsular biosynthesis [32–34].
Expression of several S. aureus prophage genes was also increased in the ΔspdC mutant strain: 11 for phage Φ13 and 17 for Φ12, (Table 1).
As shown above, spdC expression is controlled by the WalKR system. The RNA-Seq data analysis of SpdC-regulated genes reveals that 25 of these belong to the WalKR regulon (indicated by an asterisk in Table 1). In particular, the expression of 4 WalKR-dependent cell wall hydrolase genes (sceD, ssaA, lytM and atlA) is increased in the ΔspdC mutant, suggesting that SpdC negatively controls WalKR activity (Table 1). It is interesting to note that among the genes positively controlled by SpdC, many are also WalKR-activated genes (Table 1). All of these are classified as virulence genes, however they are not preceded by the WalR consensus binding site, and we have previously shown that several of these are not directly regulated by the WalKR system but through the SaeRS two-component system instead [6].
We used qRT-PCR to verify SpdC-dependent regulation for 3 positively (spa, hlgC, sdrD) and 3 negatively (atlA, sceD, lytM) controlled genes, in the ΔspdC strain compared to the HG001 parental strain, grown under the same conditions as for the RNA-Seq analysis (Fig 2B). We observed a perfect correlation with the RNA-Seq data: expression of the spa gene encoding protein A was very strongly lowered in the ΔspdC mutant strain (about one hundred-fold), while sdrD and hlgC expression levels were 5- to 10-fold less. The atlA, sceD and lytM cell wall hydrolase genes were more highly expressed in the absence of SpdC (2-, 12- and 5-fold, respectively) in agreement with the RNA-Seq analysis (Fig 2B).
In order to confirm SpdC-dependent regulation at the protein level, we chose two genes that were positively or negatively controlled by SpdC, spa and lytM, respectively, and performed Western blot analyses. Whole cell extracts were prepared from cultures of strains HG001, the ΔspdC mutant and the complemented mutant strain (ΔspdC/pMK4Pprot-spdC) and subjected to SDS-PAGE and immunoblotting. As shown in the top panel of Fig 2C, LytM levels are higher in the ΔspdC strain compared to the parental strain (lane 2), and in the complemented strain the LytM level is reduced to a level lower than in the parental strain (lane 3), likely reflecting the higher production of SpdC in the complemented strain. Indeed, under these conditions, spdC mRNA levels were increased more than 100-fold as measured by qRT-PCR as compared to the parental HG001 strain.
For studying levels of protein A, known to be covalently anchored to the cell wall (LPxTG sortase motif), identical quantities of cell wall fractions of the HG001 parental strain, ΔspdC deletion mutant, and complemented strain (ΔspdC/pMK4Pprot-spdC) were subjected to SDS/PAGE and compared by Western blot. As expected, protein A levels were significantly lower in the ΔspdC mutant than in the parental and complemented strains (Fig 2C, lower panel), in agreement with the RNA-Seq and qRT-PCR results.
These data indicate that SpdC is involved in controlling gene expression through potential interactions with regulatory systems, and the WalKR two-component system in particular.
As shown above, we have uncovered a regulatory link indicating that SpdC negatively controls activity of the WalKR two-component system, strongly suggesting that the proteins interact. In order to test possible interactions between SpdC and the WalKR proteins, we used the bacterial adenylate cyclase two-hybrid system (BACTH) [35]. We fused the full-length membrane-bound WalK histidine kinase or the WalR cytoplasmic response regulator to the C-terminal domain of the T25 subunit of the Bordetella pertussis adenylate cyclase and full-length SpdC to the C-terminal domain of the T18 subunit, using plasmids pKT25 and pUT18c respectively. To probe putative interactions, E. coli strain DHT1 was co-transformed with combinations of the pKT25 and pUT18c derivatives carrying the translational fusions. Upon protein-protein interactions, the close proximity between the T18 and T25 subunits restores adenylate cyclase activity, leading to cAMP synthesis and activation of the lactose operon. Interactions were tested both by spotting the resulting strains on LB plates containing X-Gal and by measuring β-galactosidase activity. To determine pair-wise interactions, we chose a cut-off value of 100 Miller Units as indicating a positive interaction between the protein fusions. As shown in Fig 3A, strong β-galactosidase activity was only observed for the plasmid combination co-producing the membrane anchored proteins SpdC and WalK while no interactions between SpdC and the cytoplasmic regulator WalR could be detected. SpdC is annotated as an Abi domain protein (CAAX protease self-immunity) in genome databases. The S. aureus HG001 genome encodes 4 Abi domain proteins, three of which, SpdA (SAOUHSC_01900), SpdB (SAOUHSC_02587), and SpdC, have been reported as being involved in surface protein display, whereas the fourth (SAOUHSC_02256) has no known function [31]. In order to test whether WalK also interacts with the other three Abi domain proteins, we constructed translational fusions for the remaining Abi proteins with the T18 domain of adenylate cyclase. As shown in Fig 3A, the combinations of SpdA, SpdB or SAOUHSC_02256 with WalK did not generate significant levels of β-galactosidase activity, demonstrating that SpdC is the only S. aureus Abi domain protein specifically interacting with WalK.
We have previously shown that the S. aureus WalK histidine kinase is mainly localized at the division septum [30]. Since SpdC and WalK interact, we studied the subcellular localization of SpdC in S. aureus by constructing a translational fusion with the GFP fluorescent protein using the pOLSA vector (See Materials and Methods), under the control of the cadmium-inducible Pcad promoter. The resulting plasmid, pOLSA-spdC was then introduced into the HG001 strain. Expression of the gene fusion was induced by addition of CdCl2 (0.25 μM), cells were harvested during exponential growth (OD600nm = 1.5) and observed by fluorescence microscopy. As shown in Fig 3B, SpdC is preferentially localized at the division septum, with a mean septum/membrane fluorescence ratio of around 2.6. Taken together, our results indicate that SpdC and WalK interact and are localized at the division septum.
As shown above, the expression of several cell wall hydrolase genes is significantly increased in the ΔspdC mutant strain, suggesting that sensitivity to compounds targeting the cell wall might also be affected. We followed bacterial lysis during incubation in the presence of a non-anionic detergent, Triton X-100, thought to trigger cell lysis by favoring endogenous autolysin activity [36]. However, Triton X-100 induced lysis for the HG001 and ΔspdC mutant strains was not significantly different (S3 Fig).
We also tested sensitivity to lysostaphin, a glycyl-glycine endopeptidase that hydrolyzes the peptidoglycan pentaglycine interpeptide crossbridge, leading to cell lysis [37]. The HG001, ΔspdC mutant and complemented strains were grown in TSB until OD600nm ≈ 1 and cells were then harvested and resuspended in PBS in the presence of lysostaphin. As shown in Fig 4A, the ΔspdC mutant was less sensitive to lysostaphin-induced lysis than the parental HG001 and complemented ΔspdC strains, suggesting that the absence of SpdC leads to cell wall modifications.
We then tested sensitivity to antibiotics targeting the cell wall. As shown in Fig 4B the ΔspdC mutant displayed increased sensitivity to the β-lactam antibiotic oxacillin, whereas the parental and complemented strains were able to grow at the concentration tested (0.1 μg/ml). No difference in sensitivity between the strains was seen using fosfomycin, an antibiotic inhibiting MurA, which catalyzes the very first step of peptidoglycan biosynthesis (Fig 4B). These results suggested that the ΔspdC mutant strain may either be affected in the later steps of peptidoglycan biosynthesis or may exhibit a cell wall structure modification leading to a difference in accessibility of antibiotics acting extracellularly. Wall teichoic acids (WTAs) are anionic sugar rich cell surface polymers that can alter accessibility to the cell wall. We therefore tested resistance to tunicamycin, an antibiotic targeting biosynthesis of WTAs. As shown in Fig 4B the ΔspdC mutant strain was highly sensitive to tunicamycin, in contrast to the parental and complemented strains.
Taken together, these results suggest that the S. aureus cell envelope structure is altered in the absence of SpdC.
Since our results indicate that the ΔspdC mutation may modify the S. aureus cell surface, we tested whether the absence of SpdC may have an effect on biofilm formation. Strains were grown statically in TSB, supplemented with glucose and NaCl, for 24 h at 37°C in PVC microplates. As shown in Fig 5, biofilm formation was strongly decreased in the absence of SpdC (approximately 7-fold). Complementation of the ΔspdC mutant with the pMK4Pprot-spdC plasmid restored biofilm formation to levels comparable to those of the parental HG001 strain (Fig 5). These results are consistent with a modification of the S. aureus cell surface in the absence of SpdC, which could influence resistance against antimicrobial compounds targeting cell surface structures as well as the capacity to form biofilms. In order to determine which biofilm component is affected, biofilm detachment experiments were carried out (S4 Fig) by treatment with proteinase K, DNaseI and sodium metaperiodate (a carbohydrate-modifying agent). Under our conditions, biofilm production was lowered three-fold after treatment with DNaseI, and more than 10-fold when treated with proteinase K, but not significantly modified after treatment with sodium metaperiodate. Thus, biofilms formed under our conditions by the HG001 parental strain are essentially protein-based, and, to a lesser extent, due to extracellular DNA. We observed reduced biofilm formation for the ΔspdC mutant even after DNAseI treatment, but not after proteinase K treatment (S4 Fig), suggesting that SpdC affects the production of proteins important for biofilm formation.
Cell surface modifications are known to impact virulence. Likewise, the capacity to form robust biofilms favors bacterial colonization of the host. Additionally, our RNA-Seq analysis revealed that the expression of at least 10 genes directly involved in bacterial virulence upon infection was lowered in the ΔspdC mutant, strongly suggesting that SpdC plays a role in virulence. We used a murine sepsis model to compare virulence of the HG001 and ΔspdC strains. SWISS mice were infected intravenously with 5.107 cfu and mortality was monitored over 9 days post infection. As shown in Fig 6 virulence of the ΔspdC mutant was strongly diminished. Indeed, following infection with the HG001 parental strain, significant mortality occurred in the first 3 days post-infection (greater than 60%), whereas only a single mouse out of 14 died within five days after infection with the ΔspdC mutant. After the sixth day, a moderate mortality was observed for the group infected with the ΔspdC mutant, with only 36% mortality at the end of the assay (compared to 72% mortality for mice infected with the parental HG001 strain). This significant difference indicates that SpdC is a novel virulence factor in S. aureus.
As shown above, SpdC localizes at the division septum and interacts with the WalK histidine kinase, negatively controlling WalKR activity. However, many of the SpdC-regulated genes identified by RNA-Seq do not belong to the WalKR regulon, but are known to be controlled by other two-component systems (Table 1) suggesting SpdC may interact with other TCS regulatory pathways. The S. aureus HG001 genome encodes 16 two-component systems [33] and we constructed translational fusions for each of the histidine kinase genes with the carboxy-terminal region of the adenylate cyclase T25 domain. Each of the resulting plasmids was co-transformed in combination with the pUT18c-spdC plasmid into E. coli strain DHT1. As shown in Fig 7, in addition to WalK, we detected interactions between SpdC and nine additional histidine kinases: YesM, GraS, SaeS, DesK, ArlS, SrrB, PhoR, VraS, and BraS. These interactions appear to be specific, since no interactions were detected between SpdC and the remaining six histidine kinases, as shown in Fig 7 (LytS, AirS, AgrC, KdpD, HssS, and NreB).
Among the genes positively regulated by SpdC (Table 1), the expression of at least fifteen is also activated by the SaeSR TCS, including the spl operon, sak, the hlgBC operon and chp [7, 38–40]. This suggests that SpdC activates the SaeSR TCS, in contrast to its role in negatively controlling activity of the WalKR system.
Histidine kinases have different combinations of signaling domains such as HAMP or PAS domains [41] in addition to the conserved H, N, G1, F and G2 boxes of the phosphoacceptor/dimerization (HisKA) and catalytic (HATPase_C) domains [42–44]. We have shown that SpdC interacts with 10 of the 16 S. aureus histidine kinases, which do not share any strong amino acid sequence similarities other than the conserved histidine kinase HisKA and HATPase_C domains. Since SpdC interacts with some, but not all of the S. aureus histidine kinases, this protein-protein contact must involve some other domain.
We focused our analysis on the WalK and SaeS histidine kinases. WalK has two transmembrane domains (amino acids 14–34 and 183–203), with an extracellular loop of 148 amino acid residues, a HAMP domain involved in signal transduction (204–256), a PAS domain (261–331) a PAC domain (314–378) and a histidine kinase domain (382–600; Fig 8A). In contrast, SaeS is a member of the intra-membrane sensing kinases [45], with two transmembrane domains (amino acids 9–29 and 40–60) separated by only ten amino acids, as well as a HAMP domain (61–114) and a histidine kinase domain (129–348; Fig 8A). We tested the interactions of SpdC with the N-terminal domains of WalK and SaeS containing the transmembrane regions (WalK1-203 and SaeS1-64, respectively). The truncated proteins were fused to the T25 domain of adenylate cyclase, and the resulting plasmids were co-transformed into E. coli strain DHT1 together with the pUT18c-spdC plasmid. As shown in Fig 8B, the first 203 amino acids of WalK are sufficient to allow stable interactions with SpdC. For SaeS, the N-terminal domain containing only the two transmembrane segments (SaeS1-64) did not lead to interaction with SpdC, but a longer fragment of the protein (SaeS1-120) gave rise to a stable interaction with SpdC and high β-galactosidase activity (Fig 8B). These results suggest that WalK and SaeS interact with SpdC through their transmembrane domains. The negative interaction results obtained with the truncated SaeS protein containing only the transmembrane domains suggest that since this kinase lacks an extracellular loop, the HAMP domain may be required for proper membrane insertion of the fusion protein.
In order to identify which domain of SpdC interacts with the WalK and SaeS kinases, we compared interactions with full-length SpdC (pUT18c-spdC) and a carboxy-terminal truncated SpdC consisting only of the eight transmembrane domains (SpdC1-252; pUT18c-spdC1-252). We noted self-interaction of SpdC following co-transformation of pKT25-spdC with either pUT18c-spdC or pUT18c-spdC1-252 (Fig 8B), indicating that the SpdC transmembrane domains are involved in these self-interactions. Similar results were obtained when testing interactions with WalK and SaeS, i.e. the transmembrane domains of SpdC are sufficient to allow interactions with the histidine kinases (Fig 8B).
Taken together, these results indicate that SpdC is a membrane-bound protein that interacts with itself and several histidine kinases through transmembrane domain contacts.
Abi-domain proteins constitute a large family whose functions are mostly unknown. SpdC was initially designated LyrA for Lysostaphin resistance A and identified by screening a bursa aurealis transposon mutant library for increased lysostaphin resistance [46]. An independent study aiming at characterizing proteins involved in the display of surface proteins led to the identification of three proteins, SpdA, SpdB and SpdC (Surface protein display A, B and C), playing a role in protein A levels at the staphylococcal cell surface [31]. These proteins share a similar structural organization with 6 to 8 transmembrane domains and an Abi-domain embedded within the hydrophobic region. An additional protein with a similar organization is encoded by the S. aureus genome, SAOUHSC_02256, but appears to have no role in controlling protein A levels [31].
We previously characterized the essential WalKR two-component system in S. aureus and highlighted its major role in controlling cell wall homeostasis [9, 47]. Transcriptome analysis revealed that spdC expression is positively controlled by WalKR [6]. We show here that the SpdC membrane protein and the WalK histidine kinase interact and that SpdC negatively controls WalKR activity and expression of WalKR-regulated genes. Interaction of the WalK histidine kinase with SpdC is specific since no interaction was seen with the other Abi-proteins (SpdA, SpdB and SAOUHSC_02256). Accordingly, we also showed that SpdC and WalK are both localized preferentially at the division septum.
Interestingly, an RNA-Seq analysis of a ΔsdpC mutant revealed that the expression of 107 genes varied compared to the parental strain. Among these, 24 (more than 20%) are controlled by the WalKR system. Since SpdC appears to negatively control WalKR activity by interacting with WalK, this septal localization is consistent with the previously suggested cell wall metabolism-related activation signal of the WalK histidine kinase in Bacillus subtilis [11, 48]. Indeed, in cocci, cell wall synthesis has been shown to exclusively occur at the division septum in an FtsZ-dependent manner [49], suggesting that a peptidoglycan metabolism related signal at the septum may relieve negative control of WalK activity by SpdC.
Histidine kinases often act as phosphoprotein phosphatases towards their associated response regulator. WalK was previously classified as a kinase/phosphatase «bifunctional sensor» [23] and the PAS domain of Streptococcus pneumoniae WalK plays a role in its phosphatase activity [50, 51]. In S. aureus, we have previously shown that WalK acts as a WalR phosphoprotein phosphatase upon entry into stationary phase in order to shut off WalR activity [6]. Interactions between SpdC and WalK can either interfere with signal perception by the sensor histidine kinase, inhibit its kinase activity or increase its phosphatase activity towards WalR, thus negatively controlling WalKR-dependent gene expression. SpdC is unlikely to directly regulate gene expression since it is a membrane protein lacking any typical DNA-binding domain.
To understand how the other SpdC-dependent genes were controlled (83 of the genes identified by RNA-Seq are not regulated by the WalKR system), we tested interactions of SpdC with the other S. aureus histidine kinases and found that it interacts with 10 of the 16 encoded in the genome. No obvious structural motifs or domain sequences were common to those that interacted with SpdC compared to those that did not. Two histidine kinases, WalK and SaeS, were chosen for further analysis of their interactions with SpdC. Our results indicate that the two transmembrane domains of WalK are sufficient to allow interaction with SpdC, whereas a greater amino-terminal fragment of SaeS was required, encompassing the HAMP domain. This result suggests that although the transmembrane domains of SaeS are likely involved in interactions with SpdC, a longer fragment may be necessary to ensure proper membrane insertion of the truncated protein. We also showed that a truncated form of SpdC, containing only the eight transmembrane domains, was sufficient for self-interaction and interaction with WalK and SaeS, indicating that transmembrane domains are involved in the interactions between SpdC and the histidine kinases. Interestingly, the only two S. aureus cytoplasmic histidine kinases which lack transmembrane domains, AirS and NreB, did not interact with SpdC under our conditions, in agreement with our results indicating transmembrane segments are involved in the interactions. The only other example of an Abi domain protein interacting with a histidine kinase is Abx1 of Streptococcus agalactiae, which interacts with the CovS kinase [25]. The two transmembrane domains of CovS were shown to be necessary and sufficient for these interactions [25].
At least fifteen genes belonging to the SaeSR regulon, including the spl operon, sak, the hlgBC operon and chp [7, 38–40], are also positively controlled by SpdC (Table 1), indicating that SpdC likely activates the SaeSR TCS, in contrast to its role in negatively controlling activity of the WalKR system. Since WalR controls spdC expression, this is consistent with our previous results showing that constitutive activation of WalR generates a signal leading to activation of the SaeSR TCS and a corresponding increase in SaeSR regulon expression [6].
The localization of SpdC at the division septum and its role in gene regulation through interactions with sensor kinases of two-component systems led us to speculate that SpdC may interfere with bacterial division sensing and impact cell wall metabolism. Accordingly, the ΔspdC mutant displays increased resistance against lysis when treated with lysostaphin, in agreement with the original phenotype characterized by transposon insertion [46]. Additionally, the absence of SpdC was reported to lead to increased cross wall abundance and thickness [31]. We tested sensitivity of the ΔspdC mutant to antibiotics targeting the cell wall. The ΔspdC mutant is highly sensitive to oxacillin and tunicamycin, but not to fosfomycin, which inhibits the first step of cell wall biosynthesis. In agreement with the sensitivity of the ΔspdC mutant to tunicamycin, which inhibits wall teichoic acid synthesis, the spdC gene was also identified as a candidate using a screen to identify synthetically lethal mutations with teichoic acid biosynthesis defects [52]. We have shown that expression of cell wall hydrolase genes is increased in the ΔspdC strain (sceD, ssaA, lytM, atlA). This may lead to increased cell wall degradation, which could explain the lowered resistance to oxacillin. This could also lead to the mutant’s increased sensitivity to tunicamycin. Indeed, teichoic acids are key elements for the proper localization of AtlA to the division septa, where cell wall biosynthesis takes place, since cell wall plasticity is essential for cell division [53]. In the presence of tunicamycin, the absence of wall teichoic acids results in a delocalized distribution of AtlA across the cell surface. Since atlA is more highly expressed in the ΔspdC stain, this could explain why this strain is more sensitive to tunicamycin.
Taken together with the strong links to the WalKR TCS, these results indicate that SpdC is involved in bacterial cell envelope homeostasis. The importance of the bacterial cell envelope in host-pathogen interactions cannot be over-emphasized: it is the first layer of contact between the bacterium and its host, containing an array of cell wall-linked or associated toxins and virulence factors, the first and major bacterial line of defense against threats from the host or environment, and is also both the target of choice for antibiotic treatment and the source of many antibiotic-resistance pathways. We have shown that SpdC is required for biofilm formation, an important step in S. aureus pathogenesis. We also showed that SpdC is a novel S. aureus virulence gene, required for the infectious process in a mouse septicemia model. The loss of virulence observed with the ΔspdC mutant may involve both its diminished capacity to form biofilms as well as the lowered expression of multiple virulence genes as shown by RNA-Seq analysis.
The regulatory mechanism mediated by SpdC remains to be determined. Abi-domain-containing proteins have been extensively studied in eukaryotes. They are involved in CAAX-protein maturation by cleaving the C-terminal AAX tripeptide after addition of an isoprenyl group on a cysteine, the last step consisting in methylation of the new C- terminus. These three modification steps are termed prenylation [54]. This post-translational maturation has a crucial role in maintaining cellular homeostasis by controlling the localization and activity of a large range of proteins. In particular, by adding a lipid group at the carboxy-terminal end of proteins, it favors their interactions with membranes, which have a high concentration of signaling molecules [55]. Prenylation has been recently described in prokaryotes and a geranyltransferase, IspA, has been identified in S. aureus [56]. Putative methyltransferase and CAAX-protease encoding genes (including spdC) are also present in the S. aureus genome. Deletion of the ispA gene has pleiotropic effects such as a growth defect, increased sensitivity to oxidative stress and an altered cell envelope [56]. Of note, the absence of IspA or SpdC both lead to increased cell wall antibiotic sensitivity. The RNA-Seq transcriptome analysis of the ΔispA strain [56] shows similarities with that of the ΔspdC mutant. Indeed, one of the most regulated genes in both cases is spa, encoding the immunoglobulin G binding protein A. We also noted that 21 Φ13 genes are up-regulated in the ΔispA mutant whereas we characterized 11 Φ13 genes up-regulated in the ΔspdC mutant, with several in common. These data suggest that SpdC and IspA could be involved in the same cellular pathway.
Interestingly, the transcriptome analysis of the ΔispA mutant revealed modified expression of a large number of genes involved in regulatory circuits and particularly increased expression of the walR, walH, walI and walJ genes of the wal locus [56]. While no direct link between prenylation and Abi domain proteins has been shown in prokaryotes, there are several links with the WalKR system, either by physical interaction and negative control of activity, for SpdC, or at the transcriptional level for the IspA geranyltransferase.
This study identifies a membrane-bound protein with an Abi domain, SpdC, at the core of an interaction network that coordinates bacterial division with cell envelope metabolism and host interactions. Further studies are required in order to decipher the molecular mechanism and consequences of these interactions.
Escherichia coli K12 strain DH5α (Invitrogen, Thermo Fisher Scientific, Waltham, MA) was used for cloning experiments. Staphylococcus aureus strain HG001 [57] was used for genetic and functional studies. Plasmids were first passaged through the restriction deficient S. aureus strain RN4220 before introduction into the HG001 strain. E. coli strains were grown in LB medium with ampicillin (100 μg/ml) added when required. S. aureus strains and plasmids used in this study are listed in Table 2. S. aureus strains were grown in Trypticase Soy Broth (TSB; Difco; Becton, Dickinson and Co., Franklin Lakes, NJ) supplemented with chloramphenicol (10 μg/ml) or erythromycin (1 μg/ml) as required. E. coli and S. aureus strains were transformed by electroporation using standard protocols [58] and transformants were selected on LB or Trypticase Soy Agar (TSA; Difco) plates, respectively, with the appropriate antibiotics. Expression from the Pcad promoter was induced by adding cadmium chloride (CdCl2) at a final concentration of 0.25 μM. Expression from the Pspac promoter was induced by addition of isopropyl β-D-1-thiogalactopyranoside (IPTG).
Oligonucleotides used in this study were synthesized by Eurofins Genomics (Ebersberg, Germany) and their sequences are listed in Table 3. S. aureus chromosomal DNA was isolated using the MasterPure Gram-positive DNA purification Kit (Epicentre Biotechnologies, Madison, WI). Plasmid DNA was isolated using a QIAprep Spin Miniprep kit (Qiagen, Hilden, Germany) and PCR fragments were purified using the Qiaquick PCR purification kit (Qiagen). T4 DNA ligase and restriction enzymes, PCR reagents and Q5 high-fidelity DNA polymerase (New England Biolabs, Ipswich, MA) were used according to the manufacturer's recommendations. Nucleotide sequencing of plasmid constructs was carried out by Beckman Coulter Genomics (Danvers, MA).
For construction of the ΔspdC mutant strain, two 800 bp DNA fragments were generated by PCR using oligonucleotide pairs OP375/OP376 and OP377/OP378, respectively (see Table 3), corresponding to the DNA regions located immediately upstream and downstream from the spdC gene. These DNA fragments were cloned in tandem in two consecutive steps, between the BamHI and NcoI restriction sites of the pMAD vector. The resulting plasmid was introduced by electroporation into S. aureus and transformants were selected at 30°C on TSA plates containing erythromycin (1 μg/ml) and 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (X-Gal, 100 μg/ml). Integration and excision of the plasmid were then performed as previously described [59], yielding mutant strain ST1317 (ΔspdC). A complementation plasmid pMK4Pprot-spdC was constructed by cloning the entire spdC coding sequence (amplicon OP404/OP405) in plasmid pMK4Pprot, under the control of the constitutive Pprot promoter [60]. The plasmid was introduced into the ST1317 ΔspdC mutant, generating the ST1375 complemented strain. Expression of the walKRHI in strain HG001 was placed under the control of the IPTG-iducible Pspac promoter by Φ80α phage transduction [61] using strain ST1000 (RN4220 PspacwalRKHI; [9]) as a donor and strain HG001 as the recipient, yielding strain ST1017 (HG001 PspacwalRKHI).
Plasmid pSA14 [62] is a derivative of shuttle vector pMK4 [63], carrying a promoterless E. coli lacZ gene and was used to construct transcriptional lacZ reporter fusions. The spdC promoter region was amplified by PCR using oligonucleotides OSA512/OSA513 (see Table 3) and cloned between the PstI/BamHI restriction sites of the pSA14 vector, yielding plasmid pSD3-41 (Table 2).
For β-galactosidase assays in S. aureus, strain ST1386 carrying the spdC’-lacZ fusion was grown in TSB at 37°C and cells were harvested by centrifuging 2 ml culture samples (2 min; 5,400 x g). Assays were performed as previously described [21] and β-galactosidase specific activities expressed as Miller units mg−1 protein [64]. Protein concentrations were determined using the Bio-Rad protein assay (BioRad, Hercules, CA) [65].
Strains were grown in TSB, supplemented with IPTG when specified, at 37°C with aeration until OD600nm = 1. Cells were pelleted by centrifugation (2 min, 20,800 x g) and immediately frozen at -20°C. RNA extraction was then performed as previously described [66], followed by DNaseI treatment with the TURBO DNA-free reagent (Ambion, Austin, TX) in order to eliminate residual genomic DNA.
cDNA synthesis was carried out as previously described [47]. Oligonucleotides were designed with the BEACON Designer 7.91 software (Premier Biosoft International, Palo Alto, CA) in order to synthesize 100–200 bp amplicons (see Table 3). Quantitative real-time PCRs (qRT-PCRs), critical threshold cycles (CT) and n-fold changes in transcript levels were performed and determined as previously described using the SsoFast EvaGreen Supermix (Bio-Rad, Hercules, CA) and normalized with respect to 16S rRNA whose levels did not vary under our experimental conditions [47]. All assays were performed using quadruplicate technical replicates, and repeated with three independent biological samples.
Three independent biological replicates were used for RNA-Seq analysis of the parental HG001 and ΔspdC strains. Strains were grown in TSB until OD600nm = 1. Total RNA was isolated as described above, and 7 μg were treated using the MicrobExpress kit (Ambion, Austin, TX) in order to remove rRNA. The rRNA depleted fraction was used for construction of strand specific single end cDNA libraries using the Truseq Stranded Total RNA sample prep kit according to the manufacturer’s instructions (Illumina, San Diego, CA). Libraries were sequenced using an Illumina Hiseq2000 sequencer (multiplexing 6 samples in one lane) according to the manufacturer’s instructions (Illumina, San Diego, CA).
Sequences were demultiplexed using the Illumina alignment and sequence analysis pipeline (GERALD, included in CASAVA version 1.7) giving FASTQ formatted reads. Reads were cleaned by removing adapter and low quality sequences using an in-house program (https://github.com/baj12/clean_ngs). Only sequences with a minimum length of 25 nucleotides were considered for further analysis. Bowtie (version 0.12.7, -m50—chunkmbs 400 -a—best -q -e50) [67] was used for alignment with the reference Staphylococcus aureus subsp. aureus genome (gi|88193823). Only uniquely aligning reads where considered for counting. HTseq-count (version 0.5.4p5, parameters: -m intersection-nonempty, -s yes, -t CDS -I locus_tag) was used for counting genes [68].
Statistical analysis was performed with R version 3.0.2 [69] and DESeq2 version 1.2.10 [70]. Data were first normalized with DESeq2 and the default parameters. Dispersion estimation and statistical testing were performed using the Generalized Linear Model with default parameters. Independent filtering was performed with default parameters to exclude transcripts with very low count values. Raw P-values were then adjusted according to the Benjamini and Hochberg procedure [71] and transcripts were considered differentially expressed when their adjusted P-value was lower than 0.05.
For testing protein interactions using the Bacterial Adenylate Cyclase Two-Hybrid System (BACTH), genes encoding the proteins of interest were cloned into plasmids pKT25 and pUT18c leading to translational fusions with the T25 or T18 domains of the Bordetella pertussis adenylate cyclase [35]. DNA fragments corresponding to the coding sequences were amplified by PCR using chromosomal DNA from strain HG001 and specific oligonucleotide pairs (see Table 3). Fragments were digested with BamHI and EcoRI or KpnI (indicated in Table 3) for cloning into plasmids pKT25 or pUT18c. The resulting plasmids were co-transformed into E. coli strain DHT1 [72] to detect protein-protein interactions and transformants were selected on kanamycin (50 μg/ml) for pKT25 derivatives and ampicillin (100 μg/ml) for pUT18c derivatives.
The resulting strains carrying combinations of pKT25 and pUT18c derivatives were tested for cyclic AMP-dependent activation of lacZ expression. For tests on plates, strains were grown in LB liquid medium supplemented with ampicillin (100 μg/ml) and kanamycin (50 μg/ml). Overnight cultures were then spotted on LB-agar plates containing IPTG (0.5 mM), ampicillin (100 μg/ml), kanamycin (50 μg/ml), and X-Gal (100 μg/ml). Plates were incubated for 24 H at 30°C and examined for appearance of the characteristic blue color indicative of β-galactosidase activity through X-Gal hydrolysis. Quantitative β-galactosidase assays were performed on exponentially growing E. coli liquid cultures. Cells were grown in LB with IPTG (0.5 mM), ampicillin (100 μg/ml) and kanamycin (50 μg/ml) at 30°C under aeration until OD600nm = 1 and assays performed on SDS/chloroform permeabilized cells as previously described [64]. Enzymatic activities are represented relative to negative and positive controls, respectively a strain carrying the empty pKT25 and pUT18c vectors (activity = 0, arbitrary unit), and a strain with the pKT25-zip and pUT18c-zip plasmids [35] (activity = 1000, arbitrary unit).
The pOLSA plasmid was used to produce a fluorescent SpdC-GFP fusion protein [30]. The translational fusion was constructed by PCR amplification using HG001 chromosomal DNA and oligonucleotide pair OSA417/OSA404 (Table 3). The amplicon was cloned into pOLSA between the SalI and XmaI restriction sites, allowing transcription from the Pcad promoter and production of the SpdC-GFP fusion protein.
Subcellular protein localization of SpdC was performed in S. aureus HG001 transformed with pOLSA-spdC. Fluorescence microscopy was carried out on cells grown in liquid cultures in TSB supplemented with CdCl2 (0.25 μM) to induce gene fusion expression. When cells reached OD600nm ≈ 1.5 (exponential growth phase), they were harvested and concentrated 20 times in PBS. Cell suspensions were mixed with Vectashield mounting media (Vector Laboratories, Burlingame, CA) and used for microscopic observations with a Nikon Eclipse E600. Images were acquired with a Nikon DXM1200F Digital Camera. ImageJ software was used for quantifying fluorescence (http://imagej.nih.gov/ij/index.html; [73]). Fluorescence ratios were calculated by measuring fluorescence at the division septa divided by the fluorescence at the lateral wall after subtracting background fluorescence. Quantification was performed for 33 cells and two independent biological replicates and plotted using GraphPad Prism (GraphPad Software, San Diego, CA; http://www.graphpad.com).
The HG001, ΔspdC and ΔspdC/pMK4Pprot-spdC strains were grown overnight at 37°C with aeration in TSB medium, with chloramphenicol (10 μg/ml) when required. Bacterial suspensions diluted from 10−2 to 10−7 were spotted (3 μl) onto TSA plates with the indicated antibiotic concentrations and incubated for 15 hours at 37°C.
S. aureus whole cell lysates were prepared as previously described [47]. Briefly, 5 ml of a cell culture grown to stationary phase were harvested by centrifugation (10 min; 3,000 x g), resuspended in 2X Laemmli SDS sample buffer (0.2 ml) and heated at 99°C for 10 min. Supernatants containing SDS-soluble proteins were collected following centrifugation (10 min; 20,800 x g), and used for further analysis. Cell wall extracts were prepared from 50 ml of the same cultures. Cells were pelleted and resuspended in 4 ml of digestion buffer (50 mM Tris-HCl pH 8, 145 mM NaCl, 30% sucrose, 160 ng/ml DNaseI, 250 μg/ml lysostaphin) and incubated for 60 min at 37°C. Supernatants corresponding to cell wall extracts were then harvested by centrifugation (10 min; 3,000 x g). Protein extracts were separated by SDS-PAGE on a 12% polyacrylamide gel, followed by Coomassie Brilliant Blue staining to verify that the quality and quantity of loaded extracts was equivalent for the different strains. For immunoblotting experiments, protein extracts were transferred to a nitrocellulose membrane after SDS-PAGE using a semi-dry blotter (Bio-Rad, Hercules, CA) and the following buffer: 25 mM Tris, 192 mM glycine, 20% ethanol. The LytM protein was detected using a purified polyclonal rabbit antibody [74] and horseradish peroxidase-coupled anti-rabbit secondary antibodies (Zymed, ThermoFisher, Waltham, MA) and the Pico chemiluminescence Western blot kit (Pierce, ThermoFisher, Waltham, MA). Detection of Spa was carried out directly using the secondary antibodies. Purified Staphylococcus aureus Protein A was obtained from Sigma-Aldrich (St. Louis, MO).
Strains were grown in TSB, with 10 μg/ml chloramphenicol for the complemented strain, at 37°C under aeration. When the OD600nm reached 1, bacteria were harvested by centrifugation, (10 min; 3000 x g), washed in PBS, and resuspended in the same volume of PBS (control) or PBS containing 200 ng/ml lysostaphin followed by incubation at 37°C. Lysis was monitored by measuring the decline in OD600nm over time and indicated as a percentage of the initial OD (measured OD600nm/initial OD600nm).
Biofilm assays were performed by growing cells in PVC microtiter plates (200 μl per well) in TSB with 0.75% glucose and 3.5% NaCl. After 24 h static growth at 37°C, adherent biomass was rinsed twice with PBS, air dried, stained with 0.1% crystal violet for 15 min, resuspended in ethanol-acetone (80:20) and quantified by measuring OD595nm, normalized to the OD600nm of each culture (growth rates for the different strains were the same).
Seven-week-old female RjOrl:SWISS mice (Centre d’Elevage Roger Janvier, Le Genest-St.-Isle, France) were inoculated intravenously with the S. aureus HG001 parental strain and the otherwise isogenic ΔspdC mutant. Groups of seven mice were infected with 5.107 cfu per mouse in 0.2 ml. Survival was monitored daily over 9 days post-infection and three independent experiments were carried out. Virulence of the complemented strain could not be carried out since we have shown that in vivo, in the absence of selection pressure, the complementation plasmid was lost over the assay period. Indeed, after nine days post-infection with the ST1375 complemented strain, animals were sacrificed and the kidneys removed and homogenized for determination of bacterial CFU (total and chloramphenicol resistant) per kidney, revealing that 97% of the bacteria had lost the pMK4Pprot-spdC complementation plasmid.
Animal experiments were conducted at the Institut Pasteur in compliance with French legislation (Decree N° 2001–464 05/29/01) and European Union guidelines on handling of laboratory animals:
(http://ec.europa.eu/environment/chemicals/lab_animals/index_en.htm).
Animals were sacrificed by increasing carbon dioxide concentrations. Protocols were approved by the Institut Pasteur ethics committee (Authorization N° 2013–0032).
The complete RNA-Seq dataset was deposited in the EMBL European Nucleotide Archive (accession number PRJEB11849) and is accessible at the following URL:
http://www.ebi.ac.uk/ena/data/view/PRJEB11849
|
10.1371/journal.pgen.1001078 | Consistent Association of Type 2 Diabetes Risk Variants Found in Europeans in Diverse Racial and Ethnic Groups | It has been recently hypothesized that many of the signals detected in genome-wide association studies (GWAS) to T2D and other diseases, despite being observed to common variants, might in fact result from causal mutations that are rare. One prediction of this hypothesis is that the allelic associations should be population-specific, as the causal mutations arose after the migrations that established different populations around the world. We selected 19 common variants found to be reproducibly associated to T2D risk in European populations and studied them in a large multiethnic case-control study (6,142 cases and 7,403 controls) among men and women from 5 racial/ethnic groups (European Americans, African Americans, Latinos, Japanese Americans, and Native Hawaiians). In analysis pooled across ethnic groups, the allelic associations were in the same direction as the original report for all 19 variants, and 14 of the 19 were significantly associated with risk. In summing the number of risk alleles for each individual, the per-allele associations were highly statistically significant (P<10−4) and similar in all populations (odds ratios 1.09–1.12) except in Japanese Americans the estimated effect per allele was larger than in the other populations (1.20; Phet = 3.8×10−4). We did not observe ethnic differences in the distribution of risk that would explain the increased prevalence of type 2 diabetes in these groups as compared to European Americans. The consistency of allelic associations in diverse racial/ethnic groups is not predicted under the hypothesis of Goldstein regarding “synthetic associations” of rare mutations in T2D.
| Single rare causal alleles and/or collections of multiple rare alleles have been suggested to create “synthetic associations” with common variants in genome-wide association studies (GWAS). This model predicts that associations with common variants will not be consistent across populations. In this study, we examined 19 T2D variants for association with T2D risk in 6,142 cases and 7,403 controls from five racial/ethnic populations in the Multiethnic Cohort (European Americans, African Americans, Latinos, Japanese Americans, and Native Hawaiians). In racial/ethnic pooled analysis, all 19 variants were associated with T2D risk in the same direction as previous reports in Europeans, and the sum total of risk variants was significantly associated with T2D risk in each racial/ethnic group. The consistent associations across populations do not support the Goldstein hypothesis that rare causal alleles underlie GWAS signals. We also did not find evidence that these markers underlie racial/ethnic disparities in T2D prevalence. Large-scale GWAS and sequencing studies in these populations are necessary in order to both improve the current set of markers at these risk loci and identify new risk variants for T2D that may be difficult, or impossible, to detect in European populations.
| Multiple common risk alleles have been identified as reproducibly associated with risk of type 2 diabetes (T2D) [1]–[13]. With the exception of the KCNQ1 locus which was identified in the Japanese population [1], [2], all of the well-replicated risk variants were first identified in populations of Northern European ancestry [3]–[13]. T2D morbidity varies widely across racial/ethnic groups; the prevalence is more than twice as high among African Americans, Japanese Americans, Latinos and Native Hawaiians as European Americans [14], [15]. It is important to evaluate whether and how genetic variation may contribute to health disparities between populations. For example, genetic variation at 8q24 may contribute to population differences in risk of prostate cancer [16], [17], and genetic variation at MYH9 contributes substantially to the higher rates of kidney disease in African Americans [18].
It has recently been argued that single rare causal variants and/or collections of multiple different rare variants on unrelated haplotypes may create “synthetic associations” of common variants with disease risk [19]–[21]. One prediction of this model is that the associations with common variants will not be consistent across populations (since many of the mutations will be young in age, and post-date the migrations that led to the founding of modern continental populations). Type 2 diabetes has been specifically discussed as a possible case in which synthetic associations might be operative, based on the lack of statistical significance in very small studies that examined allelic associations for T2D in multi-ethnic samples.
Testing the association of each validated risk allele for T2D in multiple populations is an important step to determine (a) whether these genetic markers can be used to better understand population risk in non-European populations, (b) to measure their association with racial/ethnic variation in disease risk, and (c) to test a prediction of the Goldstein “common SNP, rare mutation” hypothesis [19]–[21].
To allow for comparability of estimates of genetic risk among racial/ethnic groups requires large studies comprised of cases and controls defined using identical criteria and sampled ideally from the same study population. In the present study, we, as part of the Population Architecture using Genomics and Epidemiology (PAGE) Study, examined genetic associations with 19 validated risk alleles for T2D in European American, African American, Latino, Japanese American, and Native Hawaiian T2D cases (n = 6,142) and controls (n = 7,403) from the population-based Multiethnic Cohort study (MEC). We also evaluated whether these variants can be utilized to model the genetic risk of T2D in each population and their association to disparities in risk.
The age of the cases and controls ranged from 45 to 77 at cohort entry, with the mean age of cases (mean 59.0 years) being essentially the same as the controls (mean 58.8 years), and African Americans being on average the oldest (mean 60.2 years) and Native Hawaiians the youngest (mean 55.6 years). Compared to controls, cases were heavier, more likely to be a current or former smoker, less physically active and had fewer years of education (Table 1). Compared to the other groups, the Japanese were leaner (for cases and controls, men and women).
The established T2D risk SNPs were polymorphic (frequency>0.05) in all racial/ethnic groups (Figure 1), except for rs10923931 (NOTCH2) in Japanese and Native Hawaiians and rs7903146 (TCF7L2) in Japanese (Table 2). In European populations these 19 SNPs have very modest odds ratios (1.1–1.3 per copy of the risk allele), and required studies of more than ten thousand cases and controls to reach genome-wide significance [3]–[13], [22]. Our sample sizes, although substantial, provided limited power to detect these modest effects (Table S1; power to achieve nominal significance (P = 0.05) of 34%, 47%, 67%, 54%, and 33%, in European Americans, African Americans, Latinos, Japanese Americans, and Native Hawaiians, respectively).
We first assessed whether the “risk allele” of each SNP was associated in the same direction (odds ratios>1) in each ethnic group. Whereas the null hypothesis is that 50% of “risk” alleles would trend in the same direction, we observed from 12 (63%; P = 0.18; binomial probability) in European Americans to 19 (100%; P = 1.9×10−6) in Japanese Americans. The number of these associations that reached nominal significance (P<0.05) ranged from 3 (P = 0.067; binomial probability) in Native Hawaiians to 10 (P = 5.9×10−9) in Japanese (Table 2). For the majority of alleles with positive associations, odds ratios for homozygous carriers were greater than for heterozygous carriers in each population, which provides support for their associations and allele dosage effects (Table S2). In African Americans, results were similar after adjustment for percent European ancestry (Table S3). Adjustment for education, a proxy for socio-economic status (SES) and European ancestry, did not influence the results (Table S4) [23].
We next performed analyses that combined evidence for association across the five ethnic groups. In this analysis the power to achieve nominal significance for the allelic effects reported previously was >80% for 18 out of 19 alleles (average 94%; Table S1). In this analysis all 19 (100%; P = 1.9×10−6, binomial probability) variants were associated with risk in the same direction as the initial report (odds ratios>1) and 14 (P = 5.7×10−15; binomial probability) with nominal statistical significance (P<0.05). All 19 associations remained in the same direction as previous reports (OR>1) and 13 of the variants were significantly associated with T2D risk when the European American subjects were excluded from the analysis. The association of rs8050136 in FTO was attenuated by adjustment for BMI (odds ratio (95% confidence interval), 1.06(1.00–1.11) prior to adjustment; 1.02(0.96–1.08) after adjustment). Only 5 of the 19 risk variants showed nominal evidence for heterogeneity in the odds ratio across ethnic groups, and only one of these (CDKAL1) was significant after correction for having performed 19 tests (PPARG, rs1801282, Phet = 0.048; WFS1, rs10010131, Phet = 0.032; CDKAL1, rs7754840, Phet = 6.2×10−4; HHEX, rs1111875, Phet = 0.037; and, HNF1B, rs4430796, Phet = 0.043; Table 2).
We next calculated a summary risk score comprised of an unweighted count of the 19 risk-associated alleles. The average increment in risk per allele was generally similar in all populations, except Japanese Americans, where the effect of each allele was nearly double that observed in Europeans ((odds ratio, 95% confidence interval): African Americans, 1.09, 1.05–1.12; (P = 3.0×10−6); Native Hawaiians, 1.10, 1.06–1.15 (P = 1.2×10−5); European Americans, 1.11, 1.06–1.17 (P = 1.2×10−5); Latinos, 1.12, 1.09–1.14 (P = 7.5×10−19); and, Japanese, 1.20, 1.17–1.24; (P = 7.0×10−32); Phet = 3.8×10−4). Results were similar when limiting the analysis to individuals with complete genotype data for all variants and when including only those markers associated with risk (at P<0.10) (Table S5). Individuals in the top quartile of the risk allele distribution were at 1.6 (African Americans, P = 5.3×10−4) to 3.1-fold (Japanese Americans, P = 7.9×10−26) greater risk of diabetes compared to those in the lowest quartile (Table 3).
Using these ethnic-specific per allele odds ratio estimates and the aggregate risk allele counts, we built a quantitative risk model to compare the distribution of genetic risks between populations associated with these marker alleles. The higher average number of risk alleles in African Americans caused their distribution to be slightly right shifted (towards higher log ORs) compared to European Americans, however their relatively smaller per allele odds ratio resulted in wide overlap with the European American distribution (Figure 2). The Japanese Americans had a wider distribution of risk because of the large per allele odds ratio, but the low average risk allele counts caused the Japanese distribution to be left-shifted (towards lower log ORs) compared to European Americans. The distributions for Latinos and Native Hawaiians were very similar to the European Americans.
We tested 19 common genetic risk markers that were discovered in European populations. We found that association with all 19 of these SNPs trended in the same direction in this large multiethnic study, and the majority of these variants were nominally significant in their association with diabetes risk. A risk score comprised of these alleles was significantly associated with diabetes risk in all five racial/ethnic groups, with the only significant heterogeneity being larger effect sizes in Japanese Americans. However, in comparing the distribution of risk conferred by these alleles between populations we found that they explain little, if any, of known differences in the prevalence of diabetes between these populations.
These observations indicate that most, if not all, of these alleles show directionally similar association to T2D across many populations. Such a pattern indicates that the causal alleles at these validated risk loci (which have yet to be found) likely predate the migrations that separated these populations now residing in Europe, Africa, East Asia, the Pacific Islands and the Americas. We note that this pattern is unexpected under the recently described “common SNP, rare mutation” model of Goldstein that suggests that GWAS signals with common alleles for T2D and other diseases may be “synthetic associations” created by one or more rare alleles [19]–[21]. Under the Goldstein Hypothesis the consistent associations that we noted at these loci across populations would only be observed if, in each population, one or more distinct rare alleles arose at each locus, and they happened to arise each time on the same haplotype background. Although possible, this scenario seems unlikely, and a more parsimonious explanation would be the “synthetic association” hypothesis of Goldstein does not apply to a majority of these T2D SNPs.
The modest number of cases and controls in this study (as compared to the initial discovery studies) likely underlies the lack of statistically significant associations in some groups. Weaker associations in some racial/ethnic groups may also be due to differences in allele frequencies, linkage disequilibrium, and environmental and genetic modifiers. In two cases (WFS1 and CDKAL1), significant heterogeneity by race/ethnicity reflected a lack of association in African Americans, perhaps because of lower linkage disequilibrium between the marker and the biologically relevant allele.
It is interesting that the odds ratios observed for these marker SNPs were larger in Japanese Americans than in the original discovery cohorts, and in the other ethnic groups in our study. A meta-analysis of 7 association studies in Japanese populations replicated associations from studies in European populations for 7 loci under study (TCF7L2, CDKAL1, CDKN2B, IGF2BP2, SLC30A8, KCNJ11, and HHEX) [24]. A recent GWAS in Japanese observed significant associations in KCNQ1 as well as these same 7 loci and, similar to our observations, noted magnitudes of effect that were generally stronger than previously observed in European populations [25]. Additional studies in other Asian populations have replicated associations with many of these loci as well [24], [26]–[28].
In the Multiethnic Cohort, we have found the prevalence of T2D to be at least 2-fold higher in African Americans, Latinos, Japanese and Native Hawaiians compared to European Americans, with these differences being independent of body weight [14]. We examined the extent to which the known genetic risk alleles for diabetes could explain these disparities by quantifying and comparing the relative risk distributions between populations. Compared to European Americans, we did not observe evidence of greater genetic risk in any population. Our findings therefore indicate that these risk markers explain little, if any, of racial/ethnic disparities in T2D prevalence. It remains possible that the actual causal alleles in these regions may be more common in frequency and/or have larger effects than the index signals in non-European populations. As seen with KCNQ1 [1], [2], GWAS in non-European populations are effective in discovering risk loci that are important in multiple populations but difficult to identify in European populations where the alleles are rare.
This study had a number of limitations. First, a self-report of diabetes and use of medication for diabetes was used to define cases and controls. We observed that approximately 1% of a random sample of the controls in this study had HbA1C levels above 7.0%, which suggests that only a small portion of controls had undiagnosed diabetes (see Materials and Methods). Also, our case definition did not differentiate between T1D and T2D, however we expect this misclassification to be minor as <3% of T2D cases had a previous diagnosis of T1D based on other sources (see Materials and Methods). The highly consistent findings of this study, as compared to the discovery GWAS reports, argue that our phenotypic characterization is adequate to observe the association to T2D.
Some caution should also be given to the interpretation of the risk modeling conducted in each ethnic group, as the genetic markers included are unlikely to be the causal alleles. Future fine-mapping and sequencing studies to identify the functional variants (common and/or rare) and large-scale testing of each allele will be required to more precisely model risk as well as assess differences in the distribution of genetic risk across populations.
Another limitation is that we did not account for the potential confounding effects of population stratification. However, odds ratios were essentially unchanged after adjusting for global European ancestry in a subset of African Americans (336 cases 397 controls) for whom ancestry markers were available, suggesting that effects due to population substructure were not substantial, at least in this group. We also noted that controlling for education, a proxy for SES which has been shown to be significantly associated with Native American ancestry in Latinos [23], had little effect on the associations with these risk alleles. Furthermore, the risk alleles were not generally more frequent in Latinos than in European Americans which would be likely if these alleles were proxies for more general ancestry differences. While population stratification is unlikely to fully explain these findings, it remains possible that at some loci, the causal alleles may be more correlated with ancestry than the index SNPs.
In summary, our data provide strong support for common genetic variation contributing to T2D risk in multiple populations. Our findings in T2D do not support the theory that GWAS signals are due to rare alleles. Nonetheless, GWAS and sequencing studies in these and other racial/ethnic populations are needed to reveal a more complete spectrum of risk alleles that are important globally as well as those that may contribute to risk disparities.
The Institutional Review Boards at the University of Southern California and University of Hawaii approved the study protocol.
The MEC consists of 215,251 men and women, and comprises mainly five self-reported racial/ethnic populations: European Americans, African Americans, Latinos, Japanese Americans and Native Hawaiians [29]. Between 1993 and 1996, adults between 45 and 75 years old were enrolled by completing a 26-page, self-administered questionnaire asking detailed information about dietary habits, demographic factors, level of education, personal behaviors, and history of prior medical conditions (e.g. diabetes). Potential cohort members were identified through Department of Motor Vehicles drivers' license files, voter registration files and Health Care Financing Administration data files. In 2001, a short follow-up questionnaire was sent to update information on dietary habits, as well as to obtain information about new diagnoses of medical conditions since recruitment. Between 2003 and 2007, we re-administered a modified version of the baseline questionnaire. All questionnaires inquired about history of diabetes, without specification as to type (1 vs. 2). Between 1995 and 2004, blood specimens were collected from ∼67,000 MEC participants at which time a short questionnaire was administered to update certain exposures, and collect current information about medication use.
Cohort members in California are linked each year to the California Office of Statewide Health Planning and Development (OSHPD) hospitalization discharge database which consists of mandatory records of all in-patient hospitalizations at most acute-care facilities in California. Records include information on the principal diagnosis plus up to 24 other diagnoses (coded according to ICD-9), including T1D and T2D. In Hawaii cohort members have been linked with the diabetes care registries for subjects with Hawaii Medical Service Association (HMSA) and Kaiser Permanente Hawaii (KPH) health plans (∼90% of the Hawaiian population has one of these two plans) [15]. Information from these additional databases have been utilized to assess the percentage of T2D controls (as defined below) with undiagnosed T2D, as well as the percentage of identified diabetes cases with T1D rather than T2D. Based on the OSHPD database <3% of T2D cases had a previous diagnosis of T1D. We did not use these sources to identify T2D cases because they did not include information on diabetes medications, one of our inclusion criteria for cases (see below).
In this study, diabetic cases were defined using the following criteria: (a) a self-report of diabetes on the baseline questionnaire, 2nd questionnaire or 3rd questionnaire; and (b) self-report of taking medication for T2D at the time of blood draw; and (c) no diagnosis of T1D in the absence of a T2D diagnosis from the OSHPD (California Residents). Controls were defined as: (a) no self-report of diabetes on any of the questionnaires while having completed a minimum of 2 of the 3 (79% of controls returned all 3 questionnaires); and (b) no use of medications for T2D at the time of blood draw; and (c) no diabetes diagnosis (type 1 or 2) from the OSHPD, HMSA or KPH registries. To preserve DNA for genetic studies of cancer in the MEC, subjects with an incident cancer diagnosis at time of selection for this study were excluded. Controls were frequency matched to cases on age at entry into the cohort (5-year age groups) and for Latinos, place of birth (U.S. vs. Mexico, South or Central America), oversampling African American, Native Hawaiian and European American controls to increase statistical power.
Fasting glucose (FG) and HbA1C measurements were used to validate the case-control selection criteria. Among 185 T2D cases and 1,048 controls who met the T2D case-control definitions above and with FG measurements available from ongoing studies in the MEC, 57% of cases (ranging from 43% in European Americans to 63% in Japanese Americans) and 3% of controls (ranging from 1% in African Americans to 6% in Latinos) had a FG value >125 mg/dl. We also measured HbA1C (ARUP Laboratories, Salt Lake City, Utah) in 50 cases and 50 controls per each sex-ethnic group. Just over 1% (6/500) of controls were likely to have unreported T2D (HbA1C value ≥7%). In contrast, ∼47% (234/500) of T2D cases had HbA1C ≥7% (ranging from 41% in European Americans to 57% of Native Hawaiians). Since hypoglycemic medication use was part of the case selection criteria, some cases were expected to have FG and HbA1C levels in the normal range.
Altogether, this study included 6,142 T2D cases and 7,403 controls (European American (533/1,006), African American (1,077/1,469), Latino (2,220/2,184), Japanese American (1,736/1,761) and Native Hawaiian (576/983)). Genotyping was conducted by the TaqMan allelic discrimination assay (Applied Biosystems, Foster City, CA) [30]. For all SNPs, genotype call rates were >95% among case and control groups in each population and HWE p-values among controls were >0.05 in at least 4 of the 5 ethnic groups and none of the values were <0.01 (Table S6). Subjects missing data for >5 SNPs (n = 82) were removed from the analysis.
Odds ratios and 95% confidence intervals were calculated for each allele in unconditional logistic regression models while adjusting for age at cohort entry (quartiles), body mass index (BMI, kg/m2, quartiles), sex, and race/ethnicity (pooled analysis) in ethnic-stratified and pooled analyses. Associations with the two variants at KCNQ1 were examined adjusting for the other allele. Potential confounding factors including, smoking history, education, physical activity, and history of hypertension were evaluated but did not influence the results. Potential confounding by percent European ancestry was examined in a subset of African American men (336 cases, 397 controls) with available genetic ancestry information [16], [31], [32].
We also modeled the cumulative genetic risk of T2D using these markers. We summed the number of risk alleles for each individual and estimated the odds ratio per allele for this aggregate unweighted allele count variable as an approximate risk score appropriate for unlinked variants with independent effects of approximately the same magnitude for each allele. We also examined a second model where each allele was weighted and multiplied by the log of the published odds ratio prior to summing all alleles. The results of the more parsimonious unweighted risk score is presented as the two risk scores were highly correlated in each ethnic group (Pearson r≥0.92) and similar associations with T2D risk were observed for each score. For individuals missing genotypes for a given SNP, we assigned the average number of risk alleles within each ethnic group (2× risk allele frequency) to replace the missing value for that SNP. We used these ethnic-specific per allele summary odds ratios and the total number of risk alleles among control subjects to estimate the distribution of relative risks conveyed by all risk alleles. To avoid making the reference group carriers of zero risk alleles (a group which does not exist) we centered the distribution on the mean number of risk alleles observed in the control population (18.5). The log relative risk for each subject was calculated as logRR = (RA−18.5)×log(ORi) (where RA is equal to the subject's total risk alleles and log(ORi) is the log of the ethnic specific per allele odds ratio. A spline function was used to capture the shape of the distributions of log OR for display purposes. Two variants in KCNQ1 were included in the risk modeling because both were significantly associated with T2D when co-modeled (results were similar when only the most significant of the two, rs2237897, was included). The variant in FTO was excluded from risk modeling procedures, as we found (as have others) that it is not a risk factor for diabetes independent of its effect on obesity.
|
10.1371/journal.pntd.0002437 | Concurrent Infections of Giardia duodenalis, Enterocytozoon bieneusi, and Clostridium difficile in Children during a Cryptosporidiosis Outbreak in a Pediatric Hospital in China | Over 200 cryptosporidiosis outbreaks have been reported, but little is known if other enteric pathogens were also involved in some of these outbreaks. Recently, an outbreak of cryptosporidiosis linked to poor hygiene by two Cryptosporidium hominis subtypes occurred in a pediatric hospital ward (Ward A) in China, lasting for more than 14 months. In this study, the concurrence during the outbreak of three other enteric pathogens with a similar transmission route, Giardia duodenalis, Enterocytozoon bieneusi, and Clostridium difficile, was assessed.
The occurrence of G. duodenalis, E. bieneusi, and C. difficile in 78 inpatients from Ward A and 283 and 216 inpatients from two control wards (Wards C and D) in the same hospital was examined using molecular diagnostic tools. Significantly higher infection rates were found in children in Ward A for all study pathogens than in Wards C and D (P<0.01): 9.5% versus 1.4% and 0% for G. duodenalis, 10.8% versus 2.8% and 3.7% for E. bieneusi, and 60.8% versus 37.8% and 27.8% for C. difficile, respectively. These differences were mostly seen in children ≤12 months. Enteric pathogen-positive children in Ward A (31/58 or 53.4%) were more likely to have mixed infections than those in Ward C (4/119 or 3.4%) or D (5/68, 7.4%; P<0.01). Having cryptosporidiosis was a risk factor for G. duodenalis (OR = 4.3; P = 0.08), E. bieneusi (OR = 3.1; P = 0.04), and C. difficile (OR = 4.7; P<0.01) infection. In addition, a lower diversity of G. duodenalis, E. bieneusi, and C. difficile genotypes/subtypes was observed in Ward A.
Data from this study suggest that multiple pathogens were concurrently present during the previous cryptosporidiosis outbreak. Examination of multiple enteric pathogens should be conducted when poor hygiene is the likely cause of outbreaks of diarrhea.
| The transmission of Giardia duodenalis, Enterocytozoon bieneusi, and Clostridium difficile is poorly understood in developing countries despite their wide occurrence. Because they are transmitted by the same fecal-oral route as Cryptosporidium, in this study, we have examined the occurrence of these enteric pathogens in children during a cryptosporidiosis outbreak in a pediatric hospital in China. Using molecular diagnostic tools, we have detected significantly higher infection rates of these enteric pathogens in the outbreak ward than in two control wards in the same hospital. We have also shown a much higher occurrence of these pathogens in children having cryptosporidiosis than those having no cryptosporidiosis. We have demonstrated that the genetic diversity of enteric pathogens is much lower in the outbreak ward than in control wards. Therefore, other enteric pathogens are concurrently present during the cryptosporidiosis outbreak, and examinations for multiple enteric pathogens should be conducted when poor hygiene is considered the likely cause of outbreaks of diarrhea.
| Cryptosporidium is a significant cause of diarrhea in humans worldwide [1]. Humans can acquire Cryptosporidium infections through the fecal-oral route via direct person-to-person or animal-to-person contact, or ingestion of contaminated water or food [2]. Thus far, over 200 waterborne, foodborne, person-to-person, and zoonotic cryptosporidiosis outbreaks have been reported [3], [4]. However, whether other co-pathogens were involved in some of these outbreaks remains largely unexamined.
Similar to Cryptosporidium, pathogens like Giardia duodenalis, Enterocytozoon bieneusi, and Clostridium difficile are also significant causes of diarrhea in humans worldwide and can be transmitted from persons to persons by the same fecal-oral route involved in cryptosporidiosis occurrence [1], [5], [6]. All of these pathogens are major causes of healthcare-associated infections, especially Clostridium difficile [7]–[9]. Despite their wide occurrence, the epidemiology of these enteric pathogens is largely unclear in developing countries. Only limited data exist on the molecular epidemiology of these pathogens in China [10]–[16].
In one recent molecular epidemiologic study on Cryptosporidium in in-patients from three pediatric hospitals, P. R. China, we identified an extended outbreak of cryptosporidiosis in a pediatric hospital ward (Ward A, Hospital I), with more than 50% (38/74) children affected by two C. hominis subtypes (IaA14R4 and IdA19) during a 14-month period (Sep. 2007–Oct. 2009) [17]. The infection rate in Ward A was significantly higher than the overall rates in Hospitals I (2.8%), II (0.6%) and III (0.4%). The diversity of Cryptosporidium species and C. hominis subtypes were significantly lower in Ward A than in other wards/hospitals, with only one species (C. hominis) and two C. hominis subtypes (IaA14R4 and IdA19) being found in 38 patients in Ward A while four species of Cryptosporidium and six C. hominis subtypes being found in 62 patients in other wards and hospitals [17].
Because concurrent infections of multiple pathogens are sometimes involved in gastroenteritis in hospitalized children [18], [19], in the present study, we retrospectively compared the infection rates and subtype distribution of G. duodenalis, E. bieneusi, and C. difficile in hospitalized children in Ward A with those in two control wards in the same hospital: Ward C for patients having hemophilia, anemia, and neurological diseases, and Ward D for patients having general surgeries. This was the first study to use genotyping and subtyping tools to investigate the transmission of multiple enteric pathogens during a cryptosporidiosis outbreak.
Written informed consent was obtained from the parents or guardians of the children. This study was approved by the Ethics Committee of the East China University of Science and Technology.
All specimens for this study were collected from in-hospital children during September 2007–October 2009 as described [17]. These children were hospitalized mostly due to non-gastrointestinal illness: Ward A for patients with various congenital or inherited diseases from a local welfare institute; Ward C for children attending the Department of Endocrinology, Hematology and Neurology; and Ward D for children attending the Department of General Surgery.
In this study, Ward A (Cryptosporidium infection rate = 51.4%), where the cryptosporidiosis outbreak occurred, was regarded as the case ward, while two other wards (Wards C and D; Cryptosporidium infection rates = 1.8% and 2.3%, respectively) in the same hospital (Hospital I in Shanghai, China) without cryptosporidiosis outbreak were regarded as the control wards. Overall, 573 children, including 74 from Ward A (age range: 1–192 months; mean age: 20.7 months), 283 from Ward C (age range: 1–168 month; mean age: 41.3 months), and 216 from Ward D (age range: 1–216 months; mean age: 43.8 months), were examined for the occurrence and genotype/subtype distribution of G. duodenalis, E. bieneusi, and C. difficile. In addition, 2,672 children from other known or unknown wards in Hospital I (age range: 0–228 months; mean age: 46.9 months), 489 children from Hospital II (age range: 0–192 months; mean age: 37.2 months age), and 311 children from Hospital III (age range: 1–159 months; mean age: 40.4 months) in the same city, were also examined for G. duodenalis. Information on age, gender, and the occurrence of diarrhea as defined by the attending physicians was collected for each patient as previously described [17].
Genomic DNA was extracted from 0.2 ml of fecal materials using a FastDNA SPIN Kit for Soil (BIO 101, Carlsbad, CA). To detect G. duodenalis, a 532-bp fragment of the triosephosphate isomerase (tpi) gene was amplified by nested PCR [20]. A 511-bp fragment of the β-Giardia (bg) and a 530-bp fragment of the glutamate dehydrogenase (gdh) gene were further amplified from DNA of the tpi-positive specimens [21], [22]. Giardia duodenalis genotypes and subtypes were determined using the established nomenclature system based on multilocus sequence data [22].
A ∼392-bp fragment of the rRNA gene containing the entire internal transcribed spacer (ITS) was amplified and sequenced to detect and identify E. bieneusi genotypes [23]. Genotypes of E. bieneusi were named according to established nomenclature [23], [24]. A PCR based on the tcdB gene was used to detect C. difficile [25]. Clostridium difficile in tcdB-positive specimens was subtyped by sequence analysis of the slpA gene as previously described [9].
All positive PCR products generated in the study were directly sequenced using Big Dye Terminator v3.1 Cycle Sequencing Kits (Applied Biosystems, Foster City, CA) and an ABI 3130 Genetic Analyzer (Applied Biosystems). Sequences were assembled using ChromasPro (version 1.5) software (http://technelysium.com.au/?page_id=27). The accuracy of the sequencing reads was confirmed by bidirectional sequencing. The nucleotide sequences of G. duodenalis, E. bieneusi, and C. difficile genotypes/subtypes obtained were aligned with reference sequences of each genetic locus downloaded from GenBank using ClustalX (http://www.clustal.org/). A neighbor-joining analysis of the aligned sequences was performed with the program Mega 5 (http://www.megasoftware.net/). Unique nucleotide sequences generated from the study were deposited in GenBank under accession numbers JX994231-JX994292.
The χ2 test was used to compare infection rates between Ward A and the control wards. The same method was used to analyze the association between infection and age, gender, or diarrhea status. The strength of the association was measured using the odds ratio (OR). Differences were considered significant at P≤0.05. All statistical analyses were performed using the SPSS Statistics 17.0 (SPSS Inc, Chicago, IL).
Only four parasites including Cryptosporidium, G. duodenalis, E. bieneusi and C. difficile were analyzed in this study and no examinations of bacteria or viruses were conducted. The Cryptosporidium infection rates were 51.4% (38/74) in case Ward A while 1.8% (5/283) and 2.3% (5/216) in control Wards C and D, respectively [17]. Among the 74 specimens from the case Ward A, seven (9.5%) were positive for G. duodenalis at the tpi locus (Figure 1). In contrast, only 4 of 283 (1.4%) and 0 of 216 (0%) specimens from the control Wards C and D were positive (Figure 1). The difference in G. duodenalis infection rates between the case (A) and control wards (C and D) was significant (P<0.01; Table 1). In addition, 4 of 1,019 (0.4%) children from other wards in Hospital I, 17 of 1,653 (1.0%) children from unknown wards in Hospital I, 3 of 489 (0.6%) children from Hospital II, and 5 of 311 (1.6%) children from Hospital III were also positive for G. duodenalis. The prevalence of giardiasis in children from these locations was significantly lower than the prevalence in Ward A (P<0.01).
Among the 573 specimens examined, 24 (4.2%) were positive for E. bieneusi at the ITS locus, with eight positives in each ward. The infection rate of E. bieneusi in Ward A (10.8%) was significantly higher than those in Ward C (2.8%) and D (3.7%) (P = 0.01; Figure 1; Table 1). Altogether, 212 of the 573 specimens were positive for C. difficile at the tcdB locus, with 45/74 (60.8%), 107/283 (37.8%), and 60/216 (27.8%) in Wards A, C, and D being positive, respectively (Figure 1). The infection rate of C. difficile was significantly higher in Ward A than in Wards C and D (P<0.01; Table 1).
Concurrent infections of multiple pathogens, including Cryptosporidium, G. duodenalis, E. bieneusi, and C. difficile, were detected in both the case and control wards. Comparing with the control Wards C and D, Ward A had a significantly higher overall infection rate of enteric pathogens (58/74 or 78.4% versus 119/283 or 42.0% and 68/216 or 31.5%; P<0.01). Over half of children with enteric pathogens in Ward A (31/58 or 53.4%) were concurrently infected with two or more pathogens, while only a small number of children with enteric pathogens in Ward C (4/119 or 3.4%) or D (5/68, 7.4%) were infected with multiple pathogens (P<0.01).
In this study, children who had cryptosporidiosis during the outbreak were more likely to be infected with other enteric pathogens (Table 1). Among the 573 children examined for all four pathogens, 48 were previously diagnosed as having cryptosporidiosis. These Cryptosporidium-positive children had higher infection rates of G. duodenalis (6.3% versus 1.5%; P = 0.08), E. bieneusi (10.4% versus 3.6%; P = 0.04), and C. difficile (70.8% versus 33.9%; P<0.01) than Cryptosporidium-negative children (Table 1).
The age distribution of G. duodenalis, E. bieneusi, and C. difficile infections in 573 children from Wards A, C, and D is shown in Table 2. Infection rates of G. duodenalis were similar among all age groups (P>0.05). In contrast, children ≤6 months were more likely infected with E. bieneusi (11/99 or 11.1% versus 12/473 or 2.5% for other age groups, P<0.01), and children ≤12 months were more likely infected with C. difficile (124/277 or 44.8% versus 88/295 or 29.8% for other age groups, P<0.01; Table 2). Among children under 12 months, infection rates of all three study pathogens were significantly higher in Ward A than in control wards (5/59 or 8.5% versus 1/218 or 0.5%, P<0.01 for G. duodenalis; 7/59 or 11.9% versus 8/218 or 3.7%, P = 0.03 for E. bieneusi; 38/59 or 64.4% versus 86/218 or 39.4%, P<0.01 for C. difficile). In contrast, in children older than 12 months, only G. duodenalis was significantly more prevalent in Ward A than in the controls (2/15 or 13.3% versus 3/280 or 1.1%, P = 0.01 for G. duodenalis; 1/15 or 6.7% versus 7/280 or 2.5%, P = 0.88 for E. bieneusi; 7/15 or 46.7% versus 81/280 or 28.9%, P = 0.14 for C. difficile; Table 2). No gender difference was seen in the occurrence of G. duodenalis, E. bieneusi, and C. difficile infections in this study (P>0.05; Table 2).
The distribution of G. duodenalis multilocus subtypes was different between case and control wards. In Ward A, six of the seven specimens positive for G. duodenalis at the tpi locus were also positive at the bg and gdh loci, and all of them belonged to the multilocus subtype AII (Table 3; Figure 2). In contrast, both multilocus subtype AII and subtypes belonging to the assemblage B (2 cases each) were found in Ward C. Similarly, both AII and B were detected in other known or unknown wards in Hospital I, and in Hospitals II and III (Table 3).
Four known genotypes of E. bieneusi were found in this study, with Peru 11 as the dominant one (in 6 cases). The other three includes EbpC (1 case), EbpA (2 cases), and D (1 case). Twelve novel genotypes (SH1–12) were found in this study, with SH2 in three cases and all other genotypes in one case each (Table 3). All 16 E. bieneusi genotypes except SH5 belonged to Group 1 phylogenetically, while genotype SH5 belonged to Group 2 (Figure 3). Higher occurrence of E. bieneusi genotype Peru 11 was seen in Ward A (4/8 genotyped) than in Wards C and D (1/8 genotyped each; Table 3).
Among the 212 C. difficile-positive specimens based on PCR analysis of the tcdB gene, 160 specimens were subtyped at the slpA locus successfully. In total, 20 slpA subtypes were obtained, including 8 novel ones (Table 3). Most of the novel subtypes were genetically close to subtypes previously reported, although two of them, sh-01 and sh-02, had very different sequences and formed an independent clade in the phylogenetic tree (Figure 4). The most common subtype in Ward A was fr-01 (15/40 slpA-positive cases), compared to kr-03 in control Wards C (23/74 slp-A positive cases) and D (18/46 slp-A positive cases; Table 3).
A significantly higher diarrhea rate was observed in Ward A than in control Wards C and D (43/74 or 58.1% versus 180/499 or 36.1%, P<0.01). Infection with Cryptosporidium was significantly associated with the occurrence of diarrhea (OR = 1.95, p = 0.002). However, a large number of asymptomatic G. duodenalis, E. bieneusi, and C. difficile infections were observed in both case and control wards in this study(Table 2). None of the three pathogens were significantly associated with the occurrence of diarrhea in the pediatric inpatients (P>0.05; Table 2). None of the dominant genotypes/subtypes of the three study pathogens were significantly associated with the occurrence of diarrhea (P>0.05; data not shown). In addition, in Ward A, the difference in diarrhea rates between children with multiple infections and children with single infection was not significant (17/31 or 54.8% versus 17/27 or 63.0%, P = 0.51). This was also the case in the control Wards C and D (3/9 or 33.3% versus 56/178 or 31.5%, P = 1.0).
Molecular epidemiological investigations have improved our understanding of the transmission of enteric pathogens, including those examined in the present study [2], [7]–[9], [14]. They are especially useful in identifying the occurrence of outbreaks, linking seemingly un-associated cases, and tracking infection sources. In the present study, using genotype and subtype tools, we retrospectively identified concurrent transmission of G. duodenalis, E. bieneusi, and C. difficile during a cryptosporidiosis outbreak previously identified in Ward A of Hospital I in Shanghai, China. This is reflected by higher infection rates and lower genetic diversity of these enteric pathogens in Ward A than in control wards.
The low infection rates of G. duodenalis in wards other than Ward A in Hospital I (0–1.4%), Hospitals II (0.6%), and III (1.6%) are similar to those reported in out-patients and inpatients (0.2–0.6%) and the general population (2.5%) in China [11], [26]–[29]. Likewise, infection rates of E. bieneusi in hospitalized children in Wards C (2.8%) and D (3.7%) are also very low. In contrast, significantly higher infection rates of G. duodenalis (9.5%; P<0.01) and E. bieneusi (10.8%; P = 0.01) were seen in Ward A, indicating that these pathogens were transmitted frequently within this ward during the cryptosporidiosis outbreak. Although a high carriage of C. difficile was found in control Wards C (37.8%) and D (27.8%), this is similar to the 16–35% carriage of C. difficile in hospital inpatients in other countries [30], and significantly lower than the infection rate of C. difficile in Ward A (60.8%; P<0.01). Previously, no data existed on the prevalence of C. difficile in hospitalized children in China, although an infection rate of 9.5% was reported in adults in Shanghai [14].
The low genetic diversity of G. duodenalis, E. bieneusi, and C. difficile found in Ward A versus in control wards also supports the hypothesis of concurrent transmission of these enteric pathogens during the cryptosporidiosis outbreak. For G. duodenalis, AII was the only subtype seen in Ward A, although assemblage B and both AI and AII subtypes of assemblage A are commonly found in humans around the world (7), and they were all found in other wards and hospitals in the present study (Table 3). For E. bieneusi, Peru 11 was the dominant genotype in Ward A, being found in half of the genotyped specimens, while it was only found in one of eight specimens each genotyped in Wards C and D (Table 3). For C. difficile, the fr-01 subtype was dominant in Ward A and accounted for 1/3 of all C. difficile infections in this ward, whereas the most prevalent genotype kr-03 in control Wards C and D accounted for less than 1/5 of all C. difficile infections (Table 3).
Outbreaks involving multiple enteric pathogens have been infrequently reported and, in the investigations of the few such outbreaks, sewage contamination of water or food was often the main cause for concurrent transmission of multiple enteric pathogens. For example, a waterborne outbreak of Shigella sonnei, Giardia, and Cryptosporidium infections on a Lake Michigan dinner cruise was caused by contamination of potable water with diluted sewage as the result of storm runoff in the cruise ship [31]. Another waterborne outbreak of gastroenteritis with multiple etiologies in resort island visitors and residents in Ohio in 2004 was caused by sewage contamination of groundwater [32]. Likewise, a national multi-pathogen outbreak of diarrheal illness in Botswana in 2006 was caused by sewage contamination of the environment during heavy rains in late 2005 and early 2006 [33]. Similarly, contact with manure from calves was responsible for two multi-pathogen outbreaks at a farm day camp in Minnesota [34]. In the present study, over half of the patients with enteric pathogens (31/58) in Ward A were infected with more than one pathogen, compared to a very limited number of cases (9/187) in control wards (P<0.01). Of note is the significant association of the three enteric pathogens examined in this study and occurrences of cryptosporidiosis in these children (Table 1).
As suggested in our previous investigation of cryptosporidiosis in these children [17], poor diaper changing and hand washing practices by caregivers were probably responsible for this multi-pathogen outbreak among pediatric inpatients in Ward A, Hospital I. Children in Ward A were orphans from a welfare institute. They were taken care of by hired caregivers. In contrast, children in other wards were primarily from the general community and cared for by their parents [17]. Considering the fact that most infections in Ward A occurred in children younger than 12 months (Table 2), who mostly stayed in cribs and beds, hired caregivers in Ward A might have acted as vehicles for the disease transmission among pediatric inpatients. This is also supported by the finding that in children under 12 months, Ward A had significantly higher infection rates of all study pathogens than Wards C and D, but in children older than 12 month, Ward A had only significantly higher infection rates of G. duodenalis than Wards C and D.
Very few studies have been conducted on molecular epidemiology of G. duodenalis, E. bieneusi, and C. difficile in China [10]–[16]. The occurrence of both assemblages A and B of G. duodenalis in non-outbreak children is in accordance with previous findings of near equal distribution of the two genotypes in 18 Giardia-positive humans in Henan [11] and 8 in Anhui [10]. In contrast, the dominance of Group 1 E. bieneusi genotypes in children in this study is different from the dominance of Group 2 genotypes in children in Jilin [13], although we also detected a novel Group 2 genotype in a child from Ward C (Figure 3; Table 3). The high diversity of known and novel E. bieneusi genotypes reported in this study and previous studies [12], [13] suggests that there is a need for more studies to examine the characteristics of E. bieneusi transmission in humans in China.
Ribotypes 027 and 078 are recognized as leading causes of nosocomial outbreaks of C. difficile infection in the world [9], [15]. However, neither has been reported in China thus far [14]–[16]. Interestingly, the dominant C. difficile slpA subtypes fr-01 in Ward A was previously characterized as toxin A-negative and toxin B-positive (A−B+), whereas the dominant subtype kr-03 in control wards was toxin A-positive and toxin B-positive (A+B+) [9]. In a previous study, A−B+ strains were the dominant ones (24.0%) in patients in three hospitals in Beijing, Shandong and Guangzhou in China [16]. The high prevalence of A−B+ strains in China indicates that toxin B, rather than toxin A, is probably a key virulence determinant as previously suggested [35]. Nevertheless, in the present study, no significant association was found between any of the C. difficile subtypes and the occurrence of diarrhea, although we previously showed a link between cryptosporidiosis and diarrhea in these children [17]. A new group of slpA subtypes including sh-01 and sh-02 were found in many children from all three wards (Figure 3; Table 3). Further studies are needed to better understand the public health importance of this new group of subtypes.
The results of this study and our previous study [17] showed that although Cryptosporidium infection was associated with the occurrence of diarrhea, single-pathogen infection with G. duodenalis, E. bieneusi, or C. difficile was not. None of the dominant genotypes/subtypes of G. duodenalis, E. bieneusi, and C. difficile were significantly associated with the occurrence of diarrhea, and concurrent infections of multiple pathogens were not more associated with occurrence of diarrhea than infections with single pathogens. The lack of differences in occurrence of diarrhea between children with single-pathogen infection and children with mixed infections in this study was probably attributable to the already high diarrhea rates in Ward A (58.1%) and low occurrence of mixed infections in control Wards C and D (9 cases of mixed infections versus 178 cases of single-pathogen infection). In addition, with the exception of Cryptosporidium, none of the other pathogens examined in this study were among the recently identified major pathogens for moderate-to-severe diarrhea in the Global Enteric Multicenter Study [1]. This has probably also made it difficult to use attributable fraction calculation in estimating the role of mixed infections in the occurrence of diarrhea in a hospital study setting with high occurrence of diarrhea.
In conclusion, using genotyping and subtyping tools we retrospectively identified a multi-pathogen outbreak in a pediatric hospital ward. As reported previously [17], this outbreak lasted ≥14 months, with ∼60 inpatient children affected by cryptosporidiosis. Most of the Cryptosporidium-positive children were co-infected with G. duodenalis, E. bieneusi, or C. difficile. The young age of affected children and concurrent infections with multiple enteric pathogens clearly implicated poor diaper changing and hand washing by hired caregivers as the cause of the outbreak. Thus, better training of caregivers on hygienic practices such as hand washing and proper use of disposable gloves and disinfectants is needed to reduce the risk of pathogen transmission in healthcare facilities. Results of this study also highlight the importance of molecular epidemiologic investigations in understanding the transmission of enteric pathogens in hospitals.
|
10.1371/journal.pgen.1002719 | Karyotypic Determinants of Chromosome Instability in Aneuploid Budding Yeast | Recent studies in cancer cells and budding yeast demonstrated that aneuploidy, the state of having abnormal chromosome numbers, correlates with elevated chromosome instability (CIN), i.e. the propensity of gaining and losing chromosomes at a high frequency. Here we have investigated ploidy- and chromosome-specific determinants underlying aneuploidy-induced CIN by observing karyotype dynamics in fully isogenic aneuploid yeast strains with ploidies between 1N and 2N obtained through a random meiotic process. The aneuploid strains exhibited various levels of whole-chromosome instability (i.e. chromosome gains and losses). CIN correlates with cellular ploidy in an unexpected way: cells with a chromosomal content close to the haploid state are significantly more stable than cells displaying an apparent ploidy between 1.5 and 2N. We propose that the capacity for accurate chromosome segregation by the mitotic system does not scale continuously with an increasing number of chromosomes, but may occur via discrete steps each time a full set of chromosomes is added to the genome. On top of such general ploidy-related effect, CIN is also associated with the presence of specific aneuploid chromosomes as well as dosage imbalance between specific chromosome pairs. Our findings potentially help reconcile the divide between gene-centric versus genome-centric theories in cancer evolution.
| Aneuploidy, the state of harboring an unbalanced number of chromosomes, has long been hypothesized to be at the basis of malignant transformation. Recent studies have also shown that aneuploidy is an important form of genome alteration underlying adaptive evolution of cells in response to harsh environments or genetic perturbations. In addition to the profound effect that aneuploidy has on gene expression and phenotype, another feature thought to contribute to aneuploidy's role in cancer and cellular evolution is the heightened chromosome instability of aneuploid cells. Since chromosome instability is the condition of gaining and losing chromosomes at a high frequency, this could lead to a vicious cycle in which aneuploidy could lead to further enhanced genetic diversity. Given the ever-changing and heterogeneous aneuploid cell populations, and the difficulty of separating the effect of aneuploidy from other types of genetic aberrations, the molecular mechanisms underlying aneuploidy-driven chromosome instability have remained largely unexplored. Here we describe the first unbiased and systematic investigation of chromosome instability associated with aneuploid genomes in the budding yeast Saccharomyces cerevisiae. Our results revealed both genome-level and chromosome-specific determinants of chromosome instability in aneuploid yeast. Our findings potentially help explain the molecular mechanism underlying a major source of genome instability in cancer.
| The nature of the genetic changes driving cellular evolution has been a central issue in both adaptive evolution of unicellular organisms and somatic evolution of cancer cells. Phenotypic variation, acting as a substrate of Darwinian selection and as an origin of phenotypic innovation, can be driven by sequence-based mutations as well as copy number changes [1], [2], [3], [4], [5], [6], [7]. In cancer, the gene-centric theory posits that cancer progression is driven by sequence alterations in specific genes playing key roles in cell cycle control and genome stability, leading to malignant growth and accumulation of further genetic aberrations [8]. Under this perspective, aneuploidy is more likely to be an innocent byproduct than a driver of the evolutionary process leading towards malignant transformation. The chromosome theory, on the other hand, emphasizes the cytogenetic diversity in cancer and proposes that it is the abnormal chromosome copy numbers, or aneuploidy, rather than variation in specific gene sequences, that accounts for both the loss of growth control and the remarkable adaptability of tumor cell populations toward restrictive tissue environments or chemotherapy [9]. According to this theory, aneuploidy would lead to increased rates of various types of genomic instability, including chromosome instability (CIN), and therefore a continuous ability to generate new adaptive aneuploid genomes. This potential snowballing effect has been termed “genome chaos” and has been hypothesized to be at the basis of malignant transformation [9], [10], [11]. While the gene-centric theory of cancer is widely accepted, understanding the mechanisms by which aneuploidy could lead to CIN might reconcile the two theories. For example, is the increased CIN in an aneuploid genome a result of the abnormal chromosome numbers per se or of aneuploidy-driven alteration of the expression of specific genes?
Whereas studying the contribution of aneuploidy to CIN in cancer cells is complicated by the fact that most cancer cells possess both numerous point mutations and other kinds of chromosome abnormalities [6], [12], [13], simple model organisms such as the yeast Saccharomyces cerevisiae represent valuable systems for assessing independent effects of individual types of genetic changes. Budding yeast cells are especially suitable for these types of studies because they tolerate aneuploidy relatively well [14], [15], most likely because their relatively small haploid genome (∼6,000 ORFs over ∼12 million base pairs) is segmented into a relatively large number of chromosomes (N = 16). Several studies have shown that aneuploid yeast cells not only are characterized by phenotypic variation but also exhibit genome instability [15], [16], [17], [18]. For example, two independent studies with congenic aneuploid strains obtained by sporulation of triploid or pentaploid yeast found that, while some of the aneuploid strains were relatively stable, the majority of the strains were chromosomally unstable [15], [17]. Another paper recently reported decreased artificial chromosome transmission fidelity and elevated mitotic recombination in a set of disomic yeast strains compared to the haploid parent [18]. However, none of these studies investigated into the cellular mechanisms by which an aneuploid karyotype causes CIN.
Insights into the mechanisms by which aneuploidy leads to CIN are important for understanding the dynamics of the cellular adaptation process and may ultimately enhance our ability to predict or modulate cancer progression. For example, as stable phenotypes are likely to require a certain degree of genetic stability, there may exist metastable aneuploid constellations amidst the genome chaos. If this was true, what may be the determinants underlying stable or unstable aneuploidy? Formally, aneuploidy could cause CIN through three conceptually distinct though not mutually exclusive mechanisms. First, as aneuploidy leads to varying degrees of growth impairment compared to euploids under standard culture conditions [14], [15], CIN may be induced by the cellular stress present under such conditions. If this hypothesis were correct, then CIN would correlate with the level of growth impairment in specific aneuploid strains. Second, according to the genome chaos theory, the more chromosomes are in aneuploidy in a genome the more unstable that karyotype is expected to be [19]. If this hypothesis were true, then CIN would correlate with how far a strain deviates from the nearest euploid state. Third, it is possible that aneuploid chromosome stoichiometry leads to dosage imbalance for specific genes encoding structural or regulatory components that ensure chromosome stability. This possibility was previously proposed based on imbalance of mitotic spindle components directly involved in chromosome segregation [20]. If this were correct, then correlations might be found between CIN and the relative copy numbers of specific chromosomes or combinations thereof.
In this paper we used an unbiased approach to examine the karyotypic features underlying CIN by generating random aneuploid karyotypes through triploid meiosis and following the dynamics of aneuploid populations with distinct original karyotypes. Our results support a model in which CIN is promoted by both genome-level and chromosome/gene-specific determinants.
Two methods were instrumental for the analysis explained below in this study. First, we used a high-throughput flow cytometry-based (FACS) assay to determine the overall genome content (referred to as ploidy) of a population of yeast cells. A non-integer ploidy revealed by FACS is likely to correspond to an aneuploid genome. However, FACS data is insufficient to reveal copy number for each chromosome. For this we used a recently established method for determining the relative copy number for each of the 16 yeast chromosomes that is based on quantitative polymerase chain reaction (qPCR) [15]. Combining the ploidy information revealed by FACS and chromosome stoichiometry revealed by qPCR allows determination of an aneuploid karyotype [15].
In our previous work [15], we generated isogenic aneuploid yeast strains with random chromosome stoichiometries as meiotic products from sporulated homozygous triploid or pentaploid strains of the S288c background. During meiosis I, chromosome segregation of an odd number of homologs leads to highly frequent aneuploid spore progenies with random karyotypes [16]. In this work, we took a similar approach to generate fully isogenic aneuploid yeast strains, and consistent with previous studies [15], [17], 45% of the aneuploid meiotic products were viable and gave rise to colonies (52 viable spores out of 116 expected from 29 tetrads). Unlike our previous study [15], however, where aneuploid strains with stable karyotypes were chosen for phenotypic comparison and gene expression analyses, in this study our goal was to follow karyotype changes for all (within our experimental limitations, see below) viable aneuploid spores resulting from triploid meiosis. Due to a lack of established methods for single-cell karyotyping in yeast, however, we devised a population-based approach (Figure 1A) that was predicated on the assumption that the modal karyotype of the population within a small colony reflects the karyotype of the cell that seeded the colony.
As illustrated in Figure 1A, the colony grown from each of the 52 viable spores was picked in its entirety after the spore had undergone ∼20 cell divisions, resuspended in liquid media and the actual cell number and thus the number of cell divisions was estimated (see Materials and Methods for details). As the aneuploid colonies grew at different rates, the colonies were picked at different times after tetrad dissection, and the time of colony picking was recorded. Each resulting culture at this time point was called generation 20 (g20) population sample. Due to contamination, only 47 g20 populations were obtained and further analyzed (see Figure S1). A small aliquot of the g20 population sample for each spore colony was used for FACS and qPCR karyotyping analysis, giving rise to the g20 population data (see below). If the initial aneuploid karyotype of a growing spore colony were unstable, karyotype heterogeneity would be expected within the g20 population. This karyotype heterogeneity could in turn allow us to estimate the level of CIN of the initial aneuploid karyotype (see below). To observe it, ∼200 cells from each g20 population sample were spread onto a YPD plate. As soon as the resulting colonies were visible, 11 colonies were randomly chosen from each plate (see Materials and Methods), harvested and frozen for prospective karyotype analysis by FACS and qPCR. These were referred to as the g20 colony samples, and a total of 47×11 = 517 such samples were harvested and stored.
Since some aneuploid karyotypes were more stable than others, to allow for more cell divisions that could give rise to karyotypic deviants, a part of each g20 population sample was further cultured in liquid for 5 and 10 more generations to yield g25 and g30 population samples, respectively, and the time when each of the cultures reached these numbers of generations was recorded. To estimate the number of cell divisions required for detecting karyotypic deviants, we performed computer simulations to calculate the fraction of cells with deviant karyotypes based on different chromosome mis-segregation rates at different generation times. We found that >10% of a cell population is expected to display a deviant karyotype after 20 generations (cell divisions) in the presence of a very high CIN level (1×10−3 chromosome mis-segregation per generation) or after 30 generations with a lower CIN level (5×10−4 chromosome mis-segregation per generation, Figure S2A and S2B). For a comparison, wild-type diploid and tetraploid yeasts were reported to have an artificial chromosome loss rate of 2.6×10−4 and 5.7×10−2 per generation, respectively [21]. Again, to determine karyotype diversity within each population, ∼200 cells from each g25 or g30 population were spread onto YPD plates, and 11 colonies from each plate were randomly selected and stored for later karyotyping by FACS and qPCR (Figure 1A). These were referred to as the g25 and g30 colony samples, and a total of 2×47×11 = 1034 such samples were obtained. We did not extend the above procedure beyond generation 30 due to the increasing effect of growth competition on karyotype diversity within each population. Nevertheless, our experimental design was likely to somewhat under-estimate the karyotype diversity for the aneuploid populations as a result of growth competition.
Having obtained and stored away all samples as described above, we first performed FACS analysis on all g20 population samples (Figure S3). One population (strain 221) displayed an apparent ploidy of 2.1 by FACS but subsequent qPCR karyotyping indicated that it was a hypo-diploid. Another g20 aneuploid population (strain 242) had a ploidy over 2N, possibly due to a whole-genome duplication event, and was not included in further analysis (Figure S1). Five of the g20 populations exhibited FACS profiles suggesting an extreme level of DNA content heterogeneity in the population, characterized by the presence of multiple broad peaks with no clear G1 and G2 peaks (Figure 1E and Figure S3). These 5 strains were not included in further analysis due to the difficulty to determine the karyotype makeup of the population (see Figure S1). Conversely, the FACS profiles of the remaining 41 g20 population samples displayed a more homogeneous, albeit aneuploid, DNA content between 1N and 2N with clearly identifiable G1 and G2 peaks (Figure 1B–1D and Figure S3). The ploidy distribution of these 41 populations are more uniform compared to the binomial distribution expected from triploid meiosis (Figure 1B), with significantly fewer than expected viable strains with a ploidy ∼1.5 (Kolmogorov-Smirnov test between observed and expected cumulative distribution function P = 1.58×10−2, Figure S4). This result suggests that the viability of aneuploid strains may be biased toward those with ploidy close to a euploid state (haploid or diploid) compared to those with ploidy equidistant to the two nearby euploid states.
We next subjected the above 41 g20 population samples to qPCR karyotyping analysis in order to determine the modal karyotype of each population by combining the ploidy estimate from FACS with the chromosome stoichiometry data from qPCR karyotyping assays. This was successfully accomplished for 36 of the g20 populations where the dominant karyotype could be clearly determined (Figure 1C, showing one such example, and Figure S5, showing all qPCR data for the g20 populations). The remaining 5 g20 populations were simply too heterogeneous in qPCR profiles for us to determine the modal karyotype (Figure S5 “too heterogeneous”). These initial observations already indicate that different aneuploid strains exhibit different levels of CIN.
In order to associate specific CIN level with specific aneuploid karyotypes, we observed karyotype dynamics in each of the aneuploid strain populations by determining the karyotypes of the 11 randomly selected colonies plated from the population culture at one of the three (g20, g25 and g30) time points (see Figure 1A). Because qPCR karyotyping was of significant cost, only 27 of the 36 strains described above were subjected to this analysis while 9 were excluded due to contamination or redundancy in karyotype with other aneuploid strains in the collection (see Figure S1). We first used FACS data from the colonies to help select the time point most appropriate for karyotype analysis. For those strains that appeared to be most stable (g20 and g25 colonies showing ploidy variation similar to a wild-type control), colonies from the g30 populations were chosen to maximize the chance of observing some karyotypic deviants, whereas for those strains displaying the greatest apparent instability by FACS (colonies showing ploidy variation substantially larger than a wild-type control) colonies of populations from the earlier time points (g20 and g25) were chosen for qPCR analysis. Figure S6 displays representative examples of a haploid control (A), relatively stable (B) and unstable (C) aneuploid strain. The karyotyping data from the g20 population samples and g20, g25 or g30-derived colony samples allowed us to examine the relationship among the observed aneuploid karyotypes of all analyzed samples originated from the same spore.
We next used a haplotype mapping approach to determine the minimum number of chromosome gain or loss events sufficient to explain the diverse karyotypes revealed by the 11 analyzed colonies of a given population. Haplotype maps are typically used to display genetic variation based on SNP loci and help to study the genotypic variation between populations of individuals [22]. A parsimony approach is used during the reconstruction of the relationship between the observed genotypes and do not require a priori information regarding the phylogeny of the individuals in the population. In adopting this approach, the 16 yeast chromosomes were treated as independent loci, each of which can exist in, and change between, two states defined by copy numbers (analogous to alleles) – 1 or 2 copies (as the ploidy of these strains varied between 1N and 2N). The algorithm proceeded by attempting to connect all 12 karyotypes (the g20 population karyotype +11 colony karyotypes) in a single haplotype map (in this case, a ‘karyotype map’) by minimizing the total number of mutational events (in this case, copy number changes) in the entire map. It has to be noted that the parsimony approach underlying this algorithm allows distinguishing those colony karyotypes that most likely originated directly from a CIN event in the ancestral spore karyotype from those that arose as secondary events from already deviant karyotypes accumulated in the population. This step was important to not over-estimate the level of CIN that could be attributed to the original spore karyotypes.
Figure 2 and Figure S7 display the resulting karyotype maps in the 27 aneuploid strains. Based on the number of independent CIN events that could be directly linked back to the original karyotypes, we qualitatively divided the 27 aneuploid strains into three classes (Figure 2A): (i) ‘stable’ (S) strains (n = 8), referring to those in which karyotype changes were not observed during our experiments (e.g. strain 245 at g30, Figure 2B); (ii) ‘mildly unstable’ (MU) strains (n = 10), in which only one chromosome copy number change event (involving either a single or multiple chromosomes) was observed at later generations (e.g. strain 226 at g30 in Figure 2C); and (iii) ‘highly unstable’ (HU) strains (n = 9) in which more than one CIN event was observed at early generations (e.g. 225 at g20 in Figure 2D). Interestingly, strains belonging to each of the three CIN categories displayed a wide range of apparent growth rates, estimated by regressing the absolute cell counts measured at the three different time points (g20, g25, g30; Figure 3A). Even though S strains exhibited a slightly higher average growth rate than HU strains, this difference was not statistically significant (P = 0.498, Figure 3A). This observation is consistent with our previous finding that stable aneuploid strains exhibit a wide range of growth abilities [15] and a recent report showing a lack of correlation between cell cycle delay and yeast artificial chromosome loss rate in yeast disomic strains [23]. Thus, it is unlikely that the observed CIN in aneuploid strains is a consequence of aneuploidy-associated fitness impairment under standard laboratory growth conditions.
To ask if the aneuploidy-associated CIN might be a consequence of certain global karyotypic features, we examined the correlation between CIN and parameters such as the total number of chromosomes or base pairs in the genome, or the total number of chromosomes or base pairs in aneuploidy, etc. This analysis led to two observations. First, S strains tended to have a ploidy lower than 1.5, whereas MU or HU strains tend to have a ploidy around 1.5 or higher. Compared to HU strains, S strains showed a significantly lower base pair content (P = 3.15×10−2), a significantly smaller number of total chromosomes in the genome (P = 2.07×10−2) and a significantly lower basal ploidy (P = 8.52×10−3) (Figure 3B, 3C and 3D). Because the analyzed aneuploid strains had ploidy between 1N and 2N, these correlations suggest that haploid genomes with a few extra chromosomes tend to be more stable than diploids missing a few chromosomes. Second, CIN did not correlate with the number of aneuploid chromosomes: S and HU strains were significantly different neither in the total number of chromosomes in aneuploidy (P = 0.231) nor in the number of megabases in aneuploid chromosomes (P = 0.907) (Figure 3E and 3F). For this analysis, we defined basal ploidy as the integer number corresponding to the most frequently-appearing chromosome copy number in an aneuploid genome and aneuploid chromosomes as those with a copy number deviating from the basal ploidy (be it gains or losses). These observations suggest that the genome does not necessarily become more unstable as it departs further from the euploid state, however, there is a genome-level impact on CIN related to the degree of departure of the aneuploid chromosome number from the lower euploid state (in this case, the true haploid state).
We next examined the correlation between CIN and the presence of specific chromosomes in aneuploidy using the karyotyping data of the 27 strains characterized. For this analysis we focused on 21 strains for which aneuploid chromosomes could be assigned unambiguously based on basal ploidy assignment as explained above, but excluded 6 strains that had eight chromosomes with a copy number of 1 and eight chromosomes with a copy number of 2 (thus impossible to assign which chromosomes are in euploidy and which in aneuploidy). As expected, ChrVI aneuploidy was rarely found across the 21 aneuploid strains (only a single strain with ChrVI monosomy and a basal ploidy of 2N), consistent with previous reports of low tolerance of copy number imbalances of this particular chromosome, most probably due to the presence of several major cytoskeletal genes (e.g. ACT1, TUB2) on this chromosome [23], [24], [25]. We calculated the frequency at which each chromosome was present in aneuploidy across the 21 strains and searched for over- or under-representation across the three different classes of CIN (S, MU and HU) (Figure 4). In general, the frequency at which each of the 16 chromosomes was found in aneuploidy was not uniformly distributed across the three different classes of CIN (Fisher test P = 2.75×10−2). In particular, ChrVII aneuploidy was significantly associated with S strains (Fisher test P = 4.81×10−2) and ChrV aneuploidy was significantly associated with HU strains (Fisher test P = 2.03×10−2).
To gain molecular insights into the chromosome features associated with different levels of CIN, we analyzed enrichment of genes potentially linked to CIN on ChrV and ChrVII in comparison to other chromosomes. We used several published datasets obtained from different types of screens for chromosome instability genes, such as “genes causing increased colony sectoring when deleted” (source: Saccharomyces Genome Database, SGD), “genes causing increased colony sectoring when overexpressed” (SGD) or “genes causing decreased chromosome/plasmid maintenance when deleted” (SGD), as well as more comprehensive datasets such as “genes associated with chromosome instability” [26] or genes annotated with gene ontology (GO) term “chromosome segregation”. As shown in Figure S8, ChrV and ChrVII do not show exceptional or consistent over or under-representation of genes in any of these datasets.
A lack of clear insights from the above analysis based on individual aneuploid chromosomes led us to consider the possibility that it is the relative dosage of two or multiple chromosomes rather than any particular aneuploid chromosome per se that affects karyotype stability. We thus performed a systematic analysis of all pair-wise combinations of the 16 chromosomes to determine if any imbalanced pairs (copy number ratio to be either 0.5 or 2, but not 1) were significantly associated with CIN. Using a hypergeometric test between two groups, relatively stable (S+MU) vs. highly unstable (HU), the imbalance between three different pairs of chromosomes were observed to distinguish the two CIN groups: ChrII vs. ChrVIII, ChrIII vs. ChrIX, and ChrVII vs. ChrX (Figure 5A, right panel). The last pair, ChrVII vs. ChrX (Hypergeometric test P<0.02), was surprising as the single chromosome analysis found ChrVII aneuploidy to be associated exclusively with karytoypically stable strains (Figure 4). A closer scrutiny found that in 2 of the 3 cases where ChrVII was in aneuploidy (strains 220 and 230, having 1 extra ChrVII with 1N basal ploidy), ChrX was also in aneuploidy with 1 extra copy, and thus their numbers were balanced. On the other hand, of the 8 aneuploid strains where ChrVII and ChrX had unequal copy numbers, 5 were highly unstable, 1 mildly unstable, and only 2 stable. These findings suggest that the effect of an individual aneuploid chromosome on CIN is dependent on the karyotypic context in which the aneuploid chromosome is present. We note that if the hypergenometric test was performed between S vs (MU+HU) groups, different pairs of chromosomes were observed whose imbalance distinguished the two groups (Figure 5A left panel). This observation suggests that there are potentially many different pairs of chromosomes whose copy number imbalance could lead to CIN.
A recent study found that heterozygosity of MAD2, located on ChrX and encoding a key component of the spindle assembly checkpoint (SAC) [27], leads to partial SAC inactivation and elevated CIN in a diploid background [28]. This effect can be rescued by restoring a 1∶1 stoichiometry for the gene copy number of MAD2 vs. MAD1, located on ChrVII and encoding another SAC component that physically interacts with Mad2 protein [28]. This finding implies a ratio of 0.5 for the copy numbers of MAD2:MAD1 to be sufficient to induce CIN and led us to hypothesize that the association of ChrX and ChrVII imbalance with CIN may be attributed to an imbalanced MAD2:MAD1 ratio of 0.5. To increase the statistical power for testing this hypothesis over the data from the 27 karyotyped strain populations, we isolated another 56 fresh aneuploid strains as the meiotic products of the same triploid strain used before. These aneuploid strains were again collected ∼20 cell divisions after tetrad dissection and an aliquot of each population was used for analysis of the relative gene copy number of MAD1 vs. MAD2 by qPCR on genomic DNA. Each g20 population was also plated to single colonies and after three days of growth at 23°C, 11 colonies were randomly picked from each and analyzed by FACS to determine ploidy variation (Figure 5B). This analysis found 44 of the 55 aneuploid strains to display unstable ploidies with obviously divergent ploidy profiles between the 11 colonies or Coefficient of Variation (CV) of the G1 peak positions much larger than that of the control haploid sample, whereas 11 strains were found to be ploidy-stable where the 11 picked colonies showed identical G1 peak position and CV similar to or smaller than that of the haploid control (Table S3). Genomic qPCR analysis of the g20 population samples using probes against MAD1 and MAD2 gene sequences found that: (i) strains with different MAD2:MAD1 ratios were not uniformly distributed across the stable and unstable strains (P = 0.027, Figure 5C); and (ii) 18 aneuploid strains with a MAD2:MAD1 ratio of 0.5 all fell into the ‘ploidy unstable’ category, whereas all 11 ‘ploidy stable’ aneuploid strains had a MAD2:MAD1 ratio of 1 or higher, indicating a highly significant association of a MAD2:MAD1 ratio of 0.5 with CIN (P = 0.01, Figure 5D). We note that 25 strains with unstable ploidy did not have the MAD2:MAD1 ratio to be 0.5, indicating that a MAD2:MAD1 ratio of 0.5 was sufficient but not required to cause elevated CIN. qPCR analysis using probes against the ChrVII or ChrX arm opposite to the MAD1 or MAD2 locus, respectively, indicated that the MAD2:MAD1 gene copy number imbalance was indeed due to copy number imbalance between these two chromosomes (Table S3). These results demonstrate a specific case where gene dosage imbalance affecting two components of the mitotic system underlies the association between chromosome imbalance and CIN.
The results described above support the notion that aneuploid genomes are in general less stable than euploid genomes and prone to further karyotype changes. These findings in yeast are in agreement with recent observations that chromosomally stable pseudo-diploid human cells that accumulate aneuploid chromosomes frequently become chromosomally unstable [29]. However, our results also indicate that different aneuploid karyotypes can exhibit different degrees of CIN, with some being more stable than others, suggesting that CIN is not a necessary outcome of aneuploidy. In other words, CIN does not appear to be caused by some general property of being aneuploid per se but rather by determinants associated with specific aneuploid karyotypes. An advantage of our study is the carefully controlled genome variability among the strains analyzed. Because all aneuploid strains were derived from the same homozygous triploid parent and underwent minimal passage before their analysis [15], the different strains only differed in chromosome stoichiometry, minimizing the possibility that the observed CIN was due to other genetic variations or aberrations between different strains. Another advantage of our study was that CIN was assessed by examination of a wide range of spontaneously occurring karyotypic changes that include copy number gains and losses of native chromosomes. By analyzing aneuploid strains with randomly generated chromosome stoichiometries and the possibility of multiple chromosomes in aneuploid copy numbers, we were able to investigate the determinants underlying CIN in an unbiased manner and the effect of combinations of chromosomes in aneuploidy. We note, however, that our method presently does not allow analysis of those highly unstable karyotypes that quickly lead to considerable karyotype diversity within even a small population, and thus our results may not shed light on the determinants underlying extreme CIN. In addition, the qPCR-based karyotype method does not faithfully distinguish between whole-chromosome aneuploidy and partial chromosome aneuploidy and does not report on recombination events that may also be elevated in aneuploids [18].
Consistent with a recent report [18], we did not observe any correlation between fitness and CIN among the aneuploid strains. Whereas some aneuploid strains with non-observable CIN grew relative poorly, some strains with high CIN grew relatively well. This finding suggests that CIN is not necessarily a consequence of the growth defect caused by aneuploidy under standard laboratory growth conditions, or driven by the selection for improved fitness, but may be more intrinsic to specific features of an aneuploid genome. Analysis of the correlation between CIN and different parameters associated with specific karyotypes allowed us to observe two potential determinants of CIN. On the more global level, it was surprising to find that CIN was not necessarily linked to the distance of an aneuploid karyotype from the nearest euploid state. Instead, given that the analyzed aneuploid strains had a ploidy between 1N and 2N and that each chromosome exists in a copy number of either 1 or 2, we found CIN to be significantly linked to the distance of the karyotype from the haploid state. In other words, haploids with a few extra chromosomes tend to be more stable than diploids missing a few chromosomes. As the number of aneuploid chromosomes increases from 1N towards 2N, the level of CIN tends to increase until the ploidy reaches 2N, when the level of CIN is reset to a low level (Figure 6). Future work will be necessary to test whether this trend continues beyond 2N. We propose to explain this trend by the disparity between the burden of segregating an increasing number of chromosomes and a lack of linear scaling of the capacity of the mitotic system with the aneuploid genome size. In this model, certain complex machineries, such as the kinetochore, or the mitotic spindle and the associated checkpoint mechanism, are composed of stoichiometric protein components encoded by genes distributed on all 16 chromosomes. This predicts that the functional scaling to increase the capacity of such machinery to segregate an increasing number of chromosomes from, e.g., a true haploid number might occur in a discrete rather than continuous manner and requires gaining of an entire chromosome complement (Figure 6). As such, near-diploid karyotypes are predicted to be highly unstable owing to the largest disparity between the burden of having to segregate many extra chromosomes and the capacity of the mitotic machinery that, despite the near-diploid genome size, is still working with an efficiency close to that in a haploid genome (‘functional deficit’ in Figure 6). Only upon the chromosome number reaches a true diploid state, stoichiometries are reset to their basal level and mitosis can proceed with high fidelity. We note that “scaling” in our model differs from that in a previous study on CIN in polyploid yeast cells [21]. “Scaling” in the Storchova model refers to a lack of scaling in the size of the pre-anaphase spindle with a euploid genome size (1N, 2N, 3N, 4N etc). The model intends to explain why polyploids are less chromosomally stable than haploids or diploids. Scaling in our model, on the hand, refers to the discrete increase in the functionality of the mitotic system with a linearly increasing number of chromosomes and intends to explain why certain aneuploid karyotypes are particularly unstable and why aneuploids are in general more karyotypically unstable than euploids.
Although the global trend discussed above was statistically significant, exceptions to the rule could be found when comparing instability between specific karyotypes. This suggests that karyotype-specific effects may be superimposed on the global trend. Consistent with this idea, our analysis of relative dosage between pairs of chromosomes revealed an association of CIN with dosage imbalance between specific chromosome pairs. Because the level of gene expression largely scales with gene dosage at both the transcriptome and proteome levels [15], [30], [31], chromosome copy number imbalance is likely to directly lead to altered stoichiometry of proteins that interact physically or functionally. It has been shown that an unbalanced stoichiometry in specific proteins affecting mitotic spindle function is sufficient to drive chromosome mis-segregation in cancer cell lines [20]. In yeast, one example is represented by the imbalance of MAD1 and MAD2 mitotic checkpoint genes [28]. Although the precise molecular explanation remains unclear, it was shown that when MAD2 gene dosage was reduced relative to MAD1, such as in the case of heterozygous gene deletion, chromosome instability ensued. Stability could be restored by further deletion of a copy of MAD1 to revert their ratio back to 1. Indeed, our data indicate that a ChrX (carrying MAD2) to ChrVII (carrying MAD1) ratio of 0.5 strongly predicts CIN. That dosage imbalance may be a prominent cause of CIN is also supported by the observation that many SAC components are deregulated at the gene expression level in several cancer cell lines without harboring sequence mutations in the corresponding genes [32]. We note that there are likely to be many gene pairs whose imbalance could lead to CIN. For example, an imbalance between ChrII and VIII is also a predictor of CIN (Figure 5A), and the chromosome passenger proteins Sli15 (INCENP) and Nbl1 are encoded on ChrII and VIII, respectively. Whereas Sli15 and Nbl1 both interact with the Aurora kinase Ipl1, Nbl1 is the yeast borealin-like and bridges the interaction between Bir1 (survivin) and Sli15 [33], [34]. It is conceivable that these chromosome passenger complex components need to be balanced in dosage to ensure proper chromosome segregation.
The flip side of the above finding is that relatively stable karyotypes may result from fortuitous but possibly complex balancing of certain key modules of the mitotic machinery. In an adaptive landscape, such metastable karyotypes may correlate with relatively stable, thus selectable, phenotypic states. This is consistent with the observation in mouse models or cancer cells that whereas moderate levels of CIN promote tumor formation or emergence of drug resistance, extremely high CIN could abate both processes [35], [36]. A recent large-scale analysis of aneuploid karyotypes in cancer cells revealed a high rate of co-occurrence of specific chromosome gains or losses [37]. While this may be explained by a requirement for balanced gene function to maintain fitness, chromosome co-gain or co-loss may also be important for achieving relatively stable cancer karyotypes in order for persistent expression of cancer phenotypes given a certain tissue microenvironment. Further, the existence of relatively stable karyotypic and phenotypic states may explain why certain chromosome aberrations in cancer are clonal [38], [39].
Finally, the observation of both global and chromosome-specific determinants of CIN may help to reconcile the chromosome/genome-centric theory vs. gene-centric theory in cancer evolution. First, our model of discrete and genome-dependent scaling of accurate chromosome segregation is consistent with the notion that complex cellular behaviors are non-linearly related to the sum of the function encoded by individual genes or even chromosomes. At the same time, the observation of different degrees of CIN associated with different aneuploid karyotypes, and more importantly with specific chromosome imbalances, highlights the exceptional impact of certain molecular components, such as Mad1 and Mad2, on the function and stability of the genome. However, even in this latter scenario, the impact of specific gene dosage is context-dependent, i.e. dependent on the dosage or activity of its partners in a manner that is potentially difficult to decode without a better knowledge of the entire cancer genome.
Aneuploid strains were generated as meiotic products of a homozygous triploid yeast strain as previously described [15]. All strains were cultured in either liquid or solid YEPD (Yeast Extract Peptone +2% Dextrose) media at 23°C. A list of all analyzed aneuploid strains with their karyotype information is provided in Table S1.
Aneuploid spores were grown into colonies of ∼106 cells based on preliminary experiments correlating colony size with cell number. Then the spore-derived colonies were entirely picked and resuspended into 2 mL of YEPD media. The actual cell concentration was measured using a hemocytometer and ∼200 cells were plated onto 15 cm YEPD plates. 600 µL of the culture was immediately fixed with 70% Ethanol for DNA content analysis by FACS (see below). Concomitantly a biomass corresponding to an OD600 = 0.3 in 300 µL was immediately frozen at −80°C for qPCR assays (see below). A part of the culture was used to prepare glycerol stocks. The remaining culture was diluted 200× with YEPD media and grown at 23°C. Cell numbers in the growing cultures were regularly monitored using the hemocytometer and ∼200 cells were spread on to YEPD plates once the cultures reached ∼25 and ∼30 cell divisions after germination. After 3–6 days incubation at 23°C, 11 colonies from each YEPD plate were randomly picked as previously described [15]. The picked colonies were inoculated into a 96-well deep-well block containing 1.5 mL YEPD media and grown overnight at 200 rpm. Each culture was harvested for FACS and qPCR analysis as described above.
The DNA content analysis and qPCR-based karyotyping were performed essentially as previously described [15] with the only exception that FACS samples from to MAD2:MAD1 ratio experiment were acquired using a MACSquant Analyzer (Miltenyi) and 6,000 events were collected for each sample. The data were analyzed using FlowJo 7.6.1. Ploidy variation analysis was performed by extracting the mode of the G1 peak position from the DNA profile of original spore and from the corresponding 11 colonies. The CV (Coefficient of Variation) was calculated between these 12 G1 peak positions and compared to the CV of 12 randomly-picked haploid colonies processed in parallel.
A new set of aneuploid strains was generated as described above. When the spore colony reached ∼106 cells, the colony was picked, partially harvested for FACS analysis and for MAD2:MAD1 ratio determination (see below), its cell number was determined at the hemocytometer and ∼200 cells were plated into fresh YEPD plates. Once colonies became visible, 11 randomly selected colonies were picked and processed for FACS analysis as described above. When aligning the DNA profile of the original spore to the DNA profiles of the corresponding 11 colonies, the following criteria were applied: (i) those strains with obvious heterogeneous and noisy DNA profiles were classified as unstable; (ii) strains with clean DNA profiles and similar ploidy were subjected to ploidy variation analysis as described above. A strain was classified as ‘ploidy stable’ only if it showed similar or smaller CV compared to that of a haploid control strain processed in parallel. Samples harvested for MAD2:MAD1 copy number ratio determination were resuspended in 20 µL PBS (pH = 7.4) containing 50 µg/µl Zymolyase 100T (US Biological) and incubated at 37°C for 30 min. These samples were diluted 1∶200. To prepare qPCR reactions, 2 µl of these dilutions were combined with 8 µl 1× Perfecta SYBR Green Mix (Quanta) at 500 nM for forward and reverse primers in technical triplicates on a CAS-4200 robot (Corbett) and run on an ABI 7900HT cycler with the following cycling conditions: 95°C for 5 min, then 40 cycles of 95°C for 15 s followed by 60°C for 1 min. Ct values were obtained using SDS 2.4 software (ABI). The ratio of MAD2:MAD1 copy number presented in Figure 5C and 5D were calculated using the NRQ method in qbasePLUS version 2.0 software (Biogazelle) by setting either MAD1 or ChrVII as the endogenous control and scaling all samples to wild type haploid. All primers used in this study are listed in Table S2.
All statistical analyses and computer simulations were performed in the R environment for statistical computing. Difference in means was evaluated by means of two-sided unpaired Welch's t-test, association between categorical data by Fisher's Exact Test for count data, overlap between subsets by Hypergeometric test and difference between empirical cumulative distribution functions by Kolmogorov-Smirnov test. Results were considered significant if P<0.05.
Simulation of the expected fraction of cells with deviant karyotype as a function of generation time and chromosome mis-segregation rate was performed as follows. A seeding cell was represented as a vector of length 16, each element of which represented the copy number of one of the 16 yeast chromosomes. At each generation, each cell was allowed to self-duplicate and during this process each of the 16 duplicated chromosomes was allowed to mis-segregate with a given probability, resulting in one daughter inheriting both copies and the other daughter not inheriting any copy of that particular chromosome. At the end of each generation, cells that lost all copies of any given chromosome were discarded as dead cells. Every time the simulated colony reached >100,000 cells, a random sampling of 100,000 cells was used to simulate the next generation to limit computational complexity. As a control, the same simulation was performed using different cell number cutoffs without significant differences (data not shown).
Simulation of the expected distribution of apparent ploidy of spores from triploid meiosis was performed as follows. As above, random spores were represented as vectors of length 16, in which each element represented one of the 16 yeast chromosomes and having identical and independent probability of being of copy number ‘1’ or ‘2’. The apparent ploidy of each random spore was calculated based on the known length in base pairs of each of the 16 yeast chromosomes. 10,000 independent simulations were performed, in each of which 41 random spores were generated, i.e. the same number as the experimentally determined ones.
Chromosome copy number data from both the g20 population and from the 11 colonies analyzed at either g20, g25 or g30 were combined into a matrix of size 12×16, in which each row represented one of the 12 ‘individuals’ and each column represented one of the 16 ‘loci’ carrying one of two ‘alleles’, corresponding to the two copy number states (i.e. ‘1’ and ‘2’). This matrix was used as input for the Network software (version 4.5.1.6.), which reconstructed the most likely karyotype network by minimizing the number of allelic changes (here: chromosome copy number gain/loss events) across the entire map. Karyotype changes involving more than one chromosome copy number change were scored conservatively as a single event, as we ignored whether the multiple chromosome mis-segregations occurred in a single erroneous mitosis or multiple subsequent mitotic events. Also, karyotype changes unlikely to have originated directly from the inferred original spore karyotype, but more likely to have originated from one of its karyotypically deviant progeny according to the reconstructed karyotype network, were not counted for the determination of the level of CIN of the ancestral karyotype, as they would be more reflective of the level of CIN of one of its karyotypic deviants as opposed to the level of CIN of the ancestral karyotype.
|
10.1371/journal.ppat.1002267 | Exposure to the Viral By-Product dsRNA or Coxsackievirus B5 Triggers Pancreatic Beta Cell Apoptosis via a Bim / Mcl-1 Imbalance | The rise in type 1 diabetes (T1D) incidence in recent decades is probably related to modifications in environmental factors. Viruses are among the putative environmental triggers of T1D. The mechanisms regulating beta cell responses to viruses, however, remain to be defined. We have presently clarified the signaling pathways leading to beta cell apoptosis following exposure to the viral mimetic double-stranded RNA (dsRNA) and a diabetogenic enterovirus (Coxsackievirus B5). Internal dsRNA induces cell death via the intrinsic mitochondrial pathway. In this process, activation of the dsRNA-dependent protein kinase (PKR) promotes eIF2α phosphorylation and protein synthesis inhibition, leading to downregulation of the antiapoptotic Bcl-2 protein myeloid cell leukemia sequence 1 (Mcl-1). Mcl-1 decrease results in the release of the BH3-only protein Bim, which activates the mitochondrial pathway of apoptosis. Indeed, Bim knockdown prevented both dsRNA- and Coxsackievirus B5-induced beta cell death, and counteracted the proapoptotic effects of Mcl-1 silencing. These observations indicate that the balance between Mcl-1 and Bim is a key factor regulating beta cell survival during diabetogenic viral infections.
| The global prevalence of type 1 diabetes (T1D) is approximately 20 million individuals, and projections indicate that this will double in the coming decades. This increase in T1D incidence is probably related to modification in the exposure to environmental factors. Viruses are one of the putative environmental agents that trigger T1D. The mechanisms leading to beta cell loss during viral infection, however, remain to be clarified. The present study aimed to clarify the mechanisms of dsRNA- and virus-induced beta cell demise. Thus, dsRNA produced during viral replication in beta cells induces activation of the kinase PKR. This kinase phosphorylates the elongation factor eIF2α promoting inhibition of protein synthesis which decreases the short-life antiapoptotic protein Mcl-1. Decline of Mcl-1 allows Bim to exert its proapoptotic effects by activation of the intrinsic pathway of apoptosis. In conclusion, we demonstrate that the balance Mcl-1/Bim is central for dsRNA- and enterovirus-induced beta cell apoptosis. This clarifies one of the key questions in early T1D pathogenesis, namely the mechanisms behind viral-induced beta cell apoptosis, and the subsequent triggering of insulitis.
| Type 1 diabetes (T1D) is a chronic autoimmune disease characterized by the progressive and selective destruction of the insulin-producing pancreatic beta cells [1]. It mainly affects individuals during childhood or adolescence and requires a life-long treatment with insulin, which at the US represents a cost of $14.4 billion per year [2]. Triggering of diabetes results from an interaction between genetical and environmental factors. Most candidate genes for T1D are supposed to act at the immune system level [3], but we have recently shown that nearly 30% of these candidate genes are also expressed in beta cells and may modulate their responses after exposure to potential environmental factors [4]. These findings indicate that beta cells are not only targets, but also actors of T1D pathophysiology [4].
The incidence of T1D is increasing 3.9% per year in Europe, especially among the youngest population (<5 years-old) in which a doubling in the number of new cases is expected between 2005 and 2020 [5]. An explanation for this rapid augment in T1D incidence may be the increased exposure to diabetogenic environmental factors. Viruses are among the potential environmental causes of T1D [6], as suggested by epidemiological [7], experimental [8] and clinical data [9].
Enterovisuses (EV) such as Coxsackievirus B (CVB) [10], are the main candidates. Coxsackievirus B4 was isolated from a child who died of diabetic ketoacidosis and this virus induced diabetes in mice [11]. Among the Coxsackievirus, CVB5 is the most deleterious prototype strain for human pancreatic islets [12]. An increase of T1D incidence has been described after epidemics of CVB5 [13] and these epidemics are frequent in Finland, the country with the highest incidence of T1D [14].
The pathogenic role of viruses in T1D might involve damage to beta cells and the local induction of proinflammatory mediators [1]. CVB-infected pancreatic beta cells can be phagocyted by both macrophages [15] and dendritic cells [16], leading to activation of the immune system, presentation of islet autoantigens and release of cytokines/chemokines. Local injury of beta cells induced by dsRNA, a by-product of viral replication, promotes autoimmune diabetes in animal models [17]. Both the viral mimetic dsRNA [18] and Coxsackievirus [19] induce beta cell apoptosis, but the mechanisms involved remain to be clarified.
Apoptotic cell death can be initiated by two signaling cascades, namely the intrinsic and the extrinsic pathways. The intrinsic pathway is the main pathway for endoplasmic reticulum (ER) stress and cytokine-induced beta cell apoptosis [20], [21], [22], [23]. In this pathway, a disequilibrium between the antiapoptotic and the proapoptotic Bcl-2 family members [24] leads to activation of Bcl-2-associated X protein (BAX) and Bcl-2 antagonist or killer (BAK) and the consequent permeabilization of mitochondrial outer membrane. This allows cytochrome c release to the cytoplasm and the assembly of the apoptosome. The apoptosome then recruits and activates the initiator caspase 9, which cleaves and activates the effector caspase 3, leading to execution of apoptosis [25].
We have presently investigated the molecular pathways involved in dsRNA and viral-induced beta cell apoptosis. Similar to ER stress and cytokine-induced apoptosis [24], the viral by-product dsRNA activates the intrinsic pathway of apoptosis. The nature of the Bcl-2 family members involved is, however, different. Thus, dsRNA-dependent protein kinase (PKR) activation by cytoplasmic dsRNA leads to phosphorylation of the elongation factor eIF2α. The eIF2α phosphorylation promotes inhibition of protein translation, which is followed by an early and progressive decrease in the expression of the antiapoptotic myeloid cell leukemia sequence 1 (Mcl-1) protein. Decreased expression of Mcl-1 allows the proapoptotic protein Bim to activate BAX, leading to cytochrome c release, caspase 9 and 3 activation and beta cell apoptosis. Similar findings were observed during viral infection of pancreatic beta cells by CVB5. These observations clarify the mechanisms involved in viral-induced apoptosis in pancreatic beta cells, and suggest that usage of Bcl-2 proteins is context-dependent during beta cell apoptosis initiated by different stimuli.
Transfection with the synthetic dsRNA polyinosinic-polycytidylic acid (PIC) induced beta cell apoptosis already after 12h, with progressive increase up to 24h (Figure 1A). To characterize the pathways involved in beta cell apoptosis, BAX translocation to the mitochondria was analyzed by immunocytochemistry. Intracellular dsRNA promoted the translocation of BAX (Figure 1B) and the release of cytochrome c from the mitochondria to the cytoplasm (Figures 1C and D). This activated the initiator caspase 9 and effector caspase 3 (Figure 1E), characterizing induction of the intrinsic mitochondrial pathway of apoptosis.
The antiapoptotic Bcl-2 protein Mcl-1 is an important regulator of beta cell death after exposure to proinflammatory cytokines and ER stressors [20]. Mcl-1 protein expression gradually decreased from 6h until 24h after PIC transfection (Figure 2A). This reduction in Mcl-1 protein was a post-transcriptional effect since Mcl-1 mRNA expression was in fact increased at 24h (Figure 2B). Mcl-1 is targeted for degradation at the proteasome [26]. INS-1E cells treated with the proteasome inhibitor MG-132 (same experimental conditions as in [20], [22]) had increased basal Mcl-1 expression, and the inhibitor partially prevented dsRNA-induced decrease in Mcl-1 expression (Figure 2C), suggesting that proteasomal degradation contributes at least in part for the observed Mcl-1 downregulation. Mcl-1 knockdown by two previously validated siRNAs [20] exacerbated cleavage of caspases 9/3 (Figure 2D) and PIC-induced apoptosis in INS-1E cells (Figure 2E) and FACS-purified primary beta cells (Figure 2F). On the other hand, overexpressing rat Mcl-1 by the use of an adenoviral vector (adMcl-1 [20]) prevented the activation of caspases 9/3 and reduced by > 40% apoptosis induced by dsRNA (Figures 2G, H and I). This is a specific effect, since infection with a control adenovirus encoding luciferase (adLuc) did not modify apoptosis (Figures 2H and I) and Mcl-1 silencing or overexpression did not modify expression of Bcl-XL or Bcl-2 (data not shown). These observations indicate a key role for Mcl-1 decrease in dsRNA-induced apoptosis.
The two other key antiapoptotic Bcl-2 proteins in beta cells are Bcl-XL and Bcl-2 [24]. Bcl-XL but not Bcl-2 protein expression was decreased after PIC exposure (Supplemental Figures S1A and B). This downregulation, however, was only detected at later time-points (after 12h and 24h) and to a lesser extent than observed with Mcl-1 (compare Figure 2A and Supplemental Figure S1A). Bcl-XL and Bcl-2 knockdown (Supplemental Figures S1C and D) increased basal beta cell apoptosis from 7% to more than 30% and further increased PIC-induced apoptosis, as compared to siControl (Supplemental Figures S1E and F). As observed with siMcl-1, inhibition of Bcl-XL and Bcl-2 promoted caspases 9 and 3 cleavage in parallel with beta cell apoptosis (Supplemental Figures S1G and H). Similar results were observed with second independent siRNAs against Bcl-XL and Bcl-2 (Supplemental Figure S2). Nevertheless, when the effects of Bcl-XL or Bcl-2 silencing on PIC-induced apoptosis were corrected by their basal levels of apoptosis using the apoptotic index (i.e. siBcl-XL or Bcl-2 + PIC correct by siBcl-XL or Bcl-2 alone [27]) they were not different from the control group (siControl + PIC) (Supplemental Figure S3). This finding suggests that knocking down these proteins has only additive effect to PIC-induced apoptosis. On the other hand, Mcl-1 knockdown using two different siRNAs resulted in increased apoptotic indexes as compared to siControl, indicating that silencing of Mcl-1, different from Bcl-XL and Bcl-2, potentiates apoptosis caused by PIC (Supplemental Figure S3A).
The BH3-only protein Bim was previously shown to be an important regulator of glucose-mediated beta cell apoptosis [28] and virus-induced cell death in other cell types [29]. Against this background, we evaluated its role in dsRNA-induced beta cell apoptosis. Time-course experiments with INS-1E cells exposed to intracellular PIC did not show modifications in the expression of the Bim isoforms EL, L and S (Figure 3A). Since Bim may modulate apoptosis independently of changes in its expression [30], we next evaluated whether Bim knockdown affects dsRNA-induced apoptosis. The use of a specific siRNA against Bim inhibited by > 70% its protein expression at both basal and PIC-induced conditions (Figure 3B). Bim knockdown prevented PIC-induced apoptosis in INS-1E cells (Figure 3C) and in primary beta cells (Figure 3H). These results were confirmed with a second siRNA against Bim (Supplemental Figure S4). In agreement with the effects on cell viability, PIC-induced activation of caspases 9 and 3 was also reduced after inhibition of Bim (Figure 3D). We studied two other proapoptotic BH3-only proteins, namely Death Protein-5 (DP5) and p53 Up-regulated Modulator of Apoptosis (PUMA), previously shown to regulate cytokine-induced beta cell apoptosis [21], [31]. We obtained a knockdown of >80% for DP5 and >75% for PUMA mRNA expression (Supplemental Figures S5A and C); however, differently from what was observed following Bim silencing, neither siDP5 nor siPUMA prevented PIC-triggered apoptosis (Supplemental Figures S5B and D).
One of the mechanisms by which the antiapoptotic Bcl-2 proteins act is via binding and inactivating specific proapoptotic BH3-only proteins [32]. We thus performed co-immunoprecipitation studies to evaluate the association between Bim and Mcl-1. Bim was immunoprecipitated using a specific polyclonal antibody and the presence of Mcl-1 in the precipitated material was determined with an anti-Mcl-1 antibody. We observed that at basal condition Bim was bound to Mcl-1, but after 24h of PIC exposure there was a clear decrease in the association between these two proteins, confirming that Bim is liberated from Mcl-1 following dsRNA exposure (Figure 3E).
To evaluate whether the PIC-induced decrease in Mcl-1 protein expression contributes to beta cell apoptosis by hampering Bim inactivation by Mcl-1, we performed double-knockdown (siMcl-1 + siBim) and evaluated apoptosis and cleavage of caspases 9 and 3 in comparison to single knockdown (siMcl-1 or siBim). The combined use of siMcl-1 plus siBim produced a >70% suppression of the target genes, as observed with the use of these siRNAs individually (data not shown). Interestingly, the double-silencing of siMcl-1 + siBim prevented the proapoptotic effect caused by Mcl-1 knockdown in INS-1E cells (Figures 3F and G) and primary beta cells (Figure 3H), suggesting that the release of Bim from the Mcl-1/Bim complex induced by both dsRNA and siMcl-1 is a key effector of beta cell apoptosis.
To assess whether the above described pathways are also relevant in the context of a viral infection of beta cells, we infected INS-1E cells and FACS-purified rat beta cells with Coxsackievirus B5. First, we confirmed that this enterovirus infects rat beta cells at the different MOIs tested by determining the presence of the viral capsid (VP1 and 2) (Figures 4A, E and F). The viral capsid was already detected in the cells at 8h post-infection and its expression increased further up to 24h (data not shown; Figures 4A and F). Infection of beta cells at these MOIs induced cell death mainly by apoptosis (Figure 4B) via the intrinsic pathway (Figures 4C and F), as observed with internal dsRNA (Figure 1). In line with these observations, Bim silencing significantly decreased virus-induces apoptosis (Figures 4D and G) and caspase 9 and 3 activation (Figure 4E). Importantly, inhibition of virus-induced apoptosis by knocking down Bim did not modify the amount of viral capsids (VP1 and 2; Figure 4E), indicating that, under the present experimental conditions, inhibition of apoptosis did not exacerbate viral replication in beta cells. These observations suggest that both dsRNA and the potentially diabetogenic virus CVB5 induce a similar pathway of apoptosis in beta cells. Interestingly, both dsRNA and CVB5 induced expression of interferon β (Supplemental Figure S6A and B) and interferon α (data not shown), but with different kinetics. Expression of interferon β mRNA was several-fold higher in primary beta cells as compared to INS-1E cells (Supplemental Figure S6C). This increased expression of type I interferons may also contribute to the observed in vitro beta cell apoptosis, as suggested by previous studies [33].
Two important regulators of Mcl-1 protein expression have been reported in pancreatic beta cells, namely phosphorylation of eIF2α and c-Jun N-terminal Kinase (JNK) activation [20]. Taking this into account we evaluated the time course phosphorylation of eIF2α and JNK in INS-1E cells transfected with PIC. Both eIF2α and JNK were activated by PIC, but with different profiles. Thus, eIF2α phosphorylation started earlier (2h), had a peak at 6h then slowly decreased up to 24h (Figure 5A); JNK phosphorylation started later (6h), reached a plateau at 12h and remained upregulated up to 24h (Figure 5B).
JNK inhibition by a specific chemical inhibitor [34], did not prevent Mcl-1 protein decrease (Supplemental Figure S7). We also evaluated the expression of the BH3-only protein Noxa previously shown to induce proteosomal degradation of Mcl-1 [35], but did not detect its expression in beta cells (data not shown).
PKR-like endoplasmic reticulum kinase (PERK) and PKR were evaluated as two possible mediators of the eIF2α phosphorylation induced by PIC [36], [37]. PERK is a kinase activated by endoplasmic reticulum (ER) stress [36], while PKR is a classic downstream response to dsRNA/viruses [37]. Thapsigargin, a well-know ER stressor [34], was used as positive control for PERK activation. Differently from thapsigargin, PIC did not induce PERK phosphorylation, while it clearly induced eIF2α phosphorylation (Supplemental Figures S8A and B), indicating that another upstream kinase is responsible for this effect. We thus evaluated the role of PKR in this process by using a specific siRNA against PKR which inhibited PKR protein expression by >80% (Figure 6A). Knockdown of PKR almost completely prevented PIC-induced eIF2α phosphorylation and the dsRNA-induced Mcl-1 decrease (Figures 6B and C); there was even an increase in Mcl-1 expression under this condition. In line with the observed increase in Mcl-1 protein expression, silencing of PKR reduced caspases 9 and 3 cleavage (Figure 6D) and PIC-induced apoptosis (Figure 6E). Similar results were obtained with a second siRNA against PKR (data not shown), confirming the key role for PKR in dsRNA-induced Mcl-1 decrease and consequent beta cell apoptosis.
We presently show that intracellular dsRNA activates the intrinsic pathway of apoptosis in pancreatic beta cells. Triggering of apoptosis by dsRNA is secondary to an early and sustained downregulation of Mcl-1 protein expression (Figure 7), resulting in the release of the proapoptotic protein Bim. The unbound Bim then activates BAX translocation to the mitochondria, mitochondrial permeabilization, cytochrome c release, caspases 9 and 3 activation and beta cell apoptosis. Key results obtained with dsRNA, an intermediary product of viral replication, where partially confirmed in the context of a beta cell infection by the diabetogenic enterovirus CVB5.
Apoptosis is the main form of cell death in T1D [38]. Several virus-induced human diseases are associated with increased apoptosis in the target cells, including HIV1-associated dementia [39], cytomegalovirus encephalitis [40] and viral myocarditis [41]. Coxsakieviruses can induce beta cell death by different mechanisms, depending on the strain and multiplicity of infection (MOI) used [19]. At higher MOIs (>100) necrosis is the preferential mechanism of cell death, but at lower MOIs a shift towards apoptosis is observed [19]. Viral triggering of apoptosis depends on both the host and the virus [42]. The host regulates viral-induced apoptosis trough the local production of pro-inflammatory cytokines and chemokines [42]. Viral factors leading to apoptosis include dsRNA [43]. This viral nucleic acid is recognized by receptors such as the toll-like receptor 3 (TLR3), the kinase PKR, the helicases melanoma differentiation-associated gene 5 (MDA5), (which is also a candidate gene for T1D [44]) and retinoic-acid-inducible protein 1 (RIG-I); these receptors activate genes involved in both antiviral responses and apoptosis [45]. All these receptors are expressed in pancreatic beta cells [4], [18]. The pathways downstream of viral recognition/signaling leading to beta cell apoptosis, however, remained to be clarified. The present observations provide a coherent hypothesis for the mechanisms triggering beta cell apoptosis following a viral infection (Figure 7).
Mcl-1 is a pro-survival Bcl-2 family member with a short half-life (30–180 min), which makes it specially susceptible to changes in protein translation [35]. Mcl-1 is degraded in the proteasome [26], and use of a proteasome inhibitor prevented, at least in part, dsRNA-induced decrease in Mcl-1 expression (present findings). We also observed that intracellular dsRNA induces an early and sustained phosphorylation of eIF2α which, as previously shown by our group, leads to a decrease in total protein translation [18]. This eIF2α phosphorylation is mediated by PKR, a protein kinase mainly activated by dsRNAs produced during viral replication [37], but not by PERK (present data). Protein translation inhibition contributes to an early and progressive decrease in Mcl-1 protein levels, which is reverted by knocking down PKR. PKR silencing actually increases Mcl-1 protein expression, probably due to the “release” of protein translation combined with an increase in Mcl-1 mRNA induced by dsRNA (Figure 2B). PKR silencing also prevents dsRNA-induced apoptosis, at least partially due to this increase in Mcl-1 protein. Mcl-1 protein stability is partly regulated by multiple sites of phosphorylation, and JNK-mediated phosphorylation of Mcl-1 increases the rate of protein degradation [35]. We recently described that JNK activation contributes to Mcl-1 degradation in beta cells exposed to the cytokines interleukin-1β + interferon-γ [20]. Here, we observed that dsRNA induces JNK phosphorylation in beta cells, but JNK inhibition does not prevent Mcl-1 degradation. These findings indicate that dsRNA mainly regulates Mcl-1 protein expression via inhibition of protein translation.
Mcl-1 functions as a prosurvivor factor by neutralizing specific propapoptotic BH3-only proteins; Bim has been described as a preferential target of Mcl-1 in other cell types [32]. Bim is a BH3-only protein that presents features of both a “sensitizer”, i.e. it binds to antiapoptotic Bcl-2 protein and displaces the effectors BAX and BAK [32], and also as an “activator”; i.e. through directly binding to BAX and BAK it promotes their activation and the induction of the mitochondrial apoptosis pathway [46]. An imbalance between Mcl-1 and Bim has been shown to trigger apoptosis in other cell types in the context of cytokine deprivation [47] and granzyme B [48] activation. We presently observed that Bim is neutralized by Mcl-1 under basal condition. After dsRNA exposure or the use of specific siRNAs against Mcl-1, however, there is an increase in “free” Bim that subsequently activates BAX and triggers apoptosis (Figure 7). This is suggested by the observation that a double knockdown of Bim plus Mcl-1 (Figure 3) reverts the proapoptotic effects of Mcl-1 silencing. Bim can also mediate the pro-apoptotic effects of type I interferons produced during viral infections [49] and it is targeted by viruses to evade apoptosis in the host cells [50]. Type I interferons were induced during both dsRNA and CVB5 exposure (Supplemental Figure S6). Importantly, inhibition of Bim by specific siRNAs prevented caspases 9/3 activation and apoptosis in beta cells infected with the diabetogenic virus CVB5, confirming the key role of Bim during viral infection of beta cells.
Apoptosis is one of the mechanisms by which the host eliminates virus-infected cells and blocks viral spread [51]. In postmitotic and poorly proliferating cells such as neurons and beta cells, apoptosis might also promote functional loss and disease. Indeed, studies in viral-induced neuronal disease demonstrate that inhibition of apoptosis may reduce disease severity without changes in virus titers [52], [53]. In line with these findings, apoptosis prevention in beta cells by knocking down Bim did not modify viral replication in vitro.
In conclusion, we presently show that decreased expression of the anti-apoptotic protein Mcl-1, coupled to activation of the pro-apoptotic protein Bim, contributes to beta cell apoptosis during in vitro exposure to the viral mimetic dsRNA. In case these observations can be confirmed in future in vivo experiments, novel strategies to increase Mcl-1 or decrease Bim expression in pancreatic beta cells [24] may turn to be interesting approaches to protect beta cells during infection by putative “diabetogenic” viruses.
This study was carried out in strict accordance with the recommendations in the Belgian Regulations for Animal Care guidelines. The protocol was approved by the Commission d'Ethique du Bien-Être Animal (CEBEA) on the Ethics of Animal Experiments of the Université Libre de Bruxelles (Permit Number: LA 1230351). All procedures was performed under Isoflurane anesthesia, and all efforts were made to minimize suffering.
The rat pancreatic beta cell line INS-1E (passages 57–75; a kind gift from Dr. C. Wollheim, Centre Medical Universitaire, Geneva, Switzerland) was cultured in medium containing RPMI 1640 GlutaMAX-I and 5% heat-inactivated fetal bovine serum (FBS) [54].
Male Wistar rats (Charles River Laboratories, Brussels, Belgium) were housed and used according to the Belgian Regulations for Animal Care guidelines. Rat islets were isolated by collagenase digestion of the pancreases. In order to obtain purified beta cells, the islets were dispersed and submitted to autofluorescence-activated cell sorting (FACS Aria, BD Bioscience, San Jose, USA) as previously described [33], [55]. The preparations used in the present study contained 93±2% of beta cells (n = 10). FACS-purified beta cells were cultured in Ham's F-10 medium containing 10 mM glucose, 5% FBS, 0.5% charcoal-absorbed bovine serum albumin (BSA Fraction V, Boehringer, Indianapolis, USA) [33]. During dsRNA and siRNAs transfection cells were cultured in medium without antibiotics and BSA.
The sequences of the siRNAs used are provide in Supplemental Table S1. As control we used Allstars Negative Control siRNA (Qiagen, Venlo, Netherlands). We have previously shown that this negative control siRNA does not affect beta cell gene expression or insulin release as compared to non-transfected cells [20], [56]. The optimal conditions for siRNAs transfection were determined by using a FITC-coupled siRNA (Thermo Scientific) and functional/viability tests [57]. Cells were transfected overnight with 30 nM of siRNAs using Lipofectamine 2000 or RNAiMAX (Invitrogen, CA, USA) according to the manufacturer's instructions [56]. After a recovery period of 24 to 48 h, cells were treated with a synthetic dsRNA or infected by an enterovirus.
The synthetic dsRNA polyinosinic-polycytidylic acid (PIC; Sigma, St. Louis, USA) was used at the concentration of 1 µg/ml [18]. All experiments were made with intracellular PIC, achieved via cell transfection under the same conditions as for siRNAs. Cells exposed to the transfectant alone were used as Control. Since cytoplasmic helicases can recognize dsRNA molecules based on their length [58], a PIC formulation with broad lengths of dsRNAs (10–10,000 bp) was used for the experiments.
The JNK inhibitor SP600125, the ER stressor thapsigargin (Sigma) and the proteasome inhibitor MG-132 were used at concentration of 10 µM, 1 µM and 1 μM respectively [20], [34]. SP600125 was added to the cell medium 4h before PIC transfection and maintained during the whole period of PIC exposure.
The prototype strain of enterovirus (CVB5/Faulkner) was obtained from American Type Culture Collection (Manassas, VA). This virus was passaged in Green Monkey Kidney cells. The identity of the enterovirus preparations used was confirmed using a plaque neutralization assay with type-specific antisera [12].
Cells were infected with different multiplicity of infection (MOIs) of the virus preparation diluted in Hanks' Balanced Salt Solution (HBSS, Invitrogen). After adsorption for 1h at 37°C, the inoculum virus was removed, and the cells were washed twice with HBSS. Culture medium was then added to the plates and the virus was allowed to replicate for the indicated periods.
INS-1E cells and primary FACS-purified rat beta cells were infected for 2 h at different MOIs with an adenovirus encoding Luciferase (Ad-Luc), or with an adenovirus encoding rat Mcl-1 (Vector Biolabs, Philadelphia, PA, USA) [20]. After 48 h of recovery cells were exposed to intracellular PIC for 24h.
Cells were incubated for 15 minutes with the DNA-binding dyes Propidium Iodide (PI, 5 µg/ml, Sigma) and Hoechst 33342 (HO, 5 µg/ml, Sigma). Subsequently, two independent observers determined the percentage of viable, apoptotic and necrotic cells. One of the observers was unaware of sample identity and the agreement between the results obtained was >90%. At least 500 cells were counted in each experimental condition. Results are presented as percentage of apoptosis, calculated as number of apoptotic cells/total number of cells x 100. This method is quantitative and has been validated for use in primary beta cells and INS-1E cells by systematic comparison with electron microscopy, caspase 3 activation and DNA laddering [33], [34], [57], [59], [60]. The apoptotic index was calculated against the percentage of apoptotic cells in the untreated condition as previously described [27]. In selected experiments, apoptosis was confirmed by analysis of activated (cleaved) caspase 3 and 9, cytoplasmic cytochrome c release and BAX translocation to the mitochondria (see below).
Cells were harvested in cold PBS, centrifuged (500 g for 2 min), resuspended with 50 µl lysis buffer (75 mM NaCl, 1 mM NaH2PO4, 8 mM Na2PO4, 250 mM sucrose, 21 µg/µl aprotinin, 1 mM PMSF and 0.8 µg/µl digitonin) and vortexed for 30 s. Following centrifugation at 20,000 g for 1 min the supernatant was retrieved as the cytoplasmic fraction. The remaining pellet was resuspended in 50 µl lysis buffer containing a higher digitonin concentration (8 µg/µl), centrifuged at 20,000 g for 1 min and the supernatant retrieved as the mitochondrial fraction. Equal amounts of proteins were then resolved by 14% SDS-PAGE gel [22]. In addition to anti-cytochrome c (BD Biosciences), the immunoblots were probed with antibodies against β-actin and apoptosis-inducing factor (AIF) (Cell Signaling, Danvers, USA) corresponding to a cytoplasmic and a mitochondrial control, respectively.
mRNA was obtained using the Dynabeads mRNA DIRECT kit (Invitrogen Dynal, Oslo, Norway), and reverse transcribed to obtain the complementary DNA [33]. This material was analyzed by real time PCR reaction using SYBR Green fluorescence on a LightCycler instrument (Roche, Manheim, Germany) and correlated to a standard curve. Expression of the gene of interest was then corrected for the housekeeping gene glyceraldehyde-3-phosphate dehydrogenase (GAPDH). The values are normalized by the highest value of each experiment considered as 1. GAPDH was chosen as housekeeping gene because its expression is not modified by PIC treatment in insulin producing cells [33], [61]. Primers sequences are described in Supplemental Table S2.
Cells were washed with cold PBS and lysed with either Laemmli buffer or Phospho lysis buffer [4]. Equal amounts of proteins were then resolved by 10–14% SDS-PAGE gel and transferred to a nitrocellulose membrane. Immunoblot analysis was performed with antibodies targeting Mcl-1 (Biovision, CA, USA), PKR, total eIF2α (Santa Cruz Biotechnology, CA, USA), phospho-JNK, cleaved caspases 3 and 9, phospho-eIF2α, phospho-PERK, total PERK, Bcl-XL, Bcl-2, Bim (Cell Signaling), enterovirus-specific polyclonal rabbit antiserum (1/600; KTL-510) [12] or α-tubulin (Sigma), used as housekeeping protein. Horseradish peroxidase-conjugated donkey anti-rabbit or anti-mouse IgG were used as secondary antibodies (Lucron Bioproducts, De Pinte; Belgium). Immunoreactive bands were revealed using a chemiluminescent substrate (Thermo Scientific) and detected by a LAS-3000 CCD camera (Fujifilm, Tokyo, Japan). The densitometry of the bands was evaluated using the Aida Analysis software (Raytest, Straubenhardt, Germany).
For the immunoprecipitation, INS-1E cells were lysed on ice using immunoprecipitation buffer [31]. Total proteins were quantified and used as the starting material for immunoprecipitations. Equal amounts of proteins were incubated with rabbit anti-BIM antibody overnight at 4°C with gentle rocking. Antibody-protein complexes were retrieved with a 50% protein A-agarose slurry (Santa Cruz Biotechnology), washed with immunoprecipitation buffer, resuspended in SDS sample buffer and then boiled to separate antibody-protein complex from the protein A. Samples were subjected to 14% SDS-PAGE gel and immunoblotted with anti-α-tubulin (Sigma), anti-Mcl-1 (Biovision) or anti-BIM antibody (Cell Signaling).
Beta cells were plated on polylysine-coated glass culture slides (BD Biosciences). Cells were fixed for 15 min in 4% paraformaldehyde, washed with PBS and permeabilized for 5 min in Triton X-100 0.1%. Slides were then blocked using goat serum 5% and incubated overnight at 4°C in the presence of rabbit anti-Bax (1/1000; Santa Cruz Biotechnology) plus mouse anti-ATP Synthase β (mitochondrial marker) (1/2000; BD Biosciences) or an enterovirus-specific rabbit antiserum (1/600; KTL-510). Cells were washed next morning with PBS and incubated for 1 h with the appropriate Alexa fluor 488 or 555-conjugated antibodies (1/1000; Invitrogen). After, cells were stained with Hoechst, mounted and photographed using fluorescence microscopy (Axio Imager, Carl Zeiss, Zaventem, Belgium) [20].
Data are presented as mean ± SEM. Comparisons were performed by two-tailed paired Student's t-test or by ANOVA followed by Student's t test with Bonferroni correction, as adequate. A P value < 0.05 was considered as statistically significant.
Numbers were taken from GenBank at Pubmed: myeloid cell leukemia sequence 1 (Mcl-1) - 60430; BCL2-like 11 (Bim) - 64547; B-cell lymphoma 2 (Bcl-2) - 24224; Bcl-2 extra-large (Bcl-XL) - 24888; dsRNA-dependent protein kinase (PKR) - 54287; p53 Up-regulated Modulator of Apoptosis (PUMA) - 317673; Death Protein-5 (DP5) - 117271; c-jun NH2-terminal kinase (JNK) - 116554; eukaryotic translation initiation factor 2A (eIF2α) - 502531; PRK-like endoplasmic reticulum kinase (PERK) - 29702.
|
10.1371/journal.pcbi.1003572 | Contribution of Network Connectivity in Determining the Relationship between Gene Expression and Metabolite Concentration Changes | One of the primary mechanisms through which a cell exerts control over its metabolic state is by modulating expression levels of its enzyme-coding genes. However, the changes at the level of enzyme expression allow only indirect control over metabolite levels, for two main reasons. First, at the level of individual reactions, metabolite levels are non-linearly dependent on enzyme abundances as per the reaction kinetics mechanisms. Secondly, specific metabolite pools are tightly interlinked with the rest of the metabolic network through their production and consumption reactions. While the role of reaction kinetics in metabolite concentration control is well studied at the level of individual reactions, the contribution of network connectivity has remained relatively unclear. Here we report a modeling framework that integrates both reaction kinetics and network connectivity constraints for describing the interplay between metabolite concentrations and mRNA levels. We used this framework to investigate correlations between the gene expression and the metabolite concentration changes in Saccharomyces cerevisiae during its metabolic cycle, as well as in response to three fundamentally different biological perturbations, namely gene knockout, nutrient shock and nutrient change. While the kinetic constraints applied at the level of individual reactions were found to be poor descriptors of the mRNA-metabolite relationship, their use in the context of the network enabled us to correlate changes in the expression of enzyme-coding genes to the alterations in metabolite levels. Our results highlight the key contribution of metabolic network connectivity in mediating cellular control over metabolite levels, and have implications towards bridging the gap between genotype and metabolic phenotype.
| Regulation of metabolic activity in response to environmental and genetic perturbations is fundamental to the growth and maintenance of all cells. A primary regulatory process used by cells to control the activity of their metabolic network is the alteration in the expression of enzyme-coding genes. How these alterations regulate metabolite concentrations is an important question in the quest towards unraveling the genotype-phenotype relationship. The link between the expression levels of enzymes and metabolite concentrations is governed by the kinetics of individual reactions, which in turn are interlinked with each other due to the complex connectivity structure of metabolic networks. Although the enzyme-metabolite relationship is relatively well studied at the level of individual reactions, our understanding of the regulation of metabolite levels in complex networks has remained incomplete. In this study, we show that the constraints imposed by the network connectivity are key determinants of the relationship between gene expression and metabolite concentration changes. Our results provide mechanistic insight into the function of complex metabolic networks and have implications for health and biotechnological applications.
| Cellular metabolic networks provide basic biochemical building blocks and a thermodynamically favorable environment for growth and maintenance. Due to this crucial role of metabolism, cells have evolved various mechanisms to regulate metabolic reactions in response to genetic and environmental changes. Metabolic reactions can be regulated either by modulating the availability of the corresponding enzymes, e.g. through altered transcription and/or translation, or, by modulating the enzyme activities through post-translational modifications or through binding of small molecules. Our knowledge of the landscape of transcriptional, translational and post-translational regulation of metabolism is expanding with the increasing availability of datasets that provide genome-wide views of the abundance and interactions between mRNAs, proteins and metabolites [1]–[6]. Although the relative contribution of each of these regulatory layers is still unclear and is likely to be context dependent, it has long been clear that the adjustments in the cellular metabolic phenotype (i.e., rates of reactions, or fluxes, and metabolite levels) often involve changes at the level of gene expression [7]–[9]. For example, previous studies have shown that the gene expression changes in metabolic networks are centered on metabolites that are crucial for adjusting the network state in response to specific perturbations [10], [11]. Despite successful outcomes of these and other studies suggesting a strong link between transcriptional regulation and changes in metabolite levels [7], [10], [12], [13], the relationship between the two has remained elusive.
The task of developing models for describing the relationship between gene expression and metabolite concentrations is challenging due to the multiple layers of regulation involved in between (Figure 1A). Several of the regulatory mechanisms involved, such as translational control or allosteric regulation, are currently poorly understood at the scale of the whole network. Measurement of protein abundances or enzyme activities is also currently difficult to perform at the network scale and in complex systems such as human tissues. Thus, in the absence of data for intermediate molecular players, a detailed investigation of the link between gene expression and metabolite levels has both a fundamental and a practical appeal. In particular, it is of interest to estimate the degree to which the changes at the level of gene expression affect changes in metabolite concentration and to uncover the underlying mechanisms determining their relationship. In this study, we explore the hitherto poorly understood role of network connectivity constraints in controlling metabolite concentrations in a eukaryotic model organism, Saccharomyces cerevisiae. We postulate that two primary mechanisms will largely determine the association between the changes in mRNA and metabolite levels: reaction kinetics (which are non-linear by nature [14]–[16]) and the mass balance constraints imposed by the network, i.e. the balance between production and consumption of metabolites. Although we here focus on mRNA levels due to genome-wide coverage of the available transcriptomics datasets, the proposed model can also be readily applied to enzyme abundance or activity data.
The role of reaction kinetics in controlling metabolite concentration has been previously examined mostly from the perspective of the isolated reaction-metabolite pairs. With such a reaction-centric approach, a previous study on yeast metabolism was able to partially explain changes in the intracellular metabolite levels when using protein abundance as a measure of enzyme availability [3]. However, no correlation was observed in the same study when using gene expression data instead of protein abundances. One possible reason for the lack of strong correlation between gene expression and metabolite levels when looking at the isolated enzyme-metabolite pairs is that the large connectivity inherent to metabolic networks is not taken into account. A large fraction of intra-cellular metabolites participate in multiple reactions. For example, over 25% of the yeast metabolites participate in more than three reactions [17]. Consequently, abundance of an enzyme catalyzing a particular reaction cannot completely determine the concentrations of the participating metabolites or the rate of the reaction. Indeed, correlations between mRNA and fluxes, and even between enzyme activities and fluxes, have been often found to be poor [18]–[21]. Approaches accounting for the network connectivity of metabolites have been successful in linking gene expression to metabolites in an empirical or qualitative manner [10], [12], [22]–[24], but have achieved only a limited success on the quantitative front. Advantages of both reaction-centered kinetics approaches and network topology-based approaches can be combined in network kinetic models that include detailed kinetics of all involved reactions [25]–[28]. However, application of kinetic models to large metabolic networks is difficult due to their reliance on a large number of parameters. Such parameters are either currently unavailable, or their estimation requires comprehensive measurements of intra-cellular states of interest (e.g. metabolite concentrations, enzyme abundances, and fluxes) in the vicinity of the perturbation to be modeled.
In this study, we propose a steady-state model of the transcriptional control of metabolite concentrations. Our model integrates reaction kinetics and metabolic network connectivity constraints without requiring the knowledge of kinetic parameters. In essence, the model uses mass balance constraints to bridge the individual reaction kinetic constraints with those of the other reactions in the network. The resulting equations provide a log-linear relationship between the fold-change in the concentration of a given metabolite to the fold-changes in the expression of its neighboring genes, as well as topologically more distant genes.
By analogy to flux coupling analysis [29], which describes how steady-state fluxes are linked with each other, we termed our approach Concentration Change Coupling Analysis (CoCCoA). Starting with a classical reaction kinetics model, which treats each reaction as an isolated system consisting of a single enzyme and its substrate, we developed a network kinetics approach by accounting for the interactions between different reactions through their shared metabolites. As there is currently a lack of information on in vivo enzyme kinetics mechanisms at the network-scale, we used the single-substrate Michaelis-Menten (MM) kinetics for all reactions. In essence, MM kinetics describes the flux or reaction rate V as a function of three parameters: i) concentration of the substrate, S; ii) maximum capacity of the enzyme pool, Vmax; and iii) a parameter reflecting the enzyme's kinetic properties, KM (Figure 1A). The central idea of CoCCoA is to use mass balance constraints on the flux term V to link single-reaction kinetics to the other reactions in the network. We considered MM kinetics in the fold-change space, which allowed us to eliminate the need to know the KM values. For each metabolite, CoCCoA provides an overall transcriptional change score (CoCCoA score) according to the CoCCoA equations, which are developed in the subsequent sub-sections. To assess the proportion of variance in metabolite changes that can be attributed to transcriptional regulation, we compared the calculated CoCCoA scores with the experimentally measured metabolite concentration changes. The overall workflow used is depicted in Figure 1B.
The first step in our analysis is to calculate a representative transcriptional fold-change for each reaction. As the yeast metabolic network consists of several reactions that are each governed by multiple proteins, we classified all reactions into three types: i) reactions catalyzed by a single enzyme, ii) reactions catalyzed by two or more isoenzymes, and iii) reactions catalyzed by enzyme complexes. We then applied the following rules to calculate the representative fold-changes for all reactions: in the case of isoenzymes, we averaged the fold changes of the related transcripts, while in the case of complexes, we picked the transcript with the lowest fold change (Figure 1C). We used only significantly changed transcripts (P-value≤0.05) in the presented analysis. Relaxation of this filtering criterion did not change the overall results (Figure S1).
We used four published experimental datasets for evaluating the proposed CoCCoA models. These case studies included three pairwise comparisons – one genetic [3] and two environmental perturbations [23], [30] – and a time-course dataset obtained during the yeast metabolic cycle [9], [31]. In all pair-wise comparison studies, both gene expression and metabolite concentration data were obtained from the same experiment. In the case of the metabolic cycle data, although the sampling for transcriptome and metabolome was performed in two separate experiments, the experimental setups were identical and the sampling was performed at comparable time-points spanning all phases of the metabolic cycle. While the metabolic cycle is fundamentally a non-steady-state phenomenon, the observed transcript oscillation period of about 300 minute means that a reasonable degree of pseudo-steady-state can be assumed for applying our model. In the case of the three pairwise comparison studies, the correlations between CoCCoA scores and metabolite concentrations provided a perturbation-centered perspective wherein the responses of different metabolites were analyzed jointly. The metabolic cycle case study allowed us to additionally evaluate the gene expression-metabolite relationship from a metabolite-centered perspective, wherein the response of each metabolite was assessed individually for its conformity to the proposed model.
A genome-scale metabolic reconstruction of S. cerevisiae [17] was used to obtain the metabolite-reaction-gene connectivity information and to estimate the reaction directionalities (Methods). For each case study, we used the experimental measurements of exchange fluxes (uptake and secretion rates of metabolites) to constrain and simulate a flux balance model. Accordingly, we removed all blocked reactions and reactions for which the flux directions could not be unambiguously assigned. We also excluded the reactions for which the predicted flux directions did not agree between the two conditions being compared.
Depending on the extent to which the network connectivity information is included in the calculations, we term the CoCCoA models as 0th degree, 1st degree, 2nd degree, and so on (Figure 1D). 0th degree CoCCoA relies on the enzyme kinetics alone and thus considers only the consumption reaction(s) of any given metabolite. 1st degree CoCCoA additionally considers the production of the metabolite by using mass balance constraints. 2nd degree CoCCoA further expands the degree of network connectivity accounted for in the model by including the producing reactions of the precursors of the metabolite in question. Alternatively, the 2nd and higher degree models can also be expanded on both the consumption and production sides of the metabolite as described in the following sub-sections (also see Text S1).
We consider metabolite concentration changes relative to a reference condition that can be arbitrarily chosen from the conditions pertaining to the experiment under investigation. Assuming that the enzyme properties (represented by KM) remain unchanged in the experiment, by using MM kinetics one obtains (Text S1):(1)
Where * denotes the reference condition. The relative nature of this formulation allows circumventing the problem of the lack of availability of in vivo KM values. Furthermore, by assuming that &, and that the ratio can be approximated by the gene expression ratio, equation (1) simplifies to a log-linear relationship (equation (2), Text S1). Both of these assumptions are critically examined in the next sub-section. The model represented by equation (2) is hereby termed 0th degree coupling, meaning that the metabolite S is not considered to be directly coupled to any other metabolite and is connected only to the enzyme that uses it as a substrate (Figure 1D).(2)
The first assumption used in deriving equation (2) implies that the enzyme is not saturated. The opposite situation, i.e. an enzyme approaching saturation, is not amenable for establishing the metabolite-gene expression relationship (or metabolite-enzyme abundance relationship in general), as the reaction velocity will then be only a weak function of the substrate concentration. Recent studies have shown that in vivo concentrations for several metabolites, especially from central carbon metabolism, are close to the corresponding KM values [32]. At these concentrations, reaction rates V are close to half of the Vmax. Although the assumption of V≪Vmax is not strictly applicable in this flux regime, numerical simulations showed modest errors (around 20%) due to this approximation (Figure S2). Moreover, if the saturation level does not change drastically between the two conditions being compared, the error remains close to zero (Figure S2). Given the advantage that this approximation brings, namely elimination of the need for knowing the in vivo kinetic parameters, the cost of the approximation error appears to be acceptable.
The second major assumption is that the fold-change in mRNA level can be used as a proxy for the fold-change in enzyme abundance and ultimately for the fold-change in Vmax. Critical examination of this assumption is of particular importance as the role of translational efficiency and post-translational modifications in regulating metabolic enzymes is becoming increasingly evident [18], [33]–[36]. We examined our assumption by analyzing published experimental data for S. cerevisiae where genome-wide mRNA and protein fold changes were simultaneously measured. In support of the assumption, the correlations between the mRNA and the protein fold changes corresponding to the metabolic genes were found to be both significant and strong (Dataset 1–3 [37], [38], R2 = 0.77, P = 0.04; R2 = 0.66, P = 0.0365, R2 = 0.76, P = 0.0036; dataset 4 [39], R2 = 0.4, P = 0.296; dataset 5 [40], R2 = 0.57, P = 0.0681; dataset 6 [41], R2 = 0.43, P = 0.0001) (Figure 2A). We note that these correlations involving only metabolic genes are stronger than the correlations calculated by including the non-metabolic genes (Figure 2A, Figure S3). As mRNA and protein levels have recently been demonstrated to be in good agreement in mammalian systems as well [42], we expect that the assumption of proportionality between gene expression and protein abundance fold changes will be valid in a broad range of organisms.
Under the condition of flux homeostasis, i.e. no flux change between the two conditions being compared, the metabolite concentration ratio in equation (2) becomes dependent only on the transcript change. The resulting 0th degree CoCCoA model is equivalent to the analysis of the transcript/protein-metabolite relationship reported by Sauer and co-workers [3]. According to this model, we observed a significant correlation between transcriptional and metabolite changes in the glucose pulse case study (r = 0.81, P = 0.048). In the other two pairwise comparison case studies, the 0th degree model failed to correlate with the experimentally observed metabolite changes (Figure 3B). In case of the metabolic cycle dataset, around 31% of the measured metabolites showed significant correlations (FDR 10%) (Figure 3C, Figure S1). In all four case studies investigated here, flux homeostasis cannot be assumed as the growth rate as well as the substrate uptake and product secretion rates were affected by the corresponding perturbations. Our attempts to obtain reliable flux estimates by using flux balance analysis were not fruitful since only a limited number of physiological measurements were available to constrain the model, resulting in a high degree of uncertainty in the flux estimates. Thus, in the absence of reliable intra-cellular flux estimates, the 0th degree model was found to be insufficient for relating gene expression changes to metabolite levels.
At steady state, the sum of fluxes producing a particular metabolite must be equal to the sum of fluxes through the reactions that use it as a substrate. For a metabolite with a single production reaction and a single consumption reaction, the steady state assumption combined with equation (2) leads to equation (3) (Supplementary Text S1).(3)
Tprod and Tcons denote expression levels of the genes corresponding to the enzymes producing and consuming S, respectively. R refers to the concentration of the metabolite that is the precursor of S. The relation described by equation (3) implies a coupling between the concentration changes in R and S, and is here defined as 1st degree coupling. In comparison to the 0th degree coupling, the flux term ln(V/V*) is eliminated in the 1st degree coupling equation and is replaced by two new terms, ln(Tprod/Tprod*) and ln(R/R*). Equation (3) brings a new network perspective to enzyme kinetics, whereby gene expression and metabolite concentration changes in the adjacent reactions are linked through the mass balance constraint. Each metabolite pool is thus linked to the reactions consuming it as well as on the reactions producing it (Figure 3A). When multiple reactions are consuming (or producing) the same metabolite S, the consumption (or production) term can be approximated by the geometric mean of the transcript ratios of all the consumption (or production) reactions (Supplementary Text S1).
To evaluate the 1st degree model, we compared the experimentally measured metabolite concentration ratios with the 1st degree CoCCoA scores based on the transcript fold changes – the first two terms on the right-hand side of equation (3). We note that, although the strict application of our model requires the use of the ln(R/R*) term, these measurements are often not available. Moreover, a model that is completely independent of the metabolite concentration data will likely be of more practical value. Omitting the ln(R/R*) term equates to assuming that the preceding metabolite's concentration does not change between the two conditions; an alternative to omitting this term is explored below, in the 2nd degree CoCCoA model, in which the ln(R/R*) term is estimated by use of the 1st degree CoCCoA model. The effect of omitting the ln(R/R*) term on CoCCoA scores is further discussed in the later sub-section “Post-transcriptional regulation”.
Comparisons of the 1st degree CoCCoA scores to metabolite concentrations yielded significant correlations in the two environmental perturbation case studies (r = 0.88, P = 0.0099 and r = −0.96, P = 0.038), and a reasonably good correlation (r = −0.61, P = 0.06) for the genetic perturbation case study (Figure 3B, Figure S1). For the metabolic cycle case study, around 25% of the measured metabolites showed significant correlations (FDR 10%) (Figure 3C, Figure S1). Although our model suggests positive correlation between CoCCoA scores and metabolite changes, we observed negative correlations in the cases of two of the pairwise comparisons and for some of the metabolites in the metabolic cycle case study. The possible reasons underlying this discrepancy are discussed in the subsequent sub-section “Negative correlations in CoCCoA”. We maintain that the negative slopes do not invalidate the significance of the observed correlations, but rather hint at the existence of unaccounted parameters/mechanisms leading to the reversal of slope in some cases.
The number of transcripts that can be used for the calculation of the CoCCoA scores typically increases as more distant reaction nodes in the network are included with the increasing CoCCoA degree. Consequently, the number of metabolites that could be assigned CoCCoA scores varied between the coupling degrees. For example, in the C-source change study [30] (Figure 3B), only 4 metabolites have significant transcript changes corresponding to their consuming reactions and hence only these could be compared against the experimental data for the 0th degree analysis. In contrast, 2nd degree CoCCoA scores could be calculated for 7 metabolites.
In a similar manner as going from the 0th to the 1st degree coupling, the CoCCoA equations can be further extended to include more distant nodes in the metabolic network. By replacing the concentration ratio in the right-hand side of Equation 3 (i.e. R/R*) with the 1st degree CoCCoA relationship for the corresponding precursor metabolite (in this case, R), we obtained the 2nd degree coupling relationship. This 2nd degree model accounts for the gene expression changes corresponding to the precursor's production reactions (Figure 3A) (Text S1). In all case studies, the 2nd degree correlations remained as strong as for the 1st degree. This result is notable since the 2nd degree coupling score includes expression data from the genes that are further away from the metabolites in question. With further extension of the CoCCoA model in a similar manner, we observed significant correlations up to the 6th degree coupling (P≤0.05, Fendt et al. case study, Figure S4).
To gain further insight into the metabolite concentration control at different network distances, we examined this problem from a metabolite-centric perspective by taking advantage of the broad metabolite coverage of the metabolic cycle case study. First, we extended the higher degree CoCCoA formulation so as to include information from all intermediate reaction steps up to the desired degree (Text S1). For example, the calculation of the 3rd degree CoCCoA score includes fold changes from the genes associated with the reactions involving all metabolites that are three steps upstream or downstream from the metabolite in question. The inclusion of genes within a desired network distance can either be restricted to the consumption or the production side of the metabolite, or both included simultaneously. This formulation also allowed us to include, if available, measured concentrations of the neighboring metabolites within the desired distance of a given metabolite, and thereby to assess the effect of changes in neighboring metabolites over its concentration. The algorithm used for calculating CoCCoA scores using this formulation is described in Text S1. In brief, this algorithm first enumerates all paths starting from the metabolite of interest to identify genes that are within a given network distance. Next, it uses graph topology-based heuristics to weight and incorporate the expression fold-changes corresponding to these genes into the CoCCoA equations by using mass balance considerations. Following this, we evaluated the ability of these CoCCoA scores to explain the concentration changes observed during the metabolic cycle. Overall, positive correlations are apparent for most of the metabolites (∼66%, Figure 4A,C). For the long distance scores, a slightly lower number of metabolites showed positive correlations. The contrast between the close and distant neighbors, however, should be interpreted in light of the highly connected nature of the metabolic network. The numbers of genes that are included in the calculation of CoCCoA scores already reach a plateau at the 4th degree (Figure 4B), and thus, even relatively modest distances can mean inclusion of a very large fraction of the network. Consequently, the CoCCoA scores for a given metabolite will become ‘diluted’ due to the noise stemming from the inclusion of gene expression changes pertaining to the reactions that are only indirectly affecting the metabolite of interest. These results from the metabolic cycle case study, together with the results from the pairwise comparison studies, suggest that the close neighbors in the metabolic network exert the majority of the control over metabolite concentrations.
The correlations between the CoCCoA scores and the metabolite concentration changes were further strengthened when the experimentally determined fold-changes in the concentration of the upstream and/or downstream metabolites were also used in the calculation (Figure 4C). This improvement further supports the CoCCoA theory, as the inclusion of concentration changes for the upstream/downstream metabolites stems from the joint mass balance and kinetic considerations, e.g. as illustrated in equation (3).
CoCCoA is not applicable in the case of perturbations that are likely to drastically affect the kinetic properties (KM values) of several enzymes in the network, or if the metabolite concentrations are considerably above the corresponding KM values (saturated enzymes). Furthermore, the CoCCoA model needs to exclude reactions for which the flux directionality is ambiguous, and it assumes that the flux directions do not change for the rest of the reactions. Post-translational regulatory mechanisms, which can affect the kinetic parameters, are also not included in the current CoCCoA formulation, as sufficient data are not available to enable their modeling. The latter is perhaps the most restrictive limitation of our model. Post-translational regulation is known to play a crucial role in the yeast central metabolism, wherein several enzymes are controlled by allosteric binding of small molecules [35], [43], [44] and/or through post-translational modifications such as phosphorylation [33]. Together, these various assumptions and limitations can lead to poor or no correlations. In the case of the pairwise comparisons, poor correlations can also result from the pooling of metabolites with positive and negative correlation with CoCCoA scores into a single plot.
According to the CoCCoA model, all examined correlations would be expected to be positive. Among the pairwise comparison case studies, positive correlations were observed only for the glucose pulse study (Figure 3B). In the case of the metabolic cycle study, a significant majority of the metabolites (∼66%, P = 2.9×10−8, exact binomial test) showed positive correlations (Figure 4A,C). Flux regulation due to allosteric binding by small molecules and post-translational modifications are likely to be the major factors underlying this discrepancy between the expected positive slopes and the observed negative slopes for the remaining 34% metabolites. The possible causes and implications of negative correlation are discussed in the subsequent sub-section “Negative correlations in CoCCoA”.
We found that the enzyme-coding genes in yeast exhibit significantly stronger correlations between mRNA and protein fold-changes than do the non-metabolic genes (Figure 2A, Figure S3). The slopes of these correlations were, however, different across different datasets examined. To evaluate the robustness of CoCCoA towards this variation, we re-performed 1st degree CoCCoA analysis multiple times (1000 simulations), adjusting the transcript ratios in each simulation by a randomly sampled correction factor to account for the expected difference between the mRNA and protein fold changes. The sampling space for the correction factors was estimated based on the variance in the slopes of linear regression lines between mRNA and protein abundance fold changes across different datasets (Figure S3, Text S1). We then examined the number of simulations in which the correlation between the 1st degree CoCCoA scores and the metabolite fold changes remained significant. For all three pairwise comparison case studies, the correlations remained significant (P≤0.05) in a large fraction of these simulations (99%, 86% and 91%) (Figure 2B).
The correlation between metabolites concentrations and transcript fold changes becomes evident only when including network connectivity constraints (1st and higher degree CoCCoA). Thus, inclusion of gene expression changes associated with the both upstream and downstream reactions was critical for explaining metabolite concentration changes. For all three pairwise comparison case studies, CoCCoA models explained more than 60% of the variation in metabolite changes based on gene expression. For the metabolic cycle case study, CoCCoA could explain variation in about 33% of the measured metabolites, with correlation coefficients as strong as 0.90.
The use of a genome-scale metabolic model was crucial in CoCCoA analysis in order to capture the large connectivity inherent to metabolic networks. Even for a sparsely connected metabolite such as D-Ribose 5-phosphate, for which the 0th degree score accounted for only 4 transcripts, the 2nd degree CoCCoA score accounted for transcriptional information from 47 genes in the Fendt et al. study [3] (Figure 3B). With the increasing degree of CoCCoA equations, larger numbers of genes become part of the CoCCoA score (Figure 3E, Figure 4B). We also observed that, in general, the inclusion of new genes when moving from the 1st to the 2nd degree CoCCoA maintains the significance of the correlation (Figure 3B, Figure 3C). This observation implies strong co-regulation of genes that are linked through common substrates/products. Indeed, co-expression of metabolic genes at short network distances has been observed in earlier studies [45]. The CoCCoA theory suggests that homeostasis of metabolite concentrations is one of the objectives of such topology-oriented co-regulation in metabolic networks.
There are several layers of regulation that the CoCCoA model does not take into account: translational and post-translational regulation, and the kinetic effects of neighboring metabolites' concentrations. With regard to translational regulation, we observed strong correlations between transcript and protein fold changes (Figure 2A, Figure S3), and the CoCCoA results were found to be relatively robust in light of the known variability (Figure 2B). However, it remains to be seen whether these correlations extend through post-translational regulation to enzyme activity. Increasing availability of the genome-wide protein phosphorylation/acetylation data may aid in addressing this question. The CoCCoA framework can be used with enzyme abundance or enzyme activity fold changes in place of transcript fold changes. Using this information could lead to more accuracy in CoCCoA scores, and could furthermore aid in identifying the relative contribution of the different regulatory layers in controlling metabolite concentrations.
The effects of neighboring metabolites' concentrations could be examined thanks to the large coverage of metabolite measurements in the metabolic cycle dataset. The power of CoCCoA in explaining variance in metabolite concentration changes was substantially improved following the inclusion of data for the neighboring metabolites, representing ln(R/R*) terms in the CoCCoA equations (2.54 fold increase in the median correlation coefficient; P = 6.6×10−8, Wilcox test; Figure 4C). In accordance with the CoCCoA theory, this improvement indicates that a metabolite's neighbors in the metabolic network play an important role in determining its level.
Formulation of CoCCoA in a relative manner, i.e. in the fold-change space, allowed us to circumvent the problem of the unavailability of in vivo kinetic parameters. A major advantage following this relative formulation is that the CoCCoA models do not need any parameter fitting. Indeed, the applicability of CoCCoA was found to be quite broad in terms of the perturbation or experimental design underlying the data. The four datasets considered in this study represent three different biological perturbations, namely gene knockout [3], nutrient pulse [23], change in carbon source [30], as well as a fundamental rhythmic phenomenon associated with the cell cycle [9], [31]. These case studies also span two distinct cultivation types, batch [3] and chemostat [9], [23], [30], [31]. We also verified the differences in the nature of these perturbations at the level of gene expression changes: the three pairwise comparison studies were found to have only a small overlap in terms of the significantly responding genes (Figure 3D). Additionally, the CoCCoA model was found to be applicable over a broad range of concentration changes displayed by different metabolites during the yeast metabolic cycle (Figure 4D).
Intriguingly, several of the observed negative correlations were found to be not only significant but also quite strong, with R2 values as high as 0.86 (Figure 4). These negative correlations are indicative of the mechanisms that are unaccounted for in the CoCCoA model and/or highlight cases in which the assumptions of the model do not apply. We observed that the number of positive correlations in the metabolic cycle case study increased considerably when using the concentration change data from the neighboring metabolites (the ln(R/R*) term in the right-hand side of the 1st degree CoCCoA equation) (Figure 4C). This observation suggests that the kinetic effect due to changes in the neighboring metabolites is an important factor contributing to negative correlations. The results from the analysis of the metabolic cycle data also hint that the negative slopes might be characteristic to certain metabolites, for example, those for which the producing/consuming enzymes are regulated predominantly and/or prevalently at the post-translational level. Indeed, we found that the metabolites with poor or negative correlations in the metabolic cycle case study are enriched in the metabolites with previous evidence for post-translational regulation of at least one of their neighboring enzymes (metabolites marked with † in Figure 4A, data for post-translationally regulated enzymes from [34], P = 0.0006). In these cases, the post-translational regulation may be counteracting the transcriptional change. Post-translational regulation of an enzyme can change its KM value and can thereby directly affect CoCCoA scores. Consider, for example, 1st degree CoCCoA score. When changes in KM values are included in the 1st degree CoCCoA formula, the score becomes: . If the post-translational modifications counteract the transcriptional changes, the KM ratios in this new score will partially cancel out or even override the transcript ratios. When assuming constant KM values, the discrepancy between the transcript and KM ratios in some cases might be sufficiently large to result in CoCCoA scores with opposite sign. On the other hand, the inconsistency between the directions of transcriptional and post-translational regulation (or post-transcriptional regulation in general) implies non-optimal regulation and is unlikely to be a general mechanism used by the cell. However, non-optimal regulation is a possibility for certain enzymes, with two plausible biological explanations: i) the highly non-linear scenario of regulation (resulting from the concerted action of reaction kinetics, incl. allosteric regulation, and mass balance and thermodynamic constraints) can mean that the cell needs to make some locally non-optimal choices in order to achieve a global optimality in regulating the overall metabolism (for example, to take advantage of the distinct time-scales at which post-translational and transcriptional regulations act); ii) the observed behavior is both locally and globally non-optimal in case of certain perturbations. The second scenario would imply that the perturbations in question are unknown or new to the cells in the evolutionary sense.
In addition to the unaccounted post-transcriptional regulation, the simplifying assumptions of constant flux directions and flux split ratios may also be contributing to the observed negative correlations. A wrongly considered flux direction for a reaction would mean that the corresponding fold change in the expression level would be treated in the opposite direction. Similarly, moderate changes in the flux split ratio can also cause a sign reversal in the CoCCoA score if one of the fluxes is significantly lower than the other(s). The interaction between the various missing/simplifying factors in our model can further amplify the difference between the CoCCoA scores and actual concentration changes. How these interactions lead to the reversal of correlation while retaining statistical significance is yet unclear. Further investigation into the mechanisms underlying these intriguing negative correlations would require network-wide in vivo measurements of fluxes, metabolite concentrations, protein abundances and functional post-translational modifications. Nevertheless, we note that our model revealed significant correlations in several cases, including perturbations of very different nature. We also note that, in the case of metabolic cycle dataset, significantly more positive correlations were observed than negative (P = 2.9e-08, exact binomial test). Together, the statistical and mechanistic considerations suggest that the CoCCoA model captures considerable mechanistic essence of the complex processes governing metabolite levels.
Overall, our model-guided analysis highlighted the role of metabolic network connectivity in modulating metabolite concentration changes and revealed much stronger correlations between gene expression and metabolite levels than previously appreciated. The proposed model can be extended to include translational and post-translational regulation as the data becomes available. From this perspective, we see CoCCoA as a framework with a strong mechanistic yet parameter-free basis, rather than a general relationship. We anticipate that, due to its parameter-free nature, the CoCCoA framework will be widely applicable for modeling metabolite level changes in large metabolic networks.
Four different experimental studies reporting gene expression and metabolite concentration measurements for the yeast S. cerevisiae were used. The first study, Fendt et al. [3], includes a comparison between wild type yeast and a mutant strain lacking GCR2, a transcription factor responsible for activation of glycolytic genes [46]. In the second study, Kresnowati et al. [23], yeast cultures were grown in carbon-limited chemostat cultures and subjected to a step change in glucose concentration. In the third study, Wisselink et al. [30], an evolutionarily engineered strain was grown on either glucose or arabinose as the sole carbon source. In the fourth study, oscillating chemostat cultures were sampled covering different phases of the metabolic cycle [9]. A summary of the growth conditions and descriptions of datasets from all case studies is provided in Table S1. Metabolite data was used as available in the original studies; a significance cut-off α = 10% was chosen to control for the type 1 error. For the metabolic cycle data, we considered metabolites with the periodicity P-value≤0.05 as reported in the original study [31]. As the datasets used did not distinguish between the cytosolic and mitochondrial concentration of metabolites, we regarded all metabolite concentrations as cytosolic. Since 2-phosphoglycerate and 3-phosphoglycerate are usually indistinguishable in the MS measurements, we considered only 3-phosphoglycerate gene neighbors and excluded 2-phospoglycerate.
A genome-scale reconstruction of Saccharomyces cerevisiae metabolic network by Forster et al. [17] was used to map metabolite-reaction-gene connectivity. For each of the case studies, the functional reaction directions of reversible reactions were estimated by using flux variability analysis [47]. For this purpose, the model was constrained with the physiological data obtained from the publications reporting the used datasets (Table S2, Table S3, Table S4, and Table S5) [3]. Linear programming problems were solved using the glpk solver (http://www.gnu.org/software/glpk/) accessed through a C library.
Preprocessing of the Affymetrix CEL files was carried out with the statistical software environment R/Bioconductor (www.bioconductor.org). Probe intensities were corrected for background by using robust multi-array average method (RMA) [48] using only perfect-match probes, and normalization was performed using the quantiles algorithm. Gene expression intensity values were calculated from the perfect-match probes with median polish summarization method [49]. Significance of the differential expression was calculated by using the empirical Bayes test as implemented in the limma package [50].
Pearson correlation coefficients between log2 metabolite fold changes and CoCCoA scores were calculated with the statistical software R (www.r-project.org) using the function cor.test(). Metabolite changes were used as dependent variables and CoCCoA scores as independent. P-values for the null hypothesis of no correlation (regression slope = 0) were estimated by using the same function. In addition, we performed a permutation test by shuffling gene labels before calculating CoCCoA scores. The originally paired data was randomly permuted without replacement 1000 times. For each permutation, a correlation coefficient was calculated and the P-value was estimated as a fraction of squared correlation coefficients that were larger than in the case of the original paired data. The results were similar to those estimated with the cor.test() function.
http://www.patil.embl.de/supplementary
|
10.1371/journal.ppat.1001266 | Entrapment of Viral Capsids in Nuclear PML Cages Is an Intrinsic Antiviral Host Defense against Varicella-Zoster Virus | The herpesviruses, like most other DNA viruses, replicate in the host cell nucleus. Subnuclear domains known as promyelocytic leukemia protein nuclear bodies (PML-NBs), or ND10 bodies, have been implicated in restricting early herpesviral gene expression. These viruses have evolved countermeasures to disperse PML-NBs, as shown in cells infected in vitro, but information about the fate of PML-NBs and their functions in herpesvirus infected cells in vivo is limited. Varicella-zoster virus (VZV) is an alphaherpesvirus with tropism for skin, lymphocytes and sensory ganglia, where it establishes latency. Here, we identify large PML-NBs that sequester newly assembled nucleocapsids (NC) in neurons and satellite cells of human dorsal root ganglia (DRG) and skin cells infected with VZV in vivo. Quantitative immuno-electron microscopy revealed that these distinctive nuclear bodies consisted of PML fibers forming spherical cages that enclosed mature and immature VZV NCs. Of six PML isoforms, only PML IV promoted the sequestration of NCs. PML IV significantly inhibited viral infection and interacted with the ORF23 capsid surface protein, which was identified as a target for PML-mediated NC sequestration. The unique PML IV C-terminal domain was required for both capsid entrapment and antiviral activity. Similar large PML-NBs, termed clastosomes, sequester aberrant polyglutamine (polyQ) proteins, such as Huntingtin (Htt), in several neurodegenerative disorders. We found that PML IV cages co-sequester HttQ72 and ORF23 protein in VZV infected cells. Our data show that PML cages contribute to the intrinsic antiviral defense by sensing and entrapping VZV nucleocapsids, thereby preventing their nuclear egress and inhibiting formation of infectious virus particles. The efficient sequestration of virion capsids in PML cages appears to be the outcome of a basic cytoprotective function of this distinctive category of PML-NBs in sensing and safely containing nuclear aggregates of aberrant proteins.
| Many DNA viruses, including varicella-zoster virus (VZV), a herpesvirus that causes varicella (chickenpox) and zoster (shingles), replicate in the host cell nucleus. Here, we have identified an intrinsic antiviral mechanism that specifically targets newly assembled VZV capsids and contains these essential viral structures in a nuclear “safe house”. Using immuno-electron microscopy, PML (promyelocytic leukemia) protein fibers that formed filamentous spherical cages were shown to trap virion capsids very efficiently, preventing their transport out of the nucleus and inhibiting the formation of infectious virus particles. PML cages containing virion capsids were found in VZV-infected neurons and satellite cells in human sensory ganglia and in skin cells, which are major targets during VZV pathogenesis. Similar PML nuclear bodies that sequester abnormal proteins have been reported in neurodegenerative disorders, like Huntington's disease. We found that cages formed by PML isoform IV sequestered both the virion capsids of VZV, which is a neurotropic herpesvirus, and the mutant Huntington's disease protein. This work provides the first evidence that PML, which is abundant in mammalian cell nuclei, can function both to contain potentially damaging cellular protein aggregates and as an intrinsic host defense against a herpesvirus during nuclear virion assembly.
| Promyelocytic leukemia protein (PML) is a major organizing component of structures that are referred to as PML nuclear bodies (PML-NBs) or nuclear domain 10 (ND10) bodies [1]–[3]. These nuclear bodies are heterogenous in size, shape and molecular composition [4]–[6], are prominent in most mammalian cell types and participate in many basic cellular functions, including transcriptional regulation [7], DNA repair [8] and apoptosis [9]–[12]. Human PML, located on chromosome 15, has nine exons and alternative splicing of PML transcripts produces at least 11 isoforms [13], [14]. PML isoforms share a conserved N-terminus, which has the characteristic RBCC/TRIM motif, including a RING finger domain, B boxes and a coiled coil domain. The PML N-terminus is important for PML heterodimer formation and oligomeriztion but each isoform has a unique C-terminal domain [14], [15]. The PML isoforms create PML-NBs with varying morphologies but the functional implications of these differences are not well understood [16]–[18]. While little is known about the tissue or cell specific patterns of expression of individual PML isoforms, these non-conserved PML regions are of interest for their likely involvement in isoform-dependent functions within particular cell types and in the response to changing intracellular or extracellular conditions [17].
Since their discovery, PML protein and PML-NBs have been investigated for their role in the virus-host cell interactions of DNA viruses that must replicate in the mammalian cell nucleus [19], [20]. These viruses include the very extensive Herpesviridae family as well as adenoviruses, papillomaviruses, polyomaviruses and other pathogens. During herpesvirus infection, genome copies are synthesized in nuclear replication compartments and genomic DNA is packaged into icosahedral nucleocapsids (NCs) formed by the major capsid protein and smaller capsid surface proteins. After assembly, NCs egress across the nuclear membrane for secondary envelopment in the cytoplasm and release as infectious virus particles [21], [22]. PML-NBs have been implicated in controlling the replication of herpes simplex virus (HSV) 1 and human cytomegalovirus shortly after virus entry by mechanisms that limit early viral gene transcription [19], [20], [23], [24]. To overcome these antiviral effects, HSV-1 targets PML for immediate proteosome-mediated degradation through functions of the viral ICP0 ubiquitin ligase protein [25]–[28]. PML is also an interferon (IFN)-inducible protein and must be degraded to prevent IFN-mediated inhibition of HSV-1 [29]. Like HSV-1, varicella-zoster virus (VZV) is a common human alphaherpesvirus [30]. In contrast to HSV-1 infection, PML protein and some PML-NBs persist in VZV-infected cells [31] although not in association with early nuclear replication compartments [32]. Depleting PML enhances VZV replication, suggesting a possible role for PML in the host cell defense [31].
VZV, which causes varicella (chickenpox) and herpes zoster (shingles), is a highly human-restricted pathogen [30], [33]. However, VZV pathogenesis can be investigated in vivo using xenografts of human dorsal root ganglia (DRG) and skin in a severe combined immunodeficiency (SCID) mouse model [34]–[36]. This model allows the analysis of viral and cellular mechanisms that facilitate or inhibit VZV when differentiated neural and skin cells are infected within their usual tissue microenvironments in vivo and in the absence of antiviral effects mediated by the adaptive immune response [37], [38]. The fundamental importance of intrinsic and innate cellular responses in achieving the usual balance between VZV and its human host is evident from the observations that VZV undergoes a transition to persistence in neurons within DRG xenografts without requiring VZV-specific adaptive immunity [36] and that VZV skin infection is highly regulated by type-I interferon produced by epidermal cells [39]. In these experiments, we have used the SCID mouse model to define VZV interactions with PML-NBs in human neural and epidermal cells infected with this nuclear replicating DNA virus in vivo.
The characteristics and functions of PML-NBs in human neural cells are of particular interest because VZV is a neurotropic herpesvirus [30]. Aberrant proteins generated in several neurodegenerative diseases, including Huntington's disease and spinocerebellar ataxias, accumulate in PML-NBs within neural cells and the usual nuclear distribution of PML is altered by their expression [40]–[45]. These diseases, classified as polyglutamine (polyQ) disorders, are associated with an unstable CAG repeat expansion resulting in the elongation of a polyglutamine (polyQ) tract in the abnormal gene product [46]. The mis-folded, oligomeric polyQ proteins are neurotoxic and form intranuclear aggregates, often with PML and proteins of the ubiquitin-proteasome system [47], [48]. Similar structures have been referred to as clastosomes [44], [49] and in some cases appear to protect the cell by confining the abnormal proteins to an intranuclear “safehouse” where they may be degraded [44], [50]–[52]. Here we report that endogenous PML forms distinctive, large PML-NBs consisting of spherical cages in the nuclei of neurons and satellite cells in human DRG and in skin cells during VZV pathogenesis. These nuclear organelles efficiently entrap newly assembled VZV capsids in vivo as well as in cultured cells in vitro. The PML-NBs closely resemble those that retain polyQ proteins in neurodegenerative diseases. Exogenous PML IV was the only isoform that sequestered VZV NCs and interacted with the ORF23 capsid surface protein, functions that required its unique C-terminus. PML IV-NBs also retained the Huntington's disease protein, Htt, and could simultaneously sequester NCs. Importantly, the nuclear entrapment of viral capsids in PML IV-NBs inhibited the production of infectious VZV progeny. These observations suggest that PML-NBs can retain nascent virion capsids in the infected cell nucleus, resulting in an intrinsic antiviral defense at later stages of viral replication.
Since PML protein and some PML-NBs persist in VZV infected cells [31], we first investigated the intracellular localization of VZV proteins that are found in nuclear viral replication compartments, including ORF29, the single stranded DNA binding protein and IE62, the major viral transactivating factor, as well as the ORF23 capsid protein, which is present in later stages of VZV infection when progeny virion assembly occurs [32]. As expected, analysis of uninfected human embryonic lung fibroblasts (HELF) by confocal microscopy showed numerous PML-NBs that contained both PML and SP100 protein, which are known components of PML-NBs [53](Figure 1A). In infected HELF, PML-NBs did not colocalize with viral DNA replication compartments, which were identified by ORF29 and IE62 expression [32], at this late time at 24 hr after infection (Figure S1A and B).
In marked contrast, the newly synthesized ORF23 capsid protein was present in almost all PML-NBs (94%, N = 228) at this stage of VZV infection (Figure 1B, left panel) using confocal imaging conditions that were optimized to identify bright PML-NBs associated with NCs. Based on its homology to the HSV-1 VP26 small capsid protein, ORF23 protein is predicted to be exposed on VZV NCs [54]. ORF23 protein was similarly enriched in PML-NBs in HELF infected with a VZV recombinant virus that expresses ORF23 protein tagged with the red fluorescent protein (RFP-ORF23) (Figure 1B, right panels) as well as in melanoma cells (Figure S2A). The number of PML-NBs decreased by about five-fold in both HELF and melanoma cells at 48 hr after infection compared to uninfected HELF and melanoma cells (Figure 1C and Figure S2B). Since melanoma cells have fewer PML-NBs than HELF before infection, the majority of infected melanoma cells had no PML-NBs left (Figure S2C). However, as previously reported [31], immunoblot analysis with a polyclonal anti-PML antibody showed that PML protein levels were not decreased in infected HELF or melanoma cells (Figures 1D and Figure S2D). The PML-specific bands were of varying sizes and may represent different PML isoforms or post-translational modifications of one or more of the PML isoforms. PML-NBs that colocalized with ORF23 protein in infected cells were also significantly larger than those in uninfected cells (Figure 1E). These results suggested that PML-NBs present at later stages of VZV infection sequestered ORF23 protein or possibly newly assembled NCs with surface ORF23 protein and might differ from the PML-NBs that modulate virus-cell interactions shortly after herpesvirus entry [55].
We next used cryoimmuno-electron microscopy (cryoimmuno-EM) to determine with ultrastructural precision whether the ORF23 protein that was associated with PML-NBs represented assembled NCs retained within these nuclear bodies or consisted of ORF23-PML protein co-aggregates. This question was of interest because PML has been observed to be associated with aberrant protein aggregates, termed nuclear aggresomes [56] or clastosomes [44] and with capsid proteins of other nuclear replicating viruses, including papillomavirus [57], [58], the neurotropic JC virus [59] and HSV-1 [60]. Except for the HSV-1 study, these analyses showed association of PML and capsid proteins by confocal light microscopy which does not optically resolve individual virion capsids; however some NC-like structures appeared to be associated with PML gold labeling in HSV-1 infected cells by immuno-EM [60].
Our experiments to detect both PML and ORF23 protein at ultrastructural resolution were first done with the Tokuyasu method which allows single and double-immunogold labeling of ultrathin (50–80 nm) cryosections with exquisite sensitivity [61]–[63]. These results were then correlated with patterns of PML and ORF23 protein localization observed by confocal microscopy (Figure 2). By cryoimmuno-EM, endogenous PML-NBs in uninfected cells were electron-dense subnuclear domains that exhibited extensive PML gold labeling, consistent with the punctate PML pattern seen by confocal microcopy (Figure 2A). In VZV infected cells, cryoimmuno-EM revealed VZV NCs of about 100 nm diameter (Figure 2B, arrow) that were abundantly labeled with antibody to ORF23 protein, confirming its predicted location on the capsid shell [54], [64], [65]. The viral NCs most likely corresponded to the tiny ORF23-expressing punctae seen in the nuclei of infected cells by confocal microcopy (Figure 2B, left panel). Furthermore, cryoimmuno-EM revealed clusters of VZV capsids that were surrounded by PML gold labeling (Figure 2C, right panel) and double-immunogold labeling proved the spatial association of PML protein (15 nm gold labeling) with individual NCs and clusters of VZV NCs detected with antibody to ORF23 protein (10 nm gold labeling) (Figure 2D). Therefore, the co-localization of PML-NBs and ORF23 protein detected by confocal microscopy (Figure 1 and Figure 2C, left panel) resulted from the sequestration of newly assembled VZV capsids in PML-NBs rather than from the formation of PML-ORF23 protein aggregates.
To define the ultrastructure of PML subnuclear domains harboring VZV NCs even more precisely, high-pressure freezing and freeze-substitution was used to optimally preserve both cell morphology and proteins [66] in uninfected and VZV infected HELF (Figure 3 and Figure S3). Again, endogenous PML-NBs in uninfected cell nuclei showed PML-specific labeling and were electron-dense (Figure 3A). In VZV infected cells, individual mature NC, which were identified by their dark electron-dense centers, as well as immature NCs, were associated with varying amounts of PML-positive fibrous material (Figure 3B and Figure S3). Furthermore, clusters of NCs were observed within cage-like structures that had dense PML-specific labeling (Figure 3C and Figure S3). These PML cages consisted of concentrically arranged protein fibers that tightly enclosed mature and immature NCs. Thus, the sequestration of virion capsids within PML subnuclear domains was confirmed with a second ultrastructural method. Furthermore, the presence of fibrous PML cages that entrapped both mature and immature herpesvirus capsids was demonstrated for the first time because of better contrast and superior ultrastructural preservation achieved with high pressure freezing and freeze-substitution.
Validating observations from EM studies requires quantitative analysis [67]. Examination of 30 VZV infected nuclei revealed that about 80% (N = 1,611) of all PML-specific gold particles were associated with viral NCs (Figure 3D). The concentration of NCs (number of NCs/µm2) within PML subnuclear domains was nearly 40 times higher than in other nuclear areas (Figure 3E), proving that PML and VZV capsids were highly co-enriched within PML-NBs and free PML protein was rare elsewhere in infected nuclei. In infected nuclei that contained endogenous PML-NBs an average of 62%±5 SEM (N = 30 nuclear profiles) of VZV NCs was sequestered in PML-NBs. The proportion of sequestered NCs (among all NCs present in an infected nucleus) was higher when more PML protein was assembled to PML cages and correlated positively with the amount of endogenous PML protein present in the infected cell nucleus, measured as the PML protein density (PML-gold/µm2) (Figure 3F). Importantly, the clusters of VZV NCs in these PML structures differed from the paracrystalline arrays of NCs that form nuclear inclusions in HSV-1 infected cells (compare Figure 3 and Figure S3 with Figure S4). Thus, the ultrastructural analysis unequivocally identified the association of both mature and immature virion capsids with PML protein in the nucleus and NCs were usually confined within endogenous cages consisting of PML-positive protein fibers in cells infected with VZV in vitro. Furthermore, these results demonstrated that immunogold EM is the method of choice to show the distribution of PML in infected nuclei unequivocally, including demonstrating the presence of PML on some individual NCs, that may be difficult to image by standard confocal microscopy. Immunogold EM has the significant advantage that individual gold particles constitute a stable, discrete and quantifiable signal that allows a very wide range of labeling densities within the same section to be imaged and quantified accurately.
Since mature herpesvirus NCs must egress from the nucleus and undergo secondary envelopment in the cytoplasm before infectious virus release from the host cell [22], the sequestration of VZV NCs by endogenous PML cages suggested a novel mechanism by which PML might restrict viral replication and spread. However, this observation might reflect a phenomenon unique to cultured cells because VZV is known for its poor replication in vitro. Therefore, we next explored whether this cellular response also occurred in the differentiated neurons and satellite cells of sensory ganglia that are targeted during acute VZV infection in vivo [30] and when VZV reactivates from latency in neurons [68]–[70]. For these experiments, human DRG xenografts in SCID mice were infected for 14 days (acute phase of VZV infection [36]) and then harvested and processed for semithin (500 nm) cryosectioning and confocal microscopy or cryoimmuno-EM.
Numerous small PML-NBs (<1 µm) were detected in the nuclei of neurons and satellite cells in uninfected DRG by confocal microscopy (Figure 4A). In contrast, in infected neural cells, PML-NBs were either dispersed or reorganized such that only one or two enlarged structures were present; the mean diameter was (1.9 µm±0.8 SD; N = 62). These structures were often ring-shaped and exhibited intense enrichment of ORF23 protein in their interior (Figure 4B–D). ORF23 protein was present in a majority but not all (87%, N = 116) of the ring-shaped PML-NBs in VZV-infected neural cells. Ultrastructural studies revealed that these large PML-NBs consisted of spherical PML cages containing numerous sequestered viral NCs, as demonstrated by cryoimmuno-EM with PML-specific labeling (Figure 4E).
VZV also targets skin for replication, creating cutaneous lesions from which the virus is transmitted during varicella or herpes zoster [30]. To determine whether NC sequestration by PML-NBs also occurred in skin cells, skin xenografts were infected for 21 days, harvested and processed for paraffin-sectioning and confocal microscopy. Like uninfected neural cells, uninfected skin cells contained numerous small (<1 µm) PML-NBs (Figure 5A). However, very large ring-like PML-NBs were observed in the nuclei of infected cells that formed the typical VZV polykaryons and expressed the VZV glycoprotein gE on plasma membranes (Figure 5B); their mean diameter was 1.5 µm±0.3 SD (N = 81). Importantly, as observed in VZV infected human DRG cells, the ORF23 capsid protein was enriched in many of these large PML-NBs in infected skin cells (Figure 5C and D).
These experiments established that endogenous PML formed spherical nuclear cages containing VZV NCs not only in cultured cells in vitro, but also in differentiated human cells infected in vivo. Since neural and skin cells are targeted during VZV pathogenesis, we hypothesized that PML cages might play a role in the intrinsic host defense against VZV. These findings provided a rationale to further investigate the molecular mechanisms of this cellular response and to assess whether it could interfere with production of infectious virus progeny.
PML protein exists in several isoforms that have unique C-terminal domains resulting from alternative splicing of the PML gene [14]. To define the process of NC sequestration in PML cages during VZV infection and to ask if this process might function as an intrinsic antiviral host defense, we evaluated six major PML isoforms, PML I, II, III, IV, V and VI. These experiments investigated whether exogenous expression of a particular PML isoform specifically altered the intranuclear distribution of ORF23 protein and sequestered VZV capsids. PML-NBs were formed by the exogenously expressed isoforms over the background of endogenous PML in uninfected and infected cells (Figure S5). In control cells that were not transfected but were infected with VZV, the ORF23 capsid protein showed no redistribution if endogenous PML-NBs were completely dispersed as happens in the majority of infected melanoma cells (Figure 6A, upper panel; see also Figure S2A). When melanoma cells were transfected with the PML isoforms and infected with VZV, ORF23 protein was redistributed and colocalized with more than 95% (N = 550) of PML-NBs only in cells that expressed PML IV or EGFP-PML IV (Figure 6A). None of the other five PML isoforms recruited ORF23 protein to PML-NBs (Figure 6B and Figure S5).
Photo-bleaching experiments with cells that expressed EGFP-tagged PML IV and that were infected with recombinant VZV expressing RFP-tagged ORF23 protein, revealed the striking immobilization of ORF23 protein within EGFP-PML IV-NBs. Even 45 min after photo-bleaching, RFP-ORF23 remained confined to PML IV-NBs (Figure S6), indicating that PML IV may form a physical barrier that constrains the mobility of ORF23 capsid protein and possibly of assembled capsids in the nucleoplasm. Therefore we next analyzed VZV infected cells expressing EGFP-tagged PML IV protein by immuno-EM to determine whether VZV capsids were confined in PML cages (Figure 6C). A quantitative analysis of 100 infected cell nuclei showed that more than 90% of VZV NCs (N = 4,900) were sequestered (Figure 6C) whereas other nuclear areas were essentially devoid of NCs, demonstrating a highly efficient retention of NCs in PML IV-NBs.
Furthermore, like endogenous PML cages, the exogenous PML IV cages (PML IV cages that formed in cells with endogenous PML and over-expressed PML IV) in VZV infected cells did not colocalize with the viral DNA replication compartments identified by IE62 and ORF29 protein expression or in situ hybridization for viral genomic DNA (Figure S7A–C). These experiments also demonstrated that ORF23 protein recruitment was specific, as compared to IE62 and ORF29 proteins, which showed no redistribution by PML IV.
Having shown that PML IV reorganized the nuclear distribution of ORF23 protein and sequestered NCs into large PML-NBs during VZV infection, we next investigated whether ORF23 protein was a molecular target of PML IV. ORF23 protein was considered a likely candidate for recognition by PML IV because, like the related HSV-1 VP26 small capsid protein [64], [65], ORF23 protein may decorate the capsid surface on hexons formed by the major capsid protein and is therefore likely to be accessible on NC surfaces. This prediction was supported by the dense ORF23-specific gold-labeling observed at the outer edges of VZV NCs (Figure 2B).
When expressed in melanoma cells, untagged ORF23 protein or ORF23 protein tagged with maltose-binding protein (MBP) protein was found to be distributed diffusely in both the cytoplasm and nucleus (Figure 7A). However, when cells were transfected with plasmids expressing PML IV, both ORF23 protein and MBP-tagged ORF23 protein were highly enriched within PML IV-NBs (Figure 7B). ORF4 protein, which is a VZV regulatory/tegument protein [71], [72], was used as an MBP-tagged control; it was not recognized by PML IV and did not colocalize with PML IV-NBs (Figure 7B, lower panels). The specific interaction of PML IV with MBP-ORF23 protein but not with MBP-ORF4 protein was confirmed by coimmunoprecipitation (Figure 7C). These data showed that ORF23 protein was recruited to PML IV-NBs in the absence of other VZV proteins. Therefore ORF23 protein is a potential molecular target on NC surfaces that may physically interact with PML IV during infection.
The unique C-terminal domain of PML IV, which is encoded by exons 8a and 8b of the PML gene, differentiates this isoform from the other five PML isoforms that failed to sequester NCs [14]. We therefore asked whether expression of truncated PML IV protein that had a deletion of exon 8b (PML IV-Δ8B) or of both exons 8a and 8b (PML IV-Δ8AB) would eliminate the redistribution of ORF23 protein or capsids to PML cages. EGFP-tagged PML IV-Δ8B or PML IV-Δ8AB continued to form PML-NBs when exogenously expressed in VZV infected cells (Figure 8A). However ORF23 protein no longer colocalized with these mutant PML-NBs (Figure 8A and B), suggesting that the truncated PML IV proteins could not promote NC sequestration. PML I and the truncated PML I-Δ9 protein, lacking the unique PML I C-terminus encoded by exon 9, were included as controls and also failed to cause redistribution of ORF23 protein (Figure 8B). We next investigated whether these results could be reproduced using stable melanoma cells lines that avoided EGFP-tagging and expressed untagged PML IV or PML IV-Δ8AB protein under the control of a doxycycline-inducible promoter (Figure 8C and Figure S8). These recombinant cell lines expressed similar amounts of PML IV and PML IV-Δ8AB protein at 24 hr after induction with 5 µg/ml doxycycline and formed PML-NBs that were similar in size (Figure S8A and B) and fewer than 0.5% cells exhibited apoptosis based on Annexin V staining (Figure S8C). When the induced cell lines were infected with VZV, ORF23 protein colocalized completely with PML IV at 24 hr after infection whereas ORF23 protein was not redistributed to PML-NBs in cells that expressed the truncated PML IV-Δ8AB (Figure 8C), again demonstrating that PML-NBs composed of truncated PML IV were deficient in the sequestration of ORF23 protein, presumed to be on VZV capsids.
To test this assumption, quantitative immuno-EM was used to show patterns of NC distribution in cells induced to express PML IV or PML IV-Δ8AB. For this purpose, samples were processed with high-pressure freezing and freeze-substitution to optimally preserve the ultrastructure of PML cages and virion capsids and to achieve sensitive detection of PML protein by immunogold labeling. Large spherical PML cages that harbored VZV NCs were readily observed in cells expressing PML IV (Figure 9A, left panel, and Figure 9B) and all PML IV cages entrapped both mature and immature NCs (Figure 9A, left panel; Figure 9B and Figure S9A–C). In contrast, PML IV-Δ8AB-NBs rarely sequestered any NCs even though numerous NCs were present in the infected cell nucleus (Figure 9A; right panel). Infected nuclei without any PML cages usually contained numerous randomly scattered NCs (Figure S9A; compare left and right panels).
The mean size of PML cages in infected cells induced to express PML IV was 1.6 µm ±0.5 SD (N = 100), which was similar to the sizes of PML cages in neural cells (1.9 µm ±0.8 SD; N = 62) and skin cells (1.5 µm ±0.3 SD; N = 81) in human tissue xenografts that were infected with VZV (Figures 4 and 5). However, PML IV-Δ8AB structures were significantly smaller than PML IV-NBs (1.2 µm ±0.4 SD vs. 1.6 µm ±0.5 SD; p<0.0001; N = 100) and had a more compact appearance than PML IV-NBs, presumably because fewer capsids were sequestered in these mutant PML-NBs. However, large PML IV cages that were only partly filled with sequestered NCs were also readily observed (Figure 9B) indicating that size and architecture of PML IV cages is not only determined by the sequestered cargo, but also by the assembly properties of PML-positive fibers itself. Thus, truncation of the unique C-terminal domain of PML IV influenced sequestration of NCs and the overall architecture of PML-NBs.
Analysis of PML IV cages by TEM at high magnification revealed a shell structure consisting of PML-positive fibers that appeared to form a physical barrier entrapping VZV capsids (Figure 9B); NCs were often aligned along the inner side of the fibrous shell. PML IV cages were morphologically similar to the endogenous PML cages found in VZV infected HELF in vitro (Figure 2 and Figure 3) and in neural cells in vivo (Figure 4E). Importantly, the thin PML-positive fibers exhibited rather weak electron density (Figure 10B) and could only be identified unequivocally by high magnification and after PML-specific immunogold-labeling (Figure S9B and C). Therefore, despite their size, these PML structures would be easily be missed in virus-infected cells analyzed by morphological criteria only.
Quantitative analysis of NC distribution confirmed that PML IV cages in infected inducible cell lines were highly efficient in sequestration with more than 95% of NCs (N = 5,938) being retained in these structures (Figure 9C). In marked contrast, only 15% of NCs (N = 4,395) were found within PML-NBs in cell lines induced to express PML IV-Δ8AB (Figure 9C). The density of the packing of NCs within individual PML cages was also significantly higher in PML IV-expressing cells than in cells expressing PML IV-Δ8AB, suggesting that PML IV cages had more capacity to sequester NCs than those formed by truncated PML IV (Figure 9D). Furthermore, the total number of NCs was significantly higher (approximately 20%) in cells with PML IV cages compared to those with PML IV-Δ8AB-NBs (p<0.0001), indicating that NCs accumulated in PML IV-NBs (Figure 9E).
Herpesvirus capsids may be abortive (A type), intermediate (B type) or mature (C type) containing viral DNA genomes [73]. To determine whether NC assembly or packaging was effected by PML IV and if NC sequestration by PML IV cages was selective for a specific type of capsids, the relative proportion of the three major capsid types was quantified in nuclei expressing PML IV or PML IV-Δ8AB-NBs, as well as in PML IV cages (Figure 9F). Analysis of 5,938 NCs in 100 nuclei of cells expressing PML IV, of 4,395 NCs in 100 PML IV-Δ8AB expressing cells and of 5,829 NCs within PML IV cages, respectively, revealed similar proportions of all NC types; 16–20% of all NCs were mature, 76–79% were intermediate and 4–5% were abortive in each cell line and in PML cages (Figure 9F). This finding suggests that expression of PML IV did not increase the proportion of abortive or intermediate NCs compared to PML IV-Δ8AB expressing cells and that PML IV cages may participate in the antiviral host cell response against VZV by targeting all three types of NCs that are made in the infected cell nucleus for entrapment.
Since PML IV restricted NCs to PML cages, we hypothesized that PML IV would impair the production of infectious virus. Conversely, we predicted that infectious virus yields would be unaffected in cells expressing the truncated PML IV-Δ8AB mutant, which does not recruit VZV capsids (Figures 8 and 9). Since VZV is highly cell-associated in vitro [74], viral replication is measured by plaque assays using infected cells as the inoculum. Given the critical importance of cell-cell spread for VZV infection, our objective was to assess the infectivity of VZV infected cells with and without PML-IV cages based on the transfer of virus from these cell populations to a permissive cell monolayer. We focused our analysis on the 24 hr time point after VZV infection of PML IV and PML IV mutant-expressing cells in order to avoid any potential effect of PML expression on the induction of apoptosis which could also reduce viral titers and because 24 hr is a late stage of VZV replication in vitro when NCs are present. We used two plaque assay methods to assess the antiviral activity of PML IV. In the first method, cells were transfected with an EGFP control, EGFP-tagged PML IV or PML IV-Δ8AB constructs, infected with recombinant VZV expressing RFP-tagged ORF23 protein for 24 hr and separated into subpopulations by fluorescent activated cell sorting (FACS). Only cells that emitted both green (EGFP, indicating transfection) and red (RFP, indicating infection) fluorescence were recovered and equal numbers of infected cells that expressed EGFP, EGFP-PML IV or EGFP-PML IV-Δ8AB were added to permissive cell monolayers and observed for plaque formation (Figure 10A–C). Plaques generated from inoculum cells that expressed PML IV were significantly reduced (p<0.0001) in five independent experiments; plaque numbers were determined in triplicate in each independent experiment. Plaque numbers were about 50% lower when compared to infected cells expressing PML IV-Δ8AB or the EGFP control (Figure 10A). Furthermore, the mean diameter of plaques that were created by infected cells expressing PML IV (0.95 mm2 ±0.06 SEM; N = 70) was significantly smaller than plaques produced by PML IV-Δ8AB expressing cells (1.46 mm2 ±0.09 SEM; N = 70) (p<0.0001). In contrast, the number and size of plaques from cells expressing PML IV-Δ8AB was very similar to the EGFP control (Figure 10A).
In the second method, the antiviral activity of PML IV was tested using the stable cell lines expressing inducible PML IV or PML IV-Δ8AB. These conditions avoided variables that might be associated with EGFP tagging of PML or infection with the recombinant virus in which ORF23 capsid protein was expressed as an RFP fusion protein. Inoculum cells that expressed induced PML IV and that were infected for 24 hr were added to permissive cell monolayers and observed for plaque formation (Figure 10B). Plaques generated from inoculum cells that expressed induced PML IV were significantly reduced (p<0.0018, N = 4) by about 50% when compared to infected cells expressing PML IV-Δ8AB or to control cells.
Thus, PML IV exhibited antiviral activity against VZV in transfected cells as well as in an inducible cell line and its antiviral activity was confirmed to require the unique PML IV C-terminal domain. The N-terminal region, which contains the RBCC/TRIM domain and is common to all PML isoforms and is present in the truncated PML IV mutant, was not sufficient for antiviral activity against VZV. Impaired virus production in PML IV expressing cells when compared to PML IV-Δ8AB and control cells correlated with the capacity of PML IV cages to sequester NCs as observed by immuno-EM. Conversely, the failure of the truncated PML IV-Δ8AB protein to exhibit antiviral activity correlated with impaired nuclear retention of NCs in VZV infected cells. These data together with the quantitative EM data that showed that more than 90% of NCs were present within PML cages at 48 hr suggest that NC are stably and not transiently contained in the PML IV cages.
Ring-shaped PML-NBs similar to the endogenous PML cages in VZV infected neurons and satellite cells in human DRG and to PML IV cages in vitro have been observed in neural cell nuclei in neurodegenerative disorders associated with expanded CAG repeats encoding polyglutamine (polyQ) tracts in the abnormal gene product [41], [43]. Of particular interest, PML IV, which promotes VZV capsid sequestration has been reported to retain and degrade mutant polyQ-expanded Ataxin7 and the Huntington's disease protein (huntingtin or Htt) within ring-shaped PML clastosomes in cortical neurons [44]. We therefore asked whether the PML IV cages formed in our inducible melanoma cell lines were functionally like PML clastosomes as determined by their capacity to sequester GFP-tagged Htt [75]. We used the HttQ72 construct, which expresses the aberrant Htt protein with a 72 amino acid polyQ tract [75]. When PML IV cells were transfected without induction, HttQ72-GFP was primarily distributed diffusely in the cytoplasm (Figure 11A); it was also detectable in nuclei at higher levels of expression. This pattern changed dramatically when the cells were induced to make PML IV; almost all HttQ72-GFP was redistributed into nuclear PML IV-NBs (Figure 11B). This data confirmed observations about polyQ protein interactions with PML IV [44] and showed that PML IV-NBs in melanoma cells can sequester polyQ-tract proteins, as occurs in neuronal PML clastosomes.
These experiments led us to consider whether PML IV cages could co-sequester the ORF23 capsid protein marker, indicating retention of VZV capsids, together with the mutant Huntington's disease protein. VZV infection of cells that were transfected with HttQ72-GFP but not induced showed no change in the diffuse cytoplasmic pattern of HttQ72-GFP suggesting that VZV infection was not sufficient to redistribute this polyQ-protein (Figure 11C). However, when PML IV expression was induced, HttQ72-GFP was dramatically reorganized and colocalized in PML-NBs together with the ORF23 capsid protein in infected cells (Figure 11D). To our knowledge, these experiments demonstrated for the first time that the same PML isoform has the capacity to sequester both the aberrant poly-Q protein produced in a neurodegenerative disease and a viral capsid protein associated with NC entrapment in PML-NBs and that individual PML cages can simultaneously target and sequester an aberrant polyQ-protein and viral NCs.
Diverse types of PML-NBs that are distinguishable by their size, shape and dynamic behavior are found in mammalian cell nuclei [4], [5], [11], [18]. We identify a distinctive class of endogenous PML-NBs that form large spherical PML cages in differentiated human neural and skin cells infected with VZV in vivo. These endogenous PML cages were found to sequester newly assembled mature and immature viral nucleocapsids in the nuclei of infected neurons, satellite cells and epidermal cells, and in cultured cells infected with VZV in vitro.
Similar large PML cages that retained VZV NCs were formed when the PML IV isoform was expressed in cultured cells in vitro but not by PML isoforms I, II, III, V or VI. Importantly, PML IV exhibited antiviral activity against VZV. The antiviral function of PML IV was associated not only with a unique capacity to sequester NCs but also to bind the ORF23 capsid surface protein. All of these functions required the unique C-terminal domain of PML IV. The N-terminal region, which is common to all PML isoforms and contains the RBCC/TRIM domain [14] was not sufficient for antiviral activity or NC retention. The antiviral activity of PML IV was not likely to be explained by inhibition of viral genome synthesis or packaging into capsids by PML IV since the relative proportion of empty capsids was not increased in PML IV expressing cells. PML IV-Δ8AB may lack the capacity to produce PML-NBs with the architecture of those formed by intact PML IV, rendering them ineffective for capsid entrapment. It is also possible that the failure of PML-NB in PML IV-Δ8AB expressing cells to entrap virions reflects in part a dominant negative effect of PML IV-Δ8AB, preventing any endogenous PML IV from interacting with the ORF23 capsid protein.
The correlation between PML-mediated antiviral activity, capsid sequestration and ORF23 protein binding provides strong evidence that PML cages constitute an intrinsic host defense against this common herpesvirus. This role is supported by the finding that human neural and skin cells infected with VZV in vivo contained PML-NBs with the same distinctive morphology. ORF23 capsid protein was recruited to these structures and they consisted of cages formed by concentric rings of PML fibers sequestering NCs. Notably, while VZV infection promoted the dispersal of most PML-NBs, PML protein itself was not degraded, indicating that PML nuclear cages, rather than free PML protein, may be the primary obstacle to VZV replication.
How do NCs become entrapped within PML cages? Our working hypothesis is that individual NCs are recognized by PML protein (PML oligomers or fibers) through interactions with the capsid surface. These initial interactions may result in recruitment of more PML and NCs, thus growing into larger PML-NC assemblies. Alternatively, although they were rare in our immuno-EM studies, some aggregates of NCs may form spontaneously, which are then recognized by PML and enclosed in PML cages. Pre-existing, non-disrupted PML-NBs may also constitute a favorable environment (a platform) for NC aggregation, as has been proposed to account for the accumulation of aberrant cellular proteins in PML-NBs in neurons [45]. These possibilities are not mutually exclusive and a combination of all three may occur in the infected cell nucleus when capsids are produced.
Most PML cages observed in VZV infected neural and skin cells were significantly larger than the small PML-NBs that associate with viral genomes and exert antiviral effects by interfering with the initial transcription of herpesviral genes [76], [77]. PML cages were also more prominent later in infection, at the stage of viral capsid assembly. A significant antiviral effect was predicted because sequestration of abortive, immature and mature virions into fibrous PML cages was highly efficient, leaving very few mature capsids free to egress across the nuclear membrane to form infectious particles in the cytoplasm. Thus, PML cages conferred antiviral activity at a later stage of infection by a mechanism completely different from PML suppression of early viral gene transcription. These two PML-mediated antiviral mechanisms are not mutually exclusive and would be expected to function synergistically. The absence of PML cages may contribute to the increase in VZV replication observed when PML protein was depleted by RNA interference [31].
Our experiments provide the first evidence that a specific PML isoform, PML IV, contributes to the cellular antiviral response by promoting nuclear immobilization of herpesvirus nucleocapsids. PML IV interacted specifically with the ORF23 capsid protein in the absence of other VZV proteins. Since ORF23 protein was expressed on NC surfaces, it offers a likely target for PML IV recognition, as the first step leading to the retention of assembled NCs in PML cages. PML IV, as well as PML III and PML V has been reported to be expressed less abundantly than PML isoforms I/II in some cell lines and primary cells in vitro [17]. However, the pattern of PML isoform expression in vivo in human neural and skin cells for which VZV exhibits tropism or when these cells become infected with VZV is not known. We propose that PML IV may act as a potent modulator of PML-NB architecture, since others have also observed that exogenous PML IV (expressed along with endogenous PML) promotes the formation of large spherical PML-NBs [44]. As noted, PML cages that formed upon overexpression of PML IV together with endogenous PML in vitro best mimicked (both structurally and functionally) those endogenous large PML cages observed in neural and skin cells infected with VZV in vivo and both sequester NCs with ORF23 capsid protein. If PML IV is not abundant in the small PML-NBs of uninfected neurons, satellite cells or epidermal cells, the ratio of PML isoforms may change upon infection, PML IV may become more abundant or specific modifications on PML isoforms may be altered when these differentiated cells are infected or otherwise stressed. It was apparent that PML-NBs exhibited a striking reorganization from multiple small compact structures to a few large ring-shaped cages upon VZV infection in human tissues, which could be consistent with changes in the ratio of PML isoforms and upregulation of PML IV. Of particular interest, PML IV transcription was upregulated when cultured cells were stimulated with type I IFN [44], neurons have been shown to produce type I IFN during viral infection [78] and the IFN pathway is upregulated dramatically in human tissues infected with VZV in vivo [39]. Taken together, our findings point to the need to better understand the patterns and regulation of the tissue specific expression of PML isoforms and their functions in virus-infected cells. For example, different PML isoforms might recognize the outer capsid proteins of other herpesviruses or nuclear replicating DNA viruses and retain NCs, causing antiviral effects when a specific PML isoform is either not degraded efficiently or its production is induced in the infected cell.
Since the intranuclear mobility of herpesvirus NCs via actin filaments may facilitate formation of NC assembly domains and NC transport to the nuclear membrane [79], [80], we suggest that PML-mediated entrapment of NCs is likely to prevent NC interactions with the viral export machinery at the inner nuclear membrane. This hypothesis is supported by our observation that the number of intranuclear NCs was significantly higher in cells with functional PML IV cages as compared to control cells.
An intrinsic defense mechanism in which PML binds to a viral capsid protein and sequesters NCs obviously depends on the persistence or new synthesis of PML protein in the infected cell nucleus at the stage of capsid assembly, which occurs in VZV infected cells but not in HSV infected cells. Since HSV-1 degrades PML protein shortly after virus entry and shuts off host cell protein synthesis [25]–[27], both the early effects of PML on viral gene expression and this later antiviral activity would be expected to be counteracted. Some PML-specific gold labeling associated with NC-like structures, although not in PML cages, has been reported in HSV-1 infected cells at late times [60]. Overexpression of PML as a 69 kDa protein did not inhibit HSV-1 infection and PML cages were not observed but the PML isoform was not unequivocally identified [81]. Of interest, other nuclear replicating DNA viruses, including papillomavirus and JC (polymoma) virus also have limited capacity to degrade PML protein and PML binding to their capsid proteins has been reported [57], [59]. These interactions were proposed to support infection by directing the viral genome to replication compartments or facilitating virion assembly but enhancing or inhibitory effects of specific PML isoforms on viral replication could not be assessed [82], [83]. Nevertheless, particular PML isoforms could have pro-viral functions through some PML-capsid protein interactions and antiviral consequences in others, as we observed here in VZV infected cells.
Interestingly, stimulation of PML expression by type-I interferon and PML-mediated sequestration and immobilization NCs of VZV, a DNA virus, has parallels with the antiviral mechanism of IFN-induced Mx-GTPases against RNA viruses. These proteins also form oligomeric structures and inhibit the replication and assembly of several negative-strand RNA viruses by sequestering and immobilizing viral ribonucleocapsids or nucleocapsid protein [84]. For example, the human MxA GTPase was found to sequester the LaCrosse bunyavirus nucleocapsid protein into fibrillar cytoplasmic complexes thereby inhibiting viral RNA genome replication and the formation of enveloped virion particles on membranes in the Golgi compartment [85], [86]. Thus, these two IFN-induced antiviral responses appear to depend on creating intracellular structures that act as physical barriers during the later viral assembly stage of infection.
Our observations of VZV NC sequestration in PML-NBs in neurons, satellite cells and skin cells in vivo are relevant to a better understanding of the control of VZV infection in the human host. We suggest that the role of NC entrapment by PML cages in skin is important for the modulation of the infection by the innate, interferon-induced response through restriction of cell-cell spread. Diverting a substantial majority of mature capsids to PML cages is likely to regulate lytic VZV infection in vivo. In fact, this process may benefit the virus since it is important for VZV and other herpesviruses to achieve a well-balanced virus-host interaction. Persistence of the virus in the human population requires a relatively mild infection. An incapacitating infection of the host that would limit opportunities for VZV transmission to susceptible contacts and would hinder, rather than support the persistence of the virus in the population as a whole. In addition to limiting skin infection, the consistent finding of large PML cages in neuronal cells of human DRG xenografts infected with VZV suggests the possibility that this intrinsic host cell response might restrict episodes of VZV reactivation from latency so that the episode remains subclinical [33], [87]. We hypothesize that when VZV reactivates from latency in persistently infected neurons, trapping newly formed NCs by PML cages would reduce the possibility for spread of infectious virus particles down the axons to skin and would also limit infection of adjacent satellite cells and neurons in the DRG where reactivation is occurring. This process would result in abortive VZV reactivation [88]. From a clinical perspective, it is notable that patients treated with arsenicals have an extremely high rate of clinically symptomatic herpes zoster, due to VZV reactivation [89], [90] and arsenicals are potent triggers of PML degradation [91], [92].
Since VZV is a human neurotropic virus, it is of interest that ring-like PML-NBs similar in size to endogenous PML cages in VZV infected neurons and satellite cells in DRG in vivo and PML IV cages in vitro are found in neural cell nuclei in several neurodegenerative disorders [45]. Many of these disorders are associated with expanded CAG repeats encoding polyglutamine (polyQ) tracts, including Huntington's disease and several types of spinocerebellar ataxia and PML was associated with these protein inclusions [40]–[43]. Importantly, the same PML isoform, PML IV, which promotes VZV capsid sequestration has been reported by others to form ring-like PML clastosomes that mediate the sequestration and degradation of mutant polyQ-expanded Ataxin7 and Huntington's disease protein (huntingtin or Htt) in cortical neurons [44]. The morphological similarity between PML cages found in VZV infected cells and these ring-like PML clastosomes and the finding that PML IV can co-sequester ORF23 capsid protein together with the Huntington's disease protein in VZV infected cells, suggest that the PML cages described here and PML clastosomes are structurally and functionally related subnuclear domains. Neurons are post-mitotic cells and need to safely contain toxic proteins and/or increase their clearance, whether aberrant host cell proteins or virion structures [46], [50]. Of interest, aberrant polyQ proteins tend to form large aggregates that like NCs may be particularly resistant to the celluar mechanisms for degradation of unwanted proteins and the nucleus may have limited capacity to cope with misfolded proteins compared to the cytoplasm [75]. PML IV may specialize in restricting these degradation-resistant structures to nuclear subdomains and limit their toxicity longer term.
In summary, the efficient sequestration of virion capsids in PML cages that contributes to the intrinsic antiviral host response against VZV appears to be the outcome of a basic cytoprotective function of this distinctive category of PML-NBs in safely containing aggregation-prone aberrant proteins. Both aggregation-prone proteins and virion capsids are therefore potential substrates for sequestration in these PML cages.
Human fetal tissues for SCID xenograft studies were obtained from Advanced Bioscience Resources, Inc. (Alameda, CA) in compliance with state and federal regulations for tissue acquisition for biomedical research, in accordance with FDA 21 CFR Part 1271 GTP (Good Tissue Practices), UAGA and NOTA. Animal protocols complied with the Animal Welfare Act and were approved by the Stanford University Administrative Panel on Laboratory Animal Care.
Human lung embryonic fibroblasts (HELF) and a human melanoma cell line (MeWo, ATCC number: HTB-65) were grown in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum, nonessential amino acids (100 µM) and antibiotics (penicillin at 100 U/ml and streptomycin at 100 µg/ml). HELF and the melanoma cell line (as obtained from ATCC) were passaged less than 25 times.
VZV used in these experiments was recombinant Oka (rOka) derived from the wild type low passage parent Oka strain (pOka). VZV (rOka) and the recombinant rOka-RFP-ORF23 expressing RFP-tagged ORF23 [54] were propagated in melanoma cells and HELF cells [32]. Transfection was done with Lipofectamin 2000 (Invitrogen) for 14 hr. Viral infection was done with cell-associated VZV at a ratio of 1/20 (infected cells/uninfected cells) for 24 or 48 hr.
Plasmids pEGFP-PML (I-VI) [18] were gifts from Prof. Peter Hemmerich, Leibnitz-Institute of Age Research, Jena, Germany. Plasmid pcDNA3-PML IV [44] was generously provided by Prof. Annie Sittler, Universite Pierre et Marie Curie, Paris, France. Plasmids pCR3-MBP-ORF4, pCR3-MBP-ORF23 and pENTR207-ORF23 vectors have been described [93]. The plasmid pcDNA3-HttQ72-GFP [75] was a kind gift from Prof. Ron Kopito, Stanford University, Stanford, USA. Plasmid pCR3-ORF23 that expresses untagged ORF23 was constructed using the pENTR207-entry vector and the pCR3 destination vector with the LR Clonase II gateway cloning reaction (Invitrogen). PML deletion constructs EGFP-PML I-Δ9, EGFP-PML IV-Δ8AB and EGFP-PML IV-Δ8B were made as follows: exon 9 was removed from PML I by PCR (primers PML IΔ9 5′A/PML IΔ9 5′B and PML IΔ9 3′A/PML IΔ9 3′B). pEGFP-CI PML IV was the template used to delete exon 8b (primers PML IVΔ8B 5′A/PML IΔ9 5′B and PML IV 3′A/PML IV 3′B) and exons 8ab (primers PML IVΔ8AB 5′A/PML IVΔ8AB 5′B and PML IV 3′A/PML IV 3′B). pEGFP-C1 PML I was digested with MluI and a 6.1 kb vector DNA was isolated. pEGFP-C1 PML IV was digested with BbvCI/MluI and a 5.6 kb vector DNA was isolated. Purified, digested PCR products were ligated into the appropriate vectors to generate pEGFP-PML I-Δ9, pEGFP-PML IV-Δ8AB and pEGFP-PML IV-Δ8B. Primer sequences are listed in Table S1.
Melanoma cells expressing doxycyline-inducible PML IV or PML IV-Δ8AB were made using the pRetro-X-Tet-On-Advanced vector system and pRetro-X-Tight-Pur plasmid (Clontech Laboratories) with the PML IV or PML IV-Δ8AB sequence derived from the plasmid pcDNA3-PML IV [44]. The stable cells were induced with 5 µg/ml doxycyline for 24 hr before infection with VZV.
Human fetal DRG were inserted under the kidney capsule of male CB-17scid/scid mice (Taconic Farms, Germantown, NY), inoculated with rOka-infected HELF and harvested after 14 days [36]. Human fetal skin xenografts were inoculated with VZV rOka-infected HELF and harvested after 21 days [34]. Human tissues were provided by Advanced Bioscience Resources (ABR, Alameda, CA) and were obtained in accordance with state and federal regulations. Animal protocols complied with the Animal Welfare Act and were approved by the Stanford University Administrative Panel on Laboratory Animal Care.
DRG xenografts were prepared for cryosectioning as described previously [88]. Human skin xenografts were embedded in paraffin and 5 µm sections were prepared. For confocal microscopy, sections were deparaffinized with tissue clearing agent (“Safeclear II”, Fisher Scientific), rehydrated and antigen retrieval was performed in citric acid-based antigen retrieval solution for 90 sec at 100°C. Sections were then blocked and immunolabeled as described previously [88]. Cultured cells on glass coverslips were fixed in 4% paraformaldehyde in PBS for 20 min at room temperature or were fixed with methanol at -20°C for 10 min and air-dried. Cells were blocked and immunostained as described previously [32].
Antibodies used for confocal microscopy were: rabbit polyclonal anti-PML and mouse monoclonal anti-PML (PG-M3), rabbit polyclonal anti-Sp100 (all Santa Cruz Biotech), mouse monoclonal anti-N-CAM (Zymed), rabbit polyclonal anti-VZV-ORF23 [54], rabbit polyclonal anti-VZV-IE62 and rabbit polyclonal anti-VZV-ORF29 [32]. Antibodies for secondary detection were Alexa Fluor 488, Alexa Fluor 594 or Alexa Fluor 647 conjugated donkey anti-mouse or donkey anti-rabbit antibodies (Invitrogen). Infected tissues, cultured cells and PML-NBs were imaged and quantified using a Leica TCSSP2 confocal laser scanning microscope (Heidelberg, Germany). Microscope objectives were 40/1.0 (Numerical Aperture, N.A.) or a 63/1.4 (N.A.) plan apochromate objectives. Images were scanned at 1024×1024 pixels with at least four times frame averaging and the pinhole adjusted to one airy unit. Brightness and contrast were adjusted using Photoshop CS3 (Adobe) or iPhoto (Apple). The number of PML-NBs and colocalization of PML-NBs with ORF23 protein were quantified using captured digital images and the analysis and measurement tools in Photoshop CS3 extended version (Adobe).
Samples were either prepared for cryosectioning or high-pressure frozen (HPF) and freeze substituted (FS) and then embedded in either LR-white or Epoxy resin (Embed812). Preparation and labeling of cryosections was done as described previously [88]. Cells were high-pressure frozen in a Leica EM PACT2. Frozen specimen carriers with cells were placed into frozen cryovials containing acetone with 0.1% glutaraldehyde and 0.1% uranyl acetate (for LR White embedding) or in acetone with 1% osmium tetroxide and 0.1% uranyl acetate (for Epon embedding). The frozen vials were placed into a Leica AFS for the freeze-substitution procedure and then embedded in either epoxy resin (Embed 812) or LR-White resin. Ultrathin (50–70 nm) LR-white or Epon-sections were prepared with a diamond knife (Diatome) using a ultramicrotome (Ultracut, Leica). For immunogold-labeling of Epon-sections, these sections were first etched with 5% hydrogen peroxide for 5 min and then rinsed with distilled water. Sections were pre-blocked in DIG-blocking solution (Roche) for 30 min. Primary antibodies and Protein A-gold particles were diluted in blocking solution and sections were incubated for 1 hr or 30 min, respectively, at RT. Rabbit polyclonal anti-PML antibody (Santa Cruz Biotech) and rabbit polyclonal anti-ORF23 antibody were used at 1∶10 dilution. Finally the sections were stained with 3.5% aqueous uranylacetate for five minutes and with 0.2% lead citrate for three minutes and air-dried. Sections were analyzed using a JEOL 1230 transmission electron microscope (TEM) at 80 kV and digital photographs were taken with a GATAN Multiscan 701 digital camera. Quantification of PML-specific immunogold-labeling and evaluation and counting of virion capsids was done by analyzing captured digital images with Photoshop CS3 (extended edition, Adobe) using the counting tool and the measuring tool to calculate areas of nuclear profiles and PML compartments. PML-labeling density (number of gold particles per µm2) and capsid density (number of capsids per µm2) were calculated with Microsoft Excel (2008). Statistical analysis was performed using GraphPad Prism (version 5.0) statistical software.
For plaque assays after fluorescent activated cell sorting (FACS), cells were transfected with pEGFP, pEGFP-PML IV or pEGFP-PML IV-Δ8AB expression plasmids, infected for 24 hr with rOka-RFP-ORF23 and sorted by FACS (Digital Vantage, BD Bioscience). Cells that were both green (transfected) and red (infected) were recovered, equal numbers were seeded onto melanoma cell monolayers in triplicate and plates were incubated for 72 hr. Plaques were visualized by standard methods; five independent FACS experiments were performed. Plaque numbers were counted and plaque sizes were determined from digital photographs by calculating the area of >70 randomly selected plaques per sample using Photoshop CS3 (extended edition, Adobe). Virus plaque assays with inducible PML expressing cell lines (Mel39-control, Mel39-PML IV and Mel39-PML IV-Δ8ab) were done using cells induced overnight with 5 µg/ml doxycycline, infected for 24 hr, harvested, serial diluted and seeded in triplicate onto melanoma cell monolayers. Plaques were counted and numbers expressed as percentages of the control; four independent experiments were performed.
For immunoprecipitation studies, samples were analyzed using the ExactaCruz IP-Western Kit (Santa Cruz Biotechnology, Inc.), according to manufacturer's instructions. MBP-ORF23 or MBP-ORF4 and PML IV transfected cells were grown in 75 cm2 flasks and lysed in 300 µl high salt buffer (20 mM HEPES [pH 7.2], 450 mM NaCl, 1.5 mM MgCl2, 0.5% NP-40, 20% Glycerol) supplemented with EDTA-free protease inhibitor cocktail (Roche). Cell debris was removed by centrifugation. The supernatant was combined with 1,200 µl NaCl-free low salt buffer (20 mM HEPES [pH 7.2], 1.5 mM MgCl2, 0.5% NP-40, 20% Glycerol) and centrifuged again. 4 µg of anti-MBP monoclonal antibody (New England Biolabs) was coupled to 100 µl sepharose beads. Beads were washed in PBS and then incubated with 1,000 µl pre-cleared cell lysate at 4° C for 4 hr. Beads were then washed six times with the binding buffer (20 mM HEPES [pH 7.2], 90 mM NaCl, 1.5 mM MgCl2, 0.5% NP-40, 20% glycerol), boiled with SDS sample buffer, and analyzed by Western blot using rabbit polyclonal anti-PML antibody (1∶2,000), anti-MBP antibody (1∶2,000). For Western Blot analysis of cultured cell lysates, cells grown in 6-well plates were lysed in 150 µl high salt buffer (see above). Whole lysates were then combined with 150 µl SDS sample buffer and boiled for 5 min before loading on a 7.5% SDS PAGE. MBP-tagged constructs of ORF23 were used in these cotransfection and IP experiments because they yielded higher expression levels and showed less cytotoxicity; the monoclonal anti-MBP-antibody was more effective for IP than the polyclonal ORF23 antibody.
The Vibrant Apoptosis Assay (Invitrogen) with Annexin V conjugated to Alexa Fluor 488 was used to evaluate the percentage of apoptotic cells at 24 hr after doxycycline induction of PML IV and PML IV-Δ8AB expressing cell lines. Approximately 5,500 cells per cell line were evaluated by confocal microscopy.
Student's t tests were performed for all experiments using Graph Pad Prism (version 5.0) statistical software. A p-value of <0.05 was considered statistically significant.
|
10.1371/journal.pntd.0004616 | The Burden of Typhoid and Paratyphoid in India: Systematic Review and Meta-analysis | Typhoid is an important public health challenge for India, especially with the spread of antimicrobial resistance. The decision about whether to introduce a public vaccination programme needs to be based on an understanding of disease burden and the age-groups and geographic areas at risk.
We searched Medline and Web of Science databases for studies reporting the incidence or prevalence of typhoid and paratyphoid fever confirmed by culture and/or serology, conducted in India and published between 1950 and 2015. We used binomial and Poisson mixed-effects meta-regression models to estimate prevalence and incidence from hospital and community studies, and to identify risk-factors.
We identified 791 titles and abstracts, and included 37 studies of typhoid and 18 studies of paratyphoid in the systematic review and meta-analysis. The estimated prevalence of laboratory-confirmed typhoid and paratyphoid among individuals with fever across all hospital studies was 9.7% (95% CI: 5.7–16.0%) and 0.9% (0.5–1.7%) respectively. There was significant heterogeneity among studies (p-values<0.001). Typhoid was more likely to be detected among clinically suspected cases or during outbreaks and showed a significant decline in prevalence over time (odds ratio for each yearly increase in study date was 0.96 (0.92–0.99) in the multivariate meta-regression model). Paratyphoid did not show any trend over time and there was no clear association with risk-factors. Incidence of typhoid and paratyphoid was reported in 3 and 2 community cohort studies respectively (in Kolkata and Delhi, or Kolkata alone). Pooled estimates of incidence were 377 (178–801) and 105 (74–148) per 100,000 person years respectively, with significant heterogeneity between locations for typhoid (p<0.001). Children 2–4 years old had the highest incidence.
Typhoid remains a significant burden in India, particularly among young children, despite apparent declines in prevalence. Infant immunisation with newly-licensed conjugate vaccines could address this challenge.
| Typhoid fever is an important cause of avoidable mortality in regions without adequate access to safe water and sanitation. Highly immunogenic typhoid conjugate vaccines are now licensed and under consideration as a public health intervention in India. The decision about whether and how to introduce a public vaccination programme needs to be based on an understanding of disease burden, and the age-groups and geographic areas at risk. We performed a systematic review and meta-analysis of published studies reporting typhoid and paratyphoid incidence and prevalence in India between 1950 and 2015. The estimated prevalence of laboratory-confirmed typhoid and paratyphoid among individuals with fever across all hospital studies was 9.7% (95% CI: 5.7–16.0%) and 0.9% (0.5–1.7%) respectively, with a significant decline in prevalence of the former over time. We found only three population-based studies that reported incidence. Pooled estimates were 377 (178–801) and 105 (74–148) per 100,000 person years for typhoid and paratyphoid respectively, with incidence being highest in in children between 2 and 4 years. Despite an apparent decline in prevalence, typhoid remains a significant burden in India, particularly among young children. Studies are required to evaluate the effectiveness of infant immunisation with conjugate typhoid vaccines.
| Typhoid (enteric) fever caused by Salmonella enterica serovar Typhi (S. Typhi) is an important cause of morbidity and mortality. The global annual burden was estimated at approximately 12 million cases for 2010 [1,2]. Most of these were effectively treated with antibiotics, although the case fatality rate remains at about 1% such that about 130,000 typhoid deaths occur annually. Antibiotic resistance is a challenge for effective treatment of typhoid and is likely to become increasingly problematic with the spread of multi-drug resistant strains [3]. The situation is further complicated by increased incidence in some countries of S. Paratyphi A as a cause of enteric fever [4]. This serovar is not prevented by currently available typhoid vaccines and represents an increasing threat to human health.
The incidence of typhoid and paratyphoid varies geographically, with south-central and south-east Asia having the highest incidence—typically exceeding 100 cases per 100,000 person-years for typhoid and with lower, variable rates for paratyphoid. In one multicenter study, the annual incidence of typhoid per 100,000 children aged 5–15 years was 180 in North Jakarta, Indonesia, 413 in Karachi, Pakistan and 494 in Kolkata, India [5]. In the same settings, the annual incidence of paratyphoid was considerably lower, with the highest annual incidence reported from Pakistan of 72 per 100,000 children aged 2–16 years [6].
The burden of typhoid fever shows substantial variation within as well as between countries. Commonly identified risk-factors include a lack of clean drinking water, poor sanitation, inadequate hygiene practices and low socio-economic status [2,7]. Outbreaks may occur following a defined event of food or water contamination with the bacterium, in which case locally-specific risk factors or exposures may be identified e.g. eating milk products from a sweet shop, [8]. In some instances the originating infection may be a chronic carrier who persistently sheds the bacterium as a result of infection of the gall bladder. Chronic carriage occurs following primary infection in approximately 2–5% of cases in the absence of antibiotic treatment and is strongly dependent on age and sex [9]. However, the contribution of chronic carriers to typhoid transmission in endemic regions is unknown [10].
Several safe and effective vaccines that could help reduce disease burden are licensed and available in India. Three or four doses of orally-administered, live-attenuated Ty21a provide about 50–70% protection for at least 7 years and are licensed in capsule form from 5 years of age or as a liquid formulation from 2 years of age [11,12,13]. The single-dose injectable Vi polysaccharide vaccine provides similar levels of protection for at least 3 years and is licensed from 2 years of age [11,14,15]. A Vi polysaccharide conjugated to Pseudomonas aeruginosa exotoxin A (rEPA) as a carrier protein and administered to 2–5 year old children gave approximately 90% protective efficacy against typhoid over 4 years [16,17]. More recently, two Vi polysaccharide vaccines conjugated to tetanus toxoid have been licensed in India from 3–6 months of age based on their encouraging immunogenicity [18]. The immunogenicity of conjugate typhoid vaccines in children under 2 years of age (cf. Vi polysaccharide vaccines) is an important advance, given the significant burden of disease in young children and infants [19].
The WHO recommends the programmatic use of typhoid vaccines for controlling endemic disease, although in most countries vaccinating only high risk populations is recommended [20]. In India, routine typhoid vaccination is not implemented and decision-making has been hampered by the lack of reliable disease burden data with very few prospective surveillance studies in the past two decades. The one exception we are aware of is in Delhi where each year approximately 300,000 children aged 2–5 years are vaccinated with Vi polysaccharide vaccine. With the recent development of conjugate vaccines that can be administered to children under 2 years old, the case for more widespread immunization is stronger and in 2014 the Indian Academy of Pediatrics (IAP) Advisory Committee on Vaccines and Immunization Practices (ACVIP) strongly urged the Government of India (GoI) “to include universal typhoid vaccination in its UIP [Universal Immunisation Programme] all over the country.” [21]
The GoI decisions about whether to introduce a public typhoid vaccination program, its extent and the immunization schedule, need to be based on a firm understanding of the disease burden and the age-groups and geographic areas at risk. We therefore carried out a systematic review to estimate the burden of typhoid and paratyphoid in India and to identify knowledge gaps that need further evaluation. We searched for hospital and community-based studies that reported the incidence or prevalence of typhoid and paratyphoid fever and used meta-analysis and meta-regression models to summarize our findings and identify risk factors for disease.
We searched Medline and the Web of Science literature databases for articles published between 1950 and May 2015 for studies on the burden of typhoid or paratyphoid fever in India, with no language restriction. The search consisted of terms related to typhoid or paratyphoid fever (typhoid OR Salmonella Typhi OR enteric fever OR Salmonella enterica OR paratyphoid OR Paratyphi), combined with terms for Indian geography (including a list of state names) and terms for measures of incidence and prevalence (burden OR incidence OR prevalence OR mortality, etc.). The complete search term is given in the S1 Appendix.
Titles and abstracts of articles were read to identify potentially relevant articles. Studies were considered eligible for further examination in full text if they reported incidence, prevalence, number of reported cases, mortality or the burden of typhoid or paratyphoid in India. Studies were also examined in full text if only a title was returned by the initial search. Full text articles were obtained through online publisher websites, the British Library and the Christian Medical College library in India. We excluded papers reporting a small number of cases (n < 10), no information about the number of S. Typhi or Paratyphi infections, no laboratory confirmation of infection (based on culture or serology), no distinction between Salmonella serovars, vaccine trials (unless cluster-randomized with a control arm), unknown geographical areas or outside India, or a review of the literature only. If the typhoid burden was reported multiple times for the same region, study population and time period, the study with the longest follow up time was selected.
Two reviewers (CVA and NCG) independently extracted data from the included studies and entered these into independent Excel databases. Disagreements between the two databases were resolved by consensus among all authors. Year of publication, study design, setting (hospital or community based study), study location, inclusion and exclusion criteria for study participants, start and end date of recruitment, type of samples, laboratory tests, whether the study was an outbreak report, number of participants, number of cases, age distribution of cases and sex of cases were collected. Longitude and latitude information of the study location was obtained from the US National Geospatial Intelligence Agency [22].
The outcome measures of interest were the prevalence of S. Typhi or Paratyphi among patients tested for infection in hospital settings or the incidence of typhoid and paratyphoid fever recorded in community studies. We did not publish our study protocol prior to completing the systematic review.
We calculated the proportion of patients with laboratory confirmed typhoid or paratyphoid fever together with Wald 95% confidence intervals calculated on a logit scale for each study reporting data from hospitals [23]. Pooled estimates of prevalence were obtained by combining studies in a binomial regression model with a normally distributed random effect on the intercept. Heterogeneity between the studies was assessed using a likelihood ratio test (LRT) comparing a saturated mixed-effects model (with dummy variables for the random-effects) with a fixed-effects only model. A stratified analysis was performed due to anticipated heterogeneity between studies, based on the characteristics of the patients included in the studies (either fever described in the publication as suspected typhoid fever or fever where clinical suspicion of typhoid is either not present or not reported (hereafter just termed ‘fever’)). Independent variables potentially associated with the prevalence of typhoid or paratyphoid were included as fixed-effects in univariate and multivariate binomial meta-regression models.
The incidence of typhoid and paratyphoid was calculated per 100,000 person-years of observation for prospective community-based studies and Wald 95% confidence intervals calculated on a log scale. Pooled estimates of incidence were obtained by combining studies in a Poisson regression model with a normally distributed random effect on the intercept and heterogeneity assessed as above. Independent variables potentially associated with the incidence of typhoid or paratyphoid were included as fixed effects in univariate and multivariate Poisson meta-regression models.
Analyses were all performed in the R statistical programming language using the ‘metafor’ package [24,25].
The search strategy initially yielded 1,152 records of which 361 were duplicates (Fig 1). Six hundred and eleven records were excluded after screening the title and abstract. Full text copies were retrieved for 160 of 180 potential relevant records. After excluding non-eligible articles and duplicates, we included 37 studies that reported on typhoid and 18 that reported on paratyphoid. The characteristics of the included studies are given as Table A in the S1 Appendix.
Three studies of typhoid were community cohorts while the other 34 were conducted in hospitals, with all but 3 conducted in urban settings. Among the studies conducted in hospitals, 13 included participants with fever and 21 with suspected typhoid fever while all the community based studies included participants with fever. Thirty of the hospital studies and both the community studies reported culture confirmed typhoid, while four studies reported either a combination of culture and serology or serology alone. Studies reporting typhoid based on serology were included only if the serologic diagnostic criteria was clearly described.
All of the 18 studies reporting the prevalence of paratyphoid included S. Paratyphi A, and one described both S. Paratyphi A and B. Incidence of paratyphoid was described in two community cohort studies that reported from the same location (Kolkata).
The estimated prevalence of S. Typhi detected through culture or serology across all hospital-based studies in the random effects model was 9.7% (95% confidence interval (CI): 5.7–16.0%) (Fig 2). There was significant heterogeneity in prevalence among studies (LRT p<0.001). Prevalence was higher among participants with suspected typhoid fever (estimated prevalence in separate random-effects model was 14.5%, 95% CI: 8.4–23.9%) compared with fever (estimated prevalence 4.9%, 95% CI: 1.9–12%). This was confirmed in the univariate mixed-effects, meta-regression model (Odds Ratio (OR) of laboratory confirmation for suspected typhoid fever compared with fever was 3.34, 95% CI: 1.11–10.1; p = 0.032) (Table 1). In the same analysis, study year (or midpoint for multiannual studies) was significantly associated with the odds of laboratory confirmation of typhoid. The OR was 0.95 (95% CI: 0.92–0.99) for each unit increment in the study year, although this decline is apparent in the forest plot only for studies from 1991 onwards (Fig 2). Typhoid was also more likely to be confirmed for studies that reported during an outbreak, although this was only of borderline significance in the univariate analysis (OR 3.66, 95% CI: 0.95–14.1; p = 0.060). Other study characteristics, including location (urban vs. rural, latitude) and type of laboratory assay (culture, serology or both) were not significantly associated with the odds of confirmation of typhoid. In the multivariate meta-regression model including all covariates, the year of the study and whether it reported during an outbreak were significantly associated with the odds of laboratory confirmation of typhoid (Table 2). The duration of fever among patients eligible for testing and their age distribution were available in only 9 of the 37 including studies and therefore subgroup and meta-regression analysis based on these variables were not carried out.
The estimated prevalence of laboratory confirmed paratyphoid across the hospital-based studies in the random effects model was 0.9% (95% CI: 0.5–1.7%) (Fig 3). There was significant heterogeneity among studies (LRT p<0.001). Prevalence was not significantly different according to whether studies included patients with fever or suspected typhoid fever. In the univariate and multivariate meta-regression models only location (urban vs. rural) was significantly associated with the prevalence of paratyphoid, although this was driven by a single rural study with high prevalence [26] (Table 2). In the multivariate meta-regression model, reporting during a typhoid outbreak was associated with an increased odds of laboratory confirmed paratyphoid of borderline statistical significance (OR 4.16, 95% CI: 0.91–19.0; p = 0.067).
Funnel plots of typhoid and paratyphoid prevalence against study size were strongly suggestive of publication bias, such that studies with high prevalence were more likely to be published (Figs A and B in S1 Appendix).
The incidence of laboratory confirmed typhoid fever varied between the two locations where community cohort studies were carried out, with a more than four times higher incidence in Kalkaji (Delhi) of 976 per 100,000 person-years (95% CI: 736–1250) compared with Kolkata (pooled estimate 235, 95% CI: 203–271) (Fig 4A). Although the former study reported for individuals aged 0–40 years and the latter for all ages (under 2s were excluded in [27]), this does not explain this large difference in incidence, since individuals over 40 years old made up only 24% of the population in 2000 [28]. The pooled incidence across all studies was 377 (95% CI: 178–801) per 100,000 person-years although with significant heterogeneity among studies (LRT p<0.001). It was difficult to compare the age-distribution of typhoid incidence between studies because of differences in reporting of age-categories, although incidence was typically highest in the 2–4 year age-group (Fig 4B).
The incidence of paratyphoid was only reported for two studies in Kolkata that met our inclusion criteria, which gave a pooled estimate of incidence of 105 per 100,000 person years (95% CI: 74–148) for all ages (although [27] only reported from 2 years of age).
Note that the incidence of typhoid and paratyphoid in 2004 in Kolkata was estimated using number of individuals in the relevant study area and age-group at baseline because the number of person-years of observation was not reported [5,29].
There have, surprisingly, been very few epidemiological investigations of the incidence of typhoid in India. The three community cohort studies that we identified in Kolkata and Delhi, the last of which reported data nearly a decade back, suggest a variable incidence of typhoid both over time as well as across regions. The variable burden of typhoid was also apparent in the meta-analysis of hospital-based studies, which showed significant heterogeneity in the reported prevalence of laboratory confirmed typhoid among patients with fever or suspected typhoid fever.
In the meta-regression of hospital studies, testing of patients with suspected typhoid fever or during a typhoid outbreak was more likely to lead to laboratory confirmation of typhoid fever. The meta-regression also revealed a significant decline in laboratory confirmed typhoid among patients with fever or suspected typhoid fever over time, apparent since the early 1990s (Fig 2). The odds of detecting typhoid decreased by approximately 5% each year and this remained significant in the multivariate model accounting for differences in study location, laboratory assay, case definition and whether reporting was during an outbreak. Grouping the studies by decade shows that this significant decline is largely the result of the high prevalence of typhoid in hospital studies during 1980–2000 compared with more recent studies (Table 1). The cause of the high rate of typhoid isolation at this time is not clear. Moreover, inference of a trend from hospital-based studies in different locations and with variable health-seeking behaviours must be tentative. In particular, it is possible that increased use of effective antibiotics before blood collection could have contributed to this decline. The prevalence of typhoid fever was not significantly associated with any other covariates, including study location (latitude, urban vs. rural), although the number of studies in rural areas was small, limiting the power of this analysis.
The incidence of paratyphoid was reported in only two studies that met our inclusion criteria, both in Kolkata. The estimated incidence of paratyphoid in this setting was 105 per 100,000 person years, which compared with 235 per 100,000 person years for typhoid. In Kalkaji (Delhi) the incidence of paratyphoid was not originally reported, although in a companion publication [30] the number of paratyphoid cases recorded during a slightly longer follow-up (18 months) compared with the original study (12 months) [31] was 31 compared with 98 for typhoid over the same period. These estimates suggest an incidence rate for paratyphoid in these settings that is about 30–50% of the rate estimated for typhoid. The lower incidence of paratyphoid was confirmed in the meta-regression of hospital-based studies, which found a prevalence for paratyphoid that was approximately 10-fold lower than for typhoid (estimated pooled prevalence of 0.9% vs. 10.7%). The significantly lower prevalence of paratyphoid among patients tested in hospital may also reflect the shorter duration of fever and more mild clinical characteristics of paratyphoid compared with typhoid [32]. Although significant heterogeneity in the prevalence of paratyphoid was identified among the hospital-based studies, this was not associated with case definition, laboratory assay, whether an outbreak was reported or any other study covariates, with the exception of study location. Prevalence was significantly higher in rural compared with urban locations, but this result was driven by a single study of paratyphoid in a rural area that had high prevalence.
Consistent across the community cohort studies was the finding of a high incidence of typhoid in children under five years of age, suggestive of a substantial burden in a group that would benefit from infant rather than school-based immunization. This is consistent with recent recommendations from the IAP on the creation of an immunisation slot at 9–12 months of age for typhoid vaccination [21]. Risk-factors that would allow targeting of infant immunization to high-risk groups were not identified in this systematic review and meta-analysis. Significant heterogeneity was observed among studies, but this is likely in part to reflect differences in patient inclusion criteria, laboratory methods and changes in antimicrobial use and resistance patterns. Vaccine introduction is likely to be more sustainable, equitable and to provide indirect herd effects when it is done through the universal immunisation programme.
The age-distribution of paratyphoid incidence was not reported, although the mean age in 2004 in Kolkata was reported as being significantly higher compared with typhoid (17.1 vs. 14.7 years)[29]. Paratyphoid vaccines are not yet available and currently licensed typhoid vaccines offer limited or no protective immunity against paratyphoid A and B, the predominant serotypes [33]. However, vaccines are in the development pipeline, including bivalent conjugate vaccines that could offer protection against both typhoid and paratyphoid.
There were limitations common to published studies that hampered our systematic review of the burden and risk factors for typhoid and paratyphoid fever in India. There have been only three community cohort studies of incidence, in just two locations and with highly variable findings. Although far more numerous, the hospital-based studies have a number of limitations. Firstly, hospital studies provide no information about incidence without a detailed understanding of local health-seeking behaviour—something missing from published studies. Secondly, they had varying inclusion criteria for patients, sample collection and laboratory methods, making interpretation of these data challenging. We focussed on laboratory confirmed typhoid or paratyphoid fever, mostly blood culture. However, blood culture has a poor sensitivity of about 50% and is strongly influenced by the quantity of blood, prior administration of antibiotics and culture techniques including quality of media [34]. We excluded cross-sectional community studies using serology, since these were likely to be highly non-specific for typhoid. Thirdly, detailed data on the inclusion criteria for patients including the duration of fever and age-distribution were usually missing, limiting the number of covariates we could include in the meta-analysis. In some studies, a failure to clearly define inclusion criteria along with ambiguous reporting forced us to exclude them because the denominator population was unclear. Fourthly, most studies did not report clinical outcomes and therefore we were unable to evaluate trends in severe disease or mortality. Finally, there was evidence from the funnel plots for publication bias, such that studies finding a high burden of typhoid or paratyphoid were more likely to be reported and published. Therefore, while there appears to be a declining trend in typhoid isolation in hospitals, drawing inference about the underlying burden of disease from hospital based data needs to be approached cautiously.
The limitations of the community cohort and hospital-based studies make it difficult to estimate the total burden of typhoid and paratyphoid in India. Extrapolating the estimates of typhoid incidence from Kolkata and Delhi to the rest of India is clearly problematic. A naïve approach applying the pooled estimate of the incidence rate to the 2011 census population of 1.2 billion would give an estimated annual incidence of 4.6 million cases. This could be revised upwards by approximately twofold based on the poor sensitivity of culture-based confirmation of typhoid [2]. However, the community cohort studies were deliberately planned in densely populated urban areas with poor sanitation, likely to have a high incidence of typhoid. Correcting the national estimate for access to improved water following the approach used for regional estimates by Mogasale et al. 2014 [2] would give 2.1 million cases annually, or approximately 3.4 million correcting for imperfect culture sensitivity. Correction for other risk factors, such as population density, would likely reduce this estimate further.
Strengthening surveillance across geographically representative sentinel sites is key to better disease burden estimates. Inclusion of other data sources such as large healthcare facilities, and the National Disease Surveillance Project is likely to further understanding of disease burden.
Well defined surveillance criteria combined with standardized laboratory methods will greatly enhance comparability of estimates from diverse sites. Since blood cultures are highly dependent on volume of inoculum, prior antibiotic exposure and laboratory methods, a combination of conventional, molecular and serologic diagnostics modalities would probably be optimal. Information about time trends and antimicrobial resistance patterns that arise from such a systematic surveillance will enhance our understanding of typhoid and paratyphoid in India and strengthen public health decision making.
|
10.1371/journal.pgen.1004904 | ALIX and ESCRT-III Coordinately Control Cytokinetic Abscission during Germline Stem Cell Division In Vivo | Abscission is the final step of cytokinesis that involves the cleavage of the intercellular bridge connecting the two daughter cells. Recent studies have given novel insight into the spatiotemporal regulation and molecular mechanisms controlling abscission in cultured yeast and human cells. The mechanisms of abscission in living metazoan tissues are however not well understood. Here we show that ALIX and the ESCRT-III component Shrub are required for completion of abscission during Drosophila female germline stem cell (fGSC) division. Loss of ALIX or Shrub function in fGSCs leads to delayed abscission and the consequent formation of stem cysts in which chains of daughter cells remain interconnected to the fGSC via midbody rings and fusome. We demonstrate that ALIX and Shrub interact and that they co-localize at midbody rings and midbodies during cytokinetic abscission in fGSCs. Mechanistically, we show that the direct interaction between ALIX and Shrub is required to ensure cytokinesis completion with normal kinetics in fGSCs. We conclude that ALIX and ESCRT-III coordinately control abscission in Drosophila fGSCs and that their complex formation is required for accurate abscission timing in GSCs in vivo.
| Cytokinesis, the final step of cell division, concludes with a process termed abscission, during which the two daughter cells physically separate. In spite of their importance, the molecular machineries controlling abscission are poorly characterized especially in the context of living metazoan tissues. Here we provide molecular insight into the mechanism of abscission using the fruit fly Drosophila melanogaster as a model organism. We show that the scaffold protein ALIX and the ESCRT-III component Shrub are required for completion of abscission in Drosophila female germline stem cells (fGSCs). ESCRT-III has been implicated in topologically similar membrane scission events as abscission, namely intraluminal vesicle formation at endosomes and virus budding. Here we demonstrate that ALIX and Shrub co-localize and interact to promote abscission with correct timing in Drosophila fGSCs. We thus show that ALIX and ESCRT-III coordinately control abscission in Drosophila fGSCs cells and report an evolutionarily conserved function of the ALIX/ESCRT-III pathway during cytokinesis in a multi-cellular organism.
| Cytokinesis is the final step of cell division that leads to the physical separation of the two daughter cells. It is tightly controlled in space and time and proceeds in multiple steps via sequential specification of the cleavage plane, assembly and constriction of the actomyosin-based contractile ring (CR), formation of a thin intercellular bridge and finally abscission that separates the two daughter cells [1–8]. Studies in a variety of model organisms and systems have elucidated key machineries and signals governing early events of cytokinesis [1–6]. However, the mechanisms of the final abscission step of cytokinesis are less understood, especially in vivo in the context of different cell types in a multi-cellular organism [2, 4, 5].
During the recent years key insights into the molecular mechanisms and spatiotemporal control of abscission have been gained using a combination of advanced molecular biological and imaging technologies [4, 7, 9–15]. At late stages of cytokinesis the spindle midzone transforms to densely packed anti-parallel microtubules (MTs) that make up the midbody (MB) and the CR transforms into the midbody ring (MR, diameter of ~1–2 µm) [4, 10, 16, 17]. The MR is located at the site of MT overlap and retains several CR components including Anillin, septins (Septins 1, 2 and Peanut in Drosophila melanogaster), myosin-II, Citron kinase (Sticky in Drosophila) and RhoA (Rho1 in Drosophila) and eventually also acquires the centralspindlin component MKLP1 (Pavarotti in Drosophila) [4, 16, 18, 19]. In C. elegans embryos the MR plays an important role in scaffolding the abscission machinery even in the absence of MB MTs [20].
Studies in human cell lines, predominantly in HeLa and MDCK cells, have shown that components of the endosomal sorting complex required for transport (ESCRT) machinery and associated proteins play important roles in mediating abscission [4, 7, 9–15]. Abscission occurs at the thin membrane neck that forms at the constriction zone located adjacent to the MR [9, 10, 17]. An important signal for initiation of abscission is the degradation of the mitotic kinase PLK1 (Polo-like kinase 1) that triggers the targeting of CEP55 (centrosomal protein of 55 kDa) to the MR [21]. CEP55 interacts directly with GPP(3x)Y motifs in the ESCRT-associated protein ALIX (ALG-2-interacting protein X) and in the ESCRT-I component TSG101, thereby recruiting them to the MR [13–15, 22]. ALIX and TSG101 in turn recruit the ESCRT-III component CHMP4B, which is followed by ESCRT-III polymerization into helical filaments that spiral/slide to the site of abscission [9, 11, 13–15, 23]. The VPS4 ATPase is thought to promote ESCRT-III redistribution toward the abscission site [23]. Prior to abscission ESCRT-III/CHMP1B recruits Spastin that mediates MT depolymerization at the abscission site [9, 10, 24]. ESCRT-III then facilitates membrane scission of the thin membrane neck, thereby mediating abscission [9, 10].
Cytokinesis is tightly controlled by the activation and inactivation of mitotic kinases at several steps to ensure its faithful spatiotemporal progression [7, 8]. Cytokinesis conventionally proceeds to completion via abscission, but is differentially controlled depending on the cell type during the development of metazoan tissues. For example, germ cells in species ranging from insects to humans undergo incomplete cytokinesis leading to the formation of germline cysts in which cells are interconnected via stable intercellular bridges [25–27]. How cytokinesis is modified to achieve different abscission timing in different cell types is not well understood, but molecular understanding of the regulation of the abscission machinery has started giving some mechanistic insight [25, 26, 28–30].
The Drosophila female germline represents a powerful system to address mechanisms controlling cytokinesis and abscission in vivo [29, 31]. Each Drosophila female germline stem cell (fGSC) divides asymmetrically with complete cytokinesis to give rise to another fGSC and a daughter cell cystoblast (CB) [31–33]. Cytokinesis during fGSC division is delayed so that abscission takes place during the G2 phase of the following cell cycle (about 24 hours later) [31]. The CB in turn undergoes four mitotic divisions with incomplete cytokinesis giving rise to a 16-cell cyst in which the cells remain interconnected by stable intercellular bridges called ring canals (RCs) [27, 32]. One of the 16 cells with four RCs will become specified as the oocyte and the cyst becomes encapsulated by a single layer follicle cell epithelium to form an egg chamber [34, 35]. Drosophila male GSCs (mGSCs) also divide asymmetrically with complete cytokinesis to give rise to another mGSC and a daughter cell gonialblast (GB) [33, 36, 37]. Anillin, Pavarotti, Cindr, Cyclin B and Orbit are known factors localizing at RCs/MRs and/or MBs during complete cytokinesis in fGSCs and/or mGSCs [29, 31, 36, 38–43]. Mathieu et al. recently reported that Aurora B delays abscission and that Cyclin B promotes abscission in Drosophila germ cells and that mutual inhibitions between Aurora B and Cyclin B/Cdk-1 control the timing of abscission in Drosophila fGSCs and germline cysts [29]. However, little is known about further molecular mechanisms controlling cytokinesis and abscission in Drosophila fGSCs.
Here we characterize the roles of ALIX and the ESCRT-III component Shrub during cytokinesis in Drosophila fGSCs. We find that ALIX and Shrub are required for completion of abscission in fGSCs, that they co-localize during this process and that their direct interaction is required for abscission with normal kinetics. We thus show that a complex between ALIX and Shrub is required for abscission in fGSCs and provide evidence of an evolutionarily conserved functional role of the ALIX/ESCRT-III pathway in mediating cytokinetic abscission in the context of a multi-cellular organism.
The ESCRT-associated scaffold protein ALIX promotes cytokinetic abscission in human cultured cells [13–15]. We were interested to characterize the role of ALIX in cytokinesis in vivo using Drosophila melanogaster as a model because of its power for elucidation of mechanisms of cytokinesis and abscission in different cell types in a developing organism [2, 5, 6, 29]. We first raised an antibody against Drosophila ALIX (CG12876) (Fig. 1A) and examined its subcellular localization during S2 cell division. During meta-, ana- and early telophase ALIX localized at centrosomes (Fig. 1B-E), where it co-localized with Centrosomin (S1A-S1C Fig.). ALIX localization at centrosomes has been detected in human cultured cells in interphase [15], but to our knowledge ALIX localization at centrosomes during different phases of mitosis has not previously been shown. Strikingly, at mid telophase a fraction of ALIX re-localized from the spindle poles to two pools within the intercellular bridge on each side of the MR/dark zone (Figs. 1F and S1D). Finally, ALIX localized to the central region of the intercellular bridge during late telophase/cytokinesis (Figs. 1G and S1E). Here it appeared to localize to the MR because it formed a ring-like structure around the MTs of the MB (Fig. 1G) at the dark zone (S1E Fig.). The pre-immune serum neither stained centrosomes, nor the intercellular bridge or MR (S1A-S1E Fig.). This spatiotemporal redistribution from centrosomes to the MR suggested a possible role for ALIX in cytokinesis in Drosophila cells.
We further addressed the role of Drosophila ALIX in cytokinesis in vivo by analyzing two different alix mutant alleles, alix1 and alix3 (Fig. 2A). ALIX is highly expressed in Drosophila embryos, larvae, pupae, adult females and males, as well as in ovaries and testes (S2A Fig.). Interestingly, homozygous mutant offspring of both the alix1 and alix3 mutants could survive to adulthood (even though they clearly lack the full-length ALIX protein) (S2B Fig. and see below) and we detected none or only minor bi-nucleation clearly attributed to cytokinesis failure in the somatic cell types we analyzed (S2C-S2F Fig. and S1–S2 Tables). Fertility tests of alix1 mutant flies however revealed that both female and male fertility was reduced (S3A–S3B Fig.). In particular female fertility was severely compromised, manifested by very low egg lay and hatch rates (S3A–S3B Fig.). We therefore asked whether oogenesis of alix mutant flies might be altered. Wild type egg chambers contain 16 germ cells and an oocyte with 4 RCs (Fig. 2B, 2D, 2G, 2E and 2H). Curiously, egg chambers in ovaries of both alix1 and alix3 mutant females lacking full-length ALIX (Fig. 2C and 2F) often contained exactly 32 germ cells and an oocyte with 5 RCs (Fig. 2D and 2G). Quantifying the egg chamber phenotypes of alix1 and alix3 mutant ovaries revealed that about 60% of the egg chambers in both alleles contained 32 germ cells (Fig. 2E and 2H). We also detected low percentages of egg chambers with more than 32 germ cells in both alix mutant alleles (Fig. 2E and 2H).
We next analyzed whether the increased germ cell number in egg chambers was specifically due to loss of alix gene function. Firstly, alix1 and alix3 alleles combined either with two different deficiencies lacking the alix gene or with each other gave rise to 50–60% of egg chambers with 32 germ cells, similar to homozygous alix1 and alix3 mutants (S3C-S3E Fig.). Secondly, two genomic rescue lines containing the full alix gene locus rescued the 32-germ cell phenotype of both the alix1 and alix3 alleles (S4A–S4F Fig.). Finally, RNAi-mediated gene silencing of alix specifically in female germ cells using the maternal triple MTD-GAL4 driver [44–46] resulted in about 50% of egg chambers with 32 or more germ cells (S4G-S4I Fig.) showing that absence of ALIX specifically in germ cells causes the 32-germ cell phenotype. We conclude that loss of ALIX function in the Drosophila female germline causes the formation of a high frequency of egg chambers with 32 or more germ cells.
Egg chambers with 32 germ cells may arise via encapsulation of two 16-cell cysts by the follicle cell epithelium, an extra round of mitosis in germline cysts or a delay in abscission in fGSCs [29, 32, 35, 47, 48]. The fact that the egg chambers with 32 germ cells contained one oocyte with 5 RCs excluded that they arose via defective encapsulation of two 16-cell cysts. We further discriminated between the two latter mechanisms by performing RNAi-mediated gene silencing of alix specifically in the germline using either Nanos-GAL4 (expresses in all germ cells; fGSCs, CBs and 2–16-cell cysts) or Bam-GAL4 (expresses in CBs to 8-cell cysts, but not in fGSCs) to test whether the phenotype originated from fGSCs or cell autonomously in germline cysts. Interestingly, alix-RNAi using Nanos-GAL4 (Nanos-GAL4 or UAS-Dicer; Nanos-GAL4) gave rise to 40–60% egg chambers with 32 germ cells, whereas alix depletion using Bam-GAL4 resulted in normal egg chambers with 16 germ cells only (S5A-S5D Fig.). These data linked alix depletion in fGSCs to the formation of egg chambers with 32 germ cells and suggested that they did not arise from an extra round of mitosis of germline cysts. This thus indicated a role for ALIX in abscission in fGSCs in agreement with recent work showing that a delay in abscission in fGSCs can give rise to the formation of stem cysts in which the fGSC is connected to several daughter cells [29]. If abscission eventually takes place, a 2-cell cysts may pinch off and subsequently undergo four rounds of mitosis, giving rise to a 32-cell cyst [29]. We thus investigated whether or not we could detect stem cysts following loss of ALIX function.
Stem cysts are characterized by their elongated fusomes, their weak Nanos expression as in stem cells, their lack of expression of the cyst differentiation factor Bam and that the cell in direct contact with the stem cell niche is positive for p-Mad [29]. The cells within the stem cysts are moreover found to divide synchronously [29]. Importantly, alix1 and alix3 germaria as well as germaria with alix-RNAi in fGSCs displayed chains of weakly Nanos-positive germ cells interconnected by elongated fusomes in which the most anterior cell was in contact with the cap cells in the stem cell niche of the germarium (Figs. 3A-C and S5E-S5H). The cell in contact with the cap cells in such alix-deficient cysts was moreover p-Mad-positive (S6A Fig.) and the cysts were Bam-negative (S6B-S6D Fig.). We also detected synchronously dividing cells in the anterior tip of alix-deficient germaria (Fig. 3I). Taken together, these characteristics defined the alix-deficient cysts as stem cysts and indicated a role for ALIX in abscission in fGSCs.
Each fGSC divides with complete cytokinesis giving rise to another stem cell and a daughter cell CB [31, 32] (Fig. 2B). fGSC cytokinesis progression can be monitored using markers for the fusome and RCs (hereafter referred to as MRs) [31, 32, 49, 50]. To determine the nature and frequency of the abscission defects upon loss of ALIX function we quantified fGSC morphologies in wild type, alix1 and alix3 germaria using markers for the fusome (hts-F), MRs/MBs (Cindr) [38] and nuclei (Figs. 3D-G and S7A-S7H). We categorized fGSC phenotypes as indicated in Fig. 3H (and as illustrated in S7E Fig.). Wild type fGSCs displayed only normal phenotypes: ~50% fGSCs with a spectrosome, ~40% fGSC-CB pairs with an MR and ~10% fGSC-CB pairs with an MB (Figs. 3D-E, 3H, S7A-S7B and S7E). These frequencies of different fGSC cell cycle stages are consistent with previous reports [49, 51]. alix3 and alix1 mutant germaria contained smaller fractions of fGSCs with a spectrosome (~15% for both mutants), fGSC-CB pairs with an MR (~25% in alix3 and ~10% in alix1) and fGSC-CB pairs with an MB (0% in alix3 and ~1% in alix1) compared to wild type (Figs. 3H and S7E). Importantly, more than half of the alix mutant fGSCs showed abscission defects: linear chains (~20% in alix3 and ~10% in alix1), branched chains (~40% in alix3 and ~30% in alix1) or polyploidy (~2% in alix3 and ~30% in alix1) (Figs. 3F-H and S7C-S7H). Abscission defects appeared in the majority of both alix3 and alix1 mutant germaria, and never in wild type (S3–S4 Tables). Consistently, upon alix-RNAi in the germline using Nanos-GAL4 (Nanos-GAL4 or UAS-Dicer; Nanos-GAL4) the majority of germaria contained stem cysts in which the fGSC was interconnected to multiple daughter cells via fusome and MRs (S7I-S7K and S5H Figs.). We occasionally detected MBs in stem cysts in alix1 and alix3 mutant germaria (even though MRs predominated), indicating abscission events, and cysts of exactly two cells in the process of pinching off (S7L-S7N Fig.). This is consistent with the model of how 32-cell cysts appear following delayed abscission in fGSCs as previously described [29]. Collectively, these results showed that loss of ALIX caused a delay in abscission in fGSCs with the consequent formation of a high frequency of stem cysts. The fact that cells in stem cysts were interconnected in chains via MRs (Figs. 3F-H, S7C-S7E and S7I-S7J) together with the infrequent observation of fGSC-CB pairs with an MB upon loss of ALIX function (Figs. 3H and S7E) suggested that ALIX plays a role in promoting closure of the MR to mediate fGSC abscission. We conclude that ALIX is required for completion of abscission in Drosophila fGSCs.
We further asked whether ALIX may also be required for abscission in asymmetrically dividing Drosophila mGSCs. We stained testes tips from wild type, alix1 and alix3 mutants with antibodies to visualize the hub to which the mGSCs are attached, the fusome, MRs and MBs. In alix1 and alix3 mutant testes that lack full-length ALIX (S8A Fig.) ~20% of alix1 and ~40% of alix3 mutant mGSCs were found interconnected to chains of daughter cells by MRs and fusome (S8B-S8E Fig.). These results suggest that ALIX promotes abscission in both female and male GSCs.
To examine the subcellular localization of ALIX during cytokinesis in fGSCs we generated transgenic flies with GFP-tagged ALIX under the control of the UASp promoter (UASp-GFP-ALIX) and expressed it in fGSCs and germline cysts using MTD-GAL4 or Nanos-GAL4. We then visualized the progressive stages of fGSC cytokinesis using markers for the fusome and MRs/MBs (Cindr) or MTs (α-tubulin). In fGSC-CB pairs in which a small fusome plug had formed within the MR (G1) we detected GFP-ALIX overlapping mainly with the fusome plug (Fig. 4A). At this point we detected anti-parallel MT bundles with a dark zone to which the fusome plug started localizing (S9A Fig.). Then, as the fusome adopted bar morphology in G1/S GFP-ALIX localized at the MR and at this point the MTs were largely degraded (S9B Fig.). GFP-ALIX remained at MRs throughout G1/S, S and early G2 (Figs. 4B and S9B-S9C) and then localized at MBs during abscission (G2) (Fig. 4C). We thus conclude that GFP-ALIX is recruited to the center of the MR and then moves to the MR during G1/S, is detected at MRs throughout cytokinesis progression and then finally localizes to MBs during abscission in Drosophila fGSCs. This spatiotemporal dynamics of ALIX during late stages of fGSC cytokinesis is consistent with a role for ALIX in abscission in fGSCs.
We next asked by which molecular mechanisms ALIX may act during abscission in fGSCs. The ESCRT-III component and CHMP4 orthologue Shrub (CG8055) was an interesting candidate to mediate abscission together with ALIX because of the important role of ESCRT-III in promoting membrane scission during cytokinetic abscission and because ALIX directly interacts with and recruits the ESCRT-III subunit CHMP4B to the MB to promote abscission in human cells [13, 15, 52, 53]. The interaction between ALIX and CHMP4B is mediated via a motif within the Bro1 domain of human ALIX (MxxxIxxxL, aa 199–216) and a motif in the CHMP4 C-terminus (MxxLxxW, aa 214–220) [13, 15, 54]. Importantly, these mutual consensus interaction sites are conserved in Drosophila ALIX and Shrub, respectively (ALIX: LxxxIxxxL, aa 198–215 and Shrub: MxxLxxW, aa 218–224) (Fig. 5A) [54]. We therefore tested the possible interaction by co-immunoprecipitation analyses of GFP-tagged Shrub and endogenous ALIX from Drosophila Dmel cell lysates. These analyses showed that GFP-Shrub and ALIX indeed were detected in the same complex (Fig. 5B).
We next examined the relative localization of ALIX and Shrub during fGSC cytokinesis. For this purpose GFP-Shrub was expressed using Nanos-GAL4 and ALIX was detected with our anti-ALIX antibody. Interestingly, GFP-Shrub localized at MRs and MBs during cytokinesis in fGSCs (consistent with observations by [55]) and ALIX co-localized with GFP-Shrub at MRs in G1/S and S phase and then at MBs during abscission in G2 (Fig. 5C-F). Strikingly, GFP-Shrub additionally localized at the fusome (Fig. 5C-D and [55]). Furthermore, ALIX and GFP-Shrub co-localized at bright dot-like structures on the fusome in fGSCs (Fig. 5E). These most likely represented MB remnants that have been reported to be inherited by the fGSC following cytokinesis completion [36]. Consistently, we detected that GFP-ALIX on MB remnants was preferentially retained in fGSCs following abscission (data not shown). We also noted that GFP-Shrub was weakly detected along the membrane at the point that anti-parallel MTs were detected in fGSC-CB pairs in G1 (S9D Fig.) and then accumulated at MRs from G1/S (Figs. 5C-D and S9D). Taken together our results suggested that ALIX and GFP-Shrub co-localize at MRs from G1/S and then at MRs and MBs throughout cytokinetic abscission in Drosophila fGSCs.
We next analyzed the role of Shrub as well as the possible functional relationship between ALIX and Shrub during fGSC cytokinesis. We first performed RNAi-mediated depletion of shrub using the Nanos-GAL4 driver. Control germaria displayed normal fGSC and egg chamber phenotypes (Figs. 6A, 6E, S10A and S10E). Upon alix-RNAi about 40% of fGSCs were found in stem cysts (linear or branched) (Fig. 6B and 6E) and ~50% of the egg chambers contained 32 germ cells (S10B and S10E Fig.). Importantly, following shrub-RNAi about 45% of fGSCs formed stem cysts (Fig. 6C and 6E), ~10% of the fGSCs were polyploid (Fig. 6E) and ~50% of the egg chambers contained 32 germ cells (Figure S10C and S10E Fig.). Consistently, stem cysts were also present in 70% of heterozygous shrubG5/+ mutant germaria (S10F Fig.) suggesting that the stem cysts appeared specifically due to loss of Shrub function. These results showed that loss of Shrub function caused delayed abscission in Drosophila fGSCs and that Shrub is required for completion of abscission in these cells.
To test the functional relationship between ALIX and Shrub in fGSCs we performed combined shrub- and alix-RNAi using Nanos-GAL4. We detected about 55% of the fGSCs in stem cysts (Fig. 6D-E), 15% polyploid fGSCs (Fig. 6E) as well as about 40% of egg chambers with 32 germ cells and 15% of egg chambers with more than 32 germ cells (compared to 3–4% in alix- or shrub-RNAi) (S10D-S10E Fig.). The increased frequency of egg chambers with more than 32 germ cells suggested an even more delayed abscission rate upon combined ALIX and Shrub depletion than following reduction of either ALIX or Shrub levels alone. Consistently, reducing the Shrub levels in the alix1 mutant background (shrubG5/+; alix1) gave, in addition to stem cysts, rise to the appearance ~10% of germaria with polyploid fGSCs, ~10% of agametic germaria as well as fewer stem cells per germarium than normal suggesting that reduction of the ALIX and Shrub levels in all cell types in the germarium both caused abscission defects and affected germ cell viability (S10F-S10G Fig.). The fact that reducing the levels of both ALIX and Shrub in fGSCs simultaneously caused even more delayed abscission kinetics in fGSCs as compared to decreasing the levels of either of them alone indicated that ALIX and Shrub are required for the same process to promote abscission in Drosophila fGSCs.
We next asked whether the complex formation between ALIX and Shrub is important for abscission in fGSCs. In human cells interfering with the interaction between ALIX and CHMP4 causes multi-nucleation and defective midbody morphology [13, 15]. We introduced point mutations in Drosophila GFP-ALIX (GFP-ALIX-F198D and GFP-ALIX-I211D) of residues which have previously been shown to mediate the interaction with CHMP4 in human cells [13, 15, 54]. Indeed, we could verify the importance of these residues for the ALIX-Shrub interaction since wild type GFP-ALIX co-precipitated substantially more Shrub than the two mutant proteins in GFP trap analyses (Fig. 7A). To further assess the functional importance of the interaction between ALIX and Shrub in abscission in fGSCs we generated flies expressing GFP-ALIX, GFP-ALIX-F198D or GFP-ALIX-I211D using Nanos-GAL4 either alone or in the alix1 mutant background. Both GFP-ALIX-F198D and GFP-ALIX-I211D localized at MRs and MBs like wild type GFP-ALIX (Figs. 7B and S11A) and their expression per se did not induce the formation of stem cysts (Fig. 7B-C). Importantly, wild type GFP-ALIX rescued the fGSC abscission defects in alix1 mutant germaria from 76% of fGSCs in stem cysts to 22% (Fig. 7C, p < 0.05). GFP-ALIX-F198D or GFP-ALIX-I211D could on the other hand not rescue the fGSC abscission defects as 59% and 56% of fGSCs were found in stem cysts following their expression in alix1 mutant germ cells, respectively (Fig. 7B-C, borderline significant, p = 0.05). In agreement, the stem cyst lengths upon expression of GFP-ALIX-F198D or GFP-ALIX-I211D in the alix1 mutant background were similar to the stem cyst lengths in the alix1 mutant, whereas they were shorter upon expression of GFP-ALIX (S11B Fig.). These results suggest that ALIX requires the interaction with Shrub to mediate abscission in fGSCs. Moreover, consistent with the stem cyst phenotypes the expression of wild type GFP-ALIX in the alix1 mutant background rescued the number of egg chambers with 32 cells from 49% to 13% (p < 0.05), whereas neither GFP-ALIX-F198D nor GFP-ALIX-I211D expression in alix1 mutant ovaries could rescue the 32-cell phenotype (40% and 39%, respectively, p < 0.05) (S11C Fig.). Collectively, these results demonstrate that the direct interaction between ALIX and Shrub is required for completion of abscission with normal kinetics in Drosophila fGSCs.
The mechanisms controlling the kinetics of cytokinetic abscission in different cell types in the context of a multi-cellular organism are not well understood. The Drosophila female germline has emerged as a powerful genetically amendable model system to address mechanisms of cytokinetic abscission in vivo [29]. In this study we show that the scaffold protein ALIX and the ESCRT-III component Shrub form a complex to mediate completion of cytokinetic abscission in Drosophila fGSCs with normal kinetics. Loss of ALIX or/and Shrub function or inhibition of their interaction delays abscission in fGSCs leading to the formation of stem cysts in which the fGSC remains interconnected to chains of daughter cells via MRs. As abscission eventually takes place a cyst of e.g. 2 germ cells may pinch off and subsequently undergo four mitotic divisions to give rise to a germline cyst with 32 germ cells [29]. Consistently, loss of ALIX or/and Shrub or interference with their interaction caused a high frequency of egg chambers with 32 germ cells during Drosophila oogenesis. We also found that ALIX controls cytokinetic abscission in both fGSCs and mGSCs and thus that ALIX plays a universal role in cytokinesis during asymmetric GSC division in Drosophila. Taken together we thus provide evidence that the ALIX/ESCRT-III pathway is required for normal abscission timing in a living metazoan tissue.
Our results together with findings in other models underline the evolutionary conservation of the ESCRT system and associated proteins in cytokinetic abscission. Specifically, ESCRT-I or ESCRT-III have been implicated in abscission in a subset of Archaea (ESCRT-III) [56–58], in A. thaliana (elch/tsg101/ESCRT-I) [59] and in C. elegans (tsg101/ESCRT-I) [20]. In S. cerevisiae, Bro1 (ALIX) and Snf7 (CHMP4/ESCRT-III) have also been suggested to facilitate cytokinesis [60]. In cultured Drosophila cells, Shrub/ESCRT-III mediates abscission and in human cells in culture ALIX, TSG101/ESCRT-I and CHMP4B/ESCRT-III promote abscission [9, 11, 13–16]. ALIX and the ESCRT system thus act in an ancient pathway to mediate cytokinetic abscission.
Despite the fact that we find an essential role of ALIX in promoting cytokinetic abscission during asymmetric GSC division in the Drosophil a female and male germlines, we did not detect strong bi-nucleation directly attributed to cytokinesis failure in Drosophila alix mutants in the somatic cell types we have examined. This might have multiple explanations. One possibility is that maternally contributed alix mRNA may support normal cytokinesis and development. Whereas ALIX and CHMP4B depletion in cultured mammalian cells causes a high frequency of bi- and multi-nucleation [14, 15] it is also possible that cells do not readily become bi-nucleate upon failure of the final step of cytokinetic abscission in the context of a multi-cellular organism. Consistent with our observations of a high frequency of stem cysts upon loss of ALIX and Shrub in the germline, Shrub depletion in cultured Drosophila cells resulted in chains of cells interconnected via intercellular bridges/MRs due to multiple rounds of cell division with failed abscission [16]. Moreover, loss of ESCRT-I/tsg101 function in the C. elegans embryo did not cause furrow regression [20]. These and our observations suggest that ALIX- and Shrub/ESCRT-depleted cells can halt and are stable at the MR stage for long periods of time and from which cleavage furrows may not easily regress, at least not in these cell types and in the context of a multi-cellular organism. It is also possible that redundant mechanisms contribute to abscission during symmetric cytokinesis in somatic Drosophila cells. Further studies should address the general involvement of ALIX and ESCRT-III in cytokinetic abscission in somatic cells in vivo.
Different cell types display different abscission timing, intercellular bridge morphologies and spatiotemporal control of cytokinesis [10, 26, 29]. In fGSCs we found that ALIX and Shrub co-localize throughout late stages of cytokinesis and abscission. In human cells ALIX localizes in the central region of the MB, whereas CHMP4B at first localizes at two cortical ring-like structures adjacent to the central MB region and then progressively distributes also at the constriction zone where it promotes abscission [9–11, 13, 15, 23]. ALIX and CHMP4B are thus found at discrete locations within the intercellular bridge as cells approach abscission in human cultured cells. In contrast, ESCRT-III localizes to a ring-like structure during cytokinesis in Archaea, resembling the Shrub localization at MRs we observed in Drosophila fGSCs [56, 57]. Moreover, ALIX and Shrub are present at MRs for a much longer time (from G1/S) prior to abscission (in G2) in fGSCs than in human cultured cells. Here, ALIX and CHMP4B are increasingly recruited about an hour before abscission and then CHMP4B acutely increases at the constriction zones shortly (~30 min) before the abscission event [9, 11].
How may ALIX and Shrub be recruited to the MR/MB in Drosophila cells in the absence of CEP55 that is a major recruiter of ALIX and ultimately CHMP4/ESCRT-III in human cells [13, 15]? Curiously, we detect a GPP(3x)Y consensus motif within the Drosophila ALIX sequence (GPPPGHY, aa 808–814) resembling the CEP55-interacting motif in human ALIX (GPPYPTY, aa 800–806). Whether Drosophila ALIX is recruited to the MR/MB via a protein(s) interacting with this motif or other domains is presently uncharacterized. Accordingly, alternative pathways of ALIX and ESCRT recruitment have been reported [61–64], as well as suggested in C. elegans, where CEP55 is also missing [20]. Further studies are needed to elucidate mechanisms of recruitment and spatiotemporal control of ALIX and ESCRT-III during cytokinesis in fGSCs and different cell types in vivo.
We found that the direct interaction between ALIX and Shrub is required for completion of abscission with normal kinetics in fGSCs. This is consistent with findings in human cells in which loss of the interaction between ALIX and CHMP4B causes abnormal midbody morphology and multi-nucleation [13, 15]. Following ALIX-mediated recruitment of CHMP4B/ESCRT-III to cortical rings adjacent to the MR in human cells, ESCRT-III extends in spiral-like filaments to promote membrane scission [9–11, 13, 15, 23]. Due to the discrete localizations of ALIX and CHMP4B during abscission in human cells ALIX has been proposed to contribute to ESCRT-III filament nucleation [15, 53]. In vitro studies have shown that the interaction between ALIX and CHMP4B may release autoinhibitory intermolecular interactions within both proteins and promote CHMP4B polymerization [54, 65]. Specifically, ALIX dimers can bundle pairs of CHMP4B filaments in vitro [65]. Moreover, in yeast, the interaction of the ALIX homologue Bro1 with Snf7 (CHMP4 homologue) enhances the stability of ESCRT-III polymers [66, 67]. There is a high degree of evolutionary conservation of ALIX and ESCRT-III proteins [52–54, 68, 69] and because ALIX and Shrub co-localize and interact to promote abscission in fGSCs it is possible that ALIX can facilitate Shrub filament nucleation and/or polymerization during this process.
Our findings indicate that accurate control of the levels and interaction of ALIX and Shrub ensure proper abscission timing in fGSCs. Their reduced levels or interfering with their complex formation caused delayed abscission kinetics. How cytokinesis is modified to achieve a delay in abscission in Drosophila fGSCs and incomplete cytokinesis in germline cysts is not well understood [25–27]. Aurora B plays an important role in controlling abscission timing both in human cells and the Drosophila female germline [29, 70, 71]. During Drosophila germ cell development Aurora B contributes to mediating a delay of abscission in fGSCs and a block in cytokinesis in germline cysts [29]. Bam expression has also been proposed to block abscission in germline cysts [29, 32, 72, 73]. It will be interesting to investigate mechanisms regulating the levels, activity and complex assembly of ALIX and Shrub and other abscission regulators at MRs/MBs to gain insight into how the abscission machinery is modified to control abscission timing in fGSCs.
We found that intercellular bridge MTs in fGSC-CB pairs were degraded in G1/S when the fusome adopted bar morphology. Abscission in G2 thus appears to occur independently of intercellular bridge MTs in Drosophila fGSCs. This has also been described in C. elegans embryonic cells where the MR scaffolds the abscission machinery as well as in Archaea that lack the MT cytoskeleton [20, 56, 57]. In mammalian and Drosophila S2 cells in culture, on the other hand, intercellular bridge MTs are present until just prior to abscission [9, 11].
It is interesting to note a resemblance of the stem cysts that appeared upon loss of ALIX and Shrub function to germline cysts in that the MRs remained open for long periods of time similar to RCs. Some modification of ALIX and Shrub levels/recruitment may thus contribute to incomplete cytokinesis in Drosophila germline cysts under normal conditions. Because we detected stem cysts in the case when ALIX weakly interacted with Shrub it is also possible that inhibition of their complex assembly/activity may contribute to incomplete cytokinesis in germline cysts. Abscission factors, such as ALIX and Shrub, may thus be modified and/or inhibited during incomplete cytokinesis in germline cysts. Such a scenario has been shown in the mouse male germline where abscission is blocked by inhibition of CEP55-mediated recruitment of the abscission machinery, including ALIX, to stable intercellular bridges [26, 30]. Altogether our data thus suggest that ALIX and Shrub are essential components of the abscission machinery in Drosophila GSCs, and we speculate that their absence or inactivation may contribute to incomplete cytokinesis. More insight into molecular mechanisms controlling abscission timing and how the abscission machinery is modified in different cellular contexts will give valuable information about mechanisms controlling complete versus incomplete cytokinesis in vivo.
Summarizing, we here report that a complex between ALIX and Shrub is required for completion of cytokinetic abscission with normal kinetics during asymmetric Drosophila GSC division, giving molecular insight into the mechanics of abscission in a developing tissue in vivo.
Fly crosses and experiments were performed at 25°C unless noted otherwise. w1118 was used as a wild type control. w1118, w1118; Nanos-GAL4, UAS-Dcr-2, w1118; Nanos-GAL4, y v; attP2, TRiP-alix (UAS-shRNA-alix, chr 3, TRiP# HMS00298), TRiP-shrub (UAS-shRNA-shrb, chr 2, TRiP# HMS01767) [45], MTD-GAL4 [45], yw; P{EPgy2}ALiXEY10362, Df(3R)BSC499/TM6C, Sb, w1118; Df(3R)BSC739/TM6C, Sb and w*; shrbG5 P{neoFRT}42D/CyO, P{GAL4-twi.G}2.2, P{UAS-2xEGFP}AH2.2 (shrubG5/+) were from BDSC (Indiana University) and PBac{WH}ALiXf03094 from Exelixis at Harvard Medical School (referred to as alix1). The alix3 allele was generated by imprecise excision of the P-element of the yw; P{EPgy2}ALiXEY10362 line. The breakpoints were determined by sequencing and this allele lacks 860 bp in the 5’ end of the gene (nucleotides 23534881 to 23525741 on 3R missing), thus removing the alix gene start codon, exons 1, 2 and most of exon 3. The FRT82B, alix1 and FRT82B, alix3 lines were generated by recombination of to FRT82B chromosomes by standard procedures. The generation of the genomic alix rescue lines is described below. The alix1 and alix3 alleles were kept as stocks balanced over TM6B, Tb and TM6B, dfd gfp chromosomes. UASp-GFP-Shrub and Nanos-GAL4, UASp-GFP-Shrub were generated as described in [55]. Bam-GAL4 (chr 3) was a kind gift from M. Fuller (Stanford School of Medicine, CA), hsflp, tubulin-GAL4, UAS-GFP; FRT82, tubulin-GAL80/TM6B, Tb (MARCM82) was a kind gift from M. Peifer (University of North Carolina).
ALIX (CG12876) antibodies were generated by immunizing a guinea pig with two peptides (CIQSTYNGASEEEKG-/CERLLDEERDSDNQL-amide) (BioGenes) from the Bro1 domain. Primary antibodies and dilutions for immunofluorescence (IF) or Western blot (WB) were guinea pig anti-ALIX (IF: 1:1000–3000, WB: 1:1000), mouse anti-ALIX (WB: 1:1000, a kind gift from T. Aigaki, Tokyo Metropolitan University, Japan), rabbit anti-Cindr (IF: 1:1000) (Haglund et al., 2010), mouse anti-hts-F (IF: 1:50, 1B1, DSHB), rabbit anti-Shrub (WB: 1:1000, a kind gift from F-G. Bao, University of Massachusetts Medical School, MS [74], mouse anti-α-spectrin (IF: 1:25, 3A9, DSHB), goat anti-Vasa (IF: 1:100, dC-13, Santa Cruz Biotechnology), rabbit anti-Nanos (IF: 1:1000, a kind gift from A. Nakamura, RIKEN Center for Developmental Biology), mouse anti-Bam (IF: 1:10, DSHB), mouse anti-α-tubulin (WB: 1:10,000, Sigma), sheep anti α-tubulin (IF: 1:250, Cytoskeleton), guinea pig anti-Cnn (IF: 1:500, a kind gift from T.C. Kaufman, Indiana University), rabbit anti-phospho-Histone H3 (IF: 1:500, Millipore), rabbit anti-phosphotyrosine (IF: 1:500, Sigma), mouse anti γ-tubulin (IF: 1:500, Sigma), mouse anti-Fasciclin III (FasIII, IF: 1:50, 7G1, DSHB), rabbit anti-phospho-Smad1/5 (Ser463/465) (IF: 1:100, 41D10, Cell Signaling). GFP–Booster_Atto488 (IF: 1:200) was from Chromotek. To visualize F-actin, Alexa Fluor 647 phalloidin (1:50), Alexa Fluor 488 phalloidin (1:100) or rhodamine phalloidin (1:400) (Molecular Probes) were included in secondary antibody incubations. Secondary antibodies were conjugated to Alexa Fluor 488, Alexa Fluor 594 (1:200, Molecular Probes), Cy3 or Cy5 (1:500, Jackson Immunoresearch). DNA was stained using Hoechst 33342 (1μg/μl, Invitrogen). pOT2-ALIX as well as pAGW and pPGW vectors were from the Drosophila Genomics Resource Center (DGRC) (Bloomington, IN). pAc-Shrub-GFP was a kind gift from T. Takeda and D. Glover (University of Cambridge, Cambridge, UK).
S2 GFP-α-tubulin cells were a kind gift from E. Griffis (University of Dundee, UK) and S2 cells were from ATCC (CRL-1963) (a kind gift from R. Palmer, Umeå University, Sweden). Drosophila D.Mel-2 (Dmel) cells (a kind gift from P.P d’Avino and D. Glover, University of Cambridge, Cambridge, UK) were grown in Express Five SFM medium (Gibco) containing 2 mM L-glutamine, 100 U/ml penicillin and 100µg/ml streptomycin and Drosophila Schneider 2 (S2) cells were cultured in Schneider’s Drosophila Medium (Gibco) supplemented with 10% fetal calf serum, 2 mM L-glutamine, 100 U/ml penicillin and 100µg/ml streptomycin.
S2 cells were seeded on coverslips for two hours before 12 min fixation at room temperature in 4% formaldehyde (EM grade, Polysciences) in PHEM buffer (60 mM Pipes pH 6.8, 25 mM Hepes pH 7.0, 10 mM EGTA pH 8.0, 4 mM MgSO4). The cells were then washed three times with PBS and incubated in PBS + 5% BSA + 0,1% Triton X-100) for at least 1 h. Primary antibodies were diluted in PBS + 1% BSA + 0,1% Triton X-100 (PBT) and cells incubated with primary antibodies over night at 4 degrees. Cells were then washed twice in PBT for 15 min and then incubated with secondary antibodies diluted in PBT for 2 hrs at room temperature. They were then washed twice in PBT as before followed by incubation with Hoechst 33342 diluted in PBS to 1μg/μl for 5 min. Cells were finally rinsed with PBS and mounted in Mowiol.
Ovaries or testes were dissected in PBS and fixed using 4% formaldehyde (EM grade, Polysciences) for 30 min either on ice (all samples including anti-Cindr antibodies) or at room temperature (RT) (prior to anti-α-tubulin staining). Tissues were subjected to permeabilization (3 × 15min) and blocking (30 min) in PBS + 0.3% bovine serum albumin (BSA) + 0.3% Triton X- 100 (PBT) at RT and then incubated with primary antibodies diluted in PBT at 4°C over night. Samples were then washed three times 15 min in PBT, incubated with secondary antibodies diluted in PBT for 2 hrs at room temperature followed by three 15 min washes in PBT. For DNA staining, samples were subsequently stained with Hoechst 33342 (1 μg/μl) diluted in PBS for 10 min. Samples were mounted in anti-fading mounting medium (Prolong Antifade, Molecular Probes or Vectashield, Vector laboratories). For anti-ALIX staining, ovaries were fixed in ice-cold methanol for 7 min and subsequently stained as above with the addition of GFP-Booster (1:200) in the secondary antibody solution. For p-Mad detection the ovaries were fixed for 40 min in 4% formaldehyde with phosphatase inhibitor cocktail (Sigma, 1:200) and stained according to the protocol by Luo et. al. [75].
Images were captured using Zeiss LSM 780, Zeiss LSM 710 or Zeiss LSM 5 DUO laser scanning confocal microscopes (Carl Zeiss, Inc.) equipped with NeoFluar 63×/1.4 NA and 100×/1.45 NA oil immersion and Plan Apochromat 20×/0.8 NA objectives at 20°C. Image processing and analysis were done using the Zeiss LSM 510 (Version 3.2, Carl Zeiss, Inc.) and Zen 2009 softwares and Adobe Photoshop CS4 (Adobe). Images are planar projections of sections from z-stacks of germaria unless otherwise noted.
Ovaries of 2 to 4-day-old females (unless otherwise noted) that had been fed with yeast paste and kept with a couple of males for 2 days were dissected, fixed and stained with antibodies to visualize the fusome (hts-F), MRs/MBs (Cindr), RCs (pTyr)/F-actin (fluorescently labeled phalloidin) and nuclei (Hoechst) as described above. Confocal z-stacks of germaria were acquired at the confocal microscope and fGSC phenotypes were analyzed from z-stacks and three-dimensional reconstructions of z-stacks. fGSC identity was determined based on its anterior localization in the germarium, its fusome morphology and contact with the cap cells. Phenotype scoring was based on the fusome morphology, presence, absence, number and position of Cindr-positive MRs/MBs, cell-cell boundaries and nuclei. We categorized fGSCs into normal morphologies: (i) fGSCs with a spherical spectrosome, (ii) fGSC-CB pairs with an MR (includes plug, bar, dumbbell and fusing fusome morphologies) and (iii) fGSC-CB pairs in abscission with an MB between them (exclamation point fusome) as well as abnormal abscission-defective morphologies: (iv) linear chains of cells interconnected via fusome and MRs, (v) branched chains or (vi) polyploid, bi- or multinucleate fGSCs. Egg chamber phenotypes were scored at the microscope based on the number of RCs to the oocyte and the number of germ cell nuclei.
To examine whether differences between controls and alix1 or alix3 germaria were significant within experiments, each germarium was classified as either normal or non-normal (the latter being the case if at least one non-normal phenotype was present—linear, branched or polyploid). Fisher’s exact test was then used to determine significance. To test whether differences of fGSC phenotypes (classified as above) or egg chamber phenotypes between Nos-GAL4/GFP-ALIX and alix1, Nos-GAL4/GFP-ALIX-F198D; alix1 or Nos-GAL4/GFP-ALIX-I211D; alix1 ovaries were significant we used a mixed factor model with each experiment as random factor.
To generate N-terminally GFP-tagged alix, a PCR fragment corresponding to the whole-length alix cDNA (except the START codon) was amplified from a cDNA clone from the BDGP Gold cDNA Collection (DGRC, Bloomington, IN) using the primers 5’AATGGATCCGGTCGAAGTTTCTGGGCGTGCCG3’ and 5’AATGCGGCCGCTTACCAGCCAGGTGGCTTCTG3’ and the Phusion High-Fidelity PCR Kit (New England Biolabs). The alix cDNA was purified using the QIAquick PCR Purification Kit (Qiagen), and cloned into the pENTR1A Gateway entry vector using the T4 DNA ligase (Roche). The alix gene was then transferred by LR recombination using the Gateway LR clonase II enzyme mix (Invitrogen) to the pPGW (for generation of fly lines) or pAGW (for cell lines) destination vectors (DGRC, Bloomington, IN). Site-directed in vitro mutagenesis was used to introduce point mutations in the pENTR1A-alix vector using primers containing the specific mutations and the Phusion High-Fidelity PCR Kit (New England Biolabs). For generating ALIX-F198D the primers 5’CCAAGCGCAGGAGGTTGACATTCTGAAGGCAATTAAGG3’ and 5’CCTTAATTGCCTTCAGAATGTCAACCTCCTGCGCTTGG3’ were used, and for ALIX-I211D the primers 5’CTTGAAGGACCAGGACATCGCCAAGCTTTGCTGC3’ and 5’GCAGCAAAGCTTGGCGATGTCCTGGTCCTTCAAG3’ were used. The plasmid was then treated with DpnI (New England Biolabs) for one hour at 37°C after PCR amplification. The mutated alix cDNAs were then transferred to the pPGW (for fly lines) and pAGW (for cell lines) destination vectors by LR recombination
The transgenic UASp-GFP-ALIX, UASp-GFP-ALIX-F198D and UASp-GFP-ALIX-I211D Drosophila lines were generated by P-element transformation performed by BestGene Inc. The expression of GFP-ALIX was verified by Western blot analysis.
Approximately 1 hour before transfection, 8×106 Dmel cells were seeded in 10 cm plates. The cells were transiently transfected for 48 hours with 2,5 µg pAGW (empty GFP) or 5 µg pAc-Shrub-GFP, pAGW-ALIX-wt, pAGW-ALIX-F198D or pAGW-ALIX-I211D using FuGene HD according to the manufacturer’s instructions (Promega). Enrichment of mitotic cells was obtained by MG132 treatment (25 µM, 5 hours) as previously described [76]. Cells were used for GFP trap immunoprecipitation analysis performed in line with the protocol provided by the supplier (ChromoTek). The cells were lysed in 200 µl Lysis buffer (10 mM Tris-HCl pH 7.5, 150 mM NaCl, 0.5 mM EDTA, 0.5% NP-40) supplemented with 1:50 protease inhibitor cocktail (Roche), 1:50 phosphatase inhibitor cocktail 2 (Sigma-Aldrich) and 2 mM N-ethylmalemide (Sigma-Aldrich) on ice for 30 minutes with extensive mixing every 10 minutes. Nuclei and cell debris were cleared by centrifugation (20,000g, 10 minutes, 4°C), before the lysate was diluted to 1000 µl with Washing buffer (10 mM Tris-HCl pH = 7.5, 150 mM NaCl, 0.5 mM EDTA) and incubated with pre-washed GFP trap beads (30 µl) for 1 hour at 4°C. The beads and associated proteins were washed three times using Washing buffer and next boiled in SDS sample buffer containing 100 mM DTT for 10 minutes to elute associated proteins. The eluted proteins were subjected to SDS-PAGE, followed by Western blot to detect ALIX, Shrub or GFP.
Drosophila tissues were collected and homogenized in ice-cold lysis buffer (50 mM Tris pH 8, 150 mM NaCl, 0.5% NP-40 or 50 mM Hepes, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 10% glycerol, 1% Triton X-100, 25 mM NaF, 10μM ZnCl2) containing protease inhibitor cocktail (Complete, EDTA-free, Roche). Lysates were cleared by centrifugation for 15 min at 13,000 rpm and 4°C. Equal amounts of protein were mixed with Laemmli buffer containing 50 mM DTT, denatured by boiling and subjected to SDS-PAGE and transferred to either nitrocellulose or PVDF membranes. Nitrocellulose membranes were blocked in PBS/5% milk at 4°C over night followed by incubation with primary antibodies diluted in PBS/5% BSA for 1h 30 min or over night. Membranes were then washed three times in PBS/0.01% Tween-20, followed by incubation 1 h with secondary HRP-conjugated anti-rabbit and anti-mouse antibodies (1:5000) (Jackson ImmunoResearch). Following three further washes in PBS/0.01% Tween-20 and one wash in PBS, chemiluminescent (WestPico, PIERCE) signal was detected on film (Amersham Hyperfilm). PVDF membranes were blocked (by drying), re-wet in PBS/0.01% Tween-20, incubated with primary antibodies overnight at 4°C, rinsed three times in PBS/0.01% Tween-20, incubated with fluorescently labelled secondary antibodies (LI-COR Biosciences GmbH) and washed twice in PBS/0.01% Tween-20 and once in PBS followed by scanning using the Odyssey Developer (LI-COR Biosciences GmbH).
For RNAi-mediated gene silencing in germ cells, MTD-GAL4, Nanos-GAL4, UAS-Dicer; Nanos-Gal4 or Bam-GAL4 drivers were crossed to control (yv; attP2), TRiP-alix-RNAi, TRiP-shrub-RNAi or TRiP-shrub-RNAi; TRiP-alix-RNAi flies as described. For all RNAi experiments, young female offspring were fed with yeast paste and kept with a couple of males for 2 days at 25°C. Ovaries of 2–4-day-old females were dissected, fixed, and stained as described above.
0–3 day old males were dissected and stained with antibodies as described above to label midbody rings and midbodies (Cindr), the fusome (α-spectrin), hub (Fasciclin III) and germ cells (Vasa). Confocal z-stacks of testes tips were acquired at the confocal microscope and mGSC phenotypes were analyzed from z-stacks and three-dimensional reconstructions of z-stacks. mGSC identity was determined based on proximity to the hub. Phenotype scoring was based on the fusome morphology, presence, absence, number and position of Cindr-positive MRs/MBs, Vasa staining and nuclei.
For clonal analysis in the follicle cell epithelium, MARCM82 females were crossed to FRT82 and FRT82, alix3/TM6B, Tb and FRT82, alix1/TM6B, Tb males. L3 larvae were subjected to two heat-shocks at 37°C for 1 h. Newly hatched females were fed with yeast paste for 2 days in the presence of a couple of males. Ovaries were then dissected and stained to visualize F-actin and nuclei (Hoechst) as described above.
Genomic rescue constructs (BAC CH322–119C06, comprising 20339 bp from 23513227 to 23533565 of chromosome arm 3R (short-alix-rescue, alix-s), and BAC CH321–50C24 comprising 85562 bp from 23500943 to 23586504 of chromosome arm 3R (long-alix-rescue, alix-l) in the vector attB-P[acman]-CmR-BW (http://bacpac.chori.org/home.htm) were injected into strains y1 w1118; PBac{y+-attP-9A}VK00018 (BDSC# 9736, insertion site 53B2) and y1 w1118; PBac{y+-attP-3B}VK00037 (BDSC# 9752, insertion site 22A3) and integrated into predetermined attP docking sites in the genome using PhiC31 integrase-mediated germline transformation. The methodology is decribed in “Versatile P[acman] BAC libraries for transgenesis studies in Drosophila melanogaster” [77]. The injection of the constructs into Drosophila embryos was performed by BestGene (http://www.thebestgene.com/). Males with integrated constructs were obtained from BestGene, balanced and crossed to generate CH322–119C06/CyO; alix1/TM6B, Tb (alix-s/CyO; alix1/TM6B, Tb), CH322–119C06/CyO; alix3/TM6B, Tb (alix-s/CyO; alix3/TM6B, Tb), CH321–50C24/CyO; alix1/TM6B, Tb (alix-l/CyO; alix1/TM6B, Tb) and CH321–50C24/CyO; alix3/TM6B, Tb (alix-l/CyO; alix3/TM6B, Tb) stocks.
For complementation tests the alix1 and alix3 alleles were crossed to the deficiencies and to each other. In both rescue analyses and complementation tests, young females of the indicated genotypes were collected, fed with yeast paste and kept with a couple of males for 2 days. Ovaries of 2–4 day-old flies were dissected, fixed and stained to visualize F-actin and nuclei and egg chamber phenotypes were quantified as described above.
Flies used for fertility tests were 4–7 days old and kept separately with yeast paste for a couple of days before being crossed. Wild type or alix1 mutant virgin females were crossed to wild type or alix1 mutant males as indicated. Eggs were collected on apple juice agar plates for 18 hours three times for each cross in three independent experiments. The eggs were counted after each egg lay to determine the egg lay rate. Hatch rates were determined by quantifying the hatched versus unhatched eggs under a dissecting microscope after eggs had developed for 24–30 hours. The experiments were conducted at 25°C.
Embryo collection, permeabilization and fixation were based on the protocol described by Rothwell and Sullivan [78]. The Drosophila melanogaster flies were put on apple juice agar with yeast for egg lay at 25°C overnight. The embryos were dislodged from the agar into a nylon mesh/falcon basket using PBS + 0.02% Triton X-100, and dechorionated by shaking them in a 50% commercial bleach solution until agglutination of the embryos (1–3 min). The dechorionated embryos were extensively rinsed with PBS + 0.02% Triton X-100, and blotted dry on paper towels. The embryos were transferred from the nylon mesh and to a small flask with 5 mL heptane. An equal amount of 4% formaldehyde in PBS was added, and the two-phase mixture was incubated with vigorous shaking for 17 minutes. The embryos were now between the two phases. The formaldehyde phase was removed and replaced with methanol, and the embryos were gently shaken for 1 minute with gentle heating for removal of the vitelline membrane. The heptane phase was removed along with the embryos still remaining in the interphase. The embryos that sank to the bottom of the flask were washed three times in methanol and stored at -20°C. Immunofluorescent staining of embryos was performed as follows. The embryos were rehydrated by first putting them in 3:4 methanol and 1:4 4% formaldehyde in PBS for 2 minutes, and then 1:4 methanol and 3:4 formaldehyde for 5 minutes. Post fixation was done for 10 minutes in 4% formaldehyde, before the embryos were rinsed six times using PBS with 1% BSA and 0.05% Triton X-100. The embryos were incubated with α-spectrin antibodies (1:25, DSHB) over night at 4°C. After incubation the embryos were rinsed three times and washed for one hour with PBS with 1% BSA and 0.05% Triton X-100, and then incubated with secondary antibody for two hours. The antibodies were diluted in PBS with 1% BSA and 0.05% Triton X-100. The embryos were again rinsed three times and washed for one hour with PBS with 1% BSA and 0.05% Triton X-100, before being labeled with Hoechst 33342 (2 µL/mL) for 10 minutes, and then rinsed 3 times in PBS to remove detergent. The embryos were mounted using Vectashield (Vector laboratories). For quantifications of mono- and binucleate cells, images of homozygous wild type, alix1 and alix3 mutant stage 16 embryos were captured at the confocal microscope and more than 1000 cells analyzed for each genotype.
|
10.1371/journal.ppat.1002494 | Development of Functional and Molecular Correlates of Vaccine-Induced Protection for a Model Intracellular Pathogen, F. tularensis LVS | In contrast with common human infections for which vaccine efficacy can be evaluated directly in field studies, alternative strategies are needed to evaluate efficacy for slowly developing or sporadic diseases like tularemia. For diseases such as these caused by intracellular bacteria, serological measures of antibodies are generally not predictive. Here, we used vaccines varying in efficacy to explore development of clinically useful correlates of protection for intracellular bacteria, using Francisella tularensis as an experimental model. F. tularensis is an intracellular bacterium classified as Category A bioterrorism agent which causes tularemia. The primary vaccine candidate in the U.S., called Live Vaccine Strain (LVS), has been the subject of ongoing clinical studies; however, safety and efficacy are not well established, and LVS is not licensed by the U.S. FDA. Using a mouse model, we compared the in vivo efficacy of a panel of qualitatively different Francisella vaccine candidates, the in vitro functional activity of immune lymphocytes derived from vaccinated mice, and relative gene expression in immune lymphocytes. Integrated analyses showed that the hierarchy of protection in vivo engendered by qualitatively different vaccines was reflected by the degree of lymphocytes' in vitro activity in controlling the intramacrophage growth of Francisella. Thus, this assay may be a functional correlate. Further, the strength of protection was significantly related to the degree of up-regulation of expression of a panel of genes in cells recovered from the assay. These included IFN-γ, IL-6, IL-12Rβ2, T-bet, SOCS-1, and IL-18bp. Taken together, the results indicate that an in vitro assay that detects control of bacterial growth, and/or a selected panel of mediators, may ultimately be developed to predict the outcome of vaccine efficacy and to complement clinical trials. The overall approach may be applicable to intracellular pathogens in general.
| Diseases such as tuberculosis (caused by Mycobacterium tuberculosis) or tularemia (caused by Francisella tularensis) result from infections by microbes that live within cells of a person's body. New vaccines are being developed against such intracellular pathogens, but some will be difficult to test, because disease takes a long time to develop (e.g., tuberculosis) or because outbreaks are unpredictable (e.g., tularemia). Usually such infections are controlled by activities of T cells. However, there are no accepted measures of T cell function that reliably predict vaccine-induced protection. We studied two new ways to do so. We used a group of vaccine candidates against tularemia that stimulated good, fair, or poor protection of mice against Francisella challenge. We then measured whether Francisella–immune cells from vaccinated mice controlled the growth of bacteria inside cells, and/or whether the expression of immune genes in Francisella–immune cells was increased. We found that the degree of protection was matched by the degree of the cells' function in controlling intramacrophage bacterial growth. Further, the degree was predicted by relative amounts of gene expression for several immune mediators. Thus the two new options explored here may help predict protection, without waiting for the onset of disease.
| Most vaccines against infectious diseases in clinical use today act by stimulating the production of antibodies, which block virus entry, neutralize toxins, or otherwise limit infection through a variety of mechanisms. Measurements of serum antibodies have therefore been applied to predict successful vaccine-induced protection against diseases such as rabies, tetanus, and diphtheria [1]. In contrast, cell-based immune responses provided by T lymphocytes may be more important for control of intracellular pathogens. To date, however, no predictive correlates have been established for any intracellular pathogen. Understanding T cell effector functions that control intracellular infections, and developing clinically useful predictive correlates, would greatly facilitate evaluation of new vaccines for intracellular pathogens of major public health importance such as Mycobacteria, Chlamydia, Salmonella, and Leishmania.
To address these questions, we have exploited experimental infection models that use the Live Vaccine Strain (LVS) of Francisella tularensis, a Gram-negative intracellular bacterium that causes tularemia. Although the incidence of tularemia in the U.S. is low, F. tularensis is a bioterrorism concern due to its high infectivity and mortality rates following pulmonary infection [2]. Antibiotics are effective, but difficulties with diagnosis and with prompt treatment make developing vaccines a priority [3], [4]. However, the sporadic nature of disease likely means that vaccine field trials for efficacy are impractical.
The use of live attenuated Type B Francisella strains as vaccines in the former U.S.S.R. during and after World War II had clear impact on the epidemiology of disease [5]. Successful vaccination of humans using attenuated bacteria has been mimicked experimentally; rabbits, guinea pigs, rats, and mice are all either natural hosts or are susceptible to Francisella, and provide reasonable animal models for vaccination studies [6]. LVS, an attenuated strain derived from F. tularensis subsp. holarctica (Type B) [7], is the only vaccine against tularemia currently undergoing clinical development in the U.S. [2], [4]. Human challenge studies, as well as use among laboratory workers, suggest that LVS vaccination provides at least partial protection against some forms of the disease, but specific efficacy levels have not been firmly established [2]–[5], [8]–[10]. In contrast, observations suggested minimal protection of people following vaccination with killed Francisella despite stimulating production of abundant serum antibodies [3], [8], [11]–[12], similar to experimental studies using mice [13]–[16], especially following aerosol exposure to the most virulent strains.
Examination of broth cultures of all strains of F. tularensis, including LVS, on blood agar plates reveals a variety of colony morphologies. These opacity variants suggest bacterial phase variation, a phenotype that may confound evaluation and use of LVS. Lots of LVS that include a high proportion of phase variants have been associated with reduced immunogenicity in humans [4], [17], as suggested earlier in animal studies [7]. Stable opacity variants of LVS, denoted LVS-G and LVS-R, have been isolated in vitro [18]. These isolates express alternative chemotypes of Francisella LPS, and appear to be analogous to clinical lots with reduced immunogenicity in humans. However, LVS-G and LVS-R have not been tested as vaccines in any experimental models, including mice, to date.
Murine infections with LVS provide a well-established model of infection and immunity against Francisella, and indeed for intracellular bacteria generally [19]–[21]. Similar to many intracellular bacteria, LVS infects and replicates primarily in macrophages [22], but exhibits convenient route-dependent virulence in mice [19]. Thus, LVS administered to mice subcutaneously or intradermally (ID) has a high LD50 of about 106 and establishes a vaccinating infection, eliciting strong cellular as well as humoral immune responses. However, doses of 101 or more of LVS administered to mice intraperitoneally (IP) or intravenously are lethal. BALB/c or C57BL/6 mice vaccinated ID with 104 LVS survive lethal challenge with LVS of up to 106 IP, and are at least partially protected against parenteral challenge with fully virulent Type A F. tularensis SchuS4 [3], [23]–[24]. Also similar to many intracellular pathogens, in vivo studies clearly demonstrate that this protection is dependent on T lymphocytes, and involves production of Interferon gamma (IFN-γ), Tumor Necrosis Factor alpha (TNF-α), and nitric oxide (NO). To further uncover T cell effector mechanisms, we have previously developed an in vitro tissue culture system to mimic in vivo immune responses [25]–[26], in which LVS-immune lymphocytes are co-cultured with LVS-infected bone marrow derived macrophages and intramacrophage bacterial replication is measured. LVS-immune splenocytes, liver leukocytes, and lung leukocytes all control intramacrophage LVS replication in vitro, but naive cells do not [27]. In this assay, appropriate T cell subpopulations but not B cells or myeloid cells have activity, and the model appears to faithfully reflect known in vivo T cell effector mechanisms [25]–[26].
Here, we take advantage of a panel of Francisella vaccine candidates, including LVS, LVS-G, LVS-R, and heat-killed LVS, that induced quantitatively different levels of protection in mice against Francisella challenge. These vaccines were chosen to approximate Francisella vaccines studied in humans; thus, LVS has been associated with reasonable efficacy, while lots of LVS with higher proportions of opacity variants exhibit reduced immunogenicity (modeled here by the stable variants LVS-G and LVS-R), and killed Francisella provided poor protection in humans. Using this panel, we searched for immune responses that predict successful protection. We found that the relative activity of Francisella-immune lymphocytes in vitro in the co-culture assay, as well as the relative expression of a group of immunologically-related genes in cells recovered from this assay, correlated with the degree of protection observed in vivo. This approach to identifying correlates, which couples a functional in vitro assay that detects reduction in intracellular bacterial loads with expression of relevant mediators, may be generally applicable to vaccine-induced protection against intracellular pathogens.
These studies carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All experiments performed using LVS only were conducted under protocols approved by the Animal Care and Use Committee (ACUC) of CBER. Experiments including F. tularensis SchuS4 challenge were performed at the Rocky Mountain Laboratories (RML) under protocols approved by the RML ACUC. Both sets of protocols stressed practices and procedures designed to strictly minimize any suffering.
Six to twelve week old wild type male C57BL/6J mice were purchased from Jackson Laboratories (Bar Harbor, ME). All mice were housed in sterile microisolator cages in a barrier environment at CBER/FDA, fed autoclaved food and water ad libitum, and routinely tested for common murine pathogens by a diagnostic service provided by the Division of Veterinary Services, CBER. Within an experiment, all mice were age matched.
F. tularensis strain SchuS4, provided by Jeannine Peterson, (Centers for Disease Control, Fort Collins, CO); F. tularensis LVS (American Type Culture Collection 29684); F. tularensis LVS-G and LVS-R, originally obtained from Francis Nano (University of Victoria, Victoria, British Columbia, CA); and Listeria monocytogenes strain EGD (ATCC 15313) were all grown to mid-log phase in modified Mueller-Hinton (MH) broth (Difco Laboratories, Detroit, MI), as previously described [28], harvested, and frozen in 1 ml aliquots in broth alone at −70°C. For each bacterial stock, separate experiments determined numbers of live colony forming units (CFU), confirmed typical colony morphologies, and confirmed expected LD50s and times to death using adult male BALB/cByJ mice (which are the most sensitive strain for quality control testing). Bacteria were periodically thawed for use, and viability was quantified by plating serial dilution on modified MH agar plates. Aliquots of F. tularensis LVS were heat killed at 56°C for 30 minutes immediately prior to use, and complete killing confirmed by plate count.
Groups of mice were immunized by intradermal (ID) injection with 1×104 CFU LVS, 1×104 LVS-G, 1×104 LVS-R, or an amount equivalent to 1×108 heat-killed LVS; doses of each were optimized in initial experiments for maximal protection against lethal IP LVS challenge. All vaccines were diluted in 0.1 ml phosphate-buffered saline (PBS) (BioWhittaker, Walkersville, MD) containing <0.01 ng of endotoxin/ml. Actual doses of inoculated bacteria were retrospectively determined by plate count; control groups received 0.1 ml PBS ID. As indicated, four – twelve weeks after vaccination, mice were challenged with 103 – 106 LVS intraperitoneally (IP), or 50 CFU SchuS4 subcutaneously (SC), and monitored for survival. Animals were euthanized when clearly moribund.
Single-cell suspensions of splenocytes were generated for in vitro culture, flow cytometry, and qRT-PCR analysis by standard techniques, and had no detectable CFU at the time of harvest. Viability was assessed by exclusion of trypan blue and flow cytometry (see below).
Co-cultures were performed in 24 well tissue culture plates as described previously [25]–[27], [29]–[31]. Briefly, bone marrow macrophages (BMMØ) were cultured in complete DMEM (DMEM supplemented with 10% heat-inactivated FCS [HyClone, Logan, UT], 10% L-929-conditioned medium, 0.2 mM L-glutamine, 10 mM HEPES buffer, and 0.1 mM nonessential amino acids) in 24 well plates. Confluent adherent macrophage monolayers were infected for 2 hours with F. tularensis LVS at a multiplicity of infection (MOI) of 1∶20 (bacterium-to-BMMØ), washed, treated for 60 min with 50 µg/ml gentamicin, and washed extensively with antibiotic-free medium. Single-cell suspensions of splenic lymphocytes derived from vaccinated mice (5×106/well, or as indicated) were added to confluent LVS-infected macrophages (∼1×107/well) [25]–[26]. At the indicated time points, non-adherent cells were harvested, centrifuged and assessed for changes in cell surface phenotype by flow cytometry or gene expression by qRT-PCR as described below. Supernatants from harvested cells was collected and stored at −70°C until analyzed for nitric oxide and cytokines as described below. Intracellular bacterial loads in adherent macrophages were determined as previously described. Additional macrophages were collected following incubation in 0.05% trypsin/EDTA for 5 minutes, followed by neutralization with complete media containing serum.
Cells to be assessed for changes in gene expression by qRT-PCR were pelleted by centrifugation for 10 minutes at 1000 rpm, immediately immersed in RNAlater (Ambion, Austin, TX) and stored at −70°C until further characterization. Total RNA was extracted from samples using RNeasy mini kits (Qiagen, Valencia, CA), according to the manufacturer's directions. RNA quality and concentration were assessed by Bioanalyzer, including calculation of the RNA Integrity Number (RIN) via a software algorithm that estimates RNA sample integrity from elements in the bioanalyzer electrophoretic trace, and then assigns a score to RNA quality between 0 and 10 (Agilent Technologies, Santa Clara, CA). One microgram of RNA was used to synthesize cDNA using the commercially available kit RetroScript Reverse Transcription for RT-PCR (Ambion, Applied Biosystems Foster City, CA), following the manufacturer's instructions. Semi-quantitative real-time PCR amplification was completed with an ABI Prism 7000 sequence detection system (Applied Biosystems, Carlsbad, CA). For initial screening of genes' expression, cDNA prepared from non-adherent cells was diluted and used to amplify a panel of genes of immunological interest (e.g., Th1-Th2-Th3 RT2 Profiler PCR Array System, SABiosciences, Frederick, MD), following the manufacturer's instructions. To validate the initial array qRT-PCR results, a second series of independent amplifications for selected genes were performed. Independent primers and probes were purchased from Applied Biosystems. cDNA was initially diluted to 100 µl and then two µl of each cDNA was further diluted to a volume of 25 µl PCR mix (Applied Biosystems) containing 0.1 µM and 0.2 µM of each primer and probes, respectively. Serial dilutions of each individual gene were used to generate a Glyceraldehyde phosphate dehydrogenase (GAPDH) standard curve. For PCR amplifications, the initial denaturation at 95°C for 10 minutes were followed by 40 cycles of 95°C for 15 seconds and 60°C for 1 minute. The level of mRNA of each gene relative to the GAPDH mRNA concentrations was calculated by plotting the crossing point (Ct) of each amplification in relationship to the GAPDH standard curve. Delta Ct (ΔCt), and the ratio between ΔCt of vaccines' samples and control samples (ΔΔCt), were then calculated.
Single cell suspensions prepared from spleens and splenocytes recovered from co-culture after the indicated time of culture were stained for a panel of murine cell surface markers and subjected to multiparameter analyses using a Becton-Dickinson LSR II flow cytometer (San Jose, CA) and FlowJo (Tree Star, Inc) software essentially as previously described [29]–[30]. Briefly, cells were washed and resuspended in flow cytometry buffer (PBS/2% serum), and non-specific binding of antibodies was inhibited by blocking Fc receptors with anti-CD16 (Fc Block; BD Pharmingen). To discriminate live from dead cells, a staining step for dead cells was performed using a commercially available kit and following manufacturers' instruction (Live/Dead Staining Kit, Invitrogen). The cells were then washed and stained for cell surface markers. Antibody concentrations were optimized separately for use in seven- to nine-color staining protocols as required, using appropriate fluorochrome-labeled isotype matched control antibodies. The following antibodies were used: anti-B220 (clone RA3-6B2), anti-CD19 (clone 1D3), anti-TCRβ (clone H57-597), anti-CD4 (clone RM4-5), anti-CD8β (H35-17.2), anti-NK1.1 (clone PK136), anti-CD11b (clone M1/70), anti-Gr-1 (clone RB6-8C5), and anti-CD11c (cloneHL3), each labeled with a variety of fluorochromes as needed (above antibodies were purchased from BD Pharmingen).
Kaplan Meier curves were plotted to compare time to death following lethal LVS challenge between different vaccine groups, and log-rank (Mantel-Cox) analyses calculated to compare survival of different groups (using Prism, GraphPad Software, La Jolla, CA). Linear regression models were fit to compare the effects of splenocytes on controlling bacterial growth from different vaccine groups while adjusting for splenocyte concentration. Univariate and multivariate logistic regressions were used to correlate protection against lethal LVS challenge with either fold change in gene expression at two different time points (∼6 weeks after vaccination and ∼12 weeks after vaccination), or with all data combined across both time point using standardized scores of gene expression. The results of these analyses were quite similar, and thus the results using all data combined are shown here. Standardized scores were used to protect against the possibility that the relative magnitude of gene expression for any given gene might be relatively different at the early time point compared to the late time point. Standardized scores were obtained by subtracting the average log expression level and then dividing by the standard deviation of the log expression level in the same experiment. The Akaike information Criterion was used to compare different logistic regression models. Pearson's correlation coefficients of standardized scores of expression level for pairs of genes were reported. All p-values were two-sided, and p-values<0.05 were considered to be statistically significant. Data analyses were conducted using R (R Foundation for Statistical Computing, Vienna, Austria).
Initial studies optimized conditions for these experiments; further characterization and modifications of previously published in vitro co-culture methodologies [25]–[26] were required to be most appropriate for the present purposes. Detailed information regarding data supporting the resulting modifications is provided in Supporting Information (see Supporting Information, Text S1, Figures S1–S3, and Table S1).
To determine whether the in vitro co-culture system may serve as a functional correlate of protection, we first identified a panel of vaccine candidates that provided different degrees of protection against lethal Francisella challenge in vivo. C57BL/6J mice were vaccinated ID with 104 LVS, with 104 of the opacity variants LVS-G or LVS-R, or with 108 -heat killed (HK-) LVS, and then challenged one month after vaccination with increasing lethal doses of LVS IP (Figure 1). All mice vaccinated with wild type LVS survived challenge with up to 5×105 LVS IP, and 75% survived challenge with the highest dose tested, 106 LVS IP. In contrast, mice vaccinated with LVS-G exhibited 90% survival following challenge with 5×105 CFU and 30% following challenge with 1×106 CFU. Mice vaccinated with LVS-R exhibited only 20% survival following challenge with 5×105 CFU, and none survived challenge with 106 CFU. Finally, vaccination with HK-LVS failed to protect against challenge with 5×104 CFU or higher. In later experiments, a single challenge dose of 5×105 - 106 LVS IP was chosen as appropriate for discriminating between the degree of protection provided by all vaccines in the panel.
These data indicated that this panel of candidate vaccines exhibited a hierarchy of relative protection against in vivo lethal LVS challenge, such that LVS>LVS-G>LVS-R>HK-LVS>PBS (naive control). We next examined the protective capacity of these vaccines against lethal parenteral challenge with a selected dose of fully virulent Type A F. tularensis (SchuS4). Similar to the outcome using LVS challenge, and consistent with previous reports [32], approximately 30% of mice vaccinated with LVS survived this dose of SchuS4 challenge and time to death of those that died was greatly extended, while vaccination with LVS-G protected only 10% of mice against lethal SchuS4 challenge (Figure 2). In contrast to challenge with LVS, vaccination with LVS-R did not provide detectable protection against this dose of SchuS4 challenge, and vaccination with HK-LVS only slightly extended time to death. Nonetheless, for this group of vaccines, a similar hierarchy of protective efficacy found following challenge with wild type LVS was also found using challenge with fully virulent Francisella.
In parallel with in vivo vaccination and challenge studies, the activities of splenocytes obtained from mice vaccinated with LVS, LVS-G, LVS-R, HK-LVS, or PBS (control) were compared. To determine the relative effectiveness of each type of primed cells, decreasing numbers of splenocytes were added to a constant number of LVS-infected macrophages. As seen in Figure 3A, on a per-cell basis, cells obtained from LVS-infected mice were most effective in controlling the intramacrophage growth of LVS; those from LVS-G vaccinated mice were less effective, and those from LVS-R-vaccinated mice the least effective. The relationship between relative control by cells from the different groups was then investigated. A linear regression with indicators of different vaccine groups and the cell concentration as covariates was used to compare log CFU of recovered bacteria in different vaccine groups, adjusting for the levels of cell concentrations (Figure S4). The result of regression analysis demonstrated that, for any fixed cell concentration, cells from LVS-G-vaccinated mice were significantly less effective in controlling bacteria growth, a difference of about 0.95 log, compared to those from LVS-vaccinated mice (p<0.001). Similarly, cells from LVS-R-vaccinated mice were about 1.57 log less effective than cells from LVS-vaccinated mice (p<0.001). Finally, cells from LVS-R-vaccinated mice were 0.62 logs less effective than from those from LVS-G-vaccinated mice (p<0.001). As seen previously [25]–[27], [29]–[31], the results suggested that co-cultures containing cells from naive mice exhibited minimal and inconsistent reductions in bacterial numbers compared to those with LVS-infected macrophages alone (e.g., Figure 3A, p = 0.02; Figure 3B, p = 0.35); further, there was no significant relationship between the concentration of naive cells and bacterial numbers. Finally, cells obtained from mice vaccinated with HK-LVS did not significantly inhibit bacterial growth control compared to either naive cells or cultures with LVS-infected macrophages only, even at the highest cell numbers tested (Figure 3B). Thus, the hierarchy of in vitro activities of cells from vaccinated mice was again LVS>LVS-G>LVS-R>HK-LVS. Further, because only LVS-immune T cells are active in this setting, these data estimate the relative frequencies of memory T cells.
Supernatants and cells were also recovered on day 2 from each type of co-culture. Supernatants were then analyzed as above for cytokine production by ELISA and NO production by Griess reaction; cells were characterized by flow cytometry; and mRNA prepared from recovered cells was analyzed for relative gene expression. The amounts of TNF-α, IFN-γ, and NO produced were consistent with previously published studies using LVS vaccination alone [25]–[27], [29]–[31]. Here, relative cytokine and NO production exhibited the same pattern as that observed in the survival studies and in in vitro control of intramacrophage LVS replication, such that LVS>LVS-G>LVS-R>HK-LVS and naive groups (data not shown). Flow cytometry analyses of recovered cells confirmed the relative enrichment of T cells (similar to that illustrated in Table S1), and did not reveal any obvious differences between vaccine groups (data not shown). Collectively, these data indicated that the hierarchy of protective capacity engendered by in vivo vaccination with this panel of vaccines was faithfully reflected by the relative ability of each type of Francisella-immune splenocytes to persist in culture, produce relevant cytokines and nitric oxide, and ultimately to effect control of intramacrophage bacterial growth.
Because these in vitro co-culture conditions reliably detected differences in vaccine quality, non-adherent immune splenocytes from all groups were recovered on day two from each co-culture, and then analyzed in detail for relative gene expression. For these experiments, groups of mice were vaccinated with LVS, LVS-G, LVS-R and HK-LVS. At either 6 weeks or 12 weeks after vaccination, some mice were challenged in vivo, and other mice were sacrificed at the same time to prepare splenocytes, perform in vitro co-cultures, and recover non-adherent cells for mRNA analyses. Similar to initial studies using cells from naive or LVS-vaccinated mice only (see Supporting Information text), the relative mRNA expression of genes of immunologic interest in splenocytes from mice vaccinated with LVS, LVS-G, LVS-R, or HK-LVS was compared to that of splenocytes from naive mice by RT-PCR, using commercially available arrays that included immunologically-related genes (e.g., Profiler PCR Th1-Th2-Th3 array). Data generated from initial experiments using the complete vaccine panel revealed that some genes, such as GF1 and CCR4, which in initial studies were up-regulated in LVS-primed cells compared to naive cells, were either inconsistent or up-regulated to a similar degree for all vaccines and did not exhibit a differential pattern (Table S2, “SYBR”). In contrast, other genes appeared to exhibit a range of expression that reflected the relative effectiveness of vaccination. For example, IFN-γ appeared to be highly expressed in LVS-primed cells as well as in cells from LVS-G-vaccinated mice, only moderately expressed in cells from LVS-R-vaccinated mice, and expressed very little in cells from HK-LVS-vaccinated (all compared to naive mice). Similar to IFN-γ, the relative expression of several other genes, such as IL-6, TNF-α, IL-18bp, and GM-CSF, reflected the relative level of both in vivo protection and in vitro bacterial growth control activities of the different vaccines.
The initial comparisons focused attention on a group of 15 genes with apparently differential expression patterns, either in terms of relative up-regulation or relative down-regulation. Seven other genes were also considered of ongoing interest, either because they were not included on the commercial panel used and had a known biological relationship to other genes that were correlated with protection (e.g., IL-22), and/or because results were ambiguous (e.g., IL-17A and IL-13). A set of 22 genes were therefore selected for more detailed studies, and separate primer-probe sets prepared to perform qRT-PCR analyses. This approach was applied to again analyze mRNA from previous experiments, as well as mRNA from additional independent experiments. Collectively, these included two independent experiments using splenocytes from mice vaccinated six weeks earlier (Table S2, experiments 4 and 6) and two independent experiments using mice vaccinated twelve weeks earlier (Table S2, experiments 5 and 7). For analysis of qRT-PCR data from selected genes, Ct results were normalized to a standard curve of GAPDH before calculation of ΔCt and fold changes (ΔΔCt) between samples from naive and vaccinated mice (ΔΔCt). When compared for up-regulated genes, the fold change for most of these genes in cells from LVS-vaccinated mice was similar to those in cells from LVS-G-vaccinated mice; both were greater than those in cells from LVS-R vaccinated mice, which in turn were greater than those in cells from HK-LVS vaccinated mice or naive mice (Figure 4, panels A–C). In contrast, when the relative fold change (ΔΔCt) of putatively down-regulated genes was compared across different experiments, the amounts of differences were relatively modest (Figure 4, panel D), similar to observations made using Profiler PCR arrays. Taken together, using a different detection system (including probes instead of SYBR Green) and a different normalization approach, we again found clear relationships between the relative expression levels of most members of this panel of up-regulated genes and the hierarchy of vaccine protection. Further, the latter method allowed more accurate quantification of relative gene expression.
To examine whether gene expression patterns observed are specific for F. tularensis LVS vaccination and related to Francisella vaccine efficacy, C57BL/6J mice were vaccinated ID with LVS, LVS-R, HK-LVS, or 5×103 Listeria monocytogenes. Two weeks later, mice were either challenged with lethal dose of 106 LVS IP, or used to prepare spleen cells that were co-cultured with LVS-infected macrophages, recovered after two days, and analyzed for relative gene expression. For the LVS related vaccines, the same patterns of survival and in vitro growth control were found (data not shown). In contrast, Listeria-vaccinated splenocytes did not significantly reduce the intramacrophage growth of LVS compared to control co-cultures (data not shown; and see [25]), and Listeria-vaccinated splenocytes exhibited an absence of up- or down-regulation. The fold change (ΔΔCt) values using Listeria-vaccinated splenocytes were mostly similar to those observed in cells from mice vaccinated with HK-LVS. For example, IFN-γ up-regulation in cells from LVS-vaccinated mice was about 22 fold compared to naive mice; 7.2 in cells from LVS-R-vaccinated mice; 0.8 in cells from HK-LVS vaccinated mice; and 1.7 in cells from Listeria-vaccinated mice. Thus, the working panel of genes specifically reflected activities of Francisella-immune cells.
The resulting data derived from both the Profiler arrays and from the selected genes assessed by qRT-PCR were used to examine statistical correlations between survival and in vitro gene expression in response to the different vaccines (see Table S2, all experiments). For each gene, the proportion of surviving mice was plotted against a standardized score of gene expression. Examples of these relationships are illustrated in Figure 5, in which analyses of IFN-γ, TNF-α, IL-6, and T-bet relative mRNA expression by qRT-PCR are shown; p values for the relationship between survival and all genes analyzed by Profiler array or qRT-PCR are provided as Supporting Information (Tables S3 and S4, respectively). Of the 22 selected genes, 20 were significantly related to survival according to Profiler array data, validating the initial selection, but only 16, were significantly related to survival using qRT-PCR data.
The logistic regressions analyses using data from the Profiler arrays (Table S2, SYBR) indicated that about 20 other genes, in addition to the 22 genes already selected, were also significantly related to survival. These included genes that were up-regulated (e.g., SOCS3 and Tmed) and some that were down-regulated (e.g., Jak1 and CD27). However, the fold change of each of these genes even for LVS-immune cells compared to naive cells was relatively small, <2 for up-regulated genes and >0.5 for down-regulated genes. Moreover, the range of differences for these genes across all vaccines was judged too small to be reliably useful in this context, and these candidates have not yet been pursued further.
Taken together, the genes that exhibited consistent changes of useful magnitude across all experiments, as well as significant pairwise correlations of the relative degree of expression between all possible pairs of genes (Table S5), included IFN-γ, IL-6, IL-12rβ2, T-bet (Tbx21), Socs1, and IL18bp (Table 1, Group 1; Figure 4, panel A). Genes with notable, but less universal, associations were GM-CSF, IL-27, TNF-α, IL-27, and Irf1 (Table 1, Group 2; Figure 4, panel B; remaining up-regulated genes are illustrated in Figure 4, panel C, and down-regulated genes in Figure 4, panel D).
Finally, to assess whether fold changes in mRNA expression for different genes would work in concert and better predict survival than single genes, multivariate logistic regression analyses were performed. Although specific pairs of genes displayed correlative degrees of expression, multivariate analyses of gene pairs (Table S6) did not provide stronger associations with survival than any one gene alone, based on the Akaike information criterion value. Likewise, groups of three genes did not provide any statistically significant improvement in survival prediction than did the models with one or two genes (data not shown).
Currently there are no validated options for predicting protection against intracellular pathogens. Human clinical field trials for many intracellular pathogens will be difficult, either because of the long time to develop disease (i.e., tuberculosis), or the sporadic nature of disease in nature (i.e., tularemia). A recent FDA regulation provides an option for evaluating vaccine efficacy using animal studies under special, well-defined circumstances that may most likely be applicable to biodefense pathogens [33]; but a rational means of bridging efficacy between animals and humans and extrapolating vaccine dose will be critical. Many such issues could potentially be addressed by derivation of correlates of protection that can be measured in several species. The word “correlate” has been ascribed a wide variety of definitions [1], [34]; here, we use the term to refer to a measurement that detects relevant biological functions critical for, and statistically related, to protection against an infectious disease [35].
Historically, efforts to identify and measure relevant serum antibodies have failed to successfully predict vaccine-induced protection, particularly for replicating, live attenuated vaccine. Another approach is to relate the quantity of an immune parameter to the degree of protection. Production of IFN-γ ex vivo has been extensively explored, particularly in studies of M. tuberculosis. However, there are many human clinical and experimental examples where the relative levels of IFN-γ measured do not reflect the degree of successful vaccination [36]–[41]. The collective evidence instead indicates that it is likely that local availability of IFN-γ is necessary, but not sufficient, for protection. More recently, “multi-functional T cells” that exhibit the ability to simultaneously produce IFN-γ, TNF-α, and IL-2, have been proposed as vaccine correlates [42]–[43]. While promising in the context of mouse models of Leishmania infections [44] and some studies of tuberculosis vaccines [45]–[47], in other cases no obvious correlation has been detected between MFCs and protection [48]–[51]. Efforts to develop genomic and metabolomic biomarker signatures for tuberculosis infection and vaccination, particularly in the context of HIV/AIDS, are underway [52]. Although still complex, the approach illustrated here of coupling a panel of qualitatively different vaccines, combined with in vitro re-stimulation via co-cultures and semi-quantitative mRNA analyses, clearly identified mediators that correlate strongly with the relative degree of vaccine-induced protection against lethal challenge in vivo.
The Francisella infection and immunity model offered the advantage of having a panel of different vaccine candidates, coupled with the ability to more precisely define the strength of protection in vivo by using a range of lethal challenge doses. Vaccination of humans with LVS engenders production of Francisella-specific serum antibodies, as well as memory T cells in peripheral blood that produce IFN-γ, IL-17A, and IL-22 following antigen stimulation [53]–[56], but these are currently of unknown contribution to protection of people. To date, there are limited studies regarding correlates of immunity to Francisella. Vaccination mice with static Francisella vaccine candidates, including outer membrane protein preparations or ethanol-inactivated LVS formulated with Freund's adjuvant, provided partial protection against respiratory challenge with 40 CFU of fully virulent type A F. tularensis, accompanied by production of serum antibodies and large levels of TNF-α and IL-2 in sera of vaccinated mice after challenge [57]. In studies using a mouse model to compare intradermal vaccination of mice with LVS to vaccination with genetically attenuated mutants of SchuS4, protection against challenge with fully virulent F. tularensis was not correlated with levels of serum IFN-γ or IgM/IgG antibodies [58]; only pulmonary IL-17 quantities after secondary challenge appeared to track with protection [59]. Here, LVS-G and LVS-R, spontaneous variants that express alternate LPS chemotypes, proved to provide intermediate levels of in vivo protection (Figure 1). Because antibodies to LPS likely play a minor role in protection against lethal Francisella challenge even in mice [60]–[61], the reduced protection is unlikely to be explained completely by reduced serological responses. More likely, reduced protection is explained either by reduced persistence and total antigen exposure in these serum-sensitive variants [18]; or, changes in LPS expression are a visible marker for simultaneous changes in expression of other bacterial genes that are important in protection. Of note, it is likely the mechanisms of protection provided by the live attenuated strains LVS, LVS-G, and LVS-R are similar, but the strengths of protection quantitatively different; we consider this an important feature that is critical to permitting strong interpretations across different vaccines.
Using this panel, a hierarchy of strength of protection was evident using Francisella LVS challenge of mice (Figure 1). Although it was not feasible to perform larger experiments using graded doses of challenge with fully virulent F. tularensis SchuS4, the proportion of vaccinated survivors and differences in times to death following challenge with one selected dose supported a similar hierarchy (Figure 2). Tangentially, these comparisons also suggest that lethal parenteral challenge of vaccinated mice with LVS could serve as an informative screen for vaccine efficacy prior to testing by challenge with fully virulent Type A F. tularensis. Using carefully selected conditions, we then compared the relative activity of splenocytes from differentially vaccinated mice in an in vitro tissue culture system that measures reduction of intramacrophage bacterial numbers by immune T cells, and found that the relative strength of in vivo protection was clearly reflected by the relative activity of immune splenocytes in vitro (Figure 3). The results therefore support the utility of the co-culture assay as both a relevant functional assay in its own right.
Further, despite multiple attempts to identify strong T cell antigens involved in murine responses to Francisella and develop associated reagents [62]–[64], tools such as tetramers remain lacking. Studies in mice and humans suggest that host responses do not involve a classical “immunodominant” protein but are directed to a large collection of protein antigens [65]–[67]; thus although tetramer analyses may no doubt eventually prove helpful, such approaches may detect only a small fraction of the total anti-Francisella T cell response. Because only LVS-immune are active in specifically controlling intramacrophage growth of the homologous bacteria [25]–[26], the results presented here validate that the in vitro co-culture assay is a new, and currently the only available, approach to establish the relative frequency of Francisella-specific memory T cells in a mixed population (Figure 3, Figure S4).
The in vitro co-culture assay was previously developed using both Francisella LVS and M. tuberculosis as a research model to explore mechanisms of interactions between infected host macrophages and immune T lymphocytes. In many respects, this system faithfully reflects known in vivo T cell effector mechanisms, including both IFN-γ-dependent and non-IFN-dependent mechanisms [25]–[27], [29]–[31]. However, it should be noted that studies demonstrated that T cells from LVS-vaccinated IL-12 knockout mice are quite active in co-cultures, despite the fact that IL-12 knockout mice do not clear a vaccinating LVS infection [31]. Elsewhere, this co-culture approach was recently applied by our colleagues to murine studies of a panel of vaccine candidates for M. tuberculosis [68]–[69]. In those studies, the in vitro assay successfully discriminated between vaccines with high or moderate activity, as defined by in vivo protection.
Although we find the in vitro co-culture approach promising, as well as potentially feasible in the near term, cell-based assays are difficult to implement for human clinical trials. We therefore pursued an additional strategy, by searching for genes whose differential expression was related to the hierarchy of vaccine-induced in vivo protection and in vitro cellular activity. We focused on screening immunologically-related genes. This approach is obviously biased toward analyzing known entities instead of discovering new ones, but it offered the potential advantage of identifying relevant mediators. Remarkably, cells recovered from in vitro culture differentially expressed mediators at the mRNA level (Figure 4; Figure 5; Table S2) and a number of candidates emerged (Figure 4; Table 1). Of note, among these mediators, were IFN-γ and TNF-α, as might be expected, and thus validating the overall approach.
It should be noted that we observed considerable variability in mRNA levels between experiments. For example, expression of IFN-γ was always highly up-regulated in cells obtained from mice vaccinated with LVS and LVS-G, but the fold change compared to naive cells varied between about 13 and 360 (Table 1; Table S2). We suspect that biological, in addition to technical, reasons contribute to the observed variability. To increase confidence in predictors, quantifying a panel of genes is therefore likely to be preferable over assessing a single gene mediator, even an important one such as IFN-γ. This point may be especially germane to clinical settings that lack the advantages offered by using genetically identical inbred mice.
Despite the quantitative variability, from a group of about 84 genes, 16 proved to be robust enough to yield significant correlations between the magnitude of mRNA expression and survival, as well as exhibit relatively large differences in the degree of expression (Figure 5; Table 1; Figure S3; Tables S3 – S6). These likely include those whose gene products contribute directly to mechanisms, and those that are co-regulated and only epiphenomena. We are most interested in those that are mechanistically relevant, and thus likely to serve as definitive predictors across variables such as time after vaccination, route, dose, tissues sampled, and especially different animal species. Notably, the most useful of the expression differences involved up-regulated genes, which is appealing in potentially reflecting a requirement for production of a mediator to provide a particular function during challenge. It is striking that several of the leading candidates are plausibly related to Th1 cell biology, including T-bet and IFN-γ. Although TNF-α also exhibited significant differential regulation and is clearly relevant, production of TNF-α is tightly regulated to avoid toxicity, and thus ex vivo measurements may not be among the most useful. IL-12rβ2 is only found as part of the complete receptor for IL-12 p70, which is not expressed on resting T cells but induced by T cell activation and contributes directly to Th1 lineage commitment [70]. For example, in naive transgenic CD4+ T cells, IFN-γ stimulation up-regulates expression of T-bet in a STAT-1 dependent manner and promotes IL-12Rβ2 chain expression [71]. Notably, STAT-1 was also among our candidate genes with significant associations, albeit one that exhibited considerable variability and thus was not included in our two highest priority groups (Table 1).
In contrast to IFN-γ, T-bet, and IL-12rβ2, other high priority candidates such as IL-6, SOCS-1, and IL-18bp were more surprising. IL-6 has a wide variety of sources and functions, but in adaptive immunity is most commonly associated with promoting B cell activation and IgA production and infrequently with resistance to intracellular pathogens [72]. In the Francisella infection model, our preliminary results indicate that IL-6 knockout mice are severely impaired in their ability to survive primary LVS vaccination (Kurtz and Elkins, manuscript in preparation), as are T-bet knockout mice and IL-12rβ2 knockout mice (Melillo and Elkins, manuscript in preparation). The specific contribution of SOCS-1, an important member of a large family of “suppressor of cytokine signaling” mediators that regulate T cell differentiation as well as T cell effector functions, awaits further study. Perhaps the most unexpected candidate is IL-18 binding protein (IL-18bp); although induced by IFN-γ, its production is usually associated with cells other than leukocytes [73], and to our knowledge has no reported direct link to Th1 T cell effector functions.
Taken together, the results presented here are in important step toward the identification of T cell functions and products required for survival of lethal exposure of intracellular bacteria. We propose that the candidates described as Group 1 and Group 2 (Table 1) receive high priority for detailed direct exploration, initially in animal studies, of biological relevance and mechanistic contribution. Knowledge obtained by in vitro and pre clinical studies will be the key to facilitating design of assays with formats amenable to clinical studies, such as ex vivo re-stimulation of human peripheral blood leukocytes. It is likely that defining groups of mediators will be preferable in order to overcome issues related to variability in measurements, and ensure predictive confidence for human clinical trials. The larger goal will therefore be to establish protective levels of each individual mediator, and thus select combinations that reliably predict successful vaccination against intracellular pathogens.
|
10.1371/journal.pntd.0003547 | Economic and Disease Burden of Dengue in Mexico | Dengue imposes a substantial economic and disease burden in most tropical and subtropical countries. Dengue incidence and severity have dramatically increased in Mexico during the past decades. Having objective and comparable estimates of the economic burden of dengue is essential to inform health policy, increase disease awareness, and assess the impact of dengue prevention and control technologies.
We estimated the annual economic and disease burden of dengue in Mexico for the years 2010–2011. We merged multiple data sources, including a prospective cohort study; patient interviews and macro-costing from major hospitals; surveillance, budget, and health data from the Ministry of Health; WHO cost estimates; and available literature. We conducted a probabilistic sensitivity analysis using Monte Carlo simulations to derive 95% certainty levels (CL) for our estimates. Results suggest that Mexico had about 139,000 (95%CL: 128,000–253,000) symptomatic and 119 (95%CL: 75–171) fatal dengue episodes annually on average (2010–2011), compared to an average of 30,941 symptomatic and 59 fatal dengue episodes reported. The annual cost, including surveillance and vector control, was US$170 (95%CL: 151–292) million, or $1.56 (95%CL: 1.38–2.68) per capita, comparable to other countries in the region. Of this, $87 (95%CL: 87–209) million or $0.80 per capita (95%CL: 0.62–1.12) corresponds to illness. Annual disease burden averaged 65 (95%CL: 36–99) disability-adjusted life years (DALYs) per million population. Inclusion of long-term sequelae, co-morbidities, impact on tourism, and health system disruption during outbreaks would further increase estimated economic and disease burden.
With this study, Mexico joins Panama, Puerto Rico, Nicaragua, and Thailand as the only countries or areas worldwide with comprehensive (illness and preventive) empirical estimates of dengue burden. Burden varies annually; during an outbreak, dengue burden may be significantly higher than that of the pre-vaccine level of rotavirus diarrhea. In sum, Mexico’s potential economic benefits from dengue control would be substantial.
| During the past decades, dengue fever has become the most common arthropod-borne viral disease, imposing a substantial economic and disease burden in most tropical and subtropical countries, including Mexico. Dengue incidence and severity have dramatically increased in Mexico, with transmission regularly reported in 28 of 32 states. Objective estimates of the burden of dengue are important to inform policy decisions and priorities. We merged multiple data sources to estimate (i) total episodes, (ii) costs per episode, (iii) surveillance and vector control costs, and (iv) disease burden (2010–2011). Results suggest that Mexico had about 139,000 symptomatic and 119 fatal dengue episodes per year on average. The annual cost, including surveillance and vector control, was about US$170 million, or $1.56 per capita, comparable to other countries in the Americas. Annual disease burden averaged 65 disability-adjusted life years per million population, with most of the years lost to disability corresponding to ambulatory episodes. The results show a substantial burden of dengue on the health care system and economy of Mexico. This quantification of the economic burden should help public health officials make informed decisions about current and promising new preventive and control measures to reduce dengue infections.
| Dengue fever is the most important arthropod-borne viral disease affecting humans, with about half the world’s population estimated to be at risk of infection, and epidemics increasing in frequency, magnitude, and geographical reach [1–4]. Dengue imposes a substantial economic and disease burden in most tropical and subtropical countries. Mexico is no exception [5]. Dengue is hyperendemic in Mexico, with all four dengue virus (DENV) serotypes isolated in the country, high levels of disease and an increasing impact during the last decades [5–8]. Transmission of dengue is regularly reported in 28 of the 32 Mexican states; the main mosquito vector, Aedes aegypti, has been reported in 30 states [6,8,9]. The severity of dengue episodes has also steadily increased, with a substantial increase in severe dengue episodes since 1995, although case fatality rate has remained relatively low compared to other Latin American countries [7,10,11].
Objective, comparable measures of the burden of dengue are important to inform decisions about health policy, research, and health service priorities and to increase scientific and social awareness of the disease [12–15]. Despite the need for timely and reliable epidemiological data, dengue burden estimates are sparse. The total burden imposed by a disease includes the illness or disease burden, which measures the impact of a disease on morbidity and mortality in a specific population, and the economic burden [16,17], which includes the cost of illness, prevention and monitoring or surveillance strategies, and other economic impacts (e.g., decrease in travel, seasonal overload of health systems) [18,19]. Because dengue is a reportable illness in most endemic countries, an initial approximation of the total number of dengue episodes in a year is simply the total episodes reported to the country’s Ministry of Health (MoH) through surveillance systems.
Dengue is a reportable disease in Mexico; the MoH has promulgated protocols for laboratory confirmation and collects and disseminates weekly surveillance data [20]. The MoH is responsible for setting national guidelines, rules, and procedures that the 32 state health departments need to follow, although state and local health services are responsible for daily operations. Vector control and dengue surveillance systems guidelines are defined by the MoH at the federal level, although it collaborates with the 32 state health services and other health organizations including Mexican Institute of Social Security (IMSS), Institute of Social Security and Services for State Workers (ISSSTE), Mexican Petroleum (PEMEX), and the Armed Forces medical services [9,20]. A sample of patients with suspected DENV infection is diagnosed by a public health laboratory network (all probable patients in areas with no recent dengue episodes or during low transmission periods and about 30% of patients when there is evidence of transmission and during outbreaks) using confirmatory assays (NS1, IgM, or IgG ELISA), and a subset of these samples is analyzed for virus isolation (10% of the positive samples) [8,9,21]. Patients with probable and confirmed dengue have to be reported weekly, while probable or confirmed DHF and dengue-related deaths must be reported within 24 hours [20]. The MoH estimates the number of dengue episodes in two steps: probable cases are first multiplied by the proportion of positive cases from the lab-diagnosed sample (called possible cases), and then added to the total lab-confirmed cases. The MoH assumes that all episodes are notified [8].
However, passive surveillance systems have limitations. Passive surveillance systems are adequate for monitoring general trends in DENV infections; however, they usually underreport the total episodes of symptomatic dengue [22–26]. Febrile DENV infections with relatively mild symptoms have very low reporting ratios (number of reported dengue episodes / total dengue episodes in the population), and reporting increases with severity [27,28]. Other limitations in passive surveillance systems, even in well-funded systems such as Mexico or Puerto Rico, include misdiagnosis due to limited sensitivity of diagnostic tests, cost constraints, unrecognized dengue symptoms, variation in reporting ratios by severity of symptoms, and differences in diagnosis between epidemic and non-epidemic years [22,23,29–32]. Some health-seeking behaviors also reduce reporting ratios, such as symptomatic patients visiting alternative health providers, including pharmacies or local healers, or simply staying at home. In Mexico, there is little or no reporting from the private sector [33], and there is wide variation in the quality of reporting of notified cases. Limited reporting of symptomatic DENV infections leads to conservative estimates of economic and disease burden [24–26,34], which may affect health policy decisions. Many dengue-endemic countries are transitioning to the revised WHO dengue case classification [35]; however, while the new WHO classification is used in some clinical settings in Mexico, surveillance data are still compiled as DF and DHF [36].
The new web-based Epidemiological Surveillance Platform (EPS) was implemented in 2008 [8] in which health care workers enter cases directly into the national data base. It provides real-time data to support public health decisions. Before the EPS, dengue was reported in paper forms to state epidemiological departments, entered into a local electronic system, and then emailed to the federal health authority, which many times resulted in fragmented, non-compatible data [9]. Although the EPS has improved the quality of reporting, there is still substantial room for improvement: about 40% of the reports of dengue episodes in 2009 were still considered of bad or very bad quality [8]. We overcame this limitation adjusting officially reported dengue episodes, based on reporting ratios from a prospective cohort study in Morelos, Mexico, to obtain the overall number of symptomatic DENV infections.
In addition to surveillance strategies, prevention and control are a substantial part of the economic burden of dengue. While there are various promising dengue prevention and control technologies under development [37–40], currently the only way to prevent DENV transmission is to control the vector population [41]. Vector control, prevention, and surveillance are financed through the MoH at the federal level, including the design and maintenance of the EPS. Prevention and control activities include entomological surveillance and risk assessment through mosquito ovitraps and larval indices, as well as vector control activities such as insecticide nebulization, indoor spraying, and use of larvicides. Other activities include educational and awareness campaigns, training health and vector control personnel, and community-based participatory control programs [9,33,42,43].
The objective of this study was to measure the economic and disease burden of symptomatic DENV infections in Mexico. We estimated the economic costs of dengue using a societal perspective, including vector control and surveillance costs, and the disease burden of dengue in disability-adjusted life-years (DALYs). Previous studies have estimated the economic and disease burden of dengue illness in countries from the Americas [44–52], including Mexico [44]. However, these estimates for Mexico are limited due to incomplete data and extrapolation from neighboring countries [53]. Here we addressed these limitations by combining data from multiple sources and refined estimates of the economic and disease burden of dengue in Mexico. Specifically, we estimated (i) total average annual number of dengue episodes, (ii) unit costs per episode, (iii) vector control and surveillance costs, and (iv) disease burden using DALYs. With this study, Mexico joins Panama [45], Puerto Rico [47], Nicaragua [49], and Thailand [54] as the only countries or areas worldwide with comprehensive (illness and preventive) peer-reviewed empirical estimates of the cost of dengue.
We estimated the economic burden of dengue from a societal perspective and the disease burden of dengue in DALYs, using the WHO methodology [55,56]. Specifically, we used the following equations:
Economic burden of dengue (US dollars) = total episodes x costs per episode + dengue prevention and surveillance activities + other economic impacts
Disease burden of dengue (DALYs) = years of life lost (YLL) due to premature death + years lived with disability (YLD)
An accurate estimate of the total number of dengue episodes is critical to obtain the economic and disease burden of dengue, and previous studies have found that uncertainty in the total number of dengue episodes is the main source of variability [44,57]. The costs per dengue episode include direct medical and non-medical costs and indirect costs per non-fatal and fatal case. The burden of disease was measured in DALYs, a summary measure of population health that combines information on mortality and non-fatal disease outcomes [16].
We based our burden estimates on the years 2010 and 2011 because we had access to detailed surveillance data in those years from the EPS [8,9]. The years 2010 and 2011 are, on average, relatively close to historical averages in reported cases. The average annual reported episodes were 30,941 in 2010–2011, 58,688 in 2007–2011 (which includes the 2009 outbreak), 35,091 in 2002–2011, and 32,886 in 1995–2011 [8]. Thus, if anything, our burden estimates are slightly conservative considering long-term patterns. Last, we performed a probabilistic sensitivity analysis of the economic and disease burden estimates using Monte-Carlo simulations. Monte Carlo simulations are commonly used to model phenomena with substantial uncertainty in its parameters. The method relies on running repeated trials based on random sampling from the probability distribution of each parameter in the model, and recording the results of each simulation. The results from the repeated trials were used to describe the uncertainty in the model. We report our results in 2012 US dollars (USD) using the 2012 exchange rate (USD1 = 12.88 Mexican pesos, MXN), and GDP deflators [58].
To refine the estimates of the total number of dengue episodes, officially reported dengue episodes can be adjusted for underreporting using an expansion factor (EF). An EF can be calculated as the analyst’s best estimate of the total number of dengue cases in a population divided by the number of reported cases considered dengue (EF = 1/reporting ratio) [59]. We estimated total episodes of dengue by multiplying reported episodes (41,333 episodes in 2010, and 20,548 in 2011) by an empirical EF derived from a prospective cohort study in Morelos [60,61].
The prospective cohort study was conducted in a dengue-endemic urban area in Morelos, Mexico, to assess the rate of DENV infections among the neighbors of reported dengue cases [61]. Set in the towns of Tepalcingo and Axochiapan during the 2011–2012 dengue season (June 2011-March, 2012), the study contained 1,172 participants aged 5 years and above. All participants or the parent or legal guardian of minors (5–17 years of age) gave written informed consent. The Morelos study was approved by the Ethics Commission of the National Institute of Public Health, Mexico and the Brandeis University Committee for Protection of Human Subjects.
Researchers collected 10-ml blood samples (6-ml for serological diagnosis and 4-ml for DNA extraction [62]) from all participants at baseline and 6-ml in a follow-up 3–4 months later, in addition to demographic, environmental, health-seeking behavior (e.g., number of visits to health care facilities, type of facility, private or public), and socio-cultural and entomological data. Passive and active monitoring occurred between the two rounds of data collection, including phone calls or house visits at least once a month. All dengue episodes were laboratory-confirmed by means of a paired DENV-specific IgM and IgG Capture ELISA (PanBio) at baseline and follow-up. Recent DENV infections were defined as: (i) IgM or IgG positive by capture assay, which measures recent dengue infection (2–3 months) in the baseline sample (pre-enrollment infections)[63,64], (ii) IgM or IgG positive in the follow-up sample where IgM and IgG were negative in the baseline sample (post-enrollment infections), and (iii) availability of RT-PCR/NS1/IgM/IgG positive during the follow-up months from a visit to the local health service. We used the blood samples collected at baseline or follow-up to confirm DENV infection; 12 patients were also diagnosed during the febrile episode by the state of Morelos health services (Servicio de Salud de Morelos) using NS1 or IgM/IgG capture assays. Symptomatic dengue episodes were defined as lab-confirmed dengue and reported fever.
Morelos provides a good reference value of reporting ratios of dengue episodes in Mexico. Morelos has strengthened its epidemiological surveillance in recent years, there is high level of dengue awareness and willingness to participate in dengue surveillance among the population and clinicians in the public health sector [33]. A recent study of benchmarking of effective healthcare coverage (“the proportion of potential health gain that could be delivered by the health system to that which is actually delivered”, p.1729) in Mexico based on 14 healthcare interventions [65], suggests that Morelos’ quality of healthcare provision is not too different from the country’s average. Specifically, compared to other states Morelos’ measure of effective coverage was 0.54 standard deviations below the national mean or at the 30th percentile nationally. Recent studies have used healthcare indicators to estimate reporting ratios of dengue, based on access [16] and quality [34] of healthcare, with the latter probably better reflecting the idiosyncrasies of the system that may lead to underreporting. For these reasons, we consider that using the Morelos prospective cohort study to obtain point estimates of dengue burden is reasonable, and if anything, slightly conservative. To adjust for variation in reporting ratios, we used empirical estimates of EFs from a previous study of dengue in the Americas [44] in the sensitivity analysis.
We derived costs per episode by combining patient interviews in four major hospitals in the states of Quintana Roo, Morelos, and Tabasco, macro-costing data from two major public hospitals in Tabasco, MoH health and surveillance data [66], WHO-CHOICE [67] estimates for Mexico, and previous literature on dengue burden. Indirect costs were obtained based on productivity losses by age, considering both the patient and the patient’s caregivers. We estimated vector control and surveillance costs based on MoH data.
Direct costs per episode. We estimated unit costs per dengue bed-day (inpatient episodes) and per visit (outpatient episodes) by combining macro-costing data from two major public hospitals, MoH surveillance data [66], WHO-Choice estimates [67], data from the Morelos cohort study [60], and national health statistics [66,68]. Direct medical unit costs were obtained using a macro-costing technique based on data reported by two tertiary public hospitals. To derive direct medical inpatient and outpatient unit costs that were representative of the country, we derived cost ratios for the treatment of dengue in various settings from WHO-CHOICE costs estimates for Mexico [67]. For hospitalized patients, we estimated the relative weight of treated episodes in each type of hospital based on its share total hospital beds (obtained from national health statistics) assuming that the proportion of patients who are treated in each type of hospital is equal to its fraction of total hospital beds. For ambulatory episodes, we obtained the relative weights of episodes treated in each type of setting by combining data from the Morelos cohort study (share of patients who did not visit a private or public health facility), average annual outpatients visits from health statistics and WHO-CHOICE estimates (used a proxy for its relative utilization) [67,68].
As the study did not provide any treatment to participants, it was unlikely to have any major effect on health care utilization. While receiving regular questions about febrile illness may have sensitized participants through a Hawthorne effect [69], we expect the effect to be small, if any, since there was already substantial awareness of dengue in the area [33].
The costs for homecare (including pharmacy visits) were derived from combining data sources. The share of patients with apparent dengue who did not visit a hospital or health center (about 30%, largely consistent with a previous study of healthcare use [70]) were obtained from the Morelos cohort study. Of those patients who did not visit a health center, about 37% visited a pharmacy at the onset of their febrile illness. We derived the average expenditures on medications, transport, and diagnostic tests of these early pharmacy visits from36 interviews of hospitalized dengue patients (S1 Table). We assumed that the patients who stayed at home had similar costs in medications as those who visited a pharmacy at the onset of their illness, but no transport or diagnostic costs associated with their dengue episode. The hospitalized patient interviews were also used to obtain non-medical direct costs, including transport, food, and hotel expenditures for dengue patients who visited a healthcare facility and their caregivers (S1 Text).
Indirect costs per episode. We used the human capital approach, based on work-time loss caused by dengue, to derive indirect costs per fatal and non-fatal episode [71]. Productivity loss estimates included days of work or school lost by the patient as well as relatives’ time spent caring for the patient. The breakdown by age and occupation at onset of dengue illness affect the estimates of productivity loss. Fig. 1 shows the breakdown by age of the reported cases in years 2010 and 2011. The breakdown by occupation was derived assuming that all patients aged 5–15 years old were enrolled in school, patients aged 16–17 were divided between school and work based on empirical data from school enrollment [72]. We derived the average economic value of a work day lost for economically active adults (employed or actively looking for employment) based on Mexico’s wage distribution and employment rate from the Mexican National Institute of Statistics and Geography for patients aged over 17 years old [73]. For non-economically active adults (unemployed and not actively looking for employment), the estimate was based on their reported main activity (students, household chores, retired, disabled, and non-active).
Individual and societal costs of school absence are difficult to value, but, being conservative, are at least equal to the cost of providing a day of school. We derived unit cost per-day of school lost using data on total educational expenditure at the federal, state, and municipal levels, and from the private sector, for years 2010 and 2011. A school year averages 200 schools days [74].
Economic loss from days lost was valued as the number of days lost to dengue illness times the average value per day. We took the length of hospitalization from the Mexican MoH surveillance data. We estimated the durations of illness for ambulatory and hospitalized patients as the average values from surveys of dengue patients in the corresponding setting (326 hospitalized and 834 ambulatory) across five countries in the Americas (Brazil, El Salvador, Guatemala, Panama, and Venezuela) [46]. We obtained the number ambulatory visits of hospitalized patients from interviews with them or their caregivers in Mexico. For ambulatory episodes, we used the same surveys from the Americas to derive the average duration of illness and total healthcare visits [46]. We based indirect costs of fatal cases on productivity losses by age using the age distribution of reported deaths from MoH surveillance data, the average economic value of a work day (see above), and a 3% discount rate (for consistency with international recommendations and previous studies) [71]. We estimated the years of premature life lost based on life expectancy using WHO life tables [75]. Due to the paucity of data, we assumed that the rate of reporting of deaths attributed to dengue was equal to the rate of reporting in hospitalized episodes. We relaxed this assumption in the sensitivity analysis.
Dengue prevention and surveillance. We estimated vector control and surveillance costs based on the Mexican MoH annual budget for dengue. The available data included only the years 2009 and 2010, so we imputed the vector and surveillance budget in 2011, using the average budget of the two previous years with adjustment for inflation. While vector control and dengue surveillance systems are managed by the MoH at the federal level, our estimates are conservative as they do not include possible additional spending on surveillance and vector control by state level agencies and municipalities (mainly nebulization, larvae control, and patio clean-up campaigns).
Other economic impacts of dengue. While important, data limitations did not allow us to reasonably estimate other economic costs associated to symptomatic DENV infections. These impacts include the detrimental effect of dengue outbreaks on tourism and travel [76–79], co-morbidities and complications associated with dengue infection [80–85], or the effects of health system overload [86]. When dengue outbreaks are clustered in time or location [87–90], they may worsen treatment quality and decisions or degrade performance of clinical laboratories.
The burden of disease was measured in DALYs, and is composed of the person’s years of life lost (YLL) due to premature death—based on incidence, fatality rate, and life expectancy, and a measure of the time the person lives in less than full health (years lived with disability, YLD)—based on incidence, length of illness, and impact on quality of life [14]. We estimated the burden of disease using the WHO methodology [55,56], for comparability with previous studies, and expressed burden in DALYs per million population. We obtained the age distribution of non-fatal dengue episodes and deaths from surveillance data (Fig. 1) and used model parameters (age weights, disability weight, and discount rate) based on previous studies [44,52,91]. Because the 2010 estimates of global burden of disease [16] changed their definition of DALYs (dropping age weighting and time discounting), we also provide these numbers in the results section for comparability with future estimates. Under this new definition, a child death converts to a larger number of DALYs than previously.
We used a probabilistic sensitivity analysis to address the uncertainty in our estimates of the disease and economic burden of dengue. We computed 10,000 Monte Carlo simulations simultaneously varying our parameter estimates for EFs, unit costs, days lost per episode, health service utilization, and household impact using RiskAMP [92] (iterations drew random values from the distribution of each input using the Mersenne Twister random number generator). Our results include a base-case scenario, using our best estimates for each parameter, the uncertainty around these estimates based on the sensitivity analyses, and 95% certainty level (CL) bounds.
To model variation in reporting ratios, we used a beta-PERT distribution (hereafter PERT) with the Morelos cohort empirically-derived EF as our best estimates. The range of variation in the distribution was based on a recent study of dengue in the Americas [44], which identified five field studies that included reporting ratios. We used conservative estimates, including an EF of 1.0 as the lower bound in hospitalized cases. We also used a PERT distribution for direct medical costs, with the minimum and maximum values obtained in primary and tertiary hospitals for hospitalized cases and homecare and tertiary hospitals for ambulatory episodes from combining WHO-CHOICE [67] estimates, health statistics data [66,68], macro-costing estimates, expert opinion, and patient interviews. We derived direct non-medical costs from patient interviews in four major hospitals. The variation in duration of dengue episodes and health service utilization was estimated using a normal distribution with parameters based on detailed MoH surveillance data, hospital interviews, and empirical estimates from a previous study in five countries in the Americas [46]. Last, we used a normal distribution of household impact based on weighted averages from Suaya et al.[46]
Expansion factors to adjust reported episodes. Table 1 shows a summary of the results from the prospective cohort study in Morelos, Mexico. We found a total of 253 DENV infections. Most of these infections were asymptomatic (61%), consistent with previous studies [93–96]. Only 67% of the participants with symptomatic infections visited a doctor, and most of them (74%) sought care at least once in the public sector. Of all symptomatic dengue episodes that were attended by a doctor either as outpatient or inpatient, 32% were reported to the State of Morelos surveillance system [60]. In other words, for every symptomatic episode of dengue that was treated by a health professional and reported to the surveillance system, 3.1 episodes occurred. If we considered all cases of symptomatic dengue, irrespective of whether they were attended by a healthcare professional or not, only 21% were reported to the surveillance system. Thus, there were 4.7 symptomatic dengue episodes for every reported symptomatic episode.
Reporting ratios vary considerably between the public and private sectors. In the public sector, 43% of the dengue episodes were reported (reporting ratio = 0.43), whereas no dengue episode was reported by the private sector. Limited or no reporting from the private sector has also been noted elsewhere [27,33,97]. Based on these results, Table 2 shows a summary of the expansion factors for overall (EFT), hospitalized (EFH), and ambulatory (EFA) dengue episodes needed to estimate the total cases of symptomatic dengue.
Direct medical and non-medical unit costs. Table 3 shows the estimation procedure and main data used to obtain direct medical unit costs using macro-costing [98]. Combining these data with the distribution of cases and the cost ratio relative to a tertiary hospital, we derived an average cost estimate per bed-day ($240.04) and per outpatient visit ($65.53), as shown in Table 4.
Non-medical direct costs were obtained from patient interviews. For hospitalized patients, daily non-medical costs were $25.16 for adults and $27.85 for children, and daily non-medical costs for ambulatory patients were $11.96 for adults and $9.09 for children, on average. For hospitalized patients, additional daily non-medical costs for other household members were $8.39 for adults and $6.56 for children. For ambulatory patients, the additional daily non-medical costs for other household members were $3.00 for adults and $6.00 for children.
Indirect unit costs. We estimated indirect costs based on productivity loss from the number of school-days and work-days lost. The estimated average daily unit costs for elementary education (5–14 year olds) were $7.32 in 2010, and $7.59 in 2011, and for high school education (15–18 year olds) were $9.05 in 2010 and $9.14 in 2011 [72,99]. A work-day lost for economically active adults was estimated at $10.93/day in 2010 and $11.06/day in 2011 and for non-economically active adults at $4.26 in 2010 and $4.22 in 2011. Overall, the economic value of the average work day lost was $8.20 in 2010 and $8.22 in 2011; about 1.7 times the minimum wage, which is consistent with estimates from previous studies [46,97].
Duration of dengue episodes and productivity loss. We estimated the duration of hospitalized dengue episodes at 13.9 days, including both the acute and the convalescent phases (7.4 days acute phase, 6.5 days convalescent phase). Based on hospital interviews in Mexico, we estimated that an adult had 2.4 ambulatory visits on average before being hospitalized, and a child had an average of 3.7 ambulatory visits prior to hospitalization. Ambulatory patients had a total of 3.9 healthcare visits, and illness had a total duration of 12.0 days. From the interviews, we obtained that each hospitalized patient affected on average 1.7 adults and 0.6 children in the household, and each ambulatory patient affected 2.2 adults and 0.4 children in the household on average. At the household level, school-days lost were 3.7 days for inpatients and 2.2 days for outpatients, and 6.1 work-days were lost for inpatients and 3.8 work-days were lost for outpatients.
Summary of parameters and probability distributions for sensitivity analysis. Table 5 shows a summary of the main parameters used in the analysis, assumed probability distributions, and sources. The parameters described above were used to derive base case point estimates, and the distributions and range were used in the sensitivity analysis to obtain 95% certainty levels of economic and disease burden (show in parentheses in the tables henceforth).
Total adjusted symptomatic DENV infections. Table 6 shows a summary of reported cases by setting for years 2010 and 2011, and the total estimated cases using EFs. MoH reported episodes of dengue include lab-confirmed episodes plus the proportion of positive cases from the lab-diagnosed sample multiplied by the probable cases reported (probable dengue are suspected episodes of dengue with specific clinical symptoms). Overall, we estimated a total of 195,154 (95%CL: 180,459–355,343) non-fatal and 126 (95%CL: 80–180) fatal episodes of symptomatic dengue in 2010, and 82,429 (95%CL: 75,203–142,041) non-fatal and 112 (95%CL: 75–170) fatal episodes of dengue in 2011.
Economic burden of dengue. The average cost per non-fatal dengue episode was $1,327 for hospitalized patients (direct medical: $1,010; direct non-medical: $174; indirect: $143) and $451 for ambulatory patients (direct medical: $253; direct non-medical: $92; indirect: $106). The average indirect cost per fatal dengue episode was $63,817. Altogether, the aggregate economic cost of dengue was $190 (95% CL: $165-$357) million in 2010, with a per capita costs of $1.76 (95% CL: $1.52-$3.29), and $149 (95% CL: $136-$231) million in 2011 with $1.36 (95% CL: $1.24-$2.11) per capita (Table 7). These results amount to an average economic burden of dengue of $170 (95% CL: $151-$292) million, or $1.56 (95% CL: $1.38-$2.68) per capita. The fatal episodes of dengue represent a relatively small share of the total economic burden (4.5% on average for years 2010 and 2011). Surveillance and vector control cost about $0.76 per capita ($0.71 in 2010 and $0.81 in 2011), and this represents about 48.9% of the total economic burden of dengue in Mexico (Fig. 2).
Fig. 2 shows the distribution of the economic burden of dengue in Mexico. Direct medical costs represent ~29% of the total average economic costs of dengue (34% in 2010; 23% in 2011), and direct non-medical costs sum ~8% of the total costs (10% in 2010; 6% in 2011). Fatal and non-fatal indirect costs, due to productivity loss, represent ~14% of the total economic costs of dengue (15% in 2010; 11% in 2011).
The main sources of variation for the economic burden of dengue estimates are shown in the tornado plot in Fig. 3. The vertical line shows the point estimate for the average total economic burden of dengue ($170 million). The variation for each parameter corresponds to the 95% certainty level obtained through the computation of 10,000 Monte Carlo simulations for each parameter, and for the simultaneous variation of all parameters (top bar). The diagram shows that health service utilization represents the biggest source of variation among the parameters considered in the sensitivity analysis in this study, closely followed by EFs to refine estimates of reported dengue episodes.
Disease burden of dengue. The total disease burden for the adjusted average of dengue episodes was 65.1 (95%CL: 36.0–98.7) DALYs per million population (83.5 in 2010; 46.7 in 2011). Fatal episodes represented about 27% of the disease burden of dengue (DALYs) in 2010 and 45% of the disease burden in 2011 (2010: 22.3 YLL; 2011: 20.8 YLL). The Institute of Health Metrics and Evaluation’s 2010 global disease burden study (GBD 2010) [16] dropped age weighting and time discounting from the original 1994 definition of DALYs [55,56], which results in a higher relative weight of young children compared to adults. Table 8 shows a summary of DALYs estimated for Mexico using the original definition of DALYs (WHO method) [55] for comparison with past estimates, and the new GBD 2010 method [16]. The latter method results in less conservative estimates of disease burden, and higher relative weights of fatal cases (YLL) in the total DALY estimates (YLL represented about 57% on average of total DALYs; 50% in 2010 and about 68% in 2011). Most of the years lost to disability (85%), YLD, were due to ambulatory episodes of dengue. The numbers in parentheses indicate the region of uncertainty around base-case estimates (95% certainty levels). Uncertainty in DALYs is driven by the probabilistic distribution of EFs and the duration of hospitalized and ambulatory dengue episodes (Table 5).
Extrapolation of dengue burden using historical data. If we assume that the age distribution of dengue episodes, proportion of ambulatory and hospitalized patients, overall fatality rates, and the reporting ratios of ambulatory and hospitalized cases in 2010–2011 are on average representative of the situation of dengue in Mexico in the previous years, we can estimate approximate economic and disease burden for those years. While these assumptions might be strong, the objective of this exercise is not to give precise estimates of dengue burden in previous years, but to assess how comparable are 2010–2011 data to historical data. Fig. 4 shows the total estimated number of dengue episodes and economic and disease burden for the previous 5 (2007–2011), 10 (2002–2011), and 17 (1995–2011) years. While the 5-year estimates (2007–2011) were heavily affected by the 2009 outbreak, the average of the annual burden of dengue in 2010–2011 seems a reasonable estimate of the 10-year and 17-year averages.
Dengue imposes a substantial economic and disease burden in Mexico. Because of limited data, combining multiple data sources is a key factor in achieving reliable estimates of dengue burden. Our 2010–2011 average economic and disease burden estimates ($0.80 per capita excluding costs of surveillance and vector control, and 65 DALYs per million population) are below the previous 95% confidence intervals of US$1.5–4.3 per capita and 82–147 DALYs per million population found for Central America and Mexico [44]. Reasons for our lower estimates include the use of refined, Mexico-specific reporting ratios based on the prospective cohort from Morelos, and dengue’s clustering in coastal and tropical areas [5]. Our estimates for the burden of dengue in 2010–2011 were similar to those obtained for the previous 10 and 17 years, but conservative compared to the average burden of disease in the past 5 years, driven partly by the 2009 outbreak (Fig. 4). In DALYs per million population, during this outbreak year dengue imposed a greater disease burden in Mexico (203) than pre-vaccination rotavirus diarrhea (174) [100].
Other studies have found comparable estimates of the economic and disease burden of dengue in the region. Suaya et al. [46] estimated economic burden of dengue per capita for Brazil ($0.85), Venezuela ($0.71), El Salvador ($0.30), Guatemala ($0.10), and Panama ($0.31). These numbers are underestimated since they were not adjusted for the underreporting of dengue episodes, and did not include vector control and surveillance costs. Halasa et al. [47] estimated an economic burden in Puerto Rico of $3.01 per capita without adjusting for EF, $10.84 using a refined estimate of dengue episodes, and $13.00 per capita including prevention activities, vector control, and surveillance costs. Explanations of the higher burden for Puerto Rico compared to Mexico include the island’s higher GDP per capita ($27,678) compared to Mexico ($9,747), the island-wide distribution of dengue in Puerto Rico, and higher EFs for Puerto Rico (EFH: 2.4; EFA:10). Armien et al. [45] estimated a per capita economic burden of $6.49, including surveillance and vector control costs, during the 2005 outbreak in Panama (EFT:6.0). If we assumed that the characteristics of dengue (e.g. distribution, share of hospitalized cases) during the 2009 dengue outbreak were similar to 2010–2011, our per capita cost estimate of dengue in Mexico would have been $3.99 (95%CL: $3.14-$8.72). Our estimates of the economic burden of dengue per capita in Mexico are within the range of comprehensive cost estimates for Nicaragua [49] ($0.97-$2.49; up to $5.44 in an epidemic year). The study in Nicaragua estimated that disease burden ranged from 99–805 DALYs per million population, which is comparable to our estimates for Mexico. Other estimates of disease burden in the region include Puerto Rico [52] (annual average 1984–1994: 658 DALYs per million population; range 145–1,519), and Brazil [91] (annual average 1986–2006: 22 DALYs per million population; range 14–30).
We found no notification of dengue episodes from the private sector in the cohort study in Morelos, a finding that has been confirmed by local public health officers [27]. The paucity of data from the private sector has been found elsewhere [97], and is possibly among the most critical gaps in estimating the true number of symptomatic DENV infections. One interesting finding in this research relates to the population’s health-seeking behavior. The cohort study in Morelos showed that one third of the participants had not visited a private or public healthcare facility, despite having a symptomatic DENV infection. The questionnaire for dengue patients in 4 hospitals suggests that about 11% of patients had visited a pharmacy seeking treatment at the onset of their dengue episode. These results suggest that milder symptoms of dengue go underreported, which is consistent with previous findings [27,28]. Our refinement of reported dengue episodes using EFs include unreported cases from the private sector, as well as patients with symptomatic episodes who did not seek healthcare. Overall, we found an EF for all symptomatic episodes of 4.7 or a reporting ratio of 0.21.
To check the representativeness of our estimated reporting ratio, we compared it with findings from elsewhere in Mexico and neighboring countries. A study in two cities in the state of Tamaulipas [101], near the Texas-Mexico border, suggests that the number of DENV infections represent about 20 times the number of notified cases between 1980 and 2007. Considering that 39% of the DENV infections in the Morelos cohort were symptomatic, using these numbers from Tamaulipas, we would obtain an overall reporting ratio for symptomatic cases of 0.13 (EFT = 7.8). Although the quality of the health system for Morelos is not too different from that for Mexico overall [65], a dengue awareness and education campaign in Morelos may have increased its reporting ratios compared to the rest of the country [33]. Findings from other countries in the Americas [44] found a reporting ratio of 0.08 (EFT = 11.9) for total, 0.43 for hospitalized (EFH = 2.3), and 0.07 for ambulatory (EFA = 15) cases. As the Mexican data are recent and reflect the EPS, which facilitated reporting and increased the quality of data [8,9], the higher reporting ratios are also considered reasonable. Reporting ratios vary in time and by region [26]; hence, our estimate was based on the two years of the cohort study to provide a more stable estimate. Had we considered only post-enrollment infections, we would have obtained a reporting ratio of 0.13 (EF = 8.0, derived from Table 1) and a point estimate of the economic burden of dengue of $253 million or $2.32 per capita (which is included within our 95% certainty level). We think our current estimate of $170 million ($1.56 per capita) is statistically more stable and accurate, although our expansion factor may be an underestimate in relation with other parts of the country with less dengue awareness, or lower overall quality of the health system.
Our estimates suggest that at least 48.9% of the economic burden of dengue corresponds to surveillance and vector control. This share of total costs is higher than those of previous estimates of vector control in other countries. For example, the surveillance and vector control shares of estimated annual economic burden of dengue were 17% in Puerto Rico [47], 30% in Panama [45], and 28% in Thailand [54]. However, the per capita costs of surveillance and vector control (in 2012 US dollars) were lower in Mexico ($0.76) than these other countries ($2.14 in Puerto Rico, $1.79 in Panama, and $1.15 in Thailand). This pattern is partly explained by Mexico’s lower share of the national population at risk of dengue, as the disease is clustered mainly in Mexico’s coastal and tropical regions [5]. Also, Mexico did not experience an outbreak during our study years. Reflecting these patterns, the number of dengue episodes per 1,000 population was lower in Mexico (1.29) than in the other three countries—2.87 in Puerto Rico [47], 9.76 in Panama [45], and 4.08 in Thailand [54].
Several areas of uncertainty in our estimates of disease and economic burden of dengue in Mexico deserve attention. First, estimating the total episodes dengue is difficult due to paucity of data. For example, the cohort study in Morelos showed that 26% of the participants sought care in the private sector; but this estimate may be low as data from the Mexican National Health and Nutrition Survey 2012 showed that about 39% of all outpatient visits (for any illness) were in the private sector [102]. Second, the Morelos cohort is limited in geographical range, calendar years, and age groups, and therefore not necessarily representative of all regions with dengue transmission in Mexico. Local variations in the quality of the health system [34] and accessibility to health services [16] may result in differences in dengue patients’ health-seeking behavior, thus affecting reporting rates of apparent DENV infections. Third, our direct medical costs for dengue episodes were based on macro-costing in two tertiary hospitals in Tabasco, which may not necessarily be representative of hospitals in Mexico. We partially addressed this by adjusting our estimates based on WHO-CHOICE data, and varying our estimates in the sensitivity analysis. Costs in the private sector are probably higher than the costs we used, which possibly makes our economic burden estimates conservative.
Fourth, we only considered surveillance and vector control costs from the federal level. Due to data limitations, we could not distinguish operating and capital expenditures, and did not include allocated and donated resources such as the time allocated by field personnel or volunteers to surveillance and vector control activities, as has been done elsewhere [103]. These limitations make our estimates of costs of surveillance and vector control conservative.
Fifth, despite having improved previous estimates of economic burden by including costs of illness and dengue prevention and control strategies, we did not include other impacts of dengue illness due to data limitations.
Last, our estimates of the burden of dengue were based on the acute and convalescent phases of a dengue episode (Table 5). Recent studies suggest that dengue patients may present long-term symptoms [104–109] like fatigue syndrome or depression, a possibility acknowledged by the WHO since 1997 [110]; unfortunately, there is not enough evidence or agreement on the characteristics (e.g., frequency, intensity, duration) of these persistent symptoms, and whether or not they are caused by dengue alone.
Dengue costs the Mexican economy an annual average of US$170 (95%CL: 151–292) million, or $1.56 (95%CL: 1.38–2.68) per capita. Of this, $87 (95%CL: 87–209) million or $0.80 per capita (95%CL: 0.62–1.12) corresponds to illness and $83 million or $0.76 per capita to vector control and surveillance. These estimates do not include other costs, such as long-term sequelae of dengue, comorbidities, impacts on travel and tourism, or the disruption of health services during epidemics. Mexico’s annual disease burden from dengue is 65.1 DALYs per million population.
Having objective and comparable estimates of the economic and disease burden of dengue is essential to inform health policy, increase disease awareness, and assess the impact of dengue control technologies [12,111]. More so, considering that several vaccine candidates [112] and other prevention and control technologies [37,40,42,113] are currently under development, and that Mexico might be an early adopter [12,53,114]. Results from the phase III clinical efficacy multicenter trial of a dengue vaccine candidate in the Americas suggest an overall vaccine efficacy of 60.8%, and a reduction in the risk of hospitalization of 80.3% [115]. These recent results make burden estimates even more urgent as Mexico confronts real choices. Effective dengue prevention and control strategies will probably require a combination of approaches and the involvement of various stakeholders [116]. With this study, Mexico joins Panama [45], Puerto Rico [47], Nicaragua [49], and Thailand [54] as the only countries or areas worldwide with comprehensive (illness and preventive) empirical estimates of the cost of dengue. The results from this study reaffirm that exploring approaches to control dengue further would be economically valuable.
|
10.1371/journal.pgen.1005526 | A Systems Approach Identifies Essential FOXO3 Functions at Key Steps of Terminal Erythropoiesis | Circulating red blood cells (RBCs) are essential for tissue oxygenation and homeostasis. Defective terminal erythropoiesis contributes to decreased generation of RBCs in many disorders. Specifically, ineffective nuclear expulsion (enucleation) during terminal maturation is an obstacle to therapeutic RBC production in vitro. To obtain mechanistic insights into terminal erythropoiesis we focused on FOXO3, a transcription factor implicated in erythroid disorders. Using an integrated computational and experimental systems biology approach, we show that FOXO3 is essential for the correct temporal gene expression during terminal erythropoiesis. We demonstrate that the FOXO3-dependent genetic network has critical physiological functions at key steps of terminal erythropoiesis including enucleation and mitochondrial clearance processes. FOXO3 loss deregulated transcription of genes implicated in cell polarity, nucleosome assembly and DNA packaging-related processes and compromised erythroid enucleation. Using high-resolution confocal microscopy and imaging flow cytometry we show that cell polarization is impaired leading to multilobulated Foxo3-/- erythroblasts defective in nuclear expulsion. Ectopic FOXO3 expression rescued Foxo3-/- erythroblast enucleation-related gene transcription, enucleation defects and terminal maturation. Remarkably, FOXO3 ectopic expression increased wild type erythroblast maturation and enucleation suggesting that enhancing FOXO3 activity may improve RBCs production. Altogether these studies uncover FOXO3 as a novel regulator of erythroblast enucleation and terminal maturation suggesting FOXO3 modulation might be therapeutic in disorders with defective erythroid maturation.
| Red blood cells (RBCs) are highly specialized cells that transport oxygen throughout the body and are essential for survival. However, RBCs have a limited lifespan and need to be replenished continuously by stem cells in the bone marrow. Mammalian RBCs are unique in that in order to fully mature they exclude their nucleus and other organelles. Mechanisms involved in these processes are not well understood at the molecular level. Defects in any of the these processes may lead to red blood cell defects, a decreased capacity to transport oxygen and/or a block in red blood cell production in vitro. Therefore, understanding how these processes are regulated at the molecular level can lead to promising new therapies for red blood cell defects and improved methods of generating red blood cells in a dish. Here, using an integrated computational and experimental biology approach, we found that the nuclear factor FOXO3 is a crucial regulator of red blood cell production by coordinating the expression of many of the genes specific for terminal maturation of red blood cells. Furthermore we found that FOXO3 can even increase the production of normal red blood cells in culture raising the possibility that enhancing FOXO3 may have a therapeutic use. Our studies identify FOXO3 as a novel regulator of RBC enucleation and terminal erythropoiesis.
| Erythropoiesis ensures the daily production of over 200 billion RBCs whose main function is to carry oxygen. Decreased production of RBCs is associated with many human disorders involving impaired erythroblast maturation. The generation of RBCs in vitro from embryonic stem cells or human-induced pluripotent stem cells (iPS cells) has been proposed to provide a cost-effective and safe blood supply. Despite recent development [1] achieving efficient production of functional RBCs has been hindered by incomplete knowledge of terminal erythroblast maturation.
Generation of RBCs involves the differentiation of hematopoietic stem cells into common megakaryocyte and erythroid progenitors, which give rise to lineage-restricted erythroid progenitors, erythroblasts, and ultimately erythrocytes. During the final stages of erythropoiesis, proliferation of erythroblasts is coupled with differentiation as terminally differentiating erythroblasts accumulate hemoglobin, reduce cell size, and condense their nuclei. Following enucleation, reticulocytes remodel their membrane and clear mitochondria and remaining organelles to transition into fully mature erythrocytes [2]. This complex process is controlled by integration of erythropoietin receptor (EpoR) signaling with the function of erythroid lineage-specific transcription factors including GATA–1, KLF–1 and TAL–1 (SCL) and their cofactors [3]. Despite recent progress [4–6], many questions remain unanswered regarding whether these factors function alone or together to control enucleation and/or to remove organelles, including mitochondria, during terminal erythroblast maturation. Increasing evidence suggest that FOXO3 cooperates with these factors and their requisite coregulators to control specific molecular/cellular steps that drive terminal erythroid maturation [4–6].
FOXO3 belongs to the FOXO family of Forkhead transcription factors composed in mammals of the highly related members FOXO1, FOXO3, FOXO4 and FOXO6. FOXOs are homeostatic maintaining factors implicated in many diseases including cancer, diabetes, and erythroid disorders [7–10]. FOXOs integrate fundamental biological processes through the regulation of cell cycle, oxidative stress, DNA damage responses, apoptosis, inflammatory responses, and metabolism [7,11]. FOXO genes have evolutionary conserved functions in stem cell maintenance and longevity [12–24]. Emerging evidence suggests that FOXO may also play a key role in tissue-tissue communication [7,25–27]. Among FOXO factors, FOXO3 is critical for normal and stress erythropoiesis [8–10,28–31]. This is evident as Foxo3 mutant mice die rapidly when exposed to acute erythroid oxidative challenge [29]. Notably, FOXO3 expression and function increase progressively with erythroblast maturation [29,32,33]. Despite these findings, whether FOXO3 has any function in the regulation of terminal erythroblast maturation remains unknown.
Using an integrated systems and experimental biology approach, we demonstrate that FOXO3 is critical for the correct temporal expression of at least one third of the genes differentially expressed in normal maturing erythroblasts. Dysregulation of this subset of genes due to FOXO3 loss led to defects at distinct stages of terminal erythroblast maturation and RBC production. Our data demonstrate that FOXO3 is critical for erythroblast enucleation through polarization of the nucleus expulsion direction and is required for mitochondrial clearance. The transcriptomic analyses also revealed that increasingly maturing primary erythroblasts express immune-related transcripts whose expression is highly modulated upon loss of FOXO3. Collectively, these findings demonstrate that FOXO3 is an essential component of the transcriptional program that regulates terminal erythroblast maturation and required for the erythroblast enucleation process.
To investigate the FOXO3-regulated transcriptional program during erythroblast maturation we compared the transcriptome of adult bone marrow erythroid precursor populations of Foxo3-/- mice to that of wild type (WT) mice. Since immature erythroblasts accumulate in Foxo3 mutant bone marrow [29], we reasoned that the relative accumulation of immature erythroblasts due to FOXO3 deletion might reflect a block in their terminal maturation. Immature erythroblasts were isolated using flow cytometry relying on the immunophenotype and forward scatter properties of erythroblasts according to Chen et al. [34]. The erythroid specific marker TER 119 was combined with CD44 and forward scatter, which both decrease during maturation, to resolve the progressive stages of erythroblast differentiation [34] (Fig 1A). This approach enabled the distinction and purification of three consecutive stages of maturation of erythroid populations (pro-, basophilic and polychromatophilic erythroblasts) as defined by gates I, II and III respectively (Fig 1A). Relatively pure subpopulations of erythroblasts as shown by morphological analysis (S1A Fig) were isolated using this gating strategy. RNA was isolated from the first three gates (Fig 1A, depicted in red) and deep sequencing analysis was conducted to compare wild type and Foxo3 mutant erythroblast transcriptomes. In the morphological analysis and subsequent validation experiments, Gate IV cells (depicted in black) encompassing mainly reticulocytes that are enucleated cells preceeding mature red blood cells were also included (Fig 1A).
Erythroid-specific genes of wild type erythroblasts were grouped according to low, intermediate, or high levels of expression that was validated by qRT-PCR, which faithfully reproduced the levels and patterns of gene expression in each group (S1B Fig and S1 Table). Expression of globin genes was used as a control to confirm the identity of erythroblasts. Although the expression was too high to be reliably quantified by cufflinks software used for RNA-Seq analysis, qRT-PCR analysis demonstrated globin gene upregulation during erythroid maturation (S1C Fig and S1 Table). Analysis of wild-type erythroblasts revealed 5514 genes (S2 Table) with at least two fold differential expression between gates. These genes were grouped into eight clusters using k-means clustering (Fig 1B). Using Gene Ontology (GO) term analysis with FuncAssociate 2.0, signaling pathways and biological processes enriched in each of the clusters were delineated (S3 Table). The chromatin immunoprecipitation (ChIP) enrichment analysis tool, ChEA, [35,36] was used to identify potential transcription factors that may occupy the genes within each cluster (S4 Table). Cluster A, which grouped 600 genes that were continuously downregulated from Gates I to III, is enriched in inflammatory and apoptotic genes. In agreement with RNA-Seq analysis of human erythroblasts, genes continuously up-regulated during terminal erythroid maturation and clustered in G and H (801 and 446 genes respectively) were enriched for autophagy-related genes underscoring the importance of autophagy during erythroblast maturation [4,33,37]. Genes initially upregulated and then down-regulated were enriched for cell cycle-related processes, chromatin remodeling, DNA repair and mitotic genes (Clusters E and F, 867 and 865 genes respectively) (Fig 1B and S3 Table). Erythroid-specific genes including heme biosynthetic enzymes, iron metabolism, and erythroid membrane genes were distributed in clusters F, G, and H (S3 Table). Consistent with these results, known erythroid transcription factors including GATA–1, KLF–1, TAL–1, and the transcriptional co-activator EP300 (E1A binding protein p300) were associated with the sustained up-regulation in clusters G and H (S4 Table). Expression of Clusters E and F genes which were downregulated at Gate III, was associated with E2F4, E2F1, c-MYC transcription factors, and Cyclin D1 (CCND1), all known for regulating cell cycle progression. These data support previous findings [38] that erythroblasts continue to cycle at late stages of their maturation.
Immune-related genes were enriched in clusters B and C (1355 and 295 genes respectively) which include genes initially downregulated and then slightly upregulated in gate III (Fig 1B). As anticipated many inflammation-related genes, such as S100 Calcium Binding Protein A11 (s100a11), interleukin 17 receptor α (Il17ra), and immediate early response 2 (Ier2) were downregulated with maturation (S2A Fig). Furthermore, ChEA analysis implicated MYB and PU.1, known repressors of terminal erythroid maturation, as transcription factors regulating B and C gene clusters [39–41] (S4 Table). Accordingly, mRNA expression of MYB and PU.1 decreased during erythroid maturation (S2B Fig). Interestingly and consistent with previous findings [42,43], genes linked to several immune pathways, including several genes involved in the interferon response and lymphocyte activation pathways, were upregulated upon terminal erythroid maturation (clusters G and H) (Fig 1B). Expression of interferon-related genes including interferon regulatory factor (Irf7) and radical S-adenosyl methionine domain containing 2 (Rsad2) was validated by qRT-PCR which found these to be upregulated 2–15 fold upon maturation of bone marrow erythroblasts (S2C Fig). The corresponding proteins were also expressed in TER119+ erythroblasts that are negative for CD45 (TER119+CD45-) confirming the specificity of their expression and lack of non-erythroid cell contamination (S2D Fig). These observations suggest that these immune-related genes may have a specific function in terminal erythroid maturation.
For comparing WT and Foxo3 mutant erythroblast expression profiles, genes were re-clustered. Genes with at least a two-fold differential expression between the same gate of WT and Foxo3 mutant erythroblasts were reclustered for further analysis. Expression of 3906 genes (S5 Table), approximately 35% of the total expressed genes, was strikingly altered in Foxo3-/- erythroid precursors (Gates I to III). In agreement with the progressive increase in the expression and function of FOXO3 with erythroid maturation [29,31], the majority of differences in gene expression among gates I to III of wild type and Foxo3-/- erythroblasts were detected in Gate III (S3A and S3B Fig). Loss of FOXO3 led to both repression (clusters Q and R, 926 and 509 genes respectively) and induction (clusters I and J, 444 and 1094 genes respectively) of gene expression during erythroid maturation (Fig 1C). Immune-related pathways, including macrophage and neutrophil activation pathways enriched in clusters I and J, fell within programs that are normally repressed, but were aberrantly upregulated in the absence of FOXO3 (Figs 1C and S3C and S6 Table). These results indicate that loss of FOXO3 may enhance the expression of many inflammatory-related genes in maturing erythroblasts (S3C Fig), consistent with the anti-inflammatory function of FOXO3 [27,44]. The greatest impact of FOXO3 loss was exemplified by clusters Q and R, which consist of genes normally upregulated during erythroid maturation. Cluster R was enriched for autophagy and catabolic processes, while Cluster Q was enriched for heme biosynthesis, erythroid differentiation, nucleosome assembly, and DNA packaging-related processes (Fig 1C and S6 Table). Both clusters Q and R designate genes that are continuously up-regulated from Gates I to III. However, differences arise between the clusters from the loss of FOXO3. In the absence of FOXO3, the progressive upregulation of genes between Gates I to III is abrogated in cluster R. In contrast genes in cluster Q are not upregulated from Gates II to III. The stalled gene activation in Cluster Q suggests that FOXO3 is required for the transition to the gene expression profile characteristic of Gate III erythroblasts. The distinct pattern of clusters Q and R may reflect distinct modes of FOXO3 action for each respective gene cohort. Since 40% of the genes in both clusters are established targets of GATA–1, TAL–1, and/or KLF–1 transcription factors (S7 Table) [3], these results, consistent with previous findings [6], raise the possibility that FOXO3 cooperates with these factors to sustain gene transcription during terminal erythroid maturation.
Autophagy was one of the main pathways highly up-regulated upon erythroblast maturation (Fig 1B, clusters G and H). However, autophagy-related genes were greatly dysregulated in Foxo3-/- mutant erythroblasts (Fig 1B, and S6 Table cluster R). Autophagy (or macro-autophagy) serves as a homeostatic mechanism that mediates the consumption of damaged or old cellular components, as well as the cellular remodeling that is associated with cell differentiation [45]. In erythroblasts, autophagy is implicated specifically in the clearance of mitochondria (mitophagy or selective mitochondrial autophagy) during terminal erythroid maturation [37,46,47].
To evaluate the role of FOXO3 in regulating autophagy in primary erythroblasts, we validated expression of autophagy-related genes differentially expressed between wild type and Foxo3 mutant erythroblasts (Figs 2A and S4A). The expression pattern of some of these genes, including Nix (Bnip3l) and Ulk1 (Fig 2A), was upregulated 15- to over 40-fold during late stages of bone marrow erythroblast maturation and reticulocyte formation. This finding is supported by the known function of these genes in mitochondrial removal [46–48]. The similar pattern of expression of other autophagy-related genes, including Gabarapl2 (Gate–16), p62, Atg14, and Pink1 (Figs 2A and S4A) suggests their involvement in erythroid maturation. In addition, autophagy genes, including Gabarapl1 and Map1lc3b lie within the highly expressed gene cohort at all stages of erythroblast maturation indicating these genes may be involved in the homeostatic control of erythroid maturation (S1 Table). Consistent with this, the core autophagy genes Atg5 and Atg7 were expressed at similar levels throughout erythroid maturation relative to other autophagy genes examined (S4B Fig). However, expression of several autophagy-related gene transcripts was profoundly compromised in Foxo3 mutant erythroblasts (Figs 2A and S4A), suggesting that these genes are potential direct targets of FOXO3 in erythroblasts. ChIP analysis revealed occupancy at Btg1, a known FOXO3 direct target [28] and several autophagy-related genes including Nix, Gabarapl2 and Ulk1 in wild-type, but not in Foxo3-deficient erythroblasts. Thus, FOXO3 occupied regulatory regions of these genes in primary erythroblasts in vivo (Fig 2B). Occupancy was not detected at upstream sequences lacking FOXO3 binding sites in these regulatory regions. Several of the genes shown in Figs 2A and S4A are direct FOXO3 transcriptional targets, including Map1lc3b and Atg14 in non hematopoietic cells [49–51], while others including Gabarapl2 were not known to be regulated by FOXO3 in any system. Interestingly, Nix and Ulk1 are known regulators of erythroblast mitochondrial removal [47] [48].
We reasoned that the failure to upregulate autophagy gene expression might decrease autophagy during differentiation of Foxo3-mutant erythroblasts and compromise their terminal maturation. To test this, we measured the conversion of the soluble free form of Microtubule-associated protein 1 light chain 3B (LC3B-I) to the lipidated LC3B-II form, an essential step in autophagosome activation. Autophagosomes are generated as a result of sequential assembly and activation of autophagy-related proteins [45]. Once activated, they engulf the damaged organelle or proteins and fuse with lysosomes to degrade their cargo. The LC3B-II/LC3B-I ratio, an important indicator of autophagy [52], was decreased significantly in freshly isolated primary Foxo3-mutant erythroblasts under homeostatic conditions (Fig 3A), suggesting that autophagy was compromised in these cells. Further gate-by-gate analysis suggested that defective autophagy was mainly observed within Foxo3-/- gate III and IV erythroblasts (Fig 3B). To further ensure that these results reflected impairment in autophagosome formation and not abnormalities related to events upstream or downstream [53], we evaluated the autophagic flux in erythroblasts using an autophagosome-specific fluorescent probe [54]. Addition of chloroquine to bone marrow cultures blocked degradation and induced accumulation of autophagosomes in wild type and Foxo3-/- erythroblasts, enabling their measurement over time by flow cytometry and western blotting for LC3B-II (Fig 3C and 3D). Foxo3-deficient erythroblasts exhibited significantly reduced autophagic flux as compared to their wild type counterparts specifically in Gate IV erythroblasts that are highly enriched in reticulocytes (Fig 3C). The LC3B-II accumulation was also delayed in chloroquine-treated Foxo3-/- relative to wild type erythroblasts (Fig 3D). The reduced autophagy flux was mainly detected in late stage Foxo3-deficient erythroblasts, consistent with the alteration of autophagy-related gene expression in these cells (S4A Fig). Collectively, these results are consistent with the notion that FOXO3 controls autophagy at late, but not early, stages of erythroid maturation (Fig 3B and 3C).
We considered whether low levels of autophagy that are characteristic of Foxo3-/- erythroblasts suffice to support mitochondria removal during reticulocyte maturation. The frequency of circulating reticulocytes in Foxo3-mutant mice is increased to compensate for increased RBC destruction [29]. However, in Foxo3-mutants a significant fraction of mature RBCs are devoid of the transferrin receptor CD71 but remain positive for the mitochondria-specific MitoTracker Red probe (Fig 3E). This was further confirmed using a distinct mitochondria-specific probe MitoTracker Green (S4C Fig). These results indicate that despite their increased numbers in Foxo3 mutant mice [29] reticulocytes retain mitochondria, delaying the complete maturation into RBCs. This increase is consistent with the failure to induce multiple autophagy-related genes in Foxo3-/- erythroblasts, which have been implicated specifically in mitophagy (Figs 2A and S4A). Nonetheless, the impact of impaired autophagy on mitochondrial removal in Foxo3-mutant reticulocytes appeared to be relatively modest, suggesting that compensatory mechanisms might be involved. Alternatively these results might reflect the contribution of distinct autophagy pathways controlling mitochondrial removal in erythroid cells [55–57].
Key steps in RBC formation include chromatin condensation and expulsion of the nucleus (enucleation) from late-stage erythroblasts [2] [58] [59]. We discovered that many genes involved in nucleosome assembly and DNA packaging-related processes were downregulated in Foxo3-mutant erythroblasts (Fig 1C, cluster Q, S6 Table), raising the possibility that FOXO3 controls chromatin condensation and/or enucleation. Using DRAQ5, a fluorescent probe that binds DNA in vivo, we found that there were significantly fewer enucleated Foxo3-mutant bone marrow erythroblasts than wild type cells (Figs 4A and S5A). QRT-PCR expression analysis of cluster Q showed that genes implicated in chromatin condensation and/or enucleation Mxi1, Riok3, Smarca4, Trim58, Rac GTPase I and II [60–63] were all reduced at distinct stages of Foxo3-mutant erythroblast maturation (Figs 4B and S5B). Among these, Riok3 and Trim58 transcripts were upregulated 60- and over 120-fold, respectively, during differentiation of normal bone marrow erythroblasts, suggesting these genes may function beyond enucleation during terminal erythroblast maturation (S5B Fig). In contrast, expression of RacGTPase I and II decreased with maturation of normal bone marrow erythroblasts. ChIP of endogenous FOXO3 in wild type, but not in Foxo3 mutant, bone marrow erythroblasts revealed occupancy at regulatory regions of Mxi1 and Riok3 (Fig 4C). Thus, these genes similar to the autophagy genes examined earlier may also be directly regulated by FOXO3 in erythroblasts.
We used imaging flow cytometry to develop mechanistic insights into the potential enucleation defect in Foxo3 mutant erythroblasts. As described previously, [58,64] TER119+ erythroblasts can be segregated according to their cell and nuclear size into progressive stages of maturation (pro-, basophilic, polychromatic and orthochromatic erythroblasts) (Fig 5A). While this analysis confirmed the increase in erythroid precursors described previously in Foxo3 mutant bone marrow (S5C Fig) [29], the analysis also revealed that the accumulation occurs at the orthochromatic stage of Foxo3 mutant erythroblast maturation (Figs 5A and S5C). Prior to enucleation, late-stage erythroblasts displace the nucleus from its central location. The percentage of orthochromatic erythroblasts containing an asymmetrically positioned nucleus was quantified using the delta centroid, a measure of the distance between the center of the cell body and center of the DRAQ5-stained nucleus [58]. This analysis revealed a lower percentage of enucleating erythroblasts in the Foxo3-mutant versus wild-type bone marrow (Fig 5B).
To investigate how the FOXO3 loss impacts enucleation, we conducted a high-resolution morphological analysis of bone marrow erythroblasts. As shown in Fig 6A, confocal analysis revealed wild-type enucleating erythroblasts by the presence of a gap in the bright TER119 membrane staining as described previously [65]. Wild type enucleating erythroblasts displayed a dumbbell-shaped nucleus, with a neck located at the TER119 sorting boundary of the nascent reticulocyte. This ensures that cells use a single direction for nuclear extrusion. In contrast, Foxo3 mutant erythroblasts exhibited multiple nuclear necks accompanied by multiple sorting boundaries with each lobe extruding in a different direction away from the nascent reticulocyte. This unique pattern suggested defective polarization of Foxo3 mutant erythroblasts during enucleation (Fig 6A). Accordingly, 48% of the Foxo3-/- enucleating erythroblasts exhibited abnormal enucleation morphologies. These results were verified by imaging flow cytometric analysis by distinguishing orthochromatic erythroblasts with tri-lobular nuclei from normal extruding nuclei (Fig 6A–6C). This analysis demonstrated that erythroblasts with tri-lobular nuclei were remarkably more frequent in Foxo3-/- versus wild type bone marrow (Fig 6B amd 6C). These results indicate that FOXO3 is critical for enucleation and may control erythroblast polarization/nuclear positioning as a prelude to enucleation. Consistent with this notion, several genes implicated in cytoskeleton organization, cell polarization and cytokenesis, are deregulated in Foxo3-/- erythroblasts (Figs 6D, S5C and S6A). Notably, the small RhoGTPase CDC42-related gene cluster that regulates cytoskeleton organization and cell polarization [66,67] is aberrantly upregulated in Foxo3-/- erythroblasts (Fig 6D).
To investigate whether FOXO3 might directly regulate enucleation, we examined the capacity of FOXO3 to rescue enucleation in Foxo3 mutant erythroblasts. MSCV-IRES-GFP (MIG) or MIG-FOXO3 were ectopically expressed in wild type and Foxo3-/- bone marrow-derived erythroid progenitors and subjected to in vitro maturation by adapting a previously established protocol, for bone marrow erythroblasts [68] (Fig 7A). Gene expression patterns of Gata1, Pu.1, and Hbb-b2, were examined over 3 days of maturation and validated that they resemble in vivo erythroblast maturation patterns (S7A Fig). Flow cytometry analysis enabled resolution of the most immature erythroblast populations (ex vivo P1), exhibiting high CD44 levels and large size, compared to the most mature erythroblast population (ex vivo P3), exhibiting low CD44 levels and small size. The ratio of P3 to P1, termed maturation index here, reveals the extent of maturation at day 3 of culture. Foxo3-/- erythroblasts show an overall decrease in their ability to mature (Fig 7A and 7B) [29]. Importantly, ectopic expression of FOXO3 in mutant erythroblasts rescued their defective maturation without modulating the total number of TER119 cells (Fig 7A–7C). In addition, the last stage of maturation was specifically impaired in Foxo3 mutant erythroblasts (P3/P2) at ∼ 50% (± 0.046, n = 3) of that of wild type (P3/P2) cells and rescued by ectopic expression of FOXO3 (Fig 7A, bottom panels). Furthermore, a similar degree of rescue by ectopic FOXO3 expression was observed when comparing the very last stages, S2 (enucleated) and S1 (nucleated) mature erythroblasts in culture, where the majority of enucleation occurs (S7B Fig). In fact, overexpression of FOXO3 rescued the enucleation capacity of Foxo3-/- erythroblasts to levels similar or higher than that of wild type controls in ex vivo maturation cultures (Fig 7A and 7B). In support of a potential direct FOXO3 control of enucleation the ectopic expression of FOXO3 increased the expression of Mxi1, Riok3, and Trim58 (Figs 7C and S7C). Notably, overexpression of FOXO3 further improved enucleation, the maturation index, and expression of enucleation related genes in WT erythroblasts (Figs 7A–7C and S7B). Ectopic expression of FOXO3 also improved expression of autophagy-related genes in Foxo3-/- erythroblasts (S7C Fig).
Our studies show that FOXO3 has critical physiological functions at key steps of terminal erythroid maturation. While FOXO3 is required for enucleation and for subsequent mitochondrial clearance in reticulocytes, FOXO3 loss impacts more profoundly upon the enucleation process. Enucleation is a critical limiting step in erythroid maturation [2]. The removal of the nucleus results from the culmination of a multi-step process that starts early in differentiation and encompasses chromatin condensation, cell polarization, formation of contractile actin ring and ultimate enucleation [2]. Our studies demonstrate that FOXO3 is essential for normal enucleation to occur. FOXO3 directly controls the expression of several critical regulators of enucleation, including Riok3 and Xpo7, important regulators of chromatin condensation [49,60,69], as well as Mxi1, implicated in contractile actin ring formation [60]. FOXO3 also potentially controls Trim58 expression that is a gene critical for erythroblast nuclear polarization and/or extrusion [63] (Figs 4 and S5B). Loss of FOXO3 severely compromised erythroblast polarization and resulted in multilobed nucleated erythroblasts in agreement with FOXO3 regulation of genes implicated in cell polarity and cytoskeleton organization (Figs 4–6, S5 and S6). The aberrant gene expression of the CDC42 network suggests that FOXO3 may mediate polarization or the cytoskeleton organization of enucleating erythroblasts, which are potentially important for the chromatin condensation and/or nuclear expulsion (Figs 4–6). These abnormalities are likely to be direct consequence of FOXO3 loss and not due to processes resulting from developmental lack of FOXO3, as reintroduction of FOXO3 normalized the defective enucleation within a short period of time in culture (Fig 7). Future studies will address the precise mechanism of FOXO3 function in this process. PI3-kinase, an upstream negative regulator of FOXO3 in erythroblasts [70], is required for erythroblast maturation [71,72] and polarization during enucleation in vitro [73]. However, signals other than PI3-kinase/AKT are likely to override the PI3-kinase repression of FOXO3 activity during erythroblast enucleation in vivo [18]. It is therefore probable that both PI3-kinase and FOXO3 control erythroblast terminal maturation and enucleation.
FOXO3 is an established regulator of autophagy [25,74–76]. Here we demonstrated that FOXO3 is critical for removal of mitochondria during terminal erythroid maturation. We found that FOXO3 regulates transcription of autophagy-related genes in maturing erythroblasts and FOXO3 loss compromised autophagosome formation and autophagy, reduced autophagic flux in maturing erythroblasts and impaired mitochondrial clearance in RBCs (Figs 2 and 3). Despite the profound reduction in autophagy gene expression, specifically at late stages of erythroblast maturation, and the negative impact on autophagosome formation, the influence of FOXO3 loss on mitochondria clearance was relatively modest (Fig 3E). Together these findings indicate that under homeostatic conditions, the relatively low autophagy gene expression is sufficient to maintain clearance of mitochondria in FOXO3 mutant erythroblasts. These results support the notion that multiple autophagy pathways might regulate mitochondrial clearance in erythroid cells [55–57].
In erythroid cells, FOXO3 amplifies the capacity of GATA–1 to induce autophagy gene expression [4,5]. Consistent with this, loss of FOXO3 led to notable decrease in the expression of many of autophagy-related genes in primary maturing erythroblasts (Fig 2) despite endogenous GATA–1 expression (S7A Fig). These results indicate that GATA–1 induction of autophagy gene expression in terminally maturing primary erythroblasts may be compromised in the absence of FOXO3.
Our unbiased genome-wide analysis revealed that, in addition to the approximately 600 genes that have been categorized as erythroid specific [6], more than 1000 genes, including genes involved in cell cycle, autophagy and immune-modulation are upregulated during terminal erythroid maturation. Loss of FOXO3 impaired significantly the expression of over 65% of these genes. Furthermore, FOXO3 loss limited to various degrees (clusters Q and R) the increase in expression of several GATA1, TAL1 and/or KLF–1 targets during terminal erythroid maturation. These results (Fig 1C) combined with the ChEA prediction (S7 Table), support a model in which FOXO3 is required to activate and/or enhance (cluster R), or sustain (cluster Q) gene expression during terminal erythroid maturation (Fig 8). As further support for this potential role of direct FOXO3 transcriptional activation of Cluster Q and R genes, we analyzed a list of direct FOXO3 targets determined by ChIP-seq [77] in a colon carcinoma cell-line. This comparison shows a significant number of overlap with cluster Q and R genes, suggesting that these genes are direct FOXO3 targets (S8A Fig, S8 Table). A sizeable number of genes do not overlap; this is most likely due to a difference in cell type and species. Consistent with previous findings [4–6,77–80] these results support a model (Fig 8) in which erythroid transcription factor complexes induce Foxo3 gene expression in immature erythroblasts which in turn cooperates with these factors to sustain and/or enhance the erythroid transcriptional program during terminal maturation. A similar model has been proposed for the induction of expression of interferon regulatory factor IRF2 and IRF6, which subsequently cooperate with GATA–1 and TAL1 transcription factors to establish the adult human erythroid program [81]. Collectively, these findings raise the possibility that this mode of regulation of gene expression might be a common mechanism during erythroid differentiation and maturation. Cooperation of these transcription factors with FOXO3 is likely to be critical for terminal erythroblast maturation. Nonetheless, in the absence of ChIP-Seq data, we are unable to determine with certainty all genes that are direct products of FOXO3 cooperation with other transcription factors in maturing erythroblasts.
Taken together our findings demonstrate that FOXO3 activity critically regulates progressive stages of terminal primary erythroblast maturation. As FOXO3 activity is critical for the correct temporal gene expression in maturing erythroblasts, we predict that abnormal FOXO3 expression/function may significantly influence erythroid disorders as has been reported specifically for hemoglobinopathies [8–10]. Since post-translational modifications of FOXO3 are the main determinants of FOXO3 output, analysis of FOXO3 function in addition to its expression will be critical in the context of disease. Future studies should elucidate whether modulations of FOXO3 activity influence the efficiency of RBC production in vitro or in disorders of erythroid cells in vivo.
The generation and genotyping of mice (on C57Bl6 genetic background) were performed as previously described [14,31]. WT and Foxo3-/- mouse littermates (8 to 12 weeks old) of Foxo3 heterozygous intercrosses were used in all experiments. Protocols were approved by the Institutional Animal Care and Use Committee of Mount Sinai School of Medicine.
Total bone marrow cells from WT and Foxo3-/- mice (n = 3) were collected in IMDM supplemented with 15% FBS and each genotype combined. Cells from the two samples were pre-incubated with 10% rat serum, stained for TER119 and CD44 cell surface markers and three erythroid populations comprising proerythroblasts (Gate I), basophilic erythroblasts (Gate II) and polychromatic erythroblasts (Gate III) were FACS sorted as previously described [34]. Total RNA was isolated using the RNeasy Micro Kit (Qiagen) and mRNAseq libraries were prepared using the Illumina True Seq RNA prep kit following manufacturer’s instructions. Samples were sequenced in parallel lanes in a Hi Seq™ 2000 platform (Illumina) to obtain more than 107 single end 100bp reads per sample.
The RNA-seq reads were mapped to mouse genome build mm9 using Tophat v2.0.3 [82] with default parameters. The mapped reads were processed to calculate the FPKM (fragments per kilo-base per million reads) and identify differentially expressed genes using Cufflinks v1.3.0 [83]. The statistics for data processing is shown in (S9 Table).
For WT analysis, only genes that fulfilled both of the following criteria were selected: (1) to show a FPKM > 1 in at least one of the three populations (pro-, baso- or polychromatic erythroblasts) and (2) to have an amplitude equal or above two. Amplitude was calculated as the maximum difference in expression between all three gates (e.g. Gate I vs Gate II, Gate II vs Gate III or Gate I vs Gate III). The resulted 5514 genes were then normalized using Z-score and clustered (Cluster 3.0 software k-means). For the comparison between WT and Foxo3-/- samples we followed the same strategy as stated above. Only those genes that displayed a FPKM ≥ 1 in at least one of the six samples and an amplitude ≥ 2 were considered as differentially expressed between WT and Foxo3-/- samples. In this case, amplitude was calculated by computing gene expression differences between WT and Foxo3-/- at each particular gate (e.g. WT Gate I vs. Foxo3-/- Gate I) and only genes with a two-fold difference in at least one of the comparisons were selected. The selected 3904 genes were then normalized using Z-score and clustered as indicated above.
FACS-sorted cells were washed once in PBS and directly diluted in RLT buffer after centrifugation. Total RNA was isolated using the RNeasy Micro and Mini kit (Qiagen) same method used for RNA-seq. First-strand cDNA was synthesized using the SuperScript IITM Reverse Transcriptase (Invitrogen). Quantitative RT-PCR was performed using SYBR Premix Ex TaqTm II (Tli RNase H Plus) (Takara, #RR820A) with technical duplicates using an ABI Prism 7900 HT Cycler (Applied Biosystems). Gene-specific primers spanning intron-exon boundaries were designed using Primer-Blast (NCBI). Primers used in Fluidigm microfluidics technology experiments were designed by Fluidigm. The PCR cycle parameters were as follows: 95°C for 10’ followed by 40 cycles at 95°C for 15” and 60°C for 1’. Results were obtained with the Sequence Detection System 4.2 software (Applied Biosystems) and further analyzed by the 2-ΔΔCt method. β actin was used as a loading control. Results are shown as fold-change relative to wild type controls. Primer specific sequences are listed in S10 Table.
Bone marrow single cell suspensions were obtained and washed in IMDM + 15% FBS. Cells were pre-incubated with 10% rat serum and stained with anti-CD44-APC (#559250 BD) and anti-TER119-PE (#553673 BD) antibodies. For DRAQ5 (#DR50200 Biostatus) analysis cells were stained for CD44-PE (#553134 BD) antibody and TER119-Biotin (#553672 BD) followed by Streptavidin-PE-Cy7 (#557598 BD). After washing, cells were incubated with DRAQ5 DNA binding dye (1/1000 dilution in PBS + 2% FBS) at 37°C and washed. Samples were analyzed in a FacsCanto flow cytometer (BD). Data was analyzed by FlowJo software (Treestar).
For MitoTracker Green or Red (#M7514, #M7512 Molecular Probes) analyses cells were pre-incubated with 10% rat serum, stained with anti-CD71-APC (#17–0711 eBiosciences) and incubated with mitotracker (20nM for 20 min at 37°C). Cells were then washed and analyzed by FACS.
Autophagy flux was measured by flow cytometry using the Cyto-ID autophagy detection kit (Enzo Life Sciences). Briefly, WT and Foxo3-/- bone marrow cells were preincubated with 10% rat serum and stained with TER119-PE and CD44-APC antibodies on ice. After washing, cells were resuspended at 106 cells/ml with 0.5 ml of IMDM supplemented with 15% FBS, 2 U/ml EPO and 50 ng/ml SCF in 24 well plates. Cells were cultured for 2 h at 37C° and chloroquine (50 μM) was added for 0, 10, 20, 40, 60, or 80 min. Cells were then kept on ice, washed twice in cold PBS + 5% FBS and resuspended in Cyto-ID solution consisting of 0.5 ml of PBS + 5% FBS + 0.25 μl of Cyto-ID. Cells were incubated for 30 min at room temperature in the dark, washed once with cold PBS + 5% FBS. Autophagosome content was determined by measuring the Cyto-ID (FITC) geometrical mean for each erythroid subpopulation. Autophagy flux was calculated by subtracting time 0 value from each of the other time points. All time points were analyzed for each bone marrow sample.
Bone marrow cells were pre-incubated with 10% rat serum and stained with TER119-FITC, CD71-PE, GR-1-EF450, and DRAQ5 (all from ebioscience). After washing, cells were fixed with 4% paraformaldehyde for 20 min at room temperature. Cells were centrifuged (30 seconds at 2000g), washed with PBS and centrifuged again. Images were acquired with an ImageStreamX and analyzed using IDEAS 6.0 software (Amnis/EMDmilipore). Single, focused, TER119+, Gr1- cells were further analyzed for cells size (area TER119 morphology mask), nuclear size (area DRAQ5 morphology mask). Enucleating cells were enumerated from TER119+ small cell and nuclear size cells (orthochromatic cells- See Fig 3D) by increased asymmetry of nucleus (delta centroid X/Y of TER119 morphology mask and DRAQ5 morphology mask) and a boundary of low signal between the emerging pyrenocyte and reticulocyte parts of the enucleating cell (TER119 intensity in Valley mask). Lobular shaped nuclei were identified using the 3-fold symmetry feature on the DRAQ5 morphology mask.
For Fluidigm dynamic array performance, specific target amplification (STA) was performed according to the manufacturer´s protocol (PN 100–3488 B1). Briefly, cDNA was pre-amplified using the TaqMan PreAmp Master Mix (Applied Biosystems) for the 96 genes of interest. The amplification parameters were as follows: 95°C for 10’, followed by 12 cycles at 95°C for 15” and 60°C for 4’. After STA, Exonuclease I treatment was performed as recommended by the manufacturer. Briefly, Exonuclease I and Exonuclease I buffer (New England Biolabs) were added to the STA samples, and samples were then incubated for 30’ at 37°C, followed by the enzyme inactivation at 80°C for 15’. Finally, to load the dynamic array IFC, samples were prepared with the SsoFast EvaGreen Supermix with Low ROX (Bio-Rad) and 20x DNA Binding Dye Sample Loading Reagent (Fluidigm). Primers were diluted with Assay Loading Reagent (Fluidigm) and DNA Suspension Buffer (Teknova). After priming the 96x96 chip in the IFC Controller MX, samples and primers were loaded into their respective inlets. The chip was then loaded by the IFC Controller MX. The chip was run following the GE 96x96 PCR+Melt v2.pcl protocol in the Biomark using the Data Collection Software (Fluidigm). Results were obtained with the Fluidigm Real-Time PCR Analysis software (Fluidigm) and further analyzed by the 2-ΔΔCt method. β actin was used as a loading control. Results shown as fold-change relative to Gate I wild type controls. Primer specific sequences are listed in S10 Table.
Total bone marrow cells were fixed in 4% paraformaldehyde (PFA) overnight on ice, washed with PBS, permeabilized with 0.3% Triton–100 for 20 min, and blocked in 3–4% BSA, 1% goat serum in PBS overnight at 4°C. Cells were then stained in solution as described previously [65] with anti-TER119-Alexa488, Rhodamine- Phalloidin and Hoechst. After washing in PBS, cells were cytospun into slides, coverslipped and images acquired on a Zeiss LSM780 confocal fluorescence microscope with a 100X/1.4 N.A. objective using zoom 1 or 2. The entire cytospin area was systematically scanned to avoid bias in collection. Images were processed using Volocity 6.1.1 and Adobe Photoshop, and images constructed in Adobe Illustrator.
Equal WT and Foxo3-/- cell numbers were resuspended directly in Laemmli sample buffer with DTT. Samples were then boiled at 95°C for 10 min and kept at -80°C until used. Electrophoresis was performed on a 15% SDS-PAGE and transferred to a PVDF Immobilon-P membrane (Millipore). Membranes were blocked with 5% BSA in PBS 0.1% Tween–20 and incubated overnight at 4°C in 1% BSA in PBS + 0.1% Tween–20 with either anti-LC3B (#3868 Cell Signaling) or anti-actin (sc–1616 Santa Cruz) antibodies and then incubated with the appropriate secondary antibodies conjugated to HRP at 1/5000 for 1 h at room temperature. Membranes were washed and developed using the ECL reagents (Pierce) with Blue sensitive films (Crystalgen). Films were then scanned and measurements were made using the Multi-Gauge software from Fujifilm following the manufacturer’s instructions. For flux measurements, TER119+ cells were isolated and plated in IMDM supplemented with 15% FBS, 2 U/ml EPO and 50 ng/ml SCF and kept in culture for 2 h. Chloroquine (50 μM) was then added to the cultures for the indicated time. Cells were then collected, washed in PBS and resuspended directly in Laemmli sample buffer and analyzed by Western blot.
ChIP was carried out in WT total bone marrow TER119+ cells as previously described [17]. Briefly, cells were cross-linked in 0.4% formaldehyde in PBS, and lysed (10 mM Tris-HCl pH8.0, 10 mM NaCl, 0.2% NP40). Lysate was sonicated for 30 cycles of 30 s on/30 s off at 4C° using a Bioruptor Standard sonication device (Diagenode). The cell lysate was then diluted in ChIP dilution buffer (20 mM Tris-HCl, pH 8.0, 2 mM EDTA 150 mM NaCl, 1% Triton, 0.01% SDS) and incubated at 4C° overnight with anti-FOXO3a antibody (Millipore #07–702) and Magna ChIPTM Protein A+G magnetic beads (Millipore #16–663). Beads were then washed (20 mM Tris-HCl, pH 8.0, 2 mM EDTA, 50 mM NaCl, 1% Triton, 0.1% SDS) and recovered. The antibody-protein-DNA complexes were reverse cross-linked for DNA isolation and quantitative PCR (qPCR) analysis. Foxo3-/- TER119+ cells were used as negative controls. Putative binding sites were located using MatInspector from Genomatix (http://www.genomatix.de/). Primer specific sequences are listed in S11 Table. See S12 Table for all antibodies.
Bone marrow cultures were performed using a modified protocol of [68]. Briefly, WT and Foxo3-/- bone marrow cells were enriched for hematopoietic progenitors using the EasySep Mouse Hematopoietic Progenitor Cell Enrichment Kit (#19756 Stemcell Technologies) and expanded at < 2x106 cells/ml in non-treated 24-well plates with erythroid expansion medium. Erythroid expansion media consists of Stem Span SFEM (StemCell Technologies) supplemented with 2 U/ml human recombinant EPO (Amgen), 100 ng/ml SCF (PreproTech), 40 ng/ml insulin-like growth factor–1 (PreproTech), 10−6 M dexamethasone (D2915; Sigma), 0.4% cholesterol mix (Gibco) and 1% penicillin/streptomycin (Gibco). Two days later, cells were washed twice with PBS and plated at a concentration <1x106 cells/mL in erythroid differentiation medium consisting of IMDM supplemented with 2 U/ml EPO, 100 ng/ml SCF, 10% Serum replacement (Invitrogen), 5% Platelet-Derived Serum, glutamine, MTG (1.27 μl in 10 ml of 1:10 dilution) and 10% Protein-Free Hybridoma Media. Cells were incubated in maturation medium for 3 days, with additional media added at day 3.
Retroviral supernatants of MIG and MIG-FOXO3 were produced as previously described [29,72]. Bone marrow expanding erythroid progenitors were resuspended in retroviral supernatants (multiplicity of infection of 10) on retronectin-coated dishes in expansion medium for 12 h. Suspension cells were then removed and only attached cells were collected by incubation with PBS and 5 mM EDTA at 37C° for 15 min. Cells were then washed, resuspended in expansion medium and incubated for another 24 h. Cells were then washed twice in PBS and cultured with erythroid differentiation medium as detailed above.
|
10.1371/journal.pntd.0007471 | Therapeutic efficacy of albendazole against soil-transmitted helminthiasis in children measured by five diagnostic methods | Preventive chemotherapy (PC) with benzimidazole drugs is the backbone of soil-transmitted helminth (STH) control programs. Over the past decade, drug coverage has increased and with it, the possibility of developing anthelmintic resistance. It is therefore of utmost importance to monitor drug efficacy. Currently, a variety of novel diagnostic methods are available, but it remains unclear whether they can be used to monitor drug efficacy. In this study, we compared the efficacy of albendazole (ALB) measured by different diagnostic methods in a head-to-head comparison to the recommended single Kato-Katz.
An ALB efficacy trial was performed in 3 different STH-endemic countries (Ethiopia, Lao PDR and Tanzania), each with a different PC-history. During these trials, stool samples were evaluated with Kato-Katz (single and duplicate), Mini-FLOTAC, FECPAKG2, and qPCR. The reduction rate in mean eggs per gram of stool (ERR) and mean genome equivalents / ml of DNA extract (GERR) were calculated to estimate drug efficacy.
The results of the efficacy trials showed that none of the evaluated diagnostic methods could provide reduction rates that were equivalent to a single Kato-Katz for all STH. However, despite differences in clinical sensitivity and egg counts, they agreed in classifying efficacy according to World Health Organization (WHO) guidelines. This demonstrates that diagnostic methods for assessing drug efficacy should be validated with their intended-use in mind and that other factors like user-friendliness and costs will likely be important factors in driving the choice of diagnostics. In addition, ALB efficacy against STH infections was lower in sites with a longer history of PC. Yet, further research is needed to identify factors that contribute to this finding and to verify whether reduced efficacy can be associated with mutations in the β-tubulin gene that have previously been linked to anthelmintic resistance.
ClinicalTrials.gov NCT03465488.
| During the last decade, the scale of deworming programs that aim to eliminate the morbidity caused by intestinal worms has increased to a level that is unprecedented in history. It is therefore of utmost importance to monitor any change in therapeutic efficacy that may arise from emerging drug resistance. Currently, a variety of novel methods have been described, but it remains unclear whether they can be used for monitoring drug efficacy. We applied different diagnostic methods to measure the efficacy of a commonly administered drug in deworming programs in 3 countries with different historical exposure to deworming programs. Compared to the standard diagnostic method, all diagnostic methods revealed good agreement in classifying the therapeutic efficacy according to World Health Organization guidelines, despite clear differences in diagnostic performance. We also noticed that the drug efficacy was lower in countries where drug pressure has been high. However, more research is necessary to identify factors that explain this variation in drug efficacy, including but not limited to the frequency in mutations in genes that are known to be linked with anthelmintic resistance.
| Infections with soil-transmitted helminths (STHs; Ascaris lumbricoides, Trichuris trichiura, Necator americanus and Ancylostoma duodenale) are responsible for the highest burden among all neglected tropical diseases. Recent global estimates indicate that in 2015, more than 1.6 billion people were infected with at least one of the four STH species [1], resulting in a global burden of approximately 1.9 million disability-adjusted life years [2]. Preventive chemotherapy (PC) or the periodical administration of a single-oral dose of albendazole (ALB; 400 mg) or mebendazole (MEB; 500 mg) to preschool- (PSAC) and school-aged children (SAC) is the main strategy to control the morbidity caused by STHs [3]. In 2017, global coverage of PC in at-risk populations was nearly 70%, though the target is to reach 75% coverage by 2020, and to eventually eliminate soil-transmitted helminthiasis as a public health problem [4–6]. The latter target is defined by reaching less than 1% moderate and heavy intensity infections in SAC [5].
The downside of these increased control efforts is that resistance to anthelmintic drugs, such as ALB and MEB, is likely to develop. Both drugs belong to the same drug class (benzimidazole (BZ) drugs) and share the same mode of action. Moreover, they are administered in single doses that usually do not achieve 100% efficacy [7–10]. Should anthelmintic resistance (AR) against these BZ drugs eventually emerge and spread, it will jeopardize PC-based control of STH due to the few acceptable alternative treatment options [11, 12]. All this reinforces the urgent need to promote accessibility of anthelmintic drugs with different modes of action, alone or in combination, and a thoroughly designed surveillance system that detects any changes in anthelmintic drug efficacy arising through the evolution of AR.
The World Health Organization advises to monitor drug efficacy in case treatment failure is suspected or–regardless of suspected drug failure–when drugs have been administered in PC-programs for at least four years [13]. To monitor the efficacy of anthelmintic drugs against STHs, WHO currently recommends measuring the reduction in number of STH eggs excreted in stool after drug administration (egg reduction rate, ERR) using either a single Kato-Katz thick smear or the McMaster method [13].
Recently, novel methods have been introduced in the field of STH diagnostics, including Mini-FLOTAC [14, 15], FECPAKG2 [16, 17] and the DNA-based diagnostic methods such as quantitative PCR (qPCR) [18–20]. Each of these methods offers one or more advantages over the recommended methods, pertaining to increased clinical sensitivity [15, 18, 21–23] and specificity (qPCR is able to differentiate different helminths at the species level) [24–27], quality assurance (FECPAKG2 automatically stores images of each sample which can be consulted at any time [17, 28]; qPCR includes internal controls within each run [29]), flexibility as to when samples are examined (for both Mini-FLOTAC and qPCR, stool can be preserved for analysis at a later time point [23, 25, 30–32]). Although each of these novel methods has recently been used to evaluate drug efficacy [16, 33, 34], there remains a paucity of studies that perform a head-to-head comparison of the drug efficacy obtained by different diagnostic methods. Moreover, these studies tested the hypothesis that the methods provide significantly different ERR estimates. Rather, the correct hypothesis is to assess whether these differences are within the bounds of equivalence. As illustrated in supplementary information (S1 Info), the absence of a significant difference does not imply equivalent ERR estimates nor does the presence of a significant difference rule out equivalent ERR results.
Therefore, in this study we compared the equivalence in ALB efficacy measured by duplicate Kato-Katz thick smear, Mini-FLOTAC, FECPAKG2 and qPCR in a head-to-head comparison with a single Kato-Katz thick smear. For this, a drug efficacy trial with ALB was performed in three different countries (Ethiopia, Lao PDR and Pemba (Tanzania)) with different historical levels of drug exposure.
The study protocol has been reviewed and approved by the Institutional Review Board (IRB) of the Faculty of Medicine and Health Sciences of Ghent University, Belgium (Ref. No B670201627755; 2016/0266). The trial protocol was subsequently reviewed and approved by the IRBs associated with each trial site (Ethical Review Board of Jimma University, Jimma, Ethiopia: RPGC/547/2016; National Ethics Committee for Health Research, Vientiane, Lao PDR: 018/NECHR; Zanzibar Medical Research and Ethics Committee, United Republic of Tanzania: ZAMREC/0002/February/2015 and the IRB from Centro de Pesquisas René Rachou, Belo Horizonte, Brazil: 2.037.205). The trial was retrospectively registered on Clinicaltrials.gov (ID: NCT03465488) on March 7, 2018.
Parent(s)/guardians of participants signed an informed consent document indicating that they understood the purpose and procedures of the study, and that they allowed their child to participate. If the child was ≥5 years, he or she had to orally assent in order to participate. Participants of ≥12 years of age were only included if they signed an informed consent document indicating that they understood the purpose and the procedures of the study, and were willing to participate.
The selection of the study sites was based on their experience in assessing drug efficacy, evaluating the performance of diagnostic methods, the availability of well-equipped diagnostic facilities and skilled personnel, and PC-history [35]. Based on the reported national coverage of drug administration to both PSAC and SAC for the last 5 years (2009–2014; Preventive Chemotherapy Database of the WHO), the site in Ethiopia was considered to have experienced a low drug exposure, the site in Lao PDR a medium drug exposure and the site in Pemba (Tanzania) a high drug exposure prior to the start of the trials [35] (Table 1). Note that the initial study protocol included a study site in Brazil. However, due to the low number of cases on which not all diagnostic methods were performed, the site was excluded from this report.
The trials were designed to assess an equivalence in treatment efficacy of a single oral dose of 400 mg ALB against STH infections in SAC measured by a variety of diagnostic methods. The study focused on SAC (age 5–14) since they are the major target of PC programs, and they usually represent the group with highest worm burdens for A. lumbricoides and T. trichiura [36]. Subjects were not included in the study if they could not provide a stool sample at baseline or follow-up and had active diarrhea or any other acute medical condition at baseline. Children with a known hypersensitivity to ALB or MEB, who received anthelmintic treatment within 90 days prior to the start of the trial were and did not swallow the entire drug tablet or vomited within four hours following drug ingestion were also excluded from the study.
At the start of each trial, schools were visited by the local principal investigator and a team of field officers, who explained the planned trial and sampling method to the parents and teachers and the children. At baseline, SAC were asked to provide a fresh stool sample, after which they were administered a single oral dose of 400 mg ALB under supervision. The ALB used in the different studies originated from the same production batch (GlaxoSmithKline Batch N°: 335726) and was provided by WHO. All collected stool samples were kept in a cooler with ice packs while transported to the laboratory, where they were processed on the same day of collection. Stool samples were processed to determine the fecal egg counts (FECs; expressed in eggs per gram of stool (EPG)) for each STH using Kato-Katz (single and duplicate), Mini-FLOTAC and FECPAKG2. As FECs of the latter technique could not finalized on the day of sample collection (see section Diagnostic methods), results of the FECPAKG2 technique were not used to select individuals for inclusion at follow-up. Aliquots of a subset of the baseline samples were preserved in ethanol for molecular analysis. Preliminary data has indicated that downstream analysis of STH β-tubulin genes was very challenging when egg concentration was low (data not published). Therefore, only samples with a FEC of ≥150 EPG for at least one STH species were withheld for further molecular analysis. Fourteen to 21 days after drug administration, a second stool sample was collected from all the children that were found positive for any STH by duplicate Kato-Katz or Mini-FLOTAC at baseline. Stool samples collected at follow-up were again examined by Kato-Katz (single and duplicate), Mini-FLOTAC and FECPAKG2. Aliquots from all follow-up samples were preserved for further molecular analysis regardless of the FECs.
A sample size was calculated to test the hypothesis that FECPAKG2, Mini-FLOTAC and duplicate Kato-Katz provide equivalent drug efficacy results measured by ERRs compared to a single Kato-Katz. This sample size calculation did not include the qPCR method. Given the differences in drug efficacy of ALB across the STH species [8, 9] (A. lumbricoides: ~99%, hookworms: ~96%, T. trichiura: ~65%), a level of equivalence that is acceptable for T. trichiura may not be acceptable for A. lumbricoides. We therefore applied a level of equivalence that was tailored to the different STH species. The level of equivalence for A. lumbricoides, hookworms and T. trichiura was set arbitrarily at -/+2.5, -/+5.0 and -/+10-point percentage respectively. This means that a method provides equivalent drug efficacy estimates as single Kato-Katz if the confidence intervals surrounding the mean difference in drug efficacy does not exceed these set of values (S1 Info). To calculate the corresponding sample size for each of the STH species, a simulation study was performed that considered (i) the variation in ERR and baseline FECs both across and within STH species, (ii) the variation in FECs introduced by the egg counting process, (iii) the paired ERR results across diagnostic methods, and (iv) a post-hoc correction for a pair-wise comparison. Based on this simulation, at least 110, 100 and 12 complete cases are required for T. trichiura, hookworm and A. lumbricoides, respectively. A detailed description of the sample size calculation is available elsewhere [35].
Upon arrival in the laboratory, stool samples were homogenized with a wooden spatula and subsequently subjected to microscopic examination by means of single and duplicate Kato-Katz, Mini-FLOTAC and FECPAKG2. Two aliquots of 0.5 g stool were also preserved in an Eppendorf tube containing 1 ml of absolute ethanol for later DNA extraction and qPCR analysis. Detailed standard operating procedures (SOPs) for the different diagnostic methods were published earlier [35, 37]. Here we briefly mention the most important steps for each of the methods.
For Kato-Katz, two slides were prepared (slide A and B) and examined for the presence of STH eggs within 30–60 min following preparation. The results of slide A represented the results of a single Kato-Katz and egg counts were multiplied by 24 to obtain the FECs (expressed as EPG). The sum of the egg counts obtained after reading slide A and B represented the results for duplicate Kato-Katz and were multiplied by 12 to obtain the FECs.
For Mini-FLOTAC, we homogenized 2 g of fresh stool with 38 ml of flotation solution (saturated salt solution, density = 1.20) in the Fill-FLOTAC recipient [15]. After transferring the suspension into the two chambers of the Mini-FLOTAC device, the device was placed on a horizontal surface for 10 min after which the reading disk was translated. Finally, both Mini-FLOTAC chambers were screened for the presence of STH eggs. The number of eggs counted were multiplied by 10 to obtain the FECs.
The FECPAKG2 method was performed as described by Ayana et al. [17]. Briefly, stool was homogenized in tap water in a Fill-FLOTAC device [15], after which it was transferred into a FECPAKG2 sedimenter to allow STH eggs to sediment. The following day, the supernatant was poured off and saturated saline solution (specific density = 1.2) was added to the remaining slurry. The whole content of the sedimenter was then poured into a FECPAKG2 filtration unit from which 2 separate aliquots were taken and transferred to 2 wells of a FECPAKG2 cassette. Following an accumulation step of at least 20 minutes, the cassettes were placed in the Micro-I device for image capture. The device automatically imaged both wells and stored the images prior to uploading them to the FECPAKG2 server. Finally, the mark-up technician identified and counted any STH eggs present in the images using specialized software. Mark-up of the images was not performed on the day of examination and hence the results were not used to select individuals for inclusion in follow-up. Results of the mark-up were saved automatically for reporting and analysis. For FECPAKG2 the eggs counted in both wells were multiplied by 34 to calculate the FECs.
For quality control purposes, a predefined, randomly selected subset of samples (10% of the total number of samples) was re-evaluated by each of the three egg count methods. To this end, a senior researcher, who was blinded to the initial FECs, re-counted STH eggs across all three egg count methods. A third examiner would re-count STH eggs in case of discrepancies. An in-depth analysis of these quality control results will be published in a separate manuscript.
In order to perform qPCR, DNA was extracted from the preserved stool samples and analyzed for the presence of DNA of STH at the Laboratory for Medical Microbiology and Immunology (Elisabeth-TweeSteden Hospital, Tilburg, The Netherlands) as part of two multiplex qPCR assays [35]. The variation between runs was monitored by means of the Cq values of the positive controls (DNA template for each STH species). We defined the variation between runs negligible when the difference in Cq values of the positive controls between runs did not exceed 1. The inhibition of the qPCR assay was controlled by adding a known quantity of phocine herpes virus DNA in each DNA extract and by subsequently quantifying this virus’ DNA by qPCR. Inhibition was present in the sample when the difference in Cq-value between the virus’ DNA in a clinical sample and a pure virus DNA sample did not exceed 1. We did not observe a difference in Cq across controls of more than 1 Cq, nor did we observe inhibition of the qPCRs in any of the samples (quality control results will be published in a follow-up manuscript). For each target species, qPCR results were expressed as genomic equivalents per ml of DNA extract (GE/ml). The reported qPCR results for hookworms were calculated as the sum of GE/ml of both hookworm species (Ancylostoma and Necator americanus).
A sample was considered positive for a STH infection if it tested positive on at least one diagnostic method (duplicate Kato-Katz, Mini-FLOTAC, FECPAKG2 or qPCR). The efficacy of a single oral dose of 400 mg ALB is reported separately for each STH species and for each microscopic method by means of ERR, using the following formula: ERR (%) = 100% x [1- (arithmetic mean FEC at follow-up / arithmetic mean FEC at baseline)]. For qPCR, a similar formula was used, where FEC was replaced by DNA concentration (GE/ml), yielding the genome equivalent reduction rate (GERR): GERR (%) = 100% x [1- (arithmetic mean DNA concentration at follow-up / arithmetic mean DNA concentration at baseline)]. A bootstrap analysis was used to determine the corresponding 95% confidence intervals (95%CI) around the (G)ERR point estimate for each diagnostic method and the difference in drug efficacy compared to a single Kato-Katz across diagnostic methods. A permutation test was used to assess the equivalence in (G)ERR between single Kato-Katz thick smear and either duplicate Kato-Katz, Mini-FLOTAC, FECPAKG2 or qPCR. Bonferroni’s correction was applied for multiple comparison between methods (level of significance was set at 0.0125 = 0.05 / 4 comparisons).
The (G)ERR point estimates were used to classify drug efficacy as ‘satisfactory’, ‘doubtful’ or ‘reduced’ following the WHO criteria recommended for a single Kato-Katz [13] (Table 2). In addition, the agreement between a single Kato-Katz and the other methods in the assignment of drug efficacy into ‘satisfactory’, ‘doubtful’ and ‘reduced’ was evaluated by Fleiss’ kappa statistic (κFleiss). The value of this statistic indicates a slight (κFleiss <0.2), fair (0.2≤ κFleiss <0.4), moderate (0.4≤ κFleiss <0.6), substantial (0.6≤ κFleiss, <0.8) or an almost perfect agreement (κFleiss ≥0.8).
Finally, we also assessed the distribution of individual responses measured across the five diagnostic methods. Individual ERR (iERR) were calculated using the following formula: iERR = 100% x [1- (FEC at follow-up of individual i / FEC at baseline of individual i)]. Individual GERR (iGERR) were calculated using the following formula: iGERR = 100% x [1- (DNA concentration at follow-up of individual i / DNA concentration at baseline of individual i)]. We classified the individual response for each STH species and for each method into ‘cured’ (no eggs/DNA was found in follow-up sample), ‘satisfactory’, ‘doubtful’ or ‘reduced’ (see Table 1) and ‘absent’ (drug efficacy was below zero due to higher egg counts or DNA-concentration in the follow-up sample than in baseline sample). Subsequently, the agreement between a single Kato-Katz and the other methods in the assignment of individual drug efficacy was assessed by Fleiss’ kappa statistic (κFleiss). All statistical analyses were performed in R [38]. Graphs were produced using R.
The number of children that were withheld after recruitment and at baseline or follow-up visits, and those that were eventually incorporated in the final statistical analysis are summarized in Fig 1. Complete data was available for 645 children across three of the four study sites (Ethiopia: 161 cases; Lao PDR: 239 cases; Pemba (Tanzania): 245 cases).
With the exception of Pemba (Tanzania), where more females (n = 137) were included than males (n = 108), the sex ratio (males:females) was approximately 1:1 in all study sites. The median age (25th and 75th quantile) of the children across the three study sites equaled 11.0 years (9.0; 12.0). The participants in Ethiopia (9.0 years [8.0; 10.0]) were slightly younger than those in Lao PDR (12.0 years [11.0; 13.0]) and Pemba (Tanzania; 11.0 years [10.0; 12.0]). In total, there were 441 complete cases for A. lumbricoides (Ethiopia: 137, Lao PDR: 111, Pemba (Tanzania): 193), 456 for T. trichiura (Ethiopia: 106, Lao PDR: 105, Pemba (Tanzania): 245) and 457 for hookworm (Ethiopia: 90, Lao PDR: 228, Pemba (Tanzania): 139). The qPCR results showed that all individuals who excreted hookworm eggs were infected with N. americanus. Both N. americanus and Ancylostoma DNA was detected in only 8 individuals from one school in Pemba (Tanzania). The qPCR results for hookworm used in our calculations represent the combined GE/ml detected for both species. Due to the nature of the school selection procedure (prioritization of schools where STH prevalence was expected to be moderate to high and premature discontinuation of recruitment in a school when the prevalence of STH was low), the number of complete cases is not equally distributed across the schools. A minority of the schools actually provide the majority of the infected children (S2 Info).
Tables 3–5 describe the efficacy of ALB measured by the different diagnostic methods across the three study sites for A. lumbricoides, T. trichiura and hookworm, respectively. For A. lumbricoides infections, efficacy of a single-oral dose of 400 mg ALB estimated by single Kato-Katz was high (ERR >95%) across the different study sites (Table 3). This high drug efficacy was confirmed by the three microscopic methods as well as by qPCR. The absolute point percent difference in drug efficacy did not exceed 2% (duplicate Kato-Katz: 0–0.1%; Mini-FLOTAC: 0.1%– 0.6%; FECPAKG2: 0.0%– 1.8%; qPCR: 0.0%– 0.8%). All diagnostic methods provided significantly equivalent estimates of drug efficacy compared to a single Kato-Katz (i.e. the 95%CI around the difference in drug efficacy between diagnostic methods did not include 2.5%), except for FECPAKG2 in Pemba (Tanzania) and qPCR in both Ethiopia and Pemba (Tanzania), where evidence of equivalent drug efficacy was marginal since CIs included the 2.5% bounds of equivalence (FECPAKG2: [-5.7%; 0.8%]; qPCR: [-6.5%; 5.5%] in Pemba (Tanzania) and [-0.1%; 4.3%] in Ethiopia).
For T. trichiura infections, ALB efficacy estimations varied significantly depending on the study site and the diagnostic method that was used (Table 4). Estimations obtained by single Kato-Katz were 52.9% in Ethiopia, 36.7% in Lao PDR and -11.2% in Pemba (Tanzania). There was a large deviation from the drug efficacy measured by single Kato-Katz and those based on the other diagnostic methods. The absolute point percent difference between single Kato-Katz and the other methods was the smallest for duplicate Kato-Katz (4.8%– 6.3%) and the largest for qPCR (14.6%– 60.2%). None of the methods provided equal drug efficacy results (CI around the difference in drug efficacy between diagnostic methods included the 10% bounds of equivalence). Moreover, clear difference could be noted across the different methods. In contrast to the other methods, drug efficacy measured by a duplicate Kato-Katz was marginally equivalent to those obtained by single Kato-Katz with the CI just including the 10-point percent (Ethiopia: [-0.6%; 10.4%]; Lao PDR: [-12.0%; 3.1%] and Pemba (Tanzania): [-14.9%; 1.9%]).
For hookworm infections (Table 5), the drug efficacy measured by single Kato-Katz was high (>95%) in both Ethiopia (96.3%) and Lao PDR (96.1%), but moderate in Pemba (Tanzania) (84.2%). Overall, these drug efficacy estimates were confirmed by the other diagnostic methods. The absolute point percent differences in drug efficacy did not exceed 7% (duplicate Kato-Katz: 0%– 0.6%; Mini-FLOTAC: 0.9%– 2.7%; FECPAKG2: 1.3%– 6.1%; qPCR: 1.5%– 4.6%). A duplicate Kato-Katz provided equivalent drug efficacy estimates across all study sites (the CI around the difference in drug efficacy between diagnostic methods did not include the 5% bounds of equivalence, Ethiopia: [95%CI: -0.5; 0.9], Lao PDR: [95%CI: -0.6; 0.2]); Pemba (Tanzania) [95%CI: -1.5; 3.2]). The only other significant equivalent drug efficacy result was found for FECPAKG2 in Lao PDR (95%CI: -3.7; 0.9). The remaining pair-wise comparisons in both Ethiopia (Mini-FLOTAC: [95%CI: -2.1; 9.0]; FECPAKG2: [95%CI: -8.1; 0.9]; qPCR: [95%CI: -2.4; 7.6]) and Lao PDR (Mini-FLOTAC: [95%CI: -0.6; 6.0%]; qPCR: [95%CI: -2.1; 7.7]) indicated that drug efficacies were marginally equivalent, the 95%CI just including the 5% equivalence threshold. This was in contrast to the findings from Pemba (Tanzania) where no clear evidence of equivalent drug efficacies for Mini-FLOTAC [95%CI: -11.8; 9.0%], FECPAKG2 [95%CI: -15.9; 3.0] and qPCR [95%CI: -10.8; 31.4] was observed.
In summary, a duplicate Kato-Katz provided drug efficacy significantly or marginally equivalent to a single Kato-Katz for the three STH species in all study sites. For the other methods, the equivalence of drug efficacy varied by STH species and study site.
Table 6 summarizes the agreement in classifying the efficacy of ALB between single Kato-Katz and the other diagnostic methods across the three study sites for each STH species. For both duplicate Kato-Katz (κFleiss = 0.81, p <0.001) and qPCR (κFleiss = 0.84, p <0.001) there was almost a perfect agreement (κFleiss ≥0.8). Both these methods agreed with a single Kato-Katz in 7 out of the 9 observations. For FECPAKG2 (κFleiss = 0.65, p = 0.01) there was a substantial agreement (0.61 ≤ κFleiss <0.81), for MiniFLOTAC (κFleiss = 0.60, p = 0.03) there was a moderate agreement (0.41 ≤ κFleiss <0.61).
Figs 2–4 provide an overview of individual drug responses to ALB for the five diagnostic methods across the different sites for A. lumbricoides, T. trichiura and hookworm, respectively. The top three bar plots illustrate the different classifications of individual drug efficacy for that helminth species detected in the study population in each of the three countries by five different diagnostic methods. Based on their i(G)ERR, individuals are classified into one of seven different categories, which in turn correspond with a specific color, ranging from dark green (completely cured, i(G)ERR = 100%) to dark red (eggs/DNA detected at follow-up but not at baseline, i(G)ERR = -∞). The grey part of the bar represents the individuals for whom no efficacy could be calculated since no eggs or DNA was detected for that species at baseline and follow-up. The bottom three bar plots represent the individual drug efficacy for those 770 individuals for whom i(G)ERR was measured with each of the five diagnostic methods.
Generally, these figures highlight three important findings. First, they confirm the distinct differences in sensitivity across the diagnostic methods. FECPAKG2 was previously evaluated as being less sensitive than Kato-Katz, while qPCR was found to have superior sensitivity for all STH [16, 26, 39]. This is also supported by the results of our study, where we noticed high numbers of false negative test results at baseline and follow-up for FECPAKG2. When applying FECPAKG2, drug efficacy could not be measured in 438 (32.3%) of the 1,354 individuals with STH infection, because of false negative results at baseline (A. lumbricoides: 148/441, T. trichiura: 135/456 and hookworms: 155/457). In contrast, when applying qPCR, individual drug efficacy could not be measured in only 56 (4.1%) of the 1,354 individuals because of false negative results at baseline (top graphs of Figs 2–4: A. lumbricoides: 22/441, T. trichiura: 11/456 and hookworm: 23/456).
Second, they indicate that there are a number of cases where eggs or DNA were found at follow-up, but not at baseline (this mathematically results in an infinite increase of eggs or DNA at follow-up or an individual drug efficacy of minus infinity (dark red portion of bars in upper panels). Overall, these types of cases were observed by at least one diagnostic method in 8% of the total number of cases (n = 1,354), but differences across diagnostic methods and STH were observed. They were more prevalent when the FECPAKG2 method was used (5.5% = 75/1,354). For the other diagnostic methods, the proportion of samples that were positive at follow-up but negative at baseline did not exceed 1.3% (single Kato-Katz: 1.2%; duplicate Kato-Katz: 1.3%; Mini-FLOTAC: 1.0%; qPCR: 1.2%), the majority being T. trichiura cases (17.2% = 78/456). The cases were less frequently observed for hookworm (5.3% = 24/457) and A. lumbricoides (1.8% = 8/441).
Third, they indicate that variation in individual drug response across STH and countries is similar across diagnostic methods. This is most obvious when we focus on the cases for which an individual drug efficacy response was available for all methods (bottom panels of Figs 2–4). When employing a single Kato-Katz, the highest drug efficacy was observed for A. lumbricoides followed by hookworms and T. trichiura. For A. lumbricoides, 96.9% (= 250/258) of the individuals showed a drug response that was at least satisfactory (light green + dark green). For hookworm, this proportion equaled 79.7% (= 204/256), whereas for T. trichiura this was only 34.4% (= 88/256).
For A. lumbricoides, the proportions of individuals with at least satisfactory drug efficacy (light green + dark green) as measured by single Kato-Katz were comparable across the 3 countries (Ethiopia: 100% (= 80/80); Lao PDR: 94.0% (= 47/50) and Pemba (Tanzania): 96.1% (= 123/128)) (bottom graphs of Figs 2–4). For T. trichiura, fewer individuals showed satisfactory drug efficacy to ALB in Pemba (Tanzania) compared to Lao PDR or Ethiopia (Pemba (Tanzania): 28.6% (= 59/206) vs. Lao PDR: 51.7% (= 15/29) or Ethiopia: 66.7% (= 14/21)). When investigating the individual ALB response to hookworm, fewer individuals show satisfactory drug efficacy in Pemba (Tanzania) (66.1% (= 37/56)) compared to Lao PDR (82.5% (= 137/166)) or Ethiopia (88.2% (= 30/34)). Regardless of the diagnostic method used, the same trends in individual efficacy were apparent across STH species and countries.
Cross tables were made to gain more insights into the agreement between individual drug response across the different methods (S3 Info). These tables illustrate the agreement of calculated i(G)ERR using single Kato-Katz and the four other methods for A. lumbricoides, T. trichiura and hookworms, respectively. For a duplicate Kato-Katz an almost perfect agreement (κFleiss ≥ 0.80) was observed for each of the STH (A. lumbricoides: 0.96 T. trichiura: 0.91; hookworm: 0.96, p <0.001). For Mini-FLOTAC, there was an almost perfect agreement for A. lumbricoides (κFleiss = 0.88, p <0.001) and a substantial agreement for the other two STH (T. trichiura: 0.62; hookworm: 0.79, p <0.001). For FECPAKG2, there was moderate agreement for A. lumbricoides (κFleiss = 0.55, p <0.001) and a fair agreement for the remaining STH (T. trichiura: 0.31; hookworm: 0.39, p <0.001). For qPCR, there was a substantial agreement for hookworms (κFleiss = 0.61, p< 0.001), moderate agreement for both A. lumbricoides (κFleiss = 0.59, p <0.001), and a fair agreement for T. trichiura (κFleiss = 0.36, p <0.001).
The present study evaluated the efficacy of ALB against STH infections in three different endemic study sites using five different diagnostic methods. The rationale for this study was twofold. First, we wanted to evaluate if the different diagnostic methods provide equivalent drug efficacy results compared to a single Kato-Katz (the WHO recommended method) and to ultimately make recommendations on which diagnostic methods can be used for assessing drug efficacy. The second goal was to evaluate the ALB efficacy against STH in all three study sites with varying anthelmintic drug pressure histories. The presented study is unique in a number of ways. It is the first study that performs a multi-country, standardized, head-to-head comparison of established (single and duplicate Kato-Katz) and novel microscopic (Mini-FLOTAC and FECPAKG2) and molecular (qPCR) diagnostic methods for assessing drug efficacy against STHs. This study was not designed to prove that ERR estimates differ across methods, rather it verified whether methods are equivalent in assessing drug efficacy, which, as illustrated in S1 Info, is a subtle, but important difference.
We found that none of the evaluated tests provided equivalent results to those obtained by single Kato-Katz for all three STH. However, this conclusion needs to be interpreted with some caution. First, the species-specific levels of equivalence (the predefined bounds of equivalence) are arbitrary and likely to be set too strict. For instance, setting the level of equivalence at 10% for T. trichiura might be too strict for Pemba (Tanzania) given that ALB efficacy measured by duplicate Kato-Katz was -11.2%. On the other hand, the sample size was initially determined to compare the microscopic methods only (See [35]). By adding the qPCR results to this comparison, we increased the number of comparisons from 3 to 4. Consequentially, the level at which significant equivalence could be shown was reduced (0.05/4 = 0.0125 instead of 0.05/3 = 0.0166). Moreover, the sample size calculation was performed bases on certain assumptions regarding the ERR and FECs across and within STH species, which might have resulted in an underestimation of the true variation in the population [40, 41]. However, despite the lack of equivalence, for most methods there was relatively good agreement in classifying ALB efficacy according to WHO guidelines. This suggests that each method holds promise for the assessment of drug efficacy in the context of assessing drug efficacy within STH control programs.
The results of the present study highlight that the impact of the diagnostic sensitivity on ERR results is minimal (Tables 3–5 and Figs 2–4). Although there were substantial differences in FECs across the different microscopic methods, this did not have a major impact on the equivalence of ERR. For example, the FECs for A. lumbricoides based on single Kato-Katz were at least double of those based on FECPAKG2 across all study sites (Ethiopia: 7,870 EPG vs. 1,622 EPG; Lao PDR: 13,029 EPG vs. 2,711 EPG; Pemba (Tanzania): 14,372 EPG vs. 6,322 EPG), yet at each study site both methods agreed that the efficacy is still satisfactory. This agreement in drug efficacy, despite the clear differences in diagnostic sensitivity and FECs, are in line with previous studies involving both animal [42] and human helminths [43–45], and underscore that diagnostic methods need to be to be validated for their intended-use. Moreover, it highlights that other aspects such as user-friendliness and operational costs might become pivotal factors when deciding to recommend or use any given method. Additionally, it should be noted that our findings for qPCR do not necessarily apply for other qPCR assays, given that the plethora of described qPCR assays for STHs can differ substantially in performance. It is also important to point out that these findings are based on results obtained in sites where STH prevalence and intensities of STH infections are still relatively high. It is possible that the impact of the diagnostic sensitivity of a method on ERR calculations, as illustrated in animals, will increase when working in settings with very low infection intensities [46, 47]
We strategically selected the different study sites to cover a wide range of drug pressure. In our study, the study site in Ethiopia was least exposed to BZ drugs, followed by the one in Lao PDR. On Pemba (Tanzania), BZ drugs had been most frequently administered. When focusing on the drug efficacy estimated by single Kato-Katz, there was an obvious trend between the drug pressure and drug efficacy for each of the three STH species. The ERRs dropped as a function of historic drug pressure. The declining trend was most pronounced for T. trichiura, for which ERR ranged from 52.9% in Ethiopia over 36.7% in Lao PDR to -11.2% in Pemba (Tanzania). For both A. lumbricoides and hookworm, the efficacy of ALB was highest in both Ethiopia and Lao PDR (A. lumbricoides: ~ 99% and hookworms: ~96%), and lowest on Pemba, Tanzania (A. lumbricoides: 96.8%; hookworms: 84.2%). Whether this reduced drug efficacy on Pemba (Tanzania) is indicative for the emergence of AR remains to be verified since it has been shown that other factors may contribute to a reduced efficacy. For example, it has been described that the efficacy of ALB against T. trichiura infections declines as a function of increasing infection intensity [48]. This also seems to be the case in the present study, where we notice a trend between average infection intensity (Pemba (Tanzania): 3,111 EPG; Lao PDR: 357 EPG; Ethiopia: 207 EPG) and reduced drug efficacy (Pemba (Tanzania): -11.2% ERR; Lao PDR: 36.7% ERR; Ethiopia: 52.9% ERR by single Kato-Katz). It is also possible that both these processes occur simultaneously and mutually enhance the noticed effects of reduced drug efficacy. Poor drug efficacy could result in increasing transmission and more subjects being infected with a large number of worms.
To further assess the emergence of AR, we will analyze the frequency of known single nucleotide polymorphisms (SNPs) in the β-tubulin gene at codons 167 (TTC to TAC), 198 (GAA to GCA) and 200 (TTC to TAC) in a subset of the baseline and follow-up samples [35]. Subsequently, individual-based drug efficacy models will be built to explore the association between the frequency of these SNPs and other factors, including, but not limited to, infection intensity [49]. The results of this analysis will be presented and discussed in detail in a follow-up paper. At present, only a few studies with small sample sizes originating from a limited number of endemic areas have been performed to assess the association between β-tubulin SNPs and reduced anthelmintic efficacy in human STH [33, 50–55]. In these studies, it was noted that polymorphisms were predominantly found in codon 200 of the β-tubulin gene and that these mutations were more abundant in a T. trichiura worm population following drug administration. Nevertheless, no association could ever be proven with reduced drug efficacy in any STH species. Overall, there are limited reports of declining or poor drug efficacy [9, 33, 56, 57]. Of note, some of these studies were flawed in terms of their design or analysis. For example, Krücken and colleagues reported a poor efficacy of ALB against A. lumbricoides infections in Rwandan SAC, but the study findings might have been negatively affected by the fact that follow-up sampling occurred too soon after drug administration (7–10 days), which likely led to the detection of eggs from dying or degenerating worms [58].
Interestingly, Pemba was the only site where both hookworm species were detected by qPCR. In eight children, mixed infections with Ancylostoma spp. and N. americanus were identified, confirming the finding by Albonico et al. [59]. Follow-up samples of these eight individuals were all negative for Ancylostoma spp. (cure rate (CR) of 100%), while two still excreted Necator DNA (CR = 75%). Although this was observed in only eight cases, it supports the findings on the efficacy of ALB to different helminth infections presented by Horton [60] who reported a notably lower CR for Necator infections (CR = 75%, 30 studies) compared to Ancylostoma spp. infections (CR = 92%, 23 studies). Given the seemingly differential susceptibility of both hookworm genera to ALB, it is important to differentiate hookworm infections in order to have correct efficacy estimates for each species. This is of particular interest when the possible contribution of zoonotic A. ceylanicum infections from animal reservoirs to the observed drug efficacy is investigated.
The present study investigated the equivalence of five different diagnostic tools for the evaluation of anthelmintic efficacy. None of the evaluated tests provided equivalent results to those obtained by the currently recommended single Kato-Katz for all STH, but this might be due to the number of pairwise-comparisons and the strict bounds of equivalence. Overall, there was an acceptable agreement in classifying the efficacy of ALB, suggesting that each of the investigated methods holds promise to assess drug efficacy in the context of STH control programs. The results also highlight that the clinical sensitivity or the ability to accurately estimate egg counts should not be the only parameters to determine the best diagnostic tool to assess drug efficacy. Instead, there are a number of other aspects that should also be considered to make a well-founded decision on what method(s) to recommend for monitoring drug efficacy in STH control programs, like user-friendliness and operational costs per test. We observed a decreasing trend in drug efficacy as a function of increasing historic drug pressure, yet further research is needed to identify factors that are contributing to this variation and to determine whether reduced efficacy can be linked with the known β-tubulin SNPs.
|
10.1371/journal.pntd.0002778 | Molecular Analysis of Echinostome Metacercariae from Their Second Intermediate Host Found in a Localised Geographic Region Reveals Genetic Heterogeneity and Possible Cryptic Speciation | Echinostome metacercariae are the infective stage for humans and animals. The identification of echinostomes has been based until recently on morphology but molecular techniques using sequences of ribosomal RNA and mitochondrial DNA have indicated major clades within the group. In this study we have used the ITS2 region of ribosomal RNA and the ND1 region of mitochondrial DNA to identify metacercariae from snails collected from eight well-separated sites from an area of 4000 km2 in Lamphun Province, Thailand. The derived sequences have been compared to those collected from elsewhere and have been deposited in the nucleotide databases. There were two aims of this study; firstly, to determine the species of echinostome present in an endemic area, and secondly, to assess the intra-specific genetic diversity, as this may be informative with regard to the potential for the development of anthelmintic resistance and with regard to the spread of infection by the definitive hosts. Our results indicate that the most prevalent species are most closely related to E. revolutum, E. trivolvis, E. robustum, E. malayanum and Euparyphium albuferensis. Some sites harbour several species and within a site there could be considerable intra-species genetic diversity. There is no significant geographical structuring within this area. Although the molecular techniques used in this study allowed the assignment of the samples to clades within defined species, however, within these groupings there were significant differences indicating that cryptic speciation may have occurred. The degree of genetic diversity present would suggest the use of targeted regimes designed to minimise the selection of anthelmintic resistance. The apparent lack of geographic structuring is consistent with the transmission of the parasites by the avian hosts.
| Infections by food-borne trematodes are estimated to infect over 40 million people worldwide, although infections by echinostomes make up only a portion of these cases, usually in regions where their prevalence is high. In South East Asia and in the far east of Asia, human infection is associated with cultural and dietary factors and the prevalence of infection may reach 50% in parts of Thailand, Cambodia, and Laos. Treatment is generally dependent on the use of praziquantel or benzimidazole drugs but with the occurrence of anthelmintic resistance to these compounds it would be desirable to have an understanding of the diversity present in the echinostome populations within a given locality. This study deals with the systematics of echinostomes and informs various aspects of the epidemiology of echinostomiasis which may aid the development of future control strategies.
| Echinostomes are intestinal trematodes of humans and animals that are endemic to Southeast Asia and the Far East, i.e. mainland China, Taiwan, India, Korea, Malaysia, Philippines, Indonesia, and Thailand, and present a public health problem [1]. Human echinostomiasis has been attributed to at least twenty species belonging to eight genera (Echinostoma, Echinochasmus, Acanthoparyphium, Artyfechinostomum, Episthmium, Himasthla, Hypoderaeumm, and Isthmiophora) of digenea trematodes that use snails as intermediate hosts [2], [3]. Clinical symptoms of echinostomiasis include severe epigastric or abdominal pain accompanied by diarrhea, fatigue, anorexia, and malnutrition in humans [2], [3]. Numerous cases of human echinostomiasis have been reported in Japan (E. cinetorchis, E. hortense, and E. japonicum), India (E. malayanum and Paryphostomum sufrartyfex), and Thailand (E. malayanum, E. revolutum, E. echinatum, and Hypoderaeum conoideum) and are associated with the eating of raw fresh-water fish, snails, and tadpoles [4], [5], [6], [7]. In Thailand, stool examination is used to detect echinostome eggs in Thai women. The most common parasite found in both pregnant and non-pregnant women is Opisthorchis viverrini, (hookworm) while Echinostoma spp., Strongyloides stercoralis, Taenia spp., Trichuris and Hymenolepis diminuta are more rarely found under these circumstances [8].
The identification of the species of echinostomes has been based in the past on morphology with major clades being defined on the basis of the number and distribution of the collar spines [9]. However, due to a large number of morphological similarities, this has become difficult in many cases. Molecular techniques have revealed differences among morphologically similar parasites [10], [11], [12]. An additional benefit of these techniques is that they can permit the identification of species, strains, and populations from a small quantity of tissue from any stage in their life-history [10], [13]. Generally, an investigation of the phylogenetic relationships between echinostomes uses sequence data from the mitochondrial cytochrome c oxidase subunit 1 (CO1) and nicotinamide adenine dinucleotide dehydrogenase subunit 1 (ND1) genes [10], [12], [13], [14], [15]. These have been determined to be valuable for a more accurate estimate of echinostome diversity [13], [16]. The internal transcribed spacer region (ITS) of ribosomal RNA (rRNA) has also provided a means of discriminating between species that have similar morphology [17], [18]. In this study, molecular sequencing of the ITS 2 region and ND1 gene of echinostomes were utilized.
The treatment of echinostomiasis is largely reliant on two anthelmintics: albendazole and praziquantel. Both of these drugs have been associated with the development of anthelmintic resistance (AR) [19]. It is important that their application follows a regime which will minimize the development of anthelmintic resistance. The rate of development of AR is a function of the genetic diversity of the target echinostome population [20], consequently we were interested in determining the variety and genetic diversity of echinostomes in the area from which our patient population was drawn.
Echinostomes were obtained from naturally infected fresh water snail intermediate hosts; Filopaludina martensi martensi. They were collected from permanent and seasonal ponds from eight field sites in Lamphun Province in northern Thailand (Table 1). The metacercariae were removed from the snails by crushing and the parasites were examined for the presence/absence of the collar spines. Those metacercariae found to have collar spines were taken from each snail and were frozen immediately for later DNA extraction.
DNA from all collected metacercariae was extracted as described in [21]. Briefly, 150 µl of 5% Chelex (Fluka) solution containing 10 µl of proteinase K (Sigma) at a concentration of 20 mg/ml was added to approximately 20 mg of trematode tissue. It was then heated at 55°C for 1 h, followed by gentle vortexing and heating at 95°C for 30 min, again followed by gentle vortexing. The mixture was centrifuged at 13,000 g for 10 sec. The supernatant was removed and stored at −20°C until it was to be used.
Approximately 1000 base pairs (bp) of the ITS2 region were amplified by using the primers, forward BD1 (5′-GCT GTA ACA AGG TTT CCG TA-3′) and reverse BD2 (5′-TAT GCT TAA ATT CAG CGG GT-3′). The PCR conditions used were the same as those previously described in [10] with amplification steps as follows: 2 min initial denaturation at 94°C, followed by 39 cycles of 1 min DNA denaturation at 94°C, 1 min primer annealing at 57°C, and 1 min at 72°C for extension and a final extension of 72°C for 10 min.
The amplification of ND1 and the PCR conditions used were those previously described in [10] with amplification steps as follows: 2 min initial denaturation at 94°C, followed by 39 cycles of 30 sec DNA denaturation at 94°C, 20 sec primer annealing at 48°C, and 1 min at 72°C for extension and final extension of 72°C for 10 min. Approximately 530 base pairs (bp) of the ND1 gene were amplified under these conditions by using the primers: forward JB11 (5′-AGA TTC GTA AGG GGC CTA ATA-3′) and reverse JB12 (5′-ACC ACT AAC TAA TTC ACT TTC-3′) as those described in [10].
Successful production of the amplicons and their quality was checked using agarose gel electrophoresis with ethidium bromide staining to visualize the ITS and ND1 products. All ITS and ND1 PCR products were purified using the Cleanup PCR Kit (Sigma) and were subjected to sequencing.
The raw sequencing data were assembled by Chromas Pro (Technelysium Pty. Ltd, Australia). Bio Edit software [22] was used to make sequence alignments which were compared to GenBank deposited sequences using BLASTN.
The sequence data produced in this study was combined with the data of 40 GenBank of echinostome sequences (ITS and ND1) and was aligned using BioEdit. Haplotype diversity and nucleotide diversity were both calculated by DNAsp [23]. Phylogenetic trees were generated for each gene using all sites with maximum likelihood. Branches were tested for all inferred trees using bootstrap analysis on 1,000 random trees. The relationship between the genetic diversity and the geographic distance within and among the species groups were calculated for each gene with MEGA version 5.0 [24]. The intra specific variation within each of the suggested clades and haplotype networks were constructed with statistical parsimony analysis for ND1 sequences (Network 4.6.1.1, fluxus-engineering.com, Fluxus Technology Ltd., UK, 2004). The 4× rule/K/θ ratio species criterion was applied to determine the likelihood of cryptic speciation [25].
High quality sequence data (ITS2 and ND1) was obtained for forty metacercariae. Figure 1 shows the analysis of the ITS2 data obtained in this study, along with relevant sequences from GenBank as a Maximum Likelihood bootstrap consensus tree with 1000 bootstrap iterations. There is strong support (>70%) for monophyletic clades for E. malayanum, E. revolutum, E. paraensei, E. trivolvis and Echinoparyphium spp. Three of the samples from Ban Thi were grouped with the E. malayanum clade and two from Mae Ta with the Echinoparyphium/Euparyphium clade. The remainder of the samples formed two distinct monophyletic clades. The larger of these consisted of a single haplotype and both showed 98% identity within a range of Echinostoma ITS2 sequences. A Neighbor-Joining tree gave identical topology (not shown).
Figure 2 shows a Maximum Likelihood tree based on the ND1 sequences and relevant GenBank sequences. As with the ITS2 sequences, there was good support for monophyletic clades for E. malayanum, E. revolutum, E. paraensei, E. trivolvis and Echinoparyphium spp. In this analysis, four of the samples from Ban Thi were associated with E. malayanum and nine of the samples formed a monophyletic group with the Echinoparyphium/Euparyphium clade. The remaining twenty-seven samples (labelled “Clade 3”) formed a monophyletic group containing four haplotypes. The statistics associated with these samples are shown in Table 2.
In order to determine whether the echinostome-like samples were within the limits of the genetic diversity found in the Echinoparyphium, E. trivolvis and E. revolutum clades (there are insufficient sequences of E. robustum in the database to allow it to be included in this analysis), we applied the K>4θ test. The statistics associated with this calculation are shown in Table 3. This analysis indicated that the samples Ban Thi 2–5 should be considered as E. malayanum, but that the rest of the isolates, although sharing ancestry with either the Echinoparyphium or the Echinostomatrivolvis/revolutum/robustum clades, could be regarded as separate species by this criterion.
There was considerable variability in the diversity of the species found at the different sites. Most sites had more than one species present and this parameter did not seem to be correlated with the permanence of the site. Figure 3 shows a schematic Median-Joining network constructed from the ND1 sequences. This analysis, using an alternative algorithm, confirmed the division of the samples into three clades. The genetic distances between the clades and their geographic spread are shown. The most frequent isolates were from the unidentified “echinostome-like” clade 3 grouping, which was found at six of the seven sites investigated.
The results presented in this paper provide the most extensive use to date of molecular techniques for the characterization of echinostomes from Thailand. Our results indicate that even small seasonal ponds may contain infected snails carrying a range of species. The samples in our study could be grouped into three distinct clades, E. malayanum, an Euparyphium/Echinoparyphium-like clade and an Echinostoma trivolvis/revolutum-like clade; worms identified as belonging to all of these groups have been shown to endemically infect humans in South East Asia [26]. Our findings are in agreement with those of [27], [28], who reported that E. malayanum and E. revolutum as being prevalent in Thailand. Although they recorded fixed genetic differences at 19% of the loci examined between Thai E. revolutum and those from the Lao PDR, they did not consider this as evidence of cryptic speciation as there was little divergence in the 200 bp of the mitochondrial cytochrome oxidase 1 (CO1) sequences for the worms from these two regions. In contrast, our analysis of the mitochondrial diversity of the “echinostome-like” clade 3 was based on approximately 800 bp of the ND1 region of the mitochondrial genome – this region is known to be more susceptible to changes [10], and thus may be more informative than the CO1 region. The analysis presented in Table 3 indicates that the “Echinostoma-like” clade 3 worms found in Thailand are genetically distant from E. trivolvis from North America and E. revolutum from northern Europe, and may be considered to constitute a cryptic species by the K>4θ criterion proposed by Birky [25]. Likewise the Euparyphium/Echinoparyphium-like Thai clade would appear to be a separate species from the North American Echinoparyphium spp., with which it was compared. As we were able to group some of our isolates with worms that were previously identified as E. malayanum from Thailand, this may indicate that on the continental scale there is geographical structuring of the Echinostomatidae family. It has been shown for other trematodes that the involvement of a highly motile host in the parasite's life cycle will reduce local geographic structuring [29]. All of the Echinostomatidae in this study are known to be capable of using avian species, such as ducks, as their definitive host, and Thailand is situated on the East Asian-Australasian Flyway, which has been implicated previously in the spread of zoonotic diseases [30]. Support for this suggestion may be given by the analysis of an E. revolutum isolate using the ITS 1 sequences [26], which indicated that it was more closely related to an Australian isolate than to those from North America.
In conclusion, we have shown that people living in a relatively small and homogeneous geographic area of South East Asia may be exposed to infection by at least three species of Echinostomatidae. There is sufficient genetic diversity present among these populations to allow for the selection of praziquantel resistance, as has occurred in the case of schistosomiasis [19], [31] and this finding emphasizes the need for targeted administration of chemotherapies.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.