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.1007784 | Maternal and zygotic gene regulatory effects of endogenous RNAi pathways | Endogenous small RNAs (sRNAs) and Argonaute proteins are ubiquitous regulators of gene expression in germline and somatic tissues. sRNA-Argonaute complexes are often expressed in gametes and are consequently inherited by the next generation upon fertilization. In Caenorhabditis elegans, 26G-RNAs are primary endogenous sRNAs that trigger the expression of downstream secondary sRNAs. Two subpopulations of 26G-RNAs exist, each of which displaying strongly compartmentalized expression: one is expressed in the spermatogenic gonad and associates with the Argonautes ALG-3/4; plus another expressed in oocytes and in embryos, which associates with the Argonaute ERGO-1. The determinants and dynamics of gene silencing elicited by 26G-RNAs are largely unknown. Here, we provide diverse new insights into these endogenous sRNA pathways of C. elegans. Using genetics and deep sequencing, we dissect a maternal effect of the ERGO-1 branch of the 26G-RNA pathway. We find that maternal primary sRNAs can trigger the production of zygotic secondary sRNAs that are able to silence targets, even in the absence of zygotic primary triggers. Thus, the interaction of maternal and zygotic sRNA populations, assures target gene silencing throughout animal development. Furthermore, we explore other facets of 26G-RNA biology related to the ALG-3/4 branch. We find that sRNA abundance, sRNA pattern of origin and the 3’ UTR length of target transcripts are predictors of the regulatory outcome by the Argonautes ALG-3/4. Lastly, we provide evidence suggesting that ALG-3 and ALG-4 regulate their own mRNAs in a negative feedback loop. Altogether, we provide several new regulatory insights on the dynamics, target regulation and self-regulation of the endogenous RNAi pathways of C. elegans.
| Small RNAs (sRNAs) and their partner Argonaute proteins regulate the expression of target RNAs. When sperm and egg meet upon fertilization, a diverse set of proteins and RNA, including sRNA-Argonaute complexes, is passed on to the developing progeny. Thus, these two players are important to initiate specific gene expression programs in the next generation. The nematode Caenorhabditis elegans expresses several classes of sRNAs. 26G-RNAs are a particular class of sRNAs that are divided into two subpopulations: one expressed in the spermatogenic gonad and another expressed in oocytes and in embryos. In this work, we describe the dynamics whereby oogenic 26G-RNAs setup gene silencing in the next generation. In addition, we show several ways that spermatogenic 26G-RNAs and their partner Argonautes, ALG-3 and ALG-4, use to regulate their targets. Finally, we show that ALG-3 and ALG-4 are fine-tuning their own expression, a rare role of Argonaute proteins. Overall, we provide new insights into how sRNAs and Argonautes are regulating gene expression.
| A plethora of pathways based on non-coding small RNAs (sRNAs) regulates gene expression in every domain of life. These are collectively known as RNA interference (RNAi) or RNAi-like pathways. In invertebrates, which lack adaptive immune systems and interferon response, RNAi-like pathways fulfill an immune role at the nucleic acid level, by controlling viruses and transposable elements (TEs).
MicroRNA (miRNA), Piwi-interacting RNA (piRNA) and endogenous small interfering RNA (endo-siRNA) pathways are the better described RNAi-like pathways, which differ in their biogenesis and specialized cofactors. MicroRNAs are commonly found in many, if not all, tissues and broadly regulate gene expression throughout development [1]. piRNAs are typically, but not exclusively, expressed in the metazoan germline, where they assume a central function in TE control [2–5]. Endo-siRNA pathways comprise varied classes of sRNAs expressed in the soma and germline that can regulate the expression of TEs and protein-coding genes [6–8]. A key commonality of RNAi-like pathways is the participation of Argonaute proteins. These proteins directly associate with sRNAs and Argonaute-sRNA complexes engage transcripts with sequence complementarity, typically resulting in target silencing. sRNA-directed gene silencing can occur both on the post-transcriptional level, by target RNA cleavage and degradation, and/or on the transcriptional level, via nuclear Argonautes that direct heterochromatin formation at target loci.
sRNAs can be viewed as genome guardians against “foreign” nucleic acids [9]. In this light, the germline is an important tissue for sRNA production and function to control the transmission of “non-self” genetic elements to progeny. In multiple animals, Piwi-piRNA complexes have been shown to be maternally deposited into zygotes, where they may initiate TE silencing [10–19]. Endo-siRNAs are abundantly expressed in gametes, being often required to successfully complete gametogenesis. These may also be deposited into embryos and have roles in setting up gene expression in the next generation. For example in plants, TE-derived endo-siRNAs are abundant in male and female gametes [20]. Moreover, endo-siRNAs are expressed in Drosophila ovaries [21] and in mouse oocytes [22,23] to regulate protein-coding genes and TEs. Overall, gamete expression and maternal inheritance of Argonaute-sRNA complexes seem to be a widespread phenomenon in plants and animals, presumably important to tune gene expression during early development.
RNAi was first identified in the nematode Caenorhabditis elegans [24]. Ever since, C. elegans has continuously been an important and fascinating model for studies on RNAi. C. elegans has an unprecedented 27 genomically encoded Argonaute genes, including a whole worm-specific clade of the Argonaute protein family [25]. Several sRNA species have been identified in worms: miRNAs, 21U-RNAs, 22G- and 26G-RNAs [26,27]. 21U-RNAs associate with PRG-1, a Piwi class Argonaute, in the germline and are therefore considered the piRNAs of C. elegans [28–30]. 26G-RNAs can be considered primary endo-siRNAs, in that they elicit production of the overall more abundant secondary endo-siRNA pool, termed 22G-RNAs [31–33].
26G-RNAs are produced by the RNA-dependent RNA Polymerase (RdRP) RRF-3 [31–35]. The ERI complex (ERIC) is an accessory complex that assists RRF-3 in producing 26G-RNAs [36–39]. The conserved CHHC zinc finger protein GTSF-1 and the Tudor domain protein ERI-5 form a pre-complex with RRF-3 that is responsible for tethering the RdRP to the ERIC [36,39]. Two distinct subpopulations of 26G-RNAs are synthesized in the germline and in embryos. One subpopulation is produced in the spermatogenic gonad in L4 hermaphrodites and in the male gonad, where they associate with the redundantly acting paralog Argonautes ALG-3 and ALG-4 (henceforth referred to as ALG-3/4)[31,34,35,38]. These 26G-RNAs trigger the biogenesis of secondary 22G-RNAs that have been shown to either promote gene expression through the Argonaute CSR-1 or to inhibit gene expression through unidentified WAGO proteins [31,40]. Hence, the effects of ALG-3/4-dependent sRNAs on their targets is complex: while some targets appear to be silenced, the expression of others seems to be positively affected. The regulatory effects resulting of the combined action of ALG-3/4 and CSR-1 seem to be more physiologically relevant at elevated temperatures [40]. The conditions determining regulatory outcome, either silencing or licensing, are still unclear.
In the oogenic hermaphrodite gonad and in embryos another subpopulation of 26G-RNAs is produced. These are 3’ 2’-O-methylated by the conserved RNA methyltransferase HENN-1 [41–43] and bind to the Argonaute ERGO-1 [33]. ERGO-1 targets pseudogenes, recently duplicated genes and long non-coding RNAs (lncRNAs)[33,36,44]. It has recently been shown that these targets generally have a small number of introns that lack optimal splicing signals [45]. ERGO-1 may thus serve as a surveillance platform to silence these inefficient transcripts, preventing detrimental accumulation of stalled spliceosomes. Effective silencing of these genes is achieved by secondary 22G-RNAs produced after ERGO-1 target recognition [32,33]. In turn, these secondary 22G-RNAs may associate with cytoplasmic Argonautes that mediate post-transcriptional gene silencing [33], or to the Argonaute NRDE-3, which is shuttled into the nucleus and further silences its targets on the transcriptional level [46,47].
Depletion of spermatogenic 26G-RNAs, for example in rrf-3, gtsf-1 and alg-3/4 mutants, results in a range of sperm-derived fertility defects including complete sterility at higher temperatures [31,34–38]. The elimination of oogenic/embryonic 26G-RNAs, for example by impairment of rrf-3, gtsf-1 and ergo-1, gives rise to an Enhanced RNAi (Eri) phenotype, characterized by a response to exogenous dsRNA that is stronger than in wild-type [25,36–38]. This phenotype is thought to reflect competition for common factors between exogenous and endogenous RNAi pathways [37,48]. However, the Eri phenotype lacks characterization on the molecular level. Furthermore, a strong maternal rescue was reported for Eri factors [49], suggesting that maternally deposited Eri factors or their dependent sRNAs have an important role in maintaining gene silencing. The basis for this maternal rescue was not further characterized.
In this work, we address a number of gene regulatory aspects of the 26G-RNA pathways in C. elegans. First, we genetically dissect a maternal effect displayed by the ERGO-1 branch of the 26G-RNA pathway. Our findings suggest that both maternal and zygotic sRNAs drive gene silencing throughout embryogenesis and larval development until adulthood. Moreover, we interrogate a number of aspects on gene regulation in the ALG-3/4 branch of the 26G-RNA pathway. We report that sRNA abundance, origin of the sRNAs and 3’ UTR length of target transcripts are predictors of the regulatory outcome of ALG-3/4 targets. Lastly, we find that the 26G-RNA-binding Argonautes ALG-3 and ALG-4 may regulate their own expression in a negative feedback mechanism.
rrf-3 and gtsf-1 mutants lack the two subpopulations of 26G-RNAs and display the phenotypes associated with depletion of both subpopulations: the enhanced RNAi (Eri) phenotype, shared with ergo-1 mutants [25,36–38], and sperm-derived fertility defects, shared with alg-3/4 double mutants [31,34–38,40]. S1A Fig offers a simplified scheme of these pathways. For clarity, the two subpopulations of 26G-RNAs and downstream 22G-RNAs, dependent on ERGO-1 or ALG-3/4 will be referred to as ERGO-1 branch sRNAs and ALG-3/4 branch sRNAs, respectively.
We have previously shown that germline-specific GTSF-1 transgenes could rescue the enhanced RNAi (Eri) phenotype of gtsf-1 mutants [36]. This was an intriguing result, since the Eri phenotype arises after targeting somatically expressed genes with RNAi, indicating that germline-expressed GTSF-1 is able to affect RNAi in the soma, possibly through maternal deposition of GTSF-1 or GTSF-1-dependent sRNAs. We reasoned that if maternal GTSF-1 activity can prime gene silencing in embryos then the transmission of the Eri phenotype should show a maternal rescue. To address this experimentally, we linked gtsf-1(xf43) to dpy-4(e1166) and crossed the resulting double mutants with wild-type males (Fig 1A). We then allowed for two generations of heterozygosity and assayed for RNAi sensitivity in homozygous gtsf-1 mutant F1 and F2 generations, scoring for larval arrest triggered by lir-1 RNAi. Indeed, the Eri phenotype showed a strong maternal effect, arising only in the F2 generation of gtsf-1 mutants (Fig 1A). This is consistent with a maternal effect reported for other Eri factors [49]. We have previously shown that GTSF-1 is required to silence a GFP transgene reporting on ERGO-1 branch 22G-RNA activity, referred to as 22G sensor [36,43]. Therefore, we also looked at the dynamics of derepression of this transgene upon introduction of gtsf-1 mutation. We noticed that strong GFP expression appeared only in the second generation of homozygosity of the gtsf-1 allele (Fig 1B and 1C). An identical maternal effect on the expression status of this transgene is observed after crossing in rrf-3, ergo-1 and other gtsf-1 mutant alleles (S1B Fig). Combined with our previously described rescue of the Eri phenotype using a germline promoter, these results strongly suggest that maternally provided ERGO-1 branch pathway components are sufficient to establish normal RNAi sensitivity in the soma of C. elegans.
Although the silencing of the 22G sensor used in our experiments is dependent on ERGO-1, ERGO-1 is not the Argonaute protein binding to the effector 22G-RNA [33,43]. This has been shown to be driven by the somatically expressed, nuclear Argonaute protein NRDE-3 [44,46] and maybe additional cytoplasmic WAGOs [33](S1A Fig). In absence of ERGO-1 and other 26G-RNA pathway factors, NRDE-3 is no longer nuclear, and in nrde-3 mutants the 22G sensor is activated, indicating that NRDE-3 requires sRNA input from ERGO-1 branch sRNAs [36,43,44,46]. Strikingly, loss of NRDE-3 derepressed the 22G sensor transgene in the first homozygous generation (Fig 1D), showing that in contrast to 26G-RNAs, the downstream 22G-RNA pathway is not maternally provided. MUT-16 is a factor required for the nucleation of mutator foci and 22G-RNA biogenesis [50]. Confirming the requirement for zygotically produced 22G-RNAs, absence of MUT-16 derepresses the 22G sensor in the first homozygous mutant generation (Fig 1E). These results suggest a scenario in which 1) NRDE-3 is loaded with zygotically produced 22G-RNAs that are primed by maternally provided 26G-RNAs and 2) NRDE-3 activity is maintained in somatic tissues until the adult stage, in absence of a zygotic 26G-RNA pathway.
The results presented above show that maternal 26G-RNAs are sufficient for 22G sensor silencing. We also tested whether maternal 26G-RNAs are necessary for 22G sensor silencing by crossing rrf-3 mutant males with gtsf-1; 22G sensor hermaphrodites (Fig 1F). Both of these strains lack 26G-RNAs and their downstream 22G-RNAs, therefore, their progeny will not receive a maternal and/or paternal complement of these sRNAs. The 22G sensor was silenced in all cross progeny, showing that in the absence of maternal 26G-RNAs, zygotic 26G-RNAs can induce production of silencing-competent 22G-RNAs. Thus, maternal 26G-RNAs appear to be sufficient but not necessary for target silencing.
The maternal effects described above for the Eri phenotype and for 22G sensor silencing are related to the ERGO-1 branch of the pathway. Next, we wanted to determine if the ALG-3/4 branch also displays such a parental effect. To test this, we assessed the influence of maternal GTSF-1 activity on the temperature-sensitive sterility phenotype. Using the same setup as we used for the Eri experiment (in Fig 1A), we observe that the temperature-sensitive sperm defect of gtsf-1 mutants was not rescued maternally (S1C Fig). Given that the ALG-3/4 branch of the 26G-RNA pathway is mostly active during spermatogenesis, next we asked whether a paternal effect is observed for the temperature-sensitive sperm defect. As shown in S1D Fig, we did not detect any evidence supporting a paternal effect. Overall, these results indicate that 26G-RNA-derived parental effects are likely restricted to the ERGO-1 branch.
The 22G sensor reports on the silencing activity of a single 22G-RNA that maps to the so-called X-cluster, a known set of targets of ERGO-1 [33,43]. Therefore, the experiments above using this 22G sensor have a limited resolution and our observations may not reflect the silencing status of most ERGO-1 targets. To characterize this maternal effect in more detail and in a broader set of ERGO-1 targets, we decided to analyze sRNA populations in young adult animals. Concretely, we outcrossed dpy-4; gtsf-1 and sequenced sRNAs from wild-type and two consecutive generations of Dpy young adult animals (Fig 2A). First generation gtsf-1 homozygous mutants will henceforth be addressed as “mutant F1” and second generation gtsf-1 homozygous mutants as “mutant F2” (Fig 2A). We sequenced young adult animals because they lack embryos, therefore avoiding confounding effects with zygotic sRNAs of the next generation. sRNAs were cloned and sequenced from four biological replicates. The cloning of sRNAs was done either directly (henceforth referred to as untreated samples) or after treatment with the pyrophosphatase RppH [51] before library preparation. The latter enriches for 22G-RNA species that bear a 5’ triphosphate group. Sequenced sRNAs were normalized to all mapped reads excluding structural reads (S1 Table). In our analysis we strictly looked at 26G- and 22G-RNAs that map in antisense orientation to protein-coding and non-coding genes (see Methods).
Total 26G-RNA levels are depleted in young adults lacking GTSF-1 (Fig 2B). Mutant F1s have significantly less 26G-RNAs than wild-type worms, while mutant F2s have 26G-RNA levels very close to zero (Fig 2B). For a finer analysis we looked specifically at 26G-RNAs derived from ERGO-1 and ALG-3/4 targets (as defined in reference 36, see Methods). 26G-RNAs mapping to these two sets of targets recapitulate the pattern observed for global 26G-RNAs (Fig 2C and 2D). The difference between the F1 and F2 mutants might reflect a maternal 26G-RNA pool that is still detectable in the young adult F1, but no longer in the F2. However, we point out that amongst the selected F1 Dpy animals, approximately 5.2% will in fact be gtsf-1 heterozygous, due to meiotic recombination between gtsf-1 and dpy-4 (estimated genetic distance between these two genes is 2.6 map units). Hence, another explanation for the mutant F1 pool of 26G-RNAs may be a contamination of the gtsf-1 homozygous pool with heterozygous animals. The mutant F2 was isolated from genotyped F1 animals, excluding this confounding effect. We conclude that in young adult mutant F1 animals, maternally provided 26G-RNAs (or 26G-RNAs produced zygotically by maternal proteins) are no longer detectable at significant levels.
Total levels of 22G-RNAs are slightly reduced in mutant F1 and F2 animals (Fig 2E). However, total 22G-RNA levels encompass several distinct subpopulations of 22G-RNAs, including those that do not depend on 26G-RNAs. To have a closer look on 22G-RNAs that are dependent on 26G-RNAs, we focused on 22G-RNAs that map to ERGO-1 and ALG-3/4 targets. Strikingly, compared to wild-type, the 22G-RNA population from ERGO-1 targets is moderately higher in mutant F1 animals and is subsequently depleted in the mutant F2 generation (Figs 2F and S2A). These effects are not only clear in overall analysis, but also on a well-established set of ERGO-1 branch targets, such as the X-cluster (Fig 2G). Consistent with a role of NRDE-3 downstream of ERGO-1, 22G-RNAs mapping to annotated NRDE-3 targets [47] show the same pattern of depletion as ERGO-1-dependent 22G-RNAs (S2A Fig). These results are consistent with the idea that the Eri phenotype and 22G sensor derepression are caused by the absence of NRDE-3-bound, secondary 22G-RNAs downstream of 26G-RNAs.
22G-RNAs mapping to ALG-3/4 targets behave differently in this experiment (Fig 2H). Upon disruption of gtsf-1, these 22G-RNAs are only slightly affected in both the mutant F1 and F2 (Figs 2H and S2A), despite the fact that their upstream 26G-RNAs are absent. This is illustrated in S2B Fig with genome browser tracks of ssp-16, a known ALG-3/4 target. We conclude that 26G-RNA-independent mechanisms are in place to drive 22G-RNA production from these genes.
Finally, 21U-RNAs and 22G-RNAs mapping to other known RNAi targets are not affected in this inheritance setup, supporting the notion that gtsf-1 is not affecting these sRNA species (S2A and S2C Fig). One exception are the 22G-RNAs from CSR-1 targets, which seem to be slightly depleted in both the mutant F1 and F2 generations (S2A Fig). It is not possible to dissect whether this is a direct effect or not, but we note that mRNA levels of CSR-1 targets are slightly downregulated in the analyzed mutants (S2D Fig). Given that CSR-1 22G-RNAs tend to correlate positively with gene expression [52], it is conceivable that the reduction of CSR-1 target 22G-RNAs is the result of decreased target gene expression.
The very same samples used for generating sRNA sequencing data were also used for mRNA sequencing (Fig 2A and S1 Table). First, we checked gtsf-1 expression. As expected, gtsf-1 is strongly depleted in the mutant samples (S3A Fig). In the mutant F1 we still observe a low level of gtsf-1 derived transcripts (about 9.5% of wild-type) that is absent from the mutant F2. These transcripts cover the region deleted in the gtsf-1(xf43) mutant allele, indicating they cannot represent zygotically transcribed gtsf-1 mutant mRNA. Rather, these transcripts likely originate from the above-described contamination of the homozygous F1 population with heterozygous animals.
We hypothesized that ERGO-1 branch 22G-RNAs observed in the mutant F1 generation might be competent to maintain target silencing. If this is true, we should observe strong upregulation of ERGO-1 target mRNAs only the mutant F2 generation. Indeed, the X-cluster is upregulated only in the second mutant generation (Fig 3A). When ERGO-1 targets are analyzed in bulk, we observe the same trend, with stronger upregulation only in the mutant F2, consistent with the maternal effect (Fig 3B). Regarding the mutant F1, we note that the very slight, not statistically significant, upregulation of ERGO-1 target mRNAs may account for the slight increase of 22G-RNAs observed in the mutant F1 (in Figs 2F and S2A), because of an increased number of molecules available to template RdRP activity. ALG-3/4 targets, as for instance ssp-16, were found to be upregulated already in the F1 generation (Figs 3B and S3B), supporting the notion that the maternal rescue of the 26G-RNA pathways is restricted to the ERGO-1 branch.
ERGO-1 targets comprise a very diverse set of targets consisting of pseudogenes, fast evolving small genes, paralog genes and lncRNAs [33,44,45]. Considering the maternal effect described above for ERGO-1-dependent sRNA and corresponding target, we postulated that this maternal effect may exist to counteract embryonic expression of ERGO-1 targets. To address this we sequenced mRNA of synchronized populations of all developmental stages (L1, L2, L3, L4, young adult and embryos) of both wild-type (N2) and rrf-3(pk1426) mutants (S1 Table). In wild-type worms, ERGO-1 targets are most abundant in embryos (Fig 3C, lower panel, in blue). Moreover, the effect of rrf-3 mutation on ERGO-1 target expression is stronger in embryos (Fig 3C, lower panel). These results indicate that the maternal effect reported above can reflect deposition of factors which are required to initiate silencing of targets early in development. Differential gene expression data and normalized read counts calculated from the sequencing datasets described in Figs 2 and 3, can be found in S2 and S3 Tables.
The young adult sequencing datasets we obtained in this study (Fig 2A), as well as previous datasets of gravid adults [36], are not well suited to address ALG-3/4 biology, considering that in these developmental stages ALG-3/4 are not expressed, at least not abundantly. Therefore, in order to further our understanding of the dependency of ALG-3/4 branch sRNAs on GTSF-1, as well as to explore the regulation of ALG-3/4 targets, we generated additional sRNA and mRNA datasets from wild-type and gtsf-1 male animals grown at 20°C (S1 Table).
As expected, global 26G-RNA levels are severely affected in gtsf-1 mutant males, reflecting downregulation of 26G-RNAs from both branches of the pathway (Fig 4A–4C). Consistent with the absence of ERGO-1 in adult males, ERGO-1 branch 26G-RNAs are detected in extremely low numbers in wild-type animals (Fig 4B). Global levels of 21U-RNAs seem to be moderately increased (Fig 4D), possibly resulting from the lack of 26G-RNAs in the libraries. Global levels of 22G-RNAs are not affected (Fig 4E), but consistent with a global depletion of 26G-RNAs, 22G-RNAs specifically mapping to ALG-3/4 and ERGO-1 targets are reduced in gtsf-1 mutant males (Fig 4F). Next, we probed the effects of gtsf-1 mutation on male gene expression using mRNA sequencing. ALG-3/4 and ERGO-1 targets are both upregulated in gtsf-1 mutant males (Fig 4G). These changes are illustrated for the X-cluster and ssp-16 in the genome browser tracks of S4 Fig. S4 Table includes differential gene expression data and normalized read counts calculated from the sequencing datasets described in Fig 4.
As a final note on the developmental aspects of ALG-3/4 branch, consistent with enrichment in the spermatogenic gonad [31,34–36,38,40], ALG-3/4 targets are more highly expressed and more responsive to rrf-3 mutation in the L4 and young adult stages of hermaphrodite animals (Fig 3C, upper panel). Given that the overall ALG-3/4 target mRNA levels go up upon depletion of gtsf-1 or rrf-3 (Figs 3C and 4G), bulk 26G-RNA activity during spermatogenesis seems to be repressive at 20°C.
We conclude that the activity of GTSF-1 is required in adult males for silencing of 26G-RNA targets by participating in 26G- and 22G-RNA biogenesis.
ALG-3/4 were shown to have distinct effects on gene expression, either silencing or licensing [31,40]. However, how these different effects arise is currently unknown. Even though our analysis in males did not reveal a licensing effect of 26G-RNAs, the bulk analysis of targets in Figs 3B and 4G may mask the behavior of distinct target subpopulations. Of note, our sequencing datasets were obtained from animals grown at 20°C and are therefore blind to the strong positive regulatory effect of ALG-3/4 in gene expression at higher temperatures [40].
We reasoned that sRNA abundance may be correlated with different regulatory outcomes. Therefore, we defined ALG-3/4 targets that are upregulated, downregulated, and unaltered upon gtsf-1 mutation and plotted their 26G-RNA abundance. This reveals a tendency for genes that are upregulated upon loss of GTSF-1 to be more heavily targeted by 26G-RNAs in adult males (Fig 5A, left panel). The same trend is observed for 22G-RNAs: upregulated genes are more heavily covered by 22G-RNAs (Fig 5A, right panel, and 5B). In contrast, ALG-3/4 targets that are downregulated in gtsf-1 mutant males display a relatively low-level targeting by 22G-RNAs (Fig 5A and 5B).
We conclude that, in adult males, stronger 26G-RNA targeting promotes stronger 22G-RNA biogenesis and repression of targets, whereas low-level targeting by 26G- and 22G-RNAs does not. Transcripts that are downregulated in absence of GTSF-1 might be licensed for gene expression, but may also respond in a secondary manner to a disturbed 26G-RNA pathway.
It was previously noticed that ALG-3/4-dependent 26G-RNAs mostly map to both the 5’ and 3’ ends of their targets, and that this may correlate with gene expression changes [31]. We followed up on this observation by performing metagene analysis of 26G-RNA binding using our broader set of targets. Indeed, ALG-3/4 branch 26G-RNAs display a distinctive pattern with two sharp peaks near the transcription start site (TSS) and transcription end site (TES) (Figs 6A and S5A, left panels). In contrast, ERGO-1 branch 26G-RNAs map throughout the transcript, with a slight enrichment in the 3’ half (Fig 6B, left panel). Contrary to 26G-RNAs, 22G-RNAs from both branches map throughout the transcript (Figs 6A and 6B and S5A, right panels). These patterns are consistent with recruitment of RdRPs and production of antisense sRNAs along the full length of the transcript. These findings suggest substantially different regulation modes by ERGO-1- and ALG-3/4-branch 26G-RNAs.
Conine and colleagues reported a correlation between 26G-RNA 5’ targeting and negative regulation [31]. We wanted to address whether our datasets show concrete correlations between the patterns of origin of ALG-3/4-dependent 26G-RNAs and distinct regulatory outcomes. To address this, we ranked genes by 5’ and 3’ abundance of 26G-RNAs, selected genes predominantly targeted at the 5’ or at the 3' ends and plotted their fold change upon gtsf-1 mutation. In adult males, dominant 5’ targeting by 26G-RNAs seems to be correlated with gene silencing (fold change >0 in the mutant, Fig 6C), whereas dominant 3’ targeting is accompanied with only weak upregulation and in some cases very mild downregulation (Fig 6C and 6D). In further support for a non-gene silencing, and potentially licensing role for ALG-3/4 targeting at the 3’ end, genes with predominant 3’ 26G-RNAs display an overall higher expression than genes predominantly targeted at the 5’ region (Fig 6E). The same signatures are found in young adults, with an even stronger signature of the 3’ in promoting gene expression (S5B, S5C and S5D Fig).
Finally, we interrogated if the length of 5’ and 3’ UTRs may be a predictor of regulatory outcome in ALG-3/4 targets. 5’ UTR length was not significantly different between unchanged and upregulated genes (Fig 6F, left panel). Downregulated genes have a statistically significant shorter 5’ UTR length, but these results should be interpreted with caution due to the low number of transcript isoforms analyzed. In contrast, 3’ UTR length is significantly smaller in targets that respond to loss of GTSF-1 in males (Fig 6F, right panel). Interestingly, we find the same and possibly even stronger relation between 3’UTR length and responsiveness to GTSF-1 status in young adult animals (S5E Fig). Upregulated ERGO-1 targets do not display significantly shorter 3’ UTRs (S5F Fig), indicating that the trend is specific to ALG-3/4 targets.
Altogether, our results suggest that, both in males and young adult hermaphrodites, 3’ vs 5’ targeting and 3’ UTR length are predictors of whether ALG-3/4 targets are silenced or not.
While navigating the lists of GTSF-1 targets defined by differential gene expression analysis, we noticed that alg-3 and alg-4 are targets of 26G-RNAs (this study and in reference 36). These 26G-RNAs are sensitive to oxidation (not enriched in oxidized libraries, see reference 36) and map predominantly to the extremities of the transcript (Figs 7, upper panels), indicating that these 26G-RNAs share features with ALG-3/4 branch 26G-RNAs. In addition to these 26G-RNAs, significant amounts of 22G-RNAs are found on alg-3/4 (Fig 7, middle panels). These sRNAs seem to silence gene expression, since mRNA-seq shows that alg-3 and alg-4 transcripts are 2–3 fold upregulated in gtsf-1 mutants (Fig 7, lower panels).
These results strongly suggest that alg-3/4 are regulating their own expression in a negative feedback loop. Of note, the upregulation of alg-3 and alg-4 is in agreement with the results presented above, because these genes are more heavily targeted by 26G-RNAs at their 5’ (although alg-4 also has a sharp 3’ 26G-RNA peak, upper panels). Furthermore, these same signatures of negative feedback loop are observed in young adults (S6 Fig).
Animal male and female gametes are rich in RNA. Upon fertilization, several RNA species are thus provided to the zygote. Multiple lines of evidence from several distinct organisms indicate that sRNAs are included in the parental repertoire of inherited RNA. For example, piRNAs have been reported to be maternally deposited in embryos in arthropods, fish and C. elegans [10–19,53,54]. In C. elegans other endogenous sRNA populations have also been shown to be contributed by the gametes: 1) 26G-RNAs have been shown to be weakly provided by the male, while 22G-RNAs are more abundantly provided [55]; 2) 26G-RNAs and the Argonaute ERGO-1 are co-expressed during oogenesis and in embryos [33,35,56]; and 3) 22G-RNAs are deposited in embryos via the mother and participate in transgenerational gene silencing [53,57–61].
We describe a maternal effect in the transmission of the Eri phenotype and 22G sensor derepression and characterize the subjacent dynamics of sRNAs and mRNA targets (Figs 1–3 and S1–S3). We show that both maternal and zygotic 26G-RNAs are sufficient for silencing. Absence of either the maternal or the zygotic pools can thus be compensated, enhancing the robustness of this system. We note, however, that sufficiency has only been tested with the described 22G sensor. It may be that the silencing of other targets has differential dependencies on maternal and zygotic 26G-RNA populations. The maternal effect was observed in mutants of a variety of Eri genes, including gtsf-1, rrf-3 and ergo-1, but not alg-3/4. Therefore, these defects are related to impairment of sRNA populations directly associated with and downstream of ERGO-1. These results do not exclude a parental effect for ALG-3/4. In fact, a paternal effect on embryogenesis has been described for rrf-3 mutants [34]. Such phenotype most likely arises due to ALG-3/4 branch sRNAs.
Maternal rescue of Eri genes was previously reported [49], although the genetic basis for this phenomenon was not characterized further. We demonstrate that in the first Eri mutant generation, primary 26G-RNAs are downregulated, while their downstream 22G-RNAs are still present (Fig 2). These ERGO-1-dependent 22G-RNAs, maintained in the absence of their primary triggers, seem to be competent to sustain silencing of ERGO-1 targets throughout life of the animal (Fig 3). Given that 1) ERGO-1 targets display higher expression during embryogenesis; and 2) upon disruption of endogenous RNAi by rrf-3 mutation, targets become upregulated in all developmental stages (Fig 3C); maternally deposited ERGO-1-dependent factors may be especially required to initiate target silencing during embryogenesis, and to prevent spurious expression throughout development. The ERGO-1-independent maintenance of this silencing response may be mechanistically similar to RNA-induced epigenetic silencing (RNAe), involving a self-perpetuating population of 22G-RNAs [53,62,63]. Indeed, both processes depend on a nuclear Argonaute protein: HRDE-1 in RNAe [53,62,63] and NRDE-3 for ERGO-1-driven silencing [43,46,47]. Self-perpetuating 22G-RNA signals may be also in place in the male germline (see below).
Our genetic experiments and sequencing data are fully consistent with maternal inheritance of 26G-RNAs. However, these may not be the only inherited agent. A non-mutually exclusive idea is that GTSF-1, as well as other ERIC proteins may be deposited in embryos to initiate production of zygotic sRNAs. In accordance with the latter, we have previously demonstrated that formation of the 26G-RNA generating ERIC is developmentally regulated [36]. While in young adults there is a comparable amount of pre- and mature ERIC, in embryos there is proportionally more mature ERIC. These observations suggest that pre-ERIC might be deposited in the embryo to swiftly jumpstart zygotic 26G-RNA expression after fertilization.
We show that GTSF-1 is required in adult males, potentially in the male germline, to produce 26G- and downstream 22G-RNAs (Fig 4) analogous to its role in the hermaphrodite germline and in embryos [36]. In addition, the bulk of targets from both 26G-RNA pathway branches seem to be deregulated. Interestingly, we note that although ERGO-1 and its cognate 26G-RNAs are not abundantly expressed in spermatogenic tissues (Fig 4B), gtsf-1-dependent, secondary 22G-RNAs mapping to these genes maintain gene silencing in the adult males (Fig 4F and 4G). In an analogous manner, we find that ALG-3/4 targets maintain 22G-RNAs in gravid adults [36], even though ALG-3/4 are not expressed at that stage. Mechanistically this may be closely related to how maternal 26G-RNAs can trigger 22G-RNA-driven silencing (see above). NRDE-3 is downstream of ERGO-1 and is likely to silence ERGO-1 targets throughout development. However, the Argonautes associated with 22G-RNAs mapping to 1) ERGO-1 targets in the male, and 2) to ALG-3/4 targets in gravid adults have not yet been identified.
ALG-3/4-branch 26G-RNAs map very sharply to the 5’ and 3’ extremities of the targets, very close to the transcription start and end sites. We find that stronger targeting at the 3’ end does not drive robust gene silencing, and may even license expression, while targeting at the 5’ end is associated with stronger gene silencing. Targeting at the 3’ is consistent with RdRP recruitment to synthesize antisense secondary 22G-RNAs throughout the transcript. These may associate with CSR-1 and could have a positive effect on gene expression. The sharp 5’ peak in the metagene analysis could hint at additional regulatory modes, other than 22G-RNA targeting. 5’-end-bound ALG-3/4 could recruit other effector factors, which promote RNA decay or translation inhibition, e.g. by inhibiting the assembly of ribosomes. Of note, when single targets are considered individually, 26G-RNA peaks at 5’ and 3’ can be simultaneously detected (Figs 7, S2B and S5A, left panels and S6). Hence, the resolution of a balance between Argonaute-sRNA complexes binding at 5’ and 3’ could determine regulatory outcome. Notably, we find shorter 3’ UTRs to be correlated with gene silencing (Fig 6F). In a model where predominant 3’ UTR targeting by Argonaute-sRNA complexes promotes gene expression, shorter 3’ UTRs and therefore less chance of sRNA binding may shift the balance towards gene silencing. Another possibility may be that longer 3’ UTRs contain binding sites for additional RNA binding proteins that may help to restrict RdRP activity on the transcript in question. Further work will be needed to test such ideas.
In C. elegans, primary sRNAs trigger the production of abundant secondary sRNAs. If left uncontrolled, such amplification mechanisms can be detrimental to biological systems. Endogenous and exogenous RNAi pathways in C. elegans compete for limiting shared factors and the Eri phenotype is a result of such competition [37,48]. Competition for shared factors is in itself a mechanism to limit accumulation of sRNAs. In support of this, exogenous RNAi was shown to affect endogenous sRNA populations, thus restricting the generations over which RNAi effects can be inherited [64].
We find that 26G-RNAs, likely ALG-3/4-bound, as well as 22G-RNAs map to alg-3 and alg-4 mRNAs (Figs 7 and S6). In the absence of GTSF-1, a loss of these sRNAs is accompanied by a 2–3 fold upregulation of alg-3 and alg-4 on the mRNA level. This means that ALG-3 and ALG-4 may regulate their own expression. In the future, the retrieval of alg-3 and alg-4 mRNAs, as well as of 26G-RNAs complementary to their sequence, in immunoprecipitations of ALG-3 or ALG-4 will strongly support this regulatory loop.
Such regulation is not unprecedented. Complementary endo-siRNAs to ago2 have been described in Drosophila S2 cells [65]. Since AGO2 is required for the biogenesis and silencing function of endo-siRNAs, it is likely that Ago2 regulates itself in S2 cells. In addition, other studies in C. elegans have described cases where sRNAs are regulating the expression of RNAi factors [64,66,67].
Such direct self-regulation of Argonaute genes may constitute an important mechanism to limit RNAi-related responses, but the biological relevance of this regulation will need to be addressed experimentally. These observations do suggest that the Eri phenotype is but one manifestation of intricate cross-regulation governing the RNAi pathways of C. elegans.
C. elegans was cultured on OP50 bacteria according to standard laboratory conditions [68]. Unless otherwise noted, worms were grown at 20°C. The Bristol strain N2 was used as the standard wild-type strain. All strains used and created in this study are listed in S5 Table.
Wide-field photomicrographs were acquired using a Leica M165FC microscope with a Leica DFC450 C camera, and were processed using Leica LAS software and ImageJ.
Cross outline. We first linked gtsf-1(xf43) and dpy-4(e1166). These genes are 2.62 cM apart, which does not comprise extremely tight linkage. Therefore, throughout the outcrossing scheme, worms were consistently genotyped for gtsf-1 and phenotyped for dpy-4. We started by outcrossing dpy-4;gtsf-1 hermaphrodites with N2 males (in a 1:2 ratio). dpy-4(e1166) is reported as being weakly semi-dominant (https://cgc.umn.edu/strain/CB1166). Indeed, heterozygote worms look only very slightly Dpy, therefore for simplicity, we refer to the heterozygote phenotype as “wild-type” throughout this work. Wild-type looking worms were selected in the F1 and F2 generations. The F2s were allowed to lay embryos for 1–2 days and then were genotyped for gtsf-1(xf43) using PCR. Progenies of non-recombined gtsf-1 heterozygote worms were kept for follow up. F3 progenies that did not segregate dpy worms were discarded. F3 dpys were isolated, allowed to lay embryos, and genotyped for gtsf-1(xf43). Progenies of non-homozygote mutant gtsf-1(xf43) worms were discarded.
RNAi. dsRNA against lir-1 was supplemented to worms by feeding as described [69]. L1 worms were transferred to RNAi plates and larval arrest was scored 2–3 days later. L1 F3 and F4 worms were transferred to RNAi plates blinded to genotype/phenotype (the dpy phenotype only shows clearly from L3 onwards).
him-5(e1467) and him-5(e1467); gtsf-1(xf43) worm populations were synchronized by bleaching, overnight hatching in M9 and plated on OP50 plates the next day. Worms were grown until adulthood for approximately 73 hours and 400–500 male animals were hand-picked for each sample, in biological triplicates, and used to isolate RNA (see below for RNA isolation protocol). Each sample was used to prepare small RNA and mRNA libraries (see below details on library preparation). Plates with the hand-picked worms were rinsed and washed 4–6 times with M9 supplemented with 0.01% Tween. 50 μL of M9 plus worms were subsequently frozen in dry ice.
N2 and rrf-3(pk1426) animal populations were synchronized by bleaching, overnight hatching in M9 and plated on OP50 plates the next day. L1 animals were allowed to recover from starvation for 5 hours, and then were collected. L2 worms were collected 11 hours after plating. L3 animals were collected 28 hours after plating. L4 animals were collected 50 hours after plating, and young adults were collected 56 hours after plating. Animals were rinsed off plates and washed 4–6 times with M9 supplemented with 0.01% Tween. 50 μL of M9 plus worms were subsequently frozen in dry ice. Embryo samples were collected from bleached gravid adult animals, followed by thorough washes with M9. Samples were collected in triplicate and RNA isolation proceeded as described below.
Worm aliquots were thawed and 500 μL of Trizol LS (Life Technologies, 10296–028) was added and mixed vigorously. Next, we employed six freeze-thaw cycles to dissolve the worms: tubes were frozen in liquid nitrogen for 30 seconds, thawed in a 37°C water bath for 2 minutes, and mixed vigorously. Following the sixth freeze-thaw cycle, 1 volume of 100% ethanol was added to the samples and mixed vigorously. Then, we added these mixtures onto Direct-zol columns (Zymo Research, R2070) and manufacturer’s instructions were followed (in-column DNase I treatment was included).
NGS library prep was performed with Illumina's TruSeq stranded mRNA LT Sample Prep Kit following Illumina’s standard protocol (Part # 15031047 Rev. E). Starting amounts of RNA used for library preparation, as well as the number of PCR cycles used in amplification, are indicated in S6 Table. Libraries were profiled in a High Sensitivity DNA on a 2100 Bioanalyzer (Agilent technologies) and quantified using the Qubit dsDNA HS Assay Kit, in a Qubit 2.0 Fluorometer (Life technologies). Number of pooled samples, Flowcell, type of run and number of cycles used in the different experiments are all indicated in S6 Table.
For maternal effect sequencing, RNA was directly used for library preparation, or treated with RppH prior to library preparation. RppH treatment was performed as described in reference 51 with slight modifications. In short, 500 ng of RNA were incubated with 5 units of RppH and 10x NEB Buffer 2 for 1 hour at 37°C. Reaction was stopped by incubating the samples with 500 mM EDTA for 5 minutes at 65°C. RNA was reprecipitated in 100% Isopropanol and ressuspended in nuclease-free water. NGS library prep was performed with NEXTflex Small RNA-Seq Kit V3 following Step A to Step G of Bioo Scientific`s standard protocol (V16.06). Both directly cloned and RppH-treated libraries were prepared with a starting amount of 200ng and amplified in 16 PCR cycles. Amplified libraries were purified by running an 8% TBE gel and size-selected for 18–40 nts. Libraries were profiled in a High Sensitivity DNA on a 2100 Bioanalyzer (Agilent technologies) and quantified using the Qubit dsDNA HS Assay Kit, in a Qubit 2.0 Fluorometer (Life technologies). All 24 samples were pooled in equimolar ratio and sequenced on 1 NextSeq 500/550 High-Output Flowcell, SR for 1x 75 cycles plus 6 cycles for the index read.
RNA from adult males was RppH-treated as described above with the difference that 800 ng of RNA were used for RppH treatment. Library preparation of these samples was performed exactly as described above with the following modifications: starting amount of 460 ng; and amplification in 15 PCR cycles.
A summary of the sequencing output can be found in S1 Table.
|
10.1371/journal.pcbi.1002782 | Kinetics of Amyloid Aggregation: A Study of the GNNQQNY Prion Sequence | The small amyloid-forming GNNQQNY fragment of the prion sequence has been the subject of extensive experimental and numerical studies over the last few years. Using unbiased molecular dynamics with the OPEP coarse-grained potential, we focus here on the onset of aggregation in a 20-mer system. With a total of 16.9 of simulations at 280 K and 300 K, we show that the GNNQQNY aggregation follows the classical nucleation theory (CNT) in that the number of monomers in the aggregate is a very reliable descriptor of aggregation. We find that the critical nucleus size in this finite-size system is between 4 and 5 monomers at 280 K and 5 and 6 at 300 K, in overall agreement with experiment. The kinetics of growth cannot be fully accounted for by the CNT, however. For example, we observe considerable rearrangements after the nucleus is formed, as the system attempts to optimize its organization. We also clearly identify two large families of structures that are selected at the onset of aggregation demonstrating the presence of well-defined polymorphism, a signature of amyloid growth, already in the 20-mer aggregate.
| Protein aggregation plays an important pathological role in numerous neurodegenerative diseases such as Alzheimer's, Parkinson's, Creutzfeldt-Jakob, the Prion disease and diabetes mellitus. In most cases, misfolded proteins are involved and aggregate irreversibly to form highly ordered insoluble macrostructures, called amyloid fibrils, which deposit in the brain. Studies have revealed that all proteins are capable of forming amyloid fibrils that all share common structural features and therefore aggregation mechanisms. The toxicity of amyloid aggregates is however not attributed to the fibrils themselves but rather to smaller more disordered aggregates, oligomers, forming parallel to or prior to fibrils. Understanding the assembly process of these amyloid oligomers is key to understanding their toxicity mechanism in order to devise a possible treatment strategy targeting these toxic aggregates. Our approach here is to computationally study the aggregation dynamics of a 20-mer of an amyloid peptide GNNQQNY from a prion protein. Our findings suggest that the assembly is a spontaneous process that can be described as a complex nucleation and growth mechanism and which can lead to two classes of morphologies for the aggregates, one of which resembles a protofibril-like structure. Such numerical studies are crucial to understanding the details of fast biological processes and complement well experimental studies.
| The aggregation of misfolded amyloid proteins into fibrils is a hallmark of many neurodegenerative diseases such as Alzheimer's, Parkinson's, and Huntington's diseases [1]–[5] and understanding amyloid aggregation mechanisms is crucial for controlling their destructive consequences. Fibrils are known to be ordered insoluble assemblies with a core cross- structure. They are not the only aggregated species involved, however, and oligomers, smaller intermediates on or off the fibril formation pathway, have been found to be responsible for amyloid cytotoxicity [6]–[8]. Their role in amyloid aggregation is still a matter of debate but significant efforts have gone into better understanding and characterizing their structure and dynamics both experimentally [9]–[11] and computationally [12]–[17]. Oligomers are often found to be precursors to amyloid fibrils. They could also, in some cases, appear as the product of a competition between the ordered fibrillar and amorphous globular morphologies, forming via different assembly pathways. This widespread characteristic of amyloid proteins is described as polymorphism [18]–[20] and is under kinetic control [21]. The presence of oligomers is therefore crucial for the fibrillisation process as well as the final morphology of fibrils [22] and understanding their kinetics of formation could be the key to controlling this polymorphism.
The aggregation of amyloid proteins is a highly cooperative self-assembly mechanism, which is often described as a complex nucleation and growth process [23]. The nucleation step, in a supersaturation environment, consists of a series of stochastic events leading to the formation of metastable seeds for the oligomer or fibril to grow on [24]. Nucleation kinetics display two characteristic properties: the presence of 1) a lag time before aggregates can be detected and 2) a maximum growth rate after nucleation is triggered [25], [26]. Direct experimental observations of nucleation and growth have been reported [27]–[30] and nucleation was always found to be the rate-limiting step of amyloid formation [26].
The aim of the present work is to investigate the dynamics of amyloid aggregation and the forces driving self-assembly for the 20-mer system of the amyloidogenic GNNQQNY peptide using molecular dynamics (MD) and a coarse-grained potential (OPEP). The nucleation specificity of the N-terminal region (9–39) of the budding yeast prion protein Sup35, GNNQQNY, is well understood. This small heptapeptide alone drives the entire Sup35 protein to self-assemble into amyloid fibrils [31] and, when isolated, displays the same amyloid properties and aggregation kinetics as the full-length Sup35 protein [32]. In addition, its cross- spine structure has been determined at the atomic level by X-ray crystallography [33]. It is therefore a very good candidate to the study of amyloid aggregation kinetics and numerous computations have been performed on the GNNQQNY sequence to characterize the onset of aggregation for this model [34]–[38]. This work expands on our previous multi-scale thermodynamic study of different sizes of GNNQQNY systems, where we identified the morphologies accessible to the trimer, dodecamer and 20-mer [39]. Now, we focus on the aggregation kinetics using long MD simulations of unbiased spontaneous self-assembly. We offer a full analysis of the onset of aggregation for GNNQQNY peptides at a refined coarse-grained level. A total of 16.9 of simulations have been collected to allow statistically relevant analyses. Altogether, our results indicate the presence of a nucleated-polymerization process intertwined with oligomer-involving mechanisms, thus leading to a certain degree of polymorphism that is already clearly established for the 20-mer.
Following Ref. [39], which showed that the GNNQQNY amyloid aggregates generated with the coarse-grained OPEP forcefield [40] were reasonably preserved in long explicit solvent all-atom MD simulations, we revisit this system focusing, this time, on the kinetics of the aggregation process.
As in our previous study, we perform implicit solvent coarse-grained molecular dynamics (MD) simulations using the OPEP potential version 3.0 [40]. OPEP is designed for efficient protein folding and structure prediction of large systems over long timescales and is also accurate for studying thermodynamics [41]. In OPEP, all heavy backbone atoms are fully represented (N, H, , C and O). Side chains, for their part, are reduced to a single bead with appropriate geometrical properties and van der Waals radius. The implicit effects of the solvent are included in the interaction parameters of the potential energy function, as detailed elsewhere [40], [42]. OPEP is a well tested potential and has been implemented with a palette of numerical methods such as Monte-Carlo [42]–[46], the activation-relaxation technique (ART nouveau) [47]–[52], MD [41], [53]–[55] and REMD [39], [56]–[59].
Here, two sets of single temperature MD are performed on a 20-mer of GNNQQNY, with both terminii of each peptide charged, in order to characterize in details the kinetics of aggregation. The first set consists of a total of 152 100 ns simulations (76 at 280 K and 76 at 300 K) with configurations saved every 5000 steps. The choice of temperatures is motivated by the fact that 280 K and 300 K are temperatures below and above the transition temperature previously found for the 20-mer of GNNQQNY. As explained below, the initial atomic positions taken for this set are extracted from the simulations reported in Ref. [39]. An additional 10 30 ns simulations are then carried out from a subset of the starting atomic positions of the previous simulation set (5 at 280 K and 5 at 300 K) with configurations saved every 50 steps to better monitor the detailed evolution of the system during the nucleation phase. All simulations are independent, starting with different random Boltzmann distributed velocities. In every case, we maintain simulation conditions as close as those of Ref. [39], with a Berendsen thermostat for temperature control [60], an integration time step of 1.5 fs and periodic boundary conditions applied to a box 200 Å in size containing 20 monomers of GNNQQNY, which represents a constant 4.15 mM concentration.
For simplicity, the starting random structures for our simulations were extracted from the high-temperature set generated in our previous REMD OPEP runs of the GNNQQNY 20-mer [39]. A typical starting structure for our simulations is shown in Figure 1 with all 20 peptides isolated and in random coil conformations. At the start of each run, a minimization procedure is performed using a combination of the steepest descent algorithm and the conjugate gradient method [61], followed by a thermalization of 50 000 steps (0.075 ns) to ensure that all conformations are fully thermalized.
Because of the implicit solvent treatment as well as the peptide's coarse-grained representation, that decrease the number of degrees of freedom, the aggregation kinetics is accelerated. It is therefore not possible to establish a direct connection between the aggregation time observed in the simulation and in experiments. However, as shown in Ref. [41], the thermodynamical properties are, at least qualitatively, maintained. The simulations presented here, therefore, should provide the right qualitative picture for the first steps in the kinetics of aggregation.
Most of the analysis on the nucleation and growth kinetics is carried out using a clustering tool [39] adapted to multiprotein assembly and designed to classify -sheet clusters based on strand attachment. For the purpose of this work, this procedure can also handle the calculation of kinetic association and dissociation rates. To assess strand attachment, the criterion used to define and calculate hydrogen bonds between strands is similar to the DSSP definition [62]. A peptide belongs to a cluster if it is attached to another strand of that cluster by at least two hydrogen bonds. An additional criterion is applied on dihedrals and angles to determine if a given strand in a cluster has enough amino acids in -conformation. For each amino acid the and angles are calculated and if they satisfy the region (in degrees): [−180∶−150;0∶180], (in degrees): [−180∶0;150∶180] (corresponding approximately to the region of the Ramachandran plot [63]), the amino acid is in a state. A GNNQQNY peptide is considered in a state if at least three of its residues are in the region. If a peptide is not found to be in a state, it is excluded from the cluster. This determination of secondary structure is solely used to determine cluster membership of the strands. The clustering analysis allows us to measure accurately the evolution of clusters over time based on local information and to monitor their properties such as the orientation of strands within -sheets (i.e., parallel or anti-parallel). For purposes other than cluster determination, secondary structure calculations are made using the STRIDE program [64].
In order to look at the aggregation process in more details, we also consider the association and dissociation rates of the clusters in the following way. With the concentration of , we consider aggregation as a dynamical process involving both association and dissociation that can occur either one monomer or more than one monomer at a time. The former is referred to as growth by monomer addition/monomer loss while the latter is described as being a mix of two processes, oligomer fusion/fragmentation and formation/destruction of oligomers from/into monomers, when involving more than one monomer at a time. We can then define the net rate of creation of as(1)where and are the creation rate of into and the destruction rate of into , and are the creation and destruction rates of either directly from/into monomers, or from the fusion/fragmentation of other sizes of oligomers. All the C and D rates are calculated from our clustering tool and allow us to gather statistics on the microscopic kinetic events and mechanisms.
We present a study of the aggregation kinetics of 20-mer GNNQQNY oligomers under a 4.15 mM concentration, the same concentration that was used in our previous multiscale thermodynamic study of the GNNQQNY 20-mer system [39].
We first present the general results obtained from the 100 ns MD simulations whose configurations were saved every 7.5 ps (5000 simulation steps) with an initial configuration selected as discussed in the materials and methods section. Then we discuss results from the 30-ns MD simulations whose configurations are saved every 75 fs (50 simulation steps) to better study the detailed association and dissociation kinetics of oligomers.
At the lowest temperature of 280 K, all 76 100 ns simulations lead to ordered amyloid oligomers formation. In all cases, aggregation is accompanied by a sudden drop of the total potential energy of the system, by over 600 kcal/mol over less than 10 ns, and by an increase in the -sheet content of 30%, as calculated with the STRIDE program [64]. While the exact energy value is not significative, due to the implicit-solvent coarse-grained nature of our energy model, its drop corresponds to the formation of a more stable structure. The system then stays in a minimum of energy and both the number of hydrogen bonds and the amount of secondary structure stabilize. As shown in Figure 2(a), which presents a typical aggregation run, the -content in the structures fluctuates typically around 50%, near its maximum of 60%, as the glycines and tyrosines end residues of each 20 peptides do not get involved in the -sheet hydrogen bonding. Figure 2(a) also shows the high correlation between the energy drop and the increase in the number of hydrogen bonds as a function of time, suggesting that the cooperativity between hydrogen bonds plays a crucial role in lowering the energy and stabilizing the system.
Aggregation is slower at 300 K and only 68% of the 76 100 ns simulations lead to ordered amyloids. However, as shown in the typical aggregation run in Figure 2(b), the overall ordering follows a trend very similar to that at the lower temperature : a sudden potential energy drop of over 600 kcal/mol over less than 10 ns accompanied by correlated raises in both the number of hydrogen bonds and the -sheet content. If the final number of hydrogen bonds is very similar to that at 280 K, the secondary structure is less stable and tends to fluctuate around 40% rather than 50%.
In order to describe the assembly process we represent the time evolution, the probability density and the orientation of strands in structures as a function of the number of hydrogen bonds and of the number of contacts between side chains as these two coordinates are the least correlated and are the best measure of how ordered the structures are. Figures 3 and S1 show these quantities for the trajectories plotted in Figure 2 at 280 K and 300 K, respectively.
At 280 K, we observe three distinct kinetic stages over the course of a typical simulation (Figure 3(a)). The first phase is characterized as the nucleation phase, which lasts about 5 ns after the start of the simulation and leads to the formation of the metastable critical nucleus. During this phase, small oligomers form and break under stochastic collisions of the monomers. Seeds below the nucleus size fluctuate considerably, forming and disassembling at a high rate, forming a quasi-equilibrium perfectly reversible process. Once the metastable nucleus forms, the system can move into the aggregation (or growth) phase with a 50% probability, by definition. In this dynamical phase, almost all of the monomers rapidly assemble around the nucleus to form a partially disordered globular oligomer. In general, this stage is very rapid and typically lasts less than 10 ns. During the third phase, which extends over a timescale of up to 80 ns, the aggregate rearranges itself as monomers explore their local configuration environment within the confines of the oligomers, optimizing the energy and, as a consequence, the secondary structure and the number of side chain – side chain contacts (see the last 75 ns in Figure 3 (a) and (c)). This process, which we describe as a stabilization phase, is the slowest of the three and accounts for the dense region in Figure 3(b).
This aggregation process is consistent with the “condensation-ordering” mechanism previously observed experimentally [65] and computationally [12]–[14], [66]. An interesting feature of the kinetics at 280 K is the increase and later dominance of parallel orientation in the structures over time during both the growth and stabilization phases while the structures are mostly antiparallel during the nucleation phase (Figure 3(c)). By looking at the color coding on the right axis, it appears as though the system is loosing some parallel orientation between region 1 and 2 from almost 100% to 80%. Instead our results indicate that the system continues to evolve and gain some secondary structure between region 1 and 2 of the graph. It is the newly formed -strands that adopt an antiparallel orientation while the parallel content formed during the growth process remains unchanged. As a whole, 91% of the MD simulations at 280 K lead to a final assembly dominated by parallel -sheets, in agreement with recent experimental findings [33], [67], [68] and computational studies [35], [36], [39], [69], [70].
At 300 K the kinetics globally display the same three phases for nucleation, growth and stabilization of oligomers observed at 280 K, and 95% of the final aggregated structures display a dominance of parallel orientation of the -strands (Figure S1). The main difference between the two temperatures (Figure 2) is in the lag time associated with the nucleation phase: while the average lag time is found to be 13 ns at 280 K, it increased to 56 ns at 300 K, leading to a denser nucleation region on the probability map (Figure S1(b)). Mechanistically, this increase in nucleation time can be explained by the presence of bigger thermal fluctuations that destabilize the metastable aggregates, preventing nucleation.
While most simulations at 280 K and 300 K generate a single aggregation event, we observe reversibility for 34% of aggregation events at 280 K against 40% at 300 K. In these cases, such as in the example shown in Figure 4, monomers undergo a complete aggregation process up to and including the stabilization phase before the reverse reaction takes place, leading to a completely or partially random structure. For some simulations, this reversible transition was even observed to occur a few times during the 100 ns run. The presence of reversibility tells us that even though the free energy barrier for forming a 20-mer oligomer is high, the system is not completely biased towards the formation of an ordered oligomer. Thermal fluctuations for this 20-mer are sufficient to destabilize ordered oligomers on a relatively short time scale, a process that cannot be achieved in all coarse-grained aggregation simulations [16], [71] but which is crucial in order to describe aggregation kinetics correctly.
In this section we present the analysis of the 10 30 ns MD simulations, five at 280 K and five at 300 K, whose configurations are saved every 75 fs in order to describe the details of the kinetics during the final nucleation and full growth process. Because of the tremendous size of the resulting simulation data, we concentrated our analysis on a 10 ns window centered around the drop in energy (Figure 7). Panel (a) represents the average energy taken over all five simulations as a function of time at 280 K. Trajectories are aligned, in time, at the point at which they reach −80 kcal/mol, which is roughly the midpoint in the energy drop for all simulations. Most of the energy drop associated with oligomeric growth, on the order of 600 kcal/mol100 kcal/mol, takes place over 4 ns, in agreement with our earlier observations for a typical aggregation process at 280 K. The relatively small error bars along the energy curve indicate the good reproducibility of the properties over time at 280 K. At 300 K, the growth phase associated with the energy drop, of about 450 kcal/mol200 kcal/mol, also takes on the order of 4 ns (Figure 7(b)), similar to a 280 K energy drop. The standard deviation on the 300 K curve is, however, greater than at 280 K, demonstrating a greater variability associated with larger thermal fluctuations.
At both temperatures 280 K and 300 K, aggregation is generally triggered by the formation of a small-sized metastable aggregate, which appears to be stable after a certain lag time. This suggests that we are in the presence of an assembly sequence that can be classified as a nucleated-growth process [24], [26], [84], [86]–[90], i.e., that this small metastable aggregate, which we term nucleus, serves as a nucleation center of the aggregation process. The 152 100 ns MD simulations were divided in 3 sets at both 280 K and 300 K and we computed the free energy as a function of aggregate size and secondary structure for those 3 sets of simulations in order to determine the size and amount of secondary structure of the critical nucleus (Figure 10). Performing this task on different sets of data allows us to have an idea on the order of the fluctuations in the free-energy. At 280 K the nucleus size corresponding to the maximum of free energy is found to be between 4 (Figure 10(a) - green curve) and 5 monomers (Figure 10(a) - red and blue curves) and between 5 (Figure 10(b) - red and blue curves) and 6 monomers (Figure 10(b) - green curve) at 300 K. This result is expected since larger thermal fluctuations require a bigger aggregate to survive and lead to growth. The pentameric critical nucleus identified here is also near the critical size estimated by Nelson et al. [33] and by us, in a previous thermodynamic study [39].
As was shown recently [91], [92], the critical nucleus size in a finite-size system is systematically overestimated and it is necessary to correct for this artifact. From the classical nucleation theory (CNT), Grossier et al. derive an expression for the total free energy of forming an aggregate of size monomers in an infinitely large system to be [91]:(2)where is the aggregate size, is the Boltzmann constant, is the temperature, is a dimensionless constant and represents the supersaturation and is the interfacial energy (or surface tension) taken to be a constant in the model. Due to our very small system size, a 20-mer, and the low critical nucleus, it is not possible to obtain a good fit to this continuous equation. However, the overestimation correction could explain the slight difference we observe with respect to the experimentally-derived critical nucleus of four monomers.
Looking at the free energy barrier of forming a certain amount of secondary structure, we find that a viable nucleus requires between 24 and 28 residues in conformation at 280 K while it requires between 27 and 29 residues in conformation at 300 K (Figure 10 (c) and (d)). The increase in free energy for 80 residues is due to the finite-size effects of our system. It becomes harder to have 80 residues in -conformation as no more monomers are available to the system to continue growth. Figure 10 (e) and (f) show the dominant pentamer nucleus structure having such amount of secondary structure at 280 K and 300 K. In both cases, the pentamer seed is partially ordered. In most cases, no more than a dimer is formed beside the nucleus.
To assess the microscopic mechanisms involved in the kinetics, we first identify all types of association and dissociation: growth by monomer addition (and, reversibly, loss by monomer subtraction), growth by fusing two oligomers together (and, reversibly, fragmentation of one oligomer into two smaller oligomers at least 2 monomers in size) and the direct formation/destruction of oligomers from/into monomers. In this section, we refer to any aggregate bigger than one monomer as an oligomer. It is important to point out that there is a wealth of “monomer addition” models for diverse polymer-forming proteins such as actin [93], [94], tubulin [82], the sickle cell hemoglobin [25], [74] and amyloid proteins such as A [27], [95], [96], 2-microglobulin [79] and Sup35 [78]. There also exists numerous “oligomer fusion” models for A [9]–[11], [97], -synuclein [73], [97], [98], the phosphoglycerate kinase protein [99], the lysozyme protein [100] and Sup35 [65], [101], some of which have observed both processes happening at the same time.
Association and dissociation rates were calculated, with our clustering code, every 75 fs over a 10 ns window (centered around the energy drop) for the 30 ns simulations and as described in Eq. (1). Then, for each time interval, we calculated the total number of events, originating either from monomer addition/loss, from oligomer fusion/fragmentation or from monomersoligomers events across all species such as:(3)and(4)where and are the “monomer addition/loss” and “oligomer fusion/fragmentation” + monomersoligomers components of Eq. (1).
Figure 11 shows the evolution of these two quantities for both association and dissociation events at 280 K (Figure 11(a)) and at 300 K (Figure 11(b)). We differentiate the fusion/fragmentation events from the formation/destruction of oligomers (bigger than dimers) directly from/into monomers. At both temperatures, the data clearly shows that the assembly mechanism is dominated by “monomer addition/loss” events. Then when nucleation and aggregation happen, we see a notable increase in the amount of monomer events and a trigger of “oligomer fusion/fragmentation” and “monomersoligomers” events. We notice a well-defined increase in the number of “monomer addition/loss” events just before the first “oligomer fusion/fragmentation” events appear. This increase corresponds to the start of nucleation and suggests that once nucleation is triggered and most of the monomers are recruited, they join different sites, or clusters, that will later on fuse together to form a larger oligomer. Later, when the aggregate stops growing in size, we observe no more “single monomer” or “monomersoligomers” events and observe, in some cases, the presence of only fusion and fragmentation of oligomers (Figure 11(a)). This means that further rearrangements in the structure during the stabilization phase are accomplished mainly through oligomer-involving events, if any.
We presented here a detailed study of the onset of amyloid aggregation for 20-mers of GNNQQNY. Using molecular dynamics with the OPEP coarse-grained force field, we show that nucleation of this small amyloid peptide is dominated by monomer addition/loss events, with very small contributions from larger oligomers, following closely the classical nucleation theory. It is then meaningful to extract a critical nucleus, that can be obtained from the calculation of the free-energy as a function of nucleus size. We find that, at 280 K, this critical size is between 4 and 5 monomers, while it is between 5 and 6 at 300 K, in good agreement with the experimental estimate of 4 monomers [33], especially when taking into account the finite-size bias that tends to overestimate the size of the critical nucleus [91], [92]. Correspondence with CNT stops there, however, as the kinetic process associated with aggregation and growth differs in two majors from this theory. First, while most of the structural organization takes place during the 4 ns growth process, aggregates continue to mature by collective motions, slowly dropping in energy as hydrogen bonds and -sheet content evolve. Second, nucleation does not lead to a single structure, but shows clear polymorphism with a distribution of assemblies that can be classified into two distinct categories: a compact oligomer made of a number of relatively short -sheets, typically three, and a more extended fibril-compatible two-sheet structure. These structures represent well-separated local basins and the only way to move between them, in our simulations, was through a complete dissociation and reassociation of the monomers. The well-defined polymorphic nature of GNNQQNY is in line with experimental and numerical observations in other amyloid sequences, such as amyloid-. It was shown there that the protein could adopt multiple fibrillar structures [18], [102], but also off-pathway -barrel organizations that would be responsible for at least part of the toxicity. [103] For GNNQQNY, the two polymorph families observed here are close enough that they should lead to different fibrillar structures rather than on and off-pathway organizations. Only simulations with a larger number of peptides will be able to tell.
How much of these results can be applied to experimental studies of GNNQQNY? A previous stability study of the structures predicted with OPEP using explicit SPC solvent and all-atom GROMOS96 showed that our simulations are realistic, except for the most extended structures [39]. If the growth time is not directly extendable to all-atom systems, the thermodynamics and, therefore, the critical nucleus size but also the polymorphism, which is a signature of amyloid aggregates, should be valid. Our results suggest that the specific shape, out of a family of structures, is selected very early on and that moving from one to another requires going over a very high barrier, high enough that it was never observed in our simulations, the preferred being going first through a complete dissociation. Such behavior could change with larger aggregates, and the direct rearrangement become more favorable than complete dissociation. Only further work, on larger systems, will show whether new families of structures are possible for GNNQQNY and if the CNT applies when more monomers are in play. Our results on the 20-mer of GNNQQNY are at least compatible with experiments and offer a number of insights into the onset of aggregation and polymorphism for small amyloid peptides.
|
10.1371/journal.pgen.1001377 | PDP-1 Links the TGF-β and IIS Pathways to Regulate Longevity, Development, and Metabolism | The insulin/IGF-1 signaling (IIS) pathway is a conserved regulator of longevity, development, and metabolism. In Caenorhabditis elegans IIS involves activation of DAF-2 (insulin/IGF-1 receptor tyrosine kinase), AGE-1 (PI 3-kinase), and additional downstream serine/threonine kinases that ultimately phosphorylate and negatively regulate the single FOXO transcription factor homolog DAF-16. Phosphatases help to maintain cellular signaling homeostasis by counterbalancing kinase activity. However, few phosphatases have been identified that negatively regulate the IIS pathway. Here we identify and characterize pdp-1 as a novel negative modulator of the IIS pathway. We show that PDP-1 regulates multiple outputs of IIS such as longevity, fat storage, and dauer diapause. In addition, PDP-1 promotes DAF-16 nuclear localization and transcriptional activity. Interestingly, genetic epistasis analyses place PDP-1 in the DAF-7/TGF-β signaling pathway, at the level of the R-SMAD proteins DAF-14 and DAF-8. Further investigation into how a component of TGF-β signaling affects multiple outputs of IIS/DAF-16, revealed extensive crosstalk between these two well-conserved signaling pathways. We find that PDP-1 modulates the expression of several insulin genes that are likely to feed into the IIS pathway to regulate DAF-16 activity. Importantly, dysregulation of IIS and TGF-β signaling has been implicated in diseases such as Type 2 Diabetes, obesity, and cancer. Our results may provide a new perspective in understanding of the regulation of these pathways under normal conditions and in the context of disease.
| Cells in the body respond to a variety of on/off signals that are relayed in a defined spatial and temporal manner. These signals influence several processes such as growth, fat storage, and the repair of damaged molecules. As humans age, the onset of diseases such as Type 2 Diabetes, obesity, and cancer often results from an imbalance in the levels of on/off signals in the cell. The insulin/IGF-1 signaling pathway is an important regulator of longevity, development, and metabolism across phylogeny. While the protein kinases that activate this pathway have been well studied, less is known about the protein phosphatases that tune down the signals. The roundworm C. elegans has been an excellent model system to study the role of insulin/IGF-1 signaling in the aging process. Here, we identify a new phosphatase that negatively regulates the insulin/IGF-1 pathway to enhance longevity and stress-resistance. Interestingly, the phosphatase achieves this function by tuning down the activity of a conserved TGF-β pathway, a pathway important for development. By reducing TGF-β pathway activity, this phosphatase decreases expression of insulin molecules that may stimulate the insulin/IGF-1 pathway. Our studies not only unravel a new regulator of these pathways, but also point to how they are more linked than previously thought. Both insulin/IGF-1 and TGF-β signaling have been implicated in age-associated diseases, and understanding their connection will provide us with potential therapeutic avenues.
| Insulin/IGF-1 signaling (IIS) is a conserved neuroendocrine pathway that regulates longevity, development and energy metabolism across phylogeny [1], [2]. In the roundworm Caenorhabditis elegans (C. elegans), activation of the DAF-2 insulin/IGF-1 receptor tyrosine kinase intiates an AAP-1/AGE-1 PI 3-kinase signaling cascade involving the downstream serine/threonine kinases PDK-1, AKT-1, and AKT-2 [3]–[7]. Activated AKT-1 and AKT-2 phosphorylate DAF-16, the single Forkhead Box O (FOXO) family transcription factor homolog in C. elegans [8]. Phosphorylation of DAF-16 results in its inactivation and sequestration in the cytosol [9], [10]. Under low signaling conditions, DAF-16 translocates into the nucleus, where it can transactivate/repress hundreds of target genes [9]–[13].
The dauer is an alternative survival stage that worms can enter upon poor environmental conditions such as crowding [14]. Mutations in the kinases upstream of DAF-16 such as daf-2, age-1, pdk-1, akt-1 and akt-2 result in an increase in lifespan, dauer formation, fat storage and/or stress resistance, and loss-of-function mutations in daf-16 completely suppress these phenotypes [15]–[18]. In addition to the IIS pathway, dauer formation in C. elegans is also regulated by the DAF-7/TGF-β-like signaling pathway [19]–[21]. Activation of TGF-β signaling is achieved through binding of the DAF-7 BMP-like ligand to the DAF-1/DAF-4, the Type I/II receptors, which phosphorylate and activate the downstream receptor-associated SMAD (R-SMAD) proteins DAF-8 and DAF-14, presumably through a conserved SSXS phosphorylation motif that has been shown to be important for R-SMAD activation in mammals [22]–[24]. Upon activation, R-SMADs can associate with a Co-SMAD to regulate the transcription of hundreds of genes [23], [25]. In C. elegans, DAF-8 and DAF-14 act to antagonize the transcriptional activity of the DAF-3 Co-SMAD and the DAF-5 SNO-SKI repressor [22], [24], [26]–[29]. Reduction of function mutations in daf-7, daf-1, daf-4, daf-8 and daf-14 show temperature-sensitive constitutive dauer formation and mutations in daf-3 and/or daf-5 completely suppress this phenotype [19], [21], [30]. Genetic epistasis studies have suggested that the TGF-β pathway acts in a parallel manner with IIS to modulate dauer formation [31]–[33].
The PTEN lipid phosphatase homolog DAF-18, which antagonizes signaling at the level of AGE-1/PI 3-kinase, is a major negative regulator of IIS. In contrast to the kinases in this pathway, loss-of-function mutations in daf-18 reduces lifespan, fat storage, dauer formation and stress resistance [32], [34]–[39]. Besides DAF-18, few negative modulators of the pathway have been identified. In particular, less is known about serine/threonine phosphatases that counterbalance kinase activity in the IIS pathway. We recently performed a directed RNA interference (RNAi) screen for serine/threonine phosphatases that regulate C. elegans IIS using dauer formation as an output [39]. We identified the PP2A regulatory subunit PPTR-1 as an important regulator of AKT-1 dephosphorylation as well as DAF-16-dependent phenotypes [39]. Here we characterize another candidate from this screen, pdp-1, as a positive regulator of dauer formation. PDP-1 is homologous to pyruvate dehydrogenase phosphatase (PDP) in higher organisms, an enzyme that positively regulates the pyruvate dehydrogenase enzyme complex (PDHc). RNAi of the other components of PDHc do not result in changes in dauer formation. Interestingly, we report that although PDP-1 is a robust modulator of multiple IIS-regulated processes as well as DAF-16 activity, genetic epistasis studies place pdp-1 in the DAF-7/TGF-β pathway. Through this study, we find that IIS and TGF-β signaling are more tightly connected than previously suggested, with distinct roles for the Co-SMAD DAF-3 in modulating the IIS pathway. Our data suggests that PDP-1 modulates the gene expression of several insulins, and that insulins may be a potential mediator of the crosstalk between these two pathways.
Our RNAi screen was designed to identify serine/threonine phosphatases that modulated dauer formation of daf-2(e1370), a non-null, temperature-sensitive mutant of the C. elegans insulin/IGF-1 receptor gene, daf-2 [39]. We were particularly interested in phosphatases that would negatively regulate IIS similar to DAF-18/PTEN, and for all RNAi based assays described below, daf-18 RNAi was used as a positive control [39]. From this screen, we identified pdp-1 as a modulator of daf-2(e1370) dauer formation (Figure 1A and Figure S2). BLAST analyses using amino acid sequence revealed that PDP-1 is homologous to fly and mammalian PDP (∼52% positive and ∼38% identical). pdp-1 RNAi significantly reduces dauer formation of daf-2(e1370) worms, similar to daf-18 RNAi (Figure 1A and Figure S2). This phenotype is not allele-specific, as pdp-1 RNAi results in suppression of dauer formation in a second allele of daf-2, daf-2(e1368) (Figure 1B and Figure S2). Similar to the results with the RNAi, a mutation in pdp-1 also affects dauer formation - pdp-1(tm3734); daf-2(e1370) double mutants form significantly fewer dauers when compared to the daf-2(e1370) parental strain (Figure S2).
Given its homology to PDP in higher organisms, we wondered whether the effect of pdp-1 knockdown on daf-2 dauer formation was a consequence of modulating the activity of the PDHc. The PDHc is a multi-subunit enzyme complex consisting of three major enzymes: E1 pyruvate dehydrogenase, E2 dihydrolipoyl acetyltransferase and E3 dihydrolipoyl dehydrogenase that regulate energy metabolism [40]. PDHc converts pyruvate to acetyl-coA, which can either enter the TCA cycle or be used for fatty acid synthesis. In mammals, regulation of PDHc activity is primarily achieved through reversible phosphorylation/dephosphorylation of the E1α subunit by pyruvate dehydrogenase kinase (PDHK) and PDP, with phosphorylation inactivating the enzyme complex [40]. All of the components of the PDH complex have conserved C. elegans homologs, encoded by the genes T05H10.6 (E1α), C04C3.3 (E1β), F23B12.5 (E2), LLC1.3 (E3), pdhk-2 (PDHK) and pdp-1 (PDP).
To test whether modulation of PDHc activity affects daf-2 dauer formation, we grew daf-2(e1370) worms on PDHc RNAi. Quantification the RNAi efficiency of the PDHc components revealed that we achieved 60–90% knockdown (Figure S1). To our surprise, RNAi of the E1α subunit had no effect on daf-2 dauer formation, while pdp-1 RNAi resulted in dauer suppression (Figure 1C and Figure S2). In addition, RNAi of either the other E1 subunit E1β, or the E2 subunit, did not affect daf-2 dauer formation (Figure 1C and Figure S2). Knockdown of the E3 subunit resulted in lethality (data not shown). Interestingly, pdhk-2 RNAi resulted in slight suppression daf-2(e1370) dauer formation but had no effect on dauer formation of daf-2(e1368) mutants (Figure 1C and Figure S2). Therefore pdhk-2 modulates the IIS pathway in an allele-specific manner and we did not perform further characterization of this gene.
To further evaluate the components of the PDH complex, we examined their expression patterns. The expression pattern of PDP-1 does not completely overlap with that of the E1 and E2 subunits of PDHc (Figure S3). PDP-1 expression was enriched in the head and tail neurons, head muscle and the intestine. We did not observe any expression in the pharynx. In contrast, the expression of the E1 and E2 subunits, was observed throughout the body of the worm and was significantly enriched in the pharynx. Taken together, PDP-1 modulates daf-2 dauer formation and this function is likely to be independent of its role in regulating the PDHc.
In addition to dauer formation, the IIS pathway also regulates longevity, stress resistance and fat storage [17], [18]. Mutations in daf-2 and age-1 result in a significant extension in lifespan, enhanced resistance to various stresses and increased fat storage [7], [35], [41]–[44]. These phenotypes are suppressed by loss-of-function mutations in daf-18 and daf-16 [32], [34], [35], [39]. We therefore investigated whether dosage modulation of pdp-1 would affect additional outputs of the pathway. We first tested the role of PDP-1 in regulating lifespan (Figure 2 and Figure S4). The lifespan of wild-type worms was not affected by pdp-1 RNAi and slightly reduced by a mutation in pdp-1 (Figure 2A and 2D). In contrast, the mean and maximal lifespan of long-lived daf-2(e1370) and age-1(hx546) mutants was significantly reduced by pdp-1 RNAi (Figure 2B and 2C). Similarly, pdp-1(tm3734); daf-2(e1370) double mutants lived significantly shorter than the parental daf-2(e1370) strain (Figure S4).
To examine the effect of increased dosage of pdp-1, we generated transgenic worms bearing a translational fusion containing pdp-1 fused to gfp and driven by its own promoter (pdp-1::gfp). In addition, we also crossed the pdp-1::gfp worms to daf-2(e1370) mutants to generate the daf-2(e1370); pdp-1::gfp strain. Overexpression of pdp-1 results in a significant extension in lifespan compared to wild-type worms (Figure 2D and Figure S4). Interestingly, pdp-1 overexpression further extends the lifespan of daf-2(e1370) mutants (Figure 2B and Figure S4). In both of these cases, the increased lifespan was suppressed by daf-16 RNAi (Figure S5). Therefore, dosage modulation of pdp-1 regulates lifespan in a DAF-16 dependent manner.
Next, we asked if PDP-1 modulated additional outputs of the IIS signaling pathway. We first tested whether PDP-1 regulates stress resistance by assaying the survival of pdp-1 mutants and transgenic animals when exposed to heat stress at 37°C (Figure 2E and Figure S7). Dosage modulation of pdp-1 affects the response to heat stress, with a pdp-1 mutation decreasing and pdp-1 overexpression slightly enhancing thermotolerance (Figure 2E). Importantly a pdp-1 mutation drastically reduced the thermotolerance of daf-2 mutants (Figure 2E).
To examine the role of pdp-1 in regulating fat storage, we used both Oil Red O [45] and Sudan Black [7] staining (Figure 2F and 2G and Figure S7). pdp-1 mutants had similar levels of fat compared to wild-type worms, while overexpression of pdp-1 slightly enhanced fat storage (Figure S7). In contrast, a pdp-1 mutation drastically reduced the increased fat of daf-2(e1370) mutants (Figure 2F and 2G and Figure S7). This was observed in dauers, larval stage 3 (L3) animals and adults, suggesting that PDP-1 is an important regulator of fat storage in daf-2 mutants. We did observe any further enhancement of the increased fat storage in the daf-2(e1370); pdp-1::gfp worms (Figure S7). Importantly, the increased fat storage of pdp-1::gfp and daf-2(e1370); pdp-1::gfp worms was suppressed by daf-16 RNAi, similar to daf-2 mutants (Figure S7). Thus, PDP-1 modulates all four well-characterized outputs of the IIS pathway.
In addition to these phenotypes, pdp-1(tm3734) mutants exhibit a slow movement phenotype, which we quantified using locomotion assays (Figure S6). This slow movement was rescued by the pdp-1::gfp transgene. In addition, we performed brood size analysis of wild-type, pdp-1(tm3734) mutants, daf-2(e1370) mutants, and pdp-1(tm3734); daf-2(e1370) double mutants (Figure S6). pdp-1(tm3734) worms showed a slight decrease in the number of progeny compared to wild-type worms. However, when compared to daf-2 mutants, only 5% of the pdp-1(tm3734); daf-2(e1370) eggs yielded progeny (Figure S6). daf-2 mutants have a slightly reduced brood size [46], [47], and a mutation in pdp-1 severely enhances this phenotype . Taken together, PDP-1 regulates multiple outputs of IIS and acts as a negative regulator the pathway, similar to DAF-18/PTEN.
The FOXO transcription factor DAF-16 is the major target of the C. elegans IIS pathway [2], [48]. Under conditions of reduced IIS, DAF-16 is able to translocate into the nucleus, where it regulates the expression of hundreds target genes [12], [13], [49], [50]. We therefore asked whether PDP-1 modulates DAF-16 subcellular localization as well as activity (Figure 3A and Figure S8). daf-2(e1370); daf-16::gfp worms were grown on vector, daf-18 and pdp-1 RNAi, and DAF-16 nuclear/cytosolic localization was visualized using fluorescence microscopy and quantified. Throughout the body of the worm, while DAF-16::GFP was mostly nuclear on vector RNAi, its localization was enriched in the cytosol on pdp-1 RNAi, similar to daf-18 RNAi (Figure 3A and Figure S8).
The gene superoxide dismutase 3 (sod-3) is a direct DAF-16 target [11]. To test whether PDP-1 modulates transcriptional activity of DAF-16, we used a Psod-3::gfp reporter strain in a daf-2(e1370) background [51]. daf-2(e1370); Psod-3::gfp worms were grown on vector, pdp-1, daf-18 and daf-16 RNAi and GFP expression was visualized using fluorescence microscopy and scored as low, medium or high (Figure 3B and Figure S8). GFP expression was markedly lower on pdp-1 RNAi compared to vector RNAi, similar to daf-18 and daf-16 RNAi, suggesting that PDP-1 positively modulates DAF-16 transcriptional activity. To further validate these results, we used quantitative real-time PCR (Q-PCR) to look at the expression levels of well-known DAF-16 target genes [52] in daf-2(e1370), pdp-1(tm3734); daf-2(e1370) and daf-16(mgDf50); daf-2(e1370) worms (Figure 3C). Notably, the expression of sod-3, sod-5 and hsp-12.6 was significantly reduced in pdp-1(tm3734); daf-2(e1370) mutants relative to daf-2(e1370). Therefore PDP-1 positively regulates a subset of DAF-16 targets.
Thus far our data indicates that PDP-1 regulates multiple outputs of IIS as well as DAF-16 activity. Using dauer formation as the readout, we performed genetic epistasis experiments to identify the substrate of PDP-1. We first tested whether pdp-1 acted directly through the IIS pathway by focusing on kinase mutants downstream of daf-2 (Table 1 and Figure S9). pdk-1(sa680), daf-2(e1370); akt-1(ok525) and daf-2(e1370); akt-2(ok393) mutants were maintained on vector, daf-18 and pdp-1 RNAi and dauer formation of these strains was assayed at the appropriate temperatures. Interestingly, pdp-1 RNAi resulted in suppression of dauer formation of pdk-1(sa680) mutants, daf-2(e1370); akt-1(ok525) and daf-2(e1370); akt-2(ok393) worms (Table 1 and Figure S9). DAF-16 is downstream of the AKT kinases in the pathway, but we were unable to detect a physical interaction between PDP-1 and DAF-16 (data not shown).
We next examined a TGF-β pathway that also regulates dauer formation [19]–[21] using genetic epistasis analyses with mutants of this pathway. In these assays, TGF-β pathway mutants were maintained on vector RNAi, pdp-1 RNAi and daf-3 RNAi (as a positive control; Table 2 and Figure S10). We first tested daf-7 mutants, which contain a mutation in the gene encoding the TGF-β ligand [53]. Dauer formation of daf-7(e1372) mutants was suppressed on pdp-1 RNAi similar to daf-3 RNAi, suggesting that pdp-1 does not function at the level of daf-7 (Table 2 and Figure S10). Next, we tested dauer formation with mutants of the SMADS daf-8 and daf-14 [22]. We grew daf-14(m77) mutants on vector, pdp-1 and daf-3 RNAi. Interestingly, pdp-1 RNAi had no effect on daf-14 dauer formation, while daf-3 RNAi still resulted in suppression (Table 2 and Figure S10). We next looked at dauer formation of daf-8(m85) mutants and again observed that pdp-1 RNAi had no effect, while daf-3 RNAi suppressed dauer formation (Table 2 and Figure S10). Therefore, our genetic epistasis results indicate a genetic interaction between pdp-1 and daf-14/daf-8.
To confirm these results, we investigated whether pdp-1 RNAi could suppress dauer formation of daf-2(e1370); daf-3(mgDf90) double mutants (Table 2 and Figure S10). In this strain, input from the TGF-β pathway is removed due to the daf-3 null mutation, and dauer formation is presumably mediated through activated DAF-16 [39]. Therefore, if pdp-1 was indeed acting in the TGF-β pathway, we would not see any effect of pdp-1 RNAi on daf-2(e1370); daf-3(mgDf90) double mutants. Expectedly, pdp-1 RNAi had no effect on daf-2(e1370); daf-3(mgDf90) double mutants (Table 2 and Figure S10). DAF-3 itself is unlikely to be a substrate for PDP-1, as similar to mammalian Co-SMADs, it lacks the SMAD phosphorylation motif [28]. Therefore, our genetic epistasis analysis supports a model whereby pdp-1 acts in the DAF-7 TGF-β pathway at the level of daf-8 and daf-14.
How does a phosphatase in the TGF-β signaling pathway have such robust effects on the outputs of the IIS pathway and DAF-16? A number of studies have previously identified roles for the TGF-β pathway in lifespan and fat storage [7], [54], [55]. However, genetic epistasis analysis on dauer formation placed DAF-7 TGF-β signaling and IIS as two parallel pathways where components of one pathway did not affect the other [14], [56], [57]. Yet in our studies, PDP-1 was able to regulate multiple outputs of IIS. Therefore, we decided to further investigate the potential crosstalk between the IIS and TGF-β signaling pathways. First, we focused on DAF-3 and DAF-5, which are positive regulators of dauer formation similar to PDP-1, and asked whether mutations in daf-3 or daf-5 could also affect phenotypes of the IIS pathway [14], [28], [29].
We tested lifespan, fat storage, dauer formation and stress resistance of TGF-β pathway mutants in a wild-type as well as daf-2(e1370) background. (Figure 4A–4C, Figure S11, 12, S13 and Table S1). As previously reported, the lifespan of daf-3 and daf-5 single mutants is slightly shorter than wild-type worms (Table S1) [55]. In our hands, mutations in the upstream components of the TGF-β pathway such as daf-7 and daf-14 enhance dauer formation but do not significantly extend lifespan (Table S1 and Figure S4). Intriguingly, mutations in daf-3 and daf-5 have opposite effects on daf-2(e1370) phenotypes. When compared to the daf-2(e1370) parental strain, daf-2(e1370); daf-3(mgDf90) mutants lived significantly longer. This was also observed in daf-2(e1370); daf-3(e1376) worms, which is a weaker allele of daf-3. In contrast, daf-5(e1386); daf-2(e1370) double mutants live much shorter than daf-2(e1370) worms (Figure 4A, Figure S13 and Table S1). A mutation in daf-5 also decreased the increased lifespan of age-1(hx546) worms, with age-1(hx546); daf-5(e1385) double mutants living significantly shorter than the parental strain (Figure S13). Importantly, for daf-2 worms, the effect of a daf-3 null mutation on lifespan was more pronounced at 20°C where signaling through the IIS pathway is further reduced. Therefore, under low IIS conditions, DAF-3 as well as DAF-5 can modulate longevity.
We next tested the role of DAF-3 and DAF-5 on fat storage, dauer formation and stress resistance. Oil Red O staining for fat storage showed comparable levels between daf-2(e1370) and daf-2(e1370); daf-3(mgDf90) worms, but markedly lesser amounts of fat in daf-5(e1386); daf-2(e1370) worms (Figure 4B top and bottom panel and Figure S12). Similarly, age-1(hx546); daf-5(e1385) had less fat than age-1(hx546) worms (Figure S12). Both daf-3 and daf-5 single mutants have slightly reduced levels of fat when compared to wild-type worms (Figure S12).
A similar trend was seen with our data for dauer formation. daf-2(e1370); daf-3(mgDf90) worms show significant enhancement of daf-2(e1370) dauer formation across several temperatures tested, whereas a daf-5 mutation or daf-5 RNAi results in reduced daf-2(e1370) dauer formation (Figure 4Ci, Figure 4Cii and Figure S11). In addition, daf-5(e1386); daf-2(e1370) worms fail to completely arrest at the restrictive temperature of 25°C (data not shown). A mutation in daf-5 also significantly reduces thermotolerance of daf-2(e1370) worms at 37°C (Figure S13). Taken together, similar to PDP-1, DAF-3 and DAF-5 modulate multiple outputs of the IIS pathway. Unexpectedly, we find that while DAF-3 promotes dauer formation under conditions of reduced TGF-β signaling, it negatively regulates dauer formation and longevity under conditions of reduced IIS.
To further explore the crosstalk between both pathways, we next asked whether DAF-18 and DAF-16, which are components of the IIS pathway, affect TGF-β pathway signaling. For this, we assayed dauer formation of TGF-β pathway mutants on daf-18 and daf-16 RNAi (Table 3 and Figure S10). Interestingly, dauer formation of daf-7(e1372), daf-14(m77) and daf-8(m8 5) worms was robustly suppressed by daf-16 RNAi. We observed similar results for dauer formation daf-7(e1372) and daf-14(m77) mutants on daf-18 RNAi. However, in the case of daf-8(m85) mutants, daf-18 RNAi had no effect on dauer formation of (Figure S10), suggesting a complex crosstalk between both pathways. The enhanced dauer formation of daf-2(e1370); daf-3(mgDf90) is suppressed by both daf-18 and daf-16 RNAi but not pdp-1 RNAi (Table 3 and Figure S10). Therefore, we not only observe DAF-3 and DAF-5 affecting various phenotypes of the IIS pathway, but also the converse, where DAF-16 and DAF-18 robustly regulates TGF-β dauer formation. These results unravel a more complex interaction between the two pathways, where DAF-16 is likely to be the major downstream effector regulating longevity, dauer formation and other physiological outputs.
How can these two pathways, once considered to be parallel to each other, be mechanistically linked? Thus far our data suggests that PDP-1, a component of the TGF-β pathway can modulate multiple phenotypes of IIS by positively regulating DAF-16. In addition, we observe extensive crosstalk between the two pathways at multiple levels. A feed-forward model that has been proposed to connect TGF-β signaling to the IIS pathway suggests insulins as a possible link [55], [58]. The C. elegans genome encodes 40 insulin genes [59], [60] (WormBase 215: www.wormbase.org). Studies using mutants and RNAi have characterized some of the insulins as agonists or antagonists of the IIS pathway [13], [59]–[61]. Importantly, microarray studies have identified several insulin genes that are regulated by TGF-β signaling, including ins-1, ins-4, ins-5, ins-6, ins-7, ins-17, ins-18, ins-30, ins-33, ins-35 and daf-28 [55], [57]. We tested changes in the levels of these insulins using Q-PCR in TGF-β pathway mutants such as daf-3(mgDf90), daf-14(m77) as well as pdp-1(tm3734) and compared them to wild-type worms (Figure 5A–5C, Figure S14, Tables S2 and S3). Interestingly, both pdp-1(tm3734) and daf-3(mgDf90) showed elevated levels of several insulins as compared to wild-type worms (Figure 5A and Figure S14). In contrast, expression of these insulins was markedly reduced in daf-14(m77) mutants (Figure 5B and Figure S14). We next looked at the effects of overexpressing DAF-3 and PDP-1 on insulin gene expression (Figure 5C and Figure S14). The levels of several insulins are markedly reduced in daf-3::gfp and pdp-1::gfp animals when compared to wild-type worms. Therefore, dosage modulation of DAF-3 and PDP-1 modulates insulin gene expression. INS-4, for example, has been reported as a positive regulator TGF-β pathway and a suppressor of dauer formation of daf-7 and daf-8 mutants [62]. ins-4 transcript levels were elevated in pdp-1 and daf-3 mutants but reduced in daf-14.
To investigate insulin gene expression regulated by DAF-16, we tested daf-2(e1370), pdp-1(tm3734); daf-2(e1370) and daf-16(mgDf50); daf-2(e1370) mutants. Several insulins were changed relative to daf-2(e1370) worms, with the trend between pdp-1(tm3734); daf-2(e1370) and daf-16(mgDf50); daf-2(e1370) being quite similar (Figure 5D and Figure S14). Interestingly, ins-7 levels were elevated both double mutants (Figure 5E and Figure S14). Previous studies have shown ins-7 to be an agonist of the IIS pathway as well as a DAF-16 target gene [13], [63]. In contrast, ins-1 levels were drastically reduced, and INS-1 has been characterized as a potential antagonist of IIS [59]. We did not observe a significant change in ins-18, another potential DAF-16 target [13]. We also did not detect any appreciable differences in insulin gene expression in daf-16(mgDf50) single mutants (Figure S14). In addition, we were unable to detect ins-33 and ins-35 transcripts in all the strains tested, and the trend observed with daf-28 was inconclusive (Table S2 and S3). Taken together, our results suggest the possibility that insulins downstream of TGF-β signaling mediate at least part of the cross talk between the two pathways. Therefore, PDP-1 would modulate to regulate expression of several insulins that can potentially feed into or antagonize the IIS pathway to regulate DAF-16 and its associated phenotypes.
We identified pdp-1 from a RNAi screen for serine/threonine phosphatases that modulate daf-2 dauer formation. C. elegans PDP-1 is homologous to mammalian pyruvate dehydrogenase phophatase (PDP), a metabolic enzyme that is a positive regulator of the pyruvate dehydrogenase enzyme complex (PDHc). Remarkably, other components of the PDHc in C. elegans do not affect daf-2 dauer formation. Microarray and SAGE studies on dauers have indicated that genes involved in anaerobic metabolism are upregulated while genes involved in the TCA cycle and mitochondrial oxidative phosphorylation are downregulated, suggesting that PDHc activity may not be critical for dauer diapause [64]–[66]. Further, annotations indicate that the C. elegans genome encodes approximately 60 serine/threonine phosphatases, in contrast to the 400 plus protein kinases, suggesting that phosphatases are likely to have a number of cellular substrates [39], [67]. We find that PDP-1 also regulates longevity, fat storage and stress resistance in addition to dauer formation. Interestingly, these phenotypes are more severe in mutants such as daf-2 and age-1, where IIS is reduced. Further, PDP-1 positively regulates DAF-16 activity. We reason that PDP-1 function is critical under conditions of stress or low food availability, when DAF-16 activation is required [39].
Intriguingly, genetic epistasis analyses place PDP-1 in the DAF-7/TGF-β pathway, at the level of the R-SMAD proteins DAF-14 and DAF-8. A recent functional RNAi screen for serine/threonine phosphatases that modulate BMP signaling identified PDP as a SMAD1 phosphatase in Drosophila S2 cells and mammalian 293T cells [68]. Our study complements these findings and reveals a molecular conservation in the role of PDP-1 in regulating TGF-β signaling. Early genetic epistasis studies had suggested that TGF-β signaling and IIS pathways are parallel signaling pathways that modulate dauer diapause [31]. Importantly, in these studies, the conclusion was that both these pathways acted independently, and it was the IIS pathway that regulated longevity and stress resistance [31], [32].
However, the effect of PDP-1 on DAF-16 activity led us to re-investigate the interaction between the IIS and TGF-β signaling. Previous studies have shown that DAF-3 and DAF-5 are negatively regulated by TGF-β signaling, and function similarly as repressors of gene expression to ultimately promote dauer formation [28],[29],[69],[70]. We find that under conditions of reduced IIS, DAF-3 and DAF-5 affect various outputs of the IIS pathway in opposite ways. DAF-3 in particular regulates IIS depending upon the level of signaling through the pathway (Figure 6). In our hands, mutants of the TGF-β signaling pathway do not exhibit a pronounced increase in lifespan. However, components of this pathway are important for the long lifespan of mutants in the IIS pathway, as well as other phenotypes such as dauer formation, fat storage and stress resistance. Our epistasis studies reveal that daf-18 and daf-16 RNAi can strongly suppress dauer and fat storage of TGF-β pathway mutants. Together, these results point to a feed-forward model where signals through the TGF-β pathway are relayed to modulate activity of the IIS pathway as well as DAF-16. Indeed, recent studies have suggested that TGF-β pathway regulates the expression of insulins, leading to a feed-forward model, where signals from the TGF-β pathway are relayed to modulate activity of the IIS pathway as well as DAF-16 [55], [58].
In support of this model, we find TGF-β signaling regulates the expression of several insulin genes with DAF-3 and PDP-1 negatively modulating insulin gene expression. This is in agreement with previous studies that identify DAF-3 as a repressor of gene expression [69], [70]. The expression of several insulins is also modulated by DAF-16, with pdp-1(tm3734); daf-2(e1370) and daf-16(mgDf50); daf-2(e1370) worms showing similar trends in insulin levels. Therefore, in the absence of PDP-1, increased levels of agonists or reduced levels of antagonists hyperactivate the DAF-2 pathway to negatively regulate DAF-16, thereby affecting the enhanced lifespan, stress resistance, dauer formation and fat storage of daf-2 mutants.
Our results suggest a model where under favorable growth conditions, signals through the TGF-β pathway activate the SMAD transcriptional complex to regulate the expression of insulins that activate the IIS pathway to phosphorylate and inhibit DAF-16 activity, thereby promoting growth, reproduction and normal lifespan (Figure 6, top panel). However, when food is limiting or under harsh survival conditions, TGF-β signaling is downregulated by PDP-1 to activate DAF-3 and DAF-5, to regulate the repression of insulin genes that may feed into the IIS pathway (Figure 6, middle panel). DAF-3 has also been reported to negatively regulate daf-7 and daf-8 gene expression in a feedback loop [24]. We find that pdp-1 expression is elevated in daf-3(mgDf90) mutants, suggesting a similar feedback regulation (Figure S15). Repression of TGF-β and insulin gene expression by DAF-3 results in a reduction in signaling through the IIS pathway, and promotes DAF-16 nuclear localization. DAF-16 then regulates the transcription of hundreds of target genes that ultimately modulate longevity, stress resistance, dauer formation and fat storage. Under low TGF-β signaling and IIS conditions, DAF-3 and DAF-5 regulate these outputs in an opposite manner, with DAF-5 synergizing and DAF-3 antagonizing DAF-16 function (Figure 6 lower panel). With our Q-PCR data, we found that PDP-1 affected only a subset of the DAF-16 target genes tested. These could represent genes that are regulated by DAF-16 and SMAD proteins. SMAD proteins have low affinity for binding DNA, and the orchestration of cellular signals into defined outputs requires their association with additional co-factors [71]. Mammalian SMAD proteins can bind several co-activators and co-repressor proteins to modulate gene transcription [23]. Specifically, a synergy between mammalian FOXO (FOXO1, FOXO3a and FOXO4) and SMAD2/3 was identified for the regulation of several genes involved in cell cycle regulation and the response to stress [72]. Importantly, these interactions required the function of the co-SMAD protein SMAD-4, which is homologous to DAF-3 [72]. Therefore, DAF-3 and DAF-5 could also directly modulate the IIS pathway at the transcriptional level.
A clear interpretation of our results is complicated by three main factors. First, the sheer number of insulins in the worm makes it difficult to assess whether they are functionally distinct. Secondly, the role of temperature in modulating the readouts of the pathway has not been closely explored. For example, we observe the effects of pdp-1 RNAi on daf-2 lifespan at 15°C but the effect decreases at a higher temperature, as the pathway gets more inactive. It is therefore likely that a certain level of signaling through the pathway is required to activate and target PDP-1 to its substrate(s). At higher temperatures such as 20°C or 25°C, there may be extremely low levels of phosphorylated substrate available for PDP-1. Similarly, the effect of a daf-3 null mutation on daf-2 phenotypes is more pronounced at higher temperatures but not at 15°C. Third, the lack of null alleles may provide an incomplete picture of the phenotypes observed. For example, previous studies using non-null alleles of daf-16 only partially suppressed dauer formation of TGF-β pathway mutants and therefore DAF-16 was thought to only affect the IIS pathway [31]. Therefore, temperature, level of signaling and the kind of mutants used (null versus weak) are important additional inputs that need to be considered to better understand the crosstalk between the IIS and the TGF-β pathways.
In conclusion, our studies show that PDP-1 acts through the TGF-β pathway to negatively regulate IIS and promote DAF-16 activity. PDP-1 may mediate this function in part by negatively regulating TGF-β signaling to repress expression of several insulins that feed into the IIS pathway. In humans, dysregulation of TGF-β signaling and the insulin/IGF-1 signaling axis have been implicated in the onset of age-associated diseases such as Type 2 Diabetes and cancer [73]–[77]. Future studies exploring the interactions between these two pathways as well as the factors that modulate these interactions may ultimately provide a better understanding of the pathophysiology of these diseases.
All strains were maintained at 15°C using standard C. elegans techniques [78]. For all RNAi assays, worms were maintained on the RNAi bacteria for two generations except for the assays on the PDHc RNAi. Strains used in this manuscript are listed in Table S4.
RNAi plates were prepared as previously described [39]. All RNAi clones were sequenced and verified before any assays were carried out. L4 worms were picked onto fresh RNAi plates and maintained for two generations prior to the assay, with the exception PDHc RNAi plates. Worms exhibit lethality when maintained on the following RNAi clones: T05H10.6 (E1α), C04C3.3 (E1β), F23B12.5 (E2), or LLC1.3 (E3) [79]. To circumvent this problem, strains were maintained on vector RNAi for two generations and transferred to E1α, E1β, E2 or E3 plates prior to the assay.
For the pdp-1(tm3734);daf-2(e1370) double mutant, daf-2(e1370) males were mated to pdp-1(tm3734) hermaphrodites at 15°C. A total of 30 F1 progeny were picked onto individual plates and allowed to have progeny at 25°C. From the F2 progeny on each plate, dauers were selected and transferred to fresh plates and incubated for an additional 24 hours at 25°C. The next day, the dauers were allowed to recover at 15°C until they reached adulthood. Subsequently, adult worms were picked onto individual plates and transferred to 25°C and allowed to have progeny. Among the F3 progeny, we observed that some plates had 100% dauers at 25°C, while worms in some of the plates exhibited a developmental delay and could not form complete dauers even after 5–6 days at 25°C. Worms from both sets of plates were recovered, picked to individual plates and allowed to self at 15°C. Parents were then tested for pdp-1(tm3734) deletion by PCR. As anticipated, the pdp-1(tm3734);daf-2(e1370) double mutants are unable to form 100% dauers at 25°C.
The daf-2(e1370);pdp-1::gfp strain was made by crossing daf-2(e1370) males to pdp-1::gfp hermaphrodites at 15°C. About 30 F1 animals were transferred to individual plates and allowed to have progeny at 25°C. From the progeny, F2 dauers were selected from each plate and allowed to recover at 15°C. The recovered adult worms were then checked for the presence of GFP, and GFP-positive worms were transferred to individual plates and incubated at 25°C. Plates where 100% of the progeny were dauers and GFP positive were selected and established as the strain for the assays.
Strains were maintained on RNAi plates for two generations or regular OP50 plates at 15°C. Dauer assays were performed by picking approximately 100 eggs onto 2 fresh plates and incubated at the appropriate temperature. The pdk-1(sa680), daf-7(e1372) and daf-14(m77) worms have a strong Egl phenotype. For dauer assays on these strains, gravid adult worms growing on the RNAi plates were washed off the plate with sterile PBS onto a 1.5 mL eppendorf tube. After 2 washes at 2000 g for 30 seconds, the adults were vortexed for 5 mins in 5 mL of 1 N sodium hydroxide and 3% sodium hypochlorite (final concentration). The samples were then washed twice with sterile PBS and eggs were aspirated with a glass pipette onto fresh RNAi plates. For all dauer assays, plates were scored for the presence of dauers or non-dauers after 3.5–5.5 days, depending upon the strain. Dauer assays were performed at the temperature indicated. Significance was determined by Student's t-test.
Strains were maintained at 15°C and synchronized by picking eggs onto fresh RNAi or OP50 plates. Approximately 60 young adult worms were transferred per plate to a total of three fresh RNAi or regular OP-50 plates containing 5-fluorodeoxyuridine (FUDR) at final concentration of 0.1 mg/mL [80]. All RNAi-based lifespan assays were carried out at 15°C. Lifespans on OP50 plates were performed at the temperature indicated. Survival was scored by tapping with a platinum wire every 2–3 days. Worms that died from vulval bursting were censored from the analysis. Statistical analyses for survival were conducted using the standard chi-squared-based log rank test.
Strains were maintained on RNAi or regular OP50 bacteria at 15°C, as described above. From these plates, approximately 30 young adult worms were picked onto fresh RNAi or regular plates and upshifted to 20°C for 6 hrs. The plates were then transferred to 37°C and heat stress-induced mortality was determined every few hours till all the animals died. Statistical analyses for survival were conducted using the standard chi-squared-based log rank test.
Strains maintained RNAi or on regular OP50 plates were synchronized by picking eggs on to fresh plates and grown synchronously at 15°C. The plates were then upshifted to 20°C for 8 hours, at the L2 stage to get L3 worms and at the L4 stage to get young adult worms. Worms were then washed off the plates into microcentrifuge tubes and incubated in 1× PBS buffer for 20 minutes on a shaker at RT. After 2 washes at 3000 rpm for 30 seconds with 1× PBS, the strains were fixed according to the type of staining performed. Oil Red O and Sudan black staining was performed as previously described [39], [45], [81], [82]. After incubation overnight at RT, worms were mounted on slides and visualized using the Zeiss Axioscope 2+ microscope.
For Sudan Black Staining, we used Image J software to measure the average pixel intensity for a 84-pixel radius below the pharynx of each animal in the anterior intestine area. Next, an 84-pixel radius of the background was measured, and subtracted from the values obtained for the staining. At least 10 animals were measured for each RNAi clone. Significance was determined by Student's t-test.
For Oil Red O Staining, Image J was used to separate out each color image into its RGB channel components. As previously described [45], Oil Red O absorbs light at 510 nm and therefore, the green channel was used for further analysis. We measured the average pixel intensity for a 84-pixel radius below the pharynx of each animal in the anterior pharynx area. We next measured a 84-pixel radius of the background, which was later subtracted from the values obtained from the staining. At least 10 animals was measured for each RNAi clone. Significance was determined by Student's t-test.
DAF-16 localization assays were performed as previously described [39], [52]. daf-2(e1370); daf-16::gfp worms were maintained on RNAi plates at 15°C similar to the dauer assays. Approximately 30 L4 worms were transferred to fresh RNAi bacteria and the plates were shifted to 20°C for 1 hr. The worms were visualized under a fluorescence microscope (Zeiss Axioscope 2+ microscope). Worms were classified into four categories based on the extent of DAF-16::GFP nuclear-cytoplasmic distribution: completely cytosolic, more cytosolic than nuclear in most tissues, more nuclear than cytosolic in most tissues and completely nuclear.
Quantification of Psod-3::gfp was performed as previously described [39]. daf-2(e1370);sod-3::gfp worms were grown at 15°C on RNAi as described above. Approximately 30 L4 animals were transferred to fresh RNAi bacteria and shifted to 25°C for 1 hr. The expression of sod-3::gfp was visualized using Zeiss Axioscope 2+ microscope. GFP expression was categorized as follows:
High: GFP expression seen throughout the worm
Medium: Weak expression detected in the body of the worm along with the head and the tail
Low: Low GFP expression only detected in the head and tail
Promoter and ORF entry clones of pdp-1 obtained from the promoterome and ORFeome were combined using multisite Gateway cloning (Invitrogen) into the pDEST-DD03 or the R4-R2 GFP destination vectors to create the Ppdp-1::gfp or Ppdp-1::pdp-1ORF::gfp constructs [83], [84]. All constructs contain the unc-119 minigene. The vectors were verified by sequencing as well as restriction digestion. Transgenic worms were generated by ballistic transformation into unc-119(ed3) mutant worms as previously reported (Biorad, USA) [83]. Integrated lines that were obtained were used for further analyses. For the pdp-1::gfp translational fusion strain, additional lines were generated by integration of extrachromosomal array lines by UV irradiation as previously described [85]. All translational fusion lines were backcrossed 4× to wild-type prior to analysis.
For all RT-PCR experiments, strains were maintained at 15°C. Eggs were obtained from gravid adult worms by hypochlorite treatment described earlier. The eggs were seeded onto large plates maintained at 15°C until the worms entered the L4 stage. The plates were then upshifted to 20°C for 8 hours until they became young adults. Worms were then collected with sterile 1×PBS and washed twice at 2000 g for 30 seconds. The supernatant was removed, and 0.5 mL of AE buffer (50 mM acetic acid, 10 mM EDTA), 0.1 mL of 10% SDS, and 0.5 mL of phenol was added to the worm pellet and the mixture was vortexed vigorously for 1 min, followed by incubation at 65°C for 4 min. Total RNA was purified by phenol:chloroform extraction and ethanol precipitation. The quality of the RNA isolated was determined by checking the 28 S and 18 S RNA on an agarose gel. 2 ug of total RNA was used for making cDNA using the SuperScript cDNA synthesis kit (Invitrogen, USA). The expression of the DAF-16 target and insulin genes was checked by RT-PCR using the SYBR Green PCR Master Mix and 7000 Real-Time PCR System (Applied Biosystems, USA). The relative expression of the genes tested was compared to actin as an internal loading control. Significance was determined by Student's t-test. Primers used for the RT-PCR experiments are listed in Table S5.
Young adult wild-type and pdp-1(tm3734) worms were picked onto 6 individual plates each. After 5 minutes, the worms were picked off the plate. The average distance covered was calculated by measuring the traces on the bacterial lawn using ImageJ. Significance was determined by Student's t-test.
Wild type, daf-2(e1370), pdp-1(tm3734) and pdp-1(tm3734); daf-2(e1370) worms were maintained at 15°C. 5 L4 worms were picked onto individual plates and allowed to lay eggs at 22.5°C. Worms were transferred to a new plate every 12 hours. After 22.5 hours, the parental worms were picked off the plates, and the total number of eggs laid was scored. The number of progeny from these eggs was scored again after 38 hours. The % hatched eggs was calculated as a percentage of the average number of progeny over the average number of eggs laid. Significance was determined by Student's t-test.
Statistical analyses were performed using JMP and Microsoft Excel. NIH Image J was used for quantification of locomotion and fat storage.
|
10.1371/journal.pcbi.1006098 | Topological and statistical analyses of gene regulatory networks reveal unifying yet quantitatively different emergent properties | Understanding complexity in physical, biological, social and information systems is predicated on describing interactions amongst different components. Advances in genomics are facilitating the high-throughput identification of molecular interactions, and graphs are emerging as indispensable tools in explaining how the connections in the network drive organismal phenotypic plasticity. Here, we describe the architectural organization and associated emergent topological properties of gene regulatory networks (GRNs) that describe protein-DNA interactions (PDIs) in several model eukaryotes. By analyzing GRN connectivity, our results show that the anticipated scale-free network architectures are characterized by organism-specific power law scaling exponents. These exponents are independent of the fraction of the GRN experimentally sampled, enabling prediction of properties of the complete GRN for an organism. We further demonstrate that the exponents describe inequalities in transcription factor (TF)-target gene recognition across GRNs. These observations have the important biological implication that they predict the existence of an intrinsic organism-specific trans and/or cis regulatory landscape that constrains GRN topologies. Consequently, architectural GRN organization drives not only phenotypic plasticity within a species, but is also likely implicated in species-specific phenotype.
| The translation of genotype to phenotype is a tightly regulated process that is mediated by specific interactions between a variety of cellular components. Central to this is the transcription of genes, a process regulated by proteins that bind DNA, including the transcription factors (TFs). Gene regulatory networks (GRNs) describe the web of protein-DNA interactions essential for the regulation of biological pathways and developmental processes. Here, we describe fundamental properties of the architectural organization of eukaryotic GRNs. Using protein-DNA interaction data derived from the budding yeast, the fruit fly, Caenorhabditis elegans, and Arabidopsis, we determine that GRNs are scale-free, wherein a majority of TFs bind comparatively fewer target genes, while a small number of TFs bind a large number of genes. Further, we show that the scale-free connectivity power-law coefficient is organism-specific, suggesting differential wiring patterns across GRNs. We then capitalize on the organism-specific connectivity to develop a mathematical framework to predict the number of total interactions in the complete GRNs, important for understanding how the expression of all genes in an organism is regulated, and for experimental design purposes. Finally, we demonstrate numerically and by simulations that subnetworks sampled from scale-free GRNs are scale-free, as long as edges are sampled. This finding has important real-life implications to infer properties of a network with limited experimental data.
| Complex systems are formed by large numbers of components organized into networks, and modelled by graphs in which nodes are connected by edges. Network architecture is established by topological and statistical analyses, ultimately leading to inference of functional roles played by the nodes in the network, and prescribed by the observed architecture. Efforts to infer information flow, which ultimately leads to functional outputs, have been applied to different types of networks, including social communication, electrical power [1], and biological [2–10]. A general characteristic of many real-world networks is their scale-free topologies, which exhibit a node degree distribution that can be described with a power law function:
P(k)=Ck−α
(1)
where P(k) is the probability of a randomly selected node having degree k (that is k connections), and α is the power law scaling exponent (hereafter referred to as the exponent). The constant C is a Riemann’s zeta function that normalizes the power law probability distribution, such that:
∑k=1∞P(k)=1
(2)
In scale-free networks, most nodes have comparatively few interactions manifested as a lower degree, while a small number of nodes, the ‘hubs’, have a higher degree [11, 12]. This scale-free connectivity distribution is observed at different levels of biological organization ranging from the cellular and molecular, to the ecological level. Gene regulatory networks (GRNs), characterized by the interaction of a specific type of proteins, the transcription factors (TFs) with the regulatory DNA regions in the genes that the TFs control, provide excellent examples of molecular-level scale-free networks [2, 6, 10, 13–15]. GRNs can be represented by directed graphs in which the edges have a polarity, because a TF can bind to the regulatory region of a gene (which may encode for another TF) and modulate its expression, but not vice versa. Thus, GRNs can be visualized from the perspective of incoming connectivity (i.e., how many TFs bind to a specific gene regulatory region), or from the outgoing connectivity perspective (i.e., how many regulatory regions does a TF recognize). The molecular tools available to identify incoming and outgoing connectivity are different. Incoming connectivity is usually mapped using gene-centered approaches such as yeast one-hybrid (Y1H) assays [16], and outgoing connectivity is evaluated by TF-centered approaches such as chromatin immunoprecipitation (ChIP)-based (e.g., ChIP-Seq and ChIP-chip) [17] or DNA affinity purification sequencing (DAP-Seq) methods [18]. While the integration of results derived from gene- and TF-centered procedures should ultimately converge into the same GRN, much of the unbiased data available today derives from TF-centered approaches, providing a much clearer perspective of outgoing connectivity. We anticipate that the advent of new experimental approaches to map PDIs and place them in a biological context will permit to explore the convergence of incoming and outgoing connectivity in many organisms.
Organismal phenotypic plasticity is driven in part by the underlying GRNs [19–21]. Therefore, the reconstruction and topological analysis of GRNs provides an excellent opportunity for elucidating molecular mechanisms that drive phenotypic plasticity. However, despite significant research in this area, little is known with regards to whether GRNs from different organisms have similar emerging properties that only depend on node number, or whether properties such as network connectivity, manifested for example in the exponent of the power law, are unique to each organism. Experimentally, most studies will be able to provide at best an observed network, which corresponds to a subset of the complete true network (Fig 1). It is unclear to what extent properties of the observed network can be used to infer properties of the complete network (Fig 1). Conversely, several studies have investigated the properties of subnetworks, starting from synthetic or natural networks. The conclusions derived from these studies depend on the sampling method used [22–24]. For example, it was argued that randomly selected subnets of scale-free networks are not scale-free themselves, and that therefore inferences about the complete network had to be treated with caution [25]. However, the node sampling methodology used in that study results in a loss of degrees because, by targeting nodes rather than edges, all the edges associated with a node are lost, resulting in the enhanced decrease in degrees. In addition, that study did not model the stochasticity inherent in the sampling process, thereby not capturing the possible range of degree exponents that a subnetwork can take. As described here, sampling edges while accounting for stochasticity gives a very different result.
In our study, we take four representative model organisms that represent major eukaryotic evolutionary groups (the yeast Saccharomyces cerevisiae, the worm Caenorhabditis elegans, the fruit fly Drosophila melanogaster and the flowering plant Arabidopsis thaliana) and for which a wealth of PDI data is publicly available, to reconstruct GRNs followed by network connectivity analysis. Following simulations and rigorous statistical analyses, we demonstrate that GRNs exhibit organism-specific scale-free connectivity, revealed by distinct exponents of the out-degree. Further, we show that the observed coefficients are unbiased estimates of exponents derived from the degree distribution of the inferred complete GRNs. As a result, we apply a Monte-Carlo simulation approach for the estimation of the number of PDIs in complete GRNs. To provide an interpretation of the out-degree exponent, we employ ‘inequality’ analyses using Lorenz curves. We show that the exponents describe the relationship between the proportions of TFs binding to the corresponding fraction of the target genes. The resulting GRN topologies can therefore be classified as either ‘capitalistic’, exemplified by the presence of a handful of hub TFs that bind a significant and disproportionate number of target genes, or ‘socialistic’, in which TFs bind a near corresponding proportion of targets. Collectively, these observations demonstrate the utility of the observed GRNs in predicting properties of complete GRNs, with important implications for understanding the complex regulatory repertoire of eukaryotic organisms.
We constructed GRNs using all available experimentally determined PDIs derived from ChIP-Seq, ChIP-chip, and yeast one-hybrid assays (Table 1). To determine the connectivity of these observed GRN, we enumerated the target genes bound by each TF (out-degree, Fig 2), and the number of TFs binding each target gene (in-degree, S1 Fig). We observed that, in all four organisms investigated, a majority of TFs bind comparatively to few target genes (low degree TFs), while a small number of TFs bind to a large proportion of target genes (high degree TFs). A linear relationship of the probability density function on a log-log scale was observed (Fig 2, inset), indicative of the scale-free property of the interaction distribution. To unequivocally confirm the scale-free properties of the resulting observed GRNs, we implemented a formal statistical analysis framework consisting of the following steps: (i) Fitting node-degree distribution to a power law function and estimating the power law function exponent parameter (α) using the maximum likelihood approach; (ii) testing goodness-of-fit by comparing the fitted power law distribution and the empirical node degree distribution using the Kolmogorov-Smirnov (KS) D statistic; and (iii) performing pairwise model selection by comparing the fitted power law distributions to Poisson and exponential functions. A non-nested model selection approach that uses the Kullback-Leibler information criterion (Vuong’s closeness test) was employed for the pairwise model comparisons (see Methods for details). We observed a significant fit of power law functions on the out-degree distribution (Table 2), thereby confirming the scale-free nature of the observed GRNs. Further, model selection likelihood ratio tests comparing fitted power law and Poisson distribution functions (a descriptor of non-scale-free random networks) revealed that fitted power law distributions are significantly favored (Table 2). As anticipated, given the biased nature and insufficient sampling of most gene-centered PDI determination studies, in-degree distribution of the available experimental data could not be described by a power law function. Unexpectedly, the out-degree power law exponents were different for the observed GRNs obtained from the four organisms, with values of 4.12 for C. elegans, 3.04 for the fruitfly, 2.0 for yeast and 1.73 for Arabidopsis (Table 2, first row).
To determine the difference between the empirical distributions of out-degrees for pairs of observed GRNs, the two-sample Kolmogorov-Smirnov (KS) test was employed, with the null hypothesis testing whether two samples have been drawn from the same distribution. We observed that pairs of out-degrees between organisms have distinct distributions, with the exception of A. thaliana—S. cerevisiae and D. melanogaster—S. cerevisiae comparisons (Table D in S1 Information).
To investigate the possible biological consequence of the different out-degrees in the scale-free topology of GRNs for the four organisms, we investigated how ‘inequality’ in TF-target gene binding distributions is affected by the power law degree exponent in the different GRNs, using Lorenz curves [26]. We ranked TFs based on increasing number of target genes and plotted the cumulative proportion of target genes as a function of the corresponding cumulative proportion of TFs. Interestingly, we observed an increase in degree ‘equality’ for each increase in the value of the exponent (Fig 3). This contrasts with a perfectly egalitarian distribution of degrees where all TFs have approximately the same degree, and for which the associated Lorenz curve becomes the diagonal of the plot, referred to as the line of equality. Thus, GRNs with smaller exponents, such as the S. cerevisiae GRN, have hub TFs that bind disproportionately more target genes, compared to the same number of hub TFs in GRNs with higher exponents (Fig 3). Indeed, we observed that the top 20% of TFs with the highest number of target genes in S. cerevisiae bind about 50% of the target genes. In C. elegans, a similar proportion (the top 20%) of TFs binds to 30% of target genes. In an egalitarian binding scenario, 20% of the top TFs would bind 20% of the target genes. Note that our analyses here and henceforth did not include Arabidopsis out-degrees due to the low number of target genes (37% of all coding genes) represented in the observed GRN (Table 1), corresponding to insufficient sampling of out-degrees.
Next, we investigated potential biases in the estimation of the exponents that might be driven by data sources. In this regard, we estimated exponents from subnetworks derived from specific experimental data source (ChIP-Seq, ChIP-chip or Y1H) and tissues type, where data is available. Investigating the influence of experimental data source in D. melanogaster GRN revealed similar exponents for the data source-specific subnetworks (Table A in S1 Information). For instance, with the exception of a slight increment of 0.32 for the ChIP-Seq-derived subnetwork, the exponents of ChIP-derived subnetworks are similar (rows 2, 3, and 4 of Table A). In the case of Y1H, it is evident that gene-centered approaches employed in building GRNs can inadvertently introduce bias due to insufficient sampling of out-degrees. Y1H contributed only 406 (~ 0.18% of total GRN) interactions to the D. melanogaster GRN and 136 TFs, indicating that on average one Y1H-derived TF binds 2 target genes (208/136)–an unlikely in vivo phenomenon. Indeed, fewer out degrees are sampled in Y1H because the technique depends on cloning promoters of target genes. Difficulties in cloning promoters, as well as the comparatively higher numbers of targets in a genome (unlike TFs), results in an underrepresentation of out degrees in Y1H-derived GRNs. TF-based techniques (ChIP-Seq and ChIP-chip) such are more attractive in construction of GRNs that capture the expected connectivity because of the near complete sampling of out degrees (targets). TF-centered approaches are however not immune from potential bias, primarily the inclusion of false-positives targets. The ChIP-Seq and ChIP-chip analysis pipelines attempt to account for the false positive rates by determining the false discovery rates (FDRs) in cases where biological replicates exist. It’s important to note that in our analysis, inclusion of the Y1H data did not result in a deviation of the power law exponent (compare rows 2 and 3 in Table A of S1 Information). Minimal deviations were also observed in the data-specific subnetworks of C. elegans (Table B) and S. cerevisiae (Table C).
Another potential bias in estimation of exponents is tissue- (or developmental) specific sampling of targets. To address this, we sampled a subnetwork from the D. melanogaster GRN PDIs derived from the embryo stage. The choice for D. melanogaster was largely due to availability of tissue-specific data. The analyses resulted in a subnetwork with 47 TFs, 178,224 PDIs, and a total of 15,016 nodes. Fitting a power law function on the ‘embryonic’ out-degrees resulted in an exponent of 3.10 and KS P-value of 0.86 (row 7 of Table A). It is worth noting that these deviations fall within the 95% prediction intervals of the expected range of exponents for complete GRNs (see next subsection on inference of properties of complete GRNs).
Taken together, these findings demonstrate that, while GRNs are characterized by the unifying scale-free network property as opposed to random degree distribution, the GRN connectivity is quantitatively organism-specific, suggesting intrinsic organismal properties that define TF binding landscapes.
The ‘observed’ GRNs described in the prior section correspond to a fraction of the ‘complete’ GRNs that remain to be experimentally determined. A fundamental question that this study intends to address is to what extent can the observed GRNs be used to infer properties of complete GRNs (Fig 1). The answer serves two main purposes: first, one goal of systems biology is to describe all system components and their associated interactions. Since current GRNs are incomplete, but are samples from complete yet unobserved GRNs, there is need to determine whether properties of current GRNs sufficiently describe properties of the intended complete GRNs. Second, decisions on whether additional experiments need to be performed will be based on the ability of the current GRNs in describing properties of complete GRNs, an important consideration in experimental design. Therefore, to determine whether complete GRNs are scale-free with organism-specific degree scaling exponents, we evaluated the distribution of degrees of nodes sampled from large populations of simulated node degrees whose power law exponents are known. We describe the approach below.
We implemented a Monte-Carlo (MC) simulation approach to generate large populations of simulated nodes, each population exhibiting a distinct and known power law exponent of the node degrees (see Note A in S1 Information for a detailed description on sampling). Briefly stated, we first generated three sets of a large number of simulated nodes (n = 10,000), each with population degree power law exponent, αpop, corresponding to the three observed exponents of 4.12, 3.04 and 2.00. In the MC simulation, the number of computationally-generated nodes significantly exceeded the number of TFs in any organism in order to model a theoretically large population of nodes, a condition required for the central limit theorem (CLT) to be applicable (see Note A in S1 Information). Next, we randomly drew nodes from each population (with replacement) to generate samples (r = 1,000) of different sizes, followed by estimation of the scaling exponent of each sample using the maximum likelihood method. The distribution of exponents for large sample sizes (e.g., n = 5,000) followed normality with their average corresponding to the population exponent (S2A Fig). As anticipated, we observed a marked deviation from normality coupled with increased variance whenever smaller samples (n < 30) were drawn (S2B Fig). To predict the range of exponents for the complete GRNs, we calculated prediction intervals (PIs) using standard deviation (SD) of their distribution derived from the MC sampling procedure. We specifically used the MC-derived SDs corresponding to the number of TFs in the genome to construct 95% PIs (Note A in S1 Information). We observed 95% PIs falling in the {3.51–4.73}, {2.76–3.32}, and {1.87–2.13} intervals for the starting degrees of 4.12, 3.04 and 2.00, respectively. Notably, the PIs for exponents of complete GRNs do not overlap, thereby underscoring the organism-specific nature of power law scaling exponents in GRNs of the organisms investigated here.
From this analysis, we conclude that, a scale-free observed GRN with exponent αobs is likely derived from a complete true GRN, which is also scale free with exponent αobs ± c, where c is the upper and lower bounds of the 95% PI. In the following section, we capitalize on the predicted exponents of complete GRNs to estimate the size of their corresponding complete GRNs.
There is a pressing need to infer properties of complete GRNs in order to capture the system-wide regulatory landscape of a particular organism. However, experimental limitations (such as challenges in generating TF-specific antibodies for ChIP, limitations in genome sequence and annotation) and lab-specific research questions have resulted in incomplete and often fragmented GRNs, whose properties may fail to adequately capture the intended entire regulatory repertoire. To address this challenge, we undertook a simulation approach to estimate the expected number of PDIs in the complete GRNs. Our method is predicated on the finding that the observed out-degree exponent of a GRN is an unbiased estimate of the respective complete GRN exponent. As a consequence, out-degrees of an observed GRN can be described as a random sample from a population of degrees corresponding to the number of TFs in the genome. In the observed GRNs, the number of interactions (Iobs) is obtained by the summation of the out-degrees in the network, as follows:
Iobs=∑i=1nkobs,i
(3)
where n is the total number of TFs (number of out-degree nodes) in the observed GRN, and kobs,i is the ith observed out-degree value. We extend this framework to identify the number of interactions for a complete GRN (Icomp), and posit that:
Icomp=∑i=1Nkcomp,i
(4)
where N is the total number of TFs in the genome and kcomp,i is the ith out-degree value of the complete GRN (see Note B in S1 Information for a detailed description on derivation of simulated degrees). To test the feasibility and accuracy of the simulation approach, we estimated the actual number of interactions (Iobs) for the observed GRNs. We determined that, on average, the number of PDIs estimated by the method was equal to the number of PDIs of the observed GRNs (Table 3, column 3; refer to Note B in S1 Information for description on hypothesis testing). We subsequently used this method to predict the number of interactions of the complete GRNs (Icomp) based on the total number of TFs and genes that have been described in the organisms (Table 1). The maximum possible number of PDIs (upper bound) corresponds to the product of the total number of TFs and the total number of genes, as this would imply that every TF binds to every gene in the genome (Table 3, column 5). When we computed Icomp for the organisms we investigate here, we found that the budding yeast S. cerevisiae would have a total of ~60,000 PDIs, the fruitfly Drosophila has ~1.5M PDIs and the worm C. elegans has ~2M PDIs (Table 3, column 4). These estimates suggest that the number of observed PDIs represents ~45%, 14%, and 23% of the respective complete GRNs. When we compare the predicted PDI number of the complete GRNs with the maximum possible, we find that it is only 4% for yeast, 8% for Drosophila and 10% for C. elegans (Table 3, column 5), indicating that combined, TFs are sampling only a fraction of all the possible TF-target gene combinations. This observation contrasts the continuous network model that proposes in vivo binding of each TF to essentially all target genes in an organism.
Having demonstrated that complete GRNs are scale free, we set out to determine whether subnetworks of observed scale-free GRNs are equally scale-free. By sampling edges from observed GRNs, we mimic the experimental approach involved in constructing GRNs. Indeed, construction of GRNs largely involves identifying interactions (edges) between known cellular components (TFs and potential target genes). Below, we first show analytically followed by sampling, that subnetworks of scale-free GRNs are scale-free. When drawing edges from a GRN, the probability Pr(i) of a node i in the GRN becoming node i* in the subnetwork given that its edge has been randomly selected is dependent on node i degree, ki. This relationship is described by:
Pr(i*)=kikT
(5)
where kT denotes the total number of out-degrees in a GRN. To sample subnetworks of different sizes, edges are sampled with probabilities 0<p<1. Therefore, the probability of including node i* in the subnetwork when edges are sampled with a probability p is:
Pr(ip*)=p×kikT
(6)
When sampling subnetworks of specific sizes (e.g., half the observed GRN, where p = 0.5), p and kT are constants in Eq 6, which can be rewritten as:
Pr(ip*)=pkTki+ε
(7)
where ε is an error term accounting for the pseudorandom number generator (PRNG) since the PRNG algorithm is not strictly random but depends on an initial value, the seed. It is clear from Eq 7 that the probability of a node being included in the subnetwork, when its edges are sampled, is a linear function of the degree of the node being sampled. Thus, analytically, the degree distributions of a GRN and its associated subnetworks are similar.
For computationally validating the aforementioned analytical procedure while accounting for stochasticity, we randomly sampled subnetworks of varying sizes from the observed GRNs and from one synthetic complete GRN, followed by a determination of their respective out-degree exponents (note here that sampled subnetworks do not have random degree connectivity, but are rather randomly sampled from GRNs). We discovered that for the observed GRNs, a majority of subnetworks exhibited exponents similar to the exponent of their respective GRNs (Fig 4). However, there exists a subnetwork size below which there is an increased uncertainty in the determination of the exponents. This is evident in the marked increase and overlap in variation of the subnetwork exponents across organisms at lower subnetwork sizes (Fig 4). To sample from synthetic networks, we first constructed in silico networks that capture the expected connectivity of the complete yeast GRN as prescribed by the predicted exponents and number of PDIs from the previous sections (see Methods for procedure on creating in silico GRNs). Expectedly, the out-degree distribution of the synthetic GRN was strikingly similar to the observed GRN (Fig 5A). Fitting power law functions on the out-degrees of synthetic GRN resulted in exponents ranging from 1.98 to 2.14, capturing the exponent (2.0) of the observed GRN (Fig 5B). Further, the power law fit to the out-degrees of subnetworks drawn from synthetic GRN was significant, as indicated by the large KS test P values (Fig 5C). In sharp contrast with previous studies that investigated properties of subnetworks by sampling nodes, rather than edges as done here, a majority of exponents of subnetworks were similar to the exponent of the complete GRN (Fig 5D), a further indication of organism-specific GRN connectivity. In addition, maintenance of the network connectivity in randomly sampled subnetworks demonstrated an important network property that distinguishes random network from scale-free networks: robustness. However, there is a subnetwork size threshold below which the organism-specific connectivity deviates from the expected. Our analysis revealed that whenever less than 10% of the complete C. elegans GRN; or 2% of the yeast and Drosophila GRNs are sampled, the expected scale-free property no longer holds (Fig 4).
Collectively, these observations have an important implication that is likely general to other scale-free GRNs: whenever subnetworks are randomly sampled from scale-free complete and sufficiently large incomplete GRNs that are either experimentally-determined or synthetic, the sampled subnetworks are scale-free, at least for a given subnetwork size threshold.
The ultimate goal in the characterization of TF-target gene interactions is to describe the complete genome-wide regulatory repertoire of an organism. However, current GRNs are incomplete. Two related challenges have impeded the understanding of an organism's complete regulatory repertoire: Establishing the full range of PDIs that characterize complete GRN, and anticipating the properties of complete GRNs given properties of experimentally-determined but incomplete GRNs, We have addressed these challenges by a topological analysis of current GRNs across a diversity of organisms, and discovered that observed GRNs, and their respective complete GRNs, have organism-specific topologies. This finding has profound biological implications: while GRNs are largely scale-free, there exists an organism-specific GRN architecture that drives organism-specific developmental trajectories and phenotypic uniqueness. Taking advantage of conservation of network topology between observed and complete GRNs, we predicted the possible range of PDI numbers for complete GRNs. Indeed, the forecasted complete GRN PDI numbers are just a fraction of the maximum number of PDIs that result when each TF binds all target genes. This observation deviates from the previously proposed continuous network model, whose fundamental property is that TFs have the potential to bind all genes in an organism [27]. For broader applications, our simulation method employed in estimation of the expected number of PDIs can be applied to different types of biological networks such as protein-protein interaction, protein phosphorylation, metabolic interactions, and genetic interaction.
In contrast with previous work by Stumpf and colleagues [25], we demonstrate here that subnetworks sampled from scale-free networks are scale-free. Several differences exist between our approach and the one previously published [25]. First, sampling nodes leads to a loss of out-degrees resulting in a deviation between out-degrees of the observed network and sampled subnetworks. In addition, sampling nodes leads to the generation of singletons (nodes without edges or targets) in the subnetworks. In contrast, we sampled edges (with their associated nodes) from observed GRNs, thereby capturing both a TF and its potential targets. Indeed, this approach mimics the expected experimental sampling. Second, stochasticity inherent to sampling procedures was not previously accounted for [25]. Here, we account for stochasticity in sampling procedure by repeated sampling of subnetwork and estimation of the variance of the sampling distribution of subnetwork exponents at each subnetwork size. Third, estimation of power law exponents using the graphical method of ordinary least squares (OLS, see Fig 2 of reference [25]) might not be a robust approach for parameter estimation. The OLS method is based on the following assumptions: (a) regression errors are identically and independently distributed (iid) random variables with mean zero, and (b) the standard deviation of the error is independent of the independent variable (out-degrees). The OLS method is expected to perform poorly in the estimation of the power law exponents because these assumptions are not met in empirical data of power law distributions [28, 29]. In our analysis, we fit the data (out-degrees of observed and sampled subnetworks) to the power law function using the maximum likelihood estimation (MLE) method, since MLE has been shown to be asymptotically efficient and can be applied to a wide range of data with skewed distributions [28].
Our study also provides an interpretation of the organism-specific power law exponent by use of Lorenz curves: GRNs with higher values of exponents are ‘egalitarian’ in their TF-target gene binding. Simply put, GRN architectures can either be ‘capitalistic’, exhibited by a highly skewed TF-target gene binding landscape described by low exponents; or ‘socialistic’, described by high exponents. Just like skewed distributions of incomes of individuals describe less egalitarian capitalist societies, we envisage a more skewed TF-target gene binding landscape in GRNs with comparatively low exponents, wherein the number of target genes bound by a TF is analogous to an individual’s income. An increase in the exponent value denotes a decrease in skewness of TF binding. Taken together, findings reported herein provide opportunity to understand complex regulatory mechanisms from a genome-wide perspective, while paving way for construction and analyses of GRNs in non-model organisms whose complete regulatory repertoire is yet to be deciphered.
PDIs for C. elegans, D. melanogaster, S. cerevisiae and A. thaliana were extracted from regulatory databases and literature. In cases where regulatory interactions comprised of only DNA-binding sites (such as in ChIP-Seq, DAP-Seq, and ChIP-chip binding ‘peak’ location), the target genes associated with the binding sites were located within 2 kb of the ‘peak’ location. Transcriptional GRNs were subsequently modelled using directed graphs, Gn,v, with n nodes and v vertices (edges, PDIs). Nodes in GRNs represent both target genes and their associated protein products in cases where a target gene is a TF. A PDI is represented by a directed edge emanating from the TF and ending in the target gene.
Node degrees were determined by enumerating the number of TFs binding to each target gene (out-degree) and the number of target genes bound by each TF (in degree). A formal statistical framework that tests scale-free property in GRNs was developed involving the following steps:
Note that implementations of the methods presented above can be found in the R statistical packages ‘igraph’ and ‘poweRlaw’.
Sampling subnetworks involved randomly selecting a number of PDIs from the observed GRNs, followed by construction of the subnetworks from the sampled PDIs. The sizes of the subnetworks correspond to the proportion of PDIs sampled. One thousand subnetworks were sampled for each proportion. A degree distribution was determined for each sampled subnetwork. Below is the pseudocode implemented for sampling and fitting power law function on the sampled subnetworks, from each GRN:
For i in {0,..,1} // where i is a proportion (size) of the GRN
For j in {1,..,1000} // 1000 iterations
Select edges uniformly at random from edge-list of GRN
Construct subnet Gi,j*
Fit out-degree of subnet to Power law function
Estimate exponent αi,j of subnet out-degree
END For
END For
To sample from synthetically-generated out-degrees, the following pseudocode was implemented:
Generate 10,000 degrees that follow a specified exponent (Note B in S1 Information)
For i in in {0,..,1} where i is a proportion of the number of TFs (degrees) in the GRN of interest
For j in {1,…,1000} //1000 iterations
Select j * i random TFs from the population
Construct subnet Gi,j*
Fit Gi,j* out-degree to Power law function
Estimate exponent αi,j
END For
END For
Estimation of the expected number of PDIs in complete GRNs, and the power law exponent alpha for observed GRNs, enabled creation of in silico GRNs that recapitulate the expected complete GRNs. Complete GRNs were built using the edited ‘igraph’ function ‘static_power_law_pl’ which takes exponent and number of PDIs as inputs. In order to generate a biologically comparative network, the number of TFs in the function ‘static_power_law_pl’ was edited so that only 5% of genes can have out-going edges. The scale-free property, sampling of subnetworks, and the determination of the clustering coefficients of the in silico GRNs were performed using methods outlined above.
The threshold where the exponents of samples start to deviate significantly is called the knee point of exponential function. A MATLAB code written by Dimitry Kaplan called Knee Point finds the knee point by fitting two lines (in each direction) at each bisection point and calculating the sum of errors of points along those lines. The knee is judged to be at the bisection point which minimizes the sum of errors of the two fits.
|
10.1371/journal.pmed.1002908 | Portrayals of mental illness, treatment, and relapse and their effects on the stigma of mental illness: Population-based, randomized survey experiment in rural Uganda | Mental illness stigma is a fundamental barrier to improving mental health worldwide, but little is known about how to durably reduce it. Understanding of mental illness as a treatable medical condition may influence stigmatizing beliefs, but available evidence to inform this hypothesis has been derived solely from high-income countries. We embedded a randomized survey experiment within a whole-population cohort study in rural southwestern Uganda to assess the extent to which portrayals of mental illness treatment effectiveness influence personal beliefs and perceived norms about mental illness and about persons with mental illness.
Study participants were randomly assigned to receive a vignette describing a typical woman (control condition) or one of nine variants describing a different symptom presentation (suggestive of schizophrenia, bipolar, or major depression) and treatment course (no treatment, treatment with remission, or treatment with remission followed by subsequent relapse). Participants then answered questions about personal beliefs and perceived norms in three domains of stigma: willingness to have the woman marry into their family, belief that she is receiving divine punishment, and belief that she brings shame on her family. We used multivariable Poisson and ordered logit regression models to estimate the causal effect of vignette treatment assignment on each stigma-related outcome. Of the participants randomized, 1,355 were successfully interviewed (76%) from November 2016 to June 2018. Roughly half of respondents were women (56%), half had completed primary school (57%), and two-thirds were married or cohabiting (64%). The mean age was 42 years. Across all types of mental illness and treatment scenarios, relative to the control vignette (22%–30%), substantially more study participants believed the woman in the vignette was receiving divine punishment (31%–54%) or believed she brought shame on her family (51%–73%), and most were unwilling to have her marry into their families (80%–88%). In multivariable Poisson regression models, vignette portrayals of untreated mental illness, relative to the control condition, increased the risk that study participants endorsed stigmatizing personal beliefs about mental illness and about persons with mental illness, irrespective of mental illness type (adjusted risk ratios [ARRs] varied from 1.7–3.1, all p < 0.001). Portrayals of effectively treated mental illness or treatment followed by subsequent relapse also increased the risk of responses indicating stigmatizing personal beliefs relative to control (ARRs varied from 1.5–3.0, all p < 0.001). The magnitudes of the estimates suggested that portrayals of initially effective treatment (whether followed by relapse or not) had little moderating influence on stigmatizing responses relative to vignettes portraying untreated mental illness. Responses to questions about perceived norms followed similar patterns. The primary limitations of this study are that the vignettes may have omitted context that could have influenced stigma and that generalizability beyond rural Uganda may be limited.
In a population-based, randomized survey experiment conducted in rural southwestern Uganda, portrayals of effectively treated mental illness did not appear to reduce endorsement of stigmatizing beliefs about mental illness or about persons with mental illness. These findings run counter to evidence from the United States. Further research is necessary to understand the relationship between mental illness treatment and stigmatizing attitudes in Uganda and other countries worldwide.
The experimental procedures for this study were registered with ClinicalTrials.gov as "Measuring Beliefs and Norms About Persons With Mental Illness" (NCT03656770).
| Mental illness stigma is a fundamental barrier to improving mental health worldwide.
While there has been some progress in understanding how to reduce mental illness stigma in high-income countries, it is unclear how this understanding might generalize to low- and middle-income countries.
The extent to which people perceive that mental illness can be effectively treated may be an important component of changing negative beliefs about mental illness.
We conducted a survey experiment to understand how information about successful treatment of mental illness might affect stigmatizing beliefs in rural southwestern Uganda.
This experiment involved randomly assigning different people in eight villages to be read a vignette about: a woman who had signs suggestive of one of three different types of mental illness; a woman who had these signs and was treated successfully; or a woman who had these signs and was treated successfully but subsequently relapsed.
We found that stigma toward mental illness in the community was common and was generally unaffected by descriptions of successful treatment.
If unaddressed, stigma will continue to pose a major barrier to improving population mental health in Uganda.
We need to do more research to understand the relationship between perceptions of mental illness treatment and stigmatizing attitudes in Uganda and other countries worldwide.
Engaging local etiologies, making treatment more accessible, and understanding how mental illness shapes social relationships independent of actual symptoms might be important avenues of research and program implementation to explore.
| Mental illness is heavily stigmatized worldwide. In cross-national studies, people with mental illness report experiencing discrimination in most areas of their life, including making friends, keeping jobs, or interacting with their partners and families [1–3]. Available evidence suggests that while beliefs about mental illness vary by country, negative attitudes toward people with mental illness are neither uncommon nor isolated [4,5]. Widespread negative attitudes provide an enabling environment for harmful violations of basic human rights that range in severity from prejudicial behavior and employment discrimination to chaining, caging, and killing [6,7]. These attitudes undercut efforts to improve mental health at a fundamental level because stigma undermines already low rates of mental-healthcare–seeking behavior [8–18]. Compounding this attenuating effect on treatment-seeking, stigma is also associated with reduced public support for funding toward mental health services, which erodes the availability of appropriate care within the mental healthcare system [19–21].
Attempts to reduce stigma in high-income countries have achieved some measure of success, though the durability of these results is uncertain [22,23]. Furthermore, the evidence, especially from low- and middle-income countries (LMICs), is not sufficient to understand the extent to which these interventions can be generalized [24–32]. Contrary to the hypotheses that motivated many of the early large-scale awareness campaigns in many high-income countries during the 1990s and 2000s, increasing understanding of the potential biological underpinnings of mental illness has not positively influenced attitudes toward persons with mental illness, and there are some circumstances in which this increasing biological understanding may have even worsened stigma [26,33–38]. The successes of recent contact-based interventions that promote meaningful contact with persons with mental illness alongside targeted education appear to hold some promise for long-term stigma reduction efforts [22,23,39–42]. While results from this class of interventions are somewhat mixed, a key ingredient of successful contact interventions seems to be an emphasis on recovery [43,44]. It may be, then, that one reason early interventions failed to reduce stigma is that the stigma attached to mental illness is driven in large part by beliefs about the extent to which it can be treated [45], and, despite increasing acceptance of biological causes of mental illness, many people continue to perceive mental illness to be untreatable [10,46].
Experiments that have been conducted to examine the relationship of treatment information to stigmatizing attitudes in the United States have shown that providing study participants with vignettes describing successful treatment of mental illness can reduce desire for social distance and negative attitudes about mental illness and can enhance beliefs in the effectiveness of treatment [47–49]. This treatable-illness hypothesis has not been tested internationally, but related literature on HIV stigma in sub-Saharan Africa suggests that the widespread availability of effective HIV antiretroviral treatment has enabled people with HIV to actively contribute (socially and economically) to their communities, thereby reducing internalized stigma among people with HIV and enhancing their standing in the general community [50–56]. Taken together, these studies suggest that the connection between perceptions of treatability and mental illness stigma may also generalize to contexts like sub-Saharan Africa.
To investigate the treatable-illness hypothesis, we embedded a survey experiment into a whole-population cohort study in rural southwestern Uganda. Adapting vignettes from the General Social Survey and the novel study by McGinty and colleagues [48], we aimed to determine whether the extent to which mental illness was portrayed as a treatable medical condition affected personal beliefs and perceived norms about mental illness and about people with mental illness. Specifically, we hypothesized that descriptions of mental illness alone and relapse after initially effective treatment would elicit more stigmatizing responses compared with descriptions of successfully treated mental illness.
This study was conducted in the eight villages of Nyakabare Parish, a rural administrative subunit of Mbarara District in the southwest region of Uganda. The study site is representative of rural communities in Mbarara District; it is relatively isolated, with an economy dominated by small-scale farming, animal husbandry, and petty trading, and both food and water insecurity are common [57,58]. All procedures were embedded within an ongoing, whole-population social network cohort study in which study participants are surveyed biennially [59]. The study includes all adults aged 18 years and above (and emancipated minors aged 16–17 years) who maintain stable primary residence within the roughly 11-square-kilometer area of the parish and who can give informed consent to participate. Exclusions include people who cannot communicate meaningfully with research staff, for example, because of deafness, mutism, or aphasia; people with behavioral problems thought to represent psychosis, neurological damage, or acute intoxication; and people too cognitively impaired to provide informed consent. Participants in the baseline survey, which was completed in 2015, served as the randomization sample for the present survey experiment. The second biennial survey was fielded between November 2016 and June 2018, providing the data reported in this manuscript.
A total of 1,776 participants (who were enumerated in the 2014–2015 baseline survey) were randomly assigned to receive one of 10 different vignettes describing a young woman (see S1 Text). Reporting was guided by the CONSORT checklist (see S1 CONSORT checklist). The control vignette described the demographic characteristics and basic life story of a typical Ugandan woman with no further elaboration. The remaining nine vignettes included the same basic description of the woman but also described her experiencing three different types of symptoms (psychosis, mania, and depression, suggestive of schizophrenia, bipolar disorder, and major depressive disorder, respectively), each with three different treatment outcomes (no treatment, successful treatment followed by recovery, and successful treatment followed by recovery and then relapse/recurrence). These vignettes were adapted from McGinty and colleagues [48] to fit the local context based on feedback from key informants, documented symptom presentation in Uganda, and consultation with psychiatrists at the Mbarara Regional Referral Hospital.
Each eligible person was approached in the field, typically at their home or place of employment, by a research assistant who spoke the local language (Runyankore) and who requested their participation in the study. The survey was framed in general terms as a study about the social lives and health of residents of Nyakabare Parish, not as a study about beliefs about mental illness. For persons who expressed potential interest, the study was described in detail, and their written informed consent to participate was obtained. Study participants who could not sign their name were permitted to indicate consent with a thumbprint. All research assistants received in-depth training on how to administer surveys for gathering sensitive information, including instructions on how to temporarily halt the survey if another person came within earshot.
After study participants were presented with one of the 10 vignettes, they were asked to respond to three questions regarding their personal beliefs about mental illness and about people with mental illness, and three questions regarding their perceptions of village norms about mental illness and about people with mental illness. (For the sake of parsimony in writing, hereafter we refer to the subject of these questions using the shorthand “about mental illness.”) Specifically, participants were asked whether they would allow the woman to marry a member of their family, whether they believed the woman was receiving divine punishment for engaging in immoral behavior, and whether they believed she brought shame upon her family. These questions measure three different domains of stigma that are salient to the local context: social distance [60–62], etiological attribution [34], and courtesy stigma [63–65]. In response to the three items about personal beliefs, participants were allowed to provide one of five different responses: “Yes,” “No,” “It depends on knowing more details,” “Do not know,” and “Refuse to respond.”
Paralleling these three outcome variables measuring personal beliefs, study participants were also asked about the proportion of other people in their village who would allow the woman to marry a member of their families, the proportion of other people in the village who believed the woman was receiving divine punishment for engaging in immoral behavior, and the proportion of other people in their village who believed she brought shame upon her family. These questions measure the same domains of stigma (social distance, etiological attribution, and courtesy stigma) but focus on perceived norms rather than personal beliefs. These questions were modeled after previously published studies of perceived norms about different health behaviors and health risk behaviors [66–69]. The survey questions specified “your village” so that all participants would have a similar fixed, unambiguous reference group when describing their perceptions about the norm within their villages [70]. Response options for the items about perceived norms followed a four-point Likert-type scale (in addition to “Do not know” and “Refuse to respond”): “All or almost all,” “More than half,” “Fewer than half,” and “Very few, or no one.”
The English translations of the six outcome measures are provided in S1 Text. The vignettes and associated survey questions were first written in English, translated into Runyankore, and then back-translated into English to verify the fidelity of the translated text. The translation and back-translation followed an iterative process involving in-depth consultation and pilot testing with key informants.
Participants were allocated to the 10 vignettes in equal proportions in a parallel group design in which treatment assignment was determined centrally using a computerized random number generator. Neither the research assistants administering the questionnaires nor the study participants were aware of the vignettes to which the study participants had been assigned. The research assistants were not blinded, however, so they likely perceived the differences in the vignettes being administered to different study participants. To ensure balance across sex and village strata, we generated 16 separate randomization schedules for subsets of participants defined by strata of sex and village of residence [71]. The vignettes and all associated survey questions were programmed into laptop computers running the Computer Assisted Survey Information Collection (CASIC) Builder software program (West Portal Software Corporation, San Francisco, CA, USA) so that the surveys could be administered in the field. The experimental procedures for this study were registered with ClinicalTrials.gov (NCT03656770). The protocol record was entered in April 2017 but, because of an administrative error, was not released on ClinicalTrials.gov until August 2018.
The analysis was prespecified prior to data collection (S2 Text). For the three outcome variables measuring personal beliefs, responses were coded such that 0 denoted a nonstigmatizing response and 1 denoted a stigmatizing or ambivalent response. Namely, unwillingness to allow the woman to marry a member of the participant’s family, belief that she was receiving divine punishment, and belief that she brought shame on her family were assigned values of 1, along with the ambivalent response of “it depends.” Willingness to allow the woman to marry into the family, belief that she was not receiving divine punishment, and belief that she did not bring shame upon her family were assigned values of 0. “Do not know” and “refuse” were considered missing data. We then fitted Poisson regression models specifying each outcome as the dependent variable and the vignette treatment assignment as the primary exposure of interest. Cluster-correlated robust estimates of variance were used so that the estimated incidence rate ratios could be interpreted as risk ratios [72,73].
For the three 4-level categorical outcome variables measuring perceived norms, responses were coded 1–4 such that the lowest category denoted the least perceived stigma and the highest category denoted the most perceived stigma. (For example, in response to the question about whether others in the village would permit the woman in the vignette to marry into the family, “Very few, or no one” was the highest category and “All or almost all” was the lowest category.) “Do not know” and “refuse” were considered missing data. We then fitted ordinal logit regression models specifying each outcome as the dependent variable and the vignette treatment assignment as the primary exposure of interest. The exponentiated regression coefficients were interpreted as estimated odds ratios. To confirm that the regression coefficients did not vary across the logit equations (i.e., the assumption of proportional odds), we used the omnibus Wald test by Brant [74].
To ensure accurate confidence intervals that accounted for the stratified randomization scheme, we adjusted treatment estimates for sex and village of residence by including them as covariates in the respective Poisson and ordinal logit regression models described above. Stata statistical software was used to conduct all data cleaning and analysis (version 14.0, College Station, TX, USA).
Ethical approval for this study was granted by the Partners Human Research Committee at Massachusetts General Hospital and the Research Ethics Committee at Mbarara University of Science and Technology. We also received clearance for the study from the Uganda National Council of Science and Technology and the Research Secretariat in the Office of the President of the Republic of Uganda.
Of the 1,776 participants enumerated and randomized in the 2014–2015 survey, 1,355 (76%) were successfully interviewed in 2016–2018, excluding 10 individuals who were not administered the experiment correctly because of a technical error. Of the remainder, 250 (14%) were known to have emigrated out of the study site, 57 (3%) could not be located, 37 (2%) had died, 42 (2%) refused to participate, and 25 (1%) were ineligible or could not be interviewed for other reasons (for example, incarceration or acute intoxication at each of multiple interview attempts). We summarize participant characteristics in Table 1. Respondents came from all eight villages, and just over half were women (56%). The mean age was 42 years, with good representation from all age groups. Just over half (57%) had completed primary school, and almost two-thirds were married or cohabiting (64%).
A technical error in survey administration resulted in some treatment assignments that departed from the intended randomization. (S3 Text provides further detail about the nature of the error, a comparison of the correctly versus incorrectly assigned participants, and results of a sensitivity analysis based on a data set excluding the incorrectly assigned participants. As shown in the S3 Text, neither the reported results nor final conclusions were substantively affected by the error.) Participants who received a vignette portraying any kind of mental illness reported more stigmatizing personal beliefs compared with those who received the control vignette, across all outcomes, for every variant of symptom presentation, and for every variant of treatment description (Table 2). Across outcomes, relative to the control vignette (22%–30%), substantially more study participants believed the woman in the vignette was receiving divine punishment (31%–54%) or believed she brought shame on her family (51%–73%), and most were unwilling to have her marry into their family (80%–88%). A small number of study participants provided ambivalent responses (1%–8%, depending on the outcome), and there were negligible refusals (<1%) and "don't know" (0%–4%) responses.
Compared with the responses to questions about their personal beliefs, study participants’ responses to questions about perceived norms about people with mental illness followed similar patterns, though the differences in comparison with the control vignette were not as large in magnitude (Table 3). Once again, there were negligible refusals (<1%) and "don't know" (0%–6%) responses.
Using Poisson regression models that also adjusted for the stratification variables, we found that portrayals of mental illness significantly increased the risk of stigmatizing responses compared to the control vignettes, across all outcomes, for every variant of symptom presentation, and for every variant of treatment description—except one (Table 4). Namely, participants who received the vignette describing a woman receiving effective treatment for depressive symptoms but then experiencing a subsequent relapse were no more likely to believe she was receiving divine punishment than participants in the control group (adjusted relative risk [ARR] = 1.46, 95% CI 0.89–2.51, p = 0.18). Apart from that exception, participants who received vignettes depicting a woman with a mental illness were, depending on the symptom presentation and treatment experience described, 2.64–2.98 times more likely (than those who received the control vignette) to be unwilling to allow a family member to marry her, 2.20–3.14 times more likely to believe that she brought shame upon her family, and 1.53–2.49 times more likely to believe that she was receiving divine punishment (all p-values < 0.05).
Contrary to our hypotheses motivated by the work of McGinty and colleagues [48], study participants who received vignettes depicting effective treatment were only slightly less likely to endorse stigmatizing responses compared with those who received vignettes depicting untreated mental illness. In the case of bipolar illness, participants receiving the vignette about effective treatment were still more likely to endorse stigmatizing responses than participants receiving the control vignette (ARRs ranged from 1.8–2.6, all p < 0.001), but the ARRs were smaller in magnitude compared with those receiving the vignette about untreated bipolar illness (ARRs ranged from 2.5–3.1, all p < 0.001) (p-value for comparisons all <0.05).
Results from the ordered logit regression models comparing perceived norms followed a similar pattern to that of personal beliefs (Table 5). Across all outcomes, for every variant of symptom presentation, and for every variant of treatment description, a vignette describing any type of mental illness (untreated, treated, or treated with relapse) increased the odds that participants perceived more people in their village would be unwilling to allow the woman to marry into their families, participants perceived more people would believe that she brings shame upon her family, and participants perceived more people would believe that she was receiving divine punishment. As with the data on personal beliefs, in the case of bipolar illness, participants receiving the vignette about effective treatment were still more likely to endorse stigmatizing responses than participants receiving the control vignette (adjusted odds ratios [AORs] ranged from 1.8–10.5, all p < 0.001) but the AORs were smaller in magnitude compared with those receiving the vignette about untreated bipolar illness (AORs ranged from 3.1–16.3, all p < 0.001) (p-value for comparisons all <0.01).
In this population-based, randomized survey experiment conducted in rural southwestern Uganda, portrayals of effectively treated mental illness did not appear to reduce endorsement of stigmatizing responses about mental illness. Instead, any kind of mental illness portrayal—whether untreated, successfully treated, or treated with relapse—resulted in an overwhelmingly large proportion of stigmatizing responses. Among those responses, refusal to have a woman with mental illness marry into the family was the most common, though beliefs that her mental illness created shame for the family or was the result of divine punishment were also fairly common. Perceptions of village norms followed similar patterns as individual beliefs.
Our primary finding that varying degrees of treatment success had no ameliorating effect on stigmatizing beliefs runs contrary to other similar studies conducted in the United States, which have found that portrayals of effective treatment reduced stigma for a variety of mental illnesses [47–49]. Below, we offer several possible reasons why treatment information may have produced differing levels of stigmatizing responses in this study as compared with previously published work.
Lebowitz and Ahn’s [47] experiment examining the role of etiology in stigma and treatment descriptions may provide one explanation for this discrepancy. By varying the description of mental illness etiology as well its treatment, they showed that attributing a mental illness to biological origins largely parallels the findings of other studies, with worsened stigma resulting when the vignette did not mention treatment but reduced stigma resulting when the vignette described effective treatment. By contrast, ascribing a nonbiological cause to the mental illness rendered the treatment information irrelevant. In our study, we may have observed this latter effect. Nonbiological interpretations of mental illness are widely held in Uganda [75–79] and sub-Saharan Africa generally (particularly in rural areas) [80–85], unlike many high-income countries, which have experienced steady shifts over the past several decades in the direction of a neurobiological understanding of mental illness [34,86]. One potential interpretation of our findings is that participants may have viewed the symptoms as essentially nonbiological and therefore received the information about biomedical treatment as being less salient (compared with participants in studies previously conducted in the United States). What we described as “effective treatment” may not have addressed what many of our participants understood to be the true ailment and source of stigma [87].
Apart from etiology, confidence in the treatment process itself may also explain the discrepant findings. Uganda’s mental health system, much like those in other low-income countries, lacks the resources necessary to consistently provide effective treatment across the country, with only 2.96 mental health workers (including just 0.09 psychiatrists and 2.24 mental health nurses) per 100,000 people [88]. With pharmaceutical treatment availability concentrated in urban areas and typically limited to older, cheaper, and less effective medications, participants in this study may have perceived mental health treatment to be largely ineffective or inaccessible [76,77,89]. These perceptions may have attenuated the effect that providing them with treatment information could have had on stigmatizing beliefs elicited in the survey [90].
Even with descriptions of etiologically appropriate or available treatment, participants may still have understood mental illness to be essentially permanent. Key informants pointed to a local Runyankore proverb, “one who has been mad will always scare the children,” that captures the widely held perception that mental illness is simply not treatable. This belief that mental illness is never truly eliminated even if treated to long-term remission likely interacts in significant ways with etiological attributions and experiences with the mental healthcare system. However, as Schnittker and colleagues [46] note in their reflection on the failures of genetic descriptions to reduce stigma in the US, etiological beliefs and even endorsement of treatment are distinct from the idea that a person can truly recover from mental illness. This belief that mental illness is a permanent condition independent of its symptoms is most characteristic of classic stigma in that it is a “mark” or label that relates people with mental illness to undesirable attributes intrinsically as a permanent identity [65,91,92]. The loss of status associated with this label, as the results of our study show, might never be reversed even if someone resumes all roles and social functions symptom-free. Many of the studies on mental illness and its associated stigma in Uganda, and in sub-Saharan Africa generally, have focused on causation and explanatory models [75–85]. There has been far less attention to the ways mental illness can shape enduring social identities and relationships that may be only indirectly related to actual symptoms and treatment [93,94].
A second important finding of our study is that we observed highly stigmatizing responses regardless of symptom presentation and across multiple domains. Key informant feedback suggests that participants in our study did not differentiate between the different symptom presentations and instead likely understood all variations in symptom presentation to represent a single category of mental illness. Whether or not these anecdotal observations held true throughout the study sample, our data show that stigma toward people exhibiting symptoms of mental illness was very common, consistent with the pattern of findings documented in other studies [1–3]. That most participants in our study believe mental illness brings shame on a person’s family also highlights the effects stigma can have on isolating people and undermining their ability to obtain social support [64,95–97].
Interpretation of our findings is subject to several important limitations. Our use of vignettes as the primary experimental manipulation allowed us to vary key details of each description. Nonetheless, study participants were only exposed to hypothetical scenarios that were potentially lacking in context that might influence stigma [62,98]. This limitation is especially important in light of evidence that the stigma of mental illness can interact and be compounded by other intersecting identities [99]. For example, several studies have shown that gender can interact with mental illness type to affect stigma [100,101], and recent evidence from related literature suggests that socioeconomic status may also influence participant responses [102]. Since our vignettes only described a woman with average socioeconomic status for the region, we are unable to determine whether and how descriptions of other identities could have yielded differences in stigmatizing responses. For example, if the vignettes had portrayed a similar man with mental illness, it is possible that the findings would have been different. Second, it is possible that social desirability bias could have affected study participants’ responses [103]. If negative attitudes toward persons with mental illness were understood to be socially undesirable, study participants might have overestimated the proportion of others who would endorse stigmatizing attitudes while underreporting their own stigmatizing attitudes toward persons with mental illness [62,103,104]. This phenomenon was not observed in the data. Participants generally overestimated the proportion of others who would be unwilling to have the woman marry into their families but generally underestimated the extent to which others believed the woman was receiving divine punishment or brought shame on her family. It is important to note that, while social desirability bias could have affected the overall levels of stigma in the population, it is an unlikely explanation for the differences in levels of stigma across the treatment arms (given the randomized design). A third limitation is that this experiment focused on mental illnesses but did not include descriptions of substance use disorders, for which perceptions and attitudes are likely to differ significantly [48,105–107]. Fourth, our findings may not generalize beyond this rural region of southwestern Uganda. Other studies have found important differences in mental illness stigma between rural and urban areas [108,109]. However, the study was based on a whole-population sample, and the community we surveyed shares important characteristics with the rest of the country and the East African region. Finally, it is important to note that a brief vignette portrayal of mental illness as a treatable health condition might not have an enduring educational effect (that could therefore translate into an enduring antistigma effect). It is possible that more sustained education about the treatability of mental illness could have affected study participants’ responses differently. That being said, it is notable that in Uganda, more education does not appear to have reduced the stigma attached to HIV [110] and that the evidence for efficacy of education-based interventions in reducing mental illness stigma is also mixed [111].
The results from this study have several important applications for treatment and policy in Uganda. Primarily, it is clear that mental illness remains highly stigmatized in Uganda. Given the well-established connections between stigma and undertreatment, underfunding, and even abuse of people with mental illnesses, there remains important work to be done in reducing stigma to improve the health and lives of persons with mental illness. Second, this experiment indicates that portrayals of successful treatment of mental illness did not reduce stigmatizing attitudes in Uganda the way it seemed to reduce stigmatizing attitudes in the US. Further research, particularly qualitative investigation into the stigma attached to mental illness, is needed to achieve deeper understanding of the stigma pathways associated with mental illness throughout East Africa.
In summary, we found that mental illness stigma is common in rural southwestern Uganda. Stigma erodes efforts to promote mental health, preventing people from seeking treatment and putting them at risk for further suffering [91]. Describing mental illness as treatable does not seem to have had any effect on reducing negative attitudes toward mental illness or persons with mental illness in rural southwestern Uganda. Instead, further research into stigma reduction is necessary to understand other ways to address the stigma of mental illness in East Africa.
|
10.1371/journal.pntd.0007163 | Extracellular vesicles of Trypanosoma cruzi tissue-culture cell-derived trypomastigotes: Induction of physiological changes in non-parasitized culture cells | Trypanosoma cruzi is the obligate intracellular parasite that causes Chagas disease. The pathogenesis of this disease is a multifactorial complex process that involves a large number of molecules and particles, including the extracellular vesicles. The presence of EVs of T. cruzi was first described in 1979 and, since then, research regarding these particles has been increasing. Some of the functions described for these EVs include the increase in heart parasitism and the immunomodulation and evasion of the host immune response. Also, EVs may be involved in parasite adhesion to host cells and host cell invasion.
EVs (exosomes) of the Pan4 strain of T. cruzi were isolated by differential centrifugation, and measured and quantified by TEM, NTA and DLS. The effect of EVs in increasing the parasitization of Vero cells was evaluated and the ED50 was calculated. Changes in cell permeability induced by EVs were evaluated in Vero and HL-1 cardiomyocyte cells using cell viability techniques such as trypan blue and MTT assays, and by confocal microscopy. The intracellular mobilization of Ca2+ and the disruption of the actin cytoskeleton induced by EVs over Vero cells were followed-up in time using confocal microscopy. To evaluate the effect of EVs over the cell cycle, cell cycle analyses using flow cytometry and Western blotting of the phosphorylated and non-phosphorylated protein of Retinoblastoma were performed.
The incubation of cells with EVs of trypomastigotes of the Pan4 strain of T. cruzi induce a number of changes in the host cells that include a change in cell permeability and higher intracellular levels of Ca2+ that can alter the dynamics of the actin cytoskeleton and arrest the cell cycle at G0/G1 prior to the DNA synthesis necessary to complete mitosis. These changes aid the invasion of host cells and augment the percentage of cell parasitization.
| Extracellular vesicles (EVs) are a diverse group of nanoparticles involved in intercellular communication under physiological and pathological conditions. Trypanosoma cruzi, the protozoan that causes Chagas disease, releases EVs that facilitate parasite invasion of the host cell, immunomodulate the host response, and help the parasite to evade this response. However, little is known about how the host cell is altered. In this work, we confirm that EVs of tissue-culture cell-derived trypomastigotes of the Pan4 strain increase cell parasitism. We also demonstrate that EVs affect cell permeability in Vero cells and cardiomyocytes and raise intracellular Ca2+ levels, altering the actin filaments and arresting the cell cycle at the G0/G1 phases. This work seeks to elucidate the way in which EVs influence certain aspects of the cell physiology that favour the establishment of this parasite inside the host cell.
| Trypanosoma cruzi is an intracellular protozoan parasite that causes Chagas disease or American trypanosomiasis. An estimated 8 million people are infected with this parasite worldwide, with some 300,000 new cases and 15,000 deaths annually [1]. T. cruzi has a life cycle that includes mammals and blood-sucking bugs (Hemiptera, Reduviidae) as hosts. Humans can be infected through the insects faeces, by vertical (congenital) transmission, transmission by blood transfusions, organ transplants, or oral contamination via tainted fluids and foods [2].
Chagas disease displays symptomatic and pathological variations among infected individuals [3] but is characterized by an acute as well as a chronic stage. During the chronic stage, approximately 30% of the patients develop significant complications, which may include megasyndromes of the gastrointestinal tract (such as megacolon or megaesophagus), neurological complications, and cardiomyopathy [4–7].
The pathogenesis of Chagas disease is a multifactorial process. The molecular invasion mechanisms by T. cruzi trypomastigotes (T) and the associated regulatory pathways have been intensely investigated for many years [8]. A large number of molecules have been involved and are described as part of the secretome of T. cruzi [9]. Some of them are included in extracellular vesicles (EVs). EVs are small membrane-bound vesicles classified based on their size, biogenesis, and composition [10] in: a) exosome-like EVs (20–100 nm), b) ectosome-like EVs (100–1000 nm) and c) apoptotic blebs (>1000 nm) [9,11].
The presence of EVs of T. cruzi was first described in 1979 by da Silveira et al., who demonstrated the secretion of plasma-membrane vesicles by T. cruzi epimastigote forms [12]. These vesicles were later detected by Gonçalves et al. (1991) in host-cell-derived trypomastigotes [13]. Since then, numerous publications concerning EVs have appeared, demonstrating their role in cell-to-cell communication, pathogenesis, evasion of the immune response and diagnosis.
The cargo of EVs of T. cruzi contain proteins involved in host-parasite interactions, signalling, trafficking, and membrane fusion, transporters, oxidation-reduction, etc. [9]. Small RNAs derived from tRNAs and rRNAs have also been reported [14]. Some of the functions described for these EVs include the increase in heart parasitism and the immunomodulation of the host response [15]; the evasion of innate immunity [16]; and the induction of the release of EVs by the host blood cells that are involved in inhibiting complement-mediated lysis [17]. Also, EVs may be involved in parasite adhesion to host cells and host-cell invasion [15,17–20]. Recently, EVs have proved useful in evaluating disease severity as well as vaccine and drug effectiveness against chagasic cardiomyopathy [21]. However, little is known about the capacity of EVs to modulate the host-cell conditions. In this sense, the present work seeks to elucidate certain effects exerted by T. cruzi EVs over the parasite establishment inside the host cell.
Vero (ECACC 84113001) and 3T3 cells (CRL 1658) were cultured in Nunc cell-culture flasks of 75 cm2 surface area (Thermo Fischer Scientific, USA) in Modified Eagle’s Medium (MEM) (Sigma, USA) supplemented with 10% foetal bovine serum (Gibco, USA) previously inactivated at 56°C for 30 min (IFBS) plus antibiotics (penicillin 100 U/mL, streptomycin 100 μg/mL). The cultures were maintained at 37°C, in a moist atmosphere enriched with 5% CO2. HL-1 cardiac muscle-cell line was grown as described above, using Claycomb medium supplemented with 10% IFBS, norepinephrine 0.1 mM, L-glutamine 2 mM and antibiotics (penicillin 100 U/mL, streptomycin 100 μg/mL). The cell cultures were routinely monitored for Mycoplasma by PCR.
Vero cells were initially infected with purified metacyclic trypomastigotes of the Pan4 (Tc Ia + Tc Id) strain of T. cruzi obtained in vitro, according to de Pablos et al. (2011) [22]. After 120 h of the intracellular development of the parasite, tissue-culture cell-derived trypomastigotes (TcT) were harvested by centrifugation. Parasites were collected routinely every 120 h from the infected cell monolayer. The culture medium was centrifuged at 3,000 xg for 5 min and the pellet with the parasites was washed in PBS four times.
To obtain the EVs from the TcT, we followed the procedure described previously by de Pablos et al. (2016) [18], with some modifications. Parasites were incubated for 5 h at 37°C in RPMI medium (Sigma, USA) buffered with 25 mM HEPES at 7.2 and supplemented with 10% exosome-free IFBS. Afterwards, parasites were removed by centrifugation at 3,500 xg for 15 min and the supernatant was collected and centrifuged at 17,000 xg for 30 min at 4°C. This supernatant was filtered through a 0.22 μm pore filter (Sartorius, Germany) and ultracentrifuged at 100,000 xg for 16–18 h to obtain the EVs (mostly exosomes). All the steps were performed in an ultracentrifuge Avanti J-301 (Beckman Coulter, USA) with a JA-30.50 Ti rotor. The resulting pellet was washed three times in PBS in an ultracentrifuge Sorwal WX80 (Thermo Fisher Scientific, USA) with F50L-24 x 1.5 fixed-angle rotor and resuspended in 100 μL PBS.
The isolation procedure was evaluated by Transmission Electron Microscopy (TEM), Nanoparticle Tracking Analysis (NTA) and Dynamic Light Scattering (DLS). The proteins from the EVs were quantified using the Micro-BCA protein assay kit (Thermo Fischer Scientific, USA), using bovine-serum albumin as standard. Viability of the TcT after shedding of EVs was evaluated using the trypan blue exclusion test. After 5 h, no significant death was detected and over 99% of the parasites were viable.
To demonstrate the specificity of the effects of the EVs from the TcT of T. cruzi and to evaluate the effect of the EVs of trypomastigotes of the Pan4 strain over the infection of cells with T. cruzi from a different DTU and another intracellular microorganism, we performed the isolation of EVs from Crithidia mellificae and the 3T3 cell line (a fibroblast cell line) and evaluated the effect of these EVs over the parasitization percentages of cells infected with T. cruzi Pan4. We also employed T. cruzi 4162 strain (Tc IV) and tachyzoites of Toxoplasma gondii RH for the infection of cells previously incubated with EVs of T. cruzi Pan4.
For the isolation of EVs from choanomastigotes of Crithidia mellificae, 1x107 parasites were incubated for 24 h at 28°C in LIT medium. Nunc cell-culture flasks of 75 cm2 surface area (Thermo Fischer Scientific, USA) with confluent monolayers of 3T3 cells were washed 3 times with MEM without IFBS and the cells were incubated for 24 h at 37°C with MEM (Sigma, USA). After 24 h of incubation, the culture media were collected and centrifuged at 3,500 xg for 15 min and the obtained supernatants were handled the same way as for the isolation of EVs from TcT. DLS and the quantification of the protein concentration of these samples using the Micro-BCA protein assay kit (Thermo Fischer Scientific, USA) were performed as described above.
The purification of metacyclic trypomastigotes of T. cruzi 4162 strain was also performed according to de Pablos et al. (2011) [22] and TcT were obtained after the infection of Vero cells as described. Tachyzoites of Toxoplasma gondii RH strain were maintained in our laboratory by serial passage, in semiconfluent monolayers of Vero cells and cultured in the same conditions as T. cruzi. The egressed parasites were harvested, centrifuged at 5,000 xg for 10 min, washed three times in PBS and added to the cell cultures in a ratio 5:1 (parasites:cell).
To confirm the presence of EVs in our samples, we resuspended an aliquot of the pellet in 0.1 M Tris HCl (pH 7.2) and 5 μL each sample were adsorbed directly onto Formvar/carbon-coated grids and stained with 2% (vol/vol) uranyl acetate, for the direct observation in a TEM, LIBRA 120 PLUS Carl Zeiss microscope. The diameter of the EVs was measured by ImageJ 1.41 software.
Distribution, size, and concentration of T. cruzi EVs from trypomastigotes was determined by measuring the rate of Brownian motion according to the particle size, using a Nanosight NS300 (Malvern Instruments, UK). This system was equipped with a sCMOS camera and a blue 488 nm laser beam. Samples were diluted 1/100 just before the analysis, in low-binding Eppendorf tubes with PBS and the measurements were performed at 25°C. For data acquisition and information processing, we used the NTA software 3.2 Dev Build 3.2.16. The particle movement was analysed by NTA software with the camera level at 16, slider shutter at 1200, and slider gain at 146.
To confirm the results obtained by NTA, we also performed Dynamic Light Scattering (DLS) of the EVs of trypomastigotes, choanomastigotes and cells using a Zetasizer nano ZN90 (Malvern Instruments, UK). Samples were prepared the same way as described for the NTA and the measurements were also performed at 25°C. For data acquisition and information processing, the Zetasizer Ver. 7.11 software was employed.
The presence of some molecules without orthologues in other organism and involved in the invasion process of T. cruzi was evaluated in EVs by Western blotting. Briefly, 300 μg of EVs isolated from TcT of the Pan4 strain were resolved by SDS-PAGE, transferred to a nitrocellulose membrane and blocked overnight with 5% non-fat milk in PBS-0.1% Tween 20. Primary antibodies anti-cruzipain (1:3,000), anti-TS (mAb 39) (1:1,000), and anti-MASPs (signal peptide, SP) (1:1,000) [16] were incubated overnight at 4°C. The membranes were washed with PBS-0.1% Tween 20 and incubated for 1 h with goat anti-mouse IgGs conjugated with peroxidase (1:1,000) (Dako Agilent Pathology Solutions, USA) in the case of TS and MASPs and goat anti-rabbit IgGs conjugated with peroxidase (1:2,000) in the case of cruzipain (Dako Agilent Pathology Solutions, USA). The reaction was visualized using Clarity ECL Western substrate (BioRad, Spain) in a ChemiDoc Imaging system (BioRad, Spain).
Cultures of 5x104 Vero cells were grown in MEM supplemented with 10% IFBS over round 13-mm coverslips (Marienfeld, Germany), in Nunc 24-well plates (Thermo Fischer Scientific, USA) for 24 h at 37°C and 5% CO2. After this time, coverslips with cells were washed 3 times in MEM and incubated for 2 h with 0.1, 0.25, 0.5, 1 and 2.5 μg/mL EVs in MEM. After the incubation, cells were infected with T. cruzi trypomastigotes of the Pan4 strain, at a parasite:host cell ratio of 5:1. After 4 h, parasites were removed and the cells were washed and maintained in culture for 24 h. The cultures were fixed with methanol and stained with Giemsa. Parasitization percentages and parasitization indexes (number of amastigotes per cell) were calculated after counting at least 400 cells.
The invasion assays were also performed using T. cruzi EVs submitted to thermal and chemical treatments. For the thermal treatments, EVs were incubated in a water bath at 50°C, 70°C, and 90°C for 30 min. For the chemical treatments, EVs were incubated with the proteolytic enzymes trypsin (0.5 mg/mL) and proteinase K (0.5 mg/mL) for 1 h at 37°C and with sodium periodate (10 mg/mL) for 20 h at room temperature, in the dark, to reduce the glycoconjugates surrounding the EVs. After the treatments, EVs were washed twice in PBS by ultracentrifugation at 100,000 xg for 1 h, incubated with the Vero cell cultures for 2 h. The protocol for cell infection was followed as described above.
The specificity of the effects of the EVs isolated from trypomastigotes of T. cruzi Pan4 and the effect of EVs from TcT of the Pan4 strain over the invasion of another T. cruzi strain and intracellular parasite were evaluated. For these experiments, cultures of 5x104 Vero cells were grown the same way as described for the invasion assays using the Pan4 strain. The cells were incubated for 2 h at 37°C with 0.38 μg/mL EVs from Crithidia mellificae or 3T3 cells. After this time, the cells were infected with T. cruzi trypomastigotes of T. cruzi Pan4 (parasite:host cell ratio of 5:1) and after 4 h of interaction, the parasites were removed. The cells were washed and maintained in culture for 24 h, when they were fixed with methanol and stained with Giemsa. Parasitization percentages and indexes were calculated.
The effect of EVs of the Pan4 strain of T. cruzi over the infection of cells with trypomastigotes of T. cruzi 4162 strain (Tc IV) and tachyzoites of T. gondii RH were performed. In this case, cells were incubated for 2 h with EVs of T. cruzi Pan4 and then infected with trypomastigotes of T. cruzi 4162 strain or tachyzoites of T. gondii RH in a parasite:cell ratio of 5:1. After 4 h, the parasites were removed, the cells were washed and maintained in culture for 24 h. Parasitization percentages and indexes were calculated after the evaluation of the cells by Giemsa stain.
To assess the potential capacity of EVs to permeabilize cells, we cultured 5 x 104 Vero cells in 12-well plates as described above. The potential changes in permeability during or after the EVs-cell interaction was evaluated using the Aspergillus giganteus ribotoxin α-sarcin, a ~17 kDa protein that inhibits protein biosynthesis when the cells are previously permeabilized [23–25]. Briefly, after 24 h of culture, cells were washed 3 times with MEM and incubated with 0.38 μg/mL EVs in MEM for 2 h. Cells were washed once and 20 μM of α-sarcin (Sigma, USA) was added for 4 h. After this time, cells were washed 4 times and subsequently incubated with MEM supplemented with 10% IFCS. In a parallel assay, the EVs and α-sarcin were added simultaneously to the cell culture. Viability of the cells was determined using the trypan blue exclusion test as well as the MTT viability assay (Sigma, USA). Cell viability was also determined after the incubation of the cell cultures with EVs, α-sarcin and cells without any treatment as negative controls.
The HL-1 cell line was derived from atrial cardiomyocytes is a cell line that maintains a series of cardiac characteristics such as morphological, biochemical, and electrophysiological properties in vitro [26]. On round 13-mm coverslips, 5x104 cells were grown in Claycomb medium supplemented with 10% IFBS, norepinephrine 0.1 mM, L-glutamine 2 mM and antibiotics. After 24 h of culture, cells were washed and incubated with 0.38 μg/mL EVs in MEM for 2 h. Afterwards, coverslips were washed 3 times and fixed with a solution of 2% paraformaldehyde and 1% glutaraldehyde for 1 h, washed 3 times in PBS and blocked with a solution containing 1% BSA and 0.3 M glycine in PBS, for 1 h. Cells were washed 3 times and incubated with an anti-β2-adrenergic receptor primary antibody (1:500) (Thermo Fisher Scientific, USA) for 1 h. Afterwards, cells were washed 3 times and incubated in the dark, with a goat anti-rabbit IgG antibody conjugated with Alexa Fluor 647 (1:500) (Thermo Fisher Scientific, USA) for 1 h, at 37°C. Finally, samples were washed 4 times, mounted in Vectashield mounting medium with DAPI (Vector Laboratories, USA) and imaged with a Leica DM5500B inverted microscope (Leica Microsystems, Germany).
HL-1 cells cultured and fixed as described above but treated with a solution of NP-40 in 10 mM citric acid (pH 6) were employed as positive control of permeabilization of the assay.
Vero cells were grown overnight in MEM without phenol red plus IFBS and in MEM without phenol red plus IFBS and 2.5 μM EDTA, on μ-slide ibidi multichamber dishes. Cells were washed 3 times in MEM without phenol red and incubated at 37°C for 20 min with Fluo4-AM (Thermo-Fisher, USA) in 1) MEM without phenol red, 2) MEM without phenol red plus 2.5 μM EDTA and 3) a culture medium similar to MEM but without Ca2+ and Mg2+. Fluo-4 is an indicator that exhibits greater fluorescence upon binding intracellular free Ca2+. It presents an AM grouping (acetoxymethyl ester) that, when internalized, is cleaved by intracellular esterases and released to bind to cytoplasmic calcium [20].
After 20 min of incubation of the cells with Fluo-4 AM, EVs of TcT of T. cruzi Pan4 were added to the cells and followed-up in time until 25 min of interaction. Basal controls of fluorescence in cells before the application of the stimulus with EVs were included.
Images were taken every 5 min with a confocal microscope Nikon A1 (Nikon Instruments, The Netherlands) equipped with 10x, 20x multi-immersion, 40x oil, 60x oil, and 60x water objectives and a system of cell incubation at 37°C with enriched atmosphere with 5% CO2. The Fluo4 probe was excited at 494–506 nm and light emission was detected at 516 nm. The fluorescence intensity was analysed and normalized with reference to the basal fluorescence using NIS Elements Software (Nikon Instruments, The Netherlands). The analysis of images was performed using Fiji software (Fiji is Just Image J). Controls of cells incubated with A23187 (a calcium ionophore) and 3-isobutyl-1-methylxantine (IBMX) (a non-specific inhibitor of cAMP and cGMP phosphodiesterase that induces calcium release from intracellular stores) were included.
Cultures of 5x104 Vero cells were seeded over round 13-mm coverslips (as described above) and allowed to attach to the coverslips overnight. The cells were washed 3 times with MEM and incubated in MEM during different times with 0.38 μg/mL EVs in MEM. After this incubation step, coverslips were washed 3 times and fixed with cold acetone (Scharlab, Spain) for 15 min at -20°C. After the fixation step, coverslips were washed 3 times with PBS and permeabilized in a solution of 0.1% Triton X-100 (Sigma, USA) for 10 min. The cells were washed 3 times and blocked with 1% BSA and 0.3 M glycine in PBS for 1 h in PBS. After this time, the cells were washed and incubated with 5 μg/mL of a vimentin polyclonal antibody (1:300) (Thermo Fischer Scientific, USA) for 1 h, washed 4 times and incubated in the dark, with a goat anti-rabbit IgG antibody conjugated with Alexa Fluor 647 (1:500) (Thermo Fisher, USA) for 1 h, at 37°C. The coverslips were washed 4 times and stained with a solution of phalloidin, fluorescein isothiocyanate labelled (Sigma, USA) for 30 min. Samples were finally mounted in Vectashield mounting medium with DAPI (Vector Laboratories, USA) and imaged in a Leica DM5500B inverted microscope (Leica Microsystems, Germany). Images were captured 15 min, 30 min, 120 min, and 24 h after EVs-cell contact. Cells not treated with EVs and cells incubated with the final supernatant from the EVs purification medium were employed as controls.
Vero cells (1x105) were synchronized according to the method described previously by Osuna et al. (1984) [27]. Cells were seeded in 6-well plates with a culture medium with 25 mM thymidine for 12 h, when the medium was replaced with MEM + 10% IFCS. Afterwards, cells were washed with MEM and 1 h later, EVs were added directly to each well. One hour after this EV-cell contact, cells were washed and maintained for 2 and 8 h. Afterwards, the culture medium of the corresponding wells was removed, the cells were washed with PBS, fixed with 70% cold ethanol and incubated with a solution (0.2 M Na2HPO4, 0.1 M citric acid, pH 7.8) for 15 min at 37°C. Cells were then centrifuged, washed with PBS and resuspended in 250 mL of a solution of propidium iodide (40 mg/mL) and RNAse (100 mg/mL) for 30 min at 37°C in the dark, according to Carrasco et al. (2014) [28]. Finally, the samples were analysed in a FACS Calibur (BD Biosciences, San Jose, CA, USA) flow cytometer. The results were analysed with FlowJo software (v 7.6.5, Tree Star, Inc.).
Phosphorylation of the protein Rb after the incubation of cells with EVs was evaluated by immunoblotting. Briefly, 1x105 cells were grown in 6-well plates for 24 h. Cells were washed with MEM and incubated with EVs for 5, 10, 30 and 60 min. After this session, the cells were washed with PBS and lysed in RIPA lysis buffer with a protease inhibitor cocktail (Roche, Switzerland) for 15 min. Cells were harvested with a cell scraper and centrifuged at 14,000 xg for 10 min at 4°C. Supernatants were transferred to new Eppendorf tubes and stored at -20°C. The protein from cell lysates was quantified using the Bradford reagent (Sigma, USA) and 90 μg of protein from cell lysates were resolved by SDS-PAGE, transferred to a nitrocellulose membrane and blocked overnight with 5% non-fat milk in PBS-0.1% Tween 20. Rb (1:2,000) and phospho-Rb (1:1,000) (Cytoskeleton, USA) primary antibodies (Sigma, USA) were incubated overnight at 4°C. Tubulin antibody (1:5,000) (Cytoskeleton, USA) was used as the loading control. The membranes were washed with PBS-0.1% Tween 20 and incubated for 1 h with goat anti-mouse IgGs conjugated with peroxidase (1:1,000) (Dako Agilent Pathology Solutions, USA) in the case of Rb, goat anti-rabbit IgGs conjugated with peroxidase (1:2,000) in the case of phospho-Rb (Dako Agilent Pathology Solutions, USA), and rabbit anti-sheep IgGs conjugated with peroxidase (1:5,000) (Dako Agilent Pathology Solutions, USA) in the case of tubulin. The reaction was also visualized using Clarity ECL Western substrate (BioRad, Spain) in a ChemiDoc Imaging system (BioRad, Spain).
Quantification values represent the means of two or more independent experiments, each performed in triplicate. Means and standard deviations of the EVs size (NTA), percentage of infected cells (invasion assays) and percentage of live cells (permeabilization assays) were calculated. One-factor ANOVAs were performed to detect significant differences between cells treated with EVs and control cells in the case of permeabilization and cell-cycle assays. Multiple post hoc comparisons were performed using the Tukey-Kramer test on GraphPad Prism 5 Software (USA). Values with p<0.0001 were considered statistically significant (***).
After the incubation of 1x107 trypomastigotes of the Pan4 strain for 5 h at 37°C in the culture medium for the release of EVs, 12 μg of total protein of EVs were obtained. The isolation of the EVs by the ultracentrifugation protocol described was evaluated by TEM, NTA and DLS and the presence of surface molecules of T. cruzi in these samples was confirmed by Western blotting (S1 Fig).
Most of the particles visualized by negative staining under TEM were of 30–100 nm size (S1A Fig). Analyses by NTA revealed a majority of EVs with a size of 70.7 ± 7.3 nm (S1C Fig) and a concentration of approximately 5x1010 ± 3.9x109 particles/mL. In the DLS analyses, two populations of EVs with different sizes were observed for TcTs, choanomastigotes and 3T3 cells (S1D, S1E and S1F Fig); EVs of T. cruzi showed a population of 23.05 ± 6.96 nm and a population of 55.74 ± 13.97 nm (S1D Fig). Western blotting confirmed the presence of cruzipain, trans-sialidase and MASPs (SP) in the EVs of T. cruzi Pan4 (S1B Fig).
To evaluate the effect of EVs of TcT from the Pan4 strain in host-cell invasion, we tested different doses measured in total μg/mL of protein. After 2 h of incubation of the cells with the EVs, the infection with T was performed in a ratio 5:1 (T:cell) for 4 h. Counts were performed 24 h later, as described in the Methods section. The results indicate that the parasitism increased the most when the cells were treated with 0.5 μg/mL of EVs. These values in percentage of parasitization (88.88 ± 3.73) significantly differed with respect to the other doses analysed (S2A Fig). From these results, the Effective Dose 50 (ED50) for subsequent trials was set as 0.38 μg/mL.
Regarding the effect of EVs on cells over time (the increase in cell parasitization), these were treated with 0.38 μg/mL of total protein of EVs for 2 h. After the cells were washed three times in medium without serum, they were infected at different time points after the treatment with EVs at a T:cell ratio 5:1, as described above. The results (S2B Fig) show that the difference between the percentage of parasitization of the cells treated with the EVs vs. the percentage of parasitization of non-treated control cells were statistically significant up to 8 h after the treatment. The effects were more evident at 2 and 4 h, when the increase in the percentages was higher compared to untreated cultures. This effect was not appreciated when the cells were infected 24 h after the treatment with the EVs. The parasitization indexes (number of parasites per cell) calculated also differed. For example, in the case of the cells incubated with EVs for 2 h was 2.78 ± 0.55, an index that was twice the parasitization index of the control infected cells without the previous treatment (1.33 ± 0.18).
A series of experiments were performed to study whether the thermal treatment, the treatment with proteolytic enzymes or the reduction of the glycoconjugates surrounding the EVs alters their ability to induce higher levels of parasitization in the host cells with which they interact. Results are shown in S2C and S2D Fig. The thermal treatment of the EVs at 50°C, 70°C and 90°C annuled the action of the EVs on the cells, so the increase in the percentage of parasitization was not observed. The enzymatic treatment with the two proteases employed and the treatment with sodium periodate (for the reduction in the content of carbohydrates of the EVs) also inhibited the capacity of increasing the cell parasitization by trypomastigotes of T. cruzi in the cultures.
Finally, Fig 1A, 1B and 1C show the effect of the incubation of cells with EVs of C. mellificae and 3T3 cells prior to the infection with T. cruzi Pan 4 and the increase in the infection of cells when these are incubated with EVs isolated from trypomastigotes of the Pan4 strain and then infected with TcT of the 4162 strain or tachyzoites of T. gondii RH. In Fig 1A is possible to observe that the incubation of cells with EVs from another source different than T. cruzi did not boosted an increase in the percentage of infected cells as it happens when the cells are in contact with EVs of T. cruzi Pan4 prior to the infection. When the cells were incubated with EVs of C. mellificae and 3T3 cells, the percentage of infected cells obtained were 29.3 ± 5.0 and 29.8 ± 4.3, respectively. These results did not differ significantly from the results obtained in the case of the cells infected only with TcT of T. cruzi Pan4 without the previous treatment with the EVs (36.5 ± 3.8). In these experiments, the parasitization indexes obtained were 1.31 ± 0.08 in the cells incubated with EVs of C. mellificae, 1.79 ± 0.28 in the cells incubated with EVs of the 3T3 cell line and 1.54 ± 0.31 in the cells infected with TcT without the prior incubation with EVs, but it rose to 2.60 ± 0.14 when the cells were previously treated with EVs of T. cruzi Pan4. This Figure also shows that this increase in the parasitization percentage and index of cells is independent of the strain of T. cruzi employed to infect the cells, as we proved that the infection with TcT of a strain that is classified as Tc IV almost doubled the percentage of infected cells not incubated previously with the EVs (53.5 ± 6.0 vs 32 ± 2.6; parasitization indexes: 2.10 ± 0.14 vs 1.39 ± 0.21, respectively).
In the case of the cells treated with EVs of T. cruzi Pan4 prior to the infection with tachyzoites of T. gondii RH, in Fig 1B is possible to observe a slight, non-significant increase in the percentage of infected cells in those incubated with EVs of T. cruzi prior to the infection with tachyzoites, when compared to the cells infected with the parasite without the previous incubation with the EVs of T. cruzi (56.50 ± 4.80 vs. 46.50 ± 4.93).
A toxin of Aspergillus giganteus, α-sarcin, constitutes a ribotoxin with a molecular weight of 16.8 kDa that acts at the ribosomal level, inhibiting the protein synthesis. This toxin has been used to determine the potential permeabilization induced in the cells on the entry of certain viruses [29]. The cells were treated in two ways: i) they were simultaneously incubated with EVs of T. cruzi Pan4 and α-sarcin or ii) they were incubated with EVs of T. cruzi Pan4 for 2 h and then with α-sarcin for 4 h. After 24 h of the treatment, the trypan blue exclusion assay was performed. Vero cells treated with EVs and α-sarcin registered mortality percentages of 76.10 ± 6.81 (simultaneous incubation) and 82.20 ± 10.17 (separated incubation of EVs and the toxin). The control cells incubated only with the toxin or with the EVs showed mortality percentages of 16.90 ± 4.10 and 17.23 ± 7.73, respectively, while the percentage of viability of the untreated control culture cells was 15.56 ± 7.67. These results were confirmed using the MTT cell-viability assay, as shown in Fig 2A. In this case, the percentage of viability of the cells treated with EVs and the toxin was 37.36 ± 15.39, while the same percentages in the cells treated only with the toxin or the EVs alone were 90.10 ± 7.03 and 95.30 ± 3.30, respectively. The percentage of viability in the untreated control cell cultures was 99,0 ± 0.1. Fig 2B shows the appearance of the different cell monolayers preincubated with EVs and treated with α-sarcin and the different control cells at 24 h of the treatment.
The beta-2 adrenergic receptor (β2 adrenoreceptor), also known as ADRB2, is a cell membrane-spanning beta-adrenergic receptor that binds epinephrine or adrenaline, whose signalling, via a downstream L-type calcium channel interaction, mediates physiological responses such as smooth-muscle relaxation and bronchodilation [30]. Using an antibody that recognizes an epitope between the amino acids 340–413 (which corresponds to the intracytoplasmic domain of the receptor), we demonstrated that the EVs of T. cruzi can alter and permeabilize the cell membrane, exposing the epitope to the antibody, as shown in Fig 2C. In the case of the untreated cells, this effect was not detected. An image similar to that of the cells treated with the EVs resulted in the cultures that, prior to the incubation with the antibody, were permeabilized with a solution containing NP-40.
The time-course measurements of the intracellular calcium levels of Vero cells treated with the EVs of T. cruzi Pan4 are shown in Fig 3. Results show that, when the cells were incubated in a culture medium with Ca2+ and Mg2+, the fluorescence levels increased up to 3.83 ± 0.62 times the initial values as soon as 10 minutes of interaction. Moreover, when the interaction was performed in a culture medium depleted of Ca2+ and Mg2+, there was also a progressive increase in the fluorescence levels. For example, at 10 min of the interaction, there was a 1.40 ± 0.39-fold increase in the fluorescence levels of the cells, which could correspond to a mobilization of the ions from their intracellular deposits to the cytoplasm.
In the case of the cells incubated in MEM with EDTA, a calcium chelator, there was also a 1.69 ± 0.01 increase in the fluorescence levels at 10 minutes of incubation, when the fluorescence intensity started to decrease. At this point, it is possible to observe a different pattern of distribution of the fluorescence, with the appearance of a more granulated cytoplasm.
The calcium ionophore A23187 was used as the control for the assays of the cells incubated in the medium containing Ca2+ and Mg2+ and prompted a 64.80-fold rise in the fluorescence levels at 25 min. The cAMP phosphodiesterase inhibitor, 3-isobutyl-1-methylxanthine (IBMX) was used as the control for the induction of ion output from the intracellular Ca2+ deposits to the cytoplasm and prompted a 2.37-fold rise in the fluorescence levels at 10 min.
The analysis of the effect of the EVs of T. cruzi Pan4 over the actin cytoskeleton is shown in Fig 4A, 4B and 4C. The disorganization of the actin filaments was visible in the cells showing greater globular actin (GA) from 15 min up to 120 min after the treatment with EVs (Fig 4A, 4B and 4C). On the other hand, vimentin appeared to withdraw from the areas in the cytoplasm where the actin is disorganized and concentrated in the parts of the cytoplasm where GA is less patent. The morphology of the treated cells appeared to be altered and filopodia (F) were visible, giving a dendritic aspect to the cell. These effects were reversible 24 h after the treatment; both the cytoskeleton images and the morphology of the cells incubated with EVs were similar to that of the control cells not treated with EVs.
The influence of the EVs of T. cruzi Pan4 on Vero cells cycle was analysed with Vero cells previously synchronized in the S phase and treated as described in Methods. The changes in the cell cycle were analysed by flow cytometry 2 and 8 h after the addition of the EVs. Fig 5A shows the percentage of cells in the different phases of the cell cycle. At 8 h after the addition of the EVs, the percentage of cells increased at phases G0/G1 and decreased at phase S.
Fig 5B and 5C show the results of the levels of the protein of retinoblastoma (pRb) in its phosphorylated and non-phosphorilated states, in the cells treated with EVs of T. cruzi Pan4 and the untreated control cells. The protein expression increased for the phosphorylated pRb from the first minutes of the interaction of the EVs with the cells, reaching the maximum phosphorylated state at 15 min of the interaction and declining to values similar to those of the control cells at 60 min of treatment.
The cell-invasion process of T. cruzi has been widely studied. Numerous mechanisms are known to be involved in preparing the cell to induce the endocytosis of trypomastigotes into non-phagocytic cells [31–36] and among the natural agents that induce massive entry and cellular infection are the EVs secreted by trypomastigotes.
EVs from trypomastigotes of the Pan4 strain were first isolated in 2016, when de Pablos et al. confirmed that the C-terminal region of MASPs proteins is present in the EVs secreted by the trypomastigotes derived from tissue-culture cells [18]. Following the methodology described above using differential centrifugation and then Nanoparticle Tracking analysis, we obtained an homogeneous population of EVs under our experimental conditions. Also, the yield was higher than that of other authors using others strains or forms of the parasite. Trocoli Torrecilhas et al. (2009) employed 5 μg of protein from 1x105 trypomastigotes of the Y strain [15]. Garcia Silva et al. (2014) have reported a protein yield of the vesicular fraction of 1.2 μg per 1x1010 epimastigotes from the DM28c strain [14]. In our case, we found 12 μg of protein of EVs after the incubation of 1x107 trypomastigotes of the Pan4 strain for 5 h at 37°C in the culture medium for the release of EVs.
To track the time course of the effect of EVs in increasing cell parasitism, we performed infections at different times after incubating the cells with EVs and we found that the cells are still susceptible to increased parasitism at least for 8 h after the treatment (p <0.0001). However, 24 h after the incubation with the EVs, the percentage of infected cells in the treated and non-treated cultures were similar (S2B Fig). On the other hand, we observed that increasing amounts of EVs can increase the percentage of infected cells with sigmoidal kinetics, reaching a maximum of 89% infected cells (0.5 μg/mL) and ED50 of 0.38 μg/mL, a dose that was employed in subsequent experiments. In 2014, Garcia Silva et al. observed that the treatment with small amounts of EVs (160 ng) can prevent cells from appearing oversaturated in a tRNAGlu-derived 5′ halves visualization analysis by FISH [14]. In the same study, the authors determined that 30 min after the treatment of cells with EVs, a diffuse fluorescent cytoplasmic pattern appeared, which becomes granular after 2 h of treatment. Regarding the interaction and posterior infection of the cells after the treatment with the EVs, Cestari et al. (2012) pre-incubated Vero cells for 30 min before adding the trypomastigotes of the Sylvio X10/6, DTU I strain [17]. This reflects that the incubation conditions with respect to the amount of EVs used and the incubation time vary among the different research groups, and it should be taken into account that the conditions selected by a researcher do not necessarily correspond to the conditions of a natural infection [14].
It has been demonstrated in vitro (in non-phagocytic cells and monocytes), as well as in vivo, that EVs of T. cruzi increase the number of infected cells [15,17,37–38]. Under our experimental conditions, cells pretreated with EVs of T. cruzi Pan4 registered infection percentages of over 3.5-fold higher than for cells infected without the prior incubation with the EVs. The parasitization index (number of parasite per cell) was also two-fold that of the parasitization index of the control infected cells without prior treatment. In 2009, Trocolli Torrecilhas et al. reported that T. cruzi trypomastigotes invade 5-fold more susceptible cells when these were preincubated with purified parasite EVs [15]. Cestari et al. (2012) also demonstrated that THP-1 derived plasma membrane vesicles (ectosomes) could simultaneously induce an increase in Vero cell invasion [17]. This invasion was dose dependent, non-specific for parasite strain or eukaryotic cell line, and dependent on the parasite infective stage. We also proved that the increase in parasitization is specific to T. cruzi trypomastigotes but non-specific for the parasite strain and that the incubation of cells with EVs from another trypanosomatid species or those from eukaryotic cells didn´t increased the percentages of parasitization.
The interaction EVs-cell and the latter activity of EVs appears to depend on their binding through lectins to the plasma membrane and on a presumably enzymatic protein activity, given that the thermal and chemical treatments of EVs with trypsin, proteinase K, and sodium periodate drastically reduced parasitism of the cells with which they interacted. During the adhesion and invasion process of T to the host cells, a number of glycosylated molecules are expressed on the surface of the parasite. Examples include mucins, trans-sialidases, MASPs and the gp85 family of proteins [35]. Glycosylated proteins have also been detected in EVs by proteomic analyses [9,39] some with important activities and biological significance, such as trans-sialidases (TS/SAPA, TC85, gp82, gp90, CRP) [40], cruzipain [41], gp63, MASPs [42], and other types of mucins [43]. As these proteins are located on the surface of the EVs, they can bind specifically to the proteins in the plasma membrane and this would explain why the reduction of the carbohydrates of the EVs after the sodium periodate treatment can affect the binding of EVs to the surface of the cells. On the other hand, some of these glycoproteins have enzymatic activities essential for the interaction of trypomastigotes during the invasion process and maybe in the interaction of EVs with the plasma membrane.
It has been reported that the infection with some type of viruses lead to a permeabilization mechanism where the plasma membrane allows the entry of some high molecular weight molecules such as α-sarcin (16.8 KDa) [44–46], toxin that lacks a membrane receptor, unlike other toxins that affect the protein synthesis and are internalized via endocytosis [45,47]. The same effect was observed incubating cells with EVs of trypomastigotes of T. cruzi, as our results demonstrate that they can permeabilize Vero and HL-1 cell lines. Cell counts with trypan blue 24 h after the preincubation of cells with EVs and subsequent incubation with α-sarcin registered as much as 76.10% mortality. The control cells incubated only with α-sarcin, only with EVs, and without either showed percentages of cell death of 16.90%, 17.23%, and 15.56%, respectively. The percentages of mortality (100 - % of viability in Fig 2) of the cells using MTT were 62.64% in the case of the cells previously incubated with EVs and 9.90% in cells incubated only with the toxin. In 1990, Castanys et al. have reported that infective metacyclic forms of T. cruzi secrete a glycoprotein involved in cell permeabilization that enabled the entrance of molecules such as α-sarcin [24].
Permeabilization was also evaluated in HL-1 cardiac muscle cells with confocal microscopy, using an antibody directed to an epitope of the β2-adrenergic receptor located in the intracytoplasmic region of the receptor anchoring. In this experiment, fluorescence was detected in the cytoplasm of cells previously incubated with EVs and in cells not treated with EVs but permeabilized with the detergent NP-40. This permeabilization may be the result of changes in the cell membrane that allow the direct entry of the antibodies into the cytoplasm or by a transient disorganization of the membrane, capable of exposing the antigens present in inside the cell, with consequent exposure to the immune system. The presence of autoantibodies against β-adrenergic receptors in the serum of chagasic patients has been reported [48–51], although authors have related such emergence to the recognition of exposed parts in the membrane due to cross reactivity to ribosomal acidic proteins P0 of the parasite [52]. A study related to the recognition of epitopes of the β-adrenergic receptors inserted into the inner side of the membrane by autoantibodies would be necessary to confirm the possible hypothesis that the permeabilization of the cardiac cells by the parasitic EVs lead to the exposure of these receptors to the immune system and then elicit the production of autoantibodies.
The recognition and the invasion processes of the trypomastigotes to the host cells involve molecules over the surface of both the trypomastigote and the host cell. A ligand-receptor recognition occurs and generates in both a series of events that raises the intracellular Ca2+ levels due to the mobilization of the ions from the endoplasmic reticulum and the mitochondria [53–57]. This boost in calcium levels is also responsible for higher cAMP levels [58], facilitating the release of EVs and their later fusion with the plasma membrane [59]. The treatment of Vero cells with EVs of T. cruzi Pan4 raised cytoplasmic Ca2+ levels and to determine the kinetics and origin of these Ca2+ ions, we treated cells with EVs and studied the result every 5 min under confocal microscopy. From 5 min of the treatment, fluorescence intensified in both culture media (with and without calcium). This implies that the contact of Vero cells with EVs raises cytoplasmic Ca2+ levels that could come both from the intracellular deposits of calcium and the extracellular medium. The fluorescence pattern detected resembles the one in control cells treated with the xanthine IBMX for the Ca2+ mobilization from the intracellular deposits [60] or when the cells were incubated with the ionophore A23187 [61], a compound that allows Ca2+ to enter cells from the culture medium.
It has been demonstrated in different types of eukaryotic cells that the intracellular levels of calcium induce an asymmetric distribution of phospholipids in the plasma membrane by the activation of the enzymes scramblase and floppase. Then, phosphatidylserine and phosphatidylethanolamine are exposed in the outer side of the membrane and contribute to the activation of Ca2+-dependent proteases, followed by the release of EVs [62]. The exposure of anionic phospholipids to EVs strengthens the fusogenic properties of these vesicles, which could be a prerequisite for the release of the EV content. This could mean that the higher intracellular calcium levels and the changes in the distribution of phospholipids could explain the permeabilization induced in the host cell after the treatment with the EVs of T. cruzi, allowing the entrance of a toxin of ~17 kDa such as α-sarcin. Moreover, increases in intracellular Ca2+ could trigger a greater release of EVs from the cells exposed to the parasite [3,17,63], as more calcium prompts a strong response of EV release in other cell lines [64–65].
Calcium ions also contribute to the reorganization of the cytoskeleton through the activation of cytoplasmic proteins such as calpain and gelsolin. These proteins cut the actin cytoskeleton protein network, allowing membrane budding and removing capping proteins at the end of the actin filaments [66–67]. A disruption in actin filaments and vimentin at the time of the invasion of cells with trypomastigotes has been demonstrated [68–69] using drugs like cytochalasin B and latrunculin, which affect the cytoskeletal structure and functions and, therefore, the entrance of the parasite in non-phagocytic cells [32–33,70]. It has been mentioned that the increase in intracellular Ca2+ leads to a rapid and transient reorganization of host-cell microfilaments, including the disassembly of the actin cytoskeleton, which is important for the entry of T into the host cells [71–73]. Studying the gene-expression changes caused by microvesicles of T. cruzi epimastigotes of the DM28c strain in mammalian host cells, Garcia Silva et al. (2014) observed an induction of a broad response, including the modification of the host-cell cytoskeleton and the extracellular matrix [74]. In fact, the regulation of actin cytoskeleton is one of the pathways identified as being affected by EV treatment in the profile of transcriptome changes [74]. Noting increased fluorescence in cells incubated with EVs in the presence of Fluo-4AM, we suspected that these changes in the Ca2+ mobilization induced by EVs could directly affect the actin cytoskeleton, as happens when the parasite begins to invade the host cell. Our results showed a clear disruption of host-cell actin from 15 min after the incubation with EVs, an effect that remains at 120 min but not 24 h after the treatment with EVs. Ferreira et al. (2006) indicated that different strains of Mt of T. cruzi can invade host cells through both actin cytoskeleton-dependent and independent routes, by engaging different surface molecules for attachment while triggering different signal-transduction pathways [34]. For example, host-cell invasion by the strain CL Mt, mediated mainly by the surface molecule gp82, is associated with F-actin disassembly whereas the G strain is gp35/50-mediated invasion by strain G depends on target-cell actin cytoskeleton [34,75].
The analysis of the cell cycle events revealed how at 8 h after the addition of the EVs the percentage of cells in each of the cycle phases significantly differed when compared to control values, showing an arrest of the cell cycle in the G0/G1 phases. Previous observations regarding the in vitro life cycle of T. cruzi in cultured cells demonstrated a low cell-division rate among cells infected with the parasite [76–78]. In this regard, Ca2+ may be responsible for the cell cycle changes, as they act as second messengers in the control of the cell cycle. Thus, Ca2+/calmodulin activate the complex CDK4/cyclin D1, which regulates the protein of Retinoblastoma (pRb1), the main inhibitor of the DNA synthesis [79]. From our results, it is evident that the phosphorylation of the protein Rb takes place from the first few min of the EVs/cell interaction. Here, phosphorylated pRb increased rapidly, while in the cells without the treatment with EVs no such change was detected (Fig 5B and 5C). However, at 60 min of treatment with EVs, these increases in phosphorylation returned to normal levels. This apparently arrested synthesis, preventing the cells from entering phase S of the cycle. Moreover, a series of “calcium sensors” present in the cell cytoplasm, such as the stromal interaction molecule 1 (STIM1), is involved in the progression of mitosis. Cells lacking this protein may arrest the cells in phases G0/G1, as occurred in our experiments. This implies that this protein is required for the progression of the cells in the phase of DNA synthesis or phase S [80]. The arrest of cells in the G0/G1 phases exerted by EVs of T. cruzi Pan4 was possibly caused by increased expression of cyclin-dependent kinase inhibitor p21 and the subsequent decrease of phosphorylated protein (pRb) [81]. The higher intracellular calcium levels were also involved in cell-cycle events. In fact, the indirect role of high levels of calcium in cell arrest in these phases has been examined by Wu et al. (2006) [82], who employed capsaicin and blocked the cell cycle in the previous phase of DNA synthesis. This effect was reversed with BAPTA, an intracellular Ca2+ chelator. Together with the rises of intracellular calcium levels, and because of these high levels, these researchers have recently questioned the role of actin networks nucleated by the complex Arp2/3 in the signalling events necessary for the progression of the cell cycle in non-transformed cells [82–83] and demonstrated that Arp2/3 is not able to act as a sensor for the start of the phase S in the cell cycle per se, such as the actin filaments. Previous studies have shown that the use of cytochalasin B at very low doses detains the cell cycle in phases G0/G1 as in our experiments with EVs while the inhibitors that act in the polymerization of actin stopped the cell cycle before the cytokinesis [84–86].
In conclusion, it has been shown that the incubation of cells with EVs of TcT of T. cruzi Pan4 strain induce a number of changes in the host cells that include 1) a change in cell permeability, and 2) higher intracellular levels of Ca2+ that can alter the dynamics of the actin cytoskeleton and arrest the cell cycle at G0/G1 prior to the DNA synthesis necessary to complete mitosis. In the end, these changes induced by the EVs aid their invasion of host cells, augment the percentage of cell parasitization, and possibly cause some characteristic manifestations of Chagas disease.
|
10.1371/journal.pbio.3000046 | Fine-tuned adaptation of embryo–endometrium pairs at implantation revealed by transcriptome analyses in Bos taurus | Interactions between embryo and endometrium at implantation are critical for the progression of pregnancy. These reciprocal actions involve exchange of paracrine signals that govern implantation and placentation. However, it remains unknown how these interactions between the conceptus and the endometrium are coordinated at the level of an individual pregnancy. Under the hypothesis that gene expression in endometrium is dependent on gene expression of extraembryonic tissues and genes expressed in extraembryonic tissues are dependent of genes expressed in the endometrium, we performed an integrative analysis of transcriptome profiles of paired extraembryonic tissue and endometria obtained from cattle (Bos taurus) pregnancies initiated by artificial insemination. We quantified strong dependence (|r| > 0.95, empirical false discovery rate [eFDR] < 0.01) in transcript abundance of genes expressed in the extraembryonic tissues and genes expressed in the endometrium. The profiles of connectivity revealed distinct coexpression patterns of extraembryonic tissues with caruncular and intercaruncular areas of the endometrium. Notably, a subset of highly coexpressed genes between extraembryonic tissue (n = 229) and caruncular areas of the endometrium (n = 218, r > 0.9999, eFDR < 0.001) revealed a blueprint of gene expression specific to each pregnancy. Gene ontology analyses of genes coexpressed between extraembryonic tissue and endometrium revealed significantly enriched modules with critical contribution for implantation and placentation, including “in utero embryonic development,” “placenta development,” and “regulation of transcription.” Coexpressing modules were remarkably specific to caruncular or intercaruncular areas of the endometrium. The quantitative association between genes expressed in extraembryonic tissue and endometrium emphasize a coordinated communication between these two entities in mammals. We provide evidence that implantation in mammalian pregnancy relies on the ability of the extraembryonic tissue and the endometrium to develop a fine-tuned adaptive response characteristic of each pregnancy.
| Implantation in mammals requires a complex crosstalk between the conceptus (the embryo and associated membranes) and the uterus. An imbalanced regulation of the factors contributing to these interactions has negative impacts on the attachment of the fetus, the progression of the pregnancy, and the progeny. Focusing on paired conceptus–endometrium analyses of individual pregnancies in cows, we have determined that communication at implantation encompasses synchronized genome-wide coregulation of genes. Gene regulatory interactions between one conceptus and the surrounding maternal tissue vary between endometrial regions containing or lacking glands. Our data reveal new insights, to our knowledge, on the coordination of molecular mechanisms that contribute to implantation and pregnancy establishment in mammals. We conclude that the biological response of the endometrium is embryo-specific, a phenomenon that deserves further investigation in the context of assisted reproductive technologies.
| In mammals, pregnancy recognition requires a tightly synchronized exchange of signals between the competent embryo and the receptive endometrium. The initiation of this signaling is triggered by key factors produced by the conceptus [1, 2], which are translated by the endometrial cells into actions that will condition the trajectory of embryo development as well as progeny phenotype. In mammalian species, including human, rodents, and ruminants, the delicate balance in embryo–maternal communication is affected by the way the embryos are generated (natural mating, artificial insemination, in vitro fertilization, or somatic cell nuclear transfer) and by the sensor-driver properties of the endometrium defined by intrinsic maternal factors (e.g., maternal metabolism, aging) and environmental perturbations (e.g., pathogens, nutrition) [3–5]. The concept of sensor property applied to the mammalian endometrium was proposed in a former study accompanied by the notion of endometrial plasticity [6]. This property was recently confirmed in vitro with an aberrant responsiveness of human endometrial stromal cultured cells in the context of recurrent pregnancy loss [7]. Nevertheless, it remains unaddressed whether the mammalian endometrium is able to develop an adaptive embryo-tailored response in a normal pregnancy.
In mammalian reproduction, sheep and cattle are research models that have relevantly contributed key insights to the understanding of molecular and physiological pregnancy-associated mechanisms, including the deciphering of embryo–endometrium interactions [8, 9]. In the bovine species, by gestation days 7–8, the blastocyst enters the uterine lumen. After hatching by days 8–9, the outer monolayer of trophectoderm cells establishes direct contact with the luminal epithelium of the endometrium [10]. On gestation days 12–13, the blastocyst is ovoid in shape (approximately 2–5 mm) and transitions into a tubular shape by days 14–15. Next, the conceptus begins to elongate via proliferation of the trophectoderm and parietal endoderm cells [11]. The bovine extraembryonic tissue reaches 30 cm or more in length by days 19–20 [11, 12], and the trophectoderm begins to attach to the luminal epithelium (LE) of the endometrium, which marks the beginning of the attachment and onset of placentation [11].
By approximately day 15, rapidly proliferating trophectoderm cells of the extraembryonic tissues synthesize and release interferon tau (IFNT) [12–16], which is the major pregnancy recognition signal in ruminants [1, 9, 17, 18]. The disrupted release of the oxytocin-dependent pulses of prostaglandin F2 alpha [19] allows maintenance of progesterone production by a functional corpus luteum [19], which is critical for the establishment and progression of pregnancy [1, 4, 9, 12, 15, 16, 20]. IFNT actions include induction of numerous classical and nonclassical IFNT-stimulated genes and stimulation of progesterone-induced genes that encode proteins involved in conceptus elongation and implantation [4]. IFNT-regulated genes have diverse actions in the endometrium that are essential for conceptus survival and pregnancy establishment [12]. Other paracrine signals such as prostaglandins and cortisol have regulatory effects on conceptus elongation and endometrium remodeling [21]. More recently, the identification of potential ligand-receptor interactions between the conceptus and endometrium [22] and the secretion of proteins and RNAs through exosomes [23, 24] have expanded the field of possibilities by which the conceptus and endometrium interact prior to and during implantation.
The crosstalk between the conceptus and the endometrium is associated with the expression and regulation of a wealth of genes in each entity [25, 26]. The nature of the conceptus modifies gene expression of the endometrium in cattle [6, 27, 28] and decidualizing human endometrial stromal cells [29]. Similarly, the endometrium from dams with different fertility potentials [30] or metabolic status [31] influences the gene expression of the conceptus. Despite the growing evidence of the interactions between conceptus and endometrium at the level of gene regulation, the pathways and the functions that result from this interaction have yet to be unveiled. Furthermore, the lack of integrated analysis between paired conceptus and endometrium has made it challenging to advance our understanding of the functional interactions between these two entities in normal pregnancies.
Here, we hypothesized that gene expression of extraembryonic tissue is not independent from gene expression of endometrium. In the present study, we carried out an integrative analysis of transcriptome profiles of paired conceptuses and endometria at the onset of implantation, aiming at the identification of regulatory pathways that have coordinated expression between the conceptus and endometrium in normal pregnancies. Surprisingly, our results show that at gestation day 18 in cattle, several hundred genes have an expression profile in the conceptus and caruncular areas of the endometrium that is unique to each pregnancy. Analyses of genes coexpressed between the conceptus and the paired-associated endometrium revealed significantly enriched gene coexpression modules for specific biological processes with critical contribution for implantation and placentation. Our data provide evidence that successful implantation in mammalian pregnancy relies on the ability of the endometrium to elicit a fine-tuned adaptive response to the conceptus.
We analyzed the RNA-sequencing (RNA-seq) data, which consisted of samples collected from five cattle pregnancies terminated at gestation day 18 (Gene Expression Omnibus database GSE74152 [27]). The conceptus was dissected, and transcriptome data were generated for extraembryonic tissue, whereas the endometrium was dissected into caruncular (gland-free) and intercaruncular (containing endometrial glands) areas, and transcriptome data were generated from both regions of the endometrium (Fig 1A). Therefore, the data set analyzed was comprised of three sample types collected from each pregnancy: extraembryonic, caruncular, and intercaruncular tissues (Fig 1B). Alignment of the sequences to the B. taurus genome (University of Maryland [UMD] assembly 3.1) resulted into an average of 22, 31.4, and 34.6 million uniquely mapped reads for extraembryonic (n = 5), caruncular (n = 5), and intercaruncular (n = 5) tissues, respectively. After filtering for lowly expressed genes, we estimated the transcript abundance of 9,548, 13,047, and 13,051 genes in extraembryonic, caruncular, and intercaruncular tissues, respectively (Fig 1C). Unsupervised clustering of the samples based on their transcriptome data separated the samples obtained from the extraembryonic tissue from the endometrial samples and further distinguished caruncular from intercaruncular endometrial samples (Fig 1D).
The associated expression between two genes can be assessed by correlative metrics [32] within [33, 34] or between tissues [34, 35]. Thus, we calculated Pearson’s coefficient of correlation (r [36]) to test whether there is association between the transcript abundance of genes expressed in extraembryonic tissue and endometrium (caruncular or intercaruncular tissues). We reasoned that under a null hypothesis, the abundance of a gene expressed in extraembryonic tissue (Gj) would have no association with the abundance of a gene expressed in endometrium (Gk, or Gl), e.g., H0:r(Gj,Gk) ≈ 0. On the other hand, under the alternative hypothesis (H1:r(Gj,Gk) ≠ 0), two genes display coexpression [36].
The distribution of correlation coefficients for all pairs of genes expressed in extraembryonic and caruncular tissues averaged 0.13 (Fig 2A), and the equivalent distribution obtained for all pairs of genes expressed in extraembryonic and intercaruncular tissues averaged 0.03 (Fig 2B). Both distributions deviated significantly from a distribution obtained from shuffled data that disrupted the pairing of the extraembryonic tissue and endometrium (P < 2.2−16, S1 Fig). We calculated the empirical false discovery rate (eFDR) and noted that absolute correlation coefficients in both distributions were highly significant when greater than 0.95 (eFDR < 0.007, S2 Fig and S1 Table). The pairs of genes presenting significant correlation on the paired data rarely reoccurred when we scrambled the pregnancy pairs (S1 Table). Of note, S3 and S4 Figs present examples of pairs of genes we identified with the highest positive and negative correlation coefficients, which fit the alternative hypothesis (H1:r(Gj,Gk) ≠ 0), and examples of pairs of genes that show correlation coefficients close to zero, fitting the null hypothesis (H0:r(Gj,Gk) ≈ 0).
The distribution of degrees of connectivity for significant correlations (|r| > 0.95, eFDR < 0.01) between extraembryonic and caruncular tissues was not equivalent to the distribution observed between extraembryonic and intercaruncular tissues (P < 2.2−16). On average, genes expressed in extraembryonic tissue were significantly correlated with 295 genes expressed in caruncular tissues (median = 101). Eleven genes were significantly correlated with over 2,300 genes in caruncular tissues (i.e., amphiregulin [AREG], early growth response 1 [EGR1], peroxisomal biogenesis factor 3 [PEX3], gigaxonin [GAN], S5A Fig). On average, genes expressed in extraembryonic tissue were significantly correlated with 266 genes expressed in intercaruncular tissues (median = 252). Eight genes were significantly correlated with over 750 genes in intercaruncular tissues (i.e., wingless/Integrated family member 5B [WNT5B], WNT7B, receptor-tyrosine-kinase–like orphan receptor 2 [ROR2], dipeptidase 1 [DPEP1], gap junction protein beta 3 [GJB3], S5B Fig). These results strongly suggest different patterns of gene coexpression between extraembryonic and caruncular or intercaruncular tissues.
When considering highly significant correlations (r > 0.99 or r < -0.99, eFDR < 0.001), notably, over 99% of the genes expressed in extraembryonic tissue were positively or negatively correlated with genes expressed in intercaruncular tissues (Fig 2C). Of the genes expressed in extraembryonic tissues, 93% and 87% were positively or negatively correlated with genes expressed in intercaruncular tissues, respectively. Of the genes expressed in caruncular tissues, 31% and 67% negatively or positively correlated with genes expressed in extraembryonic tissues, respectively. Similarly, 50% and 54% of the genes expressed in intercaruncular tissues were negatively and positively correlated with genes expressed in extraembryonic tissues, respectively (Fig 2C). These gene pairs rarely maintained their highly significant correlation when the pregnancy pair was disrupted (S1 Table). Thus, highly significant coexpression between thousands of genes is a consequence of the interaction between the conceptus and the endometrium.
We then examined whether genes coexpressed in extraembryonic tissue and endometrium have expression patterns that are unique to pregnancies. We identified 229 and 218 genes expressed in extraembryonic and caruncular tissues, respectively (|r| > 0.9999, eFDR < 0.0001, S1 Table), whose expression profiles produced equivalent dendrograms for extraembryonic and caruncular tissues independently (P = 0.008, Fig 2D). This set of genes consisted of 223 and 212 genes expressed exclusively in extraembryonic and caruncular tissue, respectively, and six genes that were expressed in both compartments. At this level of significance, no gene pairs retained their correlation in the shuffled data (S1 Table).
Gene ontology analysis of these 441 genes identified significant enrichment in the biological processes “mRNA processing” (gem-nuclear-organelle–associated protein 6 [GEMIN6]; pre-mRNA processing factor 4B [PRPF4B]; RNA-binding motif protein 39 [RBM39]; survival motor neuron domain containing 1 [SMNDC1]; SPT4 homolog, DSIF elongation factor subunit [SUPT4H1]; splicing factor U2AF 26-kDa subunit [U2AF1L4]; FDR = 0.13, Fig 2E), “chromatin organization” (codanin 1 [CDAN1], nucleoporin 133 [NUP133], SUPT4H1, FDR = 0.13, Fig 2E), and “protein autoubiquitination” (CCR4-NOT transcription complex subunit 4 [CNOT4], membrane-associated ring-CH–type finger 5 [MARCH5], ubiquitin-like with PHD and ring finger domains 1 [UHRF1]). We also interrogated the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways database and identified an enrichment for the “RNA transport” pathway (eukaryotic translation initiation factor 4E binding protein 1 [EIF4E], GEMIN6, karyopherin subunit beta 1 [KPNB1], protein mago nashi homolog 2 [MAGOHB], NUP133, NUP54, PYM homolog 1, exon-junction-complex–associated factor [PYM1], SUMO1/sentrin/SMT3-specific peptidase 2 [SENP2], small ubiquitin-like modifier 1 [SUMO1], THO complex 1 [THOC1], FDR = 0.06, Fig 2F). A bootstrapping approach (2,000 randomizations of 441 genes) showed a probability <0.001 of these two biological processes being enriched by chance (S6A Fig). Similarly, there was <0.001 probability of the “RNA transport” pathway to have been enriched by chance (S6B Fig). We did not identify groups of genes coexpressed in extraembryonic and intercaruncular tissues capable of producing dendrograms that mirrored each other. These results demonstrate that genes highly coexpressed between extraembryonic and caruncular tissues form a signature that independently distinguishes pregnancies in an equivalent manner.
Our analysis was not an exhaustive evaluation of all potential coexpression networks that exist between extraembryonic tissue and the endometrium. Thus, we developed a web interface for dynamic and interactive data visualization based on the coexpression analysis conducted in the present study [37, 38] (https://biaselab.shinyapps.io/eet_endo/). The public access to this web application allows a user to produce networks for genes of their choosing. Furthermore, each network is accompanied by supporting data such as scatter plots and heatmaps of the gene-expression values. The raw data and codes for reproduction of this interface can be downloaded from a GitHub repository (https://github.com/BiaseLab/eet_endo_gene_interaction).
We investigated the transcriptome-wide interactions between extraembryonic and caruncular or intercaruncular tissues independently. The clustering of genes based on coexpression is a powerful means to understand coordinated gene functions [39]; thus, we used the matrix with correlation coefficients to cluster extraembryonic, caruncular, and intercaruncular tissues independently.
The heatmap resulting from clustering the two data sets (extraembryonic and caruncular tissues) showed the formation of an organized coexpression network between the genes expressed in extraembryonic and caruncular tissues (Fig 3A). We identified 36 clusters formed by the genes expressed in extraembryonic tissue that presented enrichment for several biological processes (FDR < 0.2, Fig 3B), in which we identified several genes expressed in extraembryonic tissue significantly coexpressed with genes expressed in caruncular tissues (see S1 Data for a list of genes). For instance, 142 genes associated with regulation of transcription were identified across clusters 1, 12, 30, 38, and 54. Eighty-two genes were associated with signal transduction across clusters 1, 21, 27, and 71. Interestingly, 26 genes associated with “in utero embryonic development” were identified in cluster 1.
The clustering of genes expressed in caruncular tissues according to their coexpression with extraembryonic tissue genes resulted in the identification of 32 clusters presenting enrichment (FDR < 0.2) for several biological processes (Fig 3A and S2 Data). Among the genes forming significant coexpression with extraembryonic tissue, we identified 96 genes in cluster 3 associated with “intracellular protein transport,” as well as 111 and 4 genes associated with regulation of transcription on clusters 4 and 5, respectively. Notably, 10 genes on cluster 15 were associated with “defense response to virus,” and the annotated genes are known to be stimulated by IFNT (interferon-induced protein with tetratricopeptide repeats 1 [IFIT1], IFIT3, IFIT5, ISG15, MX-dynamin–like GTPase 1 [MX1], MX2, 2′-5′-oligoadenylate synthetase 1 [OAS1Y], radical S-adenosyl methionine domain containing 2 [RSAD2]; S2 Data).
Next, we intersected the results of gene ontology enrichment obtained from clustering extraembryonic and caruncular tissues. We identified several biological processes on both data sets with coexpressing genes expressed in extraembryonic and caruncular tissues (S3 Data). Based on the number of genes and direction of connections, two pairs of biological processes are noteworthy. First, five genes associated with “positive regulation of cell proliferation” in extraembryonic tissue form negative coexpression connections (x-r=-0.96, n = 22) with 14 genes associated with “regulation of transcription, DNA-templated” expressed in caruncle (Fig 3C). Second, 10 genes associated with “transmembrane transport” in extraembryonic tissue form positive coexpression connections (x-r=0.97, n = 22) with 12 genes associated with “regulation of transcription, DNA-templated” expressed in caruncle (Fig 3D). These results are coherent with a coexpression between genes expressed in extraembryonic and caruncular tissues, with biological implications to extraembryonic tissue attachment and implantation.
The independent clustering of the correlation coefficients obtained from the genes expressed in extraembryonic and intercaruncular tissues also evidenced an organized coexpression network between the two tissues (Fig 4A). Twelve clusters formed by genes expressed in the extraembryonic tissue presented enrichment for biological processes (FDR < 0.2, Fig 4B; see S4 Data for a list of genes). Interestingly, there were 85 and 27 genes associated with “mRNA processing” and “stem cell population maintenance,” respectively, on cluster 3. On cluster 5, we identified 12 genes associated with “negative regulation of cell proliferation” and seven genes associated with “regulation of receptor activity.” On cluster 8, five genes were associated with “placenta development” (adenosine deaminase [ADA], cyclin F [CCNF], distal-less homeobox 3 [DLX3], pleckstrin-homology–like domain family A member 2 [PHLDA2], and retinoid X receptor alpha [RXRA]). On cluster 17, eight genes were associated with “regulation of transcription, DNA-templated.”
The clusters formed by intercaruncular genes coexpressed with extraembryonic tissue genes also highlighted significant enrichment of biological processes (FDR < 0.2, Fig 4A; see S5 Data for a list of genes). For instance, clusters 1 and 6 contained 145 and 63 genes associated with regulation of transcription, respectively. Interestingly, on cluster 2, there were 149, 23, 22, and 16 genes associated with “oxidation-reduction process,” “cell redox homeostasis,” “electron transport chain,” and “tricarboxylic acid cycle.” Cluster 4 contained 63 genes associated with “regulation of transcription,” and cluster 7 contained 11 genes associated with “fatty acid beta oxidation.”
The intersection of the genes identified in enriched biological processes in clusters formed by extraembryonic and intercaruncular tissues revealed several coexpression networks between these two tissues (S6 Data) that have critical implications for implantation. Notably, several of the intersecting categories involved processes associated with regulation of transcription or oxidation reduction on the intercaruncular side. For instance, 28 genes associated with “stem cell population maintenance” and expressed in extraembryonic tissue presented positive coexpression (x-r=0.97, n = 305) with 83 genes associated with “regulation of transcription” and expressed in intercaruncular tissues (Fig 4D). Five genes associated with “placenta development” and expressed in extraembryonic tissue presented negative coexpression (x-r=-0.97, n = 88) with 41 genes associated with “oxidation-reduction process” and expressed in intercaruncular tissues (Fig 4E).
In mammals and particularly in cattle, a large body of gene-expression data was produced at various steps of early pregnancy derived from in vitro- or in vivo-produced embryos [6, 27, 28, 40], varied physiological status of the dam [41], and fertility-classified heifers [30]. Altogether, results based on group analyses (extraembryonic tissue or endometrium) have demonstrated different degrees of interactions between the extraembryonic tissue and endometrium at the initial phases of implantation. In the present study, our objective was to shed light on the subtle interactions between the extraembryonic tissue of a conceptus and the endometrial tissue of the uterus hosting this conceptus in normal pregnancy using paired coexpression analyses of gene transcript abundances. Our analyses were carried out using biological material collected from the single conceptus and the endometrium from the same pregnancy, a critical aspect to determine the crosstalk during implantation at the level of one individual pregnant female.
It must be noted that our study has some limitations that may reduce the extent of the insights into individual pregnancies. First, we worked with samples collected at one developmental stage (gestation day 18). Gestation is a highly dynamic process; thus, we can anticipate that the gene interactions will also be dynamic. Second, the endometrium is a tissue diverse in cell types (e.g., epithelial lumen, stromal tissue, immune cells, glandular epithelia) that were not dissected at collection. Therefore, we did not dissect the cellular origin of the signals. This work provides a snapshot of the rich and unique interactions between extraembryonic and endometrial tissues at the tissue level that will deserve to be refined at the cell-type level.
Our analyses of transcriptome data from extraembryonic tissue and endometrium pairs identified key signatures of gene expression that are likely to be linked to the success of pregnancy recognition and implantation. A large proportion of all genes quantified in extraembryonic tissue and endometrium have transcript abundances that were not independent. Furthermore, the dependency observed for the abundance of transcripts between extraembryonic tissue and endometrium varied with morphologically and physiologically distinct areas of the endometrium, namely caruncular and intercaruncular tissues. For instance, there were twice as many highly positive (r > 0.95) and approximately half the number of highly negative (r < -0.95) coexpressing connections between extraembryonic and caruncular tissues compared to extraembryonic and intercaruncular tissues. These results greatly expand previous findings that the extraembryonic tissue triggers distinct molecular responses in caruncular and intercaruncular tissues [6, 27, 42, 43].
During the elongation phase, the mural trophoblast proliferates rapidly [12, 26, 44] while maintaining its pluripotency [45]. This period of development is modulated by dynamic regulation of gene expression [44] whereby metabolically active trophoblastic cells [46, 47] rely on the uptake of nutrients from the uterine luminal fluid [48]. Our results show that caruncular and intercaruncular tissues have an active role in the programing of those functions because several genes related to gene regulation, signal transduction, cellular proliferation, maintenance of stem cell population, and transmembrane transport are also coexpressed with genes expressed in the endometrium. The importance of gene coregulation between extraembryonic tissue and endometrium was further supported by the identification of 26 genes associated with “in utero embryonic development” and five genes associated with “placenta development” coregulated with genes expressed in caruncle and intercaruncle, respectively.
Among the genes expressed in caruncular or intercaruncular tissues that were coexpressed with extraembryonic tissues, it was noticeable that several genes were associated with regulation of gene expression. This finding is in line with former publications suggesting that the regulatory network needed for endometrial remodeling [49] during attachment is extraembryonic-tissue dependent [6, 27, 28, 40]. In the caruncular tissue, we specifically identified 15 genes associated with “defense response to virus,” of which eight genes had their expression modulated by IFNT, produced by the trophoblast between gestation days 9 and 25 [50]. This result provides additional knowledge on the biological actions of IFNT and other extraembryonic-tissue–originated signaling on the remodeling of the caruncle [51].
Our findings identified genes with high levels of coexpression (|r| > 0.9999) between extraembryonic tissue (n = 229) and endometrial caruncular tissues (n = 218) whose transcript profiles independently produced equivalent discrimination of the pregnancies. Gene ontology analysis of these 441 genes revealed that highly coexpressed genes between extraembryonic and caruncular tissues are involved in regulatory functions at the chromatin, mRNA processing, and protein levels, which is a strong indication of a coordinated reprogramming of tissues driven by multiple layers of cell regulation during the conceptus–maternal recognition. These data prompt the need for additional investigation to better define the coordinated interactions between extraembryonic tissues and endometrium at the level of tissue layer including luminal epithelium, stroma, and glandular epithelium.
In the intercaruncular tissues, our analyses identified a list of genes related with “oxidation-reduction process,” a finding consistent with a recent publication reporting that proteins associated with oxidation reduction are enriched in the uterine luminal fluid on gestation day 16 in cattle [52]. Oxidative stress is a consequence of altered oxidation-reduction state [53], and transcriptional regulation of factors involved in the regulation of oxidative stress has been reported in the bovine endometrium during the estrous cycle and early pregnancy [42, 54], Furthermore, a significant increase in oxidation-reduction potential was observed in the endometrium of mice prior to implantation [55]. The results show evidence that the maintenance of oxidation-reduction status permissive to the conceptus health [56] and implantation is strongly linked to genes regulated in intercaruncular tissues of the endometrium in cattle.
The analyses carried out in this study have provided novel, to our knowledge, insights into the molecular interactions between extraembryonic, caruncular, and intercaruncular tissues, summarized in Fig 5. Gene products expressed by the extraembryonic tissue impact the endometrial function by regulating diverse cell functions including oxidative stress, chromatin remodeling, gene transcription, and mRNA processing and translation. The endometrium also exerts key regulatory roles on the extraembryonic tissue cells by modulating chromatin remodeling, gene transcription, cell proliferation, translation, metabolism, and signaling (Fig 5). Collectively, our data have shown that endometrial plasticity, a notion first suggested in cattle [6], allows unique adaptive and coordinated conceptus-matched interactions at implantation in nonpathological pregnancies.
This study presents an analysis of paired extraembryonic tissue and endometrium in a mammalian species, using an integrative systems biology approach. A more comprehensive understanding of the connection between the conceptus and the endometrium at the gene-expression level will open new venues for the development of strategies to improve term pregnancy rates when artificial reproductive technologies are used. Since the endometrial response is embryo-specific, it would be valuable to develop approaches aiming at selection of a competent embryo better suited for the establishment of a successful crosstalk with the recipient uterus of the female considered for transfer.
This work was performed on publicly available data [27]. The initial work with animals was carried out with approval of the INRA Ethics Committee.
All analytical procedures were carried out in R software [37]. The files and codes for full reproducibility of the results are listed in the file S1 Code.
The appropriate approval from institutional committees of ethical oversight for animal use in research was obtained as reported previously [27]. All five cattle (B. taurus, Holstein breed) gestations were initiated by artificial insemination using semen from a single bull and later terminated on gestation day 18 for sample collection. We analyzed RNA-seq data (100-base long reads, GSE74152 [27]) generated from samples obtained from cattle gestations interrupted at day 18 (n = 5). The samples were extraembryonic tissue (n = 5), caruncle (n = 5), and intercaruncle (n = 5) regions from the endometrium.
The reads were aligned to the bovine genome (B. taurus, UMD 3.1) using STAR aligner [57]. Reads that aligned at one location of the genome with fewer than four mismatches were retained for elimination of duplicates. Nonduplicated reads were used for estimation of fragments per kilobase per million reads (FPKM) using Cufflinks (v. 2.2.1 [58]) and Ensembl gene models [59]. Genes were retained for downstream analyses if FPKM > 1 in ≥4 samples. We employed the t-Distributed Stochastic Neighbor Embedding approach [60] to assess the relatedness of the tissues.
Three sample types were collected from the same pregnancy (extraembryonic tissue and caruncular and intercaruncular tissues); thus, the data structure (Fig 1B) allowed us to quantify the association between genes expressed in extraembryonic tissue and the endometrium (caruncular and intercaruncular tissues). We utilized Pearson’s coefficient of correlation because of its sensitivity to outliers [61] to calculate rGj,Gk and rGj,Gl, where Gj, Gk, and Gl are the transcript abundance of a gene expressed in extraembryonic tissue, caruncle, and intercaruncle, respectively.
To assess the significance of the correlations observed in the real data set, we calculated eFDR by permuting the pregnancy index (i = 1,…,5) for the extraembryonic tissue samples, thereby breaking the pairing of extraembryonic tissue and endometrium obtained per pregnancy (B = 100 permutations). The proportion of correlations resulting from the scrambled data that was greater than a specific threshold was calculated as follows: ∑b=1B#(m:|r(Gji,Gki)0b|≥|r(Gji,Gki)scrambled|,m=1,…,(k×j))+1(k×j×B)+1 [34, 62, 63] for extraembryonic and caruncular tissues. Similar calculation was executed for extraembryonic and intercaruncular tissues.
We calculated distance matrices for extraembryonic tissue and caruncle based on the Pearson’s coefficient of correlation of the expressed genes within tissues. The correlation matrix was subtracted from one to obtain a distance matrix, which was used as input for clustering using the method “complete.” We used the Mantel statistic test implemented in the “mantel” package to assess the correlation between the two dissimilarity matrices. The significance of the Mantel statistic was assessed by a permutation approach.
We clustered samples using the “flashClust” package [64]; we used the “ComplexHeatmaps” package [65] to draw annotated heatmaps and Cytoscape software [66] to visualize the networks.
We tested for enrichment of gene ontology [67] categories and KEGG pathways [68] using the “goseq” package [69]. Subsets of genes were defined according to appropriate thresholds and defined as “test genes”; the genes expressed in the corresponding tissue were then used as background for the calculation of significance values [70]. Significance values were then adjusted for FDR according to the Benjamini and Hochberg method [71]. We further tested the likelihood of significant categories of gene ontology or KEGG pathways to have been identified by chance by a bootstrapping approach. We selected tested genes randomly (2,000 rounds of randomized subsetting) from the genes expressed and carried out the procedure for detection of enrichment described above. Then, we calculated the proportion of FDR values observed from the randomizations that were lower than the result observed from the real data.
|
10.1371/journal.pgen.1002418 | The Caenorhabditis elegans Synthetic Multivulva Genes Prevent Ras Pathway Activation by Tightly Repressing Global Ectopic Expression of lin-3 EGF | The Caenorhabditis elegans class A and B synthetic multivulva (synMuv) genes redundantly antagonize an EGF/Ras pathway to prevent ectopic vulval induction. We identify a class A synMuv mutation in the promoter of the lin-3 EGF gene, establishing that lin-3 is the key biological target of the class A synMuv genes in vulval development and that the repressive activities of the class A and B synMuv pathways are integrated at the level of lin-3 expression. Using FISH with single mRNA molecule resolution, we find that lin-3 EGF expression is tightly restricted to only a few tissues in wild-type animals, including the germline. In synMuv double mutants, lin-3 EGF is ectopically expressed at low levels throughout the animal. Our findings reveal that the widespread ectopic expression of a growth factor mRNA at concentrations much lower than that in the normal domain of expression can abnormally activate the Ras pathway and alter cell fates. These results suggest hypotheses for the mechanistic basis of the functional redundancy between the tumor-suppressor-like class A and B synMuv genes: the class A synMuv genes either directly or indirectly specifically repress ectopic lin-3 expression; while the class B synMuv genes might function similarly, but alternatively might act to repress lin-3 as a consequence of their role in preventing cells from adopting a germline-like fate. Analogous genes in mammals might function as tumor suppressors by preventing broad ectopic expression of EGF-like ligands.
| Extracellular signals that drive cells to divide must be carefully restricted so that only the correct cells receive those signals. Failure to properly control the expression of signaling molecules can lead to aberrant development and cancer. Studies of vulval development in the nematode Caenorhabditis elegans have helped define various multi-step signaling pathways involved in cancer. Here we report that two groups of proteins that control the EGF/Ras/MAP kinase pathway of vulval development act by tightly repressing the spatial expression of the gene lin-3, which encodes an EGF-like signaling molecule. Using a technique that detects single mRNA molecules, we show that inactivation of these proteins causes a low ectopic expression of lin-3 in many cells. In response, the EGF/Ras/MAP kinase pathway is activated in cells normally not exposed to the lin-3 signal, and vulval development is abnormal. This process is analogous to the cancerous growth that occurs in humans when mutations cause both tumor cells and the microenvironment surrounding the tumor cells to ectopically express factors that drive cellular proliferation. We propose that mammalian genes analogous to those that repress lin-3 expression in C. elegans vulval development act as tumor suppressors by preventing broad ectopic expression of EGF-like ligands.
| Signaling by epidermal growth factor (EGF) family ligands and EGF receptor (EGFR) family tyrosine kinases controls many aspects of mammalian development and can drive cancers: EGFRs are commonly overexpressed or constitutively activated by mutations in tumor cells [1], and EGF-family ligands can be misregulated in cancer. For example, the EGF-family ligands heparin-binding EGF-like growth factor, amphiregulin, and TGF-α are upregulated in cancer cells from many different cancer types [2], [3], and TGF-α overexpression causes widespread epithelial hyperplasia in mice [4], [5]. Growth factors often signal through a Ras pathway, and approximately 20% of tumors carry a constitutively active Ras mutation [6].
In the nematode Caenorhabditis elegans the EGF-family ligand LIN-3 acts through the EGFR LET-23 and the Ras protein LET-60 to control many aspects of development, including the induction of the hermaphrodite vulva [7]–[10]. In wild-type animals, of a set of six equipotent cells, three (P5.p, P6.p and P7.p) adopt vulval cell fates, while the other three (P3.4, P4.p, and P8.p) adopt non-vulval fates [11]. The expression of vulval cell fates requires EGF/Ras signaling, and mutations that reduce EGF/Ras signaling cause a vulvaless (Vul) phenotype in which none of the six cells adopts vulval cell fates [7]–[10]. The anchor cell, located closest to P6.p, is the only cell that both expresses LIN-3 EGF and is located near the six Pn.p cells [10], and laser ablation of the anchor cell results in a Vul phenotype [12] like that seen in mutants defective in lin-3 EGF or let-23 EGFR. Overactivation of the EGF/Ras pathway, by overexpression of lin-3 EGF or by an activating mutation in either let-23 EGFR or let-60 Ras, causes a multivulva (Muv) phenotype in which all six Pn.p cells adopt vulval cell fates [8]–[10],[13].
In vulval development, EGF/Ras signaling is antagonized by the synthetic multivulva (synMuv) genes. The synMuv genes define two classes, A and B [14], [15]. In synMuv single mutants or in class A double mutants or class B double mutants, vulval development is mostly normal. By contrast, animals mutant in both a class A synMuv gene and a class B synMuv gene exhibit a strong Muv phenotype. Many class B synMuv genes have homologs that function in histone modification, chromatin remodeling, and transcriptional repression. For example, the class B synMuv genes encode a DP/E2F/Rb complex [16], [17], a nucleosome remodeling and deacteylase (NuRD) complex [18], [19], two histone methyltransferases [20], [21] and a heterochromatin protein 1 homolog [22]. Of the three molecularly-characterized class A synMuv genes, two encode proteins with a zinc-finger-like THAP domain [23]–[25]. The expression patterns of three class A synMuv proteins have been studied, and all three are localized to the nucleus, suggesting that class A synMuv proteins regulate transcription [25], [26].
The synMuv genes function at least in part by repressing expression of lin-3 EGF. Loss-of-function mutations in either let-23 EGFR or lin-3 EGF can suppress the synMuv phenotype [16], [27], [28], indicating that the synMuv genes act upstream of or in parallel to lin-3. Furthermore, lin-3 mRNA levels are increased in synMuv double mutants but not in synMuv single mutants [27], and overexpression of lin-3 EGF causes a Muv phenotype [10].
Laser ablation of the anchor cell, the source of LIN-3 in wild-type vulval development, does not fully suppress the Muv phenotype of synMuv double mutants [28], indicating that synMuv genes cannot simply prevent overexpression of lin-3 from the anchor cell. Mosaic analyses of the class B synMuv gene lin-37 and the lin-15 locus, which contains both a class A and a class B synMuv gene, did not identify a single site of action. Both experiments indicated that lin-15 and lin-37 do not act cell-autonomously in the Pn.P cells and suggested that lin-15 and lin-37 might function in the syncytial hypodermal cell hyp7 [29], [30]. Heterologous expression experiments showed that the class B synMuv gene lin-35 functions in hyp7 to antagonize vulval cell fates, and tissue-specific RNAi of lin-3 in hyp7 can suppress the synMuv phenotype, indicating that repression of lin-3 in the hypoderm is an important function of the synMuv genes [27], [31]. Another study using the same heterologous promoters found that the class B synMuv gene hpl-2 functions in both hyp7 and the Pn.p cells [32]. However, it is not known where lin-3 is overexpressed in synMuv mutants, how the synMuv genes control lin-3 expression, or if the synMuv genes control targets other than lin-3 important for vulval development.
Here we report the identification of a lin-3 EGF promoter mutation that causes a dominant class A synMuv phenotype. The effect of this mutation reveals that the only major role of the class A synMuv genes in vulval development is to repress lin-3. We find that lin-3 mRNA is ectopically expressed throughout the animal in synMuv mutants. Our results show that low levels of ectopic lin-3 expression outside the cells that normally produce and respond to lin-3 can adversely alter the development of C. elegans, and we propose that the class A and class B synMuv genes might prevent ectopic lin-3 expression by distinct mechanisms.
During a screen for new class A synMuv mutations, we identified a Muv animal in the F1 generation after ethyl methanesulfonate (EMS) mutagenesis of the class B synMuv mutant lin-52(n771). We named the mutation that caused this defect n4441. To seek additional mutations that like n4441 dominantly cause a class A synMuv phenotype, we screened approximately 492,000 F1 progeny of lin-52(n771) animals mutagenized by EMS and approximately 89,000 progeny of animals mutagenized by N-ethyl-N-nitrosourea (ENU), but we did not identify any additional class A synMuv mutants. As a single mutant, n4441 animals are wild-type at 20°C and exhibit a low penetrance Muv defect at 25°C (Table 1), comparable to that of most class A synMuv mutants [15]. Double mutants between n4441 and the class B synMuv mutations lin-15B(n744), lin-52(n771), or lin-61(n3447) exhibit a strong synMuv phenotype. n4441 causes a fully penetrant Muv defect as a heterozygote in the class B synMuv mutant background lin-15B(n744), indicating that n4441 dominantly causes a class A synMuv phenotype. n4441 causes a 97% penetrant synMuv defect in the weak class B synMuv mutant background lin-61(n3447) at 22.5°C, comparable to the previously reported phenotype of double mutants between lin-61(n3447) and the strong class A synMuv mutations lin-15A(n767) or lin-38(n751) [15].
To determine how n4441 interacts with other class A synMuv mutations, we built double mutants between n4441 and an allele of each known class A synMuv gene. We used the putative null alleles lin-8(n2731), lin-15A(n767), and lin-56(n2728) and the missense allele lin-38(n751), since a null allele of lin-38 causes lethality (A.M.S and H.R.H., unpublished results). At 20°C and 25°C, the double mutants n4441; lin-15A(n767), lin-38(n751); n4441, and lin-56(n2728); n4441 were enhanced for the Muv phenotype when compared to their respective single mutants (Table 1). The lin-8(n2731); n4441 double mutant was roughly comparable to n4441 alone when scored at 25°C and also exhibited a low penetrance Muv defect at 20°C, which neither n4441 or lin-8(n2731) did on their own (Table 1). Thus, mutations in all known class A synMuv genes can enhance the Muv phenotype of n4441, but the enhancement is much weaker than the enhancement caused by class B synMuv mutations.
Several members of a Tip60/NuA4 histone acetyltransferase complex were previously identified as class C synMuv genes [33]. Class C synMuv genes are strongly Muv in combination with class A synMuv mutations and weakly Muv in combination with class B synMuv mutations and can be considered a subset of the class B synMuv genes [15]. To test if n4441 might be a class C synMuv gene, we built a double mutant between n4441 and the partial loss-of-function class C synMuv mutation mys-1(n3681), as null mutants of mys-1 cannot be maintained as homozygous strains [33]. The mys-1(n3681); n4441 double mutant exhibited a 56% penetrant Muv defect at 20°C, which is much stronger than the 5% penetrant Muv defect of the n4441; lin-15A(n767) strain at 20°C, despite the fact that lin-15A(n767) is a null mutation. We conclude that n4441 is not a class C synMuv mutation.
By performing SNP mapping experiments using the CB4856 polymorphic strain of C. elegans, we mapped the n4441 mutation to a 661 kb region containing approximately 170 genes between SNPs dbP6 and uCE4-1148 (Figure 1A). n4441 dominantly causes a synMuv phenotype and thus might well be a gain-of-function mutation, so we sought loss-of-function mutations in the gene affected by n4441. n4441/nT1[qIs51]; lin-15B(n744) animals, which display a fully penetrant Muv defect, were mutagenized with EMS. The nT1[qIs51] translocation causes inviability when homozygous and suppresses recombination across an interval that includes lin-3 [34]. Approximately 6,800 F1 progeny were screened, and two animals were identified that were non-Muv and produced only non-Muv progeny, indicating that they contained a suppressor mutation tightly linked to n4441. We named these mutations n4929 and n4951. n4441 n4929; lin-15B(n744) animals were sterile and exhibited a very low penetrance Muv defect (Figure 1D). n4441 n4951/nT1[qIs51] animals were superficially wild-type with no Muv defect, and n4441 n4951 homozygotes died as L1 larvae with a rod-like appearance (Figure 1D). The rod-like lethal phenotype is characteristic of loss-of-function mutations in genes in the EGF/Ras pathway required for vulval induction [35]. The only known gene in the EGF/Ras pathway in the genetic interval containing n4441 is lin-3, which encodes the EGF ligand. Strong loss-of-function alleles of lin-3 cause a rod-like lethal phenotype, and lin-3 mutations can also cause sterility [36], [37]. The n4929 mutant carries a G-to-A transition in the first nucleotide of exon 8 of lin-3 and is predicted to mutate an arginine to a lysine at amino acid 347 of LIN-3 (Figure 1B). The n4951 mutant carries a G-to-A transition that results in a nonsense mutation predicted to truncate LIN-3 after only 26 amino acids, before the EGF domain (Figure 1B). The lin-3(n1059) nonsense mutation failed to complement the sterility caused by n4929 and the lethality caused by n4951, proving that n4929 and n4951 are alleles of lin-3. Since the lin-3(n4951) nonsense mutation suppressed the n4441 synMuv defect in cis, but the lin-3(n1059) nonsense mutation did not suppress the n4441 synMuv defect in trans (Figure 1D), a lin-3 loss-of-function mutation is a cis dominant suppressor of n4441, indicating that n4441 is a gain-of-function allele of lin-3.
We determined the sequences of all exons and introns of lin-3 and of approximately 11 kb of upstream DNA in lin-3(n4441) mutants. The only mutation was a G-to-A transition at nucleotide 30904 of cosmid F36H1, approximately 200 bp upstream of the lin-3 transcript F36H1.4a (http://www.wormbase.org, release WS200, 20 Mar 2009) (Figure 1C). To show that the F36H1(30904) mutation is required for the class A synMuv phenotype caused by lin-3(n4441), we sought recombinants between lin-3(n4441) and the lin-3(n4951) nonsense mutation, which is 5.3 kb downstream of F36H1(30904). We screened approximately 90,000 progeny from lin-3(n4441 n4951)/+; lin-15B(n744) animals, identified five independent Muv animals and established homozygous lines. None of the five lines contained the lin-3(n4951) mutation, and all five carried the F36H1(30904) G-to-A mutation. Thus, the lin-3(n4441) mutation that causes the class A synMuv phenotype must be to the left of lin-3(n4951), because if it were to the right then the recombinants would not carry the F36H1(30904) mutation. The 5.3 kb between F36H1(30904) and lin-3(n4951), as well as 10.8 kb of DNA upstream of F36H1(30904), carried no additional mutations in lin-3(n4441) animals. If the mutation that causes the lin-3(n4441) synMuv phenotype is not the F36H1(30904) mutation, then the lin-3(n4441) mutation must be at least 10.8 kb to the left of the F36H1(30904) mutation. However, in that case, assuming a constant recombination rate throughout the lin-3 interval, the likelihood that all five recombination events would have occurred between F36H1(30904) and lin-3(n4951) is ((5.3)/(5.3+10.8))5, or <0.004. We conclude that the G-to-A mutation at nucleotide 30904 of cosmid F36F1 is necessary for the class A synMuv phenotype caused by lin-3(n4441). However, we cannot rule out the possibility there is a second mutation more than 11 kb upstream of lin-3 that is also required along with the F36H1(30904) mutation to cause a class A synMuv phenotype. There are no known consensus transcription factor binding sites that include the site of the lin-3(n4441) mutation (Transfac database of known transcription binding sites; http://www.gene-regulation.com). The region surrounding the lin-3(n4441) mutation is moderately conserved in the related nematodes C. briggsae and C. remanei (data not shown).
lin-3(n4441) might be a class A synMuv specific allele of lin-3. Alternatively, lin-3(n4441) might cause weak overexpression of lin-3 if weak overexpression of lin-3 behaves like a class A synMuv mutation. To differentiate between these alternatives, we overexpressed lin-3 weakly using the syIs12 integrated transgene. syIs12 expresses the EGF domain of lin-3 under the control of a heat-shock promoter [38]. At 20°C in the absence of heat-shock, syIs12 did not cause a Muv phenotype (Table 2). syIs12; lin-15B(n744) animals were mostly wild-type, with only a 1% penetrant Muv defect, whereas syIs12; lin-15A(n767) animals exhibited a Muv defect with 40% penetrance (Table 2). Thus, weak overexpression of lin-3 from the syIs12 transgene was enhanced by a class A synMuv mutation but not by a class B synMuv mutation. By contrast, lin-3(n4441) was enhanced much more strongly by class B synMuv mutations than by class A synMuv mutations (Table 1). We conclude that lin-3(n4441) is a class A synMuv specific allele of lin-3 and does not simply cause weak overexpression of lin-3.
The class A and B synMuv genes redundantly repress expression of lin-3 mRNA [27]. To test if the lin-3(n4441) mutation affects lin-3 mRNA levels similarly to other class A synMuv mutations, we assayed lin-3 mRNA levels using real-time RT-PCR. As previously reported, the class B synMuv mutant lin-15B(n744) has wild-type lin-3 levels (Figure 2B). The class A synMuv mutants lin-15A(n767) and lin-3(n4441) both had slightly increased levels of lin-3 mRNA. The synMuv double mutants lin-15AB(e1763) and lin-3(n4441); lin-15B(n744) had substantially increased lin-3 mRNA levels (Figure 2B). Therefore, the lin-3(n4441) mutation behaves as a class A synMuv mutation with respect to lin-3 mRNA repression.
The lin-3(n4441) mutation is located 211 bp upstream of lin-3 and is also 465 bp upstream of F36H1.12, which is upstream of lin-3 in the opposite orientation (Figure 2A). To determine if lin-3(n4441) or other synMuv mutations also affect expression of F36H1.12, we assayed F36H1.12 mRNA levels by real-time RT-PCR. F36H1.12 mRNA levels were roughly equivalent to those of the wild type in all possible single and double mutant combinations involving lin-15A(n767), lin-3(n4441), and lin-15B(n744) (Figure 2C). Therefore, the synMuv proteins specifically repress lin-3 and do not establish a broad domain of repression.
Although lin-3 is overexpressed in synMuv double mutants [27] (also, Figure 2), it is not known where this overexpression occurs. GFP- and LacZ-tagged lin-3 repetitive transgene arrays have been used as reporters for lin-3 expression [10], [39], [40], but these reporters might not be appropriate for determining lin-3 expression in synMuv mutants: first, the level of ectopic lin-3 expression might be too low to visualize using a GFP reporter; second, many synMuv mutations affect the expression of repetitive transgene arrays, potentially confounding interpretation of the expression pattern of such reporters [41]. Instead, we assayed lin-3 expression using a fluorescence in situ hybridization (FISH) technique that has sufficient sensitivity to detect single mRNA molecules [42]. We used 48 non-overlapping probes against lin-3 (Table S1), each conjugated to a single fluorophore, to label individual mRNA molecules brightly enough to be visible as distinct fluorescent spots. Because there are 48 probes that bind independently to the target mRNA, any single probe that binds non-specifically should not cause a false-positive signal. The distribution of intensities of the spots in any given animal was unimodal, consistent with each spot's representing a single mRNA molecule (Figure S1). Furthermore, by comparing the spot intensities in different tissues and mutants, we found that the level of expression in a given cell or tissue was independent of the intensity of the spots in that cell or tissue, and if the number of spots in a cell was altered then the average intensity of spots in that cell was unchanged (Figure S2). If each spot represented multiple mRNA molecules, then as the expression level in a given cell increased the average number of mRNA molecules in each of those spots would also be expected to increase, leading to greater intensity. Because the intensity of each spot was independent of the level of expression, we conclude that each spot is likely to represent a single mRNA molecule. We also found that all tissues are accessible to FISH probes, as probes directed against ama-1 and eft-2 robustly detected mRNA in all cells (data not shown). However, we cannot know if we are detecting every mRNA molecule; it is possible that some mRNA molecules are not accessible to the oligonucleotide probes or are not detected for some other reason.
We first determined the expression pattern of lin-3 in wild-type animals at the late L2 to early L3 stage when vulval induction occurs. Previous studies found that at the early L3 stage lin-3 is expressed in the anchor cell and in the pharynx [10], [40]. We indeed observed robust expression of lin-3 in the anchor cell and throughout the pharynx. We also saw expression of lin-3 in the germline (Figure 3A and Figure S3). In some wild-type animals we also observed a few copies of lin-3 mRNA in one or more cells in the tail, on the ventral side slightly anterior to the anus. In addition, a few copies of lin-3 mRNA were seen on the ventral side of the animal, slightly behind the posterior gonad arm. We imaged several animals that were slightly older, in the late L3 stage, and observed expression of several copies of lin-3 mRNA in the region where P6.p and its descendants are located (data not shown), consistent with previous reports of expression of lin-3 in the descendants of P6.p by the L4 stage [39]. We did not consistently detect any lin-3 mRNA in other tissues, although in some animals we observed a single lin-3 mRNA molecule elsewhere. For example, in the animal shown in Figure 3A a single lin-3 mRNA molecule was observed in or near an intestinal cell close to the anchor cell. Overall, other than for those tissues that highly expressed lin-3 there was very tight repression of lin-3. The numbers of copies of lin-3 mRNA we observed in each tissue in individual animals are listed in Table S2 and are summarized in Table 3. The expression pattern we observed for lin-3 is consistent with that seen using GFP- and LacZ-tagged lin-3 reporters [10], [39], [40] and with functional studies of lin-3 [43], indicating that most if not all of the mRNA spots identified by this technique are likely to represent actual lin-3 mRNA molecules.
lin-15AB(e1763) animals expressed lin-3 in the pharynx, germline, and anchor cell at levels grossly similar to those of wild-type animals (Figure 3D and Table 3). In addition there was widespread ectopic expression of lin-3, with an average of approximately 1100 ectopic copies of lin-3 mRNA observed per animal (Figure 3D and Table 3). This ectopic expression was much weaker than the normal expression in the anchor cell; whereas an average of 29 copies of lin-3 mRNA was seen in the anchor cell in wild-type animals (Table 3), only one or a few copies of lin-3 mRNA were observed in most cells in lin-15AB(e1763) mutants. Because we could not see cell boundaries, we could not determine if every cell ectopically expressed lin-3, but there were no tissues that appeared to lack ectopic lin-3 mRNA (Figure S4). Cells around the perimeter of the animal expressed lin-3 in the lin-15AB(e1763) mutant, consistent with ectopic expression in the hypodermis (Figure 3D and Figure S4). There were also many ectopic lin-3 mRNA copies that clearly were not in the hypodermis (Figure S4).
We also determined lin-3 expression in lin-15A(n767) and lin-15B(n744) single mutants. lin-15B(n744) animals had a lin-3 expression pattern similar to that of wild-type animals (Figure 3C). In lin-15B(n744) mutants there was an extremely low level of ectopic lin-3 expression, with an average of six ectopic lin-3 mRNA molecules detected per animal (Table 3), but lin-3 was still tightly repressed outside of the germline, anchor cell, and pharynx. lin-15A(n767) animals exhibited broad ectopic expression of lin-3, but at a much lower level than that of lin-15AB(e1763) animals (Figure 3B). An average of 64 copies of lin-3 mRNA were seen outside of the pharynx, germline, and anchor cell in lin-15A(n767) animals (Table 3). Unlike in lin-15AB(e1763) animals, in any given lin-15A(n767) animal most cells did not display ectopic lin-3 expression. However, we observed no obvious cell or tissue specificity to the ectopic expression among several lin-15A(n767) animals. Rather, it appeared that in lin-15A(n767) animals lin-3 is globally derepressed, but at a very low level.
The numerous class B synMuv genes have highly similar although not identical effects on vulval development [15]. However, the class B synMuv genes have widely differing effects on other aspects of growth and development. For example, PGL-1, which is normally expressed in the germline, is misexpressed in the somatic cells of mutants of many class B synMuv genes, including lin-15B, but not in mutants of some other class B synMuv genes, including lin-36, lin-52, and lin-53 [44], [45]. We therefore investigated the role of the class B synMuv genes lin-36, lin-52, and lin-53 in controlling lin-3 expression. We determined the expression pattern of lin-3 in lin-36(n766); lin-15A(n767), lin-52(n771); lin-15A(n767), and lin-53(n833); lin-15A(n767) mutants. All three double mutants exhibited ubiquitous ectopic expression of lin-3, with lin-3 mRNA observed in most if not all tissues (Figure 4). There were no obvious differences in the spatial pattern of lin-3 expression in lin-36(n766); lin-15A(n767), lin-52(n771); lin-15A(n767), and lin-53(n833); lin-15A(n767) double mutants as compared to lin-15AB(e1763) mutants.
The lin-3(n4441) mutation could cause global derepression of lin-3 similarly to lin-15A(n767), or it could affect lin-3 expression in a subset of tissues. We examined the expression of lin-3 mRNA in lin-3(n4441) and lin-3(n4441); lin-15B(n744) animals. lin-3(n4441) animals had widespread but weak ectopic expression of lin-3, similar to lin-15A(n767) animals (Figure 5A and Table 3). lin-3(n4441); lin-15B(n744) animals exhibited ectopic lin-3 expression in most cells and were indistinguishable from lin-15AB(e1763) animals (Figure 5B and Table 3).
Identifying the biologically relevant targets of transcriptional regulators that control development is a challenging problem. The synMuv genes encode putative transcriptional repressors that prevent ectopic vulval development. Mutating a synMuv binding site in a target gene might relieve repression of that target, and if that repression were essential to prevent ectopic vulval development could cause a dominant synMuv phenotype. We isolated a mutation in the lin-3 EGF gene that derepresses lin-3 transcription and causes a dominant class A synMuv phenotype. This finding establishes that lin-3 is a functionally important target of the class A synMuv genes, consistent with a previous report that lin-3 expression is repressed by the synMuv genes and that double-stranded RNA directed against lin-3 can suppress the synMuv phenotype [27]. Importantly, the lin-3(n4441) mutation fully recapitulates the class A synMuv phenotype with regard to vulval development and lin-3 expression and causes a class A synMuv phenotype equivalent to that caused by strong alleles of class A synMuv genes. If the class A synMuv genes repressed multiple targets to prevent ectopic vulval development, then a mutation that abolished class A synMuv-mediated repression of lin-3 would recapitulate only partially the class A synMuv phenotype. We conclude that lin-3 is likely to be the only key biologically relevant target of the class A synMuv genes in vulval development.
The simplest interpretation of the effect of the lin-3(n4441) mutation is that this mutation abolishes a binding site for a transcriptional repressor consisting of or controlled by class A synMuv proteins. However, the effect of the lin-3(n4441) mutation is slightly enhanced by mutations in all other class A synMuv genes. If the lin-3(n4441) mutation completely inactivated a binding site that responds to only one of the known class A synMuv proteins, then mutation of that class A synMuv gene should not enhance the synMuv phenotype caused by lin-3(n4441). One possibility is that a complex consisting of multiple class A synMuv proteins binds to the lin-3 promoter, the lin-3(n4441) mutation strongly reduces but does not completely eliminate that binding, and removing any one class A synMuv protein does not fully abrogate the ability of the complex to bind to the lin-3 locus and repress transcription. This model is consistent with the observation that most class A synMuv mutations, including lin-3(n4441), are enhanced by class A synMuv mutations in other genes [15]. Alternatively, the class A synMuv genes might indirectly repress lin-3 by regulating the expression or activity of or by binding to another protein that binds to the lin-3 promoter to prevent ectopic transcription. Because a mutation in the lin-3 promoter can cause a class A synMuv phenotype, the class A and class B synMuv pathways must be integrated at the point of lin-3 repression, and hence it is unlikely that the class A and B synMuv genes redundantly control a transcriptional regulator which in turn controls lin-3 expression.
lin-3 expression in the germline had not been previously observed, likely because the reporters used to assay lin-3 expression were either silenced in the germline [46] or lacked distant regulatory regions necessary to drive germline expression. Mutations in the FOG and FBF translational inhibitor RNA-binding proteins cause a germline-dependent Muv phenotype, and the FBF proteins can bind to the 3′ UTR of lin-3 in vitro, suggesting that germline lin-3 mRNA is translationally repressed during the larval stage when vulval induction occurs [43]. In many class B synMuv mutants, somatic cells express normally germline-specific genes [19], [44], [45]. Given our finding that lin-3 is normally expressed in the germline, one possibility is that the class B synMuv genes repress ectopic lin-3 expression in somatic cells as a consequence of their role in ensuring that somatic cells do not inappropriately adopt germline-like fates. The class B synMuv genes might all directly repress lin-3 in somatic cells. Alternatively, as there are a large number of class B synMuv genes and their effects on vulval development are not identical, perhaps at least some class B synMuv genes indirectly repress lin-3 by preventing the ectopic adoption of germline-like fates. In class B synMuv single mutants, the somatic cells adopt a more germline-like fate that would include lin-3 expression except that the class A synMuv genes still tightly repress lin-3, mostly preventing ectopic lin-3 expression. In class A synMuv single mutants, lin-3 is not tightly repressed, but most somatic cells are not fated to express lin-3, so there is only a low level of leaky ectopic lin-3 expression. However, in class AB synMuv double mutants, the somatic cells adopt a germline-like fate that includes lin-3 expression, and there is no class A synMuv mechanism that tightly represses lin-3, resulting in widespread and substantial ectopic lin-3 expression. In short, we suggest that the synthetic Muv phenotype caused by mutations in the synMuv genes might be a consequence of two distinct functions of the class A and class B synMuv genes: the class A synMuv genes either directly or indirectly tightly repress ectopic lin-3 transcription, and the class B synMuv genes prevent somatic expression of germline-expressed genes, which include lin-3; only if both functions are lost will somatic cells ectopically express sufficient lin-3 mRNA to cause ectopic vulval induction. These findings raise the possibility that the development of some human tumors might require the loss of one tumor suppressor gene that prevents cells from adopting a fate that is permissive for the expression of a growth factor and the loss of a second tumor suppressor gene that specifically represses the expression of that growth factor.
A subset of the class B synMuv genes is required to prevent the somatic misexpression of normally germline-restricted P-granule proteins such as PGL-1 [44], [45]. We found that lin-15B mutants, which do exhibit somatic PGL-1 expression, and lin-36, lin-52, and lin-53 mutants, which do not exhibit somatic PGL-1, all have highly similar effects on lin-3 expression. These results indicate that different germline genes are broadly repressed in the soma by different sets of transcriptional repressors. The class B synMuv genes define one such group of repressors and are classified together because they have comparable effects on the germline gene lin-3, resulting in similar vulval phenotypes. Many such partially-overlapping groups of transcriptional repressors, including various subsets of the class B synMuv genes, are likely to be required for the repression in the soma of other germline-restricted genes.
Whereas lin-3 expression in most cells is tightly repressed by the synMuv genes, the anchor cell and germline exhibit robust lin-3 expression that is not substantially affected by the synMuv genes. While it has not been reported whether or not any synMuv genes are expressed in the anchor cell, several synMuv genes are expressed in the germline [17], [22], [26], [33], and we are not aware of studies that have conclusively shown any synMuv genes not to be expressed in the germline. In most cells, the synMuv genes reduce lin-3 expression from an average of one to two copies per cell to nearly zero copies per cell. The synMuv genes clearly do not have a similar fold effect on lin-3 expression in the anchor cell and germline. However, it is possible that the synMuv genes repress a similar absolute number of leaky lin-3 mRNA molecules in all cells; given the animal-to-animal variability in lin-3 expression we likely would not have been able to detect such a small increase in the anchor cell or germline. Alternatively, the synMuv genes might not repress lin-3 in the anchor cell or germline. The strong activator(s) of lin-3 that drive expression in those tissues could override the activity of the synMuv genes, or one or more synMuv genes might not be expressed in those tissues, thereby compromising synMuv repression of lin-3.
In lin-15AB mutants, lin-3 is ectopically expressed throughout the animal in a broad range of cells and tissues. Site-of-action experiments have shown that the synMuv genes function at least in large part in the hyp7 hypodermal syncytium to prevent ectopic vulval development [31]. The expression pattern of lin-3 in synMuv mutants does not directly identify the site-of-action of synMuv genes in regulating vulval development but does show that the synMuv genes function throughout the animals to keep lin-3 very tightly repressed in numerous cells and tissues. lin-3 EGF regulates non-vulval cell fates in C. elegans development, and at least some of these fates, such as the P11/P12 fate, are also regulated by the synMuv genes in a manner analogous to that of vulval development [47]. In short, the synMuv genes act throughout the animal to prevent ectopic lin-3 expression, which can cause a variety of developmental abnormalities. Mutants with a displaced anchor cell show that lin-3 can act at a distance [48], so a cell ectopically expressing lin-3 could affect fates in both nearby and distant cells. We suggest that for any given cell-fate decision, the site of action of the synMuv genes is likely to be spread across multiple cells and determined by the size and proximity of those cells to the cell being regulated by lin-3. In the case of vulval development, hyp7 plays the major role, given its large size and close proximity to the Pn.p cells, with likely lesser contributions from many other cells. The site at which the synMuv genes repress lin-3 to ensure proper vulval development is therefore probably a combination of the Pn.p cells themselves and neighboring cells that do not normally either express or respond to lin-3. This situation is similar to that in which both tumor cells and the microenvironment surrounding the tumor provide factors that drive tumor development [49]. We suggest that analogously to the synMuv genes some tumor suppressor genes function by repressing growth factor expression in both tumor cells and the surrounding microenvironment.
In synMuv double mutants, lin-3 was ectopically expressed but at a much lower level than at its major normal site of function, the anchor cell. synMuv double mutants might ectopically express as few as one to two copies of lin-3 mRNA per cell. Thus, normal C. elegans development requires lin-3 to be exceedingly tightly repressed outside of a few cells, and only slight expression of lin-3 throughout the animal can cause abnormal cell-fate transformations. Such low levels of ectopic expression would likely be missed by most techniques used to assay gene expression. We suggest it could be important to examine the expression of EGF-family ligands in tumors using highly sensitive techniques with single-molecule resolution to determine if broad low-level misexpression of EGF-family ligands plays a role in oncogenic growth. In C. elegans, the tight repression of lin-3 EGF requires both the class A synMuv gene pathway and the class B gene synMuv pathway, which includes homologs of known tumor suppressor genes, such as lin-35 Rb. Therefore, some tumor suppressor genes in mammals might function by tightly repressing low-level ectopic expression of EGF-family ligands in many cells, possibly in both the tumor and the microenvironment surrounding the tumor.
C. elegans strains were cultured by standard methods on OP50 bacteria [50]. All animals were grown at 20°C, except where otherwise noted. The wild-type strain was N2, except in SNP mapping experiments in which the polymorphic CB4856 strain was also used [51]. The following mutations were used in this study:
LGI: dpy-5(e61), lin-61(n3447), lin-53(n833)
LGII: lin-8(n2731), lin-56(n2728), lin-38(n751), syIs12
LGIII: dpy-17(e164), lin-36(n766), unc-32(e189), lin-52(n771)
LGIV: lin-3(n4441), lin-3(n4929), lin-3(n4951), lin-3(n1059)
LGX: lin-15A(n767), lin-15B(n744), lin-15(e1763)
The balancer strain nT1[qIs51] IV∶V [34] was used; qIs51 is a GFP-expressing transgene integrated onto the nT1 translocation. Table S3 lists all strains used in this study.
Synchronized animals were harvested at or near the L2-to-L3 larval transition, when vulval induction occurs. N2 animals were harvested 33 hours after starved L1 larvae were placed on plates with food. Some mutants grew more slowly and were harvested after 39 hours. Quantitative RT-PCR for lin-3 was performed as previously described [15]. lin-3 was amplified using the primers CGCATTTCTCATTGTCATGC and CTGGTGGGCACATATGACTC.
Animals were grown to the L2-to-L3 transition as in the quantitative RT-PCR experiments. Fixation and hybridization were performed as described previously [42], except that worms were fixed for one hour instead of 45 minutes. The lin-3 probes (Biosearch Technologies, Inc) were conjugated to the fluorophore Cy5 using the Amersham Cy5 Mono-reactive Dye pack (GE Healthcare). DNA was visualized using 4′,6-diamidino-2-phenylindole (DAPI). The probe sequences used are shown in Table S1. Figure 2 and Figure 3 are maximum intensity projections of a Z-stack of images processed with the Find Edges and Smooth operations in ImageJ. lin-3 mRNA spots were computationally identified with manually determined thresholds as previously described [42]. The number of molecules within each tissue were then manually counted. The anchor cell was identified based on position. lin-3 mRNA expression in the anchor cell appeared as a tight cluster of spots; molecules within that cluster were considered to be in the anchor cell. To determine which lin-3 mRNA molecules were in the pharynx, we noted the outline of the pharynx that was clearly visible as a dark boundary in the Cy5 channel of the image stacks (see Figure S3 and Figure S4). The boundaries of the germline were estimated from the positions of DAPI-labeled germline nuclei.
|
10.1371/journal.pbio.1001392 | Radar Tracking and Motion-Sensitive Cameras on Flowers Reveal the Development of Pollinator Multi-Destination Routes over Large Spatial Scales | Central place foragers, such as pollinating bees, typically develop circuits (traplines) to visit multiple foraging sites in a manner that minimizes overall travel distance. Despite being taxonomically widespread, these routing behaviours remain poorly understood due to the difficulty of tracking the foraging history of animals in the wild. Here we examine how bumblebees (Bombus terrestris) develop and optimise traplines over large spatial scales by setting up an array of five artificial flowers arranged in a regular pentagon (50 m side length) and fitted with motion-sensitive video cameras to determine the sequence of visitation. Stable traplines that linked together all the flowers in an optimal sequence were typically established after a bee made 26 foraging bouts, during which time only about 20 of the 120 possible routes were tried. Radar tracking of selected flights revealed a dramatic decrease by 80% (ca. 1500 m) of the total travel distance between the first and the last foraging bout. When a flower was removed and replaced by a more distant one, bees engaged in localised search flights, a strategy that can facilitate the discovery of a new flower and its integration into a novel optimal trapline. Based on these observations, we developed and tested an iterative improvement heuristic to capture how bees could learn and refine their routes each time a shorter route is found. Our findings suggest that complex dynamic routing problems can be solved by small-brained animals using simple learning heuristics, without the need for a cognitive map.
| Many food resources, such as flowers refilling with nectar or fruits ripening on a tree, replenish over time, so animals that depend on them need to develop strategies to reduce the energy they use during foraging. Here we placed five artificial flowers in a field and set out to examine how bumblebees optimize their foraging routes between distant locations. We tracked the flight paths of individual bees with harmonic radar and recorded all their visits to flowers with motion-sensitive video cameras. This dataset allowed us to study how bees gradually discover flowers, learn their exact position in the landscape, and then find the shortest route to collect nectar from each flower in turn. Using computer simulations, we show that the level of optimisation performance shown by bees can be replicated by a simple learning algorithm that could be implemented in a bee brain. We postulate that this mechanism allows bumblebees to optimise their foraging routes in more complex natural conditions, where the number and productivity of flowers vary.
| Animals moving in familiar environments often follow habitual routes to navigate between important locations, such as the nest and feeding sites. Most knowledge on route following behaviours has been deduced from the stereotyped paths insects [1]–[4] and birds [5] develop when travelling between home and a single other site. In contrast, very little is known about the routing decisions made by animals that must visit multiple sites before returning home. These routing challenges are common in central place foraging nectarivores and frugivores, which typically exploit familiar food resources that replenish over time. Many of these animals develop stable foraging circuits (traplines) between distant food patches [6]–[10] and must sometimes cover several kilometres to fill their crop [11].
Developing an efficient route to reduce the travelling costs between multiple foraging locations is an optimisation task analogous to the well-known travelling salesman problem (finding the shortest route to visit a set of locations once and return to the origin) [12]. The most direct approach to solve this mathematical problem is to compare all the possible routes, which often requires extensive computational power as the number of routes increases factorially with the number of locations to be visited (e.g., 5! = 120 possible routes in a problem with only 5 locations). For animals, this problem is of a different nature as they cannot plan a route in advance, using a geographic map, but must gradually acquire information about the locations and the paths linking them. Therefore many animals [13]–[16], including humans [17],[18], navigating between multiple locations are thought to find efficient routes using heuristic strategies, such as linking nearest unvisited sites or planning a few steps ahead.
Recent laboratory studies have shown that bumblebees foraging in simple arrangements of artificial flowers in indoor flight cages develop near optimal traplines after extensive exploration, based on learning and spatial memories [15],[19]–[21]. However, whether similar strategies are observed at larger spatial scales, when animals must search to localise distant feeding sites and when the costs of travelling suboptimal routes are magnified, remains largely unexplored. In addition, over the smaller spatial scales at which bees were previously tested, nearby flowers were typically visible from other flowers, which is often not the case over natural foraging scales in the field. Obtaining data about the ontogeny of traplines in the wild is challenging, since it requires the observer to have information about the spatial location of all available food patches, the complete foraging history of the animals, and their movements with sufficient accuracy to retrace their routes.
Here, taking advantage of the possibility to train bumblebees (Bombus terrestris) to forage on artificial flowers in the field [22], to track their complete flight paths with harmonic radar [23],[24], and to record all their flower visits with motion-sensitive cameras, we investigate the acquisition of long-distance traplines by animals with known foraging experience. We describe how bees develop stable routes between five feeding locations by combining exploration, learning, and sequential optimization. We then compare bees' optimization performances to those of simple heuristic algorithms and develop a novel iterative improvement heuristic replicating the observed dynamics of route acquisition.
Our first aim was to establish whether bees develop repeatable foraging circuits between stable feeding locations. We pre-trained naïve bees to collect sucrose solution rewards from a patch of five artificial flowers (Figure S1) in the middle of the experimental field (Figure 1). After a day of pre-training, bees of known foraging experience were tested individually with the five flowers arranged in regular pentagon (50 m side length). Each flower provided a sucrose reward equivalent to one-fifth of the bee's crop capacity and was refilled after each foraging bout. We tested seven bees for seven consecutive hours each on a different day. All visits to the flowers were video recorded with motion-activated webcams at each feeding station (Video S1). The first flight of an inexperienced forager and the final flight paths of five experienced foragers were recorded with harmonic radar (Videos S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13).
Bees discovered flowers sequentially and had visited all five flowers at least once after an average of eight foraging bouts (here, and throughout the text, means are reported ± s.e.m.; 8.14±2.43 bouts, n = 7 bees; Figure 2A). The two flowers closest to the nest (F1 and F5) were located first, by all individuals. The flower furthest from the nest (F3) was found last by four bees, whereas it was the penultimate flower discovered by the other three. Individual bees consistently approached flowers from the same quadrant of the landing platform (Video S1), irrespective of the flower visited and of their experience (Generalized Linear Mixed Model (GLMM), effect of quadrant on the frequency of visits, F3,1328 = 90.23, p<0.001; effect of flower identity, F4,1328 = 1.82, p = 0.08; effect of the number of bouts completed, F1,1328 = 0.07, p = 0.791; all interactions, p>0.05). Frequency distributions of approaches in each quadrant were significantly different among bees (χ218 = 996, p<0.05; Figure 2B), indicating that each bee approached and landed on flowers from a different preferred angle. Furthermore, bees departed from the same quadrant as they arrived (and thus in opposite directions) in 71.41%±1.72% (n = 7 bees) of visits. The frequency of visits when arrivals and departures occurred in the same quadrant did not vary significantly in relation to flower location or to the foraging experience of bees (GLMM, effect of flower identity on the frequency of visits where arrival and departure occurred in the same quadrant, F4,1328 = 2.27, p = 0.065; effect of the number of bouts completed, F1,1328 = 0.46, p = 0.499; interaction, F4,1328 = 4.87, p = 0.222). We also found no significant difference in the frequency of these visits among bees (χ26 = 10.29, p = 0.113). Therefore, our data suggest that each bee acquired a directional preference in arrivals to and departure from flowers before the observations began, possibly during the pre-training phase when the bees became familiar with the flower design, and used their directional preference consistently for visiting flowers in all novel locations discovered.
As they gained experience, bees increased the number of different flowers visited per foraging bout (first five bouts, 2.29±0.35 flowers; last five bouts, 4.97±0.06 flowers; n = 7 bees; GLMM, effect of the number of bouts completed on the number of flowers visited, F1,194 = 149.62, p<0.001) and reduced the frequency of revisits to empty flowers (first five bouts, 2.83±0.58 revisits; last five bouts, 1.31±0.55 revisits; n = 7 bees; GLMM, effect of the number of bouts completed on the frequency of revisits, F1,194 = 6.50, p = 0.012). In every bout, a bee's probability to link the nest and a flower or to link two flowers together was determined by its experience. Thus, transition vectors between any two locations used in previous bouts were used more often in subsequent bouts than transitions vectors never previously experienced (GLMM, effect of the cumulative frequency of all possible transition vectors in previous bouts on their frequency of usage at each bout, F1,5848 = 1,209.5, p<0.001). Among the paths already used, the probability of repeating a transition vector in two successive foraging bouts increased significantly with the optimality ratio (straight line length of the observed visitation sequence divided by straight line length of the shortest possible sequence to visit the same number of flowers) of the first bout (GLMM, effect of optimality ratio of the first bout on the frequency of transition vectors repeated in the second bout, F1,1069 = 82.64, p<0.001; Figure 2C). In other words, transition vectors that generated short routes were likely to be used again in subsequent bouts, while transition vectors producing long routes were gradually abandoned, thus limiting the number of novel transitions over time (Figure 2D).
With increasing experience, the sequence in which flowers were visited became more similar over successive foraging bouts (similarity index—see Materials and Methods—between the first two bouts, 0.2±0.05; similarity index between the last two bouts, 0.89±0.07, n = 7 bees; GLMM, effect of the number of bouts completed on similarity index, F1,187 = 78.14, p<0.001), leading to a regular repeatable sequence, or “trapline”: the most common five-flower visitation sequence excluding revisits used by each individual bee (Figure 2E; Table S1). On average, the trapline was used in 27.13%±3.46% (n = 7 bees) of each bee's foraging bouts. It first appeared after 17.57±1.79 bouts (n = 7 bees) and was stabilized (repeated in at least three consecutive bouts at the end of training) in six bees after 30±0.8 bouts. Among the 120 possible sequences to visit all five flowers once and return to the nest, each bee selected one of the two shortest possible sequences as its trapline, either by visiting the flowers in a clockwise (sequence, 12345; n = 4 bees) or an anti-clockwise order (sequence, 54321; n = 3 bees).
Radar tracks obtained from five experienced bees, near the end of the training phase, confirmed that the routes followed were highly repeatable and close to minimizing the overall travel distance (Figure 3B–F; Videos S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13). Flight paths were composed of relatively straight segments linking either the nest and a flower or two flowers together. During each bee's final foraging bout, these flight segments were on average 26.09%±0.10% (n = 30 segments) longer than a straight line. Overall, the bees travelled 458.10±29.14 m (n = 5 bees), which is 146.92±29.14 m longer than the shortest possible path to visit the five flowers (311.8 m). This value contrasts sharply with the 1,953.01 m travelled by a naïve bee during its first foraging bout in the pentagonal array (Figure 3A; Video S2; for further tracks see Figure S2). Thus, over multiple bouts, bees effectively minimized their travel distances using a relatively direct path to visit all flowers once in an optimal order.
Our second aim was to investigate how experienced bees modify their trapline in response to changes in the spatial configuration of flowers. Immediately after radar-tracking the bees in the regular pentagonal array, we removed the flower located in the top corner (location 3) and established a new flower east of the initial pentagon (location 6). This new location was chosen to maximise the probability that search flights would be performed in the catchment area of the radar (Figure 1). We recorded the flight paths of three of the seven trained bees for eight consecutive foraging bouts (Videos S14, S15, S16, S17, S18, S19, S20, S21, S22, S23, S24, S25, S26, S27, S28, S29, S30, S31, S32, S33, S34, S35, S36, S37).
After the removal of the familiar flower, the bees increased their flight duration by around five times (last bout in initial array, 245.00±32.87 s; first bout in modified array, 1221.67±894.81 s; n = 3 bees), their travel distance more than doubled (last bout in initial array, 455.75±33.91 m; first bout in modified array, 970.93±284.24 m; n = 3 bees), and they once again started to revisit empty flowers (last bout in initial array, 0 revisits; first bout in modified array, 4.33±2.33 revisits; n = 3 bees). Bees continued to follow their trapline, visiting all four familiar flowers and the empty feeding location (location 3) in the same sequence as before the spatial arrangement was modified (Figure 4, Table S1). However, as bees could not fill their crop to capacity by visiting only four flowers, they repeated the entire circuit once, sometimes twice before returning to the nest, a stereotyped pattern observed in 33.33%±15.02% of all their foraging bouts (n = 8 bouts per bee).
At the same time, bees engaged in local searching manoeuvres, exploring new areas of the experimental field (Figure 4). Azimuthal directions of the mean flight vectors (sum of all vectors of the radar track, see Materials and Methods) indicate that bees did not investigate the entire field (Watson's test for circular uniformity, p<0.01 for every bee), but each one restricted their searching activity to a different sector (average angle for individual 1, 75.09±2.91°; individual 2, −32.17±11.40°; individual 3, 32.38±13.14°; n = 8 tracks per bee; ANOVA for circular data, F2,23 = 30.31, p<0.001). Sixteen out of 24 flight paths included loops of varying length (range, 5.10–509.26 m) between immediate revisits to the same flower (Figure 4). During these loops, the bees' ground speed was significantly slower than during other nest-flower or flower-flower flight segments (speed during loop, 1.90±0.28 m.s−1, n = 25 loops; speed during segment, 3.72±0.07 m.s−1, n = 173 segments; GLMM, effect of flight type on speed, F1,197 = 41.16, p<0.001). Slow flight loops were also frequent in the paths of the naïve bee (loop length, 171.60±95.47 m, n = 12 loops; speed during loop, 1.49±0.61 m.s−1; bouts 1–4 in Figure S2), and were observed only once in the paths of experienced bees in the initial spatial arrangement (bout 36 of individual 1 in Figure 3B). This difference in flight speed suggests that bees alternated between phases of exploitation characterized by relatively fast and straight flight segments and phases of exploration characterized by slow and localised flight loops. A similar pattern has been observed in displaced honeybees, which typically exhibit fast vector flights in the expected direction of a familiar location followed by slow search curves after finding that the target is not in its expected location [25].
One bee (individual 1) found the new flower location during its first foraging bout following the rearrangement of flowers (Figure 4A), integrated it into a new optimal sequence (sequence, 12465) during the third bout, and gradually stabilized this new sequence into a trapline. The other two bees (individuals 2 and 3) confined their searching activity in different azimuthal directions and never found the new flower during the eight foraging bouts (Figure 4B and 4C). Wind direction had no significant influence on the bees' searching direction (correlation coefficient for angular variables, r = −0.21 p = 0.307). Thus, after the removal of a familiar flower, bees increased their frequency of immediate revisits to flowers exhibiting slow loops. These localised search flights might facilitate the discovery of new flowers by allowing bees to learn the spatial characteristics of new sectors of their environment, while still exploiting familiar flowers along their established trapline.
Having established that bees develop optimal traplines without trying all possible solutions and start exploring again if some flowers are removed from and/or introduced to the array, we further examined bees' optimisation strategy by comparing the observed visitation sequences to sequences generated by simple optimisation heuristics.
First, we tested the “nearest neighbour” heuristic, in which a model bee chooses the nearest unvisited flower as its next move until all flowers have been visited. This heuristic has been suggested to explain the routing behaviour of some animals [13],[14],[17],[19], including bees [19], at small spatial scales. When applied to our experimental situation (five flowers arranged in a regular pentagon) the nearest neighbour heuristic predicts that bees should always move between neighbouring flowers along the edges of the pentagon. Although a large proportion of the bees' movements involved linking nearest neighbour flowers, especially in the early bouts when all flowers were not yet discovered (77% of all transitions between flowers, n = 50 bouts) and after the stabilization of an optimal trapline (100% of all transitions between flowers, n = 19), this unique rule of thumb is not sufficient to fully explain our data since bees were observed moving between non-nearest neighbour flowers in 52% of the bouts in which all five flowers were visited (n = 42 bouts; Table S1).
Second, we tested the “discovery order” heuristic in which a model bee visits flowers in the order it discovered them. This heuristic has been previously proposed for the establishment of long-distance traplines by bees [16]. However, we found it incompatible with our observations as none of the bees used the discovery order of the flowers as their trapline sequence (Table S1). There was no significant relationship between the discovery order of the flowers and the directionality (clockwise or anti-clockwise) of final traplines (GLMM, effect of discovery order of flowers on their order in the trapline, F1,29 = 0.04, p = 0.844). For each bee, the similarity index between the discovery order sequence and the trapline sequence was not different than expected by chance (similarity index range, 0.29–0.67, n = 7 indices; p>0.05 for all bees, see Materials and Methods).
Third, we tested random optimization and implemented a simple random “k-opt” iterative improvement heuristic [12] assuming that (1) a model bee tries to improve the route between known flowers by randomly shuffling the order in which a number (k) of randomly selected flowers are visited and (2) the route change is kept if the new route is shorter than the previous one (otherwise it is rejected). This heuristic predicts the appearance of an optimal visitation sequence only after completion of around 100 foraging bouts, which is far higher than the 17.57±1.79 bouts (n = 7 bees) observed in our experiments. In general, random optimization processes do not produce stable repeatable sequences and are therefore not compatible with our data
We therefore developed an iterative improvement heuristic based on our analysis of bees' movement patterns. In this heuristic, the probability of model bees visiting a particular flower or flying back to the nest is determined by its experience, allowing them to explore, learn, and sequentially optimise their routes (Figure 5). We assume that (1) bees can uniquely identify the flower locations using information from path integration and/or the visual context (landmarks and/or panoramas) [26]; (2) bees have a finite transition probability between the nest and each flower and between any two flowers during the very first bout; (3) this initial probability is higher between nearest neighbour locations than between any other locations; (4) at each bout bees compute the net length of the route travelled (rather than the actual distance flown) by measuring the vector distance between successive flower visits and sum the lengths of all vectors comprising the route using path integration [27]; (5) if bees have visited all flowers at least once (and thus filled their crop), they compare the length of the current route to the memorised length of the shortest route travelled so far; and (6) if the new route is shorter, the probability of using the vectors composing this route are enhanced by a common factor.
According to our observations, a model bee during its first foraging bout between five flowers arranged in a regular pentagon is most likely to visit flowers 1 and 5 first because the other flowers are farther from the nest. Having found flower 1, the bee is most likely to find flower 2 next because flowers 3, 4, and 5 are more distant, and so on. The order in which flowers are discovered determines the probable order in which they will be visited during the next few foraging bouts; for example, from flower 1 a bee with aforementioned experience is most likely to visit flower 2 next (and more rarely move to flowers 3, 4, and 5). Nonetheless, as shown by our analyses on the flower visitation sequences of real bees (Figure 2C), these transition probabilities are not fixed and change whenever a shorter route is discovered. If a newly travelled route (e.g., sequence, 12453) is shorter than the shortest route experienced so far by that bee, then the probabilities linking movements between pairs of flowers within this circuit (1-2, 2-4, 4-5, and 5-3) are enhanced by a common factor. Gradual strengthening of the transition vectors forming the shortest route experienced so far allows the bee to sequentially optimise its visitation sequence and select an optimal route as a trapline (for more details about the model, see Text S1).
Simulation data from this novel heuristic predict that model bees (1) occasionally visit fewer than all flowers especially during early bouts, (2) regularly revisit empty flowers during the same bout, (3) decrease their frequency of revisits with experience, (4) establish stable optimal routes after about 20–25 bouts, and (5) can sequentially adjust their routes to incorporate newly discovered flowers in an optimal way (Figure S3). Quantitative evaluation of the simulated data with the optimisation performance of real bees in the experimental field showed full agreement for the number of bouts to the first appearance of an optimal sequence, the number of bouts to the stabilization of an optimal sequence into a trapline, the number of different routes experienced, the net route length travelled per bout, the number of revisits per bout, and the similarity indices between successive bouts (Table 1). Therefore, bees' optimization strategy can be captured in a simple iterative improvement routine in which an individual compares the net length of their current route to the net length of the shortest route experienced so far, and increases its probability of reusing the flight vectors comprising this new route if it is shorter.
We have recorded complete flower visitation sequences and successive flight paths of bumblebees foraging in field-scale conditions, allowing us to examine the learning processes underpinning multi-destination routing strategies of animals with known foraging history. Over multiple bouts, bees minimized their overall travel distance by flying relatively straight vectors between learnt feeding locations and visiting all flowers once in a stable optimal sequence. When the spatial configuration of flowers was modified, the bees engaged in localised search flights to find new flowers.
The observed dynamic of trapline acquisition in our large-scale setup is incompatible with random movements or with an extensive exploration of all possible routes. We also ruled out the hypothesis that bees rely on a single rule of thumb such as visiting all locations in the initial discovery order or moving between nearest neighbour locations. Although a large proportion of the bees' movements involved linking nearest neighbour flowers, especially in the first few foraging bouts, this strategy alone cannot explain our data. Rather, bees developed their traplines through trial and error by combining exploration, learning, and sequential optimisation, thus confirming hypotheses derived from previous observations in smaller enclosed environments [15],[19]–[21]. Interestingly, however, the optimisation performance of bees under field-scale conditions was much higher as all the bees tested selected an optimal route as their trapline, compared to a maximum of 75% in laboratory studies using comparable numbers of feeding locations (range, 4–10 flowers) and training durations (range, 20–80 foraging bouts per bee) [15],[19]–[21]. Presumably, bees' motivation to optimise their routes increases with spatial scale because the costs of travelling long (suboptimal) distances are greatly magnified. It is also possible that celestial cues, such as the position of the sun or polarized light patterns that are not typically available in laboratory settings but are known to be involved in navigation [28],[29], allow bees to orientate more accurately and develop routes faster in natural environments.
How, then, did the bees optimise their routes? Based on our detailed analysis of bee movement patterns, we implemented a simple iterative improvement heuristic, which, when applied to our experimental situation, matched the behaviour of real bees exceptionally well. The proposed heuristic demonstrates that stable efficient routing solutions can emerge relatively rapidly (in fewer than 20 bouts in our study) with only little computational demand. Our hypothetical model implies that a bee keeps in memory the net length of the shortest route experienced so far and compares it to that of the current route travelled. If the novel route is found to be shorter, the bee is more likely to repeat the flight vectors comprising this route. Hence, through a positive feedback loop certain flight vectors are reinforced in memory, while others are “forgotten”, allowing the bee to select and stabilize a short (if not optimal) route into a trapline. These assumptions are compatible with well-established observations that bees compute and memorise vector distances between locations using path integration [30]. For instance, bees visiting the same feeders over several bouts learn flight vectors encoding both direction and travel distance to each site, by associating specific visual scenes (such as salient landmarks or panoramas) with a motor command [26],[31].
The optimisation process we describe is analogous to the iterative improvement approach developed in “ant colony optimisation” heuristics, which has been increasingly used to explore solutions to combinatorial problems in computer sciences [32]. The rationale of these swarm intelligence heuristics is based on a model describing how ants collectively find short paths between a feeding location and their nest using chemical signals [33]. “Memory” in ant colony optimisation algorithms has no neurobiological basis but instead takes the form of pheromone trails marking established routes. The shortest route becomes more attractive due to increases in pheromone concentration as multiple ants forage simultaneously along it and continue to lay pheromone, while longer routes are abandoned because of pheromone evaporation. Of course, identification of a similar iterative optimisation principle in bees, although based on very different mechanisms (bumblebees forage individually and do not recruit using pheromone trails), does not imply that bees would equal the performance of swarm algorithms in finding solutions to complex combinatorial problems. However, iterative improvement heuristics are flexible, suggesting that bees can develop functional traplines in their natural environments, where the numbers of flowers, their spatial configuration, and reward values vary over time.
The question of how spatial information is encoded and processed in an insect brain is a matter of long-standing debate [25],[34]–[37]. Recent observations of honeybees using shortcuts between separately learnt foraging locations have been interpreted as evidence for “map-like” memory [25],[35], suggesting that bees acquire a coherent representation of the spatial connectivity between important locations in their environment (such as the nest, flowers, and prominent landmarks), allowing them to compute new vectors. Although our study was not conceived to test this hypothesis, our results indicate that the routing behaviour of bumblebees can be replicated without assuming such a map-like representation of space. The proposed heuristic suggests that bees can develop optimal routes by following multi-segment journeys composed of learnt flight routines (local vectors), each pointing towards target locations (flowers) and coupled to a visual context (landmarks and/or panoramas). Such a decentralized representation of space is akin to the “route-based” navigation of desert ants, where spatial information is thought to be processed by separate, potentially modular, guidance systems [4],[34],[37],[38]. The fact that trained bees continued to visit the familiar location from which a flower had been removed (location 3) further supports the hypothesis that foragers in our experimental situation relied heavily on learnt sensory motor routines as route-based navigation constrains the ability of individuals to rapidly adjust their routes, in contrast to map-like navigation that should allow for fast computation of entirely novel solutions [36]. Future studies should clarify whether similar learning heuristics apply to insect pollinators foraging at different spatial scales and configurations, and to other animals faced with similar routing problems (e.g., hummingbirds [9], bats [7], and primates [6],[10],[13],[14]). Ultimately, characterizing the neural-computational implementation of functional multi-destination routing solutions in small-brained animals holds considerable promise for identifying simple solutions to dynamic combinatorial problems in situations lacking central control.
Experiments were carried out in a flat, open area of mown pasture (approximately 700×300 m) on the Rothamsted estate (Hertfordshire, UK; Figure 1). Global landmarks (edges between different types of cut grass, tree lines) and local features (isolated trees) were available. The observation period (October 2010) was chosen because there were very few natural sources of pollen and nectar present during this time. The radar equipment was located on the south-east corner of the experimental field to allow maximum catchment area. The Bombus terrestris colony was housed in a wooden nest-box located south of the experimental field. A transparent tube with shutters was fitted at the entrance to control bee traffic. Bees were individually marked with numbered plastic tags within a day of emergence from pupae in order to monitor their complete foraging history.
Artificial flowers (Figure S1) were made of a plastic cylinder (height 8 cm) covered with a blue horizontal landing platform (diameter 6 cm). Bees could access the flowers equally well from all angles and collect a drop of 40% (w/w) sucrose solution from a yellow plastic square (2.4 cm side) in the middle of the landing platform. Each flower was positioned on top of a truncated cone-shaped support (base diameter 30 cm, top diameter 20 cm, height 18 cm) placed on the ground. A webcam (Logitech c250, Fremont, CA) was mounted directly above the centre of each flower on an independent vertical support (height 50 cm) to capture footage of bees when they visited. Webcams were fitted with light filters (neutral density 0.6) to attenuate sunlight illumination and connected to a laptop running video motion detection software (Zone Trigger 2, Omega Unfold, Quebec, Canada). A video clip (minimum duration 5 s) was recorded each time a bee moved into the camera field of view (Video S1). Recording continued until movement stopped, thus capturing complete flower visits from when the bee landed to its departure. Feeding stations were arranged sufficiently far apart (minimum distance 50 m) such that each station would be undetectable by the bee visual system from any other. The maximum dimension of a feeding station (including the flower, webcam, and laptop) was 70 cm. Given bumblebee's failure to detect targets that subtend less than ca. 3° [39], a bee should visually detect a feeding station this size from no more than 13.4 m away. Each laptop was powered by a small petrol generator (850W, length 38 cm, width 33 cm, height 32 cm) placed 10 m from the feeding station, located outside the pentagonal flower array (Figure 1). Generators provided potential local landmarks, although due to their small size they should only have been visually detectable for bumblebees at a range of 7.3 m and were therefore less prominent to bees than feeding stations. In addition, there is solid experimental evidence that bees could not visually detect feeding stations over the distances tested, since two out of three bees failed to find the new location after a displacement (see Results).
The harmonic radar and transponders have been previously described [23]. The radar equipment provided coverage over a range of 700 m and an altitude of about 3 m above the ground. Transponders consisted of a 16 mm vertical dipole (mass 0.8 mg) that does not affect bees' flight behaviour [24]. Individual bees were caught on departure from the colony, the transponder was attached using double-sided foam tape over the plastic number tag and released at the nest-box entrance tube. Coordinates of the transponder-tagged bee in the experimental field were recorded every 3 s by the radar with a spatial resolution of approximately ±2–3 m [40]. When the bee returned to the nest entrance, the transponder was removed before it re-entered the nest. Wind speed and direction were measured every 10 s by a recording anemometer fixed 2 m above the ground, located 10 m west of the nest-box (Figure 1).
Experiments were performed between 09:00 and 18:00 on days when the sun or blue sky was visible. Bees were individually pre-trained to collect sucrose solution from the five artificial flowers arranged in a linear array (150 cm length), located 50 m north-west of the nest entrance (Figure 1). Flower rewards were refilled ad libitum with 10 µL. The mean volume of sucrose solution ingested by a given bee during three successive foraging bouts was used to estimate its crop capacity (range = 75–100 µl) [20]. We tested seven bees, each on a different day.
In the first phase of the experiment, bees were observed foraging on the five flowers arranged in a regular pentagon (50 m side, Figure 1), until they visited all flowers in at least five consecutive foraging bouts. This required about 7 h of observation and 28.86±2.22 foraging bouts per bee (n = 7 bees). Each flower contained a sucrose reward equivalent to one-fifth of the test bee's crop capacity (volume range = 15–20 µl) and was refilled after each foraging bout. All departure and arrival times at the nest-box were recorded by an experimenter. Flower visits were automatically recorded using motion-activated webcams (Video S1). Flight paths of five bees were tracked with harmonic radar towards the end of training (up to four foraging bouts per bee, including the final bout).
In the second phase of the experiment, one flower was removed from location 3 and a new one was established at location 6 (Figure 1). Three bees were observed for eight additional foraging bouts in this new spatial arrangement (all these bouts were monitored by both webcam recordings and radar tracking). The five remaining bees were not tested because of insufficient daylight to pursue the observations on the day they were trained. In total, 230 foraging bouts, 1,354 video clips, and 36 radar tracks of flight paths were analysed.
|
10.1371/journal.pgen.1006215 | Control of Oocyte Reawakening by Kit | In mammals, females are born with finite numbers of oocytes stockpiled as primordial follicles. Oocytes are “reawakened” via an ovarian-intrinsic process that initiates their growth. The forkhead transcription factor Foxo3 controls reawakening downstream of PI3K-AKT signaling. However, the identity of the presumptive upstream cell surface receptor controlling the PI3K-AKT-Foxo3 axis has been questioned. Here we show that the receptor tyrosine kinase Kit controls reawakening. Oocyte-specific expression of a novel constitutively-active KitD818V allele resulted in female sterility and ovarian failure due to global oocyte reawakening. To confirm this result, we engineered a novel loss-of-function allele, KitL. Kit inactivation within oocytes also led to premature ovarian failure, albeit via a contrasting phenotype. Despite normal initial complements of primordial follicles, oocytes remained dormant with arrested oocyte maturation. Foxo3 protein localization in the nucleus versus cytoplasm explained both mutant phenotypes. These genetic studies provide formal genetic proof that Kit controls oocyte reawakening, focusing future investigations into the causes of primary ovarian insufficiency and ovarian aging.
| In mammals, oocyte reawakening controls female fertility, the onset of the menopause, and thus, overall aging. We demonstrate here through complementary genetic experiments that Kit is the upstream receptor regulating oocyte reawakening. Although other cell surface receptors have been proposed as candidates, the data have remained contradictory, and definitive genetic evidence in support of any one receptor has been lacking. We engineered two novel Kit alleles in mice, one an activating (gain-of-function) mutation, the other an inactivating (loss-of-function) mutation. These alleles permitted us to conduct elegant genetic experiments whereby Kit was activated or inactivated in the oocytes of newborn mice. The results were complementary and striking. Oocyte-specific Kit activation resulted in female sterility due to reawakening of all oocytes, leading to premature ovarian failure. In contrast, Kit inactivation also led to female sterility and ovarian failure, but through a contrasting and opposite phenotype: a complete failure of primordial follicle reawakening. Additional studies demonstrated that Foxo3, a known regulator of reawakening, was the mediator of both phenotypes, linking our findings to prior discoveries. These complementary genetic experiments thus definitively incriminate Kit as the upstream receptor regulating reawakening.
| Primordial follicles are the reserve precursor pool for maturing follicles throughout reproductive life [1]. Primordial follicles are reawakened via an ovarian-intrinsic (gonadotropin independent) process whereby they are selected from the quiescent reserve into the growing follicle pool [2, 3]. The morphologic hallmark of reawakening is oocyte growth, and this is followed by a transition of the surrounding granulosa cells from a flattened to a cuboidal shape [4]. Reawakening is irreversible, in that follicles that have initiated growth undergo atresia if not selected for subsequent stages of maturation. Primordial follicle numbers decrease with advancing age due to oocyte reawakening or apoptosis; following follicle depletion, ovulation ceases and reproductive senescence ensues. Reawakening must therefore be metered throughout reproductive life to ensure that some growing follicles are available during each estrus cycle, but at the same time, limit the number of growing follicles to forestall depletion of primordial follicles (see Fig 1 for summary schematic of follicle maturation). Characterization of the molecular mechanisms underlying reawakening remains an important challenge in reproductive biology [5–8].
The forkhead transcription factor Foxo3 functions as a switch that controls (suppresses) reawakening. In nullizygous or oocyte-conditional Foxo3 knockout mice, primordial follicles are assembled normally [9], but undergo global reawakening at birth. This leads to a characteristic sequential syndrome of ovarian hyperplasia, follicle depletion, and hypergonadotropic ovarian failure [10]. In the adult ovary, Foxo3 protein localizes to the primordial oocyte nucleus, where it restrains reawakening in a PI3K (phosphoinositide-3-kinase)-AKT dependent manner. PI3K catalyzes the formation of a lipid second messenger, phosphatidylinositol 3,4,5-trisphosphate (PIP3), from phosphatidylinositol 4,5-bisphosphate (PIP2). PIP3 in turn leads to the phosphorylation and activation of AKT. Pten, a lipid phosphatase that converts PIP3 back to PIP2, potently suppresses AKT activation [11]. Oocyte-specific deletion of Pten hyperactivates AKT, resulting in Foxo3 nuclear export and global reawakening [12, 13]. Foxo3 normally undergoes nuclear export during the primordial to primary follicle transition (followed by its degradation within the cytoplasm via unknown mechanisms), suggesting that Pten inactivation mimics a naturally-occurring PI3K-dependent signal that regulates Foxo3 localization and hence reawakening [13]. Other studies have confirmed Foxo3’s role as the molecular switch controlling reawakening in a PI3K-AKT dependent manner, and the utility of Foxo3 nuclear vs. cytoplasmic localization as a marker of oocyte maturation [14–18]. Oocyte-specific inactivation of the mTOR inhibitors Tsc1 and Tsc2 also results in oocyte reawakening, establishing an important but incompletely understood role of mTOR signaling in this process [19].
As vital effectors of PI3K-AKT signaling, the Foxos serve fundamental biological roles in aging, cancer, and stem cell maintenance [20–22]. The three canonical Foxos—Foxo1, Foxo3, and Foxo4—are coexpressed and exhibit genetic and functional redundancy in most cell types [20, 21, 23]. In contrast, in the mouse germline, Foxo1 and Foxo3 have diverged to serve complementary roles in the maintenance of the male and female germline, respectively. Whereas Foxo3 is the principal Foxo protein in the oocyte, Foxo1 is the principal Foxo within undifferentiated spermatogonia, and is the only Foxo required for the maintenance of spermatogonial stem cells [20, 24]. In both the female and male germline, genetic experiments have shown that Foxo activity is regulated by its subcellular (nuclear vs. cytoplasmic) localization via its AKT-dependent phosphorylation. When phosphorylated, Foxo proteins are functionally inhibited by their retention in the cytoplasm via interactions with 14-3-3 proteins [25].
In many physiological processes, the upstream activators of class I PI3Ks are transmembrane receptor tyrosine kinases (RTKs) such as Igf1r, insulin receptor, Ret, Pdgfr, or Kit [11]. Ligand binding activates PI3K by phosphotyrosine-mediated binding through an SH2 domain on the p85 subunit of PI3K. G protein–coupled receptors (GPCRs) can also activate PI3K through the p110γ catalytic subunit isoform. However, p110γ−/− mice are viable and fertile (but display various GPCR-mediated immunological defects), suggesting that GPCRs may not play essential roles in the regulation of the PI3K pathway during oocyte reawakening [26]. The identification of a presumptive oocyte surface RTK that acts through PI3K-AKT-Foxo3 to regulate reawakening has remained an outstanding question in reproductive biology [27]. Several candidates including Kit have been proposed, but definitive evidence about which is the bona fide receptor has been lacking [3, 8, 28–30]. The unique biological features of primordial follicle reawakening, some of which are not readily modeled in vitro, prompted us to apply genetic approaches to identify and validate this factor [3].
Kit is an RTK that acts via PI3K-AKT, and is expressed within primordial oocytes [31, 32]. To study the role of Kit in the reawakening of primordial follicles, we generated a novel murine conditional allele, KitD818V(L) through homologous recombination in murine embryonic stem cells. The Kit D818V (Asp→Val) amino acid substitution leads to constitutive Kit activity in the absence of ligand (KL), and exerts potent dominant gain-of-function effects [31, 33]. Furthermore, it corresponds to the most common known Kit gain-of-function mutation in human germ cell tumors (KitD816V), demonstrating that this mutant protein is active in germ cells [33–35]. The conditional (floxed) KitD818V(L) allele was designed to provide Kit function through a 3’ cDNA cassette encoding exons 17–21. Cre-mediated recombination excises this cDNA cassette, permitting normal splicing of an exon 17 harboring the mutation, and thus expression of mutant D818V protein (S1A Fig). With respect to nomenclature for the four new Kit alleles described in this manuscript, 1) genotypes signify somatic genotypes (per tail DNA genotyping) and 2) the floxed (i.e., latent) alleles end in “(L)”. Whereas hemizygous Kit (a.k.a. Dominant white spotting) loss-of-function mutations produce abnormal coat pigmentation [36] (see also below), mice harboring the KitD818V(L) allele were externally normal with coat pigmentation similar to sibling controls, confirming that the floxed allele indeed provided Kit function (S1B Fig). The KitD818V(L) allele could be homozygosed, and such animals were also externally indistinguishable from littermate controls (S1C Fig).
These mice were then bred to the germ cell-specific Cre driver, Vasa-Cre (a.k.a. Ddx4-cre1Dcas/J) (abbreviated VC) to generate VC; KitD818V(L)/+ females. VC becomes active during late embryogenesis, and drives Cre-mediated recombination in >99% of oocytes by birth [37]. Ovaries were harvested at postnatal day (PD) 7 and ovarian cDNA was analyzed by RT-PCR, followed by Sanger sequencing. As expected, ovaries from experimental females expressed mutant cDNA at levels close to wild-type as evidenced by electrophoretogram peak intensities (S1D Fig). By PD7, VC; KitD818V(L)/+ ovaries were consistently larger than ovaries from sibling controls. By PD14, these size differences were even more marked, but from PD28 onward (up to 16 weeks of age), ovaries were of equal size or somewhat smaller (Fig 2A). To understand the cellular basis of this increase in ovarian size (and subsequent decrease), tissue sections were analyzed. At PD7, there was an obvious increase in oocyte diameters in follicles that otherwise resembled primordial follicles (i.e., follicles without granulosa cell growth or change to cuboidal shape) (Fig 2B). Interestingly, whereas Kit protein is predominantly membranous in controls, Kit protein underwent a general redistribution to the cytoplasm in VC; KitD818V(L)/+ oocytes (S2A–S2C Fig) with overall Kit protein levels comparable to KitD818V(L)/+ controls (S2D Fig). This is consistent with prior data demonstrating that Kit protein normally undergoes ligand-dependent internalization, and that constitutively active mutant variants are internalized more efficiently than the wild-type protein [38, 39].
Of note, these morphological alterations were global, occurring in all primordial follicles. VC; KitD818V(L)/+ oocytes grew in size up to PD28, resulting in aberrant follicles with dramatically enlarged oocytes (Figs 2B and S4A). Some morphologically normal follicles and corpora lutea stage were also present, indicating that some of the reawakened follicles progressed normally to more advanced states of follicle maturation. Concordantly, most markers of primordial or primary follicles including periodic acid-Schiff (PAS) stain (labels the zona pellucida), ZP1, Inhibin, Gdf9, Sohlh1, Nobox, and Sall4 retained their typical patterns of expression in oocytes or granulosa cells, consistent with normal differentiation despite global reawakening (S3 and S4A Figs). However, in some reawakened follicles, granulosa cells remained flattened and were negative for α-Müllerian Hormone (AMH), which is normally induced at the primary follicle stage (S4A Fig). A similar spectrum of abnormalities has been documented in other global reawakening mutants such as Foxo3 and Pten [10, 12, 13].
By 16 weeks, however, oocyte atresia occurred, resulting in morphologically abnormal, “empty” follicles depleted of oocytes (note also the absence of more advanced follicles and corpora lutea) (Fig 2B). No teratomas or other ovarian tumors were identified. Thus, constitutive Kit activation in the female germline resulted in a classic, global primordial follicle reawakening phenotype identical to that described for Foxo3 and Pten [10, 12, 13]. Global reawakening occurred in all VC; KitD818V(L)/+ primordial oocytes, which ultimately underwent atresia resulting in loss of all oocytes with premature ovarian failure. Serum follicle stimulating hormone (FSH) and luteinizing hormone (LH) levels at 5 months of age were elevated in adult mutant females, consistent with hypergonadotropic hypogonadic premature ovarian failure (P<0.003 and P<0.02 respectively) (S4B Fig). These results strongly implicate Kit as the upstream RTK regulating primordial oocyte reawakening.
At birth to PD7, oocyte numbers were unaltered in VC; KitD818V(L)/+ ovaries, indicating a normal initial endowment of oocytes [3]. Oocytes were markedly depleted by PD28, and totally absent by 6 weeks of age (P<0.02 and P<0.0004, respectively) (Fig 3A). Measurements of oocyte diameters confirmed that average oocyte sizes were significantly increased by PD7, and the difference was even more marked at PD14 (P<0.0001 for both PD7 and PD14). Oocytes continued to grow through PD28, although by this timepoint, relatively few oocytes remained (Figs 3B, 3C and S4A). Sections from experimental and control ovaries embedded in plastic confirmed the absence of oocytes by 6 weeks of age as well as the presence of atretic oocytes and zona pellucida remnants (the “ghosts” of oocytes that underwent reawakening and subsequent atresia) in VC; KitD818V(L)/+ females (S4C Fig).
Next, various markers were analyzed by immunohistochemistry at PD7 to further characterize potential abnormalities in downstream effectors such as Foxo3. All oocytes were strongly Vasa-positive, indicating preservation of germline identity following Kit activation. Kit protein was present at high levels within oocytes, although some redistribution to the cytoplasm was evident in the mutant as was the case by immunofluorescence (Fig 4A). Activated early oocytes exhibited increased phosphorylated AKT (P-AKT) on their cell membranes as compared to control primordial and primary follicles, which did not contain detectable P-AKT (Fig 4A). This is consistent with the known role of PI3K-AKT as the key signaling pathway mediating oocyte reawakening [13], and argues strongly that Kit controls reawakening via this pathway. Importantly, AKT hyperphosphorylation drove translocation of Foxo3 protein from the nucleus to the cytoplasm in all VC; KitD818V(L)/+ oocytes (Fig 4A). Confocal microscopy, which allows for higher resolution than immunohistochemistry, confirmed these results; i.e. Foxo3 was predominantly nuclear in controls, but cytoplasmic in the mutant oocytes (Fig 4B).
The above results provided strong genetic evidence implicating Kit as the upstream RTK controlling oocyte reawakening via the PI3K-AKT-Foxo3 axis. However, phenotypes associated with gain-of-function mutations should generally be interpreted with caution, as even a single amino acid substitution could have multiple, distinct, and potentially unexpected effects on protein function and thus, phenotypes. To expand upon our genetic analyses of Kit in oocyte reawakening with a complementary genetic approach, we designed a new allele where exon 17, which encodes the kinase domain [31], was floxed. Kit is an essential locus due to its requirement for hematopoiesis, necessitating conditional genetic analysis. The floxed allele KitL could be homozygosed as expected, and KitL/KitL animals were born at expected Mendelian ratios (S5A and S5B Fig). For oocyte-specific Kit inactivation, VC; KitL/+ fathers were bred to KitL/KitL females. VC activity in the father’s germline converts any paternally transmitted KitL allele to Kit- and thus, VC fathers can transmit a wild-type (Kit+) or null Kit-allele but not a KitL allele [37]. Experimental females of the VC; KitL/- genotype (per tail DNA) thus harbor homozygous Kit- loss-of-function mutations in their germline. RT-PCR of ovaries (S5C Fig) and Sanger sequencing of PCR-amplified cDNAs (S5D Fig) confirmed VC-dependent deletion of exon 17. Whereas control mice harboring the floxed allele exhibited no pigmentation or other external abnormalities, VC; KitL/- mice showed striking midline hypopigmentation, consistent with hemizygous somatic loss of Kit activity (S5E Fig).
At PD7, VC; KitL/-ovaries were minute, and their small size persisted at all mouse ages analyzed, up to 12 weeks of age (Fig 5A). Histological analyses revealed a striking and complete failure of oocyte reawakening. Follicles remained small and no growing oocytes were present, although granulosa cells became cuboidal. In these follicles, oocytes typically assumed an eccentric location (Figs 5B and S6A). Electron microscopy (EM) confirmed the viability of these oocytes and their lack of physical growth. Granulosa cells exhibited some mitotic activity in the mutant follicles per Ki67 immunohistochemistry (whereas normal primordial follicles are mitotically inactive) (S6D Fig), consistent with increased granulosa cell numbers in the aberrant follicles of VC; KitL/-ovaries. Interestingly, EM also revealed abundant lipid droplets in the granulosa cells, which could contribute to their change in shape (S6B Fig). The significance of these lipid droplets is uncertain. One interpretation is that granulosa cells “sense” the lack of oocyte growth (due to known, if poorly understood, bidirectional oocyte/granulosa cell communication), and respond in some manner to promote reawakening [5, 40]. Alternatively, an abnormal hormonal milieu associated with ovarian failure could also indirectly contribute to the observed changes in granulosa cells. In any case, ZP1, which is expressed in growing oocytes, was not expressed in any VC; KitL/- oocytes, consistent with a constitutional inability to reawaken/initiate growth despite the granulosa cell changes.
Oocyte numbers were unaltered in VC; KitL/-ovaries at PD7, demonstrating that the minute ovaries were due to a complete lack of oocyte growth, and not a diminished initial endowment of primordial follicles (Fig 6A). Measurements of oocyte diameters confirmed the lack of oocyte growth (Fig 6B and 6C). These quantitative and morphometric data thus revealed a complete failure of oocyte reawakening in Kit-deficient oocytes. Somewhat unexpectedly given prior studies implicating Kit in germ cell and primordial oocyte survival, VC; KitL/- oocytes did not undergo rapid apoptosis [28, 41]. To the contrary, VC; KitL/- primary/primordial oocyte counts showed only a minor (statistically not significant) decrease even at 12 weeks of age, consistent with a remarkably specific role for Kit in oocyte reawakening (Fig 6A). However, by 6 months of age, the ovaries were entirely depleted of follicles and oocytes and contained only luteinized stroma (S7A–S7C Fig), demonstrating that Kit is required for the long-term maintenance of oocytes, in keeping with prior studies implicating Kit as an oocyte survival factor [28, 42].
At birth, oocytes are syncytial and interconnected by intercellular bridges, which are broken down by PD3 to give rise to individualized primordial follicles. Follicle individualization (also known as assembly) occurred normally in VC; KitL/-ovaries (e.g., no follicles contained more than one oocyte), demonstrating that Kit is not essential for individualization despite its abundant expression within oocytes at PD1-3 when individualization takes place (Fig 5B). Additional marker studies at 6 weeks of age showed that all oocytes retained germline identity, with normal expression of primordial oocyte markers such as Vasa, p63 (an oocyte survival factor), Foxo3, Sohlh1, and Nobox (Fig 7A). Granulosa cells continued to express inhibin and AMH (markers of female gonadal somatic cell differentiation) but did not express the Sertoli cell marker Gata-1 (Fig 7C and 7D), evidence against a sex-reversal phenotype, a possibility entertained because of the morphologic resemblance of the aberrant follicles—particularly those with eccentric oocytes—to primitive male sex cords. These results are consistent with a specific role of Kit in oocyte reawakening. Additionally, Foxo3 was constitutively nuclear and P-AKT was undetectable in Kit-deficient oocytes (Fig 7A and 7B) further supporting a critical role of the Kit signaling pathway in regulating oocyte awakening via P-AKT/Foxo3.
Primary ovarian insufficiency (POI), also known as premature ovarian failure, is a form of hypergonadotropic hypogonadic ovarian failure that causes early menopause and infertility in 1% of women before the age of 40, in addition to other important health consequences due to estrogen deficiency [43]. POI is associated with the accelerated depletion of primordial follicles, arguing that abnormalities in primordial follicle maintenance are the unifying pathophysiologic basis of POI. However, the identification and further study of key factors regulating primordial follicle maintenance and reawakening has proven difficult with human subjects; for example, ovaries are not biopsied in the clinical workup of POI, limiting available human ovarian tissue for direct analyses. Genome-wide studies have implicated several loci, but further and more detailed genome-wide investigations are needed to more fully define the genetic basis of POI [44–47]. Numerous factors have been proposed as regulators of reawakening, such as the AMH type 2 receptor (AMHR2), a member of the transforming growth factor β superfamily of growth and differentiation factors [48]. AMHR2 female mice are fertile with no overt defects in follicle maturation, however, arguing against an essential role in reawakening [49]. Similarly, while α-Müllerian hormone (AMH) has also been proposed as a regulator of oocyte utilization and reawakening, AMH-null females are fertile, suggesting that AMH plays at most a secondary role in regulating reawakening. Other factors must therefore participate in this process [50].
Kit has also been proposed as a candidate factor regulating reawakening. Kit and Kit ligand (KL, also known as SCF) serve diverse functions in the germ cell lineage, particularly in primordial germ cell migration/survival, primordial follicle assembly [51], and spermatogenesis. KL is produced by granulosa cells in primordial follicles [32] and Kit is highly expressed by primordial oocytes. Thus, patterns of Kit and KL expression within the ovary make them plausible candidates as factors regulating reawakening. In vitro studies conducted with explanted ovaries treated with KL or Kit inhibitors have also been interpreted as supportive of this hypothesis [52, 53]. On the other hand, the observed effects have been relatively small in magnitude and more importantly, these in vitro, non-genetic approaches suffer from significant potential limitations stemming from the unique biological properties of reawakening. Reawakening normally occurs at a gradual and measured pace throughout life, such that only a very small percentage of follicles are awakening at any time. Thus, available in vitro methods with explanted neonatal ovaries (which can be maintained for only one to two weeks) do not provide a completely adequate timescale for definitive studies including the identification of novel factors regulating reawakening.
Phenotypic analysis of KitY719F homozygous female mice have provided intriguing genetic evidence implicating Kit in reawakening [28]. Kit activates PI3K through a direct interaction with an SH2 domain on the p85 regulatory subunit of PI3K. This interaction is dependent on Kit tyrosine residue 719, which undergoes autophosphorylation following ligand binding. Mutation of this tyrosine residue prevents binding of Kit to p85, and thus abrogates Kit signaling via PI3K [54]. Mice engineered with this KitY719F mutation (via a “knockin” approach) have permitted genetic dissection of the contribution of PI3K signaling to diverse Kit-dependent biological processes [55, 56]. Homozygous KitY719F males are sterile with severe defects in spermatogenesis. KitY719F females, however, are fertile at 16 weeks of age, suggesting that Kit might not play an essential role in reawakening (although severe defects in early follicle maturation including an arrest at the primary/secondary follicle stage were documented) [28]. Subsequent analyses of this allele have been interpreted as supportive of a role of Kit in reawakening, rationalizing its use in studies to genetically dissect reawakening. For example, genetic inactivation of Tsc1 within primordial follicle granulosa cells leads to global reawakening via Foxo3 cytoplasmic relocalization, and this was found to occur through enhancement of Kit ligand production in primordial follicle granulosa cells. KitY719F homozygosity suppressed this Tsc1-null reawakening phenotype, arguing that the observed increase in Kit ligand was responsible for reawakening [57].
Foxo3 relocalization to the cytoplasm occurs after the primordial-primary transition, and abnormal Foxo3 relocalization has been documented in mutants undergoing global reawakening [12, 13, 16, 57]. However, for reasons that are not well understood, Foxo3 relocalization as visualized by IHC/IF does not anticipate (precede) reawakening; i.e., early primary follicles have nuclear Foxo3 [13]. It is possible that Foxo3 protein can be functionally inactivated in the oocyte but that the protein remains detectable in the nucleus for some time. Thus, in practice, the most useful, sensitive, and earliest indicators of reawakening (i.e., for scoring mutant phenotypes) remain morphologic. As emphasized in this study, oocyte diameter is the earliest and most sensitive endpoint for reawakening. In diverse investigations of mutants with global reawakening (e.g., Pten and Foxo3 mutants, also in the KitD818V mutant described here), granulosa cells remain flattened and unilayered in many follicles with massively enlarged oocytes that have clearly undergone reawakening; indeed, such oocytes can continue to grow unabated for several weeks despite the persistence of “primordial follicle-like” granulosa cell morphology [10, 13, 16, 19, 58]. Furthermore, we have here documented in VC;KitL/- ovaries a transition in granulosa cell morphology to a cuboidal shape and mitotic activity resulting in a few layers even in follicles where oocytes were constitutionally unable (due to Kit inactivation) to grow/reawaken. Thus, while granulosa cells normally undergo significant changes in morphology and arrangement during reawakening, these do not always correlate with oocyte status in individual follicles and are not as useful as morphological indicators in mutants with strong reawakening phenotypes. Thus, we stress that abnormal oocyte diameters should be considered the sine qua non for scoring oocyte reawakening phenotypes and that careful measurements of oocyte diameters (e.g., in tissue sections) should represent the gold standard in such analyses until earlier molecular markers of reawakening are identified.
That the KitY719F mutation did not lead to a complete oocyte dormancy phenotype, as documented in earlier studies (females were fertile up to at least 16 weeks of age) can be explained by the hypomorphic nature of this mutation [28]. For example, residual Kit signalling via PI3K or other surrogate signalling pathways could feed back to PI3K, resulting in some PI3K tonic activity. Finally, some hypomorphic alleles of KL (a.k.a. Steelpanda) also accumulate follicles with cuboidal granulosa cells (albeit with growth-arrested oocytes), consistent with a role of KL in reawakening, although these ovaries contain very few follicles, obscuring interpretation of phenotypes [28, 59]. The role of KL in reawakening should be further explored in future studies; i.e., through conditional genetic analysis of KL in granulosa cells. KL appears to be constitutively expressed in granulosa cells in primordial follicles [60], and although KL is elevated in the Tsc1 reawakening mutant, the nature of the signals and inter/intra-follicular communication triggering reawakening via Kit-PI3K-Foxo3 in individual primordial follicles remains uncertain. More recently, conditional deficiency of the Lim homeodomain protein Lhx8 was found to promote global reawakening, with a synergistic effect of Lhx8 and Pten on Foxo3 nucleocytoplasmic translocation. These effects were mediated by Lin28a, an RNA binding protein and regulator of the let-7 microRNAs, a finding of particular interest given studies implicating the Lin28/let-7 axis in the control of PI3K-mTOR signalling [16, 61]. Clearly, more work is needed to fully dissect the interactions of the diverse members of the PI3K and mTOR pathways and how these function together to trigger reawakening of individual primordial follicles.
One notable aspect of this study is that we have now documented global oocyte reawakening and dormancy phenotypes with our KitD818V and KitL alleles, respectively. Such complementary and contrasting phenotypes with alleles that have opposite effects on Kit activity (i.e., gain- vs. complete loss-of-function alleles) represent compelling genetic evidence incriminating Kit as the critical receptor upstream of PI3K-Foxo3 in the control of reawakening. The phenotype we describe for the oocyte conditional Kit knockout represents the first report of a pure global dormancy mutant; i.e., where female sterility was associated with minute ovaries containing a numerically normal complement of primordial follicles but where a complete failure of oocyte reawakening left oocytes incapable of growth/reawakening. Conversely, Kit serves essential roles in other aspects of follicle formation, development, and survival [3, 28, 51].
Our genetic studies do not exclude a contribution from other cell surface receptors in reawakening, but help establish Kit as the principal upstream factor regulating reawakening, and also demonstrate a remarkably specific role for Kit in reawakening. It is surprising that Kit played such a modest role in oocyte survival, with normal oocyte counts up to 12 weeks of age, a result suggesting that other factors regulate oocyte survival and prevent apoptosis [62]. Oocyte conditional ablation (Vasa-Cre) of other genes encoding cell surface receptors acting through PI3K and known to be expressed on the oocyte, such as Igf1r, Insr, Ret, Pdgfr had no effect on fertility or follicle maturation (unpublished data), further stressing the uniqueness of the Kit reawakening phenotypes.
POI is a form of hypergonadotropic hypogonadism that causes infertility in 1% of women before the age of 40 and has important health consequences. POI is due to accelerated depletion of primordial follicles [63–66], arguing that abnormalities in primordial follicle maintenance are the unifying pathophysiologic basis of POI. However, the molecular mechanisms that cause follicle depletion in POI are poorly understood. Elucidating these mechanisms is needed to develop better treatment strategies and assays predictive of POI. Our results should help focus further investigations on Kit and, by extension, Kit ligand, as keystone regulators of primordial oocyte maintenance and hence, female reproductive aging. Pharmacologic manipulation of this pathway may someday prove useful in fertility preservation by increasing the pool of actively growing follicles [14]. This is further underscored by the fact that PI3K-Foxo3 signaling regulates the egg supply throughout life [28]. For example, transient treatment of mouse or human oocytes with either a Pten inhibitor or a PI3K activating peptide results in activation of dormant primordial follicles. Pharmacologic control of this pathway either at the level of PI3K-AKT or Kit-KL may thus prove useful in fertility preservation by increasing the pool of growing follicles [14, 15, 67].
All live animal experiments followed guidelines by the UTSW Animal Care and Use Committee who also approved the experiments performed (2015–101272).
Mice were housed in a pathogen-free animal facility in microisolator cages and fed ad libitum on standard chow under standard lighting conditions. A detailed description of the targeting strategy, assembly, primer sequences, and PCR conditions for generating the KitD818V(L) and KitL alleles are provided in S1 Text.
Mice were genotyped from tail DNA by PCR with Promega GoTaq in 1.6 mM MgCl2. Genotyping primers for the KitD818V allele were as follows: (a1) 5’-ATTAGAGCCCCGATCCTGTG-3’ and (b1) 5’-GCAACAGCCATTCATTTCAGC-3’ (see S1 Fig for positions of a1/b1 primers), under the following cycling conditions: 95° x 2 min; 94° x 30 sec, 60° x 30 sec, 72° x 30 sec (35 cycles); 72° x 7 min. The product sizes are 219 bp for the floxed allele KitD818V(L) and 171 bp for the wild type Kit allele. Genotyping primers for KitL allele are as follows: (a1) 5’-AGTTCTGAAGAGACTGTCAAGGT-3’ and (b2) 5’-ACACCCCATTTCCTTATTTTTGCT-3’(see S3 Fig for positions of a1/b1 primers), under the following cycling conditions: 95° x 2 min; 94° x 30 sec, 60° x 30 sec, 72° x 30 sec (35 cycles); 72° x 7 min. The product sizes are 174 bp for the floxed KitL allele and 126 bp for the wild type Kit allele.
Total RNA was prepared from ovaries with the Tripure isolation reagent (Roche #93876820) per the manufacturer’s instructions. To validate the KitD818V(L) allele, exon 17 was amplified by one step RT-PCR (Qiagen #210210) using the primers: 5’-AGATTTGGCAGCCAGGAATA-3’ (forward) and 5’-ATTTCCTTTGACCACGTAATTC-3’ (reverse). For validation experiments of the Kit- allele, cDNA was synthesized with M-MuLV Reverse Transcriptase (New England BioLabs #28025–013). Kit cDNA sequences, spanning exons 14 to 18 were amplified by Phusion High-Fidelity DNA Polymerase (New England BioLabs #M0530S). The forward primer (close to exon 14) was 5’-GAGAAGGAAGCGTGACTC-3’; the reverse primer (close to exon 18) was 5’-AGGAGAAGAGCTCCCAGA-3’. PCR conditions were: 98°C x 3 min; 34 cycles of 98°C x 60 sec, 60°C x 60 sec, 72°C x 120 sec; then 72°C x 5 min. RT-PCR products were analyzed by gel electrophoresis and bands purified with the QIAquick gel extraction kit (Qiagen #28704). The KitD818V mutation or exon 17 deletion were confirmed by Sanger sequencing (UTSW sequencing core).
Ovaries were fixed in 10% formalin overnight, embedded in paraffin and serially sectioned (5 μm). Every fifth section was H&E stained and analyzed. For VC; KitD818V(L)/+ and control ovaries, the middle section of the series was used for relative follicle counts and diameter measurements as described in the text. For VC; KitL/- and control ovaries, all the primordial and primary oocytes were counted in every fifth section. The longest diameter of 50 oocytes for which nuclei were in the plane of section was determined with ImageJ. For IHC, sections were rehydrated in EtOH series after deparaffinization in xylene. Antigen retrieval was performed in parboiling 10 mM sodium citrate (pH 6.0) x 20 min, cooled at RT, followed by peroxidase blocking (3% H202) and blocking in 0.5–1% BSA in PBS. Antibodies and titers for IHC were: Kit 1:750 (Cell Signaling #3074S); Vasa 1:200 (Abcam #27591); P-AKT 1:75 (Cell Signalling #S473); Sall4 1:1000 (Abcam #57577); Foxo3 1:50 (Santa Cruz #sc-11351); AMH 1:500 (Serotec #MCA2246); P63 1:500 (Thermo Fisher #MS-1081); Inhibin (Biorad #MCA951ST); Gata-1 (Santa Cruz #sc-265); Zp1 (Santa Cruz #sc-23708); Gdf9 (Santa Cruz #sc-12244); Ki67 (Abcam #15580); Sohlh1 and Nobox antibodies were kindly provided by Dr. Aleksandar Rajkovic (Magee-Womens Research Institute). ImmPRESS (Vector Laboratories) was used for detection. For immunofluorescence (IF), sections were rehydrated in an EtOH series after deparaffinization in xylene. Antigen retrieval was performed in parboiling 10 mM sodium citrate (pH 6.0) x 20 min, cooled at RT, followed by PBS washes, 10 minutes of autofluorescence blocking (100 nm Tris Glycine), and blocking in 2.5% BSA+5% goat serum (Vector Laboratories) in PBS. Antibodies and titers for IF were: Foxo3 1:50; Kit 1:100. Fluorophore conjugated secondary mouse (Alexa Fluor 555) or rabbit (Alexa Fluor 488) IgGs at 1:500 (Invitrogen #A21429 and #A11029) were used for detection. Zeiss LSM 510 confocal microscopy was used for fluorescence imaging.
Serum FSH/LH levels were measured by the Ligand Assay Core at the University of Virginia. Serum was diluted 1:10 prior to analysis.
Ovaries (two/genotype) were homogenized in RIPA buffer supplemented with Complete Proteinase inhibitor cocktail (Roche) and Phosphatase Inhibitor cocktail 2 (Sigma). Following homogenization, extracts were centrifuged at 10000 rpm for 7 min. Supernatants were collected, mixed with 4x Laemmli Sample Buffer (BioRad) and boiled for 10 min. Equal amounts were loaded in a 10% SDS PAGE and run at 100V. Gel was transferred to Immobilon P membrane (Millipore). The membrane was blotted with 5% dry milk in TBS-T (BB), and probed with 1:1000 dilution of Kit antibody in BB overnight at 4°C (Cell Signaling, #3074). The membrane was stripped with Restore Stripping Buffer (Thermo Scientific), blotted and probed with 1:5000 α-tubulin in BB for 1hr at room temperature (Sigma, T9026). Blots were developed with SuperSignal West Dura Extended Duration Substrate (Thermo Scientific) and digital images acquired with a ChemiDoc system (BioRad).
Two ovaries per genotype were fixed, embedded in Embed 812 resin (Electron Microscopy Sciences) and prepared for negative staining as described [9]. Images were acquired using a FEI Tecnai G2 Spirit electron microscope. Thin sections were also stained with Toluidine blue and analyzed with light microscope.
P values and means +/- S.E.M. were calculated by two-tailed unpaired Student’s t test with GraphPad Prism 6.
|
10.1371/journal.pcbi.1002238 | Ensemble-Based Computational Approach Discriminates Functional Activity of p53 Cancer and Rescue Mutants | The tumor suppressor protein p53 can lose its function upon single-point missense mutations in the core DNA-binding domain (“cancer mutants”). Activity can be restored by second-site suppressor mutations (“rescue mutants”). This paper relates the functional activity of p53 cancer and rescue mutants to their overall molecular dynamics (MD), without focusing on local structural details. A novel global measure of protein flexibility for the p53 core DNA-binding domain, the number of clusters at a certain RMSD cutoff, was computed by clustering over 0.7 µs of explicitly solvated all-atom MD simulations. For wild-type p53 and a sample of p53 cancer or rescue mutants, the number of clusters was a good predictor of in vivo p53 functional activity in cell-based assays. This number-of-clusters (NOC) metric was strongly correlated (r2 = 0.77) with reported values of experimentally measured ΔΔG protein thermodynamic stability. Interpreting the number of clusters as a measure of protein flexibility: (i) p53 cancer mutants were more flexible than wild-type protein, (ii) second-site rescue mutations decreased the flexibility of cancer mutants, and (iii) negative controls of non-rescue second-site mutants did not. This new method reflects the overall stability of the p53 core domain and can discriminate which second-site mutations restore activity to p53 cancer mutants.
| p53 is a tumor suppressor protein that controls a central apoptotic pathway (programmed cell death). Thus, it is the most-mutated gene in human cancers. Due to the marginal stability of p53, a single mutation can abolish p53 function (“cancer mutants”), while a second mutation (or several) can restore it (“rescue mutants”). Restoring p53 function is a promising therapeutic goal that has been strongly supported by recent experimental results on mice. Understanding of the effects of p53 cancer and rescue mutations would be helpful for designing drugs that are able to achieve the same goal. The challenge is that cancer and rescue mutations are distributed widely in the protein, and experimental testing of all possible combinations of mutations is not feasible. This paper describes a simple computational metric that reflects the overall stability of the p53 core domain and can discriminate which second-site mutations restore activity to p53 cancer mutants.
| The tumor suppressor protein p53 is a transcription factor that plays a major role in preventing cancer initiation and progression. Cellular stress conditions such as hypoxia or DNA damage activate p53, which induces cell cycle arrest, DNA repair, senescence, or apoptosis [1], [2], [3]. In most, if not all, human cancers, the p53 apoptosis pathway is inactivated, and p53 itself is mutated in about half of all human cancers. About three-quarters of tumors with mutant p53 express full-length p53 with single missense mutations in the p53 DNA-binding core domain. These mutations may cause partial or global protein destabilization, loss of zinc coordination, or disruption of DNA contacts, and thus inactivate the tumor suppressor function of p53 (www-p53.iarc.fr) [4]. These missense mutations (“cancer mutations” or “oncogenic mutations”) are widely distributed throughout the core domain (Figure 1). They have been classified based on their physical location within the protein: (i) DNA-contact mutants (e.g., R248Q, R273H), (ii) structural mutants in the DNA binding surface (e.g., R175H, G245S, R249S, R282W), (iii) β-sandwich mutants (e.g., Y220C), and (iv) zinc-binding domain mutants (e.g., C242S, R175H).
Pharmacological rescue of p53 function in cancer tissues is an attractive therapeutic target [5]. Recently, two independent studies on transgenic mice demonstrated that restoration of p53 activity enables tumor regression in vivo [6], [7]. p53 reactivation is especially promising in regression of advanced stage cancers [8], [9]. The p53 function of some oncogenic mutants has been rescued in vivo by a handful of small molecules [10], [11], [12], [13], [14] as well as by second-site suppressor (“cancer rescue”) mutations [15], [16], [17], [18]. The second-site mutations provide easily-studied cases of p53 cancer rescue.
The effect of oncogenic and rescue mutations in p53 has been of great interest. Many detailed structural studies have been pursued, including X-ray crystal structures of individual oncogenic and rescue mutants of p53 [19], [20], [21]. The loss or gain of hydrogen-bonding interactions, salt bridges and other minute stabilizing or destabilizing effects upon different missense mutations have been investigated to develop a more complete understanding of the inactivation mechanisms by the oncogenic missense mutations and, correspondingly, the mechanisms by which restoration of activity for rescue mutations occur [22], [23]. At 310 K, wild-type p53 is estimated to be only 3.0 kcal/mol more stable than the denatured state [23], and thus missense mutations can easily shift the delicate balance of p53 stability.
The present study quantifies the effect of oncogenic and rescue mutations on the overall dynamics of p53 without focusing on local structural details. The core DNA-binding domain of p53 was used, as it dictates the stability of the overall protein [22]. The overall protein flexibility of the p53 DNA-binding domain for the wild-type, cancer mutants, rescue mutants and non-rescue mutants was compared in explicitly-solvated all-atom molecular dynamics (MD) trajectories, which are well suited to investigate the local conformational space sampled by each particular mutant. A single discriminating metric, the measure of flexibility of p53 in terms of the number of clusters obtained at a certain RMSD cutoff, was able to predict the functional activity of various mutant p53 proteins.
The coordinates for the starting structure were obtained from the wild-type p53 coordinates of chain B in pdbID 1TSR [4]. Each mutant system was prepared from this structure by rebuilding the mutated side chain(s) with the AMBER suite [24]. Crystallographic waters were retained. Histidine, asparagine and glutamine side chains which were mis-fit during structure characterization were determined and flipped by 180° using the Molprobity web server [25]. Histidine protonation states were determined using the Whatif Web Interface [26] and manually verified. Zinc coordination residues (Cys176, Cys238, Cys242 and His179) were modeled following the cationic dummy atom method of Pang et al [27]. Missing atoms and hydrogens were added using the Leap module of Amber10 [24]. Each system was solvated in a TIP3P [28] water box. The buffer between the protein and the periodic boundary was not closer than 8 Å in any direction. The wild-type p53 system has a charge of +1. Chloride ions were added as needed to neutralize the different mutant systems studied. The topology and coordinate files of the systems were constructed using Amber FF99SB force field [29]. The final wild-type p53 system consisted of 27,264 atoms.
Each system was first relaxed by 36,000 steps of minimization and a standard relaxation procedure using restrained MD. In the first 2,000 steps of minimization only the hydrogen atoms were relaxed, leaving all other atoms fixed. In the second 2,000 steps, all water atoms and ions were minimized in addition to the hydrogen atoms. In the third 2,000 steps, zinc-coordinating residues Cys176, Cys238, Cys242 and His179 as well as all hydrogens, water atoms, and ions were minimized. In the following 10,000 steps, all atoms were minimized except backbone atoms, which were held fixed. In the last 20,000 steps, the entire system was minimized. Following the minimizations, restrained MD simulation at 310 K was carried out for 1 nanosecond to prevent structural artifacts from introducing kinetic energy into the system. For this purpose, positional restraints for the heavy atoms of the protein backbone were gradually decreased from 4.0 to 1.0 kcal/(mol * Å2) in four consecutive 250-picosecond-long MD simulations.
Thereafter, unrestrained MD was performed in explicit solvent for 30 nanoseconds at 310 K using a time step of 1 femtosecond. Temperature was maintained constant at 310 K by Langevin dynamics with a collision frequency of 5 ps−1, and pressure was maintained at 1 atm by the Nose Hoover-Langevin piston method [30], [31] using period and decay times of 100 and 50 femtoseconds, respectively. Long-range electrostatics was treated by the Particle Mesh Ewald method [32] and a nonbonded cutoff of 10 Å was used. The interatomic distances within the water molecules were fixed using the SHAKE algorithm [33], [34]. A multiple-time step algorithm was employed, in which bonded interactions were computed at every time step, short-range non-bonded interactions were computed at every second time step, and full electrostatics was computed at every fourth time step.
All minimizations and MD simulations were performed using NAMD2.7 [35] on the Teragrid Ranger cluster. The simulations scaled as 0.10 days per nanosecond using 64 processors. Root-mean-square-deviation (RMSD) traces over the course of the MD trajectories are depicted in Figure S1.
We considered all four structural classes of p53 mutants: (i) DNA-contact mutants, (ii) structural mutants in the DNA binding surface, (iii) β-sandwich mutants, and (iv) zinc-binding domain mutants. We did not attempt to characterize any direct zinc-binding residue mutations (e.g., C242S), as rigorous parameterization of the partial charges on the metal ion and coordinating groups would be required for proper treatment of any mutations in this area.
Mutants simulated included the wild-type p53, the six most-frequent cancer mutants (R175H, G245S, R248Q, R249S, R273H, R282W), cancer mutant Y220C for which some stabilization (although not enough to restore p53 activity) was achieved recently with a small-molecule filling the location of the mutated tyrosine side chain [36], four rescue mutants and three non-rescue mutants for the R273H cancer mutant [17], [18], two rescue mutants and one non-rescue mutant for the G245S cancer mutant (G245S_N239Y, G245S_T123P and G245S_E286D) [18], [37], two rescue mutants and one non-rescue mutant for the Y220C cancer mutant (Y220C_A138G, Y220C_L137R and Y220C_L114G), the superstable quadruple mutant M133L_V203A_N239Y_N268D [38], and stabilizing mutant N239Y [38].
Conformational clustering was performed using the gromos algorithm [39] with GROMACS4.0.5 analysis software [40]. For each of the mutants, atomic coordinates were extracted at 10 ps intervals over the 30 ns MD simulation. The resulting 3000 structures were superimposed with respect to all Cα atoms to remove overall translation and rotation, then clustered at various RMSD cutoff values (i.e., 0.95, 1.05, and 1.15 Å) based on atomic coordinates of all Cα atoms of the protein. After calculating an RMSD-distance matrix of atomic positions between all pairs of MD snapshots in a trajectory, the gromos clustering algorithm counts the number of similar MD snapshots for which the calculated RMSD is less than or equal to the determined RMSD cutoff for each MD snapshot. The MD snapshot with the highest number of neighbors (e.g. the structure with the smallest RMSD between all the other structures in the cluster) is determined to be the center of the first cluster. Thus this structure is also referred to as the “cluster centroid.” Subsequently, this entire cluster (i.e. the cluster centroid and its neighbors) is eliminated from the pool of MD snapshots, and the same process is repeated until all MD snapshots are assigned to a cluster.
As another potential flexibility metric, root-mean-square-fluctuation of all Cα atoms of p53 in the trajectories were calculated using AMBER suite.
Two alternative clustering methods available in GROMACS package, namely single-linkage clustering and Jarvis-Patrick clustering, were also performed for comparison. A cutoff of 0.65 Å was used for single-linkage clustering. In Jarvis-Patrick clustering, the RMSD cutoff used to determine the number of nearest neighbors considered for Jarvis-Patrick algorithm was set to 0.80 Å, and the snaphots that have at least 3 identical nearest neighbors were assigned to the same cluster.
Functional activities of rescue and non-rescue mutants for which no published experimental p53 activity result exists (R273H_N239S, R273H_R282S, R273H_L114G, G245S_E286D, Y220C_A138G, Y220C_L137R and Y220C_L114G) were verified using yeast assays (Figure 2). Wild-type p53 and relevant cancer mutants R273H and Y220C were also included in the assays as controls. For this purpose, p53-tester yeast strain RBy379 (1cUASp53::URA3 his3Δ200 a/alpha) [17], [41] expressing the URA3 gene under control of a p53-dependent promoter was transformed with centromeric pTW300 plasmids [41] (HIS3 selection marker) expressing either wild-type human p53 or the mutants indicated under control of the ADH1 promoter. Yeast strains were grown in YEPD (10% yeast extract, 20% pepton, 20% dextrose) and transformed with the relevant plasmids using a LiAc-based transformation protocol [42]. Transformants were selected on SC plates lacking histidine and incubated at 30°C. Serial dilutions of mid-log phase cells (10,000; 2,000; and 400 cells) were spotted onto agar plates lacking either histidine or uracil. Plates were incubated for 2 days at 37°C. The growth on plates lacking histidine is selective only for the presence of the plasmid, while growth on plates lacking uracil is dependent on expression of the URA3 gene and is a measure of p53 activity [43].
In order to compare the dynamical effects of different mutations on p53, MD-generated trajectories of various p53 mutants were clustered based on overall structural similarity. Explicitly solvated MD simulations were run in the isothermal-isobaric (NPT) ensemble for 30 ns, after which RMSD-based clustering was performed on the resulting trajectories with the gromos clustering algorithm [39]. The RMSD distance matrix was computed in a pairwise fashion over all of the alpha carbons for each structure extracted every 10 ps from a particular trajectory (i.e., 3000 structures representing each trajectory). A large range of RMSD cutoff values were tested, including 0.95, 1.05, 1.15, 1.25, 1.35 and 1.60 Å. RMSD cutoffs larger than 1.15 Å caused loss of sensitivity of NOCs to the effect p53 mutations. The low optimal RMSD cutoff is an indication of a well-behaved system sampling configurations within a single energetically low-lying substate, as well as a reflection of the small size and low flexibility of the p53 core domain. The NOCs observed for p53 wild-type and its various mutants using a cutoff of 1.15 Å are shown in Tables 1 through 3. Clustering results at several other RMSD cutoff values are presented in Table S1.
The number of clusters was significantly higher for the cancer mutants (in bold in Table 1) compared to the wild-type p53, which suggests that oncogenic mutations increase the overall plasticity of the p53 core domain. This result is consistent with Rauf et al. [44], who investigated the effects of different oncogenic mutations on the flexibility of the p53 DNA-binding domain using a graph theoretical approach. Here, the oncogenic property for all four structural classes of p53 mutants has been quantified by a single metric.
The flexibility of the structural and zinc-binding domain mutant R175H, as characterized by the number of clusters, was remarkably higher compared to the wild-type and other cancer mutants (Table 1). The especially high degree of flexibility exhibited by this system may explain why so far it has not been possible to rescue the R175H mutant with second-site suppressor mutations, even though all possible single point core domain mutations of R175H were tested exhaustively for p53 function [18].
The number-of-clusters (NOC) metric presented here correctly locates the DNA-contact mutant R273H as the closest cancer mutant to the wild-type in terms of structural variability over the 30 ns trajectories. R273H has been previously demonstrated to be the easiest to rescue among the most-frequent cancer mutants [23]. The thermodynamic stability of the R273H mutant is the closest to the wild-type p53 among the 19 cancer mutants considered by Bullock et al [23]. The number of rescue mutants known to reactivate R273H mutant is large compared to few or no rescue mutants known to reactivate each of the other hot spot cancer mutants [17], [18], [37], [38].
Comparison of the number of clusters for the R273H rescue mutants (in italics in Table 2) with those for the R273H cancer mutant (in bold in Table 2) indicated a significant decrease in flexibility for the rescue mutants. The restoration of stability to the protein was especially remarkable in the case of the R273H_N263V and R273H_N200Q_D208T rescue mutants, for which the number of clusters was even lower than the wild-type p53. Although thermodynamic stability data is not available for these rescue mutations, our results suggest that such values would be lower than wild-type p53.
As there are no single point mutations that can strongly rescue cancer mutants R175H, R248Q, R249S or R282W, the generality of this finding was tested on rescue mutants for which experimental data is available (in italics in Table 2). More specifically, the known rescue mutants G245S_N239Y and G245S_T123P were considered for the class of structural mutants in the DNA binding surface. Similarly, two rescue mutants for the β-sandwich mutant Y220C, Y220C_A138G and Y220C_L137R, were investigated with the same approach. Functional activity of the latter two rescue mutants, for which no published experimental p53 activity results exist, was verified using yeast assays and depicted in Figure 2. Three out of these four rescue mutants showed decreased flexibility compared to their relevant cancer mutant (Table 2). The only exception in this test set was G245S_T123P, which exhibited more clusters than its cancer mutant G245S.
As negative controls, we tested experimentally confirmed non-rescue second-site mutations (underlined in Table 2) of relevant cancer mutants with the same approach. Functional inactivity of all the nonrescue mutants was verified using yeast assays and depicted in Figure 2. All non-rescue mutants that we simulated were more flexible, as compared to their relevant cancer mutant, indicating destabilization introduced to the cancer mutant by these second-site mutations. Thus, our method can successfully discriminate rescue mutants from non-rescue mutants.
We extended the same analysis on stabilizing mutant N239Y and the “superstable” quadruple mutant M133L_V203A_N239Y_N268D [38] (Table 3). The N239Y mutant exhibited a significant decrease in flexibility compared to the wild-type. In contrast, the superstable mutant did not follow the same trend (Table 3). This may be due to the need for longer relaxation in MD simulations in order to account for the greater extent of structural change introduced by four point mutations. To explore this hypothesis, we extended the quadruple mutant MD simulation for an additional 30 ns (for a total of 60 ns of production dynamics). In the second 30 ns, its number of clusters decreased significantly to a value much lower than that of the wild-type p53 as hypothesized.
The data presented in tables 1–3 relies on the full 30 ns trajectories of p53. In order to determine what is the shortest MD simulation necessary to discriminate between the functional and nonfunctional forms of p53 mutants, shorter segments of the full production MD trajectories were analyzed. At the RMSD cutoff of 1.15 Å, the number of clusters for each p53 mutant calculated at 5, 10, 20, 25 and 30 ns of MD simulations were separately depicted as column graphs in Figure S2. In this set of graphs, active and inactive p53 mutants were grouped and designated with a green arrow and a red arrow, respectively. In Figure 3, the percentage of mutants for which p53 function was correctly predicted by our flexibility metric are depicted for 5, 10, 20, 25 and 30 ns of MD simulations. The success of function prediction increased from 74% to 91% while our simulation time increased from 5 ns to 30 ns. This analysis indicated that at least 30 ns of MD simulation is required for a successful prediction of function of p53 mutants.
The thermodynamic stability values of several p53 cancer mutants are available in the literature (Table 1), as measured by urea-induced unfolding experiments at 283 K [23], [37], [38]. There are no comparable experimental data for the rescue or non-rescue mutants, which were not included in this part of the analysis. All cancer mutants evaluated experimentally exhibit differential experimental destabilization compared to wild-type p53 (Table 1).
Figure 4 depicts the correlation between the available thermodynamic stability values of p53 cancer mutants and the number of clusters observed in the MD simulations at the RMSD cutoff of 1.15 Å (r2 = 0.77). The r2 values for p53 single mutants at RMSD cutoff values of 0.95 Å and 1.05 Å are both 0.74. The number of clusters at these cutoffs for each mutant is tabulated in Table S1. The number of clusters in the second 30 ns MD simulation of the superstable mutant was used for this correlation analysis. If the NOCs in the initial 30 ns MD simulation of the superstable mutants was used instead, the r2 value decreased from 0.77 to 0.55. If the average of the two was used, the r2 value became 0.70. Excluding the superstable mutant form the data set gave an r2 value of 0.66. Remarkably, the number of clusters metric alone explains about three-quarters of the variance in experimentally measured thermodynamic stability values of p53 mutants.
To compare the NOC metric with another simple flexibility metric, the root-mean-square-fluctuation (RMSF) values of all Cα atoms of p53 were calculated using the AMBER suite for each mutant (Table S2). The correlation of the RMSF values with the thermodynamic data gave an r2 value of 0.62, which showed the superiority of the NOC method comparing to its r2 value of 0.77. To compare the performance of other clustering methods, single-linkage clustering and Jarvis-Patrick clustering were performed on the MD trajectories (Table S3). Both methods resulted in a lower correlation with the thermodynamic stability values versus the RMSD-based clustering method, with r2 values of 0.44 and 0.45, respectively
p53 is an inherently unstable protein, as reflected by its low melting temperature of ∼42–44°C [45]. It has been shown that the main reason of p53 instability is neither poor packing density nor the presence of unusually large void volumes in the protein. Instead, an analysis of the solution structure of p53 core domain obtained by NMR has revealed several reasons for instability of p53 [46]. First, this study indicated that p53 has buried hydroxyl and sulfhydryl groups that form sub-optimal hydrogen-bonding networks. Second, high flexibility of loop regions, especially of L1 loop, is observed in p53. Lastly, some buried tyrosine residues were found to be involved in temperature-dependent dynamic processes possibly indicating presence of alternative hydrogen-bonding networks in p53. Based on all of these factors, Canadillas et al concluded that “the p53 structure is more flexible than is apparent from crystal structures” [46].
In an effort to capture this intrinsic structural flexibility, we have focused on finding a computational method to measure the overall flexibility of the p53 core domain and the effect of mutations, be they cancer mutations, rescue mutations or non-rescue mutations, on the flexibility. This work presents a new method in which the number of structural clusters representing an explicitly solvated all-atom MD trajectory can be used as a single robust measure of overall flexibility in the p53 core domain. All hot-spot cancer mutants we studied demonstrated higher flexibility compared to the wild-type p53, in line with the results of an earlier graph-theoretical approach that assessed the flexibilities of wild-type p53 and several cancer mutants [44]. Testing rescue and non-rescue mutants for particular cancer mutants, the number of clusters for functional p53 mutants was found to differ significantly from the nonfunctional p53 mutants. Remarkably, the NOC metric is able to predict which second-site mutations may restore p53 activity to cancer mutants and which will leave p53 functionally defective.
It is also notable that such a simple metric reflecting system flexibility or entropy can account for three-quarters of the variance in experimentally measured thermodynamic stability values of p53 mutants. MD simulations thus promise to be a robust tool to predict thermodynamic stability of p53 mutants of interest. The NOC metric could further be used to discover new rescue mutants that restore p53 activity, and thus kill the cancer cell. Additionally, whether binding of a small-molecule can achieve enough stabilization to restore p53 function to cancer mutants could be tested with this metric. The computational cost of performing classical MD simulations could be decreased by using alternative methods such as accelerated MD, which may achieve increased sampling of conformational states over significantly shorter simulation timescales. Experimental efficiencies could be achieved through an integrated strategy that is guided by use of the NOC metric as a predictive measure for p53 function.
|
10.1371/journal.pgen.1001083 | Genetic Analysis of Baker's Yeast Msh4-Msh5 Reveals a Threshold Crossover Level for Meiotic Viability | During meiosis, the Msh4-Msh5 complex is thought to stabilize single-end invasion intermediates that form during early stages of recombination and subsequently bind to Holliday junctions to facilitate crossover formation. To analyze Msh4-Msh5 function, we mutagenized 57 residues in Saccharomyces cerevisiae Msh4 and Msh5 that are either conserved across all Msh4/5 family members or are specific to Msh4 and Msh5. The Msh5 subunit appeared more sensitive to mutagenesis. We identified msh4 and msh5 threshold (msh4/5-t) mutants that showed wild-type spore viability and crossover interference but displayed, compared to wild-type, up to a two-fold decrease in crossing over on large and medium sized chromosomes (XV, VII, VIII). Crossing over on a small chromosome, however, approached wild-type levels. The msh4/5-t mutants also displayed synaptonemal complex assembly defects. A triple mutant containing a msh4/5-t allele and mutations that decreased meiotic double-strand break levels (spo11-HA) and crossover interference (pch2Δ) showed synergistic defects in spore viability. Together these results indicate that the baker's yeast meiotic cell does not require the ∼90 crossovers maintained by crossover homeostasis to form viable spores. They also show that Pch2-mediated crossover interference is important to maintain meiotic viability when crossovers become limiting.
| In meiosis, sex cells that become eggs or sperm undergo a single round of DNA replication followed by two consecutive chromosomal divisions. In most organisms, the segregation of chromosomes at the first meiotic division is dependent upon at least one genetic exchange, or crossover event, between homologous chromosome pairs. Matched chromosomes that do not receive a crossover frequently undergo non-disjunction at the first meiotic division, yielding gametes that lack chromosomes or contain additional copies. Such missegregation events have been linked to Down syndrome and human infertility. This paper focuses on Msh4-Msh5, a complex required for the proper segregation of homologous chromosomes during the Meiosis I division. We performed a mutational analysis of the baker's yeast Msh4-Msh5 complex to study its role in implementing the decision to make a crossover. We identified a class of mutants that are functional in meiosis despite significant reductions in crossing over that occurred primarily on larger chromosomes. In combination with mutations (pch2Δ, spo11-HA) that disrupted early steps in crossover placement, this msh4/5 class of mutants displayed poor spore viability. Together, these data support the presence in yeast of a robust crossover distribution mechanism.
| Meiosis produces haploid gametes from diploid progenitor cells. This reduction in ploidy results from the segregation of homologous chromosomes at the first meiotic division (Meiosis I). In most organisms, the accurate segregation of chromosomes during Meiosis I requires crossing over between homologs. These crossovers provide physical linkages between homologs that enable their proper positioning at metaphase I through spindle microtubule generated forces [1]. Disruption of these forces by the loss of chromosome arm cohesion facilitates the Meiosis I division [2]. Failure to achieve at least one crossover per homolog pair results in non-disjunction of the homolog pair, leading to the production of aneuploid gametes (reviewed in [3]).
Meiotic crossing over is initiated in meiotic prophase by the formation of Spo11-dependent DNA double strand breaks (DSBs; [4]). Meiotic DSBs can be repaired as either crossovers or non-crossovers through distinct repair pathways [5], [6]. In Saccharomyces cerevisiae, approximately 60% of the 140–170 DSBs that form in meiosis (estimated from a whole genome microarray analysis of dmc1Δ and dmc1Δ rad51Δ mutants) are processed as crossovers [7], [8]. A single S. cerevisiae cell in meiosis forms approximately 90 crossovers distributed over sixteen homolog pairs [9]–[11]. In contrast, in C. elegans meiosis, only a single crossover forms between each homolog pair that ensures Meiosis I disjunction [12].
The majority of meiotic crossovers in baker's yeast display interference. Interference ensures that a crossover designation for one DSB site makes a non-crossover fate more likely at adjacent sites, and leads to the formation of widely and evenly spaced crossovers [13]–[15]. In the interference-dependent crossover pathway, DSBs are processed to form single end invasion intermediates (SEIs) that result from the invasion of a DSB end into an intact homolog. These intermediates are then thought to undergo second-end capture with the intact homolog to form double Holliday junctions (dHJs) that are ultimately resolved to form crossovers [16]–[19]. A crossover homeostasis mechanism was identified in baker's yeast that ensures crossovers are preferentially formed at the expense of non-crossovers when the number of initiating DSBs is reduced [20]. Thus crossover interference and homeostasis ensure formation of at least one crossover on all homolog pairs [20], [21]. The presence of at least one crossover per homolog pair is known as the obligate crossover. Barchi et al. [22] further define the obligate crossover “as one of the outcomes of the process(es) through which most crossovers form, not as a special type of crossover.” Control mechanisms that ensure the obligate crossover are likely to act during the crossover/non-crossover decision, an event that takes place at or just prior to SEI formation [5], [6]. It is important to note that previous work in baker's yeast suggested that ∼20% of crossovers on a large chromosome and ∼50% of crossovers on a small chromosome involved interference-independent crossovers that occurred through a distinct Mms4-Mus81 pathway [23], [24].
The ZMM proteins (Zip1-4, Spo16, Mer3, Msh4-Msh5) act as pro-crossover factors in the interference-dependent crossover pathway by coordinating crossing over with formation of the synaptonemal complex, a zipper-like structure that connects homologous chromosomes in late stages of meiotic prophase [24]–[34]. Msh4-Msh5 attracted our attention because strains defective in this complex show strong defects in Zip1 polymerization during synaptonemal complex formation [27], [29]. Msh4 and Msh5 each contain domains II–V found in the bacterial MutS family of mismatch repair proteins, but lack the N- terminal domain I that is required to interact with domain IV for mismatch DNA binding (Figure 1A; [25], [26], [35], [36]). S. cerevisiae msh4Δ and msh5Δ mutants display reduced crossing over (∼2.5 fold decreased) and spore viability (30–40%). Tetrads obtained from these mutants display an excess of zero and two viable spores compared to wild-type. This phenotype is consistent with a Meiosis I disjunction defect [24]–[27]. The equivalent mutations in male and female mice result in sterility as a consequence of chromosome pairing and synapsis defects [37]–[39]. The residual crossovers seen in yeast msh4/5Δ mutants lack genetic interference [24], [27]; however in msh4Δ mutants, Zip2 foci, which mark crossover designation sites, still display a pattern indicating that they are subject to interference [40]. These and other data suggest that Msh4-Msh5 acts after the crossover/noncrossover decision [16], [40]. Consistent with the above data, biochemical and molecular studies showed that Msh4-Msh5 is required to stabilize SEIs and is capable of specifically binding to Holliday junctions as multiple sliding clamps [16], [41].
Additional cell biological observations, primarily in the mouse, have led to a model in which Msh4-Msh5 interacts with the MutL mismatch repair homologs Mlh1-Mlh3 to resolve Holliday junctions [25], [41]–[47]. In mouse spermatocytes in zygotene, Msh4/5 foci are present at high levels (∼140 per nucleus) but decrease until mid pachytene, where they are present at roughly twice the number of crossover sites. At this stage, roughly half of Msh4/5 foci interact with Mlh1/3 foci, which localize to sites of crossing over [48]–[50]. The presence of a large number of Msh4/5 foci in zygotene suggest the possibility of early roles for Msh4/5 in meiosis; consistent with this idea is work in Sordaria which show an early role for Msh4-Msh5 during interhomolog interactions, at a time prior to when it is required for recombination progression [51].
The above information encouraged us to systematically mutagenize Msh4-Msh5 to study its role in implementing the crossover decision. We identified a class of msh4/5 threshold (msh4/5-t) mutants that displayed high spore viability despite 1.5 to 2 fold reductions in crossing over that occurred primarily on large (XV, VII) and medium (VIII) sized chromosomes. msh4/5-t mutants displayed Msh5 foci similar to wild-type; however, they showed defects in Zip1 polymerization during synaptonemal complex formation. This phenotype is consistent with defects in a crossover maturation process that occurs after Msh4-Msh5 loading onto chromosomes. A triple mutant containing a msh4/5-t allele and mutations that decreased DSB levels (spo11-HA) and crossover interference (pch2Δ) showed preferential loss of crossovers on the small chromosome III and a synthetic spore viability defect, suggesting that crossover interference is critical to maintain meiotic viability when crossovers become limiting.
Msh4 and Msh5 amino acid sequences from S. cerevisiae, H. sapiens, M. musculus, A. thaliana, and C. elegans were aligned using clustalW and CLC free Workbench software (Figure 1, Figure S1; data not shown). We selected four different classes of conserved residues to alter by site-specific mutagenesis (Figure 1B). Class 1 (Msh4/5-specific) residues were conserved in Msh4 and Msh5 but were not conserved in other Msh family members such as Msh2, Msh3, and Msh6. Class 2 (Msh4-specific) and Class 3 (Msh5-specific) were conserved only in Msh4 and Msh5, respectively (Figure 1B; Table 1). Previous work by Pochart et al. [52] showed that mutations in the ATP binding domain of Msh5 conferred a null phenotype. Based on these observations, we also mutagenized ATP and DNA binding residues conserved among all Msh family members (Class 4). Eight of these Class 4 mutations were in homologous positions in Msh4 and Msh5 (Figure S1). In total 57 residues were mutated, 29 from Msh4 and 28 from Msh5 (Table 1). All residues were mutated to alanine, with the exception of one residue in the Msh4/5 ATP binding domain that was mutated to tryptophan to allow comparison with an amino acid substitution in a homologous position in Msh2 that affected Msh2-Msh6 ATP hydrolysis [53]. All alleles were integrated into the congenic SK1 strain EAY1108 (EAY background, [24]).
msh4 and msh5 alleles were analyzed as heterozygotes over their respective deletion mutations in the SK1 congenic strain EAY1112 [24]. The mutant diploid strains were sporulated and assessed for spore viability and genetic map distances on chromosome XV (Table 1; Figure 1C). The mutations are presented relative to Thermus aquaticus MutS domains II, III (linker), IV (DNA binding) and V (ATPase) [35]. The spore viability profiles of msh4 and msh5 mutants indicated that the Msh5 subunit was more sensitive to mutagenesis (Figure 2A). A larger proportion of msh5 mutants showed ≤50% spore viability compared to msh4 (9 of 28 for msh5 versus 2 of 29 of msh4; p = 0.02, Fisher's exact test). This difference was also seen in an analysis of mutations in domain IV (DNA binding); 5 of 12 msh5 mutations conferred ≤50% spore viability compared to 0 of 11 msh4 mutations (p = 0.03, Fisher's exact test).
Five of the eight mutations in homologous positions in Msh4 and Msh5 conferred subunit-specific phenotypes. Both msh4-G639A and msh5-G648A strains contain mutations (Walker motif A) predicted to disrupt ATP binding; both of these strains displayed null phenotypes [35]–[36], [52]–[55]. In contrast, a predicted ATP hydrolysis mutation in Msh4, msh4-R676W, conferred wild-type spore viability but the corresponding mutation in Msh5, msh5-R685W, conferred a null phenotype (Figure 2B; Table 1). Similar asymmetries between Msh4 and Msh5 were observed at four residues in the DNA binding domain IV (Figure 2B; Table 1). msh4-N532A, msh4-Y485A, msh4-L493A, and msh4-L553A had spore viabilities of 89, 95, 75, and 95%, respectively; corresponding mutants msh5-D527A, msh5-Y480A, msh5V-488A, and msh5-L548A had significantly lower spore viabilities (30, 67, 40, and 50%, respectively).
Most msh4 and msh5 mutants with significant spore viability and/or crossover defects could not form stable Msh4-Msh5 complexes as assessed in the two-hybrid assay (Table 1). The only exceptions were msh4-E276A (domain II), msh4-R676W (ATP hydrolysis), msh5-D539A (domain IV), msh5-G648A (ATP binding), and msh5-R685W (ATP hydrolysis) mutants that displayed poor spore viability or crossover defects but formed stable complexes with a wild-type partner. Inability to form a stable complex in the two-hybrid assay can be explained by the disruption of an interaction domain or a loss in protein stability. Because most mutations were created in highly conserved residues that lie outside of putative interaction domains in Msh proteins [35], [36], [54], a defect in the two-hybrid assay is likely to reflect a disruption of protein structure.
Spore viability was plotted as a function of genetic map distance for all msh4 and msh5 mutants (Figure 3). This plot shows that crossing over could be reduced by up to two-fold on the large chromosome XV without affecting spore viability. msh4/5 mutations (msh4/5-t) near the threshold limit for crossovers included msh4-E276A, msh4-F491A, msh4-N532A, msh4-R676W, msh5-D76A, msh5-D250A, msh5-S416A, msh5-Y486A, and msh5-D539A (Table 1). The phenotypes conferred by these mutations were independent of their ability to disrupt the Msh4-Msh5 complex as measured in the two-hybrid assay (Table 1). A second class of msh4/5 mutants showed greater than two-fold decreases in crossing over on chromosome XV. This below-threshold class (msh4/5-bt; msh4-Y143A, msh4-F194A, msh4-R456A, msh4-L493A, msh5-R436A, msh5-Y480A, msh5-D532A, msh5-L548A, msh5-D680A) showed spore viabilities between 50 and 76%. These mutants were all defective in their ability to form stable Msh4-Msh5 complexes in the two-hybrid assay (Table 1).
The wild-type spore viability profile for the msh4/5-t mutants suggested they were able to properly segregate all sixteen homolog pairs in Meiosis I (Table 1; Figure 3, Figure 4). We further examined the phenotype of a subset of msh4/5-t mutants (msh4-E276A, msh4-R676W, msh5-S416A, msh5-D539A; all but msh5-S416A showed wild-type two-hybrid interactions) in the SK1 isogenic NHY strain background. msh4 and msh5 alleles were analyzed as heterozygotes over their respective deletion mutations. The NHY diploid strains allowed us to measure genetic map distances in large (VII), medium (VIII), and small (III) chromosomes (Figure 5A; [23]). Smaller chromosomes have higher map distances per physical distance and weaker interference relative to larger chromosomes ([40], [56], [57] but see [58]). Thus we used this strain set to determine if msh4/5-t mutations altered crossover patterns on representative small, medium, and large chromosomes.
All four msh4/5-t mutants displayed wild-type spore viability but decreased crossing over (∼1.5-fold for the sum of map distances in three chromosomes; Figure 4, Figure 5B; Table 2). The spore viabilities of wild-type and one msh4/5-t mutant, msh4-R676W, were unaffected by raising the sporulation temperature to 33°C, a condition shown previously in the SK1 background to cause coordinated defects in the formation of recombination intermediates and crossover products in msh5Δ (data not shown; [16]). This observation provides another indication that msh4/5-t alleles confer sufficient Msh4-Msh5 function in meiosis. The sum of genetic map distances calculated from tetrads (similar values were obtained from total spores) in wild-type was 147 cM; map distances for msh4-E276A, msh4-R676W, msh5-S416A and msh5-D539A were 101, 109, 99, and 100 cM, respectively.
As shown in Figure 6, msh4/5-t mutants displayed a chromosome size-dependent loss of crossovers. For three intervals on the smallest chromosome III, the four msh4/5-t mutants showed 73 to 92% of wild-type crossover levels (determined from tetrad data). In contrast these mutants showed 63 to 76% of wild-type levels for the two intervals on a medium sized chromosome VIII, and 61 to 66% of wild-type levels for the three intervals on a large chromosome (Chromosome VII). The loss of crossovers on the large chromosome VII approached that seen in msh4/5Δ strains. For the msh4Δ and msh5Δ mutants, the sum of genetic map distances calculated from tetrads was 68 and 56 cM, respectively (2.2 to 2.6-fold drop in crossovers over three chromosomes, Figure 5; Table 2). The values from total spores were 87 and 75 cM for msh4Δ and msh5Δ, respectively. The differences in map distance calculated by spore and tetrad data were likely due to the high rate of gene conversion seen in msh4Δ and msh5Δ mutants (see below). Based on tetrad data msh4Δ crossovers levels were 36, 42 and 54% of wild-type on chromosomes III, VIII, and VII, respectively. For msh5Δ crossover levels were 26, 34 and 47% of wild-type on chromosomes III, VIII, and VII, respectively (Figure 6).
Previously Stahl et al. [15] and Abdullah et al. [59] reported a greater loss of crossovers on larger chromosomes (VII) compared to smaller ones (III) in msh4Δ/msh5Δ mutants. These groups analyzed crossing over in wild-type, msh4Δ and msh5Δ strains in two intervals (HIS4-LEU2 and LEU2-MAT) on chromosome III (small) and two (TRP5-CYH2 and CYH2-MET13) on chromosome VII (large) in the congenic RHB strain background. They found that the crossover defect in msh4Δ and msh5Δ mutants was stronger on chromosome VII (23% and 27% of wild-type, respectively) compared to chromosome III (39% and 34% of wild type, respectively). We performed our analysis in the NHY SK1 isogenic strain. We do not have a good explanation for why our data differ from the Stahl et al. [15] and Abdullah et al. [59] studies. One possibility is that genetic mapping information from a limited number of intervals may yield a pattern due to localized recombination effects that is not seen when a larger number of intervals is examined.
We then looked at crossover distribution in a msh4/5-bt mutant (msh5-D532A). This msh4/5-bt mutation conferred similar spore viability levels in the NHY and EAY strain background (65% in EAY vs 69% in NHY; Figure 4). Interestingly, the sum of genetic map distances for chromosomes III, VII, and VIII in msh5-D532A (69 cM) was similar to msh5Δ (56 cM) and msh4Δ (68 cM) (Figure 5); however, msh5-D532A showed a preferential retention of crossovers on the small chromosome III. Crossovers in this mutant were 56, 39, and 48 percent of wild-type for chromosomes III, VIII and VII, respectively (determined from tetrads; Table 2; Figure 6).
Gene conversion events were analyzed at eleven marker sites in a subset of msh4/5 mutants, (msh4-E276A, msh4-R676W, msh5-S416A, msh5-D532A, msh5-D539A). The frequency of gene conversion in these strains was similar to wild-type (Table 3). As seen previously, msh4/5Δ mutants displayed an elevated frequency of gene conversions compared to wild-type [21], [25], [60].
Lastly, crossover interference was analyzed in a representative msh4/5-t mutant (msh4-R676W) by measuring the coefficient of coincidence (COC, ratio of observed double crossovers to those expected by chance; Table 4; [61]) and the NPD ratio (Table 5; [62]–[63]). Lack of interference yields COC and NPD values of 1 while strong interference yields values significantly less than 1. On the whole crossover interference appeared similar in wild-type and msh4-R676W. In COC analysis the msh4-R676W mutant showed a lack of interference for two intervals on chromosome III; wild-type showed a lack of interference for only one of these intervals (Table 4). For chromosomes VII and VIII, msh4-R676W and wild-type both showed crossover interference at two intervals and the absence of interference at another. NPD ratios, calculated for intervals where at least eight NPD events were expected, were determined using Stahl's “better way” calculator. This method performs a chi square test to determine if there is a significant difference between the observed PD, TT and NPD tetrad classes and those expected by random crossing over. This analysis showed the presence of interference in both wild-type and msh4-R676W in three intervals on chromosomes VII and VIII (Table 5).
pch2Δ mutants display elevated crossing over on medium and large chromosomes, and are defective in crossover interference, yet display wild-type spore viability [21], [64]–[66]. In addition, initial genetic analyses showed that pch2Δ mutants displayed an increased ratio of crossovers to non-crossovers [21]. These observations, combined with cytological analyses indicating that Pch2 promotes domainal axis organization in meiosis [66], [67], led Zanders and Alani [21] to propose that Pch2 acts in early steps in crossover control to promote crossover interference at the crossover versus non-crossover decision. To test if msh4/5-t mutants showed increased sensitivity to early defects in crossover control, we made double and triple mutant combinations involving the msh4/5-t, spo11-HA, and pch2Δ mutations in the NHY strain background. The spo11-HA mutation was examined because strains bearing this allele display a 20% reduction in meiosis specific DSBs but show wild-type levels of crossing over and spore viability due to crossover homeostasis [20]. pch2Δ spo11-HA strains, however, display a significant loss in spore viability (73%). One explanation for this phenotype is that when DSBs become limiting, the proper distribution of crossovers becomes even more critical to ensure that every chromosome receives at least one crossover [21],[66].
As shown in Figure 4, Figure 5B, and Table 2, msh4-R676W spo11-HA and msh4-E276A spo11-HA double mutants displayed wild-type spore viability (89 and 91%, respectively) and cumulative map distances (113 and 106 cM, respectively, from tetrads). These values were similar to those seen in msh4-R676W (109 cM) and msh4-E276A (101 cM) single mutants. However, compared to msh4-R676W and msh4-E276A single mutants, msh4-R676W spo11-HA and msh4-E276A spo11-HA double mutants showed a decrease (∼30%) in crossing over in the small chromosome III that was accompanied by modest increases in crossing over in the medium and large chromosomes (Figure 6; Table 2). We do not have a good explanation for this phenotype; one possibility is that the spo11 hypomorphs confer mutant phenotypes in addition to lowering DSBs (see Discussion; [21]).
msh4-R676W pch2Δ and msh4-E276A pch2Δ double mutants also showed wild-type spore viability (93% for both, Figure 4); however the pch2Δ mutation conferred an increase in crossing over in msh4-R676W and msh4-E276A strains that appeared specific to the medium- (VIII) and large-sized (VII) chromosomes (Figure 5B, Figure 6). The cumulative map distances from tetrads in msh4-R676W pch2Δ (194 cM) and msh4-E276A pch2Δ (190 cM), were higher than wild-type (147 cM) but lower than pch2Δ (226 cM; Figure 5B). pch2Δ msh5Δ mutants were previously shown to have higher crossover frequencies than the msh5Δ mutant [21].
The wild-type spore viability profile seen in msh4/5-t spo11-HA suggested that crossover interference and homeostasis can distribute a smaller pool of crossovers to all 16 homolog pairs. In contrast, the wild-type spore viability profile seen in msh4/5-t pch2Δ can be explained by an increased number of crossovers compensating for interference defects [21]. Such explanations predict that compromising crossover interference (pch2Δ) and limiting DSB's (spo11-HA) would decrease spore viability because a random distribution of crossovers will favor large chromosomes (Figure 6; [21]). These effects are likely to be more pronounced in a msh4/5-t pch2Δ spo11-HA mutant that is predicted to be compromised for DSB formation, crossover interference, and crossing over. To test this we created the msh4-R676W pch2Δ spo11-HA triple mutant and analyzed its phenotype with respect to spore viability, crossover distribution, and chromosome III non-disjunction.
As shown in Figure 4, the msh4-R676W pch2Δ spo11-HA triple mutant displayed 55% spore viability, which was lower than spo11-HA pch2Δ (72% spore viability). The cumulative crossover level from tetrads for chromosomes III, VII and VIII in this mutant was 135 cM, which was lower than wild-type (147 cM) and pch2Δ spo11-HA (165 cM), but significantly higher than msh4-R676W (109 cM), which displayed high spore viability (Table 2; Figure 4, Figure 5B). msh4-R676W pch2Δ spo11-HA also showed a greater reduction in crossing over on chromosome III compared to pch2Δ spo11-HA mutants (Figure 6). Although crossover levels on chromosome III in msh4-R676W pch2Δ spo11-HA were similar to msh4-R676W spo11-HA, the medium (VIII) and large chromosomes (VII) in msh4-R676W pch2Δ spo11-HA showed specific increases in crossing over compared to msh4-R676W spo11-HA as predicted by the model (Figure 6, Figure S2). Consistent with this, the triple mutant displayed a spore viability profile indicating a Meiosis I disjunction defect (Figure 4). The triple mutant showed a higher frequency of non-mater two-spore viable tetrads in the triple mutant (12.7%, n = 71 two spore viable tetrads; 1.9% of total tetrads) compared to both pch2Δ spo11-HA (6.9%, n = 130; 0.96% of total tetrads) and msh4-R676W (6.8%, n = 44; 0.37% of total tetrads). Such tetrads are indicative of nondisjunction of chromosome III because the two viable spores carry both yeast mating types (MATa and MATalpha). In addition, 82% of the two spore viable tetrads in the triple mutant were sister spores compared to 68% in pch2Δ spo11-HA and 50% in msh4-R676W. These data are suggestive of non-disjunction of other chromosomes. Together this information is consistent with the triple mutant being unable to distribute at least one crossover between all homolog pairs (see Discussion).
msh4Δ and msh5Δ mutants show strong defects in Zip1 polymerization during synaptonemal complex formation [27], [29]. Our data below indicate that fully functional Msh4-Msh5 is required for complete Zip1 polymerization along homologs. Immunostaining of Msh5 and Zip1 was performed on a subset of the msh4/5-t (msh4-E276A, msh4-R676W, msh5-S416A, msh5-D539A) and msh4/5-bt (msh5-D532A) mutants in the NHY strain background four hours after induction into meiosis (Figure 7). The number and distribution of Msh5 foci on meiotic chromosomes for wild-type, msh4/5-t, and msh5-D532A mutants were similar. The average number of Msh5 foci per nucleus (n = 30) was 122 for wild-type, 120 for msh5-D532A, and 130 for msh5-D539A. However, all mutants showed a partial defect in Zip1 elongation and accumulated Zip1-specific polycomplexes. This phenotype is reminiscent of that displayed by spo16 and zip4 null mutants with the exception that spo16 and zip4 null mutants display poor spore viability [29]. One explanation for these observations is that the msh4/5 mutants present fewer crossover sites to initiate Zip1 polymerization; thus these mutants, while capable of loading Msh4-Msh5 onto meiotic chromosomes, appeared defective in steps required to implement crossing over at designated sites. Thus complete Zip1 polymerization may require feedback from Msh4-Msh5 that is delayed or does not occur in the msh4/5 mutants.
We also measured by DAPI staining the percent of cells that completed at least Meiosis I (MI/MII) for all of the strains examined by immunofluorescence. As shown in Figure S3, wild-type and one msh4/5-t threshold mutant, msh4-E276A, displayed similar timing and efficiencies of meiotic divisions. The msh4Δ, msh5Δ, three msh4/5-t mutants (msh4-R676W, msh5-S416A, msh5-D539A), and one msh4/5-bt mutant (msh5-D532A) all showed about a 1.5 to 2 hr delay relative to wild-type.
We identified msh4 and msh5 mutants (msh4/5-t) that displayed reduced crossing over in meiosis but maintained crossover interference and wild-type spore viability. The reduction in crossing over seen in msh4/5-t mutants appeared more pronounced on large and medium-sized chromosomes that typically receive a greater proportion of Msh4/5-dependent crossovers. msh4/5-t mutants also displayed chromosome synapsis defects. These observations and the poor spore viability phenotype of the msh4-R676W pch2Δ spo11-HA triple mutant support the idea that baker's yeast form an excessive number of meiotic crossovers and that Pch2-mediated crossover interference is critical for meiotic viability when crossovers become limiting. The msh4/5-t alleles, which can be used to titrate crossover levels without reducing spore viability, provide a new tool for investigators interested in identifying factors that regulate crossover control.
S. cerevisiae maintains a high level of crossing over, an average of 5.6 per homolog pair [8]–[11], [20]. In most organisms that display crossover interference (C. elegans, A. thaliana, Zea mays, D. melanogaster, Mus musculus and Homo sapiens), the ratio of crossovers in meiosis to homolog pairs is less than or equal to three (reviewed in [68]). Why does S. cerevisiae enjoy such a high level of crossing over when a single crossover per homolog pair appears sufficient to promote Meiosis I disjunction [12], [15]? One possibility is that high crossover levels improve fitness by reducing mutational load through the segregation of deleterious alleles [69]. Consistent with this idea are simulation studies suggesting that meiotic crossover rates in S. cerevisiae are optimized for mutational robustness [69]. Another possibility is that excess crossovers are needed to ensure crossover formation on small chromosomes [8], [11], [21]. Consistent with the latter explanation is work in yeast showing that a small chromosome (I, 230 KB) has a higher than average recombination rate. Chromosome I also showed a frequency of non-disjunction (0.2–0.4%) that was lower than expected (5%) if it had recombined at the average rate [56], [57], [70]. The enhanced recombination rates on smaller chromosomes in S. cerevisiae are likely to result from DSBs that occur at a higher than average density and weak crossover interference [40], [57], [58], [71], [72].
msh4/5 mutants displayed high spore viability and a higher retention of crossovers on a small chromosome (III) compared to larger chromosomes (VIII, VII and XV). We entertain two models to explain this phenotype. Both of these are based on work showing that Msh4-Msh5 is required to stabilize SEI recombination intermediates and can bind to Holliday junctions [16], [41]. In one model, msh4/5-t mutants are defective in converting all SEI and Holliday junction intermediates into crossovers with equal probability. Such a model predicts that crossover interference would not be affected in msh4/5-t mutants, and that msh4/5-t mutants would show defects in synaptonemal complex formation. Both of these phenotypes were seen in this study. This model predicts that msh4/5-t mutants would show high spore viability despite a decrease in crossing over because smaller chromosomes have a higher frequency of crossovers and the number of crossovers in yeast is much greater than the number of chromosomes. A drawback of this model is that it cannot fully explain why msh4/5 null mutants displayed more severe crossover defects on the smaller chromosome III. Such a pattern is unexpected if crossovers on small chromosomes are present at higher density and occur primarily through a non-interfering pathway [23]. It also cannot explain how msh4/5-t pch2Δ mutants make excess crossovers. We cannot rule out the possibility that the small number of intervals examined on chromosome III is not representative of the overall pattern. In the future we would like to test this model further by examining additional intervals on this chromosome as well as on another small chromosome such as chromosome I. In addition, we would like to examine the effect of the msh4/5-t mutations on early recombination intermediates such as SEIs.
We considered a second model that proposes a prioritization mechanism for the distribution of crossovers amongst chromosomes. This model is somewhat similar to that proposed by Kaback and colleagues [56], [57]. We suggest that msh4/5-t phenotypes reflect a temporal order of crossover designation that favors a crossover on every homolog pair before additional interference-dependent crossovers are made. Such a pattern can be presented within the context of a stress relief model for crossover initiation and distribution. In this model “crossover designation with accompanying interference can be explained by imposition, relief, and redistribution of compression stress and stress relief along chromosome axes” [13]. Crossover initiation on every homolog pair would lead to the release of mechanical stress along the homolog axis of every chromosome. For shorter chromosomes, interference created from stress relief at the crossover initiation site would extend to the end of the chromosome, leading to fewer interfering crossovers as was seen experimentally [57]. For large chromosomes, interference created by stress relief that accompanies obligate crossover designation would prevent additional crossovers until mechanical stresses are re-distributed. We suggest that this redistribution of stress delays additional crossover designations on larger chromosomes. In this model the msh4/5-t phenotype can be explained if mutant Msh4-Msh5 complexes can participate in initial stress relief to form an obligate crossover but are defective, perhaps due to stability issues, in subsequent crossover initiations that are subject to interference. This model could explain the synapsis defects seen in msh4/5-t mutants if the defect is specific to long chromosomes; a single synapsis initiation site on a small chromosome could be sufficient to allow polymerization along the entire chromosome. This model, however, does not account for why Msh5 focus formation appears wild-type in msh4/5 mutants. One possibility is that subsequent crossover initiations require functions that occur after Msh4-Msh5 loading onto chromosomes.
The temporal order model outlined above predicts that spore viability would be maintained in msh4/5-t mutants due to formation of the obligate crossover and that interference would appear stronger on larger chromosomes. Such an idea is consistent with previous studies in yeast showing that multiple interfering crossovers occur more frequently on large chromosomes and with models that explain the distributions of interfering crossovers seen on different sized chromosomes (e.g. [13], [15], [40], [57], [73]). While we have shown that msh4/5-t mutants maintain high spore viability and display crossover interference on large chromosomes (Figure 4; Table 4, Table 5), our data are not robust enough to test whether interference becomes stronger on these chromosomes. A caveat in this model is that msh4/5-t mutants display crossover levels on large chromosomes that are higher than wild-type in the pch2Δ mutant background. Thus msh4/5-t mutants do not appear limited in their ability to form crossovers. One way to explain this observation is that Pch2 acts as a general factor to repress recombination that increases the temporal window over which a mutant Msh4-Msh5 complex must execute crossover decisions. Alleviation of this repression results in increased crossing over in msh4/5-t pch2Δ mutants.
Crossovers in msh4-R676W pch2Δ spo11-HA triple mutants appear to be randomly distributed, thus leading to more crossing over on larger chromosomes compared to the msh4-R676W single mutant, and increased non-disjunction on a small chromosome. Previous studies have suggested that Pch2 is essential for proper meiotic axis organization following crossover designation and that crossover distribution is mediated by changes in meiotic axis organization/assembly (e.g. [13], [67], [74]). We suggest that the triple mutant phenotype can be explained in the second model if the pch2Δ mutation disrupts stress/stress relief mechanisms so that crossover designations occur without interference and no crossovers show a temporal delay. In this scenario Pch2 maintains meiotic viability when crossovers are limiting (i.e. msh4/5-t, spo11 hypomorph mutations) because it imposes a delay on additional interfering crossovers. This delay ensures that every homolog pair has received at least one crossover. One way to test this idea in yeast is to perform a genome wide analysis of crossing over in the msh4/5-t mutant versus the triple mutant [8], [11].
The Msh family of mismatch repair proteins display asymmetric roles with respect to DNA binding and ATP hydrolysis. In MutS, residues in domain I of subunit A specifically stack with the mismatch while domain IV of subunit B makes non-specific contacts with the DNA backbone [35], [36]. Similarly in MutSα, domain I in Msh6 specifically interacts with the mismatch while domain IV in Msh2 makes non-specific contacts with DNA [54], [75], [76]. Msh subunits also display different affinities for ATP and ADP [77]–[79]. For example in the Msh2-Msh6 mismatch repair complex, Msh6 and Msh2 contain high affinity binding sites for ATP and ADP, respectively [80]. Such asymmetries in ATP binding by Msh subunits are thought to be important to induce coordinated conformational changes in Msh-mismatch DNA complexes that signal downstream repair factors [80]–[84].
Three observations support the presence of asymmetries in Msh4-Msh5 analogous to those seen for the Msh mismatch recognition factors. 1. Snowden et al. [85] reported that the Msh4 subunit of human Msh4-Msh5 appears to have reduced ATP binding activity. 2. We identified different spore viability phenotypes for matched sets of msh4 and msh5 mutations that map to the ATP and DNA binding domains (Figure 2B). 3. We also found that on the whole, msh5 mutations conferred more severe meiotic phenotypes than the equivalent msh4 mutations, though this could indicate different structural organizations for the two proteins rather than asymmetric functions. Msh4-Msh5 binds to both single end invasion and symmetric double Holliday junction substrates [41], [85]. Based on studies performed with Msh and Mlh mismatch repair factors, it is easy to imagine that asymmetric Msh4-Msh5 interactions with its DNA substrate will involve analogous signaling steps that activate downstream factors such as Mlh1-Mlh3. Biochemical analysis of some of the mutant complexes presented in this study can provide evidence to support or refute these ideas.
S. cerevisiae SK1 yeast strains were grown on either yeast extract-peptone-dextrose (YPD) or synthetic complete media at 30°C [86]. When required, geneticin (Invitrogen, San Diego) and nourseothricin (Werner BioAgents, Germany) were added to media at prescribed concentrations [87], [88]. Sporulation medium was prepared as described in Argueso et al. [24]. msh4, msh5 mutants were analyzed in either the congenic EAY1108/EAY112 background (“EAY”) described in Argueso et al. [24] or the isogenic NHY942/NHY943 background (“NHY”) described in de los Santos et al. [23]. 28 msh5 and 29 msh4 point mutants were introduced in the EAY1108 background by transformation of EAY1281 and EAY2409 with integration plasmids bearing these mutations using standard techniques [89]. A smaller subset of these msh4, msh5 point mutants were made in the NHY background by transformation of EAY2844 and EAY2848 respectively. Double and triple mutants bearing different combinations of msh4, msh5, pch2Δ and spo11-HA were made in the NHY background by crossing single or double mutant strains followed by tetrad dissection. All strains used in this study are listed in Table S1.
Msh4 amino acid sequence from S. cerevisiae (YFL003C), A. thaliana (NM_117842), C. elegans (AF178755), M. musculus (BC145838), H. sapiens (NM_002440) and Msh5 amino acid sequences from S. cerevisiae (YDL154W), A. thaliana (EF471448), C. elegans (NM_070130), M. musculus (NM_013600), H. sapiens (BC002498) were aligned using ClustalW software (www.ebi.ac.uk/clustalw) and CLC free workbench. A Msh4, Msh5 consensus sequence was generated using CLC and aligned against S. cerevisiae Msh2 (YOL090W), Msh3 (YCR092C), Msh6 (YDR097C) to check if residues conserved across Msh4, Msh5 in all five species are conserved in the other Msh family members.
The SK1 MSH4 open reading frame with 600 bp upstream sequence and 400 bp downstream sequence was amplified with pfu DNA polymerase and cloned into pRS416 with a 1.5 kb KanMX fragment inserted 90 bp downstream of the MSH4 stop codon to create the single step integrating plasmid pEAA427. The SK1 MSH5 open reading frame with 500 bp upstream sequence and 400 bp downstream sequence was similarly amplified with pfu DNA polymerase and cloned into pRS416 with a 1.5 kb KanMX fragment inserted 45 bp downstream of the stop codon to create the single step integrating plasmid pEAA424. The MSH4 and MSH5 SK1 sequences in these plasmids were confirmed by Sanger DNA sequencing.
pEAA424 and pEAA427 were mutagenized using Quick Change site directed mutagenesis method (Stratagene, La Jolla, CA) to create 28 msh5 and 29 msh4 point mutations. The entire open reading frame of MSH4, MSH5 was sequenced to ensure only the desired amino acid change was introduced. Table S1 shows a list of plasmids bearing the msh4, msh5 point mutations.
Full length SK1 MSH4, MSH5 and point mutant derivatives were amplified by pfu DNA polymerase and cloned into pGAD424 (prey) and target pBTM116 (target) vectors kindly provided by Nancy Hollingsworth. The entire open reading frame of MSH4, MSH5 was checked by DNA sequencing to ensure that no additional mutations were created. The L40 strain [90] was co-transformed with the Prey and Target vectors and expression of the LACZ reporter gene was determined by the ortho-nitrophenyl-β-D-galactopyranoside (ONPG) assay [91].
All msh4 and msh5 point mutations integrated into EAY1108 or NHY943 were mated to null strains bearing corresponding msh4Δ (EAY2411, EAY background; EAY2843, NHY background) and msh5Δ (EAY1280, EAY background; EAY2846, NHY background) alleles. The resulting diploids were sporulated using the zero growth mating protocol [92]. Briefly, the haploid strains were patched together on synthetic complete media for four hours and then spread on sporulation media and incubated for 2 days at 30°C. Tetrads were dissected on synthetic complete media for the EAY background and on YPD media supplemented with amino acids for the NHY background. Spore clones were replica plated onto selective media or minimal drop out plates and incubated overnight. Segregation data were analyzed using the recombination analysis software RANA to determine genetic map distances for tetrads and recombination frequencies for spores [24].
Time course, DAPI, and immunostaining analyses of meiotic progression were performed as described using antibodies to Zip1 and Msh5 [29], [93]. Stable SK1 isogenic diploid strains used in the time courses were created by mating the haploid strains shown in parentheses: Wild-type (NHY942×NHY943); msh4Δ (EAY2843×EAY2844); msh4-E276A (EAY2849×EAY2843), msh4-R676W (EAY2851×EAY2843); msh5Δ (EAY2846×EAY2848): msh5-S416A (EAY2855×EAY2846); msh5-D539A (EAY2857×EAY2846); msh5-D532A (EAY2785×EAY2846).
|
10.1371/journal.ppat.1000135 | Viral Paratransgenesis in the Malaria Vector Anopheles gambiae | Paratransgenesis, the genetic manipulation of insect symbiotic microorganisms, is being considered as a potential method to control vector-borne diseases such as malaria. The feasibility of paratransgenic malaria control has been hampered by the lack of candidate symbiotic microorganisms for the major vector Anopheles gambiae. In other systems, densonucleosis viruses (DNVs) are attractive agents for viral paratransgenesis because they infect important vector insects, can be genetically manipulated and are transmitted to subsequent generations. However, An. gambiae has been shown to be refractory to DNV dissemination. We discovered, cloned and characterized the first known DNV (AgDNV) capable of infection and dissemination in An. gambiae. We developed a flexible AgDNV-based expression vector to express any gene of interest in An. gambiae using a two-plasmid helper-transducer system. To demonstrate proof-of-concept of the viral paratransgenesis strategy, we used this system to transduce expression of an exogenous gene (enhanced green fluorescent protein; EGFP) in An. gambiae mosquitoes. Wild-type and EGFP-transducing AgDNV virions were highly infectious to An. gambiae larvae, disseminated to and expressed EGFP in epidemiologically relevant adult tissues such as midgut, fat body and ovaries and were transmitted to subsequent mosquito generations. These proof-of-principle data suggest that AgDNV could be used as part of a paratransgenic malaria control strategy by transduction of anti-Plasmodium peptides or insect-specific toxins in Anopheles mosquitoes. AgDNV will also be extremely valuable as an effective and easy-to-use laboratory tool for transient gene expression or RNAi in An. gambiae.
| Paratransgenesis, the genetic manipulation of mosquito symbiotic microorganisms, is being considered as a potential strategy to control malaria. Microorganisms associated with Anopheles mosquitoes could be manipulated to alter the mosquito's ability to become infected with and transmit the malaria parasites, or reduce mosquito fecundity or lifespan. We identified the first potential microorganism (An. gambiae densovirus; AgDNV) for paratransgenesis of the major malaria vector Anopheles gambiae. AgDNV is highly infectious to An. gambiae larvae, disseminates to adult tissues and is transmitted vertically to subsequent generations. Recombinant AgDNV was able to transduce expression of an exogenous gene (EGFP) in An. gambiae cells and mosquitoes. EGFP-transducing virions infected mosquitoes, expressed EGFP in epidemiologically relevant tissues and were transmitted to offspring in a similar manner to wild-type virus. AgDNV could be used as part of a paratransgenic malaria control strategy by transduction of anti-Plasmodium genes or insect-specific toxins in Anopheles mosquitoes, as well as an easy-to-use system for transient gene expression and RNAi for basic laboratory research.
| Transmitted by Anopheles mosquitoes, malaria is a disease responsible for inordinate mortality, morbidity and economic loss worldwide [1]. Failure of traditional control methodologies has stimulated efforts to develop novel genetic strategies to control the mosquito vectors, particularly An. gambiae. Transgenic manipulation of An. gambiae has proven to be especially challenging, with few published successes [2]–[3]. Paratransgenesis, the genetic manipulation of insect symbiotic microorganisms, is being considered as an alternative to traditional transgenic strategies [4]–[5]. Microorganisms associated with Anopheles could be manipulated to alter the mosquito's ability to become infected with and transmit the malaria parasites, or reduce mosquito fecundity or lifespan. A suitable microbial candidate for paratransgenic malaria control would have a symbiotic (mutualistic, commensal or parasitic) relationship with the vector, be readily propagated and stably engineered to express the gene(s) of interest without compromising microorganism fitness, and be easily delivered to wild mosquito populations [4]. Ideally, the engineered microbe would also be maintained in the environment, be passed to subsequent mosquito generations and have limited effects on non-target species.
Densonucleosis viruses, or “densoviruses” (DNVs), are non-enveloped single-stranded DNA icosahedral viruses in the family Parvoviridae (subfamily Densovirinae) that infect arthropods such as mosquitoes. Mosquito DNVs have narrow host ranges and are maintained in natural populations by a cycle that includes both horizontal and vertical transmission from infected adults to larvae. DNVs possess some of the smallest known viral genomes (4–6 kb), a trait that makes them highly amenable as molecular tools because the entire genome can be placed into an infectious plasmid, manipulated by standard cloning techniques, and used to express foreign genes (i.e. anti-parasite or toxin) upon infection in cell cultures or live mosquitoes [6]. DNV infectious clones, expression systems, and lethal biocontrol agents (based on the Aedes aegypti densovirus; AeDNV) have been developed and show promise for Aedes mosquitoes [6]–[8]. When injected into larvae, AeDNV virions can infect An. gambiae [9], but when infection by larval exposure to virions is attempted, AeDNV does not disseminate in An. gambiae [8]. Similar results were observed when researchers could only infect An. gambiae with TaDNV (isolated from a Toxorhynchites amboinensis cell line) by adult injection but not larval exposure [10]. Thus, DNVs have previously not been considered useful for paratransgenic manipulation or control of An. gambiae.
We serendipitously discovered a novel DNV capable of infection and dissemination in An. gambiae larvae (AgDNV) while investigating a PCR artifact in an unrelated experiment. AgDNV is highly infectious to An. gambiae larvae, disseminates to adult tissues, and is passed on to subsequent generations. Recombinant AgDNV genomes were able to transduce expression of an exogenous transgene (enhanced green fluorescent protein; EGFP) in cultured An. gambiae cells and mosquitoes and were transmitted to subsequent mosquito generations. AgDNV will form the foundation for the development of much-needed tools for routine manipulation of An. gambiae and paratransgenic malaria control.
In the course of verifying Wolbachia infection of An. gambiae cell line Sua5B [11], we observed a weak band at approximately 400 bp instead of the expected ∼600 bp fragment using the putatively Wolbachia-specific primers 81F and 691R [12]. We isolated the band from the gel for cloning and sequencing. We compared the 358 bp sequence to the BLAST database where it hit with high homology (87%) to a portion of the NS1 gene of the Aedes aegypti densovirus (AeDNV) (GenBank #M37899) [13], indicating that there was a DNV present in our Anopheles cell culture which we termed AgDNV. We used a densovirus-specific immunofluorescence assay (IFA) to visualize AgDNV infection in Sua5B cells, which confirmed localized AgDNV infection in cell nuclei [6] (Figure 1A). We then determined that AgDNV virions isolated from Sua5B cells were highly infectious to An. gambiae larvae in vivo. In order to evaluate both viral infection efficiency and lethality, we infected naïve first instar larvae (Keele strain) by either allowing larvae to feed on infected Sua5B cell cultures or by adding filtered infected Sua5B cell lysate to the larval rearing water. Both methods resulted in similarly high infection levels in emerging adults as determined by PCR (whole cells: 62%, N = 39; lysate: 57%, N = 53; Fishers Exact P = 0.67). Quantitative PCR indicated that larvae were exposed to approximately 2.1×1011±0.97×1011 viral genome equivalents per ml, which is well within the range that causes significant mortality for other DNV isolates [14]. However, we observed no difference in survival to adulthood between the controls and either infection treatment (control: 34%, N = 50; whole cells: 26%, N = 150, lysate: 35%, N = 150, chi-square = 3.27, d.f. = 2, P = 0.195), possibly due to adaptation of the virus to cell culture conditions.
To test for transtadial transmission and dissemination of AgDNV in adult mosquitoes, we infected first-instar An. gambiae larvae, transferring them to clean virus-free water after 2 days. Uninfected control larvae were exposed to culture media. After adult emergence, we dissected adult tissues and performed densovirus-specific immunofluorescence microscopy. AgDNV clearly disseminates and infects adult midgut and ovary (Figure 2). We then assessed whether AgDNV could be transmitted to subsequent generations. We treated mosquitoes for 24 hours as larvae with AgDNV, which were reared to adulthood, bloodfed, allowed to oviposit, and their offspring reared to adulthood and assayed for AgDNV by PCR. Fifty percent of treated mosquitoes were positive for virus by PCR (N = 42). Twenty-eight percent (N = 71) of their offspring were positive for infection, indicating that AgDNV was transmitted between generations, either by vertical transmission or by horizontal transmission from adults to larvae.
To purify AgDNV particles for microscopy and isolation of the viral genome, we fractionated crude Sua5B cell lysates in a cesium chloride gradient and examined fractions for viral particles by negative-stain transmission electron microscopy. We isolated numerous icosahedral, non-enveloped particles of the expected size (20 nm) (Figure 1B). We extracted the viral DNA from this gradient fraction and cloned the entire viral genome into the pBluescript S/K(-) cloning vector (denoted pBAg; Figure 3, Text S1). The cloned AgDNV genome is typical of mosquito DNVs. It is 4139 nt (GenBank #EU233812) in length and has 3 overlapping reading frames: the viral capsid and 2 non-structural (NS) proteins. The 5-prime and 3-prime ends of the genome consist of inverted hairpin repeats and are predicted to fold into perfect Y-shaped hairpin structures (Figure 4). Phylogenetic analysis of the entire AgDNV genome indicated that AgDNV falls within the “Asian” clade of known mosquito densoviruses [6]. Within the coding region, it is most closely related to a recently-described cluster of DNVs isolated from Culex pipiens pallens (CppDNV) in China [15] (Figure 5).
To confirm infectiousness of pBAg to An. gambiae cells, we transfected it into the An. gambiae cell line Moss55 (which lacks endogenous densovirus infection; Figure 1C) and observed DNV-specific signal in transfected cell nuclei by IFA (Figure 1D). However, when purified from the cell culture, virions produced from pBAg were unable to infect An. gambiae larvae in vivo. By sequencing fragments of directly-cloned viral DNA isolated from Sua5B cells, we identified multiple clones with point mutations in the 5-prime UTR and non-synonymous point mutations in the NS1 and NS2 genes (Table 1), suggesting that AgDNV was not homogeneous within Sua5B cells, but rather exists as a heterogeneous population of viral genomes that may differ in their ability to infect Anopheles larvae. To select for the viral genotype(s) that were infectious to An. gambiae larvae, we infected larvae as first-instars with virus isolated from Sua5B cells, reared them to adulthood and sequenced most of the coding portion of the AgDNV genome (nucleotides 403–3709) from 5 infected females. All 5 sequences were identical, indicating that within the viral population in Sua5B cells only one genotype was infectious to larvae. This genotype differed from pBAg at 3 sites: A636G (Lys to Glu in NS1), A1174C (Asp to Ala in NS1 and Ile to Leu in NS2) and A3399T, (Asn to Ile in capsid) (no synonymous mutations were detected). We used site-directed mutagenesis to reproduce these three mutations in pBAg (denoted pBAgα). Virions produced from pBAgα in Moss55 cells had similar infectivity to An. gambiae larvae as wild-type AgDNV from Sua5B cells as determined by both PCR and IFA.
We used pBAg to create a flexible gene transduction construct by deleting most of the viral genome between the hairpin sequences and inserting a multiple cloning site (pBAgMCS; Figure 3, Text S1). Using pBAgMCS, we can easily construct viral transducing genomes carrying any gene-promoter combination of interest, and by supplying the missing viral proteins in trans with pBAgα or wild-type virus, we can express the gene in An. gambiae mosquitoes simply by adding the virions to the larval rearing water. As proof-of-concept, we inserted the enhanced green fluorescent protein (EGFP) under control of the constitutive Drosophila actin5C promoter into the multiple cloning site of pBAgMCS (pAgActinGFP; Figure 3, Text S1). When pBAgα and pAgActinGFP were simultaneously transfected into Moss55 cells, we observed cytoplasmic EGFP expression 24–48 hours post-infection (Figure 1E, F). We observed fluorescent cells in the culture even after 10 passages (approximately 2 months), indicating that the helper and transducing virions were replicating in the cells. We do not believe that these results are due to integration of the viral genome into the host genomic DNA, as integration is not known to occur for DNVs in the genus Brevidensovirus (the genus AgDNV belongs to), although integration does occur for other DNV genera [6].
We purified helper and EGFP-transducer virions from transfected Moss55 cells, exposed first-instar An. gambiae larvae to them and assayed emerged adults for EGFP expression by fluorescence microscopy. EGFP expression was observed in approximately 50% of adults (N>100). We observed similar results when virus from Sua5B cells rather than pBAgα was used as helper. EGFP expression was first observed in the fat body, later disseminating to other tissues such as the eye, midgut, hindgut, malpighian tubules and ovaries (Figures 6 and 7).
EGFP-positive mosquitoes were allowed to reproduce. We observed EGFP expression in approximately 20% of F1 offspring (N>50, Figure 8) and detected EGFP DNA by PCR and sequencing from EGFP-expressing F1 mosquitoes (N = 8). We continued to breed the offspring and again assessed EGFP expression in the F3 generation, where 20% of the mosquitoes had observable EGFP fluorescence (N = 20). These data indicate that AgDNV can be used to drive expression of exogenous transgenes in An. gambiae and that transducing virions are transmitted to subsequent generations, similar to wild-type virus. While it is not clear at this point whether offspring are infected by transovarial transmission or horizontal transmission from adults to larvae, we detected EGFP in both developing ovarioles and in mature oocytes (Figure 7) suggesting that transovarial transmission may be involved.
The development of novel, efficacious malaria control methods is critical to reduce the enormous public health and economic burdens experienced in affected areas. Densovirus-based tools for control of Anopheles mosquitoes are very attractive for this purpose due to their specificity, stability, ease in engineering, ability to spread horizontally and vertically and accumulate in natural environments, and recent advances in large-scale production and purification methods [16]–[17]. Recombinant AgDNV could potentially be used to control malaria by transduction in An. gambiae of anti-Plasmodium peptides to block parasite transmission or insect-specific toxins to reduce mosquito population density or mosquito lifespan. AgDNV will also be extremely valuable as an effective and easy to use laboratory tool for transient gene expression or RNAi [6] in An. gambiae.
The Anopheles gambiae Keele strain was used for experiments in 30 cm cube cages kept in a walk-in insectary at 28°C and 80% relative humidity. Mosquitoes were allowed access to a cotton wick soaked in 20% sucrose as a carbohydrate source. Adults were allowed to bloodfeed on an anesthetized mouse 5 days post-emergence. Two days after bloodfeeding, an oviposition substrate (consisting of a filter paper cone inside a 50 ml beaker half-filled with water) was introduced into cages and filter papers containing eggs removed the next day, placed into a 41×34×6 cm rearing tray half-filled with distilled water and one pellet dry cat food, with one additional food pellet added daily after day 3. Pupae were picked with an eye-dropper, placed in a cup and introduced into cages (∼200 pupae/cage) to begin the next generation.
The Anopheles gambiae cell lines Sua5B and Moss55 were grown at room temperature in Schneider's medium (Sigma) supplemented with 10% fetal bovine serum. DNAs used for transfection were prepared using a QIAGEN Plasmid Purification Kit (Qiagen, Valencia, CA) according to the manufacturer's protocol. For the transfection of cells with different plasmids, one µg of total plasmid DNA (0.5 µg vector and 0.5 µg helper) was used with Effectene® Transfection Reagent (QIAGEN) according to the manufacturers suggested protocol.
Genomic DNA was extracted from Sua5B cells using DNEasy kits (QIAGEN, Valencia, CA) according to the manufacturer's suggested protocol. Unexpected PCR amplification of an approximately 400-bp fragment of AgDNV was amplified using Wolbachia primers wsp81F (5′-TGG-TCC-AAT-AAG-TGA-TGA-AGA-AAC-3′) and wsp691R (5′-AAA-AAT-TAA-ACG-CTA-CTC-CA-3′) [12]. PCR amplicons were separated by 1% agarose gel electrophoresis, stained with ethidium bromide, and visualized with UV light. PCR fragments were cloned into the pCR4-TOPO vector and sequenced. We detected AgDNV infection in infected mosquitoes using primers DensoVF (5′-CAG-AAG-GAT-CAG-GTG-CAG-3′) and DensoVR (5′-GCT-ACT-CCA-AGA-GCT-ACT-C-3′) using Sua5B as a positive control and water as a negative control.
Cells were grown overnight in 8-well chamber slides, then fixed with 4% paraformaldehyde. Fixed cells were washed 3 times with PBS, permeabilized with 0.01% Triton X-100 in PBS, and washed 3 times in PBS. Cells were incubated in 1% BSA, PBS pH 7.4 for 30 min to block non-specific antibody binding. Cells were incubated with primary antibody (1∶1000) in 1% BSA, PBS pH 7.4 for 60 min and washed for 10 minutes three times with PBS pH 7.4. Cells were incubated with goat anti-rabbit IgG FITC conjugate (Sigma) (1∶500), Evans Blue (1∶1000), in 1% BSA, PBS pH 7.4 for 60 min at RT, then washed for 10 minutes three times with PBS pH 7.4. Cells were stained with DAPI, mounted and visualized by epifluorescent microscopy.
First-instar larvae were either introduced directly into culture flasks containing Sua5B cells or were infected by adding Sua5B cell lysate to the rearing water. In this case, Sua5B cells were pelleted in a 50 ml conical tube by centrifuging for 10 minutes at 2,500 G, 4°C. The pellet was resuspended in 20 ml PBS. Cells were lysed by vortexing with sterile 3 mm borosilicate glass beads for 5 minutes. Approximately 20 ml cell lysate was added to 20 ml ddH20 with approximately 50 first-instar An. gambiae larvae Keele strain (4 replicates). Larvae were exposed to virus for 24 hours, then were transferred to clean water with larval food.
First-instar larvae were infected with Sua5B lysate, reared to adulthood, allowed to bloodfed on an anesthetized mouse approximately one week post-emergence, and offspring produced as described above. Adults and offspring were tested for AgDNV by PCR using primers DensoVF and DensoVR, using Sua5B as a positive control and water as a negative control.
Sua5B cells were pelleted and lysed as described above. The supernatant was removed to a new tube and cellular debris pelleted by centrifuging for 20 minutes at 10,000 G, 4°C. The supernatant was centrifuged at 35,000 rpm for 75 minutes, 4°C to pellet virion particles. The virion pellet was removed and further purified by 1 M sucrose cushion centrifugation for 120 minutes at 39,000 rpm, 4°C. The final pellet was fractionated in a CsCl (0.3 g/ml) gradient at 60,000 rpm overnight at 8°C. The virion band was removed from the gradient for DNA extraction and TEM.
Purified virus particles were applied to glow-discharged carbon-coated grids and negatively stained with 2% (w/v) uranyl acetate. Electron micrographs were recorded on Kodak SO-163 film using a Philips CM12 electron microscope at nominal magnifications of 37,000× to 52,000×.
Pure virion particles isolated from the gradient were incubated in 300 µl buffer (100 mM EDTA, 10 mM Tris-HCl, 0.1% SDS, 100 µg/ml proteinase K, pH 8.0) overnight at 55°C. The next day, the mixture was centrifuged at 14,000 rpm for 2 minutes to pellet debris. DNA was extracted from the supernatant twice using 1 volume of phenol∶chloroform (1∶1). One tenth volume of 3 M sodium acetate and 2.5 volumes of cold ethanol were added to precipitate viral DNA. DNA was pelleted by centrifugation at 14,000 G for 20 minutes, washed with 70% cold ethanol, air dried and resuspended in 5 µl 10 mM Tris-HCl (pH 8.5).
600 ng AgDNV genomic DNA was blunt-ended by incubating for 15 minutes at room temperature with 10 units Klenow fragment. Viral DNA was ethanol precipitated, cloned into the EcoRV site of plasmid pBluescript S/K(-) and transformed into SURE® competent cells (Stratagene). 20 clones were selected and sequenced to confirm viral inserts. We were unable to clone the entire AgDNV genome in one step, and thus assembled the genome from two clones that, together, contained the entire AgDNV genome. These clones were digested with NcoI and XbaI and ligated together to build a full-length infectious clone (pBAg). pBAg infectivity in Moss55 cells was confirmed by transfection and IFA as described.
pBAg plasmid was used as a copy-number standard for viral genome quantification as previously described [14]. The plasmid has an estimated mass of 7.78×10−18 g/copy. Plasmid concentrations were determined using an ND-1000 NanoDrop spectrophotometer (Thermo Fisher Scientific), and serial dilutions were made from 50 µM to 5×10−8 µM to generate a standard curve that ranged from 6.4×1010 viral genome equivalents/µL (geq/µL) to 6.4×100 geq/µL in ten-fold increments. Primers were designed based on regions within the overlapping NS1 and NS2 genes that were highly conserved amongst all known mosquito densovirus isolates, as previously described [14]. The forward primer (5′-CAT-ACT-ACA-CAT-TCG-TCC-TCC-ACA-A-3′) and reverse primer (5′-CTT-GGT-GAT-TCT-GGT-TCT-GAC-TCT-3′) produce an 183 bp amplicon. The Quantitect SYBR Green Kit (Qiagen) was used in a 25 µL reaction containing 0.3 µM of each primer, and 5 µL of a 1/100 dilution of the Sua5B viral infection prep. Real-time PCR was performed on an ABI Prism model 7300 using 96-well reaction-plates (ABI) and MicroAmp Optical Adhesive Film (ABI) with a program of: (1) 50°C for 2 min, (2) 95°C for 15 min, (3) 45 cycles of i) 94°C for 15 sec, ii) 55°C for 30 sec , iii) 72°C for 30 sec. Data was collected each cycle at step 3iii, and the 45th cycle was followed by a dissociation program to verify specific amplification.
Virions produced by pBAg were not infectious to An. gambiae larvae. We infected larvae with virus isolated from Sua5B cells, reared larvae to adulthood and screened for infected mosquitoes by PCR as described. We selected 5 individual infected mosquitoes, sequenced the coding region of the virus that infected them and identified 3 mutations that all had in common as described in the text. We reproduced these mutations in pBAg by site-directed mutagenesis using the QuikChange Multi-Site Directed Mutagenesis Kit (Stratagene) with the manufacturer's protocol.
pBAgMCS carries a multiple cloning site (MCS) flanked by the 5-prime and 3-prime AgDNV hairpin sequences. The MCS possesses 5 common unique cloning sites: NsiI, NcoI, MluI, EcoRV and BglII (and several other less common cut sites, Figure 5). NsiI, NcoI, MluI, and BglII produce sticky ends for directional subcloning, while EcoRV produces blunt ends for blunt-end ligation procedures. We used pBAg as template for PCR using primers MCSF3 (5′- CCC-AAA-CCT-ATA-TAA-GGC-AAC-TGG-AAT-CGA-AGG-A -3′) and MCSR2 (5′- CCA-ATG-CAT-CCA-TGG-ACG-CGT-GAT-ATC-AGA-TCT-TGT-ATT-GTC-TCG-GTG-CA-3′) to amplify part of the 3-prime UTR, adding the MCS to the amplicon as part of the primer. The resultant product and pBAg were double-digested with NsiI and EcoNI. The digested pBAg was CIP-treated to prevent autoligation, and the 2 products ligated together with T4 ligase. The construct was transformed into SURE Competent cells (Stratagene), clones screened and proper vector construction confirmed by sequencing.
The actin5C-EGFP-SV40 cassette was PCR-amplified from pHermes[act5C:EGFP] using primers Actin5CegfpF (5′-CCC-AAA-GAT-ATC-CGA-TCG-CTC-CAT-TCT-TG-3′) and Actin5CegfpR (5′-CCC-AAA-GAT-ATC-CGC-TTA-CAA-TTT-ACG-CC-3′) using pfuUltra II Fusion HS DNA Polymerase (Stratagene) with the manufacturers suggested protocol. The PCR product was digested with EcoRV. pBAgMCS was digested with EcoRV, and CIP-treated to prevent autoligation. The 2 products were ligated together with T4 ligase and the construct transformed into SURE competent cells. Clones were screened and proper insert confirmed by sequencing.
A combination of pBAgα and pAgActinGFP were transfected into Moss55 cells (or pAgActinGFP into Sua5B cells) as described. EGFP expression was monitored by fluorescence microscopy daily beginning 24 hours post-transfection. For mosquito infections, virion particles were purified from cells 1–2 weeks post-transfection by glass bead lysis/filtration and first-instar larvae infected directly as described above. EGFP expression in cells, dissected tissues and mosquitoes was monitored using an Olympus BX41 epifluorescent compound microscope. Images were captured using a Macrofire monochrome digital camera (Optronics).
Mosquitoes which had observable EGFP expression were allowed to oviposit, offspring reared and EGFP expression in offspring assessed as described above. DNA was extracted from positive offspring and EGFP DNA detected using primers egfpF2 (5′-TGA-AGT-TCA-TCT-GCA-CCA -3′) and egfpR2 (5′-CAG-CAG-GAC-CAT-GTG-ATC-3′). PCR was conduced using pAgActinGFP as a positive control and water as negative control. Amplicons were gel purified and directly sequenced.
|
10.1371/journal.pgen.1003398 | Dynamic Circadian Protein–Protein Interaction Networks Predict Temporal Organization of Cellular Functions | Essentially all biological processes depend on protein–protein interactions (PPIs). Timing of such interactions is crucial for regulatory function. Although circadian (∼24-hour) clocks constitute fundamental cellular timing mechanisms regulating important physiological processes, PPI dynamics on this timescale are largely unknown. Here, we identified 109 novel PPIs among circadian clock proteins via a yeast-two-hybrid approach. Among them, the interaction of protein phosphatase 1 and CLOCK/BMAL1 was found to result in BMAL1 destabilization. We constructed a dynamic circadian PPI network predicting the PPI timing using circadian expression data. Systematic circadian phenotyping (RNAi and overexpression) suggests a crucial role for components involved in dynamic interactions. Systems analysis of a global dynamic network in liver revealed that interacting proteins are expressed at similar times likely to restrict regulatory interactions to specific phases. Moreover, we predict that circadian PPIs dynamically connect many important cellular processes (signal transduction, cell cycle, etc.) contributing to temporal organization of cellular physiology in an unprecedented manner.
| Circadian clocks are endogenous oscillators that drive daily rhythms in physiology, metabolism, and behavior. In mammals, circadian rhythms are generated within nearly every cell; and, although dysfunction of circadian clocks is associated with various diseases (including diabetes and cancer), the molecular mechanisms linking the clock machinery with output pathways are little understood. Since essentially all biological processes depend on protein–protein interactions, we investigated here on a systems-wide level how time-of-day-specific protein–protein interactions contribute to the temporal organization of cellular physiology. We constructed a circadian interactome using experimentally generated protein–protein interaction data and made this network dynamic by the incorporation of time-of-day-dependent expression data. Interestingly, systematic genetic network perturbation (RNAi and overexpression) suggests a crucial role for circadian components involved in dynamic interactions. Systems analysis of a global network revealed that interacting proteins are in the liver significantly more expressed at similar daytimes likely to restrict regulatory interactions to specific circadian phases within cells. Overall, circadian protein–protein interactions are predicted to dynamically connect important cellular processes (signal transduction, cell cycle, etc.) using—very often—protein modules with components co-expressed in time, shedding new light on the daily organization of cellular physiology.
| Circadian clocks are endogenous oscillators conserved in nearly all living organisms that drive ∼24 hour cycles in physiology and behavior. In mammals, the circadian oscillator is composed of interconnected transcriptional translational negative and positive feedback-loops which generate circadian rhythms at the molecular level. Within this gene-regulatory network, a precise timing of gene expression, protein–protein interactions (PPIs) as well as posttranscriptional and posttranslational modifications is essential for sustaining circadian rhythms with normal dynamics [1]–[3]. The interaction between the transcription factors CLOCK and BMAL1, which has been discovered in a yeast-two-hybrid (Y2H) screen [4], is crucial for the activation of the Period (Per1, Per2, Per3) and Cryptochrome (Cry1, Cry2) genes. PER and CRY proteins form large complexes that inhibit their own transcription by binding directly to the CLOCK/BMAL1 complex during the late night [5].
Circadian rhythms in gene expression are pervasive – 2–10% of the transcriptome in a given tissue is under circadian control [6], [7]. Consequently, also a large fraction of the proteome is thought to be regulated in a time-of-day dependent manner, although systems-wide studies of circadian protein abundance rhythms are still rare (however, see [8]). Cellular functions are increasingly recognized to be regulated by protein complexes or ‘modules’ [9], thus PPIs and their timing are predicted to be crucial. In most cases, in which PPIs exert a regulatory function, such interactions are transient and occur only under specific conditions, e.g. as a response to a signal, after binding of a co-factor or when the expression of one or both partners is induced in response to a changing cellular condition. Circadian clock regulation of cellular functions via PPIs can be accomplished by restricting important interactions to specific times of the day. In the circadian oscillator, many of the known PPIs also happen predominantly at specific times of the day, e.g. PER/CRY complexes bind to CLOCK/BMAL1 in the late night to inhibit transactivation [5]. Here, the temporal binding profile correlates with the abundance profiles of PER and CRY proteins. While these examples demonstrate the fundamental importance of precisely timed PPIs for the circadian clockwork, we are still far from a comprehensive view of the PPI network among circadian oscillator proteins and their dynamics. Furthermore, the extent of a regulation of circadian output processes via time-of-day dependent PPIs is largely unknown.
To elucidate unknown regulatory mechanisms within the circadian clockwork we have systematically mapped PPIs among 46 circadian components using high-throughput Y2H interaction experiments. We have identified 109 so far uncharacterized interactions and have successfully validated a sub-fraction via co-immunoprecipitation experiments in human cells. Among the novel PPIs we have identified modulators of CLOCK/BMLA1 function indicating a role for protein phosphatase 1 (PPP1) in the dynamic regulation of BMAL1 abundance. Furthermore, to generate a more comprehensive circadian PPI network we have enriched and extended our experimental network with additional validated interactions and interaction partners from literature, some of which seem to be essential for normal circadian dynamics. The integration of circadian mRNA expression profiles from mouse liver allowed us to predict the interaction dynamics within our network in hepatocytes. Using systematic genetic perturbation studies (RNAi and overexpression in oscillating cells) we propose a crucial role of dynamic regulation (via rhythmic PPIs) for the molecular clockwork. Furthermore, we have extracted a dynamic modular organization as a pervasive circadian network feature possibly contributing to time-of-day dependent control of many cellular processes. Systems analysis on a global scale regarding circadian regulation of biological processes via rhythmic PPIs suggests a time-of-day dependent organization of the interactome. Altogether our data should provide a valuable resource of circadian PPIs within hepatocytes that are important not only for keeping the pace of the molecular clockwork but likely also for the control of cellular physiology.
To systematically map the PPIs within the circadian clock regulatory network, we performed a matrix-based two-hybrid screen in yeast with 46 known or assumed clock or clock-associated components (Figure 1A; for justification of our selection see Text S1 and Table S1). In this screen, each potential interaction was tested individually in six replicas to increase screening saturation thereby minimizing the number of false negatives (for details on the method, see Figure S1 and [10]). After excluding transcriptionally autoactive components, we performed 11,040 individual yeast-two-hybrid experiments monitoring growth on selective medium and β-galactosidase activity as readouts for interaction (Figure 1B). Thereby, we identified 150 interacting protein pairs that occurred at least in two independent experiments (Figure 1C). We could reproduce a large number (41 of 104) of previously described interactions (e.g., CLOCK-BMAL1, PERs-CRYs, CRYs-BMAL1; see Figure S2A and Table S1) corresponding to a rather high sensitivity (of ∼40%) for a yeast-two-hybrid assay, which is usually only about 25% [11]. Importantly, among the 150 detected PPIs we found 109 previously unknown PPIs between circadian clock proteins. For example, we detected interactions between DEC1/2 and CRY1/2, between CLOCK and RORβ/γ, between CLOCK and the α-catalytic subunit of protein phosphatase 1 (PPP1Cα) as well as between BMAL1 and WDR5 (Figure 1B, 1C).
To test whether the PPIs discovered in yeast can also occur in mammalian cells, we performed co-immunoprecipitation experiments in HEK293 cells. As representatives for the novel PPIs we focused on the interactions of the transcriptional activators CLOCK and BMAL1 – central players within the circadian clock gene-regulatory network, whose functional modulation by interacting proteins is likely to be highly relevant for normal circadian rhythms. Twelve of the 14 (i.e., 86%) novel CLOCK and BMAL1 interactions found in yeast were validated using co-immunoprecipitation (Figure 1D and Figure S2), suggesting that a substantial proportion of all interactions identified in yeast can also take place in mammalian cells.
To understand the structure and the organizing principles of the complex web of interactions occurring between circadian clock components, we created an interaction network using our novel yeast-two-hybrid interaction data together with previously published interactions among these components. This is necessary, since the sensitivity of any high-throughput PPI detection assay is limited [11] and thus the false-negative rate is expected to be rather high. In addition, we extended this network by adding known interacting proteins (direct ‘neighbors’) of our network components (except regulatory components such as kinases, phosphatases and F-box proteins, which are known to be involved in many other cellular processes) to get an idea how the circadian PPI network is embedded in the cellular interactome (Figure S3A). To this end, we used PPI data extracted by human experts from literature and stored in the UniHI database [12], however only those, for which experimental validation exists. We did not use predicted PPIs based on orthology or from computational text mining. Thereby, a large PPI network with 134 components and 625 PPIs was created consisting of a circadian clock core (24 components), regulatory components (22 components) and the neighborhood (88 components; Figure 2 and Table S1).
For this network, a mean shortest path length between any two proteins of 2.8 links was calculated, i.e. most proteins are very closely linked to each other indicating a ‘small world’ type of network [13]. Like many other PPI networks [14], the circadian network has properties of a ‘scale-free’ network, i.e. many proteins have few and few proteins have many interactions (Figure S3B). On average each component has 8.4 interaction partners, however, 11 proteins are highly connected with more than 20 interactions (e.g. CLOCK, BMAL1, PER2, CREBBP, DEC1, AR, HDAC1). Network topology analysis further revealed that our network is hierarchically organized, i.e. highly connected components (so-called ‘hubs’) link network regions with less connected components, which themselves tend to form clusters (Figure 2; Figure S3B).
Proteins in the direct network neighborhood that interact with circadian clock core components might be relevant for regulating clock output functions, but could also include yet unknown proteins important for modulating the circadian clock machinery; i.e. they might be clock components themselves. To test the latter possibility, the expression of 88 neighborhood genes was systematically downregulated by RNAi in human U2OS cells. These cells possess robust circadian rhythms in cell culture, and RNAi-mediated downregulation of canonical clock genes has been shown to copy circadian phenotypes of classical knockout mice [15], [16]. We monitored circadian rhythms via a stably integrated Bmal1 promoter-luciferase reporter construct and identified 21 components of the neighborhood that altered circadian period upon knockdown by at least 0.5 hours (Figure 3 and Table S2). For example, downregulating the cell-cycle kinase CHEK1 (that can interact with TIMELESS and CK2) significantly shortened the circadian period by more than 1 hour, while downregulating the DNA helicase binding protein CDH4 (that is reported to interact with RORγ) lengthened it. In addition, knocking down the androgen receptor (AR), which interestingly was found to interact with many proteins (including NONO, GSK3β, HDAC1, CREBBP, UBE2I and NCOR1/2), results also in a shortening of the circadian period by almost one hour. Although these results need additional in-depth validation, the relatively high number of clock modulating components in the network neighborhood suggests the presence of yet uncharacterized mechanisms in the molecular circadian oscillator, as suggested earlier [15], [16]. Future studies are needed to investigate whether these circadian phenotypes in U2OS cells are similar in other cell types and in vivo.
In addition to possibly being novel clock components, proteins in the network neighborhood might also connect specific cellular processes to circadian control by means of directly interacting with clock components. Such interactions are likely time-of-day dependent, which may be accomplished by rhythmic abundance levels of one or both of the interaction partners. Therefore, we hypothesized that the whole network but also the neighborhood alone are significantly enriched in components with rhythmic abundance levels. This is indeed the case – at least if we consider (due to the lack of protein abundance data) mRNA expression profiles of network components in mouse liver tissue – the circadian transcriptome with highest available temporal (1 h) resolution [6]. Of the 134 network components, 65 (49%) show a significantly rhythmic mRNA expression profile in liver, a highly significant enrichment when compared to a random selection of genes from this expression data set (p<10−6; Chi-squared test). This may not be surprising, since the network as a whole contains circadian oscillator components, many of which are known to be rhythmically transcribed. However, if we analyze the neighborhood separately, we still find a significant enrichment (p<10−4; Chi-squared test) in components that are rhythmically transcribed: of the 88 components in the neighborhood, 38 (43%) are rhythmically expressed in the liver (Figure 2, yellow circles) suggesting that PPIs in the hepatocyte circadian clock network might indeed be a means to mediate rhythmic control of cellular physiology.
At what time of day do the PPIs in the circadian network occur or – in other words - can we predict dynamic properties of our (still static) network? Again, hypothesizing that a PPI more likely happens at times, when the interaction partners are co-expressed, we again used transcriptome data (from mouse liver) [6] as a validated proxy for protein abundance [17] – an approach successfully used also for the yeast interactome [18]. To first test this hypothesis for PPIs in general (i.e. beyond our circadian network), we compared the Pearson correlation coefficient (PCC) of transcript levels (as a measure for co-expression) for all pairs of interacting liver proteins present in the UniHI interactome database and for which we have time-resolved expression data [6] with the PCC of randomly chosen pairs. Interestingly, we found that interacting liver proteins are significantly more likely to be expressed at similar circadian times (PCC>0.5 with PCC ranging from −1 to 1; p<10−15 Chi-squared test; Figure 4A left and Figure S5). These data suggest that circadian co-expression may be a common feature to restrict regulatory interactions to specific times of the day. This assumption is supported by the fact that liver proteins with many interaction partners – which are known to exert regulatory functions - are more likely to be rhythmically expressed (p<10−10; Wilcoxon Rank test) and vice versa, i.e. proteins with rhythmic transcripts have statistically more interaction partners than constitutively expressed proteins (p<10−5; Chi-squared test). Interestingly, also the circadian PPI network displays these properties: interaction between proteins is more likely, when both proteins are co-expressed in time (Figure 4A, right).
Based on our results above, we hypothesized that many PPIs happen at specific times of the day. Therefore, we assigned to each PPI in our network a circadian phase, at which the corresponding components are predicted to interact in the liver based on their transcript expression. To this end, we approximated the abundance of the complex of two proteins as the product of their expression profiles. Derived time series for the interaction complexes were subsequently examined for 24 hour periodicity with a stringent threshold (false discovery rate FDR<10−5) resulting in the prediction of a dynamic circadian PPI network with 193 individual protein pairs interacting at specific circadian phases (Figure 4B and Table S3). Interestingly, PPIs in the liver seem to be distributed over the whole circadian cycle. Beyond the dynamic interactions that occur among circadian core components in this network, we extract many time-of-day specific putative regulatory interactions within the neighborhood. For example, the lysine acetyltransferase KAT2B is predicted to bind to the nuclear receptor coactivator NCOA1 - two proteins involved in transcriptional regulation - during the late day, which may hint to a time-of-day specific function of these proteins. Nevertheless, it should be noted that this prediction is only valid for the liver, since the identity of rhythmic transcript is highly tissue-specific [19]. In addition, we are aware that the restriction to transcript (and not protein) profiles, the possible tissue-specificity of certain PPIs and also a potential competitive nature of the possible interactions pose limitations to this analysis (but see below for experimental validation of the daytime dependent interaction between PPP1Cα and CLOCK/BMAL1). However, such a framework offers the possibility to globally analyze processes controlled by circadian PPIs in a time-specific manner.
Network components with many interaction partners - so-called ‘hubs’- not only have important organizing properties in scale-free networks; they are also (controversially) discussed to be more essential for life (at least in yeast, Drosophila and C. elegans; [20]). In a dynamic network, two types of hubs have been proposed – ‘party hubs’, which interact with their partners predominantly at similar times, and ‘date hubs’, whose interactions mostly occur at different times or locations [18]. In yeast, especially the ‘date hubs’ are described to be global regulators for the cellular physiology suggesting a prominent role of dynamic regulation within complex networks.
To test, whether in our network ‘hub’ proteins are essential for the trait ‘circadian rhythmicity’ - i.e. for generating and maintaining circadian rhythms – we correlated circadian phenotypes obtained upon genetic perturbation (see below) with topological characteristics of network components. For perturbing the network experimentally and assigning an essentiality score (for circadian rhythmicity) to each component, we (i) systematically knocked down and (ii) overexpressed every component of the core and the regulatory part (not the neighborhood) of the network. We performed these experiments in human U2OS reporter cells (as described above) and analyzed the effect on circadian dynamics. While we could reproduce most of the phenotypes that have been known from studies with classical knockout models (e.g. the opposite period phenotypes upon Cry1 and Cry2 deletion as well as arrhythmicity upon Bmal1 and Clock knockout), we detected interesting novel phenotypes such as period lengthening for Rev-Erbβ (Nr1d2) downregulation (Figure 5A and Figure S4A). As examples for phenotypes detected upon clock protein overexpression, Dec1 or Dec2 as well as Fbxl15 (the homologue to Drosophila Jetlag) led to a substantial period lengthening (∼1.5 hours and ∼6 hours, respectively) (Figure 5B and Figure S4B). For each network component tested we combined the downregulation and overexpression phenotypes in a ‘phenotypic score’ (for rules, see Text S1) to be able to correlate it with network properties of the individual components. Surprisingly, we did not see a correlation of phenotypic score with the number of interactions as it has been observed in more global networks of yeast, Drosophila and C. elegans [20]. In other words, ‘hub’ proteins apparently are not more important for circadian rhythmicity than components with a lower connectivity. However, proteins that are predicted to be involved in dynamic interactions (at least in liver) turned out to be more essential for circadian rhythm generation (t-test: p<0.01; Mann-Whitney U test: p<0.01; Figure 5C, Table S4). For example, CLOCK, BMAL1, PER3 and CRY1 – to which we assigned 24, 21, 18 and 9 dynamic interaction, respectively – are especially important for circadian dynamics (Figure 5A–5C). Importantly, factors that have a rhythmic transcript per se (without taking PPIs into account) are not significantly more likely to be essential for circadian rhythms (independently of whether we set the rhythmicity threshold at a FDR of 0.05 or 0.01; not shown). While we did not test the importance of rhythmic PPIs for circadian dynamics directly, this correlative result suggests that the more rhythmic interactions a protein is involved in, the more important it is for normal circadian rhythmicity. In addition, 40% (10 of 25) of the 45 ‘hubs’ that qualify as ‘party hubs’ in the liver (Table S3) show circadian phenotypes upon genetic perturbation, while the only two ‘date hubs’ (CLOCK and AR; Figure S5) both are sensitive to perturbation - in line with the described prominent role of ‘date hubs’ for network organization [18].
Are dynamic PPIs also important for the regulation of cellular events? To predict such regulations, we first assigned to each network component one or more specific gene ontology (GO) categories from a reduced, less redundant and more distinct set of GO categories (for details see Text S1). Secondly, using the information whether a PPI is likely to be dynamic or not (see Figure 4B), we investigated which cellular processes (as represented by GO categories) are significantly connected via dynamic PPIs (see Text S1). In other words, we tested whether in our network dynamic interactions are over-represented in the total set of interactions between a pair of GO categories. This resulted in a “process network” with 12 dynamic links between 11 biological processes with ‘circadian rhythm’ as the central hub. This hub is rhythmically connected with GO terms such as ‘DNA repair’, ‘transcriptional regulation’ and ‘response to external stimulus’ (Figure 6A; Table S5). A strong association of the circadian clock network with these processes relevant for e.g. cancer and cell-cycle is also found by (i) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis of the network neighborhood only (Figure 6B) and (ii) the significant (p<10−8; Chi-squared test) enrichment of the network neighborhood with cancer-associated genes (as reported in the Cancer Gene Census list (www.sanger.ac.uk/genetics/CGP/Census)).
How are these rhythmically regulated processes connected in our network - by individual components or rather by functional modules consisting of interconnected components? We explored our circadian PPI network topology for clusters of highly connected proteins (structural modules) and identified 11 different modules within the circadian network (Figure 6C; Figure S6A and Text S1) often with module components co-expressed in time suggesting that modular organization within the circadian PPI network might contribute to a coherent functional regulation of hepatocyte processes by the circadian clock. This is also supported by a high cluster coefficient (0.38) of the circadian network compared to randomized networks (0.14±0.01) (Figure S3C).
Next, we analyzed whether time-of-day dependent interaction of cellular processes via PPIs can also predicted on a more global scale. To this end, we first assigned 2788 rhythmic PPIs (using the approach described above - see Figure 4B) to a global interactome derived from the UniHI database (Figure 6D left) and then searched for GO terms (‘biological process’) that are significantly connected via predicted dynamic PPIs (Figure S6B). We extracted a network of 20 biological processes with 89 dynamic links (for details see Text S1). The central ‘hub’ of this ‘process network’ constitutes the term ‘signal transduction’ (Figure 6D middle and Figure S6C) suggesting a possible time-of-day dependent modulation of hepatocyte events such as ‘protein transport’, ‘response to stress’ and ‘cell death’ by signaling pathways via rhythmic PPIs.
To characterize the underlying PPI network properties, we constructed a global dynamic interactome and found that it again has ‘scale-free’ properties with 269 dynamic ‘hubs’, i.e. proteins with at least 5 predicted dynamic interactions. The protein with the most predicted rhythmic interactions (79 of 105 in total) is heat-shock protein HSP90AA1 – a factor required for proper protein folding upon heat stress. Notably, three of the four interaction-richest proteins (with more than 40 interactions) are cell-surface receptors (estrogen receptor 1, transforming growth factor beta receptor 1 and platelet derived growth factor receptor, beta) again suggesting a central role of signaling pathways for dynamic regulation in the liver (Figure 6D right).
Our systems biology analysis of the circadian interactome points to a timely regulated action of chromatin modifying enzymes (see Figure 6A, 6C). It is known that at the heart of the circadian oscillator binding of the transcription factor heterodimer CLOCK/BMAL1 is controlled by methylation and acetylation states of histones at specific promoter regions [21]–[25]. In addition, CLOCK/BMAL1 transactivation activity is modulated by a precisely timed acting repressor complex. In our Y2H screen, we discovered 15 new interaction partners for CLOCK/BMAL1, which might play a role in modulating their function in cells (see Table S1). We systematically tested whether these interactors (and/or their paralogs – in total 28) are able to modulate CLOCK/BMAL1 transactivation measured from an E-box containing artificial promoter with firefly luciferase as reporter (Figure S7A). As expected, already characterized CLOCK/BMAL1 repressors such as CRYs, PER2 and DECs [4], [26], [27] substantially inhibited transactivation. Interestingly, among the 15 new interactors (including their paralogs) PPP1Cα (protein phosphatase 1 alpha, catalytic subunit), but not PPP1Cβ severely and RORs moderately reduced the reporter signal (Figure 7A). The effect of PPP1Cα on CLOCK/BMAL1-mediated transactivation was dose dependent (Figure 7B).
Our in silico analysis predicted that PPP1Cα binds to the CLOCK/BMAL1 complex in mouse liver in a time-of-day specific manner. We tested this by co-immunoprecipitation experiments using antibodies against endogenous proteins (Figure 7C). We selected circadian times that were predicted by our dynamic interaction analysis (Figure 4B) to correspond to maximal and minimal likelihood of PPI. Indeed, we detected an association of endogenous PPP1Cα with CLOCK/BMAL1 at CT0 (CT = circadian time) while only little PPP1Cα-CLOCK/BMAL1 complex was found at CT12 suggesting a circadian time-dependent modulation on CLOCK/BMAL1 function. As CLOCK and BMAL1 phosphorylation have been described to affect their stability [28], [29], we tested whether PPP1Cα acts on this level. We stably expressed CLOCK and BMAL1 as a GFP-fusion protein in human U2OS cells with DsRed (a red fluorescent protein) on the same transcript [30], [31]. Protein stability can be monitored via the ratio of GFP to DsRed signal using FACS analysis thereby normalizing for different transcription rates in individual cells. As proof-of-concept of this approach, we confirmed the previously reported destabilizing effect of GSK3β on BMAL1 [32] (Figure S7B–S7D). While we could not detect an effect of PPP1Cα on CLOCK stability, we saw a substantial and significant decrease of BMAL1 abundance (Figure 7D). In addition, endogenous BMAL1 levels were reduced by about 50% upon stably overexpression of PPP1Cα in U2OS cells (Figure 7E). Lower BMAL1 abundance in the presence of PPP1Cα is likely due to reduced BMAL1 stability, since cycloheximide treated cells (in which de novo protein synthesis is blocked) revealed a much faster degradation of endogenous BMAL1 when PPP1Cα is overexpressed (Figure 7F). Together, these data indicate that BMAL1 stability and probably thereby transactivation is regulated by PPP1Cα.
Protein–protein interactions among circadian clock proteins are often time-of-day dependent, which is crucial for the function of the molecular circadian oscillator. While the recent years have witnessed the identification of an increasing number of clock proteins or modulators [15], [16], [21], [33], [34] a comprehensive analysis of PPIs within the circadian clockwork - in particular with respect to the timing of the PPIs - is still missing. Here, we identified 109 so far uncharacterized interactions within the circadian clockwork in yeast and have successfully validated a sub-fraction in mammalian cells. While our matrix screen design allowed us to perform independent replica experiments thereby reducing the risk of false positives and false negatives, it is clear that due to the obvious limitations of the Y2H system [35] our network is likely still far away from saturation. For example, interactions that depend on posttranslational modifications or on more than two proteins are difficult to detect in Y2H assays. Nevertheless, our screen showed a rather high sensitivity (∼40% recovery of previously reported PPIs) compared to other Y2H reports or other PPI interaction methods [11]. In addition, we estimate to have only a low false-positive rate, since we could validate 86% of all CLOCK and BMAL1 interactions in mammalian cells.
Interestingly, many of the new interactions occurred between core clock components and regulatory components such as kinases e.g., CSNK2β, phosphatases (e.g., PPP2, PPP1), and F-box proteins (e.g. FBXW11). Hence, our data should be a valuable resource for studying molecular events within the circadian system with so far uncharacterized posttranslational mechanisms being especially interesting. Whereas phosphorylation of clock proteins are increasingly recognized as crucial for circadian dynamics, de-phosphorylation events have not been studied as extensively [36]. Therefore, we characterized the newly discovered interaction between PPP1Cα and the CLOCK/BMAL1 heterodimer. Indeed, we could validate our in silico prediction of the daytime-dependence of this PPI, which negatively regulates BMAL1 abundance (see Figure 7 and Figure S7), whereas others propose PER proteins as substrates of PPP1 [37], [38]. Further work is needed to identify the respective regulatory subunits that may mediate substrate specificity.
Our circadian PPI network is very densely connected (Figure 2) with a high clustering. How can such a network function? We analyzed both the predicted temporal organization, which separate PPIs in time as well as modular organization, which organize the network in functional complexes. To investigate temporal organization, we have integrated circadian expression profiles from mouse liver for the interacting pairs of proteins assuming that co-expression on transcript level can represent individual protein abundance probably as one limiting factor for physical interaction. De Lichtenberg et al. have pioneered the analysis of dynamic protein–protein interactions with a specific focus on cell-cycle stages in yeast also integrating transcription data [39] and Atwood et al. predicted the interaction time of circadian co- and antiphasic expressed proteins [40]. However, our analysis is not restricted to a specific process or specific circadian phases, but provides a systems-wide view of circadian PPI dynamics.
Our transcript-based analysis led to the construction of a dynamic circadian (albeit only liver-specific) PPI network, in which PPIs are formed at all circadian phases (see Figure 4B). Obviously, our analysis harbors several limitations, since PPIs in vivo depend on a variety of factors such as spatial restrains, restriction to specific tissues, relative protein abundance, mRNA processing, stoichiometry and interaction kinetics, complex formation and posttranslational modifications. All these parameters are not represented by the corresponding mRNA profiles of interaction partners. However, our assumption that indeed dynamic binding events can be approximated by such an approach is supported by (i) our finding that co-expression of transcripts at similar circadian phases more often occurs among interacting proteins (see Figure 4B and Figure S6B), (ii) known interaction dynamics between components of the circadian system can be reproduced (see Figure 4B), e.g. the circadian phase-specific CLOCK/CRY1 interaction [5], and (iii) the in silico predicted time-of-day dependent interaction between PPP1Cα and CLOCK/BMAL1 could be validated with endogenous liver components. Nevertheless, it should be noted, that on a systems-wide scale it is still largely unknown, whether and to which extent genes with rhythmically expressed transcripts also display circadian protein levels. While recent comparisons between transcript levels and protein levels have shown a rather good correlation [41], [42], our circadian PPI network should still be considered as a prediction.
Dynamic ‘hubs’ (proteins predicted in many rhythmic PPIs) seem to be especially important for circadian rhythms (see Figure 5C) as revealed by our genetic perturbation analyses. Thus, apparently not the absolute number of interactions is crucial for the importance of a clock protein but the degree of dynamic PPIs. This may be not too surprising, since precisely timed interactions between activators and their repressors is the fundamental principle of the circadian negative feedback mechanism. Interestingly, this principle may be translated to a global scale: we find that proteins with a rhythmic transcript have significantly more interaction partners than non-rhythmic proteins (p<10−10, Wilcoxon Rank test). In addition, proteins that qualify as regulatory components (as defined by their GO category ‘regulation of biological process’) have significantly more interaction partners than non-regulatory proteins (p<10−15, Wilcoxon Rank test; see also Text S1). Together, this suggests that rhythmic control of PPIs is an essential feature of biological networks. While such analyses are only of correlative nature, it would be interesting in future studies to analyze directly whether a particular PPI or the rhythmicity of a particular PPI is required for normal rhythms. To this end, however, novel (perhaps pharmacological) tools are needed to specifically disrupt the PPI without interfering with the abundance or other PPIs the component might execute.
In the last decade transcriptome analysis were successfully used to study circadian dynamics on a systems-wide level [6], [7] with mRNA rhythms serving as indicators for output control. Corresponding comprehensive studies on the level of the proteome are still largely missing. To get novel insights into the time-of-day dependent regulation of cellular processes we propose a new strategy to predict circadian regulation at the level of protein complexes rather than looking at mRNA profiles of individual components. Based on this dynamic interactome we have constructed a ‘process network’ with many processes (represented by corresponding GO terms) strongly connected by predicted dynamic PPIs (see Figure 6D and Figure S6C). While this concept has obvious limitations (ambiguous GO assignments, predictive nature of rhythmic PPIs, etc.) it allows a first, systems-wide glance on how cellular processes might be regulated in a time-of-day specific manner beyond circadian transcription. Future studies are needed to investigate to what extent and on what mechanistic bases rhythmic PPIs contribute to the dynamic modulation of cellular processes.
Overall, we propose a global view on the circadian control of protein–protein interactions important not only for the circadian oscillator but also for the temporal orchestration of many essential cellular processes.
Matrix-based Y2H interaction analyses were performed essentially as described [10], [43]. For the generation of the Y2H matrix 46 full-length entry constructs were shuttled into Y2H vectors resulting in LexA DNA binding domain fusions (bait configuration) and Gal4 transcription activation domain hybrids (prey configuration). The L40ccαMATα yeast strain was transformed with prey constructs while baits were introduced into a MATa strain carrying HIS3, URA3, and lacZ as reporter genes. All constructs were tested for auto-activation properties. For mating, liquid cultures of the MATa strain were mixed with prey colonies in 384-micro titer plates and mixtures were then spotted onto yeast complete medium agar plates. After mating at 30°C, colonies were transferred into 348 well plates containing SDII liquid (-Leu, -Trp) selective medium and then transferred to SDII agar for selection of diploid yeast (at 30°C). Diploid yeasts were spotted on solid selective SDIV agar plates (-Leu, -Trp, -Ura, -His) as well as on nylon membranes placed on SDIV agar plates. X-Gal assays were performed with the colonies that grew on membranes as described.
HEK293 cells were lentivirally transduced with Clock- or Bmal1-luciferase constructs. Cells stably expressing luciferase hybrids were transfected with MYC-tagged putative interactors. After 48 hours, lysate containing one million luciferase counts was subjected to immunoprecipitation. Pull-downs were performed with an anti-MYC or an isoform specific ideotypic antibody and agarose beads after overnight incubation. After three washes luciferase activity of pulled-down complexes was measured.
Western blot analysis was performed essentially as described [15]. Briefly, proteins were denatured via boiling in SDS-loading buffer. Separation was performed by SDS-PAGE using 4%–12% Bis-Tris gels. Proteins were transferred to nitrocellulose membrane and incubated with primary antibodies. Membranes were probed with corresponding HRP-conjugated secondary antibodies. Chemiluminescence reaction was performed for protein visualization.
RNAi and overexpression constructs were lentivirally delivered as described [15]. Briefly, filtered medium containing virus particles was used for transduction of human U2OS cells carrying the Bmal1-promoter luciferase reporter [15] in the presence of protamine sulfate. Next day, medium was exchanged to puromycin or blasticidine containing medium. After positive selection cells were synchronized by a 30 min pulse of dexamethasone. Bioluminescence was monitored for ∼6 days in a TopCount luminometer with a sampling rate of 30 min. Time series were analyzed for circadian rhythmicity correlating them to a cosine function via the ChronoStar software [15].
HEK293 cells were transiently transfected with a firefly luciferase reporter (containing six E-box enhancer elements), CLOCK/BMAL1 and individually all discovered putative CLOCK and BMAL1 interactors (including their paralogs or functional subunits) and a renilla luciferase construct for normalization [4], [27]. Signals were detected with a dual-luciferase reporter assay in a luminometer plate reader. Experiments were repeated three times.
U2OS cells stably expressing a fluorescence reporter either with BMAL1 or CLOCK as EGFP fusion protein (see Figure 7D left; [30], [31]) were transduced with lentiviruses containing PPP1Cα or GSK3β expression constructs. Cell fluorescence was analyzed using flow cytometry (FACS Canto II). Red fluorescence of DsRed and green fluorescence of EGFP intensities of DsRed positive cells were detected. The protein stability index (PSI) is defined as the maximum of the distribution curve of the ratio between EGFP and DsRed intensities. Thus, a high PSI value corresponds to a high green fluorescence intensity, i.e. highly abundant (and likely stable) fusion protein.
Firstly, standardized 48-hour transcript liver profiles taken from Hughes et al., 2009 [6] were analyzed for 24 hour periodicity using Fourier analysis:where x is the standardized expression vector (mean(x) = 0; sd(x) = 1) for the gene, T is the period (in our case 24 h), and xi is the measured expression at time point ti. Statistical significance was calculated by comparison with randomly permutated time series using the Bioconductor cycle package [44]. Secondly, abundance AC of a complex C formed by two interacting proteins P1,2 is assumed to be proportional to the expression E of P1,2. Abundance AC(t) over time is approximated by the product of expression vectors EP1(t) * EP2(t), which was then associated to the corresponding PPI. As proxy for protein abundance, the transcript levels over time were utilized, thusStatistical significance of AC(t) rhythmicity was calculated using the Fourier-score and permutated time series as background model after standardization (i.e. mean (EP1 * EP2) = 0; sd (EP1 * EP2) = 1). A phase was assigned to a periodic interaction through shifting a cosine (with 24 h periodicity) along the time axis and measuring the overlap of the expression levels with the cosine curve. The time shift leading to a maximum overlap was considered as the phase α of the PPI and ranges from 0 to 24 h.
All PPIs of the compiled human interactome in the UniHI database (N = 45775) were assessed for possible dynamic behavior [12], [45], as described above. Human proteins were mapped to their mouse orthologs and periodicity of 30413 interactions was analyzed as described above resulting in the prediction of 2788 significantly (FDR<10−5) dynamic interactions.
The discovered PPIs are listed in the IMEx (http://www.imexconsortium.org) consortium through IntAct [pmid: 19850723] and assigned the identifier IM-16832.
|
10.1371/journal.pntd.0005121 | Seasonal and Spatial Environmental Influence on Opisthorchis viverrini Intermediate Hosts, Abundance, and Distribution: Insights on Transmission Dynamics and Sustainable Control | Opisthorchis viverrini (Ov) is a complex-life-cycle trematode affecting 10 million people in SEA (Southeast Asia). Human infection occurs when infected cyprinid fish are consumed raw or undercooked. Ov requires three hosts and presents two free-living parasitic stages. As a consequence Ov transmission and infection in intermediate and human hosts are strongly mediated by environmental factors and understanding how environmental variability influences intermediate host abundance is critical. The objectives of this study were 1) to document water parameters, intermediate hosts abundance and infection spatio-temporal variation, 2) to assess their causal relationships and identify windows of transmission risk.
Fish and snails were collected monthly for one year at 12 sites in Lawa Lake, an Ov-endemic region of Khon Kaen Province in Northeast Thailand. Physicochemical water parameters [pH, temperature (Tp), dissolved oxygen (DO), Salinity, electrical conductivity (EC), total dissolved solid (TDS), nitrite nitrogen (NO2-N), lead (Pb), total coliform bacteria (TCB) and fecal coliform bacteria (FCB)] were measured. Multivariate analyses, linear models and kriging were used to characterize water parameter variation and its influence on host abundance and infection prevalence. We found that sampling sites could be grouped in three clusters and discriminated along a nitrogen-salinity gradient where higher levels in the lake’s southern region predicted higher Bithynia relative abundance (P<0.05) and lower snail and fish species diversity (P<0.05). Highest Bithynia abundance occurred during rainy season (P<0.001), independently of site influence. Cyprinids were the most abundant fish family and higher cyprinid relative abundance was found in areas with higher Bithynia relative abundance (P<0.05). Ov infection in snails was anecdotal while Ov infection in fish was higher in the southern region (P<0.001) at sites showing high FCB.
Our results indicate that water contamination and waterways configuration can influence freshwater communities’ assemblages possibly creating ideal conditions for sustained transmission. Sustainable control may require a better appreciation of the system’s ecology with wise governance and development planning particularly in the current context of SEA agricultural intensification and landscape modification.
| Opisthorchis viverrini (Ov) is a fish-borne parasite infecting humans when they consume raw or undercooked fish. Ov is endemic in Southeast Asia, particularly in rural parts of Northeast Thailand, Lao PDR, Cambodia and Vietnam. The Ov lifecycle includes three different hosts: snails, fish and humans. Transmission takes place in various environments and water habitats where many factors interact and influence Ov survival and transmissibility from host to host. Our study investigated the influences of water parameters on snail and fish abundance and Ov infection rates. We found that salinity and nitrite-nitrogen were positively correlated to Bithynia snail relative abundance and inversely correlated to snail and fish species diversity. Cyprinids were the most abundant fish family and high cyprinid abundance was found in areas with high Bithynia abundance creating ideal conditions for transmission. Ov infection in fish was consistently detected from an area with high levels of salinity and nitrite-nitrogen and characterized with high abundance of Bithynia and cyprinids. Our findings suggest that water contamination can influence freshwater communities’ assemblages possibly creating ideal conditions for sustained parasite transmission. Sustainable Ov control may require a better appreciation of ecological systems together with wiser governance and development planning.
| Opisthorchis viverrini (Ov), the Southeast Asian liver fluke, is a fish-borne complex life cycle trematode endemic in Thailand, Lao PDR, Cambodia and southern parts of Vietnam where an under-estimate of 10 million people are reported to be at risk of Ov infection [1, 2]. While most infections are asymptomatic, heavy chronic infections are associated with clinical hepatobiliary complications such as cholangitis, advanced periductal fibrosis, hepatomegaly and in some rare cases cholangiocarcinoma, a bile duct cancer associated with very poor prognosis upon diagnosis [1, 3, 4]. The northeast region of Thailand is known in particular to be a hotspot of Ov endemicity which despite nationwide public health prevention campaigns led by the government and private organizations [5], still is plagued with high infection prevalence [3, 6]. The persistence of high infection rate in the region is likely due to its cultural and ecological particularities where wet rice agrarian habitats; centuries old raw food culture and the parasite complex biology combine to create an ideal transmission arena [3, 7].
The parasite complex lifecycle begins when Ov eggs are released in the environment through the feces of a definitive human host or reservoir host, which are mostly cats and dogs [8]. Upon reaching freshwater habitats, Ov eggs are eventually ingested by freshwater snails belonging to the family Bithyniidae. Within the snail, Ov eggs hatch and the emerging miracidia develop to become sporocysts, which undergo asexual multiplication. The sporocysts develop to rediae and then finally to their free-swimming cercaria stage that will be released in the environment. Thousands of cercariae can be released as free-swimming parasites into the aquatic environment where they actively search for certain species of freshwater fish of the Cyprinidae family, the second intermediate host. Upon contact with the fish, cercariae encyst within the fish body and develop into infective metacercariae. Ov metacercariae can infect humans when the fish that contained them are consumed raw or not cooked sufficiently to alter the parasite’s infectious potential [9].
Research conducted in the fields of immunology and pathology has greatly improved our ability to punctually diagnose, treat and respond to Ov and other liver fluke infections [10–12]. However, there is a lack of robust understanding of the ecological and environmental determinants of Ov transmission and therefore strong limitations remain regarding the sustainable interruption of the transmission and effective control. The inherent complexity of the Ov lifecycle, including the need of three taxonomically different hosts with markedly different ecologies, provides ample opportunity for environmental modification at different spatial and temporal scales to modulate patterns of transmission [13].
Biotic factors such as toxic exudates produced by hosts, non-hosts, predators and decoy organisms may act simultaneously and in conjunction with abiotic factors to expose free-living endohelminth stages to a complex array of hazards on their way to the down-stream host [14]. Similarly water-related environmental parameters can strongly influence host physiological status [15], demography [16] and distribution and modulate patterns of host-parasite encounter, hence transmission dynamics and infection likelihood. For example natural environmental variables such as temperature, salinity and pH have strong species- and stage-specific effects on survival rates [17]. In the case of Ov, infection likelihood in snails has been recently shown to be temperature-dependent [18]. Environmental disturbances, which can affect freshwater snail community structure, including species diversity and relative abundance [19, 20] contribute to modulate parasite transmission. For example, more species-diverse snail communities cause a 25–50% reduction in infection among Schistosoma mansoni snail hosts (Biomphalaria glabrata) and infected snails raised alongside non-host snails (Lymnaea or Helisoma sp.) also produce 60–80% fewer cercaria, suggesting that diverse snail communities could reduce human infection risk and that environmental change impacting host ecological functioning are important overarching determinants that modulate transmission [21, 22].
In the context of the ongoing agricultural intensification, landscape modification, and livelihood changes in rural SEA [23], wetland water contamination is increasing, natural biogeochemical cycles are disrupted, natural and human hosts demography are remodeled and as a consequence infectious disease risk in general and Ov incidence in particular fluctuates spatially and temporally. A dynamic view acknowledging the spatial and temporal interactions between environmental variability, host distribution and abundance and their consequence on transmission and infection at multiple scales is thus critical to improve our ability to identify pathogenic landscapes and refine our intervention strategies [24]. This rationale is particularly relevant in the case of Ov, which transmission interruption implies understanding cultural behaviors, socio-economic shifts and environmental particularities [25, 26].
The objective of this study was two-fold: 1) to document water parameters, intermediate hosts abundance and Ov infection spatial and seasonal variation, 2) to assess their causal relationships and identify windows of transmission risk in an Ov endemic ecosystem characterized by high rate of environmental and livelihood changes.
Khon Kaen Province, in Northeast Thailand, has been acknowledged as an area of high endemicity and ongoing Ov infection with certain wetland areas characterized by particularly high infection prevalence. For instance the Kaeng Lawa Reservoir, more commonly referred to as Lawa Lake, located in Ban Phai District, is known for its high and persistent endemicity of Ov infection [6]. Conversations with local people in the area have indicated that communities are strongly relying on the lake and its tributaries for subsistence. As a result fish preparations, including raw/fermented/undercooked fish dishes, are particularly frequently consumed and appreciated. Rice cultivation is also a dominant activity in the area, which not only provides an ideal habitat for snails, the first intermediate host, but also fosters the prolonged presence of local farmers in the environment and a high likelihood for open defecation, hence sustained parasite transmission. As a consequence Lawa Lake is a suspected important focal point of transmission, but known local differences in lake geomorphology and seasonal variability in water movements may imply heterogeneous transmission risk among lakeshore localities [27, 28]. Khon Kaen Province has a tropical monsoon climate with three major seasons: hot-dry (March-June), hot-rainy (July-October) and cool-dry (November-February) seasons. The rainy season comprises severe weather events such as heavy rainfall, floods and sometimes even drought that can strongly influence humans and wildlife through seasonal habitats modification. For instance floods in Lawa Lake typically occur during the months of September and October. During this time the water level of the Chi River will rise so high that some areas on the west side of the lake such as in Ban Chikokkor (Fig 1), will be completely submerged forcing local communities to use boats to commute from place to place. These important seasonal variations in water level are likely to strongly influence Ov eggs movement from latrines to the environment and snail and fish distribution and abundance, hence Ov transmission dynamics. In 2013, during which our research was implemented, temperatures in Thailand were higher than usual from January to November and lower than usual in December. Thailand was hit by a tropical storm in October that was downgraded from typhoon Wutip [29]. These storms also brought unusual rainfall in November and December and contributed to the lower than normal temperatures that occurred in Northeast Thailand [29].
Twelve sampling sites within the main body and near the shore of Lawa Lake were selected for the systematic collection of Bithynia snails and cyprinid fish and to perform a water quality assessment over a 1-year study period. The number and location of sampling sites were chosen to provide a representation of the main aquatic habitats found in and around the lake and therefore to account for their influence on water quality variability. Among the 12 sites, 8 were located 5 meters (m) from the shore and categorized as ‘nearshore sites’ and described based on their immediate surroundings to various types of rich and aquatic vegetation. This included: ‘shallow grassland,’ indicating a water depth ≤2 m; and ‘deeper grassland’ indicating a water depth >3 m (see Table 1). ‘Nearshore sites’ are exposed to desiccation and strong water quality fluctuation. An additional four ‘offshore sites’ were selected to account for the deepest areas of Lawa Lake. The ‘offshore sites’ contrast with ‘nearshore sites’ in their tendency to remain submerged even during the hot dry season creating refuges for fish during the summer months. Nearshore sites in particular were chosen to be in close proximity to villages for which Ov infection prevalence data from the Tropical Disease Research Laboratory, Khon Kaen University [6] is available. The investigation of the snail and fish hosts’ seasonal abundance and infection trends has been intended to improve our understanding of Ov prevalence and infection intensity variation in their neighboring local communities.
Samples were collected on the third week of every month from February 2013 to January 2014. This was to create consistency and allow a week of time each month as a safety-net-period in case samples could not be collected due to outside circumstances. Data collected at the 12 sites included 10 water parameters as well as cyprinid fish. Snails were collected from 8 out of the 12 sites since they are not present in water depths exceeding 3 meters [30]. A portable Global Positional System (GPS) unit was used to record the latitude and longitude coordinates at each sampling location.
Five physicochemical water parameters were measured in situ 30 cm below the surface of the water from a steady boat to reduce sediment disturbance using a portable water meter (Extech Oyster DO 0700 Meter) that was calibrated according to instrument guidelines. The parameters measured by the Extech meter included temperature, pH, electrical conductivity (EC), total dissolved solid (TDS) and salinity. Additionally, water samples were collected in sterilized 500 mL sized polyethylene and 100 mL glass bottles and transported on ice to the Environmental Laboratory of Regional Environment Office 10 Khon Kaen for analysis. The parameters measured there were dissolved oxygen (DO), nitrite nitrogen (NO2-N), lead (Pb), total coliform bacteria (TCB) and fecal coliform bacteria (FCB). Levels of DO were measured using the Azide Modification Method; Pb, by the Nitric digestion method, GFAAS; NO2-N, by the colorimetric method; and TCB and FCB were both tested using the most probable number (MPN) technique per 100 mL sample of water.
Snails were sampled for a time period of 10–15 minutes [30], at one time and place in each of the 8 sampling sites [28, 31]. The snail collecting techniques used for this study included hand picking and scoop technique [30]. All freshwater snails in a 1-square meter quadrant were collected, placed into plastic bags and labeled by study site and time period. Snails were brought back to the Tropical Disease Research Laboratory, Khon Kaen University and identified based on their morphological characteristics [32, 33]. Particular attention was given to Bithynia siamensis goniomphalos (Bsg), the Bithynia subspecies endemic to the northeastern region of Thailand [33] and host for Ov. Bithynia snails were distributed into plastic containers with a maximum of 5 snails per cup containing 5 mL of dechlorinated water. These containers were placed under 25 Watt light bulbs for 2 hours for cercarial shedding. The cercariae shed from Bsg snails were observed under a stereomicroscope and identified based on their morphological characteristics and distinctive movement as described by Schell [34] and Kaewkes [35]. The Ov cercaria body and tail are 154 x 75 μm and 392 x 26 μm, respectively; they are covered in a brownish colored pigment and have two eyespots, oral sucker, pharynx and undeveloped ventral sucker [35]. When observed through a microscope, their morphology combined with their spinning, jerking, floating then sinking motion make them easily identifiable to trained parasitologists and field technicians. In the case of an infected snail cup, snails were separated further into individual cups to identify the infected snail. Once the infected snails were identified a 50 μL drop of water from the cups holding infected individuals was placed on a glass slide and the number of cercariae in that drop was counted allowing for a quantification of cercaria infection intensity.
The gillnet fishing technique was utilized for this study since it is the fishing method employed by northeastern Lao-Thai fishermen for catching cyprinid fish. Twelve gillnets were custom prepared with a net mesh size of 20 mm, a net length of 5 m, and a net width of 1 m. Gillnets were dropped in all 12 sampling sites between 1:00 p.m. to 4:00 p.m. and picked up the next morning from 6:00 a.m. to 9:00 a.m. All freshwater fish caught in the nets were put into plastic bags, labeled, packed in iceboxes and transported back to the Tropical Disease Research Laboratory, Khon Kaen University. The iceboxes were put in a walk-in cool room (4°C). Approximately 3 days after collection, fish were measured, weighed, and identified according to their morphological characteristics using FishBase [36] and confirmed by experts from the Khon Kaen Department of Fisheries. The fish samples were labeled, separated and digested one-by-one using the Pepsin-HCl Digestion Method, described in Sohn [37]. Ov metacercariae were identified morphologically under stereoscopic microscope [35, 37]. Ov metacercariae are identifiable by a double-layered thick cyst wall, oval shape and approximate size of 201 x 167 μm in its encysted stage [35]. At this stage and under room temperature, it is possible to see them moving vigorously with a discernible black excretory bladder, oral sucker and ventral sucker.
There was no significant difference in Bsg snail abundance by cluster groups (Table 4). However, Bsg snail abundance showed statistically significant differences by season, with the majority of Bsg snails collected during the rainy season (P<0.001; Table 4). The abundance of Bsg snails was highest in October and overall higher in sites located in the southern region of the lake, particularly in site 11, where salinity and NO2-N both reported their highest measurements all year. High Bsg relative abundance was positively associated with high salinity levels (P<0.05, Fig 4A), particularly in site 12 and during February, the end of the cool season. Higher levels of NO2-N were positively associated with high Bsg relative abundance (P = 0.004 and P = 0.1, Fig 4B) particularly in site 11 and during the month of April, the country’s hottest month of the year. Snail species diversity showed statistically significant negative association with salinity (P = 0.02) but not with site cluster or seasons.
Cyprinid fish were the most abundant family of freshwater fish and were ubiquitously distributed. Cyprinid relative abundance presented statistically significant differences by season (P<0.001) with the greatest number of fish being caught during cool season months. Additionally, there were statistically significant differences of cyprinid abundance in the three clusters, (P = 0.005; Table 4, Fig 5). The relative abundance of cyprinid fish was greatest in the deep water zone of cluster 2 characterized by strong variations of pH and DO, and the richer-vegetated-southern-region zone of cluster group 3, where Bsg abundance, as well as salinity, nitrogen and fecal matter were also at their highest. Fish species diversity was significantly greater in cluster 2 (Table 4) and negatively influenced by salinity and NO2-N (both P < 0.01).
Although Ov cercaria infection in Bsg snails was nearly undetectable, with only one Ov infected snail retrieved during the study period, other trematode infections in Bsg snails were detected from 7 out of 8 snail sampling sites. The trematodes infecting Bsg snails from Lawa Lake included: Ov, amphistome, xiphidiocercariae, parapleurolophocercous, and furcocercous cercariae morphotypes. Ov cercaria is the only form that has been specifically identified and named according to its associated trematode species, Opisthorchis viverrini. The richness of trematode morphotypes infecting Bsg snails as well as the likelihood of infection was highest in sampling site 11 (Table 5), cluster 1 near-shore zone, where we observed the greatest abundance of snails and a significant positive relationship with salinity and NO2-N (P<0.05; Fig 4C). Despite a singular peak at 0.25% in site 1 during the month of March, Bsg trematode infection rate was overall low throughout the year generally ranging from 0.04%–0.09% and consequently seasonality did not significantly influence Bsg trematode infection prevalence (P>0.05; Table 4).
Fish Ov infection prevalence was significantly higher in cluster 3 (Table 4 and 5), the southern region zone that includes sites 7, 8, and 9, exhibiting high fecal matter, salinity and nitrogen pollution, mainly coming from livestock and agriculture. The locations where Ov infected fish were collected are close to cluster 1; site 11 in which we found a statistically significant effect of Bithynia snail abundance on fish infection rates (P<0.05). Ov infection in cyprinid fish was not detected in any northern region sampling sites (Fig 5C).
In our study we investigated spatial and seasonal variation in water quality parameters and their influence on Ov hosts abundance and infection prevalence. Our objective was to provide novel environmental data to foster a refined understanding of Ov transmission dynamics in a highly endemic area and to highlight that knowledge of the ecology of the system needs to be incorporated in interventions planning, including landscape management and local government economic decisions.
We observed environmental differences and similarities between the sampling sites mostly in relation to salinity; fecal matter; pH and dissolved oxygen; lead; and nitrogen components, as suggested by the CA and PCA groupings. The observed differences in physicochemical parameters were mostly expressed along a north-south gradient and further discriminated across bathymetric zones. Among the most significant water parameters, we observed higher overall concentrations of salinity and nitrite-nitrogen in sites located in the nearshore and southern region sites, respectively. Two mutually reinforcing sets of determinants thus can potentially explain the patterns of water parameter variation observed: (1) lake hydrological and geomorphological characteristics and (2) human activity varying mostly along the north-south gradient in the region.
The sites contained in the deep-water zone, are located on higher elevations where the lake is wide, open, and deep with moderately strong water flow. Lawa Lake is a dammed reservoir that serves multiple villages for a number of predominantly agrarian services. The dam opening located near Chikokkor Village (cluster 1) facilitates the sustained accumulation of large volumes of water in this area. The deep-water zone (cluster 2) showed typical levels of salinity and nitrite-nitrogen as found in wetland or inland lake ecosystems. These levels were markedly different from those of clusters 1 and 3 where extremely high levels of salinity and nitrogen were measured. The extreme levels of salinity and nitrogen in this area can be explained by differences in water depth associated with marked water stratification, deeper water dilution and homogeneity in contaminant dissemination [57]. Lower elevation sites from cluster 1 and 3, in the southern region zone, where the lake is narrow and shallow, are more severely affected by drought and are characterized by slower water flow, which implies greater water compartmentalization and site-specific biochemical profiles. Coupled with greater intensive agricultural pressure and other human influence, sites from cluster 1 and 3 are more susceptible to contaminant, nutrient and dissolved solid accumulation [57, 58]. The southern region zone also presents a greater concentration of rice paddy fields and a higher overall agrarian pressure, which combined with a lack of human and animal waste monitoring, the unregulated use of agrochemicals and pesticides [59, 60], and general poor irrigation and domestic practices, create a highly contaminated environment, contributing to what others have called a ‘pathogenic landscape’ [13]. In particular, the rural community of Ban Pao, near site 11, has intensified its agricultural practices, mainly rice cultivation and livestock farming [6], with the use of high amounts of agrochemicals and the manipulation of water flow through man-made canals and streams connected to Lawa Lake. The Huoi-jit canal flows through Ban Pao and Ban Nonlamon villages into Lawa Lake and carries with it the waste from the urban settlements of nearby cities such as Ban Phai, an urban area of approximately 20,000 people. Among all sites, site 11 presents particular hydrologic and ecological characteristics, being not only closer to a developed urban area but also more riverine than other sites. These characteristics may have contributed to the unique water quality profile as well as snail and fish community structure in this area as discussed below. Similarly, a plot of land along the shore of Lawa Lake near cluster 3 (sites 7, 8, and 9) is used for the raising of water buffalo, which together with cows, ducks, chickens, pigs and dogs serve as major sources of fecal contamination in the area. Another major source of fecal contamination in the area are humans, who continue to practice open defecation in these aquaculture/rice paddies dominated landscapes, thus contributing to increased fecal coliform contamination and the greater likelihood of snail and fish infection and ultimately parasite transmission [61].
In the specific case of salinity, both historical contingency and current agricultural activities can explain the high levels observed in the southern region zone of the lake. Khon Kaen Province and the Korat Basin, for instance are known to exhibit soils containing high amount of rock salt as a result of pre-historic salt sedimentation in the area [30]. Adding to the inherently high salt content of regional soils, higher agricultural pressure and associated unmanaged irrigation practices in the lower region of the lake induce high rate of salt deposition as a result of intense water evaporation. For example, sugar cane, a cash crop highly prevalent in the area [62], needs about 20,000 m3/ha of water per year [63]. As a result, irrigated areas often receive more than 3,000 kg/ha of salt per year and some receive as much as 10,000 kg/ha/year [63].
Overall, we observed a high relative abundance of Bsg snails and cyprinid fish in Lawa Lake, which confirms their potentially important role in sustaining Ov transmission in the area. More specifically we observed spatiotemporal patterns in the abundance of both Bsg snails and cyprinid fish in cluster 1 and 3, respectively with sites located near the shore and in the southern region of the lake being more prone to Bsg and cyprinid fish presence, respectively. These patterns of distribution and abundance may reflect broad scale hydrological influence of the Lawa dam on freshwater snail communities as documented in several recent studies [64]. Dams can alter water flow, sediment flow, and the overall typology of downstream environments [65] causing major ecological alteration in the habitats of humans and aquatic organism diversity [66]. For example, Wu et al [67] looked at how differences in water level and water quality parameter fluctuation impacted the density of schistosoma-transmitting snails, Oncomelania hupensis, in Dongting Lake, located near the Three Gorges Dam. They found that low elevation sites saw slight increases of snail density after the completion of the dam, possibly in relation to changes in water temperature, hydrology, vegetation, soil and more indirectly salinity and nitrogen content [67].
Additionally, hydrology and flood plain morphology changes may foster more intensive agricultural practices [68] and other types of development that can lead to water contamination and elevated levels of salinity and nitrogen [69]. For instance, our results suggest that sampling sites located in the southern region of the lake, particularly in the low elevation site 11, which report high levels of salinity and nitrite-nitrogen, are characterized by higher relative abundances of Bsg snails and reduced specific diversity of freshwater snail communities. The proximity of site 11 to an urban canal and a developed area with particular hydro-ecological conditions further exemplify the the environment-degradation-pathogen-transmission link as the changes in ecological conditions in these disturbed environments can lead to an unwanted proliferation of disease vectors (sensus lato) such as Bsg. The apparent strong influence of these water contaminants on snail community structure and Bsg abundance patterns in Lawa Lake indicate species-specific differences in behavior or physiology in snail ability to cope with increased levels of salinity [70] and nitrogen [24]. Previous studies have indicated that while Bsg is not found in highly saline freshwater environments, the species prefers water with some saline content (0.05 and 22.11 parts per thousand) over low salinity freshwater [30]; providing further evidence of the possible mediating influence of salinity in freshwater snail communities in the region. Similarly, high levels of nitrogen impair the ability of aquatic animals, including snails and fish, to survive, grow and reproduce and ultimately disrupt freshwater ecological functioning [71]. While no studies have investigated formally the influence of nitrate-nitrogen on Bsg survival, elevated nitrate levels in water environments are a major cause of deleterious reproduction and behavior effects for the invasive aquatic snail, Potamopyrgus antipodarum [24].
The higher relative abundance of Bsg snails and the overall reduced species diversity in lower sites and the implication of high levels of nitrite-nitrogen and salinity in driving this pattern suggest that water contaminants may modulate local freshwater snail community structure (i.e. species richness and abundance) through environment-mediated and species-specific physiological tolerance processes [72]. While this line of interpretation needs further investigation, several recent studies confirm Bsg snails as a ‘weedy species’, characterized by high proliferation potential in disturbed/transient environments [73].
The influence of water contamination on cyprinid fish distribution, abundance and community structure is less remarkable than for snails, possibly because of the fish ability to move to or select for favorable habitat across larger distances. Nevertheless, we noticed higher relative abundance of cyprinid fish in sites located in the southern region of the lake particularly in the sites located close to the shore in densely vegetated areas. While further research is needed, this observation may also indicate that some cyprinid species may be better adapted to the ecological situations found in contaminated sites than non-cyprinid species, which may further explain their ubiquitous distribution as a dominant fish family in the area and more generally in SEA [74].
Our results suggest that environmental contamination strongly influences patterns of hosts' abundance in a way that sets the stage for increased transmission risks. Indeed we observed higher Ov infection prevalence in areas of the lake where high nitrite-nitrogen and salinity levels as well as higher Bsg snails and cyprinid fish abundance were found (Table 5). We also found higher Ov infection prevalence in fish in water that was contaminated by fecal coliforms, which from the parasite perspective is a medium through which eggs can encounter its first intermediate host. In our study, we used fecal coliform contamination as a proxy for assessing the likelihood of Ov presence in the environment. High concentrations of fecal indicators were detected from near shore sites where the boats of fishermen and local huts for shelter were observed. Open defecation is the most direct avenue for Ov egg release in the environment in rural areas of northeastern Thailand and Lao PDR where rice farmers spend the majority of their time in the rice fields especially at the onset of the rainy season. The combination of physical presence in the field imposed by the farming practice and low awareness of transmission risks is an incentive for open defecation and creates an ideal transmission arena [75]. While the origin (i.e. human or domestic/wild animals) of the fecal indicators measured in our study cannot be ascertained, it is known that farmers in both northeastern Thailand and Lao PDR tend to prefer open defecation as it can be done immediately in the field and also provides natural fertilizer, a sort of win-win situation, which suggest that human waste is likely to be found diluted in the local water [75]. Echoing [76] further research is needed to clarify the quantitative (how much) and qualitative (what type) relationships between fecal coliforms and egg presence in the environment and improve our ability to use fecal coliforms as a proxy for transmission risks in endemic wetlands.
Despite the high likelihood of Ov egg presence in the environment as suggested by high fecal coliform concentrations, and the high human infection prevalence in human communities, particularly around the lower region of the lake (Table 1), we did not find high and regular infection prevalence in Bsg snails. Most of the studies surveying Ov infection prevalence in snails in endemic areas either fail to find infection or document extremely low infection prevalence typically averaging 0.11% [74, 77]. We believe that the low prevalence usually documented in the literature and the near absence of snail infection reported in our study reflect either 1) the biological processes underlying Ov development in the snails and/or 2) the limitation in our ability to detect Ov infection in snails through shedding procedures. For instance, infected snails can shed hundreds of cercaria daily [78] and accordingly it is assumed that one infected snail is enough to trigger infection prevalence in fish. It is thus relatively easy to imagine how unlikely finding infected snails can be even in highly endemic areas. Additionally, the technical limitations of infection detectability through cercarial shedding are high [79]. Recent evidence indicates that the routinely used cercarial shedding detection method is greatly outcompeted by molecular detection methods, which identify trematode infection in snails with up to 60.1% higher efficacy. The authors of the study suggested that detection failure is most likely due to the immature and covert infections, which can result in delayed and therefore undetectable (at the time of observation) cercarial shedding [80]. Therefore, the near absence of detected infection in our study may not necessarily be indicative of low snail infection prevalence. While this argument is of speculative nature, year-round Ov infection evidence found in cyprinid fish from the same area and high chronic Ov infection and sustained transmission in neighboring human communities suggest that infection in snails is hidden and/or their prevalence’s lie below our detection capacity [74]. Adding further complexity, water parameters, such as salinity, are known to modulate greatly snail cercariae shedding patterns, cercariae encystment capacity and overall transmission potential [81] and is likely to play a role in increasing or reducing infection prevalence in fish and thus to blur the relationship between snail infection and fish infection prevalence. Further research is needed to clarify the influence of salinity as well as other water contaminants on snail shedding and cercariae transmission potential in laboratory conditions.
Our results indicate that in the southern region of the lake, which is characterized by narrow and complex landscape architecture as well as higher agrarian pressure, water contamination resulting in high nitrite-nitrogen and salinity levels are consistently higher than in other areas. We found that elevated nitrogen and salinity levels were associated with higher Bsg snail relative abundance, particularly during the rainy season, and that cyprinid fish relative abundance was higher in the southern part of the lake where vegetation is denser. While infection in snails was nearly undetectable, infection in fish was found year-round in the lower region of the lake where fecal coliform levels, snail and cyprinid relative abundance were the highest. Together, our findings suggest that transmission likelihood is higher in the southern region of the lake where high water contamination, likely associated with intensive agrarian practices and larger scale irrigation schemes affect freshwater communities’ structure and create a particularly “pathogenic landscape” [13]. Indeed, human communities around this part of the lake exhibit particularly high infection prevalence.
Our observations and analyses highlight the highly dynamic nature of Ov transmission, both spatially and temporally, and strongly identify it as an ecological process modulated by human and environmental factors [14, 82], for instance agriculture intensification which contributes to water contamination. As such, Ov transmission needs to be understood in its social-ecological context [83, 84], including not only natural wetland ecology, but also with an appreciation of the need to connect local landscape management strategies and cultural/agrarian practices/politics with regional and global economic decisions. This is particularly important considering the current context of agriculture intensification and livelihood shifts in northeastern Thailand where unregulated agrochemical use and large-scale irrigation systems disturb local wetlands ecological dynamics while seemingly improving regional and national capital.
Interdisciplinary research and transdisciplinary actions are thus needed to improve our understanding of Ov transmission dynamics and to assess carefully the direct benefits and potential direct and indirect losses (e.g. benefits of irrigation versus transmission risk) when managing wetlands (e.g. irrigation projects, agriculture intensification planning, etc.) and, in some instances, to reach compromises and agreed tradeoffs between services and beneficiaries at different administrative scales. This holistic approach can ensure that future interventions designed to reduce Ov risk incorporate adaptive management strategies utilizing decision-making and ecological risk assessment tools [15], to be more sustainable and better aligned to improve the governance of health systems [85].
|
10.1371/journal.pgen.1008176 | CREB-B acts as a key mediator of NPF/NO pathway involved in phase-related locomotor plasticity in locusts | Gene expression changes in neural systems are essential for environment-induced behavioral plasticity in animals; however, neuronal signaling pathways mediating the effect of external stimuli on transcriptional changes are largely unknown. Recently, we have demonstrated that the neuropeptide F (NPF)/nitric oxide (NO) signaling pathway plays a regulatory role in phase-related locomotor plasticity in the migratory locust, Locusta migratoria. Here, we report that a conserved transcription factor, cAMP response element-binding protein B (CREB-B), is a key mediator involved in the signaling pathway from NPF2 to NOS in the migratory locust, triggering locomotor activity shift between solitarious and gregarious phases. We find that CREB-B directly activates brain NOS expression by interacting with NOS promoter region. The phosphorylation at serine 110 site of CREB-B dynamically changes in response to population density variation and is negatively controlled by NPF2. The involvement of CREB-B in NPF2-regulated locomotor plasticity is further validated by RNAi experiment and behavioral assay. Furthermore, we reveal that protein kinase A mediates the regulatory effects of NPF2 on CREB-B phosphorylation and NOS transcription. These findings highlight a precise signal cascade underlying environment-induced behavioral plasticity.
| The migratory locust, Locusta migratoria, is a worldwide agricultural pest that displays a remarkable density-dependent phase polyphenism, where being kept in a crowd triggers individuals to transit from the sedentary solitarious phase into the high-active gregarious phase. So, the migratory locust has been regarded as an excellent study model for environment-induced behavioral plasticity. Our previous finding shows that NPF2-regulated NOS transcription plays important roles in phase-related locomotor plasticity in the locust. Here, we further demonstrate that phosphorylated CREB-B directly activates NOS transcription in the pars intercerebralis, thus mediates phase-related locomotor plasticity. Further studies show that the levels of CREB-B phosphorylation is positively correlated with the crowding treatment and suppressed by NPF2. Among several candidate kinases, protein kinase A is demonstrated to transmit the inhibitory effects of NPF2 on CREB-B phosphorylation and NOS transcription. Our study provides deep insight into the precise regulatory mechanisms underlying environment-induced behavioral plasticity.
| Animals can adjust to a changing environment by developing alternative behavioral phenotypes that improve their fitness; this phenomenon is known as “behavioral plasticity” [1–3]. Environmental stimuli acts directly on the nervous system and induces short-term changes in neural and endocrine activity, or long-term changes in gene expression, thus lead to behavioral alterations at different time scales [4, 5]. Various neural modulators with distinct profiles of molecular action are involved in this process [6–8]. In a given context, long-term behavioral plasticity is greatly shaped by transcriptional changes in key genes that are governed intricately by the interactions of neural modulators in the brain [9–11].
Transcription factors (TFs) play central roles in the regulation of behavioral plasticity through integrating neural signals and downstream transcriptional events [8, 12, 13]. Nuclear TFs primarily modify behavioral performance by binding to the regulatory region (e.g., promoter or enhancer) of their target genes [14, 15]. With distinct stimulation, TFs undergo either expression alteration or protein modification changes that affect their subcellular location, binding activity, or stability; and then result in transcriptional changes of downstream behavior-related genes [16–18]. For example, Fos family proteins can be induced rapidly or transiently in specific regions (such as nucleus accumbens and dorsal striatum) by drug abuse, thus influencing rewarding and locomotor behaviors [19]. Nuclear factor-κB (NF-κB) family members are essential for hippocampus-dependent long-term memory formation by regulating memory-associated genes [20]. Therefore, uncovering the precise signaling cascade by which TFs respond to upstream signal and regulate downstream gene transcription will provide insights into the regulatory mechanism underlying environment-induced behavioral plasticity.
The migratory locust, Locusta migratoria, displays phase-related behavioral plasticity in response to population density variation [21, 22]. Gregarious locusts are highly active and attracted to their conspecifics, whereas solitarious locusts are cryptic and repelled from other individuals. The attraction index and locomotion activity substantially decrease during the isolation of gregarious locusts but are promoted by forced crowding in solitarious locusts [23]. During the time-course processes of solitarization and gregarization, gene expression profiles in the locus brain display dynamic changes [24]. The regulatory roles of several key genes have been revealed in behavioral phase transition [23, 25]. In particular, we have recently demonstrated that a neural neuropeptide F (NPF)/nitric oxide (NO) signaling pathway plays an essential role in phase-related locomotor plasticity, which results from the sequential changes in the phosphorylation and transcriptional states of nitric oxide synthase (NOS) as regulated by NPF1a and NPF2 systems, respectively [26]. NPF1a-regulated NOS phosphorylation initiates an immediate change in locomotor activity, whereas NPF2-regulated NOS transcription is responsible for long-term locomotor plasticity. However, the mechanism underlying NOS transcription changes as regulated by the up-stream TFs in response to population density variation is still unknown.
In this study, we aim to identify the TF activating NOS transcription in response to population density change. Our results uncovered several novel components of the NPF/NO signaling cascade underlying phase-related locomotor plasticity in locusts.
Our previous work showed that the mRNA level of NOS robustly responded to the changes in population density [26]. To further uncover the regulatory mechanisms of NOS transcription, we first identified a ~1.5 kb genomic region located upstream of NOS coding sequence by genome walking based on the whole genome sequence of the migratory locust. The transcriptional activity of DNA constructs carrying the ~1.5 kb upstream region of NOS fused with luciferase gene cassette was stronger than that of the empty pGL4.1 vector (NC) in HEK293T cells (>10 fold, S1 Fig). By progressively truncating the upstream genome sequence of NOS, we found that a genomic region from -150 bp to -121 bp upstream of the ATG start codon may serve as the core promoter sequence of NOS (S1 Fig).
To identify candidate TFs responsible for NOS transactivation, we predicted cis-response elements (CREs) in the promoter/enhancer sequence of NOS gene by using two different software, MatInspector program and TANSFAC program [27, 28]. Only CREB CREs were confirmed in the regulatory region of NOS gene according to the prediction of two independent programs (Fig 1A). In the locust genome sequence, we identified three CREB family members named as CREB-A, CREB-B, and CREB3 according to their phylogenetic relationship (S2 Fig). Extremely low sequence identity was found among these three CREB proteins (13.55%, S3 Fig). We examined which one of these CREBs can enhance the transcriptional activity of NOS gene by using dual-luciferase reporter assay. Over-expression of CREB-A and CREB-B protein (fused with flag tag) can strongly increase the transcriptional activity of NOS promoter in HEK293T cells, even after the promoter region of NOS was reduced to -110 bp and contains only the fourth CREB CRE (CREB R4, -98 to -87 bp upstream of NOS translational start site, Fig 1C and S4 Fig). Whereas over-expression of CREB3 only enhanced the transcriptional activity of DNA constructs containing CREB R3 and CREB R4 in HEK293T cells.
The potential regulatory effects of these three CREBs on NOS expression in vivo were validated through RNA interference experiments. Only the knockdown of CREB-B significantly suppressed the NOS expression in the brains of gregarious locusts (Fig 2A and S5 Fig) and prevented the up-regulation of NOS transcription when solitary locusts were crowded (Fig 2B). However, the knockdowns of CREB-A or CREB3 did not affect NOS transcription in either gregarious or solitary locusts (Fig 2A and 2B). These results suggested the CREB-B is required for NOS transcription in the locust brain.
We performed electrophoretic mobility shift assay (EMSA) to identify distinct CREB-B binding sites from the four putative CREB CREs (CREB R1, CREB R2, CREB R3, and CREB R4) in the NOS promoter region. Only CREB R4 probe could bind with the nuclear proteins isolated from brain tissues (Fig 2C and S6 Fig). The shift band disappeared after the cold probe (100 × over the labeled probe) was added (Fig 2C, lane 3) but was not affected by the mutant probe, in which five nucleotides were replaced (Fig 2C, lane 4). A super shift assay was further performed by using three different antibodies recognizing locust CREB-B (S7 Fig). The shift band receded, and a super-shifted band appeared after the CREB-B antibodies were added (Fig 2D, lane 3–5). By contrast, no super-shifted band was found in the control incubated with normal rabbit IgG (Fig 2D, lane 6), indicating that the nuclear protein binding CREB R4 should be CREB-B protein.
Our previous study showed that NOS is extensively expressed in the neurons of the pars intercerebralis (PI) [26]. Here, by using double immunofluorescence staining, we found that CREB-B was also mainly localized in the nuclei of the neurons expressing NOS in PI (Fig 3 and S8 Fig). These results indicated that CREB-B may serve as a direct modulator of NOS expression in the locust brain.
To verify whether or not CREB-B is involved in the behavioral phase transition in the locust, we conducted behavioral assay after the knockdown of CREB-B gene in the locusts. The total distance moved (TDM) and total duration of movement (TDMV) were robustly reduced in the gregarious locusts injected with dsCREB-B (Fig 4A and 4B). Meanwhile, the knockdown of CREB-B also inhibited the enhancement of TDM and TDMV in the solitary locusts upon 32 h crowding, (Fig 4C and 4D). However, neither TDM nor TDMV was affected by CREB-B gene silencing in solitarious locusts (S9 Fig).
We found that there were no significant difference of the mRNA levels of CREB-B in brain tissues between gregarious and solitary locusts and between isolation and crowding (Fig 4E, 4F and 4G). A number of documents have suggested that CREB protein phosphorylation is a conserved and critical regulatory mechanism for transcriptional activation [29, 30]. Thus, we confirmed whether or not CREB-B could be phosphorylated in locust brain tissues. The phosphorylation site of human CREB1 protein and locust CREB-B share high identity in their phosphorylated kinase-inducible (KID) domains (S7A Fig). We found that only the antibody against serine133 of human CREB1 (corresponding to Ser110 of the locust CREB-B) can detect positive band with the predicted molecular weight of the locust CREB-B protein (S7B–S7D Fig). The phosphorylation of CREB-B (p-CREB-B) was further validated by RNAi experiments, in which the band intensity (detected by anti-p(S133-CREB1) was significant reduced after the knockdown of CREB-B gene (S7E Fig). Moreover, the level of p-CREB-B was higher in gregarious locust brains (Fig 4H) and dynamically changed during the time-course phase transition, in which p-CREB-B level remarkably decreased after 4 h isolation but increased after 4 h crowding (Fig 4I and 4J and S10 Fig). These results together with our previous findings suggested that CREB-B plays a regulatory role in phase-related locomotor activity.
Considering that neuropeptide F, NPF2, modulates phase-related locomotor activity by suppressing NOS transcription [26], we determine whether or not CREB-B could mediate the regulatory effect of NPF2 on NOS transcription. We found that the p-CREB-B level strongly decreased after the injection of NPF2 peptide into gregarious locusts but was elevated after the knockdown of NPF2 gene in solitary locusts (Fig 5B and 5D and S11 Fig). By contrast, the treatment with another neuropeptide F, NPF1a (either the full length or the truncated peptide) did not affect the levels of p-CREB-B (Fig 5A and 5C and S12 Fig). RNAi-mediated CREB-B knockdown completely blocked the up-regulation in both TDM and TDMV caused by NPF2 knockdown upon the crowding treatment (Fig 5E and 5F). Furthermore, RNAi-mediated CREB-B knockdown also inhibited the increase of p-CREB-B levels and NOS transcription (Fig 5G and 5H), though the knockdown of NPF2 did not affect CREB-B transcription (S13 Fig). These results revealed that CREB-B is an essential mediator of phase-related locomotion under the control of NPF2 signal at the phosphorylation level.
We predicted candidate kinases that may catalyze CREB-B phosphorylation by using NetPhos program (www.cbs.dtu.dk/services/NetPhos/). We found that the Ser110 site of CREB-B is the most potentially phosphorylated by three kinases, namely, PKC, PKA, and ribosomal protein S6 kinase (S6K) (S1 Table). By using RNAi-mediated knockdown, we further examined the effects of these three kinases and another three reported kinases on p-CREB-B and NOS expression [31, 32]. The levels of p-CREB-B and NOS expression were significantly decreased by RNAi-mediated interferences of PKA activity. Suppressing the PKA activity by the knockdown of its catalytic C1 subunit (pkac1) notably inhibited the p-CREB level and NOS transcription (Fig 6A–6D). By contrast, enhancing the PKA activity by the knockdown of its regulatory R1 subunit (pkar1) induced opposite effects (Fig 6E–6H). The silencing of other five kinase encoding genes did not significantly change the levels of p-CREB and NOS transcription (S14 Fig). These results indicated that PKA is the key main regulator of p-CREB-B in the locust brain.
We then investigated whether or not PKA is involved in the regulation of NPF2 on the CREB-B/NOS signaling. The brain PKA activity was down-regulated by NPF2 injection but was promoted by the gene knockdown of NPF2 (Fig 7A and 7B). Moreover, the administration of PKA agonist (Colforsin) rescued the levels of p-CREB-B and NOS transcription, which were reduced by NPF2 injection in gregarious locusts (Fig 7C and 7D). Meanwhile, the injection of PKA inhibitor (KT5720) can fully recover the boosted levels of p-CREB-B and NOS transcription induced by NPF2 knockdown in solitary locusts (Fig 7E and 7F). In summary, these data revealed the essential role of PKA in the transduction of NPF2/CREB-B/NOS signaling involved in locust behavioral phase change.
Our study shows that the transcription factor CREB-B acts as a key mediator of the NPF/NO signaling pathway in regulating phase-related locomotor plasticity in migratory locusts. The essential role of CREB-B in locomotor plasticity is achieved through its direct control of NOS transcription under NPF2 regulation, and PKA transmits the effects of NPF2 on p-CREB-B.
By combined using genome walking and dual-luciferase assays, we characterized the promoter sequence of NOS gene. The core promoter sequence of NOS lies at 150 bp to 120 bp upstream region of the NOS ORF that is required for basic NOS transcription. In this region, there might exist other negative regulatory elements because we found that transcriptional activity of NOS promoter decreases when more than -140 bp DNA sequence is included in pGL4.1 vector, implying a complex regulatory networks in NOS promoter [33]. Apparently, gene transcription is a complex process that not only involves different regulators, but also depends on the sequence characteristics of promoter region [33, 34]. In fact, we bioinformatically predicted multiple CREB binding sites in the promoter region of NOS. And all three locust CREB members have active effects in HEK293T cells, although there are some differences in extent probably due to the differential expression levels of three CREB proteins or their distinct transactivation efficiency. In addition, the lower transcription activation effect of CREB3 observed in longer NOS promoter might be due to some inhibitory regulators, or more negative regulatory sequence were involved.
Further in vivo studies show that only CREB-B can directly regulate NOS transcription by interacting with CREB R4, whereas the other two CREBs, CREB-A and CREB3, do not have significant regulatory effects on NOS transcription in the locust brain. CREB-B and its target NOS are extensively expressed in the neurons upon PI which is a brain region that involves in the regulation of locomotor rhythms in insects [35]. The similar localization of CREB-B and NOS provided additional evidence for their interaction and may support their roles in locomotor modulation in the locust. To our best knowledge, this result provides the first evidence showing the regulatory roles of CREB in dynamic NOS transcription in insects. A number of studies have reported that CREB family members display diverse biological functions depending on their cell- or tissue- specific distributions in insects. For example, in Apis mellifera, CREB is localized in mushroom bodies and is associated with age-dependent labor division [36]. Moreover, CREB has different regulatory effects on the feeding behavior of Drosophila when it expresses in neuronal and peripheral tissues [37]. Therefore, three locust CREB proteins may have distinct distributions and functions. Actually, similar phenomena have also been reported in mammal species [38].
Many reports suggested that NOS transcription is essential for both neural and behavioral plasticity by controlling NO content under various physiological or external stimuli [39, 40]. The transcriptional regulatory mechanisms of three NOS isoforms in vertebrates such as neural NOS (nNOS, NOS1), inducible NOS (iNOS, NOS2), and epithelial NOS (eNOS, NOS3) have been well studied. eNOS is regulated by FOXO in human umbilical vein endothelial cells [41], whereas iNOS transcription is controlled by both STAT1 and NF-κB in human fibroblasts [42]. For nNOS, multiple TFs have been proposed, such as NF-κB, SP, and ZNF family members [43–45]. To date, only one NOS isoform in insects has been characterized [46]. The phylogenetic analysis suggests that insect NOS and vertebrate NOSs evolve from the same ancestor. It seems that three vertebrate NOSs duplicate from a single gene after evolutionarily divergence with insect NOS (S15 Fig). Although the importance of NO signaling has been reported in many insects [47, 48], the regulatory mechanism underlying NO synthesis is rarely uncovered. This study provides insights into the inducibility of NOS expression in response to environmental stimuli in animals.
Our results suggested that the phosphorylation level of CREB-B, but not the transcription level, is involved in the regulation of NOS transcription in response to the changes of population density in the locusts. Except for transcription alteration, several kinds of post-translational modifications, including phosphorylation, acetylation, and ubiquitination, have also been reported to control the transcriptional activation of CREB [31, 49, 50]. Phosphorylation, especially at Ser133 in the kinase-inducible domain (KID), is the most conserved indicator for CREB activation in mammals [51]. Our results showed that among the three antibodies that we tested; only the antibody against Ser133 phosphorylation of mammalian CREB1 (homolog for locust CREB-B) can detect the predicted band in the locust brain. We also observed that the phosphorylation level of CREB-B remarkably responded to the population density change. This result is similar to those from other studies; for example, the reduced CREB activity by protracted social isolation in rats [52] and the age-dependent increase in CREB phosphorylation in honeybees [36]. These findings indicate the CREB phosphorylation serves as a common molecular signal in response to the changes in internal or environmental states.
Among the candidate kinases that catalyze the phos-Ser133 of CREB [31, 32], only PKA has been demonstrated to be required for p-CREB-B in activating NOS transcription in the locust brain in our studies. However, we cannot exclude the potential roles of other several kinases on CREB phosphorylation in other physiological processes or tissues. The PKA-stimulated CREB phosphorylation has been accepted as a critical step for both neural and behavioral plasticity [53, 54]. The importance of PKA has been reported in the behavioral phase transition of desert locusts, Schistocerca gregaria [55]. Moreover, PKA has been recognized as a conserved down-stream signal transducer of both dopamine and serotonin [56, 57]. The two neurotransmitters induce behavioral gregarization and solitariness in the migratory locust, respectively [25, 58].
The involvement of PKA in NPF2/CREB-B axis was further validated by a series of pharmacological experiments of manipulating PKA activity. The importance of PKA in NPY-induced behavioral processes has also been revealed in other species. For example, the NPY signaling displays strongly suppressive effects on PKA-sensitized stress response in Drosophila [59]. In addition, the cAMP/PKA signal inhibits NPY-induced feeding behavior in rats [58]. We previously reported that the locust NPF2 peptide displays a close evolutionary relationship with NPY [26]. Thus, the NPY/PKA/CREB-B signaling cascade may represent a common mechanism underlying behavioral plasticity in animals. Surprisingly, although cAMP has been well-known as the main upstream messenger for PKA activation [53], we did not find any changes in cAMP level after manipulating either NPF1a or NPF2 peptide when we examined the whole locust brain [26]. One reasonable explanation is that the effects of NPF peptide on cAMP levels might have cell-specific patterns in the locust brain as reported in Drosophila [60]. The cAMP level may change only in a few neurons refer to NPF-NO circuit so that we cannot detect its changes in the whole brain. So, further work is needed to validate whether cAMP can link to NPF and PKA in specific regions of the locust brain during phase transition.
Here, we show that NPF2 displays a long-term effect on locomotor activity by negatively regulating p-CREB-B level and subsequent NOS transcription in locust brain. Another locust NPF member, NPF1a gene, has been confirmed to be not involved in the regulation of p-CREB-B by the RNAi-induced knockdown, or the injection of either the full length or the truncated peptide. Apparently, p-CREB-B/NOS signal transduction may specifically respond to NPF2 manipulation. According to this work and our previous findings [26], we infer that NPF2, CREB-B, and NOS should have overlap expressions in the same neuron cells of PI in locust brains. The effects of NPF2 on CREB-B/NOS might through autocrine or paracrine manner by disperse from its expressing cells as reported in other species [61, 62].
In summary, this study extends our previous findings by uncovering two key signal components CREB-B and PKA in the NPF2/NO signaling pathway involved in the regulation of phase-related locomotor plasticity (See model in Fig 8). Our findings present a previously undefined regulatory mechanism on NOS transcription in insects and shed light on how neuromodulator/TF/effect gene network contributes to environment-induced behavioral plasticity.
All locusts were obtained from a colony maintained at the institute of Zoology, Chinese Academy of Sciences. G-phase locusts were reared in large well-ventilated cages (40 cm × 40 cm × 40 cm) under crowded condition (500–1000 insects per cage). S-phase locusts were maintained separately in small boxes (10 cm × 10 cm × 25 cm) under physical, visual, and olfactory isolation from other locusts. Both colonies were maintained at 30 ± 2°C and under 14:10 light/dark photocycle regime. The locusts were fed with fresh wheat seedling and bran [23].
The fourth instar locusts were used for the time-course analysis. For the isolation treatment, the G-phase locusts were separately reared under the solitarious condition as described above, and their brains were collected and snap frozen after 0, 1, 4, and 16 h treatments. For the crowding treatment, two S-phase locusts were reared together with 20 G-phase locusts of the developmental stage in a small cage (10 cm × 10 cm × 10 cm), and their brains were dissected and frozen in liquid nitrogen after 0, 1, 4, and 16 h treatments. Each treatment group consisted of 8–10 locust individuals split approximately between sexes. Four independent biological experiments were conducted for each treatment.
Commercially synthesized peptides (BGI, NPF1a truncated peptide: YSQVARPRF-NH2; NPF1a full length peptide (NPF1a-FL): AEAQQADGNKLEGLADALKYLQELDRFYSQVARPRF-NH2; NPF2 peptide: RPERPPMFTSPEELRNYLTQLSDFYASLGRPRF-NH2, 2.5 μg/μl, 2 μl/each locust individual) were injected into the hemolymph in the heads of fourth-instar locusts by using a microinjector. The brains of test locusts were collected 2 h following injections. PKA agonist (Colforsin, 10 μM, 2 μl) was microinjected into the heads of G-phase locusts pretreated with NPF2 peptide, and the locust brains were dissected and frozen 1 h after drug injections. PKA inhibitor (KT5720, 50 μM, 2 μl) was microinjected into the heads of S-phase locusts pre-injected with dsNPF2. Collected brain samples were used for further analysis of p-CREB-B and NOS transcription.
Genome walking experiments were performed by using the Genome Walking Kit (Takara) following the manufacturer’s protocol to obtain the upstream regulatory region of NOS gene. Three specific primers (SP1, SP2, and SP3) were designed based on the known upstream sequence of NOS (obtained from the locust genome). Three times of thermal asymmetric interlaced PCR were sequentially performed by using the universal primer together with SP1, SP2, and SP3. The PCR product was ligated into the pGEM-Teasy vector (Promega) for sequencing.
The obtained upstream genome sequence of NOS was used for the bioinformatic predication of potential cis-response elements (CREs). MatInspector program (http://www.genomatix.de/) and TANSFAC program (http://genexplain.com/transfac/) were used for TF analysis to identify the most reliable putative regulatory elements. Candidate TFs raised in both programs were validated in further experiments.
The upstream regions of NOS with different lengths were amplified using specific primers and were sub-cloned into the pGL4.1 vector fused with a firefly luciferase (Pp-luc) maker gene (Life Technologies) to validate the predicted regulatory element responding to CREB proteins. The empty pGL4.1 vector was used as a negative control, and the pGL4.1 vector containing SV40 (pGL4.13[luc2/SV40], Promega) promoter sequence that can constitutively transcript was used as a positive control. The Opening reading frame (ORF) sequences of CREB-A, CREB-B, and CREB3 were respectively cloned into pcDNA3.1 expression vector with Flag tag on the C-terminal ends of the target genes.
The HEK293T cells were seeded in 500 μl of DMEM medium (Thermo Scientific) in a 24-well plate 1 day before transfection. The pGL4.1-derived constructs (200 ng/well) were separately or co-transfected with the CREB expression vectors (200 ng/well) to the HEK293T cells using LipofectamineTM 3000 (Invitrogen, California, USA). The pRL-TK that contains a Renilla luciferase (Rr-luc) encoding sequence was co-transfected with the pGL4.1-derived vectors, and was used as an internal control to normalize the transfection efficiency and luciferase activity [63]. All reporter constructs were prepared using the TIANprep MINI Plasmid Kit (TIANGEN, Beijing, China). The cells transfected with different recombinant vectors were cultured for additional 36 h at 37°C for transcriptional activity analysis using the Dual-Glo Luciferase Assay System (Promega) with a luminometer (Promega) according to the manufacturer’s instruction. The luciferase activity was defined as the ratio of Pp-luc activity from pGL4.1-derivative to Rr-luc activity from pRL-TK.
Double-stranded RNAs (dsRNAs) of target genes were synthesized using T7 RiboMAX Express RNAi system (Premega). dsRNA was firstly microinjected into the brains of the third instar locusts to improve RNAi efficiency and specificity. A second injection was performed at day 1 of the fourth instar locusts (200 ng/locust for each injection). dsGFP was used as the control in all RNAi assays. The brains of test locusts were collected at 48 h following the second injection.
Total RNA of experimental samples (6–8 locust/sample, four samples were collected for each treatment) was extracted using TRIzol reagent according to the manufacturer’s instruction. RNase-free DNase (1 μl, 1 U/μl, Promega) was added to RNA solution and incubated at 37°C for 30 min to remove genomic DNA, the mixture was then incubated at 65°C for 10 min to inactivate DNase. RNA quantification and reverse transcription were performed as previously described [26]. Gene-specific transcript levels were detected by qPCR using the SYBR Green kit on a LightCycler 480 instrument (Roche). RP49 was used as the internal reference. The primers are shown in S2 Table.
Gel-shift assay was conducted using the LightShift Chemiluminescent EMSA Kit (Thermo Scientific, USA) to verify the binding of CREB-B to candidate regulatory regions of NOS promoter. The oligonucleotides were labeled with biotin at the 5′ end and incubated at 95°C for 10 min and then annealed to generate the double-stranded probe by natural cooling. The cold probes (unlabeled probe) or mutant probes were used as competitors of wild type probes. All oligonucleotide probes were synthesized by Invitrogen Company (Shanghai, China).
DNA-binding reactions were conducted in a 20 μl mixture containing 1 μg of nuclear extracts (isolated from the locust brains using Nuclear and Cytoplasmic Protein Extraction Kit, Beyotime), 50 ng of poly (dI-dC), 2.5% glycerol, 0.05% NP-40, 50 mM KCl, 5 mM MgCl2, 4 mM EDTA, and 0.25 μM of the biotin-labeled probe at room temperature for 20 min. For the competition assay, cold probes (100 ×) were added to the binding reaction. For the super-shift assay, 2 μg of CREB-B antibody or rabbit IgG was added and incubated for another 30 min at room temperature. The protein-DNA complexes were separated using a 6% polyacrylamide gel and transferred onto nylon membranes. The transferred complex was then exposed to UV light cross-linking for 300 s (254 nm, 1200 mJ). The membrane was incubated with a streptavidin-horseradish peroxidase conjugate and was detected by cemiluminescent nucleic acid detection module (Thermo Scientific).
Whole-mount immunohistochemistry was performed as described in Hou et al. (2017). Brains of the fourth instar locusts were dissected and fixed in 4% paraformaldehyde (PFA) at 4°C overnight. Polyclonal antibody against p(S133)-CREB1 (Cell signaling, 1:100) and monoclonal mouse antibody against uNOS (Thermo, 1:100) were used as the primary antibodies, which were incubated with locust brains for 48 h at 4°C. Alexa Fluor-488 goat anti-rabbit IgG (1:500; Life Technologies) and Alexa Fluor-568 goat anti-mouse IgG (1:500; Life Technologies) were used as secondary antibody for CREB-B and NOS staining, respectively. Hoechst (1:500) was used for nuclear staining. Fluorescence was examined under the LSM 710 confocal laser-scanning microscope (Zeiss).
Total proteins from locust brains were extracted using TRIzol reagent following the manufacturer’s instruction. Brain tissues (8–12 insects/sample) were homogenized in 1 ml of TRIzol reagent. After 200 μl of chloroform was added, the aqueous phase was used for total RNA isolation, and the DNA in the phenol phase and interphase was excluded by ethanol precipitation. Proteins were then precipitated by adding isopropyl alcohol for 20 min at room temperature. The protein pellet was washed in a solution containing 0.3 M guanidine hydrochloride in 95% ethanol, followed by 100% ethanol washing. After vacuum drying, the protein pellet was weighed and dissolved in 1% SDS sample buffer to 10 μg/μl.
The protein extracts (100 μg) were electrophoresed on 12% SDS-PAGE gel and then transferred to polyvinylidene difluoride (PVDF) membrane (Millipore). The membrane was incubated with polyclonal antibody against p(S133)-CERB1 (Cell signaling, 1:1000) or monoclonal antibody against H3 (1:5000). Goat anti-rabbit IgG (CoWin, 1:5000) and goat anti-mouse IgG (CoWin, 1:10000) were used as secondary antibody for CREB and H3, respectively. Protein bands were detected by chemiluminescence (ECL kit, Thermo Scientific). The band intensity was analyzed using Quantity One software. Briefly, the targeting bands were selected and the background signal was deducted. Band intensity was presented as the peak area value and normalized by that of nucleoprotein H3.
Locomotor activity was monitored as previously described [26]. Generally, locust behaviors were detected in a rectangular Perspex arena (40 cm × 30 cm × 10 cm). The locust behaviors were recorded for 6 min by an EthoVision video tracking system. Total distance moved and total duration of movement in the middle 300 s represent the locomotor activity of individual locusts. More than 15 locusts were tested for each experimental treatment according to the sample size used in previous studies [25, 55]. Locusts that did not move in the arena assay were excluded.
PKA activity was examined using the PKA Kinase Assay Kits, Type I (Immunechem) according to the manufacturer’s protocol. The locust brains (15–20 locusts/sample) were collected and homogenized in 200 μl 1 × PBS buffer (0.1 M phosphate buffer, 0.15 M NaCl, pH 7.4). The supernatant (50 μl/well) was added into the substrate plate containing kinase assay dilution buffer. ATP (10 μl/well) was added to initiate the kinase reaction at 30°C for 90 min. After the reaction solution was removed, anti-p-substrate antibodies (40 μl/well) were incubated for 60 min at room temperature. Goat anti-rabbit IgG HRP was used as secondary antibody. TMB solution was used to develop the color indicating reaction activity. OD450 was detected to calculate the relative kinase activity. The enzyme activity was normalized to the protein concentration, which was measured by using BCA method.
The protein sequences of Drosophila CREB-A and CREB-B were used to search their homologs in the locust genome and transcriptome database by utilizing the tblastn algorithm. CREB proteins in several representative insects and human were used to construct their phylogenetic relationship by using MEGA software. Neighbor-joining method and Poisson model was used and the number of bootstrap replications was 1000. The sequence identity of three locust CREB proteins was analyzed in GENEDOC software.
All data were statistically analyzed using GraphPad Prism 5 software and presented as mean±s.e.m. Student’s t-test was used for two-group comparison, and one-way ANOVA followed by Turkey’s post-hoc test was used for multi-group comparisons. Differences were considered as statistical significance at P < 0.05. All the experiments were performed with at least three independent biological replicates. Numerical data for main figures and supporting figures were shown in S3, S4 and S5 Tables.
|
10.1371/journal.pgen.1002116 | Genetic Analysis of Genome-Scale Recombination Rate Evolution in House Mice | The rate of meiotic recombination varies markedly between species and among individuals. Classical genetic experiments demonstrated a heritable component to population variation in recombination rate, and specific sequence variants that contribute to recombination rate differences between individuals have recently been identified. Despite these advances, the genetic basis of species divergence in recombination rate remains unexplored. Using a cytological assay that allows direct in situ imaging of recombination events in spermatocytes, we report a large (∼30%) difference in global recombination rate between males of two closely related house mouse subspecies (Mus musculus musculus and M. m. castaneus). To characterize the genetic basis of this recombination rate divergence, we generated an F2 panel of inter-subspecific hybrid males (n = 276) from an intercross between wild-derived inbred strains CAST/EiJ (M. m. castaneus) and PWD/PhJ (M. m. musculus). We uncover considerable heritable variation for recombination rate among males from this mapping population. Much of the F2 variance for recombination rate and a substantial portion of the difference in recombination rate between the parental strains is explained by eight moderate- to large-effect quantitative trait loci, including two transgressive loci on the X chromosome. In contrast to the rapid evolution observed in males, female CAST/EiJ and PWD/PhJ animals show minimal divergence in recombination rate (∼5%). The existence of loci on the X chromosome suggests a genetic mechanism to explain this male-biased evolution. Our results provide an initial map of the genetic changes underlying subspecies differences in genome-scale recombination rate and underscore the power of the house mouse system for understanding the evolution of this trait.
| Homologous recombination is an indispensable feature of the mammalian meiotic program and an important mechanism for creating genetic diversity. Despite its central significance, recombination rates vary markedly between species and among individuals. Although recent studies have begun to unravel the genetic basis of recombination rate variation within populations, the genetic mechanisms of species divergence in recombination rate remain poorly characterized. In this study, we show that two closely related house mouse subspecies differ in their genomic recombination rates by ∼30%, providing an excellent model system for studying evolutionary divergence in this trait. Using quantitative genetic methods, we identify eight genomic regions that contribute to divergence in global recombination rate between these subspecies, including large effect loci and multiple loci on the X-chromosome. Our study uncovers novel genomic loci contributing to species divergence in global recombination rate and offers simple genetic explanations for rapid phenotypic divergence in this trait.
| Meiotic recombination fulfills dual roles in genetics and evolution. In many species, including mammals, the proper segregation of homologous chromosomes at the first meiotic division is contingent on the presence of at least one well-positioned crossover per homologue pair [1]–[3]. The improper patterning of recombination events across chromosomes can lead to aneuploidy, a significant risk factor for fetal loss and developmental disability in humans [4]. In addition, recombination influences the evolutionary dynamics of populations by rearranging existing patterns of allelic variation to generate novel multi-locus genotypes. This genetic shuffling can increase the efficacy of natural selection by decoupling high fitness alleles from linked deleterious variation [5]–[7]. At the same time, recombination can facilitate the removal of deleterious variation from the gene pool [8], [9].
The amount of recombination per unit DNA (i.e. the rate of recombination) exhibits tremendous variation among species and between individuals. In mammals, the mean rate of recombination across species genomes varies by an order of magnitude [10]–[12]. Likewise, there is considerable heterogeneity in the global crossover rate among individual humans [13]–[17], house mice [18],[19], dogs [20], cows [21], and shrews [22]. Fine-scale recombination rates display similar trends: the genomic locations of recombination hotspots are not conserved between humans and chimpanzees [23], [24], and hotspots that segregate as presence/absence polymorphisms are common in human populations [25], [26] and among closely related laboratory strains of house mice [27]–[29].
Classical genetic experiments established that the rate of recombination is a complex genetic trait [30]–[33]. More recently, specific genes that influence genome-scale recombination rate variation in humans have been identified [17], [34], [35]. Prdm9, a meiosis-specific histone H3 methyltransferase, was recently found to control the genome-wide distribution of recombination hotspots in mice [36] and humans [37]–[39]. Despite these exciting advances in our understanding of population variation in recombination rate, genetic explanations for the large differences in recombination rate between species remain elusive. Fundamental questions have never been addressed experimentally: How many loci contribute to species divergence in recombination rate? What are their effect sizes and modes of inheritance? Where in the genome do species recombination rate modifiers reside? Do loci that control the positioning and activity of recombination within species also contribute to recombination rate differences between species? Answers to these questions are essential for understanding how the rate of recombination evolves.
To date, most genetic studies of recombination rate variation have measured recombination rates using patterns of marker inheritance in large human pedigrees [16], [17], [35] or in experimental crosses [29], [40], [41]. However, recombination rates are estimated from genetic data with considerable statistical uncertainty owing to the independent assortment of recombinant chromatids at meiosis [42]. In pedigree-based studies, this error is further compounded by the limited number of meiotic transmissions surveyed per individual. In addition, the inability to eliminate environmental contributors to phenotypic variation in humans adds even more noise to recombination rate estimates. These sources of error result in a marked loss of statistical power to find genomic regions contributing to recombination rate variation through linkage or association analysis.
A powerful alternative approach to the genetic dissection of recombination rate variation is to combine experimental crosses with cytological measures of recombination rate [43]. In particular, the immunolocalization pattern of the mismatch repair protein MLH1 along the mature synaptonemal complex has been shown to accurately and faithfully reproduce the distribution of meiotic crossovers in late pachytene spermatocytes [14], [44]–[46] and oocytes [47]. The MLH1 method for measuring recombination rate offers several notable advantages over traditional genetic approaches. First, because crossovers are directly observed, recombination rate estimates are not affected by binomial sampling of recombinant chromosomes at meiosis. Second, large numbers of spermatocytes or oocytes can be analyzed to yield precise estimates of the global recombination rate for single individuals. Third, though laborious, this method is inexpensive compared to the costs of genotyping many offspring from a single individual (as required by pedigree-based methods).
We use the MLH1 immunocytological method to demonstrate that males from wild-derived inbred strains of the house mouse subspecies Mus musculus musculus have markedly increased genome-scale recombination rates relative to the closely related subspecies M. m. castaneus and M. m. domesticus. We identify multiple genetic determinants of this substantial divergence in global recombination rate, providing an initial portrait of the genetic basis of species differences in this key genomic parameter.
A representative image of a late pachytene spermatocyte stained with fluorescently labeled antibodies against MLH1 and a protein component of the synaptonemal complex (SYCP3) is shown in Figure 1. We used the MLH1 immunostaining assay to measure genomic recombination rates in two wild-derived inbred strains from each of three distinct subspecies of house mice (Mus musculus musculus, M. m. castaneus, and M. m. domesticus) [48]. We observed a striking difference in mean MLH1 foci count between M. m. musculus and both M. m. domesticus and M. m. castaneus males (Figure 2; Wilcoxon signed rank test, P<10−16). On average, M. m. castaneus and M. m. domesticus have 21–23 MLH foci per meiosis, whereas M. m. musculus males undergo >26 crossovers.
Several patterns suggest that the distribution of MLH1 foci along the synaptonemal complex (SC) faithfully mirrors the distribution of meiotic crossovers in the wild-derived inbred strains we examined. First, SCs lacking a MLH1 focus were rare in our survey (0.4%), consistent with the obligate chiasma requirement for homologue disjunction in mammals [1], [2]. Second, on SCs with two or more MLH1 foci, foci were distantly spaced. This patterning matches expectations under a model of positive crossover interference, a process that is known to be important in house mice [42]. Third, we very seldomly observed cells with two or more SCs lacking a MLH1 focus. Pachytene spermatocytes nearly always had a full complement of foci, indicating that MLH1 protein loads on and off sites of recombination repair along the SC in a highly concerted fashion. Fourth, our cytology maps approximate the total male mouse genetic map length estimated from genetic data. Assuming that each MLH1 focus corresponds to a map distance of 50 cM, wild-derived inbred strains included in our survey have map lengths that range from 1085 cM–1500 cM. This range includes the estimate of total male genetic map length from the standard mouse map (1375 cM) [49]. Although the small fraction of crossover events that are not dependent on MLH1 will be missed by this method [50], our observations suggest that the total number of MLH1 foci in a meiotically dividing cell provides a reliable estimate of the genomic recombination rate.
Mus musculus subspecies radiated nearly simultaneously from a common ancestor ∼500,000 years ago [51], [52]. The striking increase in mean MLH1 foci count in inbred strains of M. m. musculus relative to the M. m. domesticus and M. m. castaneus strains suggests that considerable divergence in male recombination rate has accrued along the M. m. musculus lineage. Under a neutral model of phenotypic evolution, the expected recombination rate divergence between subspecies is ≈2Vmt, where Vm is the per-generation rate at which phenotypic variance increases via neutral mutation and t is the divergence time in generations [53]. Given that t is roughly equal for pairwise comparisons between house mouse subspecies and assuming constancy of Vm over this short evolutionary period, absolute divergence in recombination rate should be approximately equal among subspecies pairs. Clearly, our data are not consistent with this theoretical prediction. At mutation-drift equilibrium, the amount of within subspecies polymorphism for recombination rate is ≈2NeVm, where Ne is the effective population size [53]. Curiously, M. m. musculus has a smaller estimated Ne than either M. m. domesticus or M. m. castaneus (∼60,000, 100,000, and 200,000, respectively; [52]) yet displays the highest level of polymorphism for recombination rate (∼3 MLH1 foci between CZECHI and PWD). The higher polymorphism for recombination rate in M. m. musculus and greater divergence for recombination rate in comparisons with this subspecies are consistent with several evolutionary hypotheses. Recombination rates may have experienced a relaxation of selective constraint along the M. m. musculus lineage. Alternatively, recombination rates in M. m. domesticus and M. m. castaneus may have been subject to stronger purifying selection. These findings could also be explained by higher mutational variance for recombination rate in M. m. musculus. We caution that these observations derive from comparisons of just two wild-derived inbred strains per subspecies. A detailed survey of polymorphism and divergence in recombination rate in natural populations of these three subspecies will be required to determine the underlying evolutionary processes at work.
To investigate the genetic basis of the rapid divergence in genomic recombination rate in M. m. musuclus, we conducted an F2 intercross between wild-derived inbred strains PWD/PhJ (M. m. musculus) and CAST/EiJ (M. m. castaneus). We measured the global rate of recombination in 276 F2 males by averaging total autosomal MLH1 foci counts from at least 15 spermatocytes per animal (mean = 20.4 cells). F2 mice vary substantially in the global number of crossovers (Figure 3). Importantly, the variation in MLH1 foci number among cells from a single male is far less than the variation in mean MLH1 foci count among animals, indicating the presence of segregating genetic differences in this inter-subspecific F2 population. Most males have recombination rates that fall within the range defined by the two parental means, although 9% of individuals lie outside these values. The continuous nature of this variation provides evidence for multiple recombination rate modifiers segregating between the parental PWD and CAST strains. The high broad-sense heritability of recombination rate in this cross (H2 = 0.93) motivates the application of genetic mapping approaches to link variation in recombination rate with genetic variation at specific genomic loci.
We genotyped our F2 population at 222 informative SNPs distributed across the genome. Using standard interval mapping [54] with a permutation-derived threshold for declaring statistical significance (genome-wide α = 0.05) [55], we identified two genomic regions linked to variation in mean MLH1 foci count. One of these QTL localizes to the proximal half of chromosome 7 and the second QTL lies on the X chromosome (Figure 4). Although there is a clear peak in the X chromosome LOD profile centered on ∼30 cM, the entire chromosome displays strong statistical evidence for linkage to variation in global recombination rate (Figure 4). QTL genotype at this single, large-effect locus explains 46% of the variance in mean MLH1 foci count among F2 males (adjusted R2 = 0.46 from a linear regression).
Interestingly, the allele from the low recombination rate CAST parent confers the increase in recombination rate at this X-linked locus, opposite the pattern seen at the chromosome 7 QTL (Table 1). Consistent with this result, we uncover a striking difference in genomic recombination rate between reciprocal F1 males (Figure 6). Male F1 animals that receive their X chromosome from a CAST mother (CASTxPWD F1s) have ∼5 more MLH1 foci per meiosis than F1 males carrying the PWD X chromosome (PWDxCAST F1s).
Single QTL mapping approaches, including standard interval mapping, formally assume that only one QTL in the genome affects the phenotype. When QTL of moderate to large effect exist, accounting for the phenotypic variance they explain can enhance statistical power to find additional loci. The discovery of the major effect QTL on the X chromosome prompted us to use an approach that could adjust for the presence of this locus to enable the simultaneous detection of multiple additional QTL. We applied a model-based multiple QTL mapping strategy [56] to identify the set of genetic loci that best explain segregating variation in mean MLH1 foci count among F2 males. Using a forward/backward model selection approach, with model discrimination performed via penalized LOD scores to control the rate of false inclusion [57], we identify six autosomal QTL and two X-linked QTL for genomic recombination rate (Figure 5). As expected, the two QTL identified in the single QTL scan are among those recovered in the multiple QTL mapping analysis.
The six autosomal loci show mostly additive effects, with CAST alleles at the chromosome 4, 11, and 17 QTL exhibiting slight dominance over PWD alleles (Table 1). At each autosomal locus, the high recombination rate PWD parent confers the high recombination rate allele. Consistent with single QTL scan results and reciprocal F1 phenotypes, the high recombination rate allele at both X-linked QTL derives from the low recombination rate CAST parent (Table 1).
Combined, these eight QTL explain 74% of the phenotypic variance among F2s, individually accounting for 0.9–3.4 MLH1 foci (2–35% of the total phenotypic variance; Table 1). These effect sizes are probably overestimated, as the conflation of QTL detection and estimation on a common dataset leads to a systematic upward bias [58].
Although these eight QTL explain a large fraction of F2 variation in global recombination rate, they account for a lesser percentage of the observed difference between the parental PWD and CAST strains. Combined, the six autosomal QTL explain a difference of 8.5 MLH1 foci between the inbred parents – more than the observed difference of 8 foci. However, the two X-linked QTL account for ∼4 MLH1 foci in the opposite direction. Summing these effects suggests that our multiple QTL model explains approximately half of the difference in mean MLH1 foci count between PWD and CAST males. Clearly, additional QTL for mean MLH1 foci number segregate between these strains. These undetected QTL likely have small to moderate effect sizes, as power calculations indicate that our study is only sufficiently powered (80% power) to find QTL with additive effects >0.9 [59].
The early stages of female meiosis, including recombination, occur in the fetal ovary. These temporal aspects of oogenesis complicate cytological analysis of recombination in females. For this reason, we limited our genetic mapping efforts to males. However, the genome-wide rate of recombination is a sexually dimorphic trait in many mammals, including house mice. The female standard mouse genetic linkage map is 9% longer than the corresponding male map [49], and marked sex-specific recombination trends are observable on finer physical scales of measurement [29], [49], [60], [61]. The non-random concentration of QTL with transgressive effects to the X chromosome, coupled with the noted sex differences in this trait, led us to investigate variation in global recombination rate in females from the two parental inbred strains and hybrid F1s.
We applied the MLH1 immunostaining procedure to oocytes harvested from day 17–20 post-conception PWD and CASTxPWD F1 female fetuses (n = 3 and n = 2 animals, respectively). Mean MLH1 foci counts from CAST females have been reported previously [62]. Although CAST and PWD males differ in their global crossover count by 8 MLH1 foci, PWD females have only 2 foci more than CAST females (Figure 6). PWD and CASTxPWD F1 females have indistinguishable crossover counts. Overall, there is surprisingly little variation for mean MLH1 foci count among females, indicating that evolutionary divergence in recombination rate has occurred primarily in males. These findings suggest that several of the QTL detected in our inter-subspecific F2 male population may be sex-limited in their expression or have polarizing effects in males versus females [17].
Interestingly, PWD females have a lower mean MLH1 foci count than PWD males (Figure 6). This finding presents an intriguing directional reversal of the global recombination rate sex dimorphism widely observed in house mice [49], [61], nominating the PWD strain as an excellent model for understanding the causes of sex differences in this phenotype.
Our genetic study of variation in mean MLH1 foci number in an inter-subspecific panel of F2 males identifies QTL for global recombination rate divergence. Our findings provide initial clues toward the genetic mechanisms of species divergence in this trait.
First, the discovery of eight QTL jointly explaining 74% of the variance among inter-subspecific F2 males indicates that observed patterns of recombination rate evolution are dominated by loci with large phenotypic effects. The X-linked QTL at 33 cM is the strongest modifier of global recombination rate identified to date, explaining 35% of the variation in our inter-subspecific F2 panel (Table 1). The presence of such a large effect locus provides a clear genetic mechanism for rapid phenotypic evolution between species.
Second, the autosomal loci we identify display mainly additive inheritance. This finding extends studies of within species recombination rate variation in humans [34], indicating that additive alleles contribute to both within and between species differences in recombination rate.
Third, at least two of the QTL we identify exert trans effects on recombination rate. Males do not recombine along their X chromosome, indicating that the two X-linked QTL act strictly in trans. This result corroborates findings from genetic studies of fine-scale recombination rate control: Prdm9 regulates recombination hotspot activity across the mouse genome [36], indicating that trans regulatory mechanisms are important for both the fine- and broad-scale control of recombination.
Fourth, our multiple QTL map points to the presence of high and low recombination rate alleles in the two parental strains (Table 1). A similar pattern has been previously reported for recombination rate variation in Drosophila melanogaster [32] and is often observed in the evolution of complex traits [63].
Finally, our study uncovers a prominent role for the X chromosome in the evolution of recombination rate. Combined, the two X-linked loci in our multiple QTL model account for a difference of 4 MLH1 foci (200 cM) between males hemizygous for CAST versus PWD alleles. Recessive X-linked loci subject to positive selection will reach fixation more rapidly than autosomal loci because their expression is unmasked in hemizygous males [64]. We speculate that selection on the X-linked modifiers identified in our F2 male population may have played a leading role in the rapid evolution of recombination rate in this sex.
Recently, Murdoch et al. [43] used the approach applied in our study – genetic mapping of MLH1 foci count in F2 males – to identify seven QTL conferring recombination rate differences between the C57BL6 and CAST inbred mouse strains. Five of these loci map to chromosomes that harbor QTL in our study, including the X chromosome (chromosomes 3, 4, 15, 17, and X). Interestingly, the CAST genotype at the X-linked QTL was associated with only a moderate increase in F2 recombination rate in this study, as opposed to the large effect observed in our cross (males with the CAST X have ∼1 focus more than males with the C57BL6 X chromosome; in comparison, males with the CAST X have ∼2.5 foci more than males with the PWD X chromosome). If the large-effect X-linked QTL at 33 cM identified here and the X-linked QTL identified by Murdoch et al. [43] are the same locus, it would appear that genetic background strongly affects its expression. Our application of multiple QTL mapping did not identify any genetic interactions (even when the penalty to QTL inclusion was relaxed; data not shown), but we acknowledge limited power to find interacting QTL with our small sample size.
Allelic incompatibilities that decrease reproductive fitness in hybrids commonly evolve between incipient species [65]–[67]. These genetic incompatibilities often affect hybrid fitness by hindering progression through meiosis, including impairment of chromosome synapsis and recombination [e.g. 68]. Importantly, we detect no epistasic interactions between subspecies-specific alleles in our cross. Several additional observations indicate that observed F2 variation in mean MLH1 foci count is not due to hybridization-related defects in meiosis. First, our immunostaining assay allowed us to identify diplotene stage cells in all F1 and F2 animals, ruling out wide-spread activation of the pachytene meiotic checkpoint as an underlying mechanism of possible hybrid sterility [69]. Second, we observed no overt defects in chromosome pairing or synapsis in any hybrid animals. If CAST and PWD hybrids suffer fitness reductions, the molecular mechanism(s) of infertility must act after the completion of recombination at prophase I. It is difficult to imagine how any problems that surface late in meiosis (or possibly in spermatogenesis) could affect recombination. Third, the distribution of mean MLH1 foci counts in our F2 population is centered on the mid-parent mean (Figure 3), an unlikely result in the event of hybrid dysgenesis in the phenotype. Finally, we note that M. m. musculus and M. m. castaneus hybridize in nature [70] and F1s from both directions of our CAST and PWD cross were fertile. Together, these observations provide a compelling case that patterns of inter-subspecific hybrid variation in mean MLH1 foci count reflect underlying subspecies differences in the rate of recombination per se.
Our analysis uncovered two striking differences in recombination rate evolution between the sexes. First, the magnitude of evolutionary change is much greater in males than females. Second, there is a reversal in the direction of the sex-dimorphism for recombination rate between PWD and CAST. It is tempting to consider these results, combined with the localization of both transgressive QTL to the X chromosome, as inter-related findings. In particular, they seem to raise the possibility that divergence in recombination rate among house mouse subspecies has been shaped by conflicting evolutionary pressures on the sexes. If natural selection favors distinct recombination rates in males and females [71], [72], modifiers might preferentially aggregate on the X chromosome [73]. Recessive X-linked loci will differ in expression between males and females, thereby imposing a reduced fitness burden on the opposite sex. This scenario is speculative, especially given that the dominance effects of X-linked recombination rate modifiers identified here cannot be determined from hemizygous F2 males. An extension of our QTL analysis to include mean MLH1 foci counts in F2 females will offer further clues into the evolution and genetic basis of these intriguing observations.
Although the rate of recombination differs between males and females of many mammalian species [13], [16], [19], [74], [75], the causes of this pattern remain poorly understood. Sex differences in crossover interference [76], features of the meiotic cell cycle [77], [78], the strength of epistatic selection in haploid gametes [72], and the genetic architecture of recombination [17], [35] may play contributing roles. Few X-linked recombination rate modifiers have been previously identified [43], [79], but our recent findings suggest that sex-linked loci are pervasive components of the genetic architecture of recombination rate evolution in house mice [this study]; [ 48,80]. Further examination of the genetic basis of recombination rate should allow the relative importance of sex chromosome evolution and other causes of sexual dimorphism to be determined.
Prdm9 is the only gene known to contribute to species differences in recombination rate. Human and chimpanzee alleles of Prdm9 are predicted to recognize and bind distinct DNA sequence motifs that may be important for recombination hotspot initiation [37], [81], [82]. These observations have led to the hypothesis that rapid evolution at Prdm9 underlies abrupt shifts in the distribution of recombination hotspots between species [37], [81]. Although the CAST and PWD strains used in our study have different functional variants of Prdm9 [36], we do not find QTL that co-localize with this gene. Prdm9 modifies the activity of multiple hotspots in mice [36] and in humans [38], but it does not appear to have detectable effects on the global level of recombination in either species [this study; 37]. Taken together, these findings suggest that recombination rate evolution on fine and broad scales could be controlled by separate genes [83].
While examining the genetic control of hotspot activity can deliver mechanistic insights into recombination rate evolution, the total number of recombination events in a meiotically dividing cell – the phenotype examined here – is more likely to be a functionally relevant measure [84]. For example, female reproductive output is associated with global recombination rate in humans [15], [16], whereas the presence or absence of recombination activity in individual hotspots has yet to be linked to variation in fitness. In fact, the rapid evolutionary turnover of recombination hotspots within [25], [26], [39], [85], [86] and between species [23], [24] seems to argue against a selective advantage of particular hotspot locations over others. In contrast, the genomic rate of recombination is subject to evolutionary constraints imposed by its essential functions in mammalian meiosis. A minimum of one crossover per chromosome is required for the proper disjunction of homologs in mice whereas high recombination rates may elevate the frequency of deleterious rearrangements [87].
Our study nominates eight genomic regions contributing to evolutionary divergence in genomic recombination rate. Future work will be required to determine whether the causal alleles are fixed or shared between M. m. musculus and M. m. castaneus. The observation that independent wild-derived inbred strains of M. m. musculus and M. m. castaneus conform to the recombination pattern established by PWD and CAST (Figure 2) suggests that at least some of these QTL represent subspecies differences.
The QTL identified in our analysis have broad peaks, each spanning a genomic interval that includes hundreds of genes. As a first step toward the identification of the causal variant(s) underneath the large X-linked QTL at 33 cM, we assayed transcript abundance between PWD and CAST alleles at 12 candidate genes. We found a suggestive difference in allele-specific expression at one gene, Brcc3, a component of the BRCA1-BRCA2 complex involved in double-strand break repair (Figure S1; Text S1; Table S1) [88]. These considerations nominate Brcc3 as a putative candidate gene for divergence in recombination rate. However, fine-mapping strategies will be required to test this hypothesis and to further localize the genetic changes that contribute to the increased global recombination rate in PWD.
Genetic and ecological resources will facilitate the fine-mapping of QTL identified in our experimental intercross. The Collaborative Cross, an eight-way recombinant inbred line panel currently under development, includes inbred strains CAST and PWK [89], a close relative of PWD. The increased mapping resolution and ability to measure mean MLH1 foci count on multiple animals with identical genotypes are key advantages of this resource that will aid efforts to fine-map those QTL that are common between PWK and PWD. In addition, populations of M. m. musculus and M. m. castaneus hybridize in nature, forming a fourth widely recognized subspecies of house mouse, M. m. molossinus [70]. The genetic properties of these natural hybrid populations are not well characterized, but lower levels of linkage disequilibrium could allow genomic windows containing causative loci to be narrowed through association studies or admixture mapping [90]. Identifying the determinants of the marked divergence in male recombination rate between M. m. musculus and M. m. castaneus promises to reveal the mechanisms of sex-limited evolution in this important phenotype.
Wild-derived inbred strains of Mus musculus castaneus (CAST/EiJ) and Mus musculus musculus (PWD/PhJ and CZECHI/EiJ) were purchased from the Jackson Laboratory (Bar Harbor, Maine, USA) and housed in the University of Wisconsin School of Medicine and Public Health mouse facility according to animal care protocols approved by the University of Wisconsin Animal Care and Use Committee. Pups were weaned into same-sex groupings at 21 days, with males subsequently separated into individual cages prior to 56 days. Animals were provided with food and water ad libitum. A total of 315 F2 males were sacrificed at 70 (±3) days of age (305 CAST/EiJ×PWD/PhJ and 10 PWD/PhJ×CAST/EiJ).
Males from inbred strain CIM were purchased from Dr. Francois Bonhomme's stock repository at the Universite Montpellier II. Animals were sacrificed shortly after arrival to the University of Wisconsin-Madison (aged 24.5–35.5 weeks).
Spermatocyte spreads were prepared as described [91]. Briefly, the left testis of sexually mature males was removed, weighed, and rinsed in sterile 1× PBS. The outer tissue coating of the testis was punctured to allow a small volume of seminiferous tubules to be extracted. Tubules were incubated in a hypotonic solution (30 mM Tris, 50 mM sucrose, 17 mM citric acid, 5 mM EDTA, 2.5 mM dithiothreitol, 0.5 mM phenylmethanesulfonylfluoride) for approximately 45 minutes at room temperature. Tubules were then transferred to a small volume (20 µl) of 100 mM sucrose solution deposited on a clean glass slide and shredded using fine-gauge forceps. Tubular remnants were removed and an additional 20 µl of 100 mM sucrose added to the cell slurry. The solution was agitated by pipetting and 20 µl deposited onto each of 2 3×1″ glass slides coated with 100 µl 1% paraformaldehyde supplemented with TritonX-100 (0.15%; pH = 9.2). The slides were gently rocked to distribute cells across their surface and allowed to dry overnight in a room temperature humid chamber. Dried slides were then washed briefly in 0.4% PhotoFlo (Kodak), air dried, and subjected to immunostaining.
The immunostaining protocol was adapted from Anderson et al. [45] and Koehler et al. [46]. A 10× concentration of antibody dilution buffer (ADB) was prepared (2.5 mL normal donkey serum (Jackson ImmunoResearch), 22.5 mL 1× PBS, 0.75 g bovine serum albumin (Fraction V; Fisher Scientific), and 12.5 µl TritonX-100) and sterilized by vacuum filtration (0.45 µm; Millipore). Slides were blocked in 1× ADB (diluted in 1× PBS) for approximately 30 minutes then lightly drained by touching the edge of the slide to a clean paper towel. All antibody dilutions were made into 1× ADB and all incubations were performed in a 37 C humid chamber. A 60 µl aliquot of primary antibody cocktail (1∶50 rabbit polyclonal antibody against MLH1 (Calbiochem) and 1∶50 goat polyclonal antibody against SYCP3 (SantaCruz Biotechnology)) was dispensed on each slide. Slides were cover-slipped, sealed with rubber cement, and incubated overnight. The rubber cement was then carefully removed and coverslips were soaked off in 1× ADB. Slides were washed twice for 30 minutes each in 1× ADB. A 60 µl volume of 1∶100 Alexa 488 donkey anti-rabbit secondary antibody (Molecular Probes) was deposited on each slide. Slides were cover-slipped, sealed with rubber cement, and incubated overnight. After soaking off coverslips in 1× ADB, 60 µl of 1∶100 Alexa 568 donkey anti-goat secondary antibody (Molecular Probes) was applied to each slide. Slides were sealed with a parafilm “coverslip” and incubated for 2 hours. Slides were then washed three times for one hour each in 1× PBS, air-dried, and mounted in a drop of ProLong Gold antifade media (Molecular Probes).
Cells were visualized using a Zeiss Axioskop microscope equipped with an AxioCam HRc camera and a 100× objective lens. Late pachytene cells that were damaged during preparation, displayed bulbous chromosome termini (indicative of transition into diplotene), lacked clear cell boundaries, or displayed flagrant defects in synapsis were not imaged. Images were captured in AxioVision (Rel. 4.8) software and stored as moderate resolution tiff files. Images were subsequently cropped and the fluorescent intensity adjusted using ImageJ software.
Numbers of autosomal MLH1 foci in late pachytene cells were manually scored. Only cells characterized by (i) the complete merger of SYCP3 signals from the two homologues, (ii) a full complement of chromosomes, (iii) clear, brightly stained MLH1 foci, and (iv) minimal background fluorescence were scored. We retained only cells with at least one MLH1 focus on each synaptonemal complex, excepting the possibility of one achiasmate bivalent; cells with more than two synaptonemal complexes lacking a MLH1 focus were extremely rare and likely represent staining artifacts. Approximately 20 cells were scored per animal. We were unable to obtain a sufficient number of high quality images for 39 of the 315 F2 animals.
DNA from each F2 animal was extracted from liver tissue using a Wizard Genomic DNA Purification Kit (Promega) following manufacturer's protocols. 295 SNPs distinguishing PWD/PhJ and CAST/EiJ alleles were identified from Phase 4 of the Perlegen mouse resequencing project (Frazer et al. 2007) and genotyped using the Sequenom iPLEX (San Diego, CA) MassARRAY system as previously described [92]. Raw genotype data were cleansed of putative genotype errors and non-Mendelian inheritances as described [80]. A total of 222 high quality SNPs, with an average call rate of 94.2% per SNP, were retained.
A F2 genetic linkage map was constructed using the est.map function in the qtl add-on package for R [93]. Recombination fractions were converted to map distances using the Carter-Falconer mapping function [42], [94]. We assumed no genotype error for map construction. Although a few base miscalls might have survived our data cleaning procedure, including a very small number of errors will have a negligible effect on map length estimation.
Multiple QTL mapping was performed using the forward/backward model selection algorithm implemented in the R/qtl command stepwiseqtl. Model fitting was performed via extension of Haley-Knott regression [95], with genotype probabilities calculated along a 1 cM grid. Model comparisons were conducted using a penalized LOD score approach with penalties calculated from 1000 permutations of the data [93]. Because biases may be introduced in the stepwise addition of new QTL to the model [93], we repeated the model search multiple times. Each search converged on an identical model of eight QTL and zero epistatic interactions.
|
10.1371/journal.pgen.1002400 | The FGFR4-G388R Polymorphism Promotes Mitochondrial STAT3 Serine Phosphorylation to Facilitate Pituitary Growth Hormone Cell Tumorigenesis | Pituitary tumors are common intracranial neoplasms, yet few germline abnormalities have been implicated in their pathogenesis. Here we show that a single nucleotide germline polymorphism (SNP) substituting an arginine (R) for glycine (G) in the FGFR4 transmembrane domain can alter pituitary cell growth and hormone production. Compared with FGFR4-G388 mammosomatotroph cells that support prolactin (PRL) production, FGFR4-R388 cells express predominantly growth hormone (GH). Growth promoting effects of FGFR4-R388 as evidenced by enhanced colony formation was ascribed to Src activation and mitochondrial serine phosphorylation of STAT3 (pS-STAT3). In contrast, diminished pY-STAT3 mediated by FGFR4-R388 relieved GH inhibition leading to hormone excess. Using a knock-in mouse model, we demonstrate the ability of FGFR4-R385 to promote GH pituitary tumorigenesis. In patients with acromegaly, pituitary tumor size correlated with hormone excess in the presence of the FGFR4-R388 but not the FGFR4-G388 allele. Our findings establish a new role for the FGFR4-G388R polymorphism in pituitary oncogenesis, providing a rationale for targeting Src and STAT3 in the personalized treatment of associated disorders.
| Several human cancers have been associated with increased growth hormone levels. Here we show that a frequent single nucleotide polymorphism (SNP) associated with increased cancer risk and progression also deregulates pituitary function. Through recruitment of a distinct STAT3 signaling cascade, this polymorphic receptor variant drives pituitary growth hormone cell survival and hormonal output. These findings provide an example of a potentially common genetic program shared between cancer and a hormone that promotes its progression.
| Pituitary tumors occur in almost 20% of the population [1] and represent nearly 10% of surgically resected intracranial tumors [2]–[3]. They can cause significant health problems due to abnormal hormone production and invasion into surrounding brain structures [2]–[3]. However, the mechanisms underlying the development of sporadic pituitary tumors that rarely involve mutations of classical oncogenes or tumor suppressor genes remain to be clarified [2]–[3]. Indeed, the only consistent molecular event reported thus far is activating mutations of the G-protein coupled Gsα that occurs in a subset of somatotroph adenomas [4]–[5]. Germline genetic abnormalities associated with pituitary tumor pathogenesis include inactivating mutations of menin in patients with Multiple Endocrine Neoplasia type 1 [6]–[7], loss of function mutations of the aryl hydrocarbon receptor-interacting protein (AIP) tumor suppressor gene in patients with familial isolated pituitary adenomas [8], and activating mutations the Protein kinase A type I regulatory subunit PRKA [9] in patients with Carney complex, however these alterations have not been shown to mediate pituitary neoplastic growth in the more common sporadic neoplasms.
Evidence suggests that epigenetically controlled growth signals implicated in pituitary development may be relevant to the tumorigenic processes in this gland [10]–[11]. Of note members of the fibroblast growth factor (FGF) and FGF receptor families have been proposed as candidate effectors, given their recognized importance in pituitary organogenesis [12]–[13]. FGF signaling is critical in pituitary development. Deletion of FGF10 or its receptor, the FGFR2 IIIb isoform, leads to failure of pituitary development [13]. Mid-gestational expression of a soluble dominant-negative FGFR results in severe pituitary dysgenesis [14]. FGF ligands are over-expressed in pituitary tumors. FGF-2, originally described in bovine pituitary folliculostellate cells, regulates multiple pituitary hormones and is over-expressed by human pituitary adenomas tumors [15].
We identified altered FGFR4 expression in pituitary tumors [16] due to expression of an N-terminally deleted isoform, pituitary tumor-derived FGFR4 (ptd-FGFR4) [17] generated by alternative transcription initiation from a cryptic promoter [18]–[19]. Prototypic FGFR4 (FGFR4-G388) is a 110 kD membrane-anchored protein expressed in several endocrine cells including the normal pituitary. In contrast, ptd-FGFR4 is a cytoplasmic protein expressed in pituitary tumors. The invasive tumorigenic potential of ptd-FGFR4, but not full length FGFR4, was demonstrated by targeted pituitary expression in transgenic mice [17]. The basis for the contrasting functions between these FGFR4 isoforms relates to their differential ability to associate with neural cell adhesion molecule (NCAM) and engage N-cadherin [20].
These studies were all carried out with the prototypic receptor prior to the identification of a single nucleotide polymorphism (SNP) that alters the coding region of the transmembrane domain. This germ-line polymorphism substitutes a glycine with an arginine at codon 388 of FGFR4, resulting in a charged amino acid in the highly conserved and normally hydrophobic transmembrane region of the receptor [21]. This FGFR4-R388 allele has been linked with advanced [21] and treatment-resistant breast cancer [22], prostate cancer [23], sarcomas [24], and head and neck carcinomas [25]. The mechanisms underlying FGFR4-R388 actions remain unclear. In this report we identify distinct signaling and hormone regulatory properties that distinguish FGFR4-R388 from the prototypic FGFR4-G388 form. The data unmask important patho-physiologic consequences of this common SNP with therapeutic implications for related diseases.
To determine if the FGFR4 polymorphic isoforms possess distinct functional properties in hormone-producing pituitary cells, we compared the effects of FGFR4-G388 and FGFR4-R388 on pituitary hormone production in rat GH4 mammosomatotroph cells that co-express prolactin (PRL) and growth hormone (GH) and in PRL235 cells that express PRL only. These GH4 and PRL235 cells express endogenous FGFR4 (Figure S1) and are homozygous for FGFR4-G385, the rodent equivalent of the human 388 site. Expression of human FGFR4-G388 or FGFR4-R388 to comparable levels shows that FGFR4-G388 enhances PRL and suppresses GH expression whereas FGFR4-R388 increases GH production with a reciprocal effect on PRL (Figure 1a, 1b).
To determine the effects of these FGFR4 isoforms on cell growth, stably transfected cells were plated in soft agar and examined for colony formation. GH4 cells expressing FGFR4-R388 were more efficient at forming colonies in soft agar compared with their FGFR4-G388 counterparts (Figure 1c; left). Enhanced colony formation resulting from FGFR4-R388 compared to FGFR4-G388 was also noted in PRL235 pituitary cells (Figure 1c; right).
The FGFR4-R388 substitution does not alter receptor kinase activity [21] and (data not shown). Thus, to examine signaling differences induced by the two FGFR4 isoforms in pituitary cells, we compared the ability of FGF to promote phosphorylation of the immediate FGFR substrate FRS2α. In contrast to FGFR4-G388 which showed ligand-dependent stimulation of this docking protein, FGFR4-R388 cells displayed enhanced FRS2α phosphorylation (Figure 2a). Src phosphorylation at Y416 was also appreciably higher in cells expressing FGFR4-R388 compared to those expressing FGFR4-G388 while Src phosphorylation at Y527 remained unchanged.
To examine the functional significance of this finding, we compared the ability of the Src inhibitor, dasatinib, to impede pituitary tumor cell growth and hormone production. Dasatinib effectively diminished Src phosphorylation (Figure 2b) and significantly inhibited colony formation in soft agar of cells expressing FGFR4-R388 (Figure 2c). By comparison, the less efficient colony forming FGFR4-G388 cells were relatively insensitive to the Src inhibitor (Figure 2c).
In contrast to the impact on cell growth, pharmacologic Src inhibition did not alter GH or PRL hormone expression (Figure 2b). Additionally, siRNA-mediated Src down-regulation did not significantly affect GH (Figure 2d) or PRL levels (data not shown). These findings suggested that while Src may play a role in driving FGFR4-R388-mediated cell growth, Src signaling may not be intimately coupled with pituitary hormone regulation in these cells.
STAT activation is implicated in mediating the effects of FGFR3 mutations associated with thanatophoric dysplasia [26]. We, therefore, examined the ability of the two FGFR4 isoforms to activate STAT signaling. Figure 2a depicts the differential impact of the FGFR4 isoforms on their ability to phosphorylate STAT3. While FGFR4-G388 supported ligand-induced tyrosyl phosphorylation of STAT3, this effect was not shared with FGFR4-R388, which instead resulted in sustained STAT3 serine phosphorylation at S727 (Figure 2a). STAT1 and STAT5 modifications were not affected by either FGFR4 isoform in GH4 or PRL235 cells (data not shown).
In contrast to the nuclear residence of pY-STAT3, pS-STAT3 translocates to the mitochondria where it has been implicated in cellular metabolism [27]. Thus, we performed immunofluorescence to localize pS-STAT3. Figure 3a identifies the mitochondrial residence of pS-STAT3 in FGFR4-R388; FGFR4-G388 cells are almost negative (upper panels). As controls, we expressed a constitutively active serine form of STAT3 (STAT3-S727D) in GH4 cells, and this also co-localized to the mitochondria (Figure 3a). In contrast, an inactive serine form of STAT3 (STAT3-S727A) failed to show a mitochondrial signal (Figure 3a). Subcellular fractionation followed by western blotting supported these findings with prominent mitochondrial expression of STAT3 and pS-STAT3 in FGFR4-R388 and in STAT3-S727D control but not in FGFR4-G388 cells in both GH4 and PRL235 cells (Figure 3b).
To examine the impact of pS-STAT3 on mitochondrial function, we measured Cytochrome C oxidase activity in pituitary cells expressing the different FGFR4 isoforms (Figure 3c). FGFR4-R388 cells which displayed higher pS-STAT3 levels also demonstrated higher Cytochrome C oxidase activity than FGFR4-G388 cells. In addition, lactate dehydrogenase (LDH) levels in FGFR4-R388 cell lysates were higher (2612 IU/ml) than those of the FGFR4-G388 (1197 IU/ml; n = 3).
To examine the functional impact of distinct STAT3 modifications on pituitary cells we compared growth in soft agar of cells expressing various STAT3 expression vectors (Figure S2). Of the STAT3 modifications, the active serine (STAT3-S727D) form displayed the greatest positive impact on colony formation (Figure S2).
To determine whether Src and STAT3 serine phosphorylation were inter-dependent, we examined the impact of pharmacologic Src inhibition. To this end, dasatinib-mediated inhibition of Src phosphorylation also reduced pS-STAT3 in FGFR4-R388 GH4 cells (Figure 2b). Moreover, siRNA-mediated Src down-regulation also resulted in diminished pS-STAT3 levels in these cells (Figure 2d). Additionally, treatment with the protein kinase C inhibitor G06983 reduced pS-STAT3 whereas the protein kinase A inhibitor H89 had no effect (data not shown).
To determine whether differential STAT3 responses were responsible for altered hormone gene expression, we compared the hormonal responses of FGFR4-G388 and FGFR4-R388 cells to a panel of growth factors. FGFR4-G388 cells with intact pY-STAT3 responses exhibited the expected GH inhibition in the presence of IGF-1 and related growth factors and the expected PRL increase in response to growth factors (Figure 4a). In contrast, and consistent with their attenuated pY-STAT3 responses, FGFR4-R388-expressing cells failed to suppress GH or mount a PRL response (Figure 4a).
In complementary experiments we down-regulated STAT3 in FGFR4-G388 cells. This forced STAT3 reduction resulted in increased GH and reduced PRL (Figure S3). Conversely, forced expression of STAT3 induced PRL and diminished GH expression (Figure 4b, 4c). Moreover, introduction of the dominant negative STAT3-Y705F diminished IGF-1 inhibition of GH (Figure 4b, 4c). The STAT3-Y705F mutant also failed to stimulate PRL expression further underscoring the requirement for pY-STAT3 in mediating PRL induction. In contrast to the impact of pY-STAT3, introduction of the constitutively active serine STAT3-S727D or the serine inactive STAT3-S727A did not alter GH or PRL responses (Figure 4b, 4c).
Taken together, these findings suggest that pY-STAT3, but not pS-STAT3, plays a more important role in pituitary GH and PRL regulation.
To determine if the observed cellular actions of FGFR4 polymorphism translate into biologically relevant actions on pituitary cell growth and function, we examined mice with knock-in (KI) of the mouse homologue of the polymorphism, Fgfr4-R385. Importantly, introduction of this SNP does not alter Fgfr4 expression levels [28] and (Figure S1). Systematic examination of the pituitary glands from mice carrying the Fgfr4-R385 allele at different ages identified the presence of pituitary tumors by 12 months of age. This revealed increased cellularity with loss of the reticulin network representing the hallmark of true neoplasia in this gland (Figure 5a–5d). Importantly, unlike the more common prolactinomas which are seen sporadically in aging mice [29], hormone staining identified GH (Figure 5e) but not PRL (Figure 5f) production by these tumors. Further, we examined pS-STAT3 in pituitary tissue from the knock-in mouse model. Fgfr4-R385 KI mice displayed strong immunoreactivity for pS-STAT3 (Figure 5g) which was not noted in control Fgfr4-G385 mice (Figure 5h). Double staining localized this pS-STAT3 in GH- immunoreactive somatotrophs (Figure 5i) but not in other cell types such as FSH-immunoreactive gonadotrophs (Figure 5j).
The frequency of these pituitary tumors and their morphologic phenotypes according to Fgfr4 genotype are summarized in Figure 5k. No GH-containing pituitary tumors were detected in control littermates (Figure 5k).
To corroborate the pituitary phenotypic abnormalities we compared circulating levels of the GH target growth factor IGF-1. Shown in Figure 5l is the positive impact of the Fgfr4-R385 allele on circulating IGF-1 levels. In contrast, and consistent with the in vitro data, mice carrying the Fgfr4-R385 allele did not demonstrate high PRL levels, and instead showed a tendency to lower concentrations compared to their Fgfr4-G385 littermates (Figure 5m).
Given the ability of FGFR4-R388 to facilitate pituitary tumorigenesis and GH production, we sought to identify evidence linking these two processes in human disease. We first examined STAT3 serine phosphorylation in human pituitary tissue. In the normal gland, immunohistochemistry for pS-STAT3 revealed strong reactivity in vascular endothelium (Figure 6a), providing an internal positive control; adenohypophysial cells and stroma were largely negative or showed only faint staining. All but one of 8 somatotroph adenomas exhibited strong positivity (Figure 6b) whereas lactotroph adenomas (n = 4) were negative or showed focal weak positivity (Figure 6c). Four of 6 gonadotroph adenomas and all but one of 7 null cell adenomas were also either negative or weakly positive (Figure 6d).
We next compared GH levels and pituitary tumor size in 64 patients with pituitary tumors and acromegaly based on their FGFR4 genotypic status. This examination identified a positive correlation between circulating GH levels (r = 0.622, p = 0.006) and pituitary tumor size in patients (n = 30) harboring an FGFR4-R388 allele. In contrast, patients homozygous for FGFR4-G388 (n = 34) showed no relationship (r = 0.23; p = 0.468) between GH levels and pituitary tumor size. Additionally, there was no relationship between FGFR4 genotype and tumor size in non-functional gonadotroph pituitary tumors (n = 22) or lactotroph adenomas (n = 13). These data support a selective link between the FGFR4-R388 allele, GH hormone production, and clinical pituitary somatotroph tumor formation.
The FGFR4-R388 SNP is known to promote breast cancer cell motility and invasiveness [21]. It has also been associated with accelerated cancer progression and treatment resistance [21]–[25]. However, the mechanisms underlying these actions remain unclear. We show here that FGFR4-R388 significantly alters pituitary function. Compared with the prototypic form of the receptor (FGFR4-G388), the polymorphic FGFR4-R388 variant supports distinct signaling to deregulate pituitary growth hormone production and cell growth in vitro and in vivo. In the mouse model we report, Fgfr4 expression levels are not altered [28], providing relevance to the human situation. Unlike the common sporadic pituitary lactotroph adenomas in rodents or the intermediate lobe corticomelanotroph pituitary tumors associated with several mouse models of cancer [29], the Fgfr4 SNP knock-in mice develop GH-producing pituitary tumors. The resulting GH/IGF-1 excess in these animals is potentially important in the enhanced breast cancer progression associated with this model [28].
FGFR signaling relies heavily on recruitment of the immediate substrate FRS2α through tyrosyl phosphorylation [30]. Using hormone-producing pituitary cells we show that compared to FGFR4-G388, FGFR4-R388 is associated with enhanced phosphorylation of FRS2α but not with the anticipated downstream MAPK activation. Instead, FGFR4-R388 signaling is accompanied by enhanced Src and STAT3 activation in pituitary cells. Consistent with this feature, pharmacologic Src inhibition results in greater growth inhibition by pituitary cells expressing FGFR4-R388. Neither pharmacologic inhibition nor Src knockdown, however, could alter the GH excess associated with FGFR4-R388. Instead, the FGFR4 SNP variant relies heavily on serine (S727) but not tyrosyl (Y705) phosphorylation of STAT3. These findings suggested that while Src may play a role in promoting cell growth, the observed hormone dysregulation was not intimately coupled with this putative oncogene.
The current study implicates multiple consequences of altered STAT3 modifications in the control of pituitary hormonal balance. As anticipated, FGFR4-G388 supports FGF-induced FRS2 phosphorylation to promote pY705-STAT3 activation. In turn, pY705-STAT3 induces PRL, as has been shown previously [31]. Conversely, the attenuated pY-STAT3 response, likely the result of pS-STAT3 [32], associated with the FGFR4-R388 SNP relieves GH from inhibition leading to higher expression of this hormone. STAT3 is a well-recognized mediator of cytokine signaling, and is known to regulate GH produced by the pituitary gland [27]. Interestingly, STAT5 which is also implicated in GH regulation [33] is not affected by this FGFR4 SNP in pituitary cells.
Pituitary auto-feedback mechanisms are candidate pathways whose interruption has become increasingly well-appreciated [2]–[3]. PRL receptor knockout mice develop pituitary lactotroph tumors [34]. Similarly, a somatic pituitary tumor-associated mutation in the extracellular domain of the GH receptor (GHR) disrupts N-terminal glycosylation of the receptor, thereby impairing GHR trafficking to the membrane, limiting ligand binding, and disrupting auto-feedback inhibition through diminished STAT activation [35]–[36]. Insulin and IGF-1 are growth factors that are known to exert negative-feedback at the level of the pituitary to inhibit GH [37]. In our study, insulin/IGF-1 efficiently activated pY-STAT3 to inhibit GH expression. In contrast, FGFR4-R388 failed to activate pY-STAT3 following ligand stimulation. The importance of diminished pY-STAT3 in mediating increased GH expression was further gleaned from knockdown of this STAT. STAT3 down-regulation or introduction of dominant negative STAT3-Y705F resulted in augmented GH production. Taken together, these data support the importance of STAT3 in the feedback inhibition control of GH regulation. In the bi-hormonal mammosomatotroph cell line examined, this altered signaling had the reverse effect on PRL. STAT3 down-regulation reduced PRL, whereas forced expression of STAT3 increased PRL and reduced GH expression by facilitating IGF-1 action. It is noteworthy that mice lacking pS-STAT3 have reduced IGF-1 levels [38] providing a complementary model to the increased IGF-1 noted in our FGFR4-R385 KI mice with pituitary somatotroph gain of pS-STAT3.
In contrast to the impact of FGFR4-G388 on pY-STAT3, FGFR4-R388 was associated with serine STAT3 phosphorylation (S727-STAT3). STAT3 is generally regarded as a requirement for Src-mediated cell transformation as shown in many carcinomas [39]. Traditionally, STAT3 oncogenic functions have been regarded to rely on pY-STAT3 and its nuclear translocation. However, more recently the positive impact of pS727-STAT3 on cell transformation has emerged. Unlike the tyrosyl modification, serine phosphorylated STAT3 has been also been described in the mitochondria [40] and negatively modulates tyrosyl phosphorylation [32]. Mitochondrial pS-STAT3 as shown in this study has been implicated in augmented electron transport complex and ATP synthase activity to yield higher lactate dehydrogenase [27], a critical metabolic requirement for transformed cells.
Previous studies have shown that pharmacologic inhibition of wild-type FGFR4 was not effective in arresting pituitary tumor xenografts [41]. Given our newly recognized FGFR4-R388 ability to preferentially activate Src in pituitary cells, we set out to re-examine the potential role of pharmacologic interruption on pituitary tumor-associated parameters. Using the Src inhibitor dasatinib [42], we demonstrate the ability of this agent to inhibit colony formation by FGFR4-R388 pituitary tumor cells. Given the recognized oncogenic actions of Src [43], and specifically in pituitary tumorigenesis as shown here, our data provide new insights into how this kinase might be an attractive therapeutic target in patients harboring the FGFR4-R388 SNP. It is equally plausible that inhibitors of STAT3 and those targeting oxidative phosphorylation may be of potential value in modulating pituitary tumors for therapeutic purposes.
In summary, we show that the heritable FGFR4-R388 allele yields a receptor variant that signals in a distinct manner from its prototypic FGFR4-G388 form in pituitary cells. Through its preferential ability to activate Src and pS-STAT3, FGFR4-R388 facilitates pituitary cell transformation. Further, the diminished ability to respond through pY-STAT3 results in attenuated negative feedback inhibition to augment pituitary GH expression. Given the recognized impact of FGFR4-R388 [21]–[25] and of the GH/IGF-I axis on cancer progression [44], the current findings identify the common FGFR4 polymorphism as an endocrine signal participating in these processes. It also highlights Src and STAT3 as potential targets for the treatment of patients with growth disorders in the context of the FGFR4 transmembrane polymorphism.
As there are no human-derived hormone-producing pituitary cell lines, we used rat pituitary GH4 mammosomatotroph cells which were propagated in Ham F10 medium 12.5% horse and 2.5% fetal bovine serum (FBS; Sigma, Oakville, ON), 2 mM glutamine, 100 IU/ml penicillin, and 100 µg/ml streptomycin (37°C, 95% humidity, 5% CO2 atmosphere incubation). Rat pituitary PRL235 lactotroph cells were propagated in DMEM 10% FBS, 2 mM glutamine, 100 IU/ml penicillin and 100 g/ml streptomycin.
Plasmids encoding human prototypic FGFR4 (G388) or the polymorphic form FGFR4-R388 were generated and stably transfected into GH4 and PRL235 cells as previously described [17]. Construct fidelity was confirmed by DNA sequencing after introduction into pcDNA3.1. STAT3 expression vector was kindly provided by M. Minden (University of Toronto), dominant negative STAT3-Y705F, inactive STAT3-S727A, or constitutively active STAT3-S727D were kindly provided by J. Chen (University of Illinois) [42]. Cells were transfected using Lipofectamine 2000 (Life Technologies, Rockville, MD) according to the manufacturer's instructions. Stable clones were selected using neomycin (G418) at a concentration of 0.7 µg/ml. A minimum of 3 clones of each isoform were pooled for further analyses in each of the cell types examined.
Oligonucleotides complementary to the gene of interest were synthesized by Ambion and introduced by transfection using lipofectamine 2000. Scrambled sequences of equal length were used as controls.
Ligand stimulations were performed on cells grown in 100 mm plate (4×106 cells/plate), pre-incubated as indicated for 1 hr or 24 hrs in serum-free defined medium (3 µg/ml putresine, 10-6 M hydrocortisone, 10-11 M tri-iodothyronine T3, and 0.375% albumin bovine factor V), without or with dasatinib (Sequoia Research Products Ltd, Pangourne, UK, 100 nM), H-89 (Calbiochem, San Diego, CA, 1 µM), or G06983 (Sigma, 1 nM). Treatments with IGF-1 (Sigma, 13 nM), insulin (Eli Lilly, 600 nM), FGF-1 (Sigma, 25 ng/ml) with heparin (10 U/ml), and EGF (R & D systems, 25 ng/ml) were based on earlier dose and time course studies ranging from 5 min up to 24 hrs.
Twenty five hundred cells were plated in 35 mm dishes as a single cell suspension in 0.3% agar in Ham F10 medium supplemented with 15% horse serum and 10% CS over under-layer of 0.5% agar prepared in Ham F10 as above. For Src inhibition, cells were incubated with the pharmacologic inhibitor dasatinib at concentrations ranging from 10−6 to 10−4 M. Colony formation was monitored daily with a light microscope and colonies photographed 4 weeks later as previously described [43].
Cells were lysed by mechanical homogenization for mitochondrial extraction using Qproteome Mitochondria isolation kit (Qiagen). Isolated fractions were analyzed by Western blotting to detect the MnSOD mitochondrial marker. Effective exclusion of contaminating cytoplasmic or nuclear proteins was confirmed by detection of tubulin and acetylated histone 3 respectively. Cytochrome C oxidase activity was used as a measure of electron transport chain activity according to the manufacturer's (Sigma) instructions.
Cells were lysed in lysis buffer (0.5% sodium deoxycholate, 0.1% sodium dodecyl sulfate, 1% Nonidet P-40 and 1× PBS) containing proteinase inhibitors (100 µg/ml phenylmethylsulfonyl fluoride (PMSF), 13.8 µg/ml aprotinin (Sigma), and 1 mM sodium orthovanadate (Sigma). Total cell lysates were incubated on ice for 30 mins, followed by micro-centrifugation at 10,000 g for 10 min at 4°C. Protein concentrations of the supernatants were determined by Bio-Rad method. Equal amounts of protein (50 µg) were mixed with 5× SDS sample buffer, boiled for 5 mins and separated by 8, 10, or 12% sodium dodecyl sulfate (SDS)-polyacrylamide gel electrophoresis, and transferred onto PVDF membranes (0.45 µm, Millipore, US). Intracellular and secreted hormones were determined using the following antibodies: polyclonal antisera to PRL or GH [donated by the National Hormone and Pituitary Program (NHPP), National Institute of Diabetes and Digestive and Human Development, Bethesdsa, MD] applied at dilutions of 1∶8,000 and 1∶50,000, respectively. Blots were incubated with polyclonal affinity–purified rabbit antiserum against the carboxy terminus of FGFR4 (Santa Cruz, Santa Cruz, CA). Immunoblotting was performed using anti-FGFR4 (Santa Cruz, 1∶1000), a monoclonal antibody to the V5-tag (Invitrogen, Burlington, ON), anti-MnSOD (Millipore, Billerica, USA), anti-FRS2α (R&D systems, Minneapolis, USA, 1∶1000), anti-Erk1/2 (Sigma, 1∶10000), anti-pFRS2α (Y196, 1∶1000), anti-pErk1/2 (1∶1000), anti-pY-STAT3 (Y705, 1∶1000), pS-STAT3 (S727, 1∶1000), STAT3 (1∶2500), anti-pSrc (Y416, 1∶1000), Src (1∶1000), anti-Tubulin (1∶1000), anti-acetylated Histone H3 (1∶3000) were purchased from Cell Signaling (Pickering, ON). Loading was monitored by detection of actin (1∶500, Sigma). Non-specific binding was blocked with 5% nonfat milk in 1× TBST (Tris-buffered saline with 0.1% Tween-20). After washing for 3×10 mins in 1× TBST, blots were exposed to the secondary antibody (anti-mouse or rabbit IgG-HRP, Santa Cruz) at a dilution of 1∶2000 and were visualized using ECL chemiluminescence detection system (Amersham, U.K.).
Cells were grown in 2 chamber slides and pre-incubated in serum-free define medium for 16 hrs. Cells were incubated with MitoTracker Red CMXRos (Invitrogen) at 37°C for 20 minutes, washed twice with PBS, fixed with 4% formaldehyde/PBS for 10 minutes, and washed three times with PBS. Cells were permeabilized for 10 minutes in PBS with 0.2% Triton X-100 and blocked for 30 minutes with PBS containing 5% FBS. Cells were first incubated with rabbit anti-STAT3 antibody (1∶100) or anti- pS-STAT3 antibody (1∶100) for 30 minutes at room temperature, washed three times with PBS, subsequently incubated with anti-rabbit IgG Alexa Fluor 488 for 30 minutes at room temperature, and washed three times with PBS. Coverslips were mounted in Fluoromount-G purchased from Electron Microscopy Sciences (Hatfield, PA) on glass slides. Cells were examined with Two-photon microscope (Zeiss LSM 510 META NLO), equipped with a 63× water-immersion objective lens and filters optimized for double-label experiments. Images were analyzed using the LSM IMAGE browser.
Fgfr4-R385 knock-in (KI) mouse were generated using standard approaches as described previously [28]. Mice were maintained on a pure C57BL/6 background. Genotyping was performed by PCR of genomic tail-DNA [28]. The care of animals was approved by the Institutional Animal Care facilities. Serum IGF-1 (Quantikine ELISA kit) and Prolactin (Calbiotech) levels were measured according to the manufacturer's protocols. Tissues were frozen in liquid nitrogen and stored at −70°C, or fixed in formalin and embedded in paraffin for histologic and immunohistochemical analyses. At least 6 animals were included at each time point for experimental measures.
Pituitary glands were stained with the Gordon-Sweet silver method to demonstrate the reticulin fiber network. Immunocytochemical stains to localize adenohypophysial hormones were performed as previously reported [17]. Primary polyclonal antisera directed against rat pituitary hormones were used at the specific dilutions: GH, 1∶2500; prolactin, 1∶2500; ß-thyroid-stimulating hormone (ß-TSH), 1∶3000; ß-follicle-stimulating hormone (ß-FSH), 1∶600; ß-luteinizing hormone (ß-LH), 1∶2500 (National Hormone and Pituitary Program, Rockville, MD); and adrenocorticotropin pre-diluted preparation, which was further diluted 1∶20 (Dako, Carpinteria, CA).
Human pituitary tissues were retrieved from the files of the University Health Network with REB approval. All had been fully characterized according to currently accepted criteria [45]. Fasting serum growth hormone levels were obtained from patients with histologically proven GH-producing pituitary adenomas. Pituitary tumor size was based on the maximal diameter noted on magnetic resonance imaging (MRI). FGFR4 germline genotyping was performed on DNA isolated from circulating while blood cells as described [21].
The care of animals was approved by the Institutional Animal Care facilities. Human pituitary tumors were retrieved from the files of the University Health Network with REB approval.
Data are presented as mean ± standard deviation (SD). In the experimental models, differences were assessed by the unpaired, two-sided t test. P<0.05 was considered statistically significant. The analysis of surgical human tumor specimens applied Fisher's exact test.
|
10.1371/journal.ppat.1004245 | Structure of the Trehalose-6-phosphate Phosphatase from Brugia malayi Reveals Key Design Principles for Anthelmintic Drugs | Parasitic nematodes are responsible for devastating illnesses that plague many of the world's poorest populations indigenous to the tropical areas of developing nations. Among these diseases is lymphatic filariasis, a major cause of permanent and long-term disability. Proteins essential to nematodes that do not have mammalian counterparts represent targets for therapeutic inhibitor discovery. One promising target is trehalose-6-phosphate phosphatase (T6PP) from Brugia malayi. In the model nematode Caenorhabditis elegans, T6PP is essential for survival due to the toxic effect(s) of the accumulation of trehalose 6-phosphate. T6PP has also been shown to be essential in Mycobacterium tuberculosis. We determined the X-ray crystal structure of T6PP from B. malayi. The protein structure revealed a stabilizing N-terminal MIT-like domain and a catalytic C-terminal C2B-type HAD phosphatase fold. Structure-guided mutagenesis, combined with kinetic analyses using a designed competitive inhibitor, trehalose 6-sulfate, identified five residues important for binding and catalysis. This structure-function analysis along with computational mapping provided the basis for the proposed model of the T6PP-trehalose 6-phosphate complex. The model indicates a substrate-binding mode wherein shape complementarity and van der Waals interactions drive recognition. The mode of binding is in sharp contrast to the homolog sucrose-6-phosphate phosphatase where extensive hydrogen-bond interactions are made to the substrate. Together these results suggest that high-affinity inhibitors will be bi-dentate, taking advantage of substrate-like binding to the phosphoryl-binding pocket while simultaneously utilizing non-native binding to the trehalose pocket. The conservation of the key residues that enforce the shape of the substrate pocket in T6PP enzymes suggest that development of broad-range anthelmintic and antibacterial therapeutics employing this platform may be possible.
| Here, we describe the structure of trehalose-6-phosphate phosphatase (T6PP) from Brugia malayi. This enzyme is essential to the organism; deletion of the gene encoding T6PP results in toxic accumulation of trehalose 6-phosphate. Structure-guided mutagenesis coupled with kinetic analyses revealed residues important for binding and catalysis. The model for substrate binding suggests a binding mode in which shape complementarity plays a major role. Conservation of binding residues among T6PP orthologs present in pathogenic nematodes and bacteria favors T6PP as a suitable target for broad-range anthelmintic and antibacterial drug design.
| Parasitic nematodes are responsible for devastating diseases that plague the tropical, low-income areas of Africa, Asia, and the Americas. In 2011, the World Health Organization estimated that 41% of the population worldwide was affected by these organisms [1]. Among the diseases caused by parasitic nematodes is the mosquito-transmitted lymphatic filariasis. It is projected that as many as 120 million people suffer from this disease with nearly 1.2 billion people at risk [2]. Lymphatic filariasis occurs upon infection of the lymphatic system by Wuchereria bancrofti, Brugia malayi, or Brugia timori and clinically manifests as lymphedema, hydrocele, and in the most extreme cases, elephantiasis. Related filarial nematodes inhabit other tissues; such infections may result in equally severe diseases, as exemplified by onchocerciasis or “river blindness” caused by Onchocerca volvulus infection. Filarial infections are responsible for extreme infirmity, distress, and social stigma. In fact, lymphatic filariasis is a major cause of permanent and long-term disability in people worldwide [3].
Due to the nature of infection and the impact on people suffering from this disease, the Global Program for the Elimination of Lymphatic Filariasis (GPELF) was established in 1999 with the major objective of ending the transmission of the disease by 2020. Currently the recommended regimen for treatment is the administration of albendazole together with either ivermectin (where onchocerciasis is endemic) or diethylcarbamazine citrate (where onchocerciasis is not present). Though community-wide treatment programs utilizing albendazole, ivermectin and/or diethylcarbamazine citrate have been effective, they are not without drawbacks. Side effects of albendazole and ivermectin are common, and although less frequent, they have also been observed with diethylcarbamazine citrate. Furthermore, these drugs only kill microfilariae, leaving the adult worms intact. Consequently, the drugs must be administered for the entire reproductive life span of the adult worm (approximately 5 years) [4]. In addition, the drug combination administered must be tailored to the specific parasite population in a given area because ivermectin administration can lead to encephalopathy in individuals with high microfilarial loads caused by Loa loa [5]. The development of new drugs that are not subject to these limitations is needed, particularly in the face of increasing drug resistance [6]. Indeed, drug resistance in veterinary nematodes is already widespread [7]–[9] and there are indications that in Ghana O. volvulus has developed a resistance to ivermectin [10].
To facilitate drug-discovery programs, the sequence determination of genomes of parasitic nematodes having human, domestic animal or plant hosts has been initiated [11]–[18]. The complicated life cycle of nematodes increases the difficulty of laboratory-based investigation. In fact, culturing W. bancrofti for in vivo studies has to date been unsuccessful. Fortunately, B. malayi can be maintained in a jird host [19] and is amenable to in vitro studies at different stages of its life-cycle [20]. Consequently, B. malayi now serves as a plausible model for research on lymphatic filarial nematodes alongside C. elegans, which for decades, has served as a model system for research on free-living nematodes. The similarities between C. elegans and parasitic nematodes with regard to genome sequences, and the phenotypes resulting from RNAi gene knockdown [21]–[23], indicate that C. elegans, being by far the most easily maintained organism, is an important resource in the development of broad-spectrum anthelminthic drugs.
Recently, top drug target candidates were identified in B. malayi using a ranking system [24], and among the highest-ranking targets is trehalose-6-phosphate phosphatase (T6PP) (UniProt: A8NS89), an enzyme that is required for the synthesis of trehalose [25]. T6PP is present in bacteria, fungi, plants, and invertebrate animals, but not in mammals. Trehalose is used by these organisms as an energy reserve, and it can also protect against environmental insults such as oxidative and osmotic stress, anoxia, heat, cold, freezing, desiccation, and anhydrobiosis [26], [27].
Trehalose is synthesized by a two-step pathway that involves T6PP and trehalose-6-phosphate synthase (genes gob-1, tps-1 and tps-2 in C. elegans, Figure 1). The synthase catalyzes the condensation of glucose 6-phosphate and UDP-glucose, forming trehalose 6-phosphate, and T6PP catalyzes phosphoryl transfer to water, forming trehalose. RNAi knockdown of the C. elegans T6PP-encoding gene gob-1 (gut-obstructed 1) gives rise to larval lethality due to intestinal blockage and subsequent starvation [28]. Importantly, this phenotype is reversed by RNAi knockdown of the tps-1 gene, suggesting that the lethality is due to a toxic accumulation of trehalose 6-phosphate [28]. A T6PP inhibitor might therefore bring about the same result, and thus we have targeted nematode T6PP for the development of small molecule anthelmintics.
As a first step toward inhibitor design the structure determination of T6PP was undertaken. The crystal structure of a putative T6PP has been reported from Thermoplasma acidophilum (PDB: 1U02) (29). Although this ortholog has low sequence identity, its structure identified it as a HAD superfamily (HADSF) phosphatase. All HADSF phosphatases possess a conserved Rossmann-fold catalytic domain, which contains the catalytic residues and the Mg2+ cofactor binding residues that together, constitute the substrate phosphoryl-group binding site. Most HADSF phosphatases, including the T6PP, also possess a cap domain (designated type C0, C1, C2A or C2B). During catalytic turnover the cap domain moves over the catalytic site through binding interactions with the substrate-leaving group, thereby forming an encapsulated active site. The size, shape and electrostatic properties of the active site are unique to each individual phosphatase. Although the sequences of the nematode T6P phosphatases are quite similar to one another, they share little identity with the T. acidophilum T6PP (12.7%). Moreover, sequence alignments revealed that the nematode orthologs possess a ∼140 amino-acid segment preceding the predicted N-terminus of the HAD phosphatase fold that is long enough to constitute a structural domain. It was thus both out of necessity for inhibitor design, and the intrigue for discovery of the fold and function of the novel N-terminal domain, that we pursued structure determination of the T6PP from B. malayi.
Herein, we report the X-ray structure of B. malayi T6PP and the findings from structure-activity analysis aimed at elucidating the role of the N-terminal domain and the active-site residues in substrate binding and catalysis.
The structure of B. malayi T6PP bound to Mg2+ was determined at 3.0 Å resolution (residues 63–491, Figure 2A, Figures S1 and S2). Electron density was not observed for residues 1–62 and SDS-PAGE analysis of dissolved crystals revealed that this region was not lost during crystallization. These residues are not conserved among the nematode T6PP enzymes, and in the case of several Caenorhabditis species, they are absent (Figure S3). In addition, the B. malayi T6PP truncation mutant Δ59 (residues 1–59 deleted) retains the native fold (Figure S4) and the same catalytic activity as the wild type enzyme (Table 1). Thus, we conclude that residues 1–62 are probably disordered and are not required for stability or catalytic activity.
The overall T6PP structure consists of a three-helix bundle N-terminal domain (residues 63–157) connected via a short linker (residues 158–205) to the Rossmann-fold catalytic core domain (residues 206–292; 398–492) and the α,β-fold cap domain (residues 293–397), which together comprise the HADSF phosphatase structure (Figure 2A). Six parallel β-strands flanked by five α-helices form the core domain. A short β-hairpin (residues 282–292) precedes the cap domain (residues 293–397) and extends the central six-stranded β-sheet of the Rossmann-fold by two strands. The C2B-type cap domain is formed by an anti-parallel β-sheet topped by two α-helices (Figure 2A). Two short loops in the cap (residues 320–322; 367–370) were not well ordered and were not modeled.
The closest related structure is the T6PP-like enzyme from T. acidophilum (which does not possess the N-terminal domain (residues 1–205; Figure 2B) [29]. Superposition of the B. malayi and T. acidophilum structures (Figure 2C) gave a three-dimensional alignment with an overall rmsd of 2.6 Å compared to the rmsd of 2.1 Å obtained for the overlay of the individual core domains (14% sequence identity) and 2.2 Å rmsd for the overlay of the cap domains (9% sequence identity). Although the tertiary structures of the paired domains are nearly identical, the difference in global conformation resulting from distinct cap-core domain orientations observed for the two crystalline states, prevented structure determination by molecular replacement, thus necessitating phasing by single wavelength anomalous diffraction of the selenomethionine-substituted B. malayi T6PP.
The N-terminal domain (residues 63–157) consists of a three-helix bundle which is similar in topology to the Microtubule Interacting and Transport (MIT) domains of the Vps4-like ATPases from Sulfolobus acidocaldarius (PDB ID: 2W2U, rmsd 2.4 Å, Figure 3A) and Sulfolobus solfataricus (PDB ID: 2V6Y, rmsd 2.4 Å) [30], [31]. The MIT-like domain is tethered to the core domain by a helical-connector (residues 158–206) that is comprised of a broken α-helix (helices 4 and 5), which interacts with both domains and a second, longer α-helix (helix 6), which selectively interacts with the core domain (Figures 2A,3B). The presence of the MIT-like domain is striking as no other HADSF members are known to incorporate this domain. Sequence alignment analysis revealed that that the MIT-like domain is unique to the T6PP of nematodes and in organisms from the genus Mycobacterium (Figure S5 and S6).
MIT domains are protein-interacting domains typically associated with multivesicular body formation, cytokinetic abscission, or viral budding [32]. The best-characterized MIT domains are found in the essential AAA-ATPase Vps4. Vps4 is vital for endosomal trafficking to lysosomes, where it acts to dissociate ESCRT (endosomal sorting complexes required for transport) from membranes [30]. The MIT domain of Vps4 recognizes a conserved sequence in ESCRT-III termed the MIT-interacting motif (MIM). This sequence forms a protein-protein interaction site between the second and third α-helices of the Vps4 MIT-domain. However, MIT domains can interact with MIM motifs at each helical interface in the three-helix bundle, and can interact with multiple MIM motifs at once [33]. In T6PP, an intramolecular interaction occurs at the interface between the first and third α-helices of the MIT-like domain and the C1-loop of the HAD core domain (Figure 3B), leaving the solvent-exposed interface between the second and third helices free to make intermolecular interactions with other proteins. Although the protein-protein interface is the same as that seen in other MIM-MIT domain interfaces [33] the sequence of the C1-loop of T6PP does not share significant identity with any known MIM motif. The availability of a potential protein-protein interaction interface on the MIT-like domain raised the question of whether T6PP has a “moonlighting” [34] function in addition to or in conjunction with its enzymatic phosphatase activity.
The T6PP gene was originally identified in C. elegans and was named gob-1 owing to the phenotype (gut obstruction) that occurs when the gene is knocked down. Knockdown of gob-1 as well as the upstream T6P synthase genes tps-1 and tps-2 results in a normal phenotype [28]. This suggests that the gob-1 lethality results from the buildup of the intermediate trehalose 6-phosphate, rather than the absence of trehalose (Figure 1). The presence of the MIT-like domain in the structure of T6PP led us to carry out further investigation. To test if the phenotype results from the absence of a protein-protein interaction between the T6PP MIT-like domain and another unknown protein, we repeated the RNAi experiments. Studies were carried out with the RNAi hypersensitive strain of C. elegans eri-1(mg366); lin-15B (n744). We observed that feeding dsRNA of the T6PP gob-1 gene results in arrest at the early larvae stage. In contrast, RNAi-hypersensitive worms which lack the T6P synthase gene tps-1 and which are fed gob-1 dsRNA showed a wild-type phenotype (Table S1). This observation is consistent with the earlier finding in wild-type C. elegans that the accumulation of T6P, rather than the lack of trehalose, is likely responsible for the observed lethality [28]. These results reinforce the suggestion that worm death is due to a metabolic effect, and that the gut obstruction phenotype is secondary to the accumulation of trehalose 6-phosphate. Moreover, the gene knock down lethality cannot be attributed to the removal of a hypothetical protein-protein interaction (through the MIT domain) because it occurs only if T6P synthase is present.
Deletion or mutation of the MIT-like domain is highly destabilizing to the enzyme and attempts to express the protein with the domain deleted (Δ179) or with the domain deleted and the potential hydrophobic patch “repaired” with the corresponding residues from the T. acidophilum T6PP (Δ179+L229Y/V232S/V236S/A243K) resulted in an unstable, insoluble form of the enzyme. The MIT domain itself was soluble and stable (T6PP:1-178-MIT Figure S7D). Notably, the loop region (C1-loop) that flanks the conserved active site forms extensive contacts with the MIT-like domain in the B. malayi protein, but the same C1-loop conformation is retained in the T. acidophilum ortholog lacking the MIT-like domain (Figure 2B). Thus, the interaction with the MIT-like domain is not necessary to retain the C1-loop conformation, although it is possible that in MIT-domain containing orthologs there is interdependence between the two structural units. Thermal unfolding observed by CD revealed one transition, suggesting that both domains unfold concurrently (Figure 3C). Additional proteolytic analysis confirms this observation (Figure S7) Together these findings suggest that the MIT-like domain and core domain are structurally co-dependent, and that any attempt to remove this interaction is detrimental to the overall stability of the protein. At this point it is uncertain whether the MIT-like domain plays a role in binding and catalysis of trehalose 6-phosphate hydrolysis.
The active site of HADSF members lies at the interface of the cap and core domains, with the core domain comprising the phosphoryl-transfer catalytic machinery. Indeed, the B. malayi T6PP core holds the four conserved active-site motifs that, in the HADSF [35], coordinate both the magnesium cofactor and phosphoryl group of substrate for catalysis. A single magnesium ion is observed to be coordinated in the active site by two waters and four residues (Motif I: Asp213, Asp215 (mainchain C = O oxygen), Motif IV: Asp424 and Asp428). Thr253 (Motif II) and Lys398 (Motif III) presumably form a hydrogen bond with the phosphate during catalysis.
In B. malayi T6PP, the cap was found to be in an open orientation relative to the core, though the crystal contacts were consistent with stabilization of the protein in this conformation and the protein may exist in alternate conformations in solution. Notably, the B. malayi T6PP has a slightly more open conformation than the orthologous T. acidophilum enzyme (Figure 2C). Because the cap domain mediates interactions with the substrate leaving group, but only the closed conformation positions substrate-binding residues for binding and catalysis, initial analysis of the structure did not reveal obvious residues involved in substrate recognition. Moreover, attempts to obtain a liganded structure or closed conformer structure have been unsuccessful thus far.
To identify residues important for ligand binding, structure-guided single site mutagenesis coupled with enzyme kinetics was performed. Several residues near the active site or in the cap region were selected for mutagenesis (Table 1, Figure 4). These residues are present on the face of the β-sheet that is oriented toward the active site and are highly conserved among T6PP enzymes from nematodes, but were not as conserved among T6PP enzymes from other phyla including Mycobacterium, other prokaryotes and Saccharomyces (Figure S6). As expected, replacement of the catalytic Asp213 or Asp215 resulted in the loss of all detectable activity (Table 1). None of the other replacements resulted in a dramatic change in Km for T6P. However, mutations in Tyr221, Gln332, Lys334, Asp336, or Arg337 resulted in large decreases in kcat (24-, 110-, 400-, 255- or 85-fold, respectively) suggesting a role in catalysis. Each of these residues except for Tyr221 are found in the cap domain (Figure 4) and may interact with the sugar moiety in order to orient the substrate for catalysis. Since the observed changes in kcat/Km stemmed from changes in kcat, we moved to inhibitor binding studies to test if any of the protein variants differed in their affinity for substrate.
To assess substrate affinity, the substrate analogue trehalose 6-sulfate (Materials and Methods) was used to measure KI for all of the T6PP variants (Table 1). Trehalose 6-sulfate was shown by steady-state inhibition kinetics to be a competitive inhibitor of wild type T6PP with a KI of 82±7 µM against trehalose 6-phosphate as substrate. Three of the variants that showed a decrease in kcat also showed dramatic increases in KI for trehalose 6-sulfate: Tyr221, Gln332 and Arg337 (Table 1, Figure 4). From these experiments, Tyr 221, Gln332, and Arg337 are likely to be involved in binding trehalose 6-phosphate, while Lys334 and Asp336 may play other roles in catalysis. These roles may include the exclusion of solvent from the active site, steric restraint of the substrate for catalysis, or the positioning of other residues required for enzyme activity.
To determine whether homologous residues could play a role in T6PP orthologs, homology models of the enzymes from E. coli, S. cerevisiae, and M. tuberculosis were generated [36] and compared to the structure from B. malayi and T. acidophylum (because sequence identity was low, a homology model which utilizes additional constraints was the most reliable way to ensure correct alignment). The residues that affected both catalysis and KI were found to be conservatively replaced in each of these orthologs (the bulky residue Tyr221 is replaced by Ile; Gln332 is replaced by Glu/Tyr; Arg337 is replaced by Lys) (Figure S8), reinforcing the importance of the roles of the residues identified here. The presence of these residues in all T6PP enzymes examined suggests that design of broad-spectrum inhibitors may be possible.
Attempts to visualize the closed-cap form of the B. malayi enzyme using molecular dynamics simulations were unsuccessful (see Materials and Methods). To provide a platform for inhibitor design, a model of T6PP with trehalose 6-phosphate bound in the active site was constructed. To accomplish this, trehalose 6-phosphate was manually placed in the active site of T6PP such that it was positioned for “in-line” attack by Asp213 (the Asp nucleophile) and coordinated to the Mg2+ cation as has been observed in all HADSF enzyme-substrate or enzyme transition-state analogue complex structures [37]–[40]. In addition, the side chain of Asp215 (the general acid/base catalyst conserved in phosphatase members of the HAD superfamily [35]) was positioned to promote leaving-group protonation. To position the trehalose moiety, the FTMap server [41], [42] was used to identify putative hot spots in the cap and core domains of both the B. malayi and T. acidophilum T6PPs. Intriguingly, FTMap analysis of the two enzymes revealed several hot spots forming a pocket extending from the phosphate-binding site to the C1 loop (Figures 5A,B). The identified binding pocket was juxtaposed to the important residues identified by mutagenesis in B. malayi in the cap (Figure 5B). The trehalose moiety was manually rotated to fit within the hot-spot binding sites and the model was then minimized using NAMD [43]. Tyr221, Lys 334, Gln332 and Arg337 are positioned near the substrate and, consistent with the mutagenesis results, may play important roles in binding (Figure 5C). Overall, the interface between the cap and core domains forms a substrate-binding pocket for trehalose 6-phosphate that may exclude solvent, provide steric constraint through van der Waals contacts, and may form only a few specific hydrogen bonds.
Although the T6P docked in the crystal structure of B. malayi T6PP brings the proposed binding residues proximal to the substrate, the model is not in a closed conformation. FTMap analysis suggests that the T. acidophilum T6PP enzyme is in a more closed conformation compared to that of B. malayi. Superimposition of the B. malayi cap with the T. acidophilum T6PP cap and subsequent analysis by DynDom [44]–[46] predicts that the cap rotates 45.6° with respect to the core (Figure S9A). Placement of the B. malayi cap in this orientation positions the residues identified by mutagenesis within contact distance of the predicted T6P model position (Figure S9B).
To compare our model of the T6PP-T6P complex to other phosphatases the DALI server [47] was used to identify HAD members with similar structures. Because the orientation between cap and core varies and the cap is the most variable portion of HAD family members, the cap domain from T6PP was used to find similar structures with C2B-type caps. Besides the T6PP-like enzyme from T. acidophilum, one of the highest scoring structures was that of sucrose-6-phosphate phosphatase (S6PP) from the cyanobacteria Synechocystis sp. PCC 6803 [48]. Sucrose plays a similar role in cells to that of trehalose, and is synthesized in response to osmotic stress in a two-step pathway involving the dephosphorylation of a sucrose 6-phosphate (S6P) intermediate [49], [50]. The structure of the S6PP-S6P complex (PDB: 1U2T) reveals a closed cap conformation with an extensive hydrogen-bonding network providing substrate stabilization (Figure 6A). The ability of these enzymes to discriminate between T6P and S6P may be due to the different glycosidic-linkages of sucrose and trehalose (sucrose: α(1→2)β vs. trehalose: α(1→1)α). According to our model, T6P packs against Tyr221 in the C1-Loop and may be stabilized for catalysis via hydrogen bonds from Gln332 and Arg337. S6P is oriented in the opposite direction and is stabilized by several hydrogen bonds from cap-residues (Figure 6A). The C1-loop in S6PP is much shorter and less pronounced, emphasizing its importance in T6P recognition by T6PP.
Overlay of S6P in the active site of T6PP reveals a steric clash with a loop in T6PP (Figure 6B). This is consistent with the fact that T6PP does not accept S6P as a substrate as well as the inability of S6P to bind to or act as an inhibitor of T6PP. In addition, the use of hydrogen bonds as a mode of sugar recognition in S6PP versus van der Waals interactions in T6PP explains the inhibition of S6PP by the monosaccharide glucose moiety comparable to that of sucrose itself [48]. In contrast in T6PP, we found that there is no inhibition by the trehalose or glucose 6-phosphate substituents of trehalose 6-phosphate (KI>10 mM for these compounds, Figures 7C, D). Thus, it is expected that a critical issue in designing inhibitory ligands for T6PP will be the occupancy of the phosphoryl binding site in conjunction with shape recognition of the leaving group (Figures 7A, B). Indeed, T6PP acts as an excellent system for the exploration of inhibitory ligands of phosphoryl binding sites as the majority of inhibitors to phosphoryl transfer enzymes [51] take advantage of adjacent binding sub-sites (e.g. the nucleotide or peptide/protein binding subsites in kinases [52], [53] rather than the transferring phosphate site.
In order to improve our understanding of the structure/function relationship in T6PP enzymes from parasitic nematodes we determined the structure of T6PP from B. malayi. Although the cap-open, unliganded form was observed in the structure, site-directed mutagenesis coupled with substrate and inhibition kinetics identified five residues (Table 1, Figure 4) important for catalysis and/or substrate affinity. Notably, no single residue is critical to binding, consistent with a mode of binding of the trehalose moiety by van der Waals contacts and shape complementarity between the sugar and enzyme. Moreover, the trehalose binding site is not dominated by hydrophobic residues that could form ring-stacking interactions like those found in other carbohydrate binding proteins [54]–[57]. The conservation of several key binding site residues among T6P phosphatases from parasitic nematodes and pathogenic bacteria indicates that a common strategy for T6PP inhibitor design might be used in the development of antibiotics as well as anthelmintics.
An inhibitor design strategy is now afforded by the availability of the T6PP structure. Based on inspection of the structure and the inhibition kinetic analysis of wild-type and mutant B. malayi T6PP, there are two independent regions that can be simultaneously targeted for inhibitor design, namely the phosphoryl group binding site and the trehalose leaving group binding site (Figure 7). The lack of inhibition of the B. malayi enzyme by trehalose (KI>10 mM) compared to the observed tight binding of the inert substrate analog trehalose 6-sulfate (KI = 82 µM) shows that electrostatic interaction with the phosphoryl group binding site provides a significant amount of binding energy that can be captured by a strategically positioned inhibitor sulfate group. Although the T6P trehalose moiety is essential for properly orienting the substrate in the catalytic site for catalytic turnover, it does not appear to provide the amount of binding energy needed for a lead inhibitor. Thus, we envision the design of a bi-dentate inhibitor comprised of the high affinity sulfate group (or similar) for interaction with the catalytic site of the core domain with an organic moiety that maximizes the intrinsic binding energy derived from hydrophobic and electrostatic interaction with the binding region made available by the interfaced cap domain (not necessarily mimicking those made by the substrate trehalose moiety).
Sequences of T6PP enzymes were downloaded from UniProt, aligned using T-Coffee [58] and visualized using ESPript [59]. A sequence alignment of all T6PP enzymes from nematodes revealed a high degree of conservation with the exception of the N-termini which vary in length (1–60 amino acids) and sequence (Figure S3). In addition, an extra N-terminal domain consisting of approximately 150 residues was found in T6PP enzymes from nematodes and in organisms from the genus Mycobacterium (Figs. S5 and S6)
The gene Bm1_08695 encoding T6PP from B. malayi was cloned from a cDNA library by PCR using primers with the restriction sites for NdeI and BamHI embedded. Standard cloning procedures were used to place the gene into a modified pET-15b vector (thrombin site replaced with a TEV site; pET-15(TEV)). The individual domains were also separately cloned into a pETDUET-1 vector by placing the HAD core (180–492) into MCS1 using XbaI and BamHI and placing the MIT domain (1–179) into MCS2 using NdeI and KpnI. To create variants for probing residues important to activity (Table S2), the Q5 Site-Directed Mutagenesis Kit (New England Biolabs, NEB) was used following the manufacturer's protocol. NEB 10-beta Competent E. coli (NEB 3019) were used for plasmid propagation and purification, and all plasmids were sequenced to confirm accuracy before use. For expression, T6PP, T6PP-Δ59, T6PP-Δ179, T6PP:1-178 (MIT) or the T6PP variants were prepared with T7 Express lysY/Iq Competent E. coli (NEB C3013) that were transformed with appropriate expression vectors. Flask cultures were grown in LB medium containing 0.1% glucose. Fermentation medium was as follows: 2% soytone, 1% yeast extract, 171 mM NaCl, 1.87 mM KH2PO4, 5.97 mM Na2HPO4, 13.1 mM NH4Cl, 0.5 mM K2SO4, 8% glycerol, 5 mM MgSO4, 10 mM betaine, 1× Korz trace metals [60], and 0.01% antifoam 204 (Sigma). For flask cultures 100 µg/ml of ampicillin was used and 150 µg/ml ampicillin was used in fermenters. One colony was inoculated into 7 ml of LB and 0.1 ml of the suspension was inoculated into 500 ml of LB for seed cultures. The seed cultures were incubated overnight at 30°C to an OD600 of 2. Ten liter Fermentations were done in 19.7 L Bioflo 510 fermenters (Eppendorf). The pH was controlled at 7.0 with automatic addition of 30% NH4OH and 46 N H3PO4. The dissolved oxygen was controlled at 20% of air saturation using constant gas flow at 0.5 vvm and an agitation-oxygen cascade. The culture was grown at 30°C to an OD600 of 4. IPTG was then added to 0.5 mM and the temperature was dropped to 18°C for induction. The culture was induced for 18 h before harvesting. Cells containing pET-15b (TEV):T6PP-Δ179 were induced with 1 mM IPTG for 3 hours at 37°C once the OD600 reached 0.6–0.8. Cells were pelleted and frozen at −20°C until further processing. For lysis, cell pellets were thawed and resuspended (3 ml buffer per g pellet) in a solution consisting of 50 mM Tris pH 7.5, 500 mM NaCl, 20 mM imidazole, 1 mM EDTA, 1 mM PMSF, 30 µg/ml DNase I (Sigma) and 1 protease inhibitor cocktail tablet (Sigma S8830) per every 100 g of cell pellet. Cells were lysed via a single pass through a microfluidizer (Microfluidics Model #M110P) at 20,000 PSI. Lysate was clarified at high speed (100,000× g), bound to a HisTrap FF column using an ÄKTA FPLC (GE Biosciences) and washed extensively before elution using a gradient with a buffer consisting of 50 mM Tris pH 7.5, 500 mM NaCl, 500 mM imidazole. The resulting protein fraction was then diluted 20-fold in a low-salt buffer consisting of 20 mM Tris pH 7.5 and 10 mM NaCl and bound to a HiTrap Q column (GE Biosciences). The protein was eluted using a gradient with a buffer comprised of 25 mM Tris pH 7.5 and 1 M NaCl. The N-terminal His-tag was removed by TEV protease (Blommel and Fox, 2007) using a 1∶50 ratio (mg∶mg) of TEV:T6PP prior to dialysis into a storage buffer consisting of 25 mM Tris pH 7.5, 10 mM NaCl, 5 mM MgCl2. Purified T6PP was concentrated to 15 mg/ml using an Amicon Ultra concentrator (10K MWCO, Millipore), aliquoted in small volumes and stored at −80°C.
To prepare selenomethionine-incorporated protein, the same procedure was used with two minor changes. Protein was expressed in the T7 Express Crystal Competent E. coli methionine-auxotroph cell strain (NEB C3022), following the manufacturer's protocol. Second, all buffers used contained 5 mM DTT to prevent oxidation of selenium.
Initial crystal screens were set up with either full-length T6PP or T6PP-Δ59 at 15 mg/ml in 96-well microplates (Corning 3552) using the Crystal Screen HT, PegRx HT and Index HT sparse matrix crystallization screens (Hampton Research) and a final drop size of 1.5 µl at 290 K. The full-length protein yielded hexagonal crystals in several low molecular weight polyethylene glycol (PEG) solutions buffered within a pH range of 5.0–8.5 and were optimized using the Additive HT and Detergent HT screens (Hampton Research). For data collection, optimized crystals of native T6PP were grown at 290 K in hanging drops (concentration of 15 mg/ml) in 33% PEG 300, 0.1 M sodium citrate pH 5.0, 13% ethylene glycol (EG) and 10 mM CoCl2. For cryoprotection, the concentration of EG was slowly increased to 25% by adding a concentrated solution directly to the crystal drop. Crystals were harvested using CryoLoops (Hampton Research), looped through LV CryoOil (MiTeGen), frozen in liquid nitrogen and stored until data collection. Crystals of selenomethionine-incorporated T6PP were grown at 290 K in hanging drops (concentration 10 mg/ml) in 30% PEG 300, 0.1 M sodium citrate pH 5.0 and 20 mM NDSB-256. For cryoprotection, 2-methyl-2,4-pentanediol (MPD) was slowly added to crystallization drops (final concentration of 25%). Crystals were harvested and stored in the same way as the native crystals.
Native and derivative datasets were collected at either the National Synchrotron Light Source (NSLS, beamline ×25) or the Stanford Synchrotron Radiation Lightsource (SSRL, beamlines 7-1, 12-2) using automounter systems. Data was collected under nitrogen gas flow at 100 K in 1° oscillations using either a Dectris Pilatus 6M detector (NSLS ×25; SSRL 12-2) or an ADSC Quantum 315r detector (SSRL 7-1) and processed with HKL-3000 [61]. Initial electron density maps were calculated via single-wavelength anomalous dispersion (SAD) with Phenix AutoSol using a 3.4 Å selenomethionine derivative dataset. This revealed 15 anomalous selenium peaks corresponding to 15 (out of 16) selenomethionine residues. Phases were improved and extended to 3.0 Å with Phenix AutoBuild using a native dataset [62]. The model was built iteratively with Coot [63] and refined in Phenix Refine [62] with 10% of the reflections excluded for the calculation of R-free. The final model converged to an R-value of 21.5% and an R-free of 25.9% with 93.2% of residues in the favored region and 6.6% of residues in the allowed region of the Ramachandran plot (Lovell et al., 2003). Full data collection and refinement statistics are summarized in Table S3.
The quality of the completed model was assessed by MolProbity [64]. Structural similarity to known folds was determined using either the PDBeFold (http://www.ebi.ac.uk/msd-srv/ssm) [65] or DALI (http://ekhidna.biocenter.helsinki.fi/dali_server/start) [47] servers.
The eri-1(mg366); lin-15B(n744) tps-1(ok373) worms were constructed by crossing tps-1(ok373) worms with eri-1(mg366); lin-15B(n744) worms. The genes lin-15B and tps-1 are closely linked on the X chromosome. Animals (n = 600) were genotyped to identify a recombination event that occurred between lin-15B and tps-1. Progeny (n = 600) of worms heterozygous for tps-1(ok373) and lin-15B (n744) were picked individually to a petri plate. After the worms generated progeny on the new plate the genomic DNA from each parent worm was isolated as previously described [66]. The genomic DNA of each worm was screened by PCR for homozygosity of the tps-1(ok373) deletion allele and heterozygosity of the lin-15B (n744) allele. Homozygosity for tps-1(ok373) was determined by the presence of the deletion mutation and absence of the wild-type gene. Additionally, the worms were tested for the presence of the lin-15B (n744) mutation by performing PCR of the lin-15B locus and digesting the product with BccI, which only cuts the mutant n744 mutant lin-15B allele. Only animals that were homozygous for the tps-1(ok373) deletion allele and heterozygous for lin-15B(n744) would had the recombination event between the two genes that was necessary for the eri-1(mg366); lin-15B(n744) tps-1(ok373) strain construction. Finally, the homozygosity for the eri-1(mg366) mutation was established by PCR and sequencing. L4 worms (n = 45) of each genotype were picked to individual plates with HT115 bacteria expressing gob-1 dsRNA [67]. The progeny of each animal was examined for defects.
To a solution of SO3/C5H5N complex (432 mg, 2.7 mmol) in freshly distilled pyridine (5 mL) was added a solution of the benzyl-protected trehalose (1.345 g, 1.35 mmol) in freshly distilled pyridine (5 mL) was added. The mixture was stirred for 5 h at room temperature, neutralized by addition of Na2CO3(aq) (1 M, 10 mL) and concentrated under reduced pressure. The residue was titrated with anhydrous methanol and the triturate was filtered. The filtrate was concentrated in vacuo giving a residue that was subjected to silica gel column chromatography (30∶1 MeOH/EtOAc) affording the protected trehalose 6-sulfate (1.35 g, 93%) as a white powder. A solution of the white powder (0.96 g, 0.89 mmol) in MeOH/H2O (1∶1,V/V, 40 mL), 20% Pd(OH)2/C (1.26 g) was repeatedly purged with hydrogen. The resulting mixture was stirred under 1 atm of H2 at room temperature for 24 h and filtered through a Celite pad. The filtrate was concentrated in vacuo giving a residue that was subjected to silica gel column chromatography (1∶4∶4, water/isopropanol/ethyl acetate) to give trehalose 6-sulfate as a white powder (366 mg, 93%). TLC (water/isopropanol/ethyl acetate, 1∶4∶4, v/v/v): Rf = 0.40; 1H NMR (D2O): 3.39–3.52 (2H, m), 3.60–3.67 (2H, m), 3.72–3.86 (5H, m), 3.97–4.0 2(1H, m), 4.23–4.26 (2H, m), 5.16–5.18 (2H, m); 13C NMR (D2O): 60.3, 66.7, 69.1, 69.5 70.0, 70.68, 70.74, 72.0, 72.19, 72.25; HRMS(ES) m/z: [M+Na+] calculated for C12H21O14S, 421.0652; found, 421.0648.
To determine the substrate specificity profile for T6PP, purified enzyme was diluted to 10 µM into a buffer consisting of 25 mM Tris pH 7.5, 50 mM NaCl, 5 mM MgCl2 and 1 mM DTT and aliquoted into 96-well plates in 5 µl volumes. Equivalent volumes of 20 mM substrate from a 167-compound in-house screen were then added and incubated for 45 minutes at room temperature. Next, 52.5 µl of BioMol Green (Enzo Life Sciences) was added to the wells to detect enzymatically released phosphate and incubated for another 30 minutes. After incubation, the absorbance of each well at 625 nm was measured using a Molecular Devices SpectraMax M5 microtiter plate reader. T6PP was found to be highly specific for trehalose 6-phosphate, eliminating the need to measure steady-state kinetic constants for other substrates.
For steady-state kinetic characterization, purified T6PP and the variants made by site-directed mutagenesis were diluted to between 25–75 nM into a buffer consisting of 25 mM Tris pH 7.5, 50 mM NaCl and 5 mM MgCl2. The steady-state kinetic parameters (Km and kcat) for trehalose 6-phosphate were determined from initial reaction velocities measured at varying concentrations (0.0625–5 mM) using the EnzCheck Phosphate Assay Kit (Invitrogen). Absorbance measurements were performed with a Beckman DU800 spectrophotometer using quartz cuvettes (Starna Cells, 18B-Q-10) and 250 µl volumes. Data were fit to the following using SigmaPlot Enzyme Kinetics Module:Here, vo is the initial velocity, vmax the maximum velocity, [S] the substrate concentration and Km the Michaelis-Menten constant calculated for trehalose 6-phosphate. The value for kcat was calculated from the following:where [E] is the protein concentration in the assay. The steady-state kinetic constants for T6PP and its variants made via site-directed mutagenesis are reported in Table 1.
The steady-state competitive inhibition constant KI was determined for trehalose 6-sulfate by fitting the initial velocity data, measured as a function of trehalose 6-phosphate (0.5 Km to 5 Km) and inhibitor (0, KI, 2 KI) concentration to the following using SigmaPlot Enzyme Kinetics Module:where [I] is the inhibitor concentration and KI is the inhibition constant.
To generate homology models of related T6PP enzymes, FASTA sequences for E. coli (UniProt: E8Y507), S. cerevisiae (A6ZY39), and M. tuberculosis (H8EZ37) were uploaded to the Protein Homology/analogY Recognition Engine V 2.0 (Phyre2) server, where models were automatically generated using the best possible template model [36]. In this case, each model was based on 1U02 with a >95% confidence level. To model the putative MIT-like domain in M. tuberculosis, one-to-one threading on the Phyre2 server was used utilizing the structure of T6PP from B. malayi as the template.
To map substrate binding residues in the HAD domain, the FTMap server (http://ftmap.bu.edu) was used to identify hot spots. To map putative MIT interactions, PDB files consisting of only the MIT domains (residues 63–156 for B. malayi; 2W2U chain A; 2V6Y chain A) were uploaded and analyzed in PPI mode. To map putative substrate binding residues, PDB files (residues 63–491 for B. malayi; 1U02 chain A) were uploaded and analyzed using default settings.
To model the T6PP-T6P complex, trehalose 6-phosphate was built (α-D-glucopyranosyl-(1→1)-α-D-glucopyranoside 6-phosphate) and manually placed in the active site of T6PP. The phosphate was placed such that it was positioned for “in-line” attack by Asp213 (the Asp nucleophile) and coordinated to the Mg2+ cation. In addition, Asp215 (Asp+2, general acid/base residue) was positioned to promote leaving-group protonation. The sugar was manually rotated to lie against the C1-loop guided by mutagenesis and FTMap hot spot results. The model was then minimized using NAMD [43] and analyzed using UCSF Chimera [68].
Hydrogen atoms were added to the docked T6PP-T6P model, using the HBUILD function in CHARMM [69] and the coordinates gently minimized in general Born implicit solvent (GBIS) [70] by gradually removing harmonic restraints on the system heavy atoms and performing 25 steps of steepest descent followed by 250 to 500 steps of advanced basis Newton-Raphson (ABNR) steps. A final, unrestrained minimization using ABNR steps was performed until the energy gradient stabilized below 10−5 kcal/mol-step. A further 200 steps of minimization was performed in NAMD [43] and rigid bonds. GBIS solvation was employed with an effective ion concentration of 150 mM and a 12 Å cutoff for calculating the Born derivatives. All other parameters were set to NAMD defaults. Surface tension terms in the solvation model were ignored. A non-bonded cutoff distance of 15 Å was employed with a switching function employed to 16 Å. Langevin dynamics was initiated at 300 K utilizing a damping coefficient of 1/ps, a 2 fs timestep and rigid bonds enforced. This configuration was integrated for a total of 1.2 ns.
Coordinates for T6PP and structure factors have been deposited in the Protein Data Bank (PDB ID: 4OFZ).
|
10.1371/journal.pgen.1004050 | The Midline Protein Regulates Axon Guidance by Blocking the Reiteration of Neuroblast Rows within the Drosophila Ventral Nerve Cord | Guiding axon growth cones towards their targets is a fundamental process that occurs in a developing nervous system. Several major signaling systems are involved in axon-guidance, and disruption of these systems causes axon-guidance defects. However, the specific role of the environment in which axons navigate in regulating axon-guidance has not been examined in detail. In Drosophila, the ventral nerve cord is divided into segments, and half-segments and the precursor neuroblasts are formed in rows and columns in individual half-segments. The row-wise expression of segment-polarity genes within the neuroectoderm provides the initial row-wise identity to neuroblasts. Here, we show that in embryos mutant for the gene midline, which encodes a T-box DNA binding protein, row-2 neuroblasts and their neuroectoderm adopt a row-5 identity. This reiteration of row-5 ultimately creates a non-permissive zone or a barrier, which prevents the extension of interneuronal longitudinal tracts along their normal anterior-posterior path. While we do not know the nature of the barrier, the axon tracts either stall when they reach this region or project across the midline or towards the periphery along this zone. Previously, we had shown that midline ensures ancestry-dependent fate specification in a neuronal lineage. These results provide the molecular basis for the axon guidance defects in midline mutants and the significance of proper specification of the environment to axon-guidance. These results also reveal the importance of segmental polarity in guiding axons from one segment to the next, and a link between establishment of broad segmental identity and axon guidance.
| During nervous system development, once formed from neuroblasts, neurons grow axons in the direction of their synaptic partners. Genetic forces guide these axon growth cones towards the target. This is known as axon guidance or pathfinding. There are a number of proteins that regulate axon-pathfinding. The well-known examples are the Slit and its receptor Roundabout, and Netrin and its receptor Frazzled. The Drosophila embryo and the nervous system are divided into segments by segmentation genes. Mutations in segmentation genes affect axon guidance, although how they do so is not well understood. In our work described here, we show that the T-box protein Midline prevents mis-specification of neuroblast rows, in particular, it prevents row 2 from becoming row 5. Thus, in midline mutants, row 2 changes into row 5, ultimately creating a non-permissive barrier that prevents axons from following their defined path. Thus, axons stop and diverge when they reach this barrier. Our results show how mutations in segmentation genes can affect axon guidance and how significant the environment is for axon-pathfinding. Our work is also a cautionary reminder that guidance defects need to be interpreted with care and can arise due to a variety of other defects.
| In the Drosophila nerve cord, about 20 longitudinal axon tracts on either side of the midline, each consisting of axons from several neurons, connect different segments with one another. Several direct players in axon guidance have been identified. For example, previous studies have shown that mutations in two signaling pathways, the ligand Slit (Sli) and its receptors Roundabouts (Robo, Robo2 and Robo3) and the ligand Netrin (Net) and its receptor Frazzled (Fra; the vertebrate homologue is known as Deleted in Colorectal cancers or DCC) disrupt the precise positioning of these tracts by altering their growth cone guidance [1]–[7]. Whenever the Slit system is disrupted, longitudinal axon tracts inappropriately cross the midline [1], whereas with the disruption of the Net-Fra system, which primarily mediates the attraction of commissural tracts to facilitate their midline crossing [4]–[6], a large number of commissural growth cones fail to cross the midline [4], [5], [8].
There is a second set of players not linked to the direct players such as Slit-Robo or Net-Fra, but cause axon guidance defects when disrupted. In these mutants, the pioneering axon growth cones fail, either due to the absence of the neurons themselves or due to a mis-specification of their identity. As a result, follower neurons fail to properly project their growth cones along the correct trajectories. For instance, when the pioneering neurons pCC or vMP2 are either ablated [9] or mis-specified [10], the follower axon tracts cross the midline, ignoring the guidance cues mediated by Slit and Robo [10].
It is obvious that the environment in which growth cones travel would have an impact on axon guidance. However, it is not clear in what specific way the environment in which axons travel influence axon guidance or how specific the influence would be on axon guidance. The environment is defined by cells, which express guidance determinants on their surface or release cues into the extracellular matrix. Segmentation genes, in particular segment polarity genes, broadly define the environment in which axons travel by specifying cellular identity, which then by expressing specific genes regulate guidance of specific growth cones. Segment polarity genes are expressed in rows and columns within the nerve cord and mutational analysis indicates that they specify the initial NB identity along the rows and columns [11]–[15]. For instance, row 5 identity is set mainly by the expression of Wg and Gsb (all row 5 cells express these genes), whereas row 4 is determined by the expression of Patched (Ptc) in row 4, Wg in row 5, and the absence of expression of Gsb in row 4 [reviewed in ref. 15]. Loss of function for these genes alters the identity of NBs along the entire rows. Thus, loss of function for Ptc changes row 4 into row 5, loss of Gsb changes row 5 into row 4, and loss of Wg alters rows 5, 6 and 4 identities (non-cell autonomous function of Wg also confers row identity to adjacent rows) [11]–[15]. Their expression persists in successive divisions of NBs, even as NB-specific expression of transcription factors changes following each division of a NB [16], [17].
Loss of function for these genes also cause axon guidance defects [18]. However, we do not know if the axon guidance defects in segmentation mutants are due to mis-specification of a pioneering neuronal identity, or broad changes in the environment in which axons travel (or both). Given that growth cones interact with the environmental niche along their path, it is reasonable to suppose that broad changes in the local environment can affect axon pathways. However, separating neuronal identity from changes in the environment in influencing axon guidance has been experimentally difficult.
We have been studying a gene called midline (mid), which belongs to a class of transcription factors known as T-box binding (Tbx) proteins. Tbx proteins are highly conserved among metazoans and are defined by the presence of a T-box domain, a 180–230 amino acid DNA-binding domain. Tbx proteins bind to a T-Box element (TBE), a 20-bp degenerate palindromic sequence [19]. However, TBEs are highly variable in sequence, number and distribution within promoters and Tbx proteins diverge significantly in their sequence preference [20]. Tbx proteins are known to repress transcription [21]. Moreover, mutations in Tbx genes can be haploinsufficient, i.e. developmental processes are sensitive to the levels of some Tbx proteins. For example, upper limb malformation and congenital heart defects in Holt-Oram syndrome are due to haploinsufficiency for TBX5 [22], [23]. Haploinsufficiency for mouse brachyury and human TBX3 and TBX1 genes causes dominant phenotypes such as short tails/tailless, Ulnar-Mammary syndrome and DiGeorge syndrome, respectively [23], [24].
In Drosophila, loss of function for mid (also known as lost in space or los, or extra) was initially shown to cause cuticle defects in the midline region of the embryo, thus the name midline [25]. Subsequently, it has been shown that mid mutants also cause heart defects [26], defects in the lateral chordotonal axons, and shorter and defasciculated dorsally routed axons in the peripheral nervous system (PNS) [5]. We recently showed that Mid ensures ancestry-dependent fate specification of a GMC, i.e, fate of a GMC is changed without affecting the parent NB identity, thus overriding the GMC's ancestry [20]. Thus, in mid mutants, a GMC from an unrelated NB (we have named it the M-lineage, M for Mid) changes into GMC-1 (also known as GMC4-2a) of the RP2/sib lineage without altering the parent NB identity. Also, this occurs several hours after the window of time in which the bona fide GMC-1 of the RP2/sib cells is formed.
A subsequent study by Liu et al. [27] reported that ectopic expression of mid in salivary gland can ectopically induce expression of robo, slit, Netrin and frazzled. The implication is that Mid regulates axon guidance via regulation of these guidance genes and that the axon guidance defects observed in mid loss of function mutants are due to loss of expression of these guidance genes. However, the regulation of these genes by Mid in salivary glands, where none of these axon guidance genes including mid are normally expressed, is of no functional significance. Mid must regulate these genes in the nerve cord to be of relevance. Moreover, the mostly non-overlapping expression patterns of mid, robo and slit in the developing CNS, save a few cells in the lateral region of the nerve cord as reported by Liu et al [27] does not make sense with a functional direct transcriptional regulatory role for Mid on these genes during axon guidance.
We sought to explore these issues, including the possibility of an indirect regulation of axon guidance genes by mid, with the aim to understand the molecular basis for the guidance defects in mid mutants. We show here that the primary axon guidance defect in mid mutant embryos is stalling of axon tracts midway between the posterior commissure (PC) and the anterior commissure (AC) of the next segment, with tracts often crossing the midline, or projecting peripherally outward, perpendicular to the midline. This defect is due to the transformation of row 2 NBs and their precursor neuroectodermal (NE) cells, which are located midway between the PC and the AC of the next segment, into row 5 cells. Row 5 is normally located at the level of the PC and defines the parasegmental boundary (PSB). The fact that axon tracts stall or project across the midline or towards the periphery precisely along this transformed row, indicates that these newly re-specified row 5 cells creates an unsuitable or inhibitory niche for these pioneering axons to navigate along the midline. These results argue that the role of Mid in regulating axon guidance is indirect and via proper specification of row identity within the nerve cord. Our results also show that Mid does not regulate transcription of frazzled, sli or robo, directly or indirectly, in cells where their expression matters. These results provide novel insight into how segmentation or row identity facilitate axon guidance later in neurogenesis and distinguishes how broad environmental identities, as opposed to individual neuronal identity, influence axon guidance.
Previous results have indicated that embryos mutant for mid show axon guidance defects [5]. We sought to examine in detail the axon guidance defects in mid mutants in the embryonic CNS during development and compare those defects to the defects at corresponding developmental stages in slit and robo mutants. As shown in Fig. 1, embryos of different developmental stages were stained for Fasciclin II (Fas II) positive axon pathways using an antibody against Fas II. Fas II staining of ∼9 hours old embryos reveals the nascent medial tract, which is closest to the midline and is pioneered by the growth cone from pCC (arrows in Fig. 1A, wild type). In ∼9 hours old mid deficiency embryos the pCC growth cones were the same as in wild type, projecting slightly outward and then parallel to the midline (Fig. 1D, arrows). However, in ∼9 hours old slit mutant embryos, the pCC growth cones projected directly towards the midline (arrows in Fig. 1G). In ∼9 hours old robo mutant embryos, the pCC growth cones also projected towards the midline, although the defects were less severe than in slit embryos (Fig. 1J, arrows).
By 10 hours of development, the growth cones from pCC in mid embryos were projecting outward and away from the midline as if they had encountered an inhibitory zone (Fig. 1E, arrows, compare with wild type, Fig. 1B), whereas in ∼10 hours old slit mutant embryos, the growth cones from pCC were fasciculated with each other at the midline (Fig. 1H). By ∼14 hours of development in mid embryos, the three different Fas II tracts, the medial tract, the intermediate tract and the lateral tract, all run parallel to the midline, could be seen with similar spacing between each other and from the midline, as in wild type (Fig. 1F to 1C). However, as shown in Fig. 1F, in ∼14 hours old mid mutant embryos we could observe tracts inappropriately projecting outward (thick arrow), breaks or missing tracts along the longitudinal axis (arrowhead) and crossing the midline (midline arrow). We could also observe stalled growth cones forming a blob of axon tracts along the nerve cord (Fig. 1C, star, see also Table 1).
While in ∼14 hours old slit mutant embryos the three tracts were all collapsed at the midline (Fig. 1I), in robo mutants, the medial tract was mostly fused at the midline, with the other two tracts more or less normal (Fig. 1K)(the partial penetrance of the guidance defects is due to redundancy with Robo2 and Robo3 receptors) [2], [3], . The frequency of various guidance defects in mid, slit and robo mutants are presented in Table 1. These results indicate that axon guidance defects in mid mutants are significantly different from axon guidance defects in slit and robo mutants. If Mid regulates axon guidance via regulating slit and robo, the guidance phenotypes in all the three mutants should fall more or less into the same general category. Our above results show that this is not the case and argues against the possibility that Mid regulates slit and robo and that the axon guidance defects in mid mutants are due loss of function for these axon guidance genes.
We next sought to determine the growth cone projections from vMP2, dMP2, MP1, pCC and aCC neurons in mid mutant embryos using more selective markers. We chose to examine the growth cones from these neurons since these neurons send out pioneering growth cones. For example, the anteriorly projecting growth cones from vMP2 and pCC pioneer the medial Fas II tract to meet the homologous axons from the next anterior segment [9], [10]. Similarly, the posteriorly projecting growth cones from MP1 and dMP2 pioneer the lateral Fas II tract to meet up with the homologous axons from the next posterior segment. Therefore, first we stained mutant embryos with a monoclonal antibody 22C10, which is raised against MAPIB. In a ∼10 hours old embryo, vMP2 (Fig. 2A) projects its growth cone anteriorly (arrow), while dMP2 projects posteriorly (arrow)(Fig. 2A). By ∼11.5 hours of development, 22C10 antibody staining revealed a fasciculated, more mature medial tract (Fig. 2B, upper arrow) and lateral tract (lower arrow), as well as several other axon pathways including the motor pathway of the aCC and RP2 neurons, both fasciculated together to form the intersegmental nerve bundle before exiting the CNS (smaller arrow).
In mid mutant embryos, both vMP2 and dMP2 neurons are normally formed, but we observed two key defects in their projection pattern: the growth cones often projected away in a posterior-lateral pathway similar to and/or sometimes part of the aCC-pioneered intersegmental nerve bundle (Fig. 2C, top, left arrow with star). The projections were either stalled or projected away like a motor pathway (Fig. 2C, D, E, arrow and arrow with a star). These aberrant projection patterns suggest that these growth cones have come upon a non-permissive barrier or a zone of repulsion and cannot travel in their normal path. They either stall and or project away. We further examined the projection pattern from vMP2 by expressing mCD8-GFP (mCD8 targets GFP to membrane) using the achaete (ac)-GAL4 driver. While in the wild type the axon tract from vMP2 is projected along the midline (Fig. 2F, arrow), in the mutant the projection is diverted away and perpendicular to the midline in a pathway towards the periphery, often exiting the nerve cord (Fig. 2G, arrow with star).
We next examined the projection from MP1 by expressing tau-GFP (tau directs GFP to microtubules) using the sim-GAL4 driver. While in the wild type the axon tract from MP1 is projected along the midline (Fig. 2H arrow), in the mutant the projection is diverted towards the periphery, perpendicular to the midline (Fig. 2I, J, arrow with star). This aberrant projection defect in MP1 was highly penetrant and severe.
We next examined the growth cone projection from pCC by expressing UAS-tau-lacZ transgene in pCC neuron using the RN2-GAL4 driver. This driver drives the tau-lacZ in pCC neuron (Fig. 2K, L; it also drives in aCC and RP2, Fig. 2M, N). As shown in Fig. 2K, in the wild type the pCC projects its axon anteriorly along the midline (arrow). However, in the mutant, the projection is diverted away towards the periphery perpendicular to the midline (Fig. 2L, arrow with star). We also examined the two motor pathways from neurons aCC and RP2, but did not observe any defects in their pathfinding (Fig. 2N). These results indicate that the defects are mostly confined to axon tracts from interneurons. These defects are unlikely due to a negative effect on axon growth, instead, the projections appear to encounter a barrier in their normal path and travel in an aberrant path as defined by this barrier.
The above results show that the axon guidance defects in mid are fundamentally different from those in slit or robo mutant embryos. However, given the recent report that Mid ectopically regulates sli and robo transcription in salivary glands [27], we sought to examine mid mutant embryos for the expression of these genes in cells where they are normally expressed. If one of the functions of Mid in wild type is to regulate expression of slit and robo genes, a significant reduction in the levels of Sli and Robo proteins should be observed in their respective domains in loss of function mid mutant embryos. First, we stained mid mutant embryos with a Slit antibody. As shown in Fig. 3 (A, C and E) in wild type, Slit is present at high levels in midline glial cells where the axon tracts of AC and PC cross the midline. It is also present in commissural and longitudinal tracts due to movement of Slit from the midline to the axon tracts via the commissural tracts [7]. We examined the two alleles of mid (mid1 and los1)(Fig. 3B and D) and the mid H15 deficiency (which removes both mid and its sister gene H15)(Fig. 3F). We have reported previously [20] that the mid1 allele has a stop codon at amino acid (aa) position 128 (this allele is likely the strongest loss of function mid mutant allele) and los1 has a 22 base-pair deletion resulting in a deletion of 7 aa at position 321, as well as a frame shift leading to a stop codon at aa 350 (thus, in addition to the truncation the mutant protein in this allele has 28 amino acids that are entirely different from the wild type gene; this has the potential to cause gain of function/neomorphic effects in addition to loss of function effects).
The Slit protein level was not significantly affected in homozygous mid1 allele nor was it affected in the homozygous mid H15 deficiency embryos (Fig. 3B and 3F); a marginal reduction in the levels of Slit protein was observed in los1 embryos in the PC region (Fig. 3D). Whether this is due to a los1-specific gain of function effect or a background effect is not known. We further examined if the levels of the Slit protein is affected in younger stage embryos from mid1, los1, and the mid H15 deficiency. However, no reduction in the levels of the Slit protein was observed in these alleles (data not shown). To quantify the level of Slit between wild type and the mutant embryos, we performed Western analysis of Slit in the mid H15 deficiency embryos. The results reproducibly showed only a marginal reduction in the amount of Slit (Fig. 3G). One possibility for this slight reduction in Slit protein levels is that Liu et al [27] had reported that there is an overlapping expression of Mid and Slit in a few neurons located laterally within the nerve cord. It is possible that Mid regulates slit expression in these cells and that the slight reduction on Westerns reflect this regulation. Alternatively, the slight reduction in the levels as seen in Western blots is due to indirect effect of loss of function for mid and H15 genes, such as mis-specification of relevant neurons/glia.
Since Mid is a transcription factor, we next sought to determine if the transcription of the slit gene is affected in mid mutant embryos by performing whole mount slit RNA in situ. If Mid regulates slit transcription at least in the PC region, where mid is expressed, we should observe loss of slit transcription in these midline cells in mid mutant embryos. However, as shown in Fig. 3J, K and L), no such effect on the transcription of the slit gene by loss of function for mid was observed.
We next examined the expression of Robo in mid1, mid H15 deficiency, and in embryos transheterozygous for the mid H15 deficiency and mid1 alleles using an antibody against Robo (Robo levels were also examined in los1 allele, see later section). As shown in Fig. 4A, in wild type Robo is expressed in longitudinal pathways and is also present very weakly in AC and PC (due to incomplete down-regulation of Robo by a Commissureless protein-mediated process in commissural tracts [1]). In mid mutant embryonic CNS, the levels of Robo was not affected in any significant way (Fig. 4B); the lack of Robo staining in tracts (arrows, Fig. 4B) is due to the absence of axon tracts themselves. We also examined the expression of Robo in mid H15 deficiency embryos by Western analysis, which indicated a slight reduction in the levels of the Robo protein relative to wild type (Fig. 4G). This reduction is likely due to a secondary effect originating from the breaks in axon tracts or loss of Robo-expressing cells [due to identity changes, see ref. 20] as opposed to a direct Mid regulation of robo.
We also examined the transcription of sli, robo and frazzled (fra) in mid H15 deficiency embryos using the sensitive qRT-PCR method. As shown in Fig. 5, no significant differences were detected in the transcription of any of these genes in mid H15 mutant embryos compared to wild type. These qPCR results were reproducible using three different samples of embryo RNA preparations prepared separately in three different days, and qPCR was done in triplicates for each of the samples (the averages with standard errors were shown in Fig. 5). These results suggest that Mid has no role in the transcription of these genes during neurogenesis (note that there is no maternal contribution of mid).
Finally, Liu et al [27] had suggested that mid and fra show transheterozygous genetic interaction since they found that embryos transheterozygous for mid and fra have strong axon guidance defects. We re-examined if the two mutations show such an interaction by staining mid/+, fra/+ embryos from mid/CyO and fra/CyO with Fas II and BP102 antibodies. However, we did not observe any axon guidance defects in these embryos. Sometimes, balancer-bearing parents generate a few embryos that show axon guidance (or other) defects. We have previously named this ‘balancer-induced parental effect’ [7], [20]. This effect can also be suppressive. Therefore, we generated transheterozygous embryos from mid and fra parents that do not carry any balancers (mid/+ and fra/+). The transheterozygous embryos from this cross also did not have any axon guidance defects (Table 1). We did not find any axon guidance defects in embryos transheterozygous for the mid H15 deficiency and fra as well (Table 1). Similarly, no transheterozygous interaction between slit and mid was observed (Table 1). Therefore, we conclude that no transheterozygous interaction occurs between mid and fra or between slit and mid.
We next sought to determine the molecular basis for the axon guidance defects in mid mutant embryos. Our results show that in mid mutant embryos some of the interneuronal pathways that normally project along the midline stall between PC and AC of the next segment and then get redirected across the midline or away towards the periphery, perpendicular to the midline (there are variations to this phenotype but the spectrum of such variations are all within this category). NBs are formed in waves (S1–S5) and in rows (1–7) under the control of neurogenic and proneural genes. Previous studies have shown that many of the segmentation genes, especially segment polarity genes, are expressed row-wise in NE and NB cells. These genes play a crucial role in the row-wise specification of NB identity [reviewed in ref. 15].
To determine if the row-wise cellular identity within the nerve cord is altered in mid mutants, which might underlie the inhibitory zone and the associated guidance phenotype, we sought to examine the expression of some of the segment polarity genes. First, we examined mid mutant embryos for the expression of Wingless (Wg or W in Fig. 6K) and Gooseberry (Gsb and G in Fig. 6K). In wild type, Wg is present in row 5 NBs and the corresponding NE cells (Fig. 6A, B, K, see also Bhat, 1998). In mid mutant embryos, row 5 NB or NE expression of Wg was not affected, however, we observed ectopic expression of Wg in row 2 NBs and the corresponding NE cells (Fig. 6C–H). This ectopic expression was often stronger in alternate segments (see Fig. 6C, D, E). We note that the extent of ectopic expression of Wg was variable from segment to segment. For example, we found hemisegments or segments in which large patches of cells in the region between row 5 and row 7 (of the preceding segment) expressing ectopic Wg (Fig. 6G and H), which can also explain some of the variations in the guidance defects. Nonetheless, these results indicate that cells in row 2 behave as if they were row 5 cells. That this mis-expression occurs during segmentation is also indicated by the cuticle defects seen in mid mutant embryos, with missing denticle belts in the corresponding region (see Text S1 and Fig. S1).
Consistent with the above interpretation of Wg results, Gsb expression was also mis-expressed in mid mutant embryos. In wild type, Gsb is expressed in rows 5, 6 and one NB in row 7 (NB7-1). In mid mutant embryos, while the normal expression of Gsb in rows 5, 6 and NB7-1 was not affected, we observed ectopic expression of Gsb in the same cells expressing ectopic Wg (Fig. 6E). However, unlike the ectopic Wg stripe, which was always present in the mutant embryos at detectable levels, the ectopic expression of Gsb in the stripe was often incomplete and at times undetectable. Occasionally, we observed strong ectopic Gsb expression corresponding to both row 5 and row 6 cells suggesting that in mid mutants in addition to row 2 cells changing into row 5 cells, some row 3 cells may change into row 6. Though infrequent, in such segments it appears there is a reiteration of row 5 and 6 (rows 1, 5, 6, 4, 5, 6, 7) to varying degrees within the nerve cord in mid mutant embryos.
We next stained the mutant embryos for the expression of Sloppy-paired (Slp). We decided to examine Slp expression since in wild type Slp is expressed in rows 4 and 5 [Fig. 6I; see also ref. 13] and a change in Slp expression in mid mutants would allow us to confirm the results from the Wg and Gsb staining. This would also help us determine if cells corresponding to row 4 have changed to some other row of cells. In mid mutant or deficiency embryos, we observed ectopic expression of Slp in cells corresponding to row 2 cells (possibly some cells from row 3)(Fig. 6J). However, the ectopic expression of Slp was stronger in those segments where ectopic Wg was also strongly expressed. Again, the ectopic Slp expression was incomplete compared to ectopic Wg. Nonetheless, these results show that multiple row 5-specific segmentation genes are expressed in row 2 cells in mid mutant embryonic CNS.
Our previous results have shown that Mid is strongly expressed in row 7 and row 1 cells as well as in corresponding midline cells [20]. Since the expression of key genes can change quickly from division to division in NBs, and is highly time-sensitive [16], [17], we re-examined wild type embryos with an antibody against Mid. As shown in Fig. 6L–O, we found that Mid is indeed expressed at low levels in a large number of NBs, including in rows 2, 3 and 4 (perhaps also in one NB in row 5). Except for the strong expression in row 7 and row 1, which remained unchanged during neurogenesis, the expression pattern of Mid in other NBs changed as neurogenesis proceeded (Fig. 6L–O).
If we stain wild type Drosophila embryos with a monoclonal antibody BP102, we can clearly visualize commissural architecture with the longitudinal axon tracts (LC) and the anterior and posterior commissures (AC and PC; see Fig. 7A, green and Fig. 7B). Unlike the Fas II or other markers examined in the preceding sections, which are all directed against a small number of axon tracts, BP102 recognizes many more CNS axons and provides a more complete picture of axon tracts within the nerve cord. Therefore, we double stained embryos with BP102 and an antibody against Even-skipped (Eve) to determine the position of certain Eve-positive neurons (and therefore their parent NBs) in relation to the commissural architecture. Eve staining shows that an RP2 neuro, which is generated by NB4-2 (a row 4 NB), is located at the inner armpit of AC [Fig. 7A; see also ref. 28], affirming the position of row 4 NBs at the level of the posterior border of AC. While RP2 undergoes a complex migration within the nerve cord during development, ultimately it settles down in the same row where its parent NB is formed [28]. Similarly, the Eve-positive aCC/pCC neurons are located at the inner armpit of PC (Fig. 7A), which are generated by NB1-1, thus, fixing the location of row 1 to the posterior border of PC. Unlike the RP2, aCC/pCC neurons do not undergo much migration and stay in the same row 1 [28]. Thus, although NBs generate numerous progeny and there is both germ band retraction and condensation of the nerve cord, the relative position of commissures at a later point in neurogenesis to early NB rows remain more or less stable.
We next stained embryos from different alleles of mid (los1, mid1, mid H15 df and los1/mid, H15 df) with BP102 (Fig. 7C–H) to visualize the commissural architecture in mutant embryos (we did not double stain mutant embryos with BP102 and Eve in order to be able to flatten the nerve cord to fully visualize the commissural architecture and also to maximize the number of mutant embryos examined; the double staining shown in Fig. 7A was done to determine the relative position of AC and PC to rows of NBs). As shown in Fig. 7C–H, significant disruption of the commissural tracts was observed in all the mutant alleles of mid. We could clearly observe blobs of tracts or tracts projecting laterally (black arrow in 7C, D–G) at the level of AC, breaks in LC, as well as mis-projection of commissures between two adjacent neuromeres, creating a criss-cross phenotype (panel 7E; this criss-cross phenotype was observed in other alleles/combinations as well, data not shown). These tracts appear to encounter a non-permissive region for LC projection at row 2/3 (which lies just above AC). Furthermore, the posterior commissural (PC) tracts are reduced to only a few axon pathways in nearly all commissures, indicating a loss of axon tracts in PC. This may be a secondary effect of stalling of axon tracts in preceding neuromeres, thus, reducing the number of axon tracts that cross the midline through PC. The anterior commissural (AC) tracts were also affected but to a lesser degree. In general, in all mid alleles, more than 80% of the hemisegments had longitudinal tracts stalled at the AC level, resulting in breaks above AC. However, it appears that los1 has slightly more severe overall CNS defects compared to other alleles or the deficiency and this appears to be the case in embryos transheterozygous for los1/Df as well. This may be consistent with the possibility that this specific allele has some gain of function effects given the molecular lesion in the gene in this allele [20]. Nevertheless, the defects were similar in all alleles.
Although Robo is present at very low levels in commissural tracts [Robo is down–regulated in commissures, see ref. 1], the Robo-staining pattern closely resembles that of BP102, minus the strong commissural staining of BP102 (Fig. 7I). With Robo staining of mid mutants, we could observe that the longitudinal axon tracts stall at the AC level in all mid mutants (see Fig. 4 also). This corresponds to row 3 NBs in wild type, which is just before row 2 (re-specified as row 5 in the mutant). It appears that, when longitudinal axon tracts encounter the re-specified row 5 cells in mid mutants, they stop and simply congregate at this position, forming a blob (Fig. 7J, see also C–H, indicated by star). This is also evident by the breaks in the continuity of longitudinal tracts (Fig. 7J, arrowhead). These defects are consistent with the presence of a region or a barrier above AC that is not permissive to longitudinal axon projection. These results argue that loss of function for mid alters the identity of rows of NE and NB cells. By the time neurons begin to project their growth cones, this change of row identity creates a zone which is either non-permissive or lacks signals for growth cones to continue in their usual path. Thus, these growth cones either collapse on to themselves or project laterally outward, or in some segments/hemisegments cross the midline in this region (see Fig. 2 also).
Guiding axon growth cones towards their synaptic targets is one of the most fundamental processes during neurogenesis. Axon growth cones navigate through different regions by responding to cues from the environment to ultimately find their synaptic targets. While the two major signaling pathways, Slit-Robo and Net-Fra, provide a larger architecture for axon guidance within the nerve cord, local environment is expected to influence axon guidance as well. However, this particular aspect has not been examined in detail. A given local environment will be determined by the identity of the neuroectoderm, neurons, ectoderm and perhaps by the identity of the mesoderm as well. Thus, broad changes in local environment in which axon growth cones have to navigate is likely alter the route or guidance of these axons.
It is well established that segment polarity genes determine the broad identity of cells within the nerve cord just as they do later during development to determine the segmental identity within the epidermis [reviewed in ref. 15]. Segmentation genes, specifically the segment polarity genes such as Wg, Ptc, Hh, En are expressed in rows of NBs to define specific and row-wise NB identities. That these segmentation genes also play a role in axon guidance is indicated by the fact that mutations in many segmentation genes alter axon guidance [18]. However, given that these mutations also alter NB identity, the effect of mis-specification of neuronal identity versus broad changes in the environment in which axons navigate, on axon guidance has been experimentally difficult to separate.
Our work described in this paper, however, attempts to separate the role of identity versus environment and reveal the significance of local environmental niche to normal axon guidance. Our results show that in mid mutants, there is an ectopic expression of segment polarity genes such as Wg, Gsb, Slp (and perhaps many more) in row 2 cells within the developing nerve cord, thus, re-specifying this row of cells into more like row 5 cells (and a second PSB). This re-specification appears to ultimately create a zone or a barrier that prevents axon growth cones from progressing further in their normal route. Instead, such growth cones either stall or project peripherally or across the midline but along this zone of non-permissive barrier. The highly specific nature of the phenotype(s) in response to a specific change in the environment in mid mutants presents a classic example of the specificity of the environment to axon guidance. Our results also show that the identity of some of the pioneering neurons, whose axon projections are misrouted, is not affected by loss of function for mid.
It would have been ideal if we were able to identify a single molecule that makes this re-specified row of cells non-permissive to longitudinal tracts extension in their usual path. We do not know if such a molecule exists, or the mechanism that created the barrier. But the barrier is unlikely due to the ectopic expression of genes such as Wg, instead it must be due to the change in the row-identity, activating a distinct genetic program that does not permit axonal extension in their normal path. Ectopic expression of Wg, or Slp or Gsb simply reflects this change. We also point out that this re-specified row 5 cells may not have all the features/genetic programs of a bona fide row 5 cells and more likely have a mixed identity. This is suggested by the fact that Slp and Gsb or even Wg expression in the re-specified row is not exactly like in a bona fide row 5. Similarly, we do not know at what point in development this zone or barrier is put in place, but it indeed originates with the altered row identity, and certainly becomes active by the time of pathfinding. This barrier might be due to signals from other neurons generated by the transformed NB row, or the transformed neuroectoderm/ectoderm. We further point out that a clonal analysis experiment would have been desirable to show that a broad identity-change is necessary for the observed guidance defects. However, mid is expressed extensively in the germline, both in the soma and the germ cells, and there is a requirement for Mid in these cells. Ultimately, it may require isolation of a temperature sensitive allele in mid to address this question in an unequivocal way.
Our results show that Mid does not regulate the expression of slit, robo or fra genes in the CNS. Consistent with this, the axon guidance defects in mid are distinct from the defects in slit or robo mutants. We confirm this by several different ways: immunostaining, RNA whole mount in situ, Western analysis, qPCR and genetic interaction studies. A previous study has suggested that Mid regulates sli, robo and fra [27]. They based their conclusion on finding a strong transheterozygous genetic interaction between mid and sli, and mid and fra, detected using BP102 staining of embryos that are transheterozygous for mid1 and sliGA20 and mid1 and fra3. Furthermore, they reported that levels of fra mRNA and Fra and Robo proteins in mid mutant embryos were down regulated, and that this can be completely rescued by expressing mid using elev-GAL4 driver [27]. They also reported that ectopic expression of mid in salivary glands induces expression of robo and slit. We have not found most of these effects reported by Liu et al [27]. For example, we failed to observe any genetic transheterozygous interactions between mid and sli mutants (Table 1). Transheterozygous interactions are rare given the negative evolutionary impact of such interactions to survival, but when observed, it is usually with mutations in receptor-ligand pairs, or with gain of function/neomorphic situations). We used stronger allelic combinations than the ones used by Liu et al [27] with mid and sli. For mid, we used not only mid1, but also a deficiency that removes both mid and its sister gene H15, as well as los1. For slit, we used sli2, which is the strongest loss of function allele and genetically behaves as a null. Furthermore, we also failed to observe any such transheterozygous interactions between mid and fra (Table 1).
Secondly, we found that while the ectopic expression of mid in salivary glands induced robo expression as was reported by Liu et al [27], no such induction was observed with slit (Text S1 and Fig. S2). One should also consider the fact that Mid, Robo (and Slit) have mostly non-overlapping domains of expression in the CNS, therefore, the direct regulation of robo by Mid in the salivary gland has little relevance in the CNS or CNS development mediated by Robo or Mid. Our results indeed bears this out. Not only the axon guidance defects are different between mid and slit or robo, the transcription of robo, slit or fra are also unaffected in mid mutants (Figs. 3–5, Table 1). There was some reduction in the levels of Slit in los1 allele in the midline in the PC region. But, the molecular lesion in los1 is complex and might have some allele-specific gain of function effects that alters cellular identity or function of the corresponding midline glia to mediate reduction in the Slit level in this region. Since no such reduction in the levels are seen in other mid alleles and more importantly the transcription of slit is unaffected in the deficiency that removes mid (and H15), we think that the slight reduction in the levels of Slit in los1 is allele-specific.
In the case of robo, the promoter has three TBEs. With three sites, Mid is more likely to be able to induce robo in an ectopic site. However, within the CNS, we did not find any significant loss of Robo expression in mid mutant embryos by immunohistochemistry (either in los1 or mid H15 deficiency embryos) or by Western analysis (Fig. 4) or robo transcription by qPCR (Fig. 5). A slight reduction in the levels of Robo seen in Westerns is likely due to a secondary effect of loss of tracts and perhaps loss of some of the Robo-expressing neurons perhaps due to identity changes [19]. The reason for the significant reduction in the expression of Robo in mid observed by Liu et al [27] is not clear. We think that this may be due to some technical reasons such as variability from embryo to embryo to fixing and staining. Because of this possibility, we follow a simple rule: in this case, we focused on mid mutant embryos that had strong guidance defects to determine if such mutant embryos also had a strong or weak expression of Robo and Slit. We found that embryos with strong guidance defects also had strong Robo (or Slit) expression. Thus, we avoided selecting sub-stained mutant embryos and comparing them to optimally stained wild type embryos. Finally, Liu et al [27] reported that there is overlapping expression of Mid and Slit in a small number of cells located laterally within the nerve cord. It is possible that Mid regulates slit expression in these cells, however, the contribution of Slit or such a regulation of slit to the overall axon guidance mediated by Slit is not clear and likely very minimal, if there is any. We have also not examined if mid affects netrin gene expression.
Our work shows that in mid mutants, the majority of axon growth cones of the longitudinal tracts stall and club together at the level of AC, creating a blob of axons, thus leading to interruptions between neuromeres. Interestingly, some of the tracts project outward towards the periphery or inward across the midline (see Table 1). This outward projection route is quite revealing: the projection path is mostly perpendicular to the midline and just below the transformed row of cells. The transformed row of cells corresponds to the region right above the AC or where the tracts stall (Figs. 6 and 7). The most consistent change is seen with row 2 cells, changing into row 5 cells. How does these changes relate to wild type? In wild type, row 5 cells normally separate one neuromere from the next and also define PSB. Thus, the change from row 2 to row 5 in mid must be creating an environment that either lacks the necessary permissive/attractive cues or possess cues that are inhibitory to the projection of these axon tracts, causing the tracts to stall.
For example, in wild type row 5 cells are located between pCC and vMP2, the two axons that pioneer the Fas II-positive medial tract. The growth cone from vMP2 in wild type only marginally encounters row 5 cells but does not necessarily traverse row 5. This is due to the fact that vMP2 is located in row 5 and the growth cone from a vMP2 stops at row 5 region and fasciculates with the vMP2 of the next hemisegment. However, it does encounter row 2 cells midway through the projection path. In mid mutants, since row 2 cells change to row 5, creating a region that vMP2 growth cone is perhaps normally programmed to stop. For the proper guidance of medial tract, normal projection of vMP2 and pCC is necessary and loss of either of the two pioneer neurons causes aberrant medial tract guidance [9], [10]. Therefore, it seems likely that vMP2 stalls and the pCC projection, along with several other follower projections, also stalls; or some of the tracts project across the midline or away towards the periphery. In fact, these abnormal projection patterns, especially towards the periphery appear to be guided by the newly created barrier. This situation is also the same for MP1 or dMP2.
What is the mechanism within the re-specified row of cells that eventually mediates the block for axon projection? The re-specified rows of cells would have a whole set of new (row 5-specific) genetic programs that may simply not conducive to longitudinal tracts. Additionally, the role of Ephrin pathway in axon guidance may be relevant here. The Drosophila Ephrin (Eph), which is a transmembrane protein, is shown to prevent interneuronal axons from exiting the Drosophila embryonic CNS [29]; some of the interneuronal pathways in mid mutant exit the nerve cord (Fig. 2G). Ephrin/Eph signaling is via cell-to-cell contact and depends on the clustering of Eph receptors and their ligands [30]–[33]. This multimerization activates the kinase activity of the receptor and leads to the phosphorylation of the receptor within the cytoplasm-exposed tail region and the binding of downstream effectors [34]. This triggers the depolymerization of actin in growth cones, modifying the Integrin-based cell adhesion [29], [35]. The CNS-exiting phenotype of interneuronal pathways in mid mutants suggests a possible de-regulation of the Eph-pathway. But, it may also be that changes in Eph or similar cell-adhesion mechanism mediate the formation of the barrier and exiting of some of the interneuronal pathways from the CNS.
Our previous results show that Mid acts as a transcriptional repressor of gsb-n [20]. However, in mid mutants the transformation of row 2 into row 5 also activates Gsb expression (Fig. 6). The ectopic activation of Gsb in these cells in mid, however, is not a direct de-repression of gsb, but an indirect consequence of the transformation of cell identity from row 2 (a Gsb-negative row of cells) to row 5, a Gsb-positive row. Finally, our results provide clear evidence that segmentation genes can regulate axon guidance via broadly defining cellular identity, creating a permissive and non-permissive boundaries or niche. We also emphasize that extrapolating expression relationships to functional relevance from induction in ectopic sites, in vitro and tissue culture experiments, bioinformatics or other similar in vitro studies carry inherent risks and should be done with caution.
mid mutant alleles used were mid1, mid2 and los1. We also used a deficiency that removes both mid and H15 genes (mid H15df; BL# 7498: breakpoints: 25D5-25E6). The other lines used were: sli2, robo4, robo-deficiency [Df (2R) BSC787, breakpoints: 58F4-59B1; BL#27359], UAS-mid, sim-GAL4, sgs-3-GAL4 (to induce mid in the salivary gland), ac-GAL4, UAS-mCD8-GFP, RN2-GAL4 (eve-GAL4) and UAS-tau-lacZ. For wild type, we used Oregon R flies. All the mutant lines were balanced using GFP-bearing balancer chromosomes to facilitate identification of the mutant genotype.
Transgenic UAS-mid fly lines containing one or two copies of the UAS-mid were previously generated in the lab [20]. The transgenic flies were crossed to sim-GAL4. Embryo collection was done overnight at 28°C. The embryos were fixed, divided into three portions and stained separately with antibodies against Mid, Robo and Slit.
ac-GAL4 driver (BL#: 8715) and the UAS-mCD8-GFP (BL# 41803) were introduced into the mid H15 deficiency background and the embryos were stained for GFP and Odd-skipped. ac-GAL4 drives the UAS-mCD8-GFP in vMP2 and dMP2 and their axons. Odd-skipped is expressed in dMP2 and MP1.
RN2-GAL4 (portion of the eve promoter that drives expression in aCC/pCC and RP2 neurons) and UAS-tau-lacZ were introduced into mid H15 deficiency background and the embryos were stained for LacZ.
sim-GAL4 and UAS-tau-GFP introduced to mid H15 deficiency background and stained for GFP and Eve (Eve to identify the mutant). Tau and MCD8 targets GFP to axon tracts.
The embryo collection, fixation and immunostaining were performed according to the standard procedures. The following antibodies were used: anti-Sli C (1∶20, DHSB), anti-Robo (1∶5, DHSB), anti-Mid [1∶50, generated in the lab, see ref. 20], anti- Fas II (1∶5, DHSB), 22C10 (1∶1, DHSB), anti-Wg (1∶5), anti-Gsb (1∶3), anti-Slp (1∶400), anti-GFP (1∶300), anti-Odd (1∶500), anti-Eve (1∶2000), anti-Lac Z (1∶500). For color visualization, either AP-conjugated or HRP-conjugated secondary antibodies were used. For double staining, secondary antibodies conjugated with AlexaFluor488 and AlexaFluor635 were used. Whole-mount RNA in situ hybridization for sli expression was done following the standard procedure using a digoxigenin-labeled slit probe, synthesized by PCR.
Cuticle preparation was done as per standard procedure by fixing embryos and dissolving organic embryonic material on slides using Hoyer's solution at 65°C for 24 hours.
For western blot analysis, 30 embryos were collected (homozygous mutant embryos were identified by the lack of GFP expression under microscope), homogenized in 37.5 µL lysis buffer (0.15 M NaCl, 0.02 M Tris pH,7.5, 0.001M EDTA, 0.001 M MgCl2, 1% Triton-X-100 and PIC) and kept on ice for 10 minutes. The lysed protein is centrifuged for 5 minutes at 13,000 rpm, the supernatant is collected and diluted with 12.5 µL 4× Laemelli sample buffer. The protein sample is boiled in water for 10 minutes and kept in 4°C for 10 minutes. Equal amount of lysed protein 20 µL (15 embryos per lane) was loaded on to a 4–12% SDS-PAGE gel. The separated proteins were transferred to a Nitrocellulose membrane. The efficiency of transfer was determined by Ponceau S staining. The membrane is blocked in 5% milk for 2 hours at room temperature, and incubated with primary antibodies (anti-Slit N 1∶50000 or anti-Robo 1∶40) overnight at 4°C and washed with PBST (PBS+0.02% Tween 20). The Nitrocellulose membrane was then incubated with HRP-conjugated secondary antibody (anti-Rabbit 1∶20000 or anti-mouse 1∶20000) for 2 hours at room temperature and washed with the washing buffer. Proteins were detected by the chemi-luminiscent ECL reaction method (Thermo Scientific). The autoradiographs were scanned and intensities of bands were analyzed using the software AlphaEaseFC. Anti-Tubulin antibody (Abcam, 1∶4000) was used for determining loading control.
Embryos from Oregon R (wild type) and mid H15 deficiency lines were collected and aged for 12–14 hours. They were dechorionated in 50% bleach and washed with water. Approximately 150 embryos were selected under microscope for each sample. Total RNA isolation from these embryos were performed using the RNaqueous Kit (Ambion). The isolated RNA was DNase treated and quantified using Nanodrop Spectrophotometer (Nanodrop Technologies) and qualified by analysis on RNA Nanochip using Agilent 2100 Bioanalyzer (Agilent Technologies). Synthesis of cDNA was performed with 1 µg of total RNA in a 20 µL reaction using the Taqman Reverse Transcription Reagents Kit (ABI). Reaction conditions were as follows: 25°C, 10 minutes, 48°C, 30 minutes and 95°C, 5 minutes. Primers for real-time PCR were designed and made by the Molecular Genomic Core facility at UTMB. Real-time PCR were done using 1.0 µL of cDNA in a total volume of 20 µL using the Faststart Universal SYBR green Master Mix (Roche, #04913850001). RpL32 was used as endogenous control. All PCR assays were performed in the ABI Prism 7500 Sequence Detection System and the conditions were as follows: 50°C, 2 min, 95°C, 10 min, 40 cycles of 95°C, 15 sec and 60°C, 1 min. Primers used: slit: Forward: 5′-GCGTTATGCCCGGTTCC-3′, Reverse: 5′ TCCACAACGTGCCGCTC-3′); robo: Forward: 5-CAGCATTAGTCTTCGTTGGGC-3, Reverse: 5-AATCCAACCAGTTTGCAGATTC-3); fra: Forward: 5-AGACCCCAGAGCATCCTTATG -3, Reverse: 5-TCTTTAGAGGATGGCCACGC-3. The qRT-PCR was done on three seperate embryo collections for each genotype and in triplicates for each collection.
|
10.1371/journal.pntd.0002401 | Mediational Effects of Self-Efficacy Dimensions in the Relationship between Knowledge of Dengue and Dengue Preventive Behaviour with Respect to Control of Dengue Outbreaks: A Structural Equation Model of a Cross-Sectional Survey | Dengue fever is endemic in Malaysia, with frequent major outbreaks in urban areas. The major control strategy relies on health promotional campaigns aimed at encouraging people to reduce mosquito breeding sites close to people's homes. However, such campaigns have not always been 100% effective. The concept of self-efficacy is an area of increasing research interest in understanding how health promotion can be most effective. This paper reports on a study of the impact of self-efficacy on dengue knowledge and dengue preventive behaviour.
We recruited 280 adults from 27 post-outbreak villages in the state of Terengganu, east coast of Malaysia. Measures of health promotion and educational intervention activities and types of communication during outbreak, level of dengue knowledge, level and strength of self-efficacy and dengue preventive behaviour were obtained via face-to-face interviews and questionnaires. A structural equation model was tested and fitted the data well (χ2 = 71.659, df = 40, p = 0.002, RMSEA = 0.053, CFI = 0.973, TLI = 0.963). Mass media, local contact and direct information-giving sessions significantly predicted level of knowledge of dengue. Level and strength of self-efficacy fully mediated the relationship between knowledge of dengue and dengue preventive behaviours. Strength of self-efficacy acted as partial mediator in the relationship between knowledge of dengue and dengue preventive behaviours.
To control and prevent dengue outbreaks by behavioural measures, health promotion and educational interventions during outbreaks should now focus on those approaches that are most likely to increase the level and strength of self-efficacy.
| Dengue fever is one of the most rapidly increasing vector-borne diseases of humans in the tropics. There is currently no treatment and no vaccine, so control of the disease depends on controlling the mosquito vector. Unfortunately health promotional campaigns aimed at encouraging people to reduce mosquito breeding sites have not always been 100% effective. Self-efficacy is an area of increasing research interest and can be thought of as people's confidence in their ability to engage in health behaviours. We report a study of the impact of self-efficacy on dengue preventive behaviour. We conducted face to face interviews in villages in the state of Terengganu, Malaysia that had been affected by dengue outbreaks. A structural equation model was tested and fitted the data well. Mass media, local contact and direct information-giving sessions significantly predicted level of knowledge of dengue. However, self-efficacy fully mediated the relationship between knowledge of dengue and engagement in dengue preventive behaviours. We conclude that educational components of community dengue control programmes should focus on interventions.
| Dengue fever is transmitted by the bite of an Aedes mosquito infected with any one of the four dengue viruses. Although most infections are self-limiting a proportion of cases develop severe complications such as dengue haemorrhagic fever which can carry a significant risk of death. The incidence of dengue has risen dramatically around the world in recent decades. Since no vaccine is currently available, primary prevention is regarded as the most effective measure in controlling dengue. Each time an outbreak occurs, the local health authority will plan and carry out various types of promotional and educational activities that aim to increase knowledge of dengue and change dengue preventive behaviour among communities at the centre of the outbreak. These promotional activities can be carried out through various methods such as individual home visits, or at the population level through the mass media. Health promotion and educational intervention like, ‘search and destroy’ activities, advice on the need to seek immediate medical attention in patients with fever, and proper disposal of rubbish are usually the focus of behavioural-change promotion activities. The promotional and educational messages are usually delivered using small group discussion, public lecture, live public announcement, demonstration, distributing printed materials, putting up posters, bunting and billboards, community source reduction and community dengue-cleanliness program (in Malay: Gotong-Royong) and health exhibition [1].
There have been a number of systematic reviews of public health interventions aimed at reducing the risk of dengue fever in recent years [2]–[6]. However, as pointed out by Bouzid et. al. authors have often reached different conclusions regarding the effectiveness of interventions, even when reviewing the same primary studies [7]. Health promotion campaigns that appear to have some benefit are those aimed at encouraging local people to engage in activities that reduce the number of mosquitoe breeding sites close to home [2]. However, such campaigns are not totally effective and the impact on vector presence may only be short-live. Achieving sustainable change in dengue preventive behaviours remains difficult [8], and may not necessarily lead to dengue prevention [9], [10], [11].
Dengue fever is endemic in Malaysia with frequent major outbreaks in the urban areas. Since dengue was first documented in Malaysia in 1902 and was made notifiable in 1973, the disease pattern has changed from major outbreaks every four years to one of increasing trend yearly. The largest outbreak was seen in 1996 with 14,255 dengue cases reported and 32 deaths. The fever is the number one disease in the top 10 listed communicable diseases in Malaysia as compared to other diseases like Tuberculosis, Malaria and HIV/AIDS in 2010 and 2011 [12]. The number of dengue cases reported also increased from 27,381 cases in 1998 to 46,171 cases in 2010. Estimate of an economic burden of dengue in Malaysia is USD102.25 (95%CI: 77.94–310.66) million per year which is approximately USD3.72 (95%CI: 2.83–11.30) per capita [12]. There is evidence that despite the fact that Malaysians generally have good knowledge of dengue fever and its prevention [13], dengue incidence rate has substantially increased from 31.6/100,000 population in year 2000 to 163/100,000 in year 2010 [12].
There is a growing body of literature concerning the concept of self-efficacy, which is considered to be people's belief or confidence in their capabilities to achieve different levels of performance attainment [14]. Self-efficacy perceptions are viewed as important determinants of behaviour and affect, and the potency of these perceptions in predicting behaviours in many domains has been shown [15]. The concept of self-efficacy is commonly used in studies of health behaviours [16], [17]. including area such as smoking cessation [18], [19], weight loss and body weight control [20]–[25], exercise [26], [27], [28], nutrition intake [29], [30], alcohol use [31]–[34], and AIDS prevention [35], [36], [37]. Self-efficacy may also function as a mediator between cognitions, feelings and behaviours and the adoption of lifestyle behaviours such as healthy diet [38]–[41]. Although the effects of health promotion and educational interventions to control dengue fever have been investigated in previous studies, none of the studies have investigated the impact of self-efficacy dimensions (level and strength of self-efficacy) as mediators between level of dengue knowledge and effective behavioural actions to control dengue outbreak and transmission. Strength of self-efficacy refers to a person's perceived assurance that they ‘can do’ or ‘cannot do’ something reflected in their affirmative answers to questions about whether they can perform particular dengue preventive behaviours. Level of self-efficacy is a person's judgement about whether or not they can accomplish a given performance which reflects their perceived capability as measured against task demands (dengue preventive behaviour) at various levels of challenge (scenarios) to successful control of dengue fever during outbreaks [42].
We argue that understanding the relationship between knowledge, self-efficacy and behavioural change may be a route towards improved and sustainable dengue control. This paper reports on work that was conducted to study the impact of self-efficacy on dengue preventive behaviours. We conducted a survey in villages that was subsequently examined with analyses based on predictions from Bandura's Social Cognitive Theory and Maibach's path model [42], [43]. We specifically examined the potential mediating effects that level and strength of self-efficacy may have on the relationship between knowledge of dengue and dengue preventive behaviour after being exposed to health promotion and educational interventions during the outbreaks.
We recruited heads of families or their spouses aged above 18 years old from 27 villages that had recently experienced an outbreak of dengue fever. These villages were located in the state of Terengganu, on the east coast of peninsular Malaysia. Using the method by Woodward, we calculated that we needed a sample size of about 280 respondents [44]. This was based on a requirement to detect a Pearson correlation coefficient of 0.4 with a power of 80% and alpha 5% (200 samples). This sample size was then inflated by 20% to account for possible non-parametric tests and 20% for potential impact of clustering within village.
The population of the study included all the villages of the outbreak localities from July to December 2010 in the state of Terengganu. The list of outbreak sites was obtained from Terengganu Vector Borne Disease Division and the Terengganu Crisis Preparedness Resource Centre (CPRC) database. In total there were 32 outbreak locations for that 6-months period, but only 27 locations were included in the study as five others were not actually villages but higher education institutions and schools. Figure 1 shows a simplified illustration of sampling procedures used in this study. The households that were interviewed in those selected villages or sites were randomly selected based on the current outbreak list obtained from the Terengganu Crisis Preparedness Resource Centre (CPRC) database. The households were selected randomly from 9,959 houses or premises included in the study using SPSS. A total of 149 premises were excluded because they were non-owner premises or abandoned houses. The selected households were not changed or replaced with other households even if the first and second visits resulted in failure to meet some the participants for interview.
Research ethics approval for this study was granted by the National Medical Research Register, Ministry of Health Malaysia (NMRR-10-206-5412) and the Faculty of Health Research Ethics Committee of the University of East Anglia (2010/2011–13), the author's institution, prior to data collection. Written informed consent was obtained from all participants prior to completion of the survey.
The data collection was carried out from January to March 2011. The recruitment and training of the 32 interviewers from the local State Health Department staff was undertaken in December 2010. The interviewers were dedicated staff from the State Health Department whose usual tasks involved running health promotion and education activities during the outbreak. The training was conducted by the lead researcher assisted by the Head of the Health Promotion Unit, Terengganu Health State Department. All respondents were recruited after they had given informed consent. The interviewers read the questionnaires and the respondents gave their answers to those questions. The interviewers then ticked the answers in the column provided and recorded any subjective answers not listed in the answer scripts. A copy of the questionnaire is given in supplementary Text file S1. The questionnaire used in the field was in Malay. Correct translation was checked through a parallel back-translation by members of Malaysia Institute of Translation. The questionnaire was also tested in a pilot study in both languages.
Pearson or Spearman correlation coefficients used in analyses between the three main outcome measures: exposure to health promotion and education, knowledge of dengue and self-efficacy dimensions. Regression modeling was carried out using Generalized Estimating Equations of SPSS 18 to account for cluster sampling at the village level. Based on consideration of Bandura's theory, the research model by Maibach et al. [42], and previous empirical findings on related public health issues and significant correlations from the investigation, an initial proposed model was constructed [38]–[41]. A structural equation model (SEM) was developed using AMOS version 18 [50]. The model reflects the relationships between variables obtained in the study in order to predict the dengue preventive behaviour change resulting from an increased dengue knowledge level and self-efficacy dimensions after being exposed to health promotion and educational intervention during the outbreak. SEM was used to test the proposed model against the observed dataset. SEM is a combination of factor analysis and path analysis and it is a confirmatory rather than an exploratory technique, because it compares a hypothesized model's covariance matrix with that of the observed data. Since the proposed model of this study involved observed variables, SEM allows us to determine significant paths between those variables in deriving a better explanation of their significant relationship findings based on the research hypotheses and proposed model.
There are several steps in analysing SEM using AMOS: 1) to develop a model based on research theory; 2) identify unique values that can be used for the parameters to be estimated in the proposed model; 3) apply various estimation techniques, for example in this study, maximum likelihood; and 4) test the fit of the model against the data. According to the results, the researcher might 5) modify the measurement model based on theoretical justifications; revise the model by adding, deleting, or modifying relationships between variables; or use measures indicating lack of fit for specific parts of the model when theoretically justified in the Modification Indices table [51]. Goodness of fit indices were used as indicators of model fit. Chi-square tests were used as an index of the significance of the discrepancy between the original (sample) correlation matrix and the (population) correlation matrix estimated from the model. Because the significance of chi-square tests is dependent on the number of subjects, the comparative fit index (CFI) and the root mean square error approximation (RMSEA) were further considered. CFI values are derived from the comparison of the hypothesized model with the independence model. RMSEA values help to answer the question of how well the model with unknown but optimally chosen parameter values would fit the population covariance matrix if it were available [52]. The lower the discrepancy measured by the RMSEA the better, with an RMSEA of 0.0 indicating a perfect fit. Acceptable values are CFI>.90 and RMSEA<.08.
Once the model fitted the data well, the next step was to test the mediation effect of self-efficacy dimensions on the relationship between knowledge of dengue and dengue preventive behaviour by comparing a Full Mediation Model, Direct Model and Indirect Model as recommended by Baron and Kenny [53], and Hayes [54]. The post-hoc probing test for mediation effect significance was performed to determine if the drop in the total effect (i.e. level of dengue knowledge to dengue preventive behaviour) was significant upon inclusion of mediator (level or strength of self-efficacy) in the model [55].
We aimed to test two primary hypotheses, namely (i) knowledge of dengue is directly associated with dengue preventive behaviours and (ii) both strength and level of self-efficacy are associated with dengue preventive behaviour.
We recruited 280 participants as per the sample size calculation. More than half of the respondents were female (58.9%). Their mean age was 42.7 years, and the majority (57.5%) were aged between 36 to 55 years old, and were married (96.1%). The ethnic background of the respondents was 98.6% Malay with the remainder being Chinese. Nearly half of the respondents were housewives (45%).
Table 1 presents the means, standard deviations, percentiles and ranges for all the principle scores. The level of health promotion and educational intervention exposure was low, with only 20% of the respondents receiving a moderate to high level of exposure. The respondents seemed to have moderate (40.7%) to good (38.6%) knowledge of dengue. In general, the self-efficacy of the respondents was at the moderate level. Although it was also found that 62.2% of respondents perceived they were relatively confident in performing dengue preventive behaviours, only 1.1% of them reported having excellent strength of self-efficacy. About half of respondents (45.4%) showed moderate levels of self-efficacy, while 36.4% had little confidence and felt uncertain how to perform these kinds of dengue preventive behaviours. 2.5% of them reported below the average confidence (mean = 2.99).
Most of respondents said they had received health information on dengue fever from Public Announcements (57.5%), Television (57.9%) and the Newspaper (44.6%). In term of respondents' participation in the health promotion and educational interventions, most of them tended to be involved in the Community Source Reduction Program or Gotong-Royong (60%) as compared to the Public Lecture (24.3%). Only 4.6% were involved in Demonstration activities. Regarding respondents' recent behaviour to control dengue outbreak and transmission, 73.2% of them failed to perform a 10-minute search and destroy exercise to eradicate Aedes mosquitoes breeding sites within the last 14 days.
With regards to dengue preventive behaviour, about half (45.5%) of the respondents did not comply with correct behaviours to control dengue fever transmission. Moreover, 30.4% of them had carried out only 5 minutes of cleanliness activity within the past 14 days. Only 23.9% of the respondents were found to comply with the correct behaviours to prevent dengue fever transmission as promoted in the educational interventions during the outbreaks.
Four factors were extracted from the data on information sources and together these four factors represented 60.1% of the variance in the original variables. Table 2 shows the rotated component. Factor 1 was associated with obtaining information through Television, Radio and Newspaper (regression score >0.5). We named this factor Mass Media. Factor 2 was associated with participation in Gotong-royong and obtaining information from public announcements and outdoor media. This factor was named Local Contact. Factor 3 was named Small Group Contact (Small Group Discussion and Demonstration) and Factor 4 was named Direct Information-Giving Session (Public Lecture and Individual Advice).
A correlation matrix was generated that included each of the variables in the study (see Table 3). Overall, the bivariate relationships between the majority of independent and dependent variables were weak. The relationship between Factor 1 from the health promotion and educational intervention (Mass Media) and mean dengue knowledge scores was the strongest (r = 0.326, p<0.01) as compared to other factors. Significant bivariate relationships were evident between dengue knowledge and level (r = 0.262) and strength of self-efficacy (r = 0.363) at p<0.01. Level of self-efficacy was significantly correlated with strength of self-efficacy (r = 0.383) at p<0.01 and dengue preventive behaviour (r = 0.212) at p<0.05. There was a significant correlation between dengue knowledge and dengue preventive behaviour. There was no significant difference in level of dengue knowledge between those respondents who were exposed to different levels of health promotion and educational intervention. However, there were different degrees of strength in self-efficacy among those who were exposed to different levels of health promotion and educational interventions (p = 0.022). The level of self-efficacy however was no different among them. Although not included in the SEM, we found a significant correlation between proportions of villages who didn't undertake at least 10-minutes-cleanliness behaviour per week with the duration of the outbreak in the village (p = 0.044).
Examination of our proposed model using SEM of AMOS 18 indicated that adjustment could be made to improve the match between the data and model (χ2 = 75.622, df = 41, p = 0.189, CFI = 0.870, TLI = 0.895, RMSEA = 0.098). To identify the sources of error in the proposed model as indicated in Modification Indices, we eliminated paths that were not significant one at a time in order to find the most parsimonious model. First we eliminated the path between Factor 3 from the health promotion and educational intervention (small group contact) and knowledge of dengue. Second, we dropped the path between knowledge of dengue and dengue preventive behaviours. Table 4 contains the models' goodness of fit indices.
Our final model fitted the data well (χ2 = 71.659, df = 40, p = 0.002, CFI = 0.973, TLI = 0.963, RMSEA = 0.053). Bentler [56], and Chou [57] both recommend CFI and TLI scores of greater than 0.90 as indicators of good fitting models. Browne & Cudeck [58], (1993) and Byrne [52] recommend that models with an RMSEA of 0.08 or less and preferably 0.05 or less are good fitting models. Figure 2 shows the entire final model with accompanying path coefficients. Overall, the structural model contains relatively weak influences on the dengue preventive behaviours, with path coefficients ranging from 0.092 to 0.271.
Our main objective was to investigate the self-efficacy dimensions as mediators of the relationship between dengue knowledge and dengue preventive behaviours in relation to control of dengue outbreaks. Assessment of the mediation effects was done by comparing the full mediation model (which includes a direct model) and indirect model from the final structural model that we created earlier. Table 5 shows the mediation effect findings of this models comparison.
From the initial analysis, knowledge of dengue did not have a direct effect on dengue preventive behaviour (standardized β weight = 0.092, p = 0.082). However, knowledge significantly predicted the level of self-efficacy as expected (standardized β weight = 0.172, p<0.001), and this level of self-efficacy also significantly predicted dengue preventive behaviour (standardized β weight = 0.179, p = 0.036). Knowledge had a direct effect on strength of self-efficacy (standardized β weight = 0.291, p<0.001) and this strength of self-efficacy also significantly predicted dengue preventive behaviours (standardized β weight = 0.149, p<0.001). Analysis for model comparison as recommended by Baron and Kenny [53], and Hayes [54] found that the Beta for the Indirect Model was reduced from 0.092 to 0.090 in the Full Mediation Model (in both for level of self-efficacy and strength of self-efficacy as mediators). Therefore, knowledge on dengue was found to have significant indirect effect on dengue preventive behaviour with a mediation effect of level of self-efficacy or strength of self-efficacy on the relationship.
Post-hoc probing of significant mediation effects was performed using the Sobel Equation of computing to determine if the drop in the total effect (i.e., knowledge on dengue) is significantly dependent upon inclusion of the mediator (level of self-efficacy and strength of self-efficacy) in the model [53], [55], [59]–[60]. This strategy indicated that level of self-efficacy (z = 4.77, p<0.05) and strength of self-efficacy (z = 2.38, p<0.05) did function as mediators. According to Holmbeck [55], p<0.05 is the absolute value of z>1.96. In addition since the Beta for the total effect of the relationship between knowledge on dengue and dengue preventive behaviours was 0.37, thus roughly 65% of the path was accounted for by level of self-efficacy as a mediator (Beta for indirect effect was 0.2405). Likewise, the path between knowledge of dengue and dengue preventive behaviours was 97% accounted for by strength of self-efficacy as a mediator in the relationship (Beta for indirect effect was 0.3575) [61]. Therefore, since the direct path between knowledge on dengue and dengue preventive behaviours was not significant, level and strength of self-efficacy did function as full mediators of that relationship. This result showed that self-efficacy has a complete mediation effect on the relationship between knowledge on dengue and dengue preventive behaviour.
As we hypothesized, strength of self-efficacy and level of self-efficacy significantly predicted dengue preventive behaviours (p<0.001). In addition, strength of self-efficacy significantly predicted level of self-efficacy (standardized β weight = 0.282, p<0.001). Later analysis for models comparison found that, the Beta for the Indirect Model was reduced from 0.179 to 0.168 in the Full Mediation Model. Therefore, strength of self-efficacy was also found to have a significant indirect effect on dengue preventive behaviour with a mediation effect of level of self-efficacy on the relationship.
Once again, post-hoc probing of significant mediation effects was performed using the Sobel equation of computing as a follow-up to the findings for the structural equation model. The post-hoc strategy was conducted to determine if the drop in the total effect (i.e., strength of self-efficacy) is still significant upon inclusion of the mediator (level of self-efficacy) in the model. This strategy indicated that level of self-efficacy did function as mediator (z = 2.020, p<0.05). (Note that p<0.05 if the absolute value of z>1.96). In addition since the Beta for the total effect of the relationship between strength of self-efficacy and dengue preventive behaviours was 0.3695, thus roughly 43% of the path was accounted for by level of self-efficacy as a mediator (Beta for indirect effect was 0.1596) [61]. In this case, level of self-efficacy partially mediated the association between strength of self-efficacy and dengue preventive behaviours. This result showed that level of self-efficacy has a partial mediation effect on the relationship between strength of self-efficacy and dengue preventive behaviour.
This investigation is one of the first within the public health and health psychology research literature to concentrate on health promotion and educational interventions designed to reduce the risk of dengue fever. With regards to our first primary hypothesis, knowledge was found not to be independently associated with dengue preventive behaviour other than through the impact of knowledge on self-efficacy. For the second hypothesis, both level and strength of self-efficacy were predictive of dengue preventive behaviours. Our hypothesis regarding the mediational effect of self-efficacy on the relationship between knowledge on dengue and dengue preventive behaviours was supported through post-hoc probing, as recommended by Holmbeck [55]. Thus, our work would suggest that increasing the public's knowledge about dengue fever is an essential first step towards encouraging people to engage in dengue preventive behaviours. However, increasing knowledge alone would not be sufficient unless it results in increasing the level and/or strength of people's confidence in performing these behaviours. These findings are consistent with previous work on self-efficacy and healthy lifestyle behaviours [15], [38]–[41]. Our findings hold a number of important implications for health promotion authorities and planners. This is because, since both self-efficacy level and strength are modifiable and reliable mediators of health behaviours [62], health promoters should design dengue educational interventions and campaigns that promote self-efficacy as well as knowledge.
With regards to knowledge generation, we have shown that mass media campaigns such as TV, radio and newspapers and local contact (Gotong-royong, public announcements, and outdoor media) and direct face-to-face communication session (public lecture and individual advice) are the most effective. For self-efficacy, various authors have suggested that one of the most effective means of promoting self-efficacy is through modelling socially relevant enactments of the behaviours in the mass media [15], [42], [62]–[65]. For example, in order to empower people to perform dengue preventive behaviours, the health authorities could produce video-taped material with trained role-players/actors performing the desired behaviours. The specially made videos could be shown to people who do not feel confident performing the specific preventive behaviours. People could then be asked to perform the desired behaviours until a level of competence and confidence was achieved. These activities could be promoted in groups or at individual level.
We would also add that since the self-efficacy dimensions are a cognitive response to direct and vicarious experiences with the behaviours, health promotion and educational interventions should use persuasive messages on dengue prevention. This is to enable the community to translate the messages into the anticipated or actual dengue prevention behaviour (example of a persuasive message: “If it breeds, we bleed, take action! Only 10 minutes to destroy Aedes breeding sites”). This accords with Bandura's idea on health campaign messages that successfully encourage the target audience to engage in simply enacted interim behaviour which will serve to enhance self-efficacy through direct experience. In relation to dengue prevention, such interim behaviour might include both trial performances of the behaviour such as not purchasing flower pots that accumulate water, as well as low-level versions of the target dengue preventive behaviour such as putting garbage in a closed bin instead of accumulating it in a group for incineration. Tailored messages tend to be more personally relevant and thus attract more attention [66]. When recipients received messages tailored to their personal information processing style, they were later more likely to engage in the desired behaviour advocated in the message [67]–[69].
Therefore, in health promotion and educational interventions during dengue outbreaks, guidelines and protocols should outline vividly the specific persuasive messages to be conveyed to specific target audiences during the outbreaks in order to increase self-efficacy dimensions. The health promotion and educational interventions should advocate dengue preventive behaviour that takes account of the psychological characteristics of the desired behaviour and of the information-processing style of the target population. This simple strategy may lead to better crafting of persuasive messages, which in turn, increase adoption of dengue preventive behaviour so that outbreaks and transmission are reduced.
In considering the generalizability of our study it should be noted that our respondents were predominantly based in rural villages that had recently experienced a dengue outbreak. Our findings may not be applicable in areas where there had been no prior experience of dengue fever. However, given the fact that the vast majority of the world's currently at risk population will have had prior experience, our findings ought to remain applicable. Clearly our study was conducted in a rural Malaysian population and so there are issues about whether the results can be generalised to urban populations or to rural populations in other countries with different cultures. It should also be noted that because the structural equation model was based on cross-sectional data, there should be some degree of caution on the interpretation of causal inferences. Nevertheless, it should be noted that these causal paths were hypothesized based on available research concerning predictors of dengue preventive behaviours during the outbreaks [12], [70], [71]. There remains an issue regarding possible reporting bias in the data collection, especially as the interviewers were also involved in the health promotion activities. The importance of not leading the respondents in any questions was stressed during the interviewers training sessions in order to minimise this source of bias. Although we cannot say definitively that there was no bias in our data collection, we would argue that the complexity of the model and the relationships between self-efficacy and knowledge independent of any particular health promotion activity would suggest that interviewer bias would have been unlikely to have played a major role in the main findings of this study.
Our findings address the important need for studies that generate empirically sound and theoretically relevant data to identify variables likely to be effective for designing interventions. Further research should aim to describe other aspects of psychological variables related to behaviour changes and maintenance in relation to control of dengue fever, such as the complex role of motivation as well as perceived barriers and perceived benefits to engaging in the target behaviours [72]–[76]. Furthermore, we should also consider future intervention studies that evaluate different mediation effects of level and strength of self-efficacy as separate psychological components in predicting other health behaviours to control or prevent public health diseases.
In conclusion, our research indicates that the impact of public health campaigns designed to increase the adoption of behaviours by the public to control dengue fever by increasing knowledge is mediated by the impact on self-efficacy. We argue that to be most effective public health campaigns should be designed to maximise the impact on self-efficacy. There is a strong need for further research on how to design public health campaigns for the control of vector-borne disease that maximise self-efficacy and not just knowledge.
|
10.1371/journal.pcbi.1006641 | Bayesian inference of protein conformational ensembles from limited structural data | Many proteins consist of folded domains connected by regions with higher flexibility. The details of the resulting conformational ensemble play a central role in controlling interactions between domains and with binding partners. Small-Angle Scattering (SAS) is well-suited to study the conformational states adopted by proteins in solution. However, analysis is complicated by the limited information content in SAS data and care must be taken to avoid constructing overly complex ensemble models and fitting to noise in the experimental data. To address these challenges, we developed a method based on Bayesian statistics that infers conformational ensembles from a structural library generated by all-atom Monte Carlo simulations. The first stage of the method involves a fast model selection based on variational Bayesian inference that maximizes the model evidence of the selected ensemble. This is followed by a complete Bayesian inference of population weights in the selected ensemble. Experiments with simulated ensembles demonstrate that model evidence is capable of identifying the correct ensemble and that correct number of ensemble members can be recovered up to high level of noise. Using experimental data, we demonstrate how the method can be extended to include data from Nuclear Magnetic Resonance (NMR) and structural energies of conformers extracted from the all-atom energy functions. We show that the data from SAXS, NMR chemical shifts and energies calculated from conformers can work synergistically to improve the definition of the conformational ensemble.
| Proteins are commonly built up by folded domains connected by regions with higher flexibility. The interdomain orientations encoded by such hinges or linkers can play central roles in controlling the function of multidomain proteins, which makes them important to characterize. Small Angle X-ray Scattering (SAXS) is uniquely suited to study the conformational ensembles adopted by these kinds of proteins. However, because of the limited information provided by SAXS, ensemble models must be built by combination with other information sources and care have to be taken to avoid constructing ensembles that are more complex than data can support. We developed a method based on Bayesian statistics that combine data from molecular simulation with experimental data from SAXS and Nuclear Magnetic Resonance while automatically balancing the complexity of ensemble model with information in the data. We demonstrate that this method is capable of accurate inference of ensembles even in the presence of high levels of experimental noise. The method represents a general approach to combine data and simulation in the modeling of protein ensembles and can be extended to employ additional sources of experimental information.
| Proteins are highly dynamic systems [1] often with large scale conformational dynamics facilitated by regions of flexible or disordered amino acid sequence linking stably folded structured domains [2]. Close to half to the proteins coded in the human genome contain significant disordered regions of greater than 30 residues [3] and there is a multitude of multi-domain proteins with shorter flexible linkers or hinges that are important for their biological function (e.g.: in enzyme catalysis [4, 5], DNA damage signalling and repair [6], DNA binding and allosteric signalling [7], mechanical properties in the giant protein muscle protein titin [8, 9], target recognition by the intracellular regulatory Ca2+-receptor calmodulin [10], and ubiquitin-mediated regulatory mechanisms [11, 12]. These multi-domain proteins connected by flexible regions are difficult to characterize structurally as they tend to be resistant to crystallization, too large for NMR solution structure techniques and often present ambiguous results for microscopy techniques.
The small-angle scattering (SAS) from proteins in solution samples the time and ensemble average of the randomly oriented structures present. For mono-dispersed macro-molecules of uniform size, one can reliably extract accurate structural parameters such as the radius of gyration (Rg), molecular weight (M), the probability distribution of inter-atomic distances (P(r) vs. r), and an estimate of the molecular volume [13, 14]. Advances with 3D structural modelling against SAS data have further provided more detailed structural interpretation and yielded important biological insights (reviewed in Trewhella et al. [15]). This success has been achieved in spite of the fact that the SAS profile from a protein in solution represents the rotationally averaged 3D structure, hence directional information is lost leaving only 1D distance information that generally can fit multiple 3D solutions. Further, the SAS profile is a smooth function that decays rapidly and can be adequately defined by as few as 10–15 points [16]. When experimental errors are taken into account, the information content is further reduced and it is not uncommon that only 5–10 parameters can be extracted from a SAS profile [17]. Successful 3D modelling against SAS data thus depends upon restraining the conformational space to be sampled by a priori knowledge of protein structure and wherever possible by other experimental data.
In the event that a structural ensemble is present, the values of the structural parameters determined and any optimized individual 3D model will represent a population weighted average. Given the abundance of multi-domain proteins with structurally undefined linking sequences, and the difficulty in characterizing them, ensemble or multi-state modelling against SAS data is an increasingly popular choice (see reviews [18–20]. However, the problems arising from the limited information content of the SAS profile are many times amplified with the ensemble model. An ensemble model of 3D structures will have many more degrees of freedom than a single 3D model. As a result, ensemble modelling against a SAS profile is even more vulnerable to over-fitting and over-interpretation, even considering limits to the conformational space to be sampled via restraints such as knowledge of domain structures, specific flexible regions, contact information from NMR, cross-linking or FRET measurements, etc.
The objective of ensemble modelling is to return a set of structural models and their corresponding population weights. Conceptually, we can divide this process into two steps: model selection and weights inference. In model selection we determine the size of the ensemble and which members of the structural library to include. In weight inference the population weights of the selected ensemble is determined. In practice, these steps are often done simultaneously, using minimization of the difference between observed and predicted experimental data as guiding principle (often measured as χ or χ2). A number of different approaches has been presented to limit ensemble sizes and overfitting. MultiFoXS [21] optimizes χ for a given number of conformers (usually in the range 1–5) from which a minimal ensemble can be defined. The Sparse Ensemble Selection (SES) method [22] finds an optimal ensemble using linear least squares with a regularization term to obtain a sparse ensemble of conformations. Overfitting can also be combatted by using model comparison metrics like Aikake Information Criteria (AIC), an approach used by Bowerman et al. [23] to select optimal ensembles in their Bayesian ensemble modelling method. For highly flexible systems such as intrinsically disordered proteins, a small number of conformers cannot realistically describe the ensemble. Methods like EOM [24] result in sizable shrinkage of the initial structural library but do not explicitly limit the ensemble size. The use of discrete protein conformations can also be avoided altogether in the modelling of flexible proteins by using a generative probabilistic model of protein structure in Bayesian modelling [25]. A more extensive discussion of approaches for model selection and weight inference is found in the review by Bonomi et al. [26].
Because SAS data does not contain enough information to infer the full ensemble as it is sampled in solution, we choose to find an ensemble that is “optimal” in the sense that it is the simplest model that explains the available experimental data while avoiding fitting to noise. In this study we use model evidence [27] or marginal likelihood, to select ensembles with optimal sets of members. Model evidence (ME) is widely used in Bayesian model comparison and provides an automatic Occam’s razor effect [28] by balancing between fit to data and model complexity, thereby providing a rigorous approach to combat overfitting. However, ME is a multidimensional integral that can be very difficult to evaluate, which is a significant barrier to its use in ensemble selection. Our ensemble selection method is based on an approximate, variational Bayesian inference (VBI) method for model selection pioneered by Fisher and colleagues who used the method to infer ensembles of intrinsically disordered protein from NMR chemical shifts and residual dipolar couplings [29]. The VBI approach has two major benefits. First, it is significantly faster than complete Bayesian inference, which enables the use of large structural libraries. Second, VBI implicitly leads to maximization of ME without the need for evaluation of a multidimensional integral. A downside of the VBI approach is that it involves a few approximations in the probabilistic model. Hence, after arriving at the optimal ensemble with VBI we carry out a complete Bayesian inference of weights which we use to quantify uncertainties in the ensemble model and population weights.
Here, we first demonstrate the feasibility of Bayesian inference based on large structural libraries from detailed all-atom simulations. By inferring ensembles from synthetic data we show that the method is capable of accurate recovery of population weights and ensemble sizes. We then investigate how noise in the experimental data impacts the accuracy of ensemble inference, showing that information encoded in energy functions can compensate for noisy SAS data. The inference machinery is then applied to evaluate conformational ensembles of two well-characterised proteins, previously studied by SAXS and NMR, each having two domains connected by a flexible linker: calmodulin (CaM) and a two-domain construct, designated ΔmC2, from the cardiac myosin binding protein C. A significant benefit of Bayesian methods is that multiple experimental observations along with simulations and force fields can be rigorously combined in both model selection and weight inference to gain insight into the underlying ensemble. This approach is exemplified in the study of our two example proteins where we demonstrate how data from SAXS, NMR and structural energy values of individual conformers can be combined into one probabilistic model for improved ensemble inference.
We seek to determine optimal structural ensembles from experimental data by selecting conformers from a structural library and inferring their population weights. The experimental measurements generated by a discrete ensemble of conformers can be modelled as a weighted sum of measurements expected from each conformer
m→(x)=∑i=1nwiM→(x)
(1)
where M→(x) is the expected measurement for a single conformer i over a sampling point x and wi is the population weight of conformer i. For SAXS measurements M→(x)=I→(q) where I→(q) is intensity defined for scattering vector amplitude q. The objective of the Bayesian methodology is to infer the population weights wi on the basis of experimental measurements m→ and a set of structural models, which can be done by employing Bayes’ theorem
f(w→|m→,S)=f(m→|w→,S)f(w→|S)f(m→|S)
(2)
where f(w→|S) is the prior probability of weights w→=[w1,…,wn], S = {S1, …, Sn} is a structural library, f(m→|w→,S) is the likelihood of observing the measurements given the weights and set of structures, and f(w→|m→,S) is the posterior probability of the weights given the experimental measurements.
The likelihood function measures how well a given model matches experimental data. In our modeling, we assume that the experimental errors are normally distributed with standard deviations that can be estimated from the data, and that the individual data points are independent. We primarily focus on experimental data from SAXS but also employ chemical shift data from NMR. SAXS and NMR data can easily be combined by multiplying their respective likelihood functions.
Finally, we need to define a prior distribution over the weights w→. It is convenient to use Dirichlet distribution, which guarantees that weights sum up to 1
g(w→|α→,S)=Γ(α0)∑i=1nΓ(αi)∏i=1nwiαi-1
(3)
where αi are the parameters of the Dirchlet distribution and α0 is the sum of αi’s. At this stage we assume that all conformers are equally likely in the modeling and chose αi’s as the non-informative Jeffrey’s prior. However, if a more realistic energy function has been used to generate the structural library it is possible to bias the inference towards those conformers with favorable energies. In a scenario where several structurally different conformers have very similar scattering curves, such energy data can be used to select a more realistic ensemble. There are several different approaches that could be used to employ structural energy data in the ensemble inference. Our preference is to bias the prior probability distribution over weights by energy values from simulations. The structural energy values can be used to predict the population weights based on the Boltzmann distribution
wi=e-(Uref+Ui)/kT∑ine-(Uref+Ui)/kT=e-Ui/kT∑ine-Ui/kT
(4)
where Ui is the energy of conformer i. Uref can be thought of as a variable that shifts the energy measured by the energy function onto the absolute energy scale but does not affect the populations. By using a Dirichlet distribution with concentration parameters αi=e-(Uref+Ui)/kT, the prior can be centered around the Boltzmann values, with Uref controlling the sharpness of the distribution. We assign a uniform prior to the hyperparameter Uref and treat it as sampling parameter.
Once likelihood and prior distributions are defined it is possible to evaluate the posterior probability distribution by employing Markov Chain Monte Carlo sampling. However, when large structural libraries are used there can be thousands of parameters in such probabilistic models, which make complete Bayesian inference computationally intractable. We therefore use variational Bayesian inference to shrink the size of the ensemble to a more tractable size range, at which point a complete Bayesian inference is applied to infer population weights.
The goal of model selection is to determine the size of the ensemble and which members of the structural library to include. In variational Bayes, the true posterior probability distribution is approximated by a distribution with a favorable mathematical form. The parameters of this approximate distribution are found by minimizing the difference to the true posterior. This can be achieved by minimizing the Kullback-Leibler (KL) divergence between the true and approximate distribution: two identical distributions have zero KL-divergence. The KL-divergence cannot be easily evaluated, but it turns out that minimizing the KL-divergence is equivalent to maximizing a lower bound on the value of the model evidence (ELBO, denominator in Eq 2):
f(m→|S)=∫f(m→|w→,S)f(w→|S)dw→
(5)
We can find an analytical form for ELBO, which means that the inference problem can be turned into an optimization problem that is much more computationally tractable than sampling.
Maximizing ELBO thus also leads to maximization of the model evidence function, which is a central property in Bayesian model selection. Consider two possible subsets of structures (or, mathematical “models”) S(1) and S(2) from a structural library. To compare the models, we can calculate the ratio of likelihoods of the competing models given experimental data (the Bayes factor)
f(S(1)|m→)f(S(2)|m→)=f(m→|S(1))f(m→|S(2))f(S(1))f(S(2))=f(m→|S(1))f(m→|S(2))
(6)
where the second identity comes from assuming that each model is equally probable a priori. Thus, finding the most likely model given the experimental data is identical to selecting the ensemble with the highest model evidence. As demonstrated by Fisher and colleagues [29], the variational approach can be used to build a straightforward model selection approach along these lines: with a given structural library the KL-divergence is minimized by maximizing the ELBO. Members of the ensemble with lowest population weights (below preset wcut threshold) are pruned and the calculation is repeated on the reduced ensemble until the ELBO no longer increases, at which point the optimal ensemble has been identified.
To carry out the inference we need to approximate the posterior distribution over the weights w→. In the variational approach we assume that the posterior probability distributions over the weights can be well described by a Dirichlet distribution (Eq 3) and ELBO is maximized by optimizing the concentration parameters αi. The choice of the Dirichlet distribution to approximate the posterior results in a closed-form solution for ELBO [29]. Simulated annealing is then used to maximize with the respect to the concentration parameters αi. The population weights are then calculated as
wi=αi∑iαi
(7)
These weight estimates are compared to the cutoff value in the model selection algorithm.
Our method enables optimal ensemble selection from large structural libraries using variational Bayesian inference. Before we demonstrate the full potential of the model selection, we first demonstrate the power of model evidence to identify optimal ensemble sizes when it can be accurately calculated (not approximated). To illustrate the concept, we generated synthetic data and a structural library of ten members from discrete structural models of the two-domain construct ΔmC2 from cardiac Myosin Binding Protein C (which will be described in more detail below in the context of the applications with real experimental data). We created an ensemble of 3 arbitrarily selected models from the set of ten and simulated a combined scattering curve for these models. Using these simulated data and a structural library of 10 members, we calculated the model evidence for all possible ensembles with 2,3 and 4 members. Fig 1A shows the maximal model evidence as a function of model size. As expected, model evidence picks out 3 as the most optimal ensemble size.
We then investigated the ability of VBI to accurately recover the correct ensemble and population weights using synthetic data based on a structural library of ΔmC2. From a larger structural library of 1000 conformers a smaller library of 100 was generated by selecting structures that covered a similar distribution of radius of gyration (Rg) values to the larger library. From this subset a handful of structures (5 models) was selected, each with an arbitrarily chosen population, to generate synthetic experimental data. Gaussian statistical errors were added to the data according to the method described by Karaca et al. [30].
A key challenge in ensemble inference is to identify the optimal set of members. This step can be very difficult because even with a relatively small structural library of 100 members the number of possible ensembles is staggering; e.g. there are 1010 unique ensembles available having 1–7 members. S1 Fig illustrates the process of ensemble inference by the algorithm on synthetic data generated with 5 members and added synthetic noise starting from the 100-member structural library. VBI recovers the correct members of the ensemble and their corresponding weights. Although the recovery of weights in this example is impressive, there are a couple of caveats. One is that ensemble members with small population weights may be prematurely pruned during iterations of the ensemble selection algorithm. This simple algorithmic issue could be corrected by optimizing the threshold used to cull members from the ensemble. But there is also a more fundamental issue with uniqueness of the ensemble. In a bigger structural library, there will be conformers with nearly identical scattering profiles. As the size of the structural library increases, the exact identity of members in the ensemble may not be recovered. When we expand the library from 100 to 1000 members this behavior is indeed observed. However, the alternative ensembles recovered in this case have similar model evidence to the simulated ensemble and are thus equally optimal.
Synthetic ensembles allow us to characterize the effects of experimental noise on the ensemble selection, such as reduced accuracy of population weights inference or a reduction in information content in the data that leads to a smaller number of members of the ensembles that can be supported by the data. Information content in a SAS curve has traditionally been estimated using information theory by calculating the number of Shannon channels needed to completely recover the data [31]. However, this approach does not take into consideration the effect of noise. Such effects can be evaluated by calculating the “number of good parameters”, Ng, instead. Ng provides the number of parameters that can be determined from measurements and can be estimated from data using maximum entropy regularization [32]. Vestergaard and Hansen [33] have developed a Bayesian approach to evaluate Ng for SAXS data, an approach we employ here. Based on the synthetic ensemble with 5 members we increased the amount of synthetic noise applied to the data and calculated Ng. VBI was then applied to these data to recover optimal ensembles. Fig 1B shows the size of the ensemble as a function of added noise. Ng for the simulated data is around 6 and drops down to 4 at the highest levels of noise. At lower noise levels all 5 ensemble members are recovered. However, increasing noise leads to smaller inferred ensembles with only two members at the highest noise levels. A second effect of increasing noise is a change in the identity of the recovered ensemble members. As the noise increases and the size of the ensemble is reduced, the original ensemble members are not necessarily part of the optimal ensembles.
To further investigate how noise affects the accuracy of inference we repeated the above model selection with synthetic data and signal-to-noise levels set with reference to the experimental data for ΔmC2 (described below). In Fig 1C the accuracy of the inferred weights, characterized by the root mean square deviation (rmsd) between simulated and inferred weights, is plotted as a function of increasing noise in the data. The results demonstrate that the inference is still very accurate up to three times the experimentally observed noise in our example ΔmC2. As the added noise increases beyond this value the number of inferred ensemble members decreases, which is the primary reason for the rapid increase in error in rmsd.
So far, we have assumed that all conformers are equally likely in the modeling. However, we can also bias the inference with the energies generated for conformers from the structural library. In our simulations, Uref, which controls the strength of the prior, is selected by optimizing evidence using a variational Bayes approach. In this way, the uncertainty in the experimental data will automatically control the strength of the energy prior. This effect is demonstrated by carrying out inference with an energy prior that is centered around Boltzmann weights whose values differ from the simulated values. When the noise level is low and the information content high in the experimental data, the inference relies strongly on the experimental data with small rmsd differences between inferred and simulated weights. As the noise levels increase and the information content is reduced, the energy prior takes over and the weights move towards the values predicted by the Boltzmann distribution (S2 Fig).
By establishing the impact of inference with structural energies on the fixed set of models, we further investigate the power of using structural energies on model selection in the presence of experimental noise. In Fig 1D we show the result of the inference of a synthetic ensemble of 5 lowest energy conformers from a library of 100 members as a function of noise. In the absence of the energy prior, the number of recovered members from the simulated ensemble is reduced to 4 and 3 as the noise increases. With the energy prior turned on, the full ensemble is recovered at much higher levels of noise. This result is obtained even when the Boltzmann weights did not exactly match the simulated population weights. However, due to the different weights, the rmsd relative to the simulated weights is slightly higher with the energy prior turned on.
In order to demonstrate that the introduction of energy priors does not steer the resulting ensembles excessively towards the lowest energy structures, we added an energy refined conformer with substantially improved energy to the library. With this addition, there was little effect on the identity of recovered models (S3 Fig) and the trend observed in Fig 1D is retained.
Once a smaller subset of models has been selected using VBI, we subject the optimal ensemble to Complete Bayesian Inference (CBI) to determine the population weights and their distributions. In general, a strong benefit of Bayesian inference is that we can go beyond single values (point estimates) for population weights and characterize the complete posterior probability distributions of inferred parameters. This step provides probability distributions over the individual weights in the ensemble, together with credibility intervals if requested.
It is also possible to characterize the uncertainty of the complete ensemble. Fisher and colleagues [29] developed a useful metric to measure the uncertainty of ensembles, the expectation value of the Jensen-Shannon divergence (JSD) relative to the optimal weights over the posterior distribution
σw→B,S=∫JSD(w,→w→B,S)f(w→|m,→S)dw→
(8)
where JSD(w→,w→B,S)=12∑i=1nw→ilog2(2w→iw→i+w→B,Si)+12∑i=1nw→B,Silog2(2w→B,Siw→i+w→B,Si) and ranges between 0 and 1 for two maximally identical and different vectors, respectively, which means that also σw→B,S falls within this range.
We carry out the complete Bayesian inference using the No-U-Turn sampler (NUTS) [34] implemented in the Stan software library [35]. NUTS is an extension of Hamiltonian Monte Carlo, an MCMC algorithm that avoids the random walk behavior and sensitivity to correlated parameters that often plague MCMC inference.
To validate the inferred ensembles, it is useful to carry out posterior predictive checks [36]. This check can be achieved by repeatedly simulating scattering curves with the inferred ensemble model and then comparing these to the experimental data. As seen in S4 Fig, experimental curves simulated by our statistical model closely match experimental data. For example, when ensembles are inferred using an unsuitable error model, it is immediately obvious in these predictive checks.
Having characterized the performance of Bayesian inference methods on synthetic data sets with relatively small structural libraries, we now apply the method to two experimental systems from our previous work: a two-domain protein calmodulin (CaM) [14], and the two-domain construct, ΔmC2, from the cardiac myosin binding protein C (cMyBP-C) [37].
CaM is the major intracellular Ca2+ receptor that binds to a diverse array of target proteins (numbering in the 100s) to regulate their activities in response to Ca2+ signals (reviewed by Tidow et al. and Crivici et al. [10, 38, 39]). The crystal structure of CaM [40] shows a mostly α-helical structure with an unusual dumbbell shape formed by two globular, cup-shaped domains connected by an extended α-helix of 7–8 turns. Upon Ca2+-binding at the base of each cup-shaped domain a hydrophobic cleft, which is essential for target binding, opens via the concerted movements of pairs of helices. NMR studies showed the interconnecting helix is broken in solution by a short sequence of four highly mobile amino acids [41] that allow CaM to orient and position the hydrophobic clefts and additional contact regions to accommodate structurally diverse targets. Thus CaM’s structure encodes for both structural diversity and specificity for target binding. CaM was chosen as a test case because it is an extensively characterized protein and understanding the nature of the conformations present in solution for uncomplexed CaM and how that conformational equilibrium is influenced by the presence of binding partners is thus of considerable interest. It is also a popular target for molecular dynamics (MD) simulation, including studies aimed both to gain insight into CaM dynamics (e.g. [42–47] and to test MD results against experiment (e.g. [48]).
To generate a library of structurally and energetically reasonable conformers of CaM (which herein refers to the Ca2+-saturated form with the four Ca2+ sites fully occupied, and thus primed for target binding) we developed a Monte Carlo based simulation of linker flexibility. A sampling protocol was developed in the Rosetta macromolecular modeling package where the torsion angles in the linker segment were sampled in a Monte Carlo simulation followed by an all atom energy refinement of the linker segment and the neighboring residues. In addition, the 3 N-terminal residues and the last C-terminal residue (lysine 148) missing in the crystal were modelled de novo as well. Around 10000 models were generated by this procedure and a structural library was created by taking the lowest energy 1000. The distribution of Rg-values in the structural library for all 10000 models and after applying energy filter is shown in S5A and S5B Fig. The Rg distribution for the lowest energy subset models covers the same Rg range as for the complete library but is slightly more peaked.
Using a high quality SAXS data set of CaM obtained using in-line SEC (size exclusion chromatography) at the Australian Synchrotron [14] and NMR chemical shift data [49], we performed model selection using VBI with the 1000 lowest energy conformers. We evaluated four inference scenarios using: 1) SAXS data only, 2) SAXS data + Rosetta energies, 3) SAXS data + chemical shifts and 4) SAXS data + chemical shifts + Rosetta energies. Once VBI converged and the ensemble consisting of a few members was selected, we used CBI to infer population weights and their distributions. While condensing the probability distributions into point estimates (single values) of parameters is undesirable in general, it is sometimes convenient in comparison with alternative methods to easily summarize error residual plots and evaluate other figures of merit. For this purpose, we calculate scattering curves for inferred ensembles using point estimates of parameter taken from the VBI inference. These point estimates are found as the parameters (e.g. population weights) that maximizes the ELBO metric.
Each of the ensembles inferred with the prior distribution unbiased by the inclusion of energies (scenarios 1 and 3) consists of 4 members (Fig 2A and 2C), while the scenarios with the Rosetta energies included for the prior distribution (2 and 4) result in 3 members (Fig 2B and 2D). The drop in the number of members upon inclusion of energy priors is due to the peaked energy landscape, which reduces the number of possible solutions and also results in faster convergence of selection algorithm (S1 Table). Inferred weights for each scenario have relatively peaked distributions (Fig 2E–2H) and JSD ranges from 0.05 to 0.08, which means that there is high certainty in the predicted parameters given the ensemble of models and the experimental errors.
The predicted scattering profile from each of the ensembles for the different inference scenarios matches the SAXS data well, as illustrated in Fig 2 (panels I-L) and a number of statistical measures. The reduced χ2 value obtained for the predicted scattering profile for each ensemble is in the expected range for an excellent model fit to the data (i.e. near 1; in this instance in the range 0.81–0.87). The use of energy priors leads to a small increase in χ2 in the presence and absence of the CS data. The addition of CS data slightly improves the fit to the SAXS data compared to when SAXS data is used alone, indicating the data sources are at least not in conflict and potentially may be reinforce each other. The absolute value of χ2 depends critically on accurate counting statistics and error propagation. Further as a global parameter, χ2 will not identify significant regions in q-space of mis-fit. The predicted scattering profiles were therefore also assessed (1) using an error weighted difference plot over the measured q-range and (2) with the recently developed correlation map (CorMap) test [50] that is independent of the errors and identifies regions of misfit with a significance test. Simply put, CorMap identifies the longest stretch of data points that lie on one side of the model profile and provides a probability (P) for that occurrence given the number of points in the data set. Consistent with the observed flatness of the error weighted model versus experiment intensity difference plots (Fig 2M–2P) over the entire q-range, CorMap gives P-values indicating high confidence in the model fit (0.53–0.96). Thus by all measures each of the inferred ensembles are in excellent agreement with the SAXS data, have high certainty in the predicted parameters. Arguably, one could conclude that the “best-fit” to the SAXS data is obtained for scenario 3 (SAXS data + CS) as assessed by the lowest χ2 value combined with the highest P-value and the fact that the longest stretch of points on one side of the model profile lies, uniquely among the four scenarios, in the high-q background scattering region. All parameters for the inferred ensembles are summarized in S2 Table.
Examining the CaM conformers in each selected ensemble, with a single exception, the Rg values are all in the relatively narrow range 20.6–23.0 Å (S2 Table). This range is consistent with the original SAXS study of CaM in solution [51] that concluded that the CaM lobes are on “average” reoriented and closer together in solution compared to the crystal structure (PDB 1CLL) with its fully extended helical inter-domain connector (Rg = 22.7 Å). The main distinction among the inferred Rg distributions is that the inclusion of Rosetta energies results in a higher proportion of more compact structures within this range, although the SAXS + Rosetta energies inference also yields the most extended conformer with an Rg value 26.0 Å, albeit with a relatively low population weight (0.06 ± 0.1).
The conformers of the inferred CaM ensembles all show variable orientations of the N- and C-terminal target-binding hydrophobic clefts and variable degrees of extension in the flexible linker (Fig 2A–2D). Inspection of known crystal or NMR solution structures of CaM complexed with target binding proteins or domains also reveals conformers with highly variable domain dispositions (reviewed in Tidow et al. [10]). They also include CaM conformers that are significantly more compact or more extended than either the crystal structure or those present in the majority conformers from inferred ensembles; e.g. CaM with its binding domain in myosin light chain kinase has an Rg of 17 Å with its two globular lobes wrapped tightly around the helical binding domain (PDB 2LV6) while the 20 lowest energy NMR structures for CaM complexed with its binding domain from Munc13 (PDB 2KDU) includes CaM conformers with Rg values as large as 26.4 Å. A systematic comparison of all CaM conformers represented in complexes with binding partners in the PDB identified 1 crystal structure (4DJC) and 3 NMR solution structures (1CFF, 2KDU and 1L53) with similar dispositions of the CaM domains as assessed by rmsd values for Cα coordinates in the range 4.6–7.3 Å (S3 Table). Of this set of structures, only the 2KDU structure has both CaM binding domains involved in the target domain interaction, the remaining three only involve C-terminal domain binding, and the 1CFF crystal structure has the fully extended helical inter-domain connector, similar to the Ca2+-CaM 1CLL structure. A library of CaM structures was generated from all the structures in the PDB of CaM complexed with a target involving interactions with both of CaM’s N-and C-terminal domains. When inference is carried out with this structural library, the resulting ensemble cannot describe the experimental data well.
In sum, each of the inferred ensemble models show variable dispositions of the target-binding hydrophobic clefts and includes some conformers that have similar dispositions to conformers observed in crystal or NMR solution structures of CaM complexes. Further, the Rg values for the ensemble model conformers are all in a range that is within the range observed in these structures. However, each inference scenario results in distinct set of conformers in an ensemble that fits the available data more-or-less equally well. Thus, while the model evidence justifies an ensemble model of 3–4 models, the solution is not uniquely defined by the available experimental data.
This ambiguity can be potentially removed by introducing additional experimental data that informs on inter-domain orientation. Such information is found in data from NMR Paramagnetic Contact Shifts (PCS) and Residual Dipolar Couplings (RDCs) measurements for example, and has proven to be useful in combination with SAXS [52, 53]. Developing methods required to incorporate this type of data into our statistical framework is beyond the scope of this study. However, we can test how well the ensembles inferred in this study explain experimental PCS values from paramagnetic data. We compared predicted values from inferred ensembles with available paramagnetic data for Tb (terbium(III)), Dy (dysprosium(III)) and Tm (thulium(III)) bound to the N-terminal domain of CaM derivatives [54]. The predicted ensembles do not fit particularly well with the PCS data for the C-terminal domain. This could be because PCS reports on orientational information not available in SAXS and chemical shift data. However, the conditions at which the PCS data is significantly different than used for SAXS (pH (6.5 vs 7.5) and ionic strength (300 vs 400 mM)). Since CaM is very negatively charged [55], it cannot be ruled out that the ensembles are different at these two conditions.
It is the hydrophobic cleft in the C-terminal lobe of CaM that is generally the initial recognition site for target binding in a two-step binding process whereby subsequent N-terminal lobe binding is necessary for full cooperative target binding. Further, it is not unusual for the CaM binding sequences to be anchored via other interactions within the target proteins; e.g. in myosin light chain kinase the CaM-binding domain has to be released and translocated away from the kinase’s catalytic cleft [56], and in CaM’s interaction with the MA protein from HIV-1 the two-tryptophan’s that bind to the C- and N-terminal domains of CaM are deeply buried in the helical head domain of MA [57, 58]. The ensemble models thus support the idea that the flexible linker in CaM primarily allows the hydrophobic clefts to reorient independently. This mobility enables target recognition and binding by the C-terminal hydrophobic cleft of CaM that in turn triggers the unfolding and folding events required to form the interaction surfaces. Such a process is consistent with the conclusions of Liu and colleagues from their molecular dynamics study of CaM binding to its binding domain in skeletal muscle myosin light chain kinase, that the binding process is “quite complex with the mixture of induced fit, conformational selection, and simultaneous binding–folding.” [42].
Our second example of the application of VBI to experimental data considers ΔmC2 from cMyBP-C, which has never been crystallized but our NMR solution structure (PDB:2KDU) [37] reveals it to have a two-domain structure with a 7-residue flexible linker. The cMyBP-C is a modular protein with eleven predominantly β-structured immunoglobulin (Ig) or fibronectin (Fn) domains (designated C0 through C10) and a 100-amino acid sequence between C1 and C2 that contains cardiac specific phosphorylation sites and is mostly unstructured (referred to as the “motif” or m-domain) [59, 60]. Found in the cross-bridge bearing C zone of the A band of the muscle sarcomere, cMyBP-C interacts with both thick and thin filaments and has both structural and regulatory functions [61]. It exercises its regulatory function via alternate myosin/actin interactions with its N-terminal domains (C0-C1-m-C2), with phosphorylation of the motif implicated in the switching [62–64]. The ΔmC2 construct includes the loosely structured C-terminal region of the m-domain that is a tri-helix bundle [65] with a tightly structured C2 that has an Ig-type fold [66]. Our NMR structure showed the same folded tri-helix bundle as previously determined by NMR and the C2 domain connected by a 7-amino acid linker that is highly mobile, and yet there is a surprisingly high degree of sequence conservation in this linker sequence across all known chordates [37]. Further, the linker includes sites of severe disease-linked mutations and also forms part of the interface of a stable, Ca2+-dependent interaction with CaM. These observations, combined with evidence implicating ΔmC2 in actin binding, led us to postulate that, like CaM, the flexible linker region of ΔmC2 may facilitate its role as a polymorphic binding domain that interacts with multiple proteins to regulate muscle action in the sarcomere [37].
SAXS and NMR chemical shift data for highly purified ΔmC2 were from [37]. The SAXS data were of good quality, also from the Australian Synchrotron, but measured in a typical batch mode without the benefit of in-line SEC. A small concentration dependence was observed in the lowest-q data that, while corrected by a linear extrapolation to zero concentration, amplified the errors in this region.
Following the procedure described for CaM, and assuming two stable folded domains connected by a 7-residue linker, we generated a structural library of 1000 lowest energy conformers using the Rosetta protocol and ran the same 4 inference scenarios: 1) SAXS data only, 2) SAXS data + Rosetta energies, 3) SAXS data + chemical shifts and 4) SAXS data + chemical shifts + Rosetta energies. The ensembles inferred in scenarios 1 and 3 consist of 5 members (Fig 3A and 3C), while scenarios 2 and 4 (Fig 3B and 3D) yield 3 and 4 members, respectively. Similar to CaM, model selection when Rosetta energies are included in the prior leads to a smaller subset of inferred models.
The Rg range in each of the inferred ensembles is similar (~17–27 Å). As was observed for CaM, inclusion of Rosetta energies distributions significantly alters the weighting of more compact structures to more extended ones (0.80–0.84 and 0.42–0.58 without and with energy priors, respectively). In contrast to the CaM, however, the change in weights with energy priors shifts the distribution to an increase in the proportion of more extended structures. The most highly extended conformer (Rg = 27.0 Å) appears in all four variants (model 1 (green) in Fig 3A–3D) though its population weight with the inclusion of both CS and energy priors (inference 2) is significantly smaller than in the other ensembles. In all scenarios except 3, which is the only one for which a conformer with the intermediate Rg value (24 Å) is absent, inferred weights have a peaked distribution over the weights and JSD ranges from 0.04 to 0.07. The JSD is slightly higher for variant 3 (0.11), primarily due to the long tail of the lowest weight, even so it still corresponds to an ensemble that is well-defined.
Similar to CaM we can assess the fit to data based on a point estimate of weights from VBI (Fig 3I–3L). Compared to CaM χ2 are considerably higher (ranges from 3.55 to 3.81), although error weighted difference plots (Fig 3M–3P) and CorMap P-values values (0.19–0.81) indicate good fits to the data over the measured q-range with no statistically significant specific region of mis-fit. We can thus conclude that the errors propagated from counting statistics were on this occasion underestimated, which has been a common issue for SAS data. χ2 drops when simulations include Rosetta energies in the prior over weights (variant 2 and 4). In interpreting this result, it is important to highlight that the ensemble and population weights are not selected by minimizing χ2. The drop in χ2 is the result in improved quality of the ensemble and highlights how multiple data sources can work together to provide a better-defined ensemble. Inference with chemical shift data leads to slightly increased χ2 for SAXS of 3.81, suggesting that the ensemble observed by SAXS and NMR chemical shift may differ somewhat, potentially due to subtly different solution conditions. The detailed values of inferred parameters can be found in S2 Table. To further investigate this issue, we ran CBI with the ensemble only selected from SAXS data with the four different data scenarios as presented above (results found in S4 Table, which also presents values for CaM). With the SAXS ensemble, inference of SAXS+NMR data is essentially identical to when only SAXS data is used. However, no improvements in the inference is observed when the Rosetta energy is used in this scenario. This highlights that differences with or without NMR and Rosetta energies is a consequence of identifying different conformers from the structural library with the additional data.
The ensemble members in each of the scenarios 1, 2 and 4 adopt 3 distinct conformations that upon aligning the tri-helix bundle form an approximate cross-like configuration, while those from scenario 3 form an approximate T-shaped configuration (Fig 3A–3D). However, given that the inference with energy priors have better match to SAXS data as well as lower JSD values we can conclude that the ensemble with cross-like conformation is more likely. Much less is known about ΔmC2 and its putative binding partners. The measured binding affinities are moderate (~100 nM) compared to CaM (~nM) [37] and, to date, there is no evidence for a common recognition motif. The ensemble modelling indicates that the longer flexible linker in ΔmC2 compared to CaM allows for significantly greater flexibility and relative positioning of its two domains, and more highly extended conformers are favored. Such an ensemble may be optimized for binding targets with moderate affinity where there is not a common initial recognition motif, and the binding process will also involve a mixture of induced fit, conformational selection, and simultaneous binding–folding.
Many methods have been proposed for building conformation ensembles from SAS data. Typically, ensembles have been optimized by minimizing χ2. The fits are then characterized by visualization of fitting residuals. We compared the results from point estimates of weights from VBI with two popular methods for conformational ensembles modeling from SAS data: Ensemble Optimization Method or EOM [67] and MultiFoXS [21]. The results were summarized in terms of Rg distributions, number of ensemble members, χ2 and CorMap P-values (S5 Table). Focusing on the CaM ensembles obtained with SAXS-only data, with and without energies for the VBI ensembles, we see a striking similarity between the Rg-values of conformers and weights between MultiFoXS and VBI for SAXS-only results. In contrast, the EOM and SAXS+Rosetta energies ensembles are more similar to each other, differing from the MultiFoXS results in the relative proportions of the more compact and more extended conformers. The inclusion of CS data does not significantly alter the VBI results in terms of Rg values and weights. For MultiFoXS, the minimal number of conformers required to minimize χ2 is selected and all structures that have correct stereochemistry, while for EOM a genetic algorithm is used to find an ensemble that minimizes χ2 and flexible regions are treated simply as a self-avoiding polyglycine chain. Thus, as might be expected, the number of ensemble members selected by VBI is much smaller than the number of representative structures selected by EOM but larger than for MultiFoXS. In the case of EOM the Rg distribution for the ensemble is a continuous double-peaked distribution that is represented by 13 conformers from this distribution, which is more than twice the number from the other methods.
While we have compared the χ2 values for the ensemble model fits to the SAXS data here, it is important to keep in mind that in contrast to EOM and MultiFoXS, the Bayesian approach does not select ensembles and weights based on direct minimization of χ2/χ and uses chemical shift and energy data in addition to data from SAXS in the inference. Nonetheless, by this comparison we see that the resulting χ2 values for the SAXS data fits are similar those obtained using EOM and MultiFoXS.
Small angle scattering data can provide structural insights into conformationally heterogeneous biological samples. Due to its inherently low information content, SAS data typically must be complemented with structural modeling to draw biologically relevant conclusions. While we want to extract as much information from the data as possible, care must also be taken to avoid overfitting. In ensemble inference there are two areas where overfitting may become a problem. First, with structural libraries containing thousands of members the number of modeling degrees of freedom significantly exceeds the information content in the data and this can result in inferences of overly complex ensembles. Second, by optimizing model parameters directly with respect to χ2 there is a risk of fitting to noise rather than signal in the experimental data.
Model evidence provides a principled approach to balance model complexity with fit to experimental data. We demonstrate that the approach can identify the optimal number of members using simulated ensembles with a known ensemble size. Model evidence also enables investigation of how experimental noise affects the inference of optimal ensembles. Our results show that although the ensemble inference is robust to high levels of noise, increasing noise eventually leads to the reduction of the information content in the data and smaller ensembles sizes that can be supported by data. Encouragingly, the analysis of the experimental data sets reports optimal ensemble sizes that are similar to the values obtained from the analysis of the number of good parameters (Ng) suggesting that a good balance between model complexity and fit to data is reached. Model evidence is only one of several approaches for model selection employed in Bayesian inference. We have also employed model selection using WAIC and PSIS-LOO [68] but found that they did not result in stable ensemble inference.
In the simulation experiments with synthetic data, the exact identity of members in the optimal ensemble could be inferred from SAS data alone, except when the added noise became high. However, in scenarios with experimental data and large structural ensembles we do not necessarily expect there to be single optimal solution and many competing ensembles may equally well describe the experimental data. This result is not surprising as many different conformations can give rise to the same scattering profile. This is a fundamental consequence of the three-dimensional averaging of coordinates in SAS and not something that can be tackled with improved inference methods.
Bayesian approaches have some inherent properties that provide protection against overfitting to noise by balancing the fit to experimental data with information encoded in prior distributions over model parameters. The protection from the prior is particularly important in situations where the amount of experimental data is limited. Another benefit of the Bayesian methodology is that it returns probability distributions over modeling parameters rather than point estimates. Point estimates of population weights are a convenient approach to summarize results but represents an unnecessary reduction of information. The posterior probability distributions provide information about uncertainty of individual population weights. This can be complemented by the JSD metric that summarize uncertainty over the complete ensemble. We find small JSD values overall, suggesting relatively well-defined ensembles. Altogether, the posterior probability distributions and the JSD metric gives a full picture of the uncertainties in the ensemble inference given the available data.
Our approach for ensemble inference involves two separate stages. First, fast model selection is carried out using a variational approach that enables Bayesian inference with structural libraries consisting of thousands of members. This is followed by a complete inference the selected set using a full Bayesian inference. Comparison of the weight inference for CaM and ΔmC2 using the variational and complete suggests that the two approaches gives highly similar results, indicating that the approximations used in the variational method do not lead to any significant inaccuracies.
A powerful approach to better define ensembles is to include additional data into the inference and thereby increasing the information content. An additional benefit is that different data sources can provide different types of structural information. SAS provides information about relative positions of atoms in a structure. NMR chemical shift data on the other hand provides information about local structure of the protein while energies calculated through a force field or energy function provides information about stereochemistry and intermolecular interactions in the protein. The Bayesian approach straightforwardly enables the use of several information sources simultaneously in the inference. Our study of the two-domain proteins CaM and ΔmC2 with data from SAXS and NMR chemical shifts as well as Rosetta structural energies shows that for ΔmC2 that had higher levels of noise in the low-q SAXS regime, the use of Rosetta energy information leads to a significant improvement of the inference. The resulting ensembles have more peaked population weights distributions, better fit to the SAXS data (measured through χ2), fewer members and the Monte Carlo simulations converge faster. For the more ideal CaM data, we also observe more peaked probability distributions, fewer member and faster simulation convergence but see no improvement with the inclusion of energy priors in model fit to SAXS data measured through χ2. The inferred ensembles using SAXS only, SAXS+chemical shifts and SAXS+chemical shifts+structural energy have some conformers in common, but are different enough to present an alternative view of the conformational states of the proteins. Because the different inference scenarios are based on different data input, it is not straightforward to compare them statistically. Nonetheless, the ensemble inferred from the SAXS+chemical shifts+structural energies has the strong benefit that the conformers are consistent with the distance distributions measured through SAXS, the torsional preferences of the linker assessed by NMR and are energetically and stereochemically realistic through the use of the Rosetta energy values. When SAXS data is used alone, there are many ensembles with almost identical model evidence. Because of the lack of orientational information in the SAXS data, such ensemble can be quite different. The additional information from NMR and Rosetta can then tip the balance between these competing ensembles.
In reality we do not expect proteins with flexible linkers to populate only a discrete number of conformational states. The inferred ensembles represent a simplified model for explaining the dominant conformational states adopted by the protein. The small ensemble sizes are a reflection of the limited information content in the data which is not sufficient to infer more detailed picture of the conformational landscape. A fuller picture of the conformational ensemble could emerge if discrete structural library is replaced by a continuous model for structure. Antonov et al. have developed a probabilistic model for protein structure that enables sampling of conformations of the protein during ensemble inference [25], a method that does not rely on structural libraries. The challenge in employing such approaches is the development of probabilistic models over structure that samples energetically realistic protein conformations. For this reason, the use of structural libraries generated by atomistic force fields and energy functions still represent a useful strategy for inference of structural ensembles. Further research is necessary to develop approaches that combines the rigor of complete Bayesian inference with the structural and energetic realism encoded in force fields and energy functions.
In order to apply Bayes’ theorem (Eq 2) to infer the population weights wi on the basis of experimental measurements m→ and a set of structural models S, we need to state the prior probability f(w→|S), and the likelihood function f(m→|w→,S). We define a prior probability over the weights w→ as Dirichlet distribution (Eq 3). The αi parameter that defines Dirichlet distribution is either chosen to assume that all conformers are equally probable (non-informative Jeffrey’s prior) or to bias toward lower energy conformations from Rosetta simulations. For the non-informative prior the probability density function is defined as:
f(w→|S)=Γ(n/2)nΓ(1/2)∏i=1nwi-1/2
(9)
However, when Rosetta energies are used the prior probability equals to:
g(w→|S)=Γ(β0)∑i=1nΓ(βi)∏i=1nwiβi-1
(10)
where βi=e-(Uref+Ui)/kBT, Uref is the Boltzmann reference energy, kB Boltzmann constant and β0=∑i=1nβi.
The likelihood function describes uncertainty in experimental data. For SAXS data with normally distributed errors it is defined for each measurement mj as a Gaussian density function:
f(mj|w→,λ)=12πεSAXS2exp(-(mj-λ∑inwiIij)22εSAXS2)
(11)
where λ is a scaling factor, Iij is a SAXS intensity calculated from the ensemble and εSAXS is the experimental error. We assume that measurements are independent and the joint likelihood is the product of individual likelihood functions:
fSAXS(m→|w→)=∏j=1Nf(mjw→,λ)
(12)
where N is the number of experimental measurements.
The Bayesian framework provides an easy approach to add structural information from different experimental sources. In the case of NMR chemical shifts measurements, we also assume that measurements are normally distributed and uncertainty of theoretical prediction of chemical shifts εCS can be summed up with experimental errors εpre.
fNMR(mj|w→)=12πεCS2+εpre2exp(-(mj-∑inwiCij)22(εCS2+εpre2))
(13)
where Cij are chemical shifts calculated from the ensemble. Similar to SAXS data we assume that NMR chemical shift measurements are independent and joint probability fNMR(m→|w→) is the product of individual likelihood functions.
The overall goal of variational Bayesian inference is to maximize the model evidence f(m→|S). This is typically intractable, but we can find a lower bound for model evidence (ELBO) by introducing an approximate posterior g(w→|α→,S) and applying Jensen’s inequality to the model evidence and maximize that instead [29]:
logf(m→|S)=log∫g(w→|α→,S)f(m→|w→,S)f(w→|S)g(w→|α→,S)dw→≥∫g(w→|α→,S)logf(m→|w→,S)f(w→|S)dw→g(w→|α→,S)dw→≡-L(α→|S)
(15)
ELBO is determined by maximization of -L(α→|S) or minimization of L(α→|S) (Eq 16) through the choice of the parameters of the approximate distribution g(w→|α→,S). In this way the parameters of g(w→|α→,S) are chosen to minimize the KL divergence to the true posterior f(w→|α→,S). The choice of g(w→|α→,S) as a Dirichlet distribution enables a closed form solution for L(α→|S). The derivation for NMR chemical shift data can be found in Fisher et al. [29]. We modified the method to accommodate SAXS data:
L(α,S)=logΓ(α0)Γ(n2)+∑i=1nlogΓ(12)Γ(αi)+∑i=1n(αi−1/2){ψ(αi)−ψ(α0)}+1/2∑j=1Nεi−2(mj−λ/α0∑i=1nIijαi)2+12∑i=1n∑j=1n(∑k=1NIikIjkεk2)αi(α0−αi)δij−αiαj(1−δij)α02(α0+1)
(16)
where δij is Kronecker delta function, ψ(·) is digamma function and λ is a scaling factor between experimental and ensemble averaged inferred measurements calculated according to the formula described in Svergun et al. [69]:
λ=∑j=1Nεj-2α0-1mj∑i=1nIijαi∑j=1Nεj-2(α0-1∑i=1nIijαi)2
(17)
When the Rosetta energies are used in the inference, L function has the following form:
L(α,S,Uref)=logΓ(α0)Γ(β0)+∑i=1nlogΓ(βi)Γ(αi)+∑i=1n(αi−βi){ψ(αi)−ψ(α0)}+1/2∑j=1Nεj−2(mj−λ/α0∑i=1nIijαi)2+12∑i=1n∑j=1n(∑k=1NIikIjkεk2)αi(α0−αi)δij−αiαj(1−δij)α02(α0+1)
(18)
In each round of the model selection algorithm the L function is minimized for the current set of conformations S by identifying the optimal set of parameters αi and Uref (when Rosetta energies are available) using simulated annealing. After finding the optimal weights through the αi parameters, the conformers with lowest weights are removed from the ensemble by applying a cut, wcut (fixed at the start of the simulation, explicit values are provided in S1 Table), so that conformers with wi < wcut are culled from the set. This procedure is repeated, and the simulation stops when the L function does not improve in 10 iterations. In the case of SAXS-only data and SAXS with NMR chemical shifts we restart optimization several times, starting from the set of structures from previous run until the L function did not improve (see S1 Table). When running simulations with structural energies this was not necessary because the algorithm converged in a single run. Because of the stochastic nature of the algorithm the inferred ensemble may not always converge to the same set of structural models and population weights. We repeated entire procedure 2 to 4 times depending on the data type used in the inference to monitor convergence and selected solutions with the lowest L. We implemented VBI using openmp library allowing for parallel computation, which provides considerable speed up compare to original method by Fisher et al. [29].
Once the small subset of models has been selected using VBI, we determine corresponding population weights with complete Bayesian inference. We based CBI implementation on the Stan library—platform for statistical modeling and high-performance statistical computation [35]. The weights w→, scaling factor λ and parameter defining the shape of Boltzmann distribution Uref are sampled using Markov Chain Monte Carlo (MCMC) simulations. In each run we performed 2000 simulations with No-U-Turn sampler [34] using 4 chains and 4 jobs. We monitored MCMC simulations by inspecting the effective sample size and split R^ parameter, which are diagnostics available directly from Stan. In addition to these metrics, we used a few statistics from stan_utility (https://github.com/betanalpha/jupyter_case_studies/blob/master/pystan_workflow/): trajectory tree depth, energy Bayesian fraction of missing information and posterior parameters divergence.
VBI and CBI were implemented with python and C++ and are available from: https://andre-lab.github.io/bioce/ as well as through web-server: https://andre-lab.github.io/bioce/webserver.html.
In the case when model evidence was explicitly evaluated and not approximated we performed numerical integration of the integral from Eq 5:
∫f(m→|w→,S)f(w→|S)dw→
(19)
This was calculated by determining the expectation value of the likelihood function f(m→|w→,S) evaluated on the weight values sampled from prior distribution f(w→|S) (Dirichlet distribution).
We used the FoXS [21] program to calculate scattering profiles from atomic coordinates of conformers. In cases where experimental data was available scattering profiles were calculated for experimental q values, otherwise we used equally spaced q values ranging from 0 to 0.5 1/nm (default in FoXS). Scattering profiles calculated on experimental q values were subsequently descaled by dividing intensities with the c1 scaling parameter (returned by FoXS) to have equally scaled intensities for the Bayesian inference. To predict NMR chemical shift data Cij and their uncertainties εCS from the set of structural models we used the SHIFTX2 program [70]. Python scripts for generating scattering profiles and chemical shift data and converting them to the required input format are available with the software.
To generate a library of energetically reasonable conformers of ΔmC2 and CaM we developed a sampling protocol in Rosetta macromolecular modeling package [71]. The protocol samples torsion angles in the linker segment using Monte Carlo simulations (1000 iterations) and subsequently repacks side chains. The linker modeling was followed by all atom energy refinement of the linker segment and the neighboring residues with fast relax protocol [72]. Around 10 000 models were generated by this procedure and the 1000 lowest energy conformers constituted the lowest energy structural library.
In order to demonstrate that presence of low energy conformer does not considerably bias simulations towards Boltzmann weights, we used the Rosetta Relax protocol to optimize energy of one of the ΔmC2 models. Constraints on atomic coordinates were introduced to ensure that model did not substantially deviate from its starting conformation so that the scattering pattern of the energy-refined model is highly similar.
NMR chemical shifts measurements for CaM were described in [73] and the data was obtained from Biological Magnetic Resonance Data Bank (BMRB Entry 547). This data was recorded for CaM from Drosophila, which differs from human CaM in three amino acid positions: Y99F, N143T, and T136S. We excluded these three substitutions in our simulations by omitting them in experimental and predicted chemical shift data. SEC-SAXS data for CaM are deposited in the SASBDB (https://www.sasbdb.org/), identifier SASDCQ2, and fully described in [14] an open access article for which the CaM data are publicy available under the uniform resource identifier https://creativecommons.org/licenses/by/2.0/uk/legalcode. SAXS data for ΔmC2 are deposited in SASBDB (identifier SASDDD9), while NMR chemical shift data was taken from [37].
The web version of MultiFoXS (https://modbase.compbio.ucsf.edu/multifoxs/ [21]) and the ATSAS on-line version of EOM (https://www.embl-hamburg.de/biosaxs/atsas-online/ [67]) were used to obtain the multi-state and ensemble optimization modelling results, respectively, for CaM and ΔmC2 shown in S5 Table. The crystal structure coordinates of CaM (PDB:1CLL) and Model 1 from the NMR ensemble for ΔmC2 were the starting structures (PDB:5K6P). In the case of CaM the 3 missing N terminal amino acids (Ala1, Gln2, Asp3) from the crystal structure and the flexible linker (Lys77, Asp78, Thr79, Asp80, Ser81) were assigned unknown structure. In the case of ΔmC2 the 7-amino acid flexible linker (Arg356, Arg357, Asp358, Glu359, Lys360, Lys361, Ser362) was assigned unknown structure. MultiFoXS generates structures for the unknown regions that have correct stereochemistry, while for EOM the random coil option was chosen to model the missing amino acids. The SAXS data used for modelling were for CaM SASBDB ID SASDCQ2, q = range 0.0066–0.3104 Å-1, and for ΔmC2 SASBDB ID SASDDD9.
The amount of structural information covered by SAXS experimental data was assessed using the BayesApp program [33]. We included all data points in the analysis and used default input parameters.
The radius of gyration for individual models was calculated using CRYSOL program from ATSAS package [74].
|
10.1371/journal.pcbi.1005720 | Limb-Enhancer Genie: An accessible resource of accurate enhancer predictions in the developing limb | Epigenomic mapping of enhancer-associated chromatin modifications facilitates the genome-wide discovery of tissue-specific enhancers in vivo. However, reliance on single chromatin marks leads to high rates of false-positive predictions. More sophisticated, integrative methods have been described, but commonly suffer from limited accessibility to the resulting predictions and reduced biological interpretability. Here we present the Limb-Enhancer Genie (LEG), a collection of highly accurate, genome-wide predictions of enhancers in the developing limb, available through a user-friendly online interface. We predict limb enhancers using a combination of >50 published limb-specific datasets and clusters of evolutionarily conserved transcription factor binding sites, taking advantage of the patterns observed at previously in vivo validated elements. By combining different statistical models, our approach outperforms current state-of-the-art methods and provides interpretable measures of feature importance. Our results indicate that including a previously unappreciated score that quantifies tissue-specific nuclease accessibility significantly improves prediction performance. We demonstrate the utility of our approach through in vivo validation of newly predicted elements. Moreover, we describe general features that can guide the type of datasets to include when predicting tissue-specific enhancers genome-wide, while providing an accessible resource to the general biological community and facilitating the functional interpretation of genetic studies of limb malformations.
| The majority of the human genome does not code for proteins. Regulatory roles have been ascribed to a growing fraction of the non-coding genome. Enhancers, short stretches of non-coding DNA, confer spatial and temporal specificity to gene expression patterns. These regions are essential to the proper development of multi-cellular organisms and, when mutated, can give rise to congenital malformations and contribute to human disease. In line with these observations, a large fraction of the genetic variation associated to developmental abnormalities cannot be narrowed down to protein-coding defects, suggesting these variants actually reside in functional non-coding regions, such as enhancers. However, identification of enhancers in the mammalian genomes that are functional in vivo remains a difficult task. Here we combine chromatin and DNA-sequence data from the mouse genome in a machine learning framework resulting in the Limb-Enhancer Genie (LEG), an accurate and easily accessible collection of predicted enhancers active in the developing limbs. LEG outperforms state-of-the-art approaches, as testified by the high fraction of newly tested elements that we validated in the developing mouse embryo in vivo. To grant the community access to our predictions and their mappings to the human genome, we established a user-friendly web-interface.
| Mammalian body plans are shaped by the precise spatiotemporal execution of transcriptional programs [1], which have been shown to rely on the coordinated activity of enhancers [2]. Despite the increased availability of epigenomic data, the identification of these short, cis-regulatory DNA sequences in the vast non-coding portions of mammalian genomes has proven to be a difficult task. Indirect measurements suggest that hundreds of thousands of enhancers populate mammalian genomes [3], but only a few thousand of them have been validated for their activity in vivo [4]. A wide range of experimental and computational approaches have been applied to the prediction of regions showing enhancer activity in vivo, including: 1) Evolutionary conservation [5]; 2) Chromatin signatures, such as the binding of the co-activator p300 [6] or the acetylation of lysine residue 27 of histone H3 (H3K27ac) [7,8]; 3) Chromatin accessibility to DNase I digestion [9]; 4) Genomic sequence signatures, such as the presence of binding sites for relevant transcription factors (TFs) [10]; 5) Combinations of the former strategies. Despite significant advancements in enhancer identification, through the generation of genome-wide datasets and their integration using supervised [11–15] or unsupervised [16,17] models, all available approaches to date suffer from one or more of the following limitations: 1) Lack of integration of chromatin and sequence features that are immediately relevant to the tissue(s) and the developmental stage(s) under consideration; 2) Lack of thorough, biological interpretation of the features driving the prediction, which in turn is a key requirement to instruct experiments and more refined models in the future; 3) Lack of appropriate negative controls–e.g. the use of random genomic intervals instead of regions showing (at least partially) a known signature of enhancers but failing to display tissue-specific activities when tested in vivo; 4) Lack of user-friendly access to the de novo predictions, limiting the value of the resulting resources for the community of experimental as well as computational biologists.
In this work we integrate multiple machine learning approaches in order to produce robust predictions of enhancers active in the developing limbs of mouse embryos at embryonic day 11.5 (E11.5). By focusing on this well-studied developmental system [18,19], we are able to overcome the limitations described above and outperform previously described state-of-the-art approaches [11,12]. First, we exclusively considered datasets generated from embryonic limbs (with one exception, a DNase I hypersensitivity dataset from headless embryos) at the relevant developmental time points (E10.5, E11.5 and E12.5), including the binding profiles for CTCF, the cohesin complex, and a large panel of histone modifications [3,6,8,20,21]. Among the latter we also included recently published ChIP-seq data from specific limb compartments [22]. Importantly, we trained statistical models that provide intrinsically interpretable measures of feature importance (LASSO and random forests). This allowed us to identify a previously unreported feature capable of significantly improving predictions, namely limb-specific DNase I enrichment. The predictive power of this feature was confirmed using data from other tissues (central nervous system and facial prominence) at the same developmental time point. We additionally trained models based on clusters of evolutionary conserved binding sites for those TFs expressed in the developing limbs, and formally integrated these results with the chromatin features described above. We used a set of >200 validated limb-enhancers and ~2,000 negatively tested regions corresponding to either validated elements active in tissues other than limb or that were previously selected based on chromatin or sequence features of active enhancers but failed validation due to absence of reproducible reporter activity [4]. Based on our results, we were able to confirm the in vivo activity of three out of four newly predicted enhancers in the vicinity of the Hand2 gene, an important regulator of limb morphogenesis [23,24].
Importantly, our genome-wide predictions can be queried through a user-friendly web interface named LEG (Limb-Enhancer Genie), which is available at http://leg.lbl.gov/. Since a large fraction of limb developmental enhancers are evolutionarily conserved between human and mouse [6], the user can also input regions from the human genome. The complete set of predictions along with all the sequencing datasets re-analyzed in this study are available for browsing via a public track hub (see Methods) on the UCSC genome browser [25].
By providing the community with significantly improved genome-wide maps of the enhancer landscape underlying limb development, our results will assist the functional interpretation of genetic studies assessing human developmental diseases. Moreover, the analysis of the feature importance in the trained models provides novel generalizable insights into the chromatin signature of developmental enhancers that will help guide the design of predictive models in tissues other than limb.
DNase I accessibility and H3K27ac are routinely used to identify tissue-specific putative enhancers in the human and the mouse genomes [7,8,26,27]. We first aimed to determine the sensitivity and specificity of these marks, using the developing limb as a test case (Fig 1A). Even when used in combination, these marks suffer from both high false positive and false negative rates. More than 50% (1,094/1,967) of the limb-negative elements in the VISTA collection [4] overlapped H3K27ac-enriched or DNase I accessible regions (false positives). At the same time, a fraction of enhancers truly active in the limbs at E11.5 (18/234) were still missed by both assays (false negatives). These results prompted us to set up a more integrative approach towards more effective limb enhancer discovery.
To this aim, we took advantage of >50 recently published limb-specific, genome-wide datasets (S2 Table) from four major categories of chromatin features (namely DNase I accessibility, six histone modification and co-activator p300, CpG methylation, and the binding of CTCF/cohesin, Fig 1B). We chose to use the limb as a model based on the extensive available chromatin data [3,8,20,21], which includes robust time series spanning three closely spaced developmental time points (E10.5, E11.5 and E12.5) and hundreds of in vivo validated elements [4]. These datasets comprise two subregion-specific sets (representative of two important signaling centers in the developing limb, namely the Apical Ectodermal Ridge, or AER, and the Zone of Polarizing Activity, or ZPA, [22]). Importantly, including DNase I digestion patterns from whole embryos whose heads had been removed [28] allowed the estimation of tissue-specific DNase I enrichment scores. In addition to the chromatin state, we also incorporated one class of sequence features, i.e. the predicted clusters of evolutionary conserved binding sites for those TFs expressed in the developing limbs.
In order to better understand the relative contribution of chromatin and sequence features, our strategy first considered them separately, and then in combination (Fig 1C). To improve robustness, we partitioned our set of 234 limb enhancers (positive examples) and 1,967 regions negative for activity in the limbs (1,025 negatively tested regions and 942 showing activity in another tissue, Fig 1A and S1 Table) into ten equally sized bins with constant ratio of positive to negative observations (where one bin was used in turn as a test set, whereas the remaining nine constituted the training set). The model parameters were learnt using 10-fold cross-validation over the training set. The models trained include least absolute shrinkage and selection operator (LASSO, [29]), support vector machines (SVM, [30]) and random forests (RF, [31]). Model performances were compared across the ten independent test sets. Predictions from different models and distinct sets of features were then combined using ridge regression or a weighted sum of ranks approach (SOR). For the final prediction of enhancers genome-wide, the training step was re-iterated using the entire dataset, and the resulting models and their combinations were used to call enhancer regions using a sliding window (see Methods).
LASSO and RF allow evaluation of the importance of each feature for model performance. This enabled us to gain insights into the biological relevance of the most predictive features. A notable novelty in our method is the use of a previously unappreciated feature, a score measuring increased tissue-specific DNase I accessibility (DNase I enrichment), into our predictive models. Unexpectedly, we found the headless embryo DNase I accessibility pattern to be well correlated with the DNase I specific for fore- and hindlimbs (r = 0.87 and 0.85, respectively, within the VISTA-dataset). This prompted us to hypothesize that including the ratio between the DNase I signal from limbs and the headless embryos would better capture the limb-specific changes in DNase I accessibility, as compared to the limb DNase I signal alone. Unsupervised clustering of the training set followed by visual inspection of the results provided additional evidence to support this hypothesis and revealed further interesting groups (S1 Fig). Next to clusters of negative elements showing predominantly low values across all features, the set contained groups of negatively tested elements showing known chromatin signatures associated with regulatory elements other than enhancers. One group contained mostly promoter-like elements (high H3K4me3 and H3K9ac), while others resembled insulators (high CTCF and Smc1a) or Polycomb-associated heterochromatin (high H3K27me3) [32,33]. These qualitative observations prompted us to include these chromatin features, along with the DNase I enrichment score as described above, into our machine learning strategy.
We first built models of increasing complexity, starting from p300 alone and incrementally adding H3K27ac, DNase I, DNase I enrichment and all the remaining chromatin features (Fig 2A and 2B). The median AUROC (Area Under the Receiver Operating characteristic Curve) as well as the median AUPRC (Area Under the Precision Recall Curve) steadily increased by including more features (as assessed on the independent test sets). This was observed consistently across the different models. Considering the one showing the highest performances (namely the radial SVM), the median AUPRC when training only on p300 was 0.372, a figure that increased to 0.412 when including H3K27ac, to 0.502 if considering also DNase I, and finally to 0.542 and 0.545 if adding the DNase I enrichment or all the features, respectively (Fig 2B, S3 and S5 Tables). Interestingly, models trained only on H3K27ac/p300 and DNase I (including the enrichment over the head-less embryo samples) reached a performance almost as high as the full set of chromatin features on the VISTA dataset. However, these additional features are well-known marks for categories of regulatory elements–e.g. insulators and promoters–that are under-represented in our training set but are widespread genome-wide. In fact, by overlapping the CTCF-bound sites in the developing limbs (which are enriched for insulator sequences) [21] with the de novo enhancer predictions genome-wide using different feature subsets, the effect of including the additional features became more evident. While models trained on H3K27ac/p300 and DNase I alone showed 10 to 20% overlap with CTCF-bound sites across a wide range of predicted values, models trained on the complete set of features showed less than 5% (S2 Fig), potentially removing a fraction of false positive predictions.
A qualitative evaluation of complex loci (Fig 2C) indicated that validated limb enhancers show a higher DNase I enrichment (DNase I versus Headless embryo) as compared to nearby regions that tested negative in vivo. Given these observations, we then sought to systematically and quantitatively assess the relative importance of the different chromatin features. Specifically, the estimated coefficients from the LASSO and the mean decrease in accuracy estimated by the RF were evaluated. We also estimated a selection probability for each predictor by Bootstrap LASSO (see Methods). The results are summarized in Fig 3 and S9 Table. DNase I accessibility as well as the limb-specific DNase I enrichment were systematically co-selected (>0.98 selection probability by Bootstrap LASSO) and showed the highest coefficients as well as importance in the RF. Co-selection of these two features further supports a substantial rather than incremental role of tissue-specific DNase I enrichment in identifying active enhancers in vivo. Interestingly, while both the DNase I enrichment values from hind- and forelimbs were often selected and assigned positive coefficients, the forelimb DNase I signal was only selected in 2 out of 1,000 bootstrap samples, in contrast to the the hindlimb DNase I signal which was selected every time. We re-ran the Bootstrap LASSO after removing the hindlimb DNase I and found that the DNase I from forelimbs was selected with a probability of one and almost identical performances. This indicates that the dataset from hindlimb might be favored for technical rather than biological reasons. Other features that were selected with high probabilities, but less often than DNase I associated features, were p300, CTCF, H3K27ac, H3K27me3 and H3K4me3 (all showing a selection probability >0.6). The small size of the training set combined with the multiplicity of classes of regulatory elements, each represented only by a few examples (S1 Fig), was likely responsible for the lower selection probabilities of these features. As expected, p300 and bulk H3K27ac are the most important predictors after DNA accessibility (selection probability >0.75). On the other hand, when we assessed the contribution of the H3K27ac datasets specific for different sub regions of the limb, we found that the ZPA-specific profile was very unlikely to be selected, in contrast to the AER-specific one, which was often included in the LASSO models with a positive coefficient. This region-specific feature is selected >75% of the time, often together with H3K27ac profiles from whole limbs. Thus, our findings highlight the importance of gathering chromatin information at a finer scale in order to be able to identify enhancers with more sub-regional-specific activity. The histone modification H3K4me3 (and to a lesser extent H3K9ac, which are both usually found at promoter regions [33]), was assigned a negative coefficient, as was the mark H3K27me3, which has been associated with inactive, poised enhancers [34] or more generally with Polycomb-associated heterochromatin. The two proteins CTCF and Smc1a, while often co-bound to DNA on the same genomic elements (mainly insulators or promoters) [32], are assigned coefficients of opposite sign (negative for CTCF, positive for Smc1a), indicating that cohesin but not CTCF is more generally associated with enhancer function. Smc1a was assigned a high importance for the prediction of limb-enhancers by the RF, while it was selected only in 35% of the bootstrap-samples by LASSO. CpG methylation was also found to be rather important in the RF predictions, but very unlikely to be included in the LASSO models. A possible explanation could be the implicit accounting for feature interactions in the RF, which remain unappreciated by the LASSO. Nevertheless, the small size of the training set impinged our ability to explicitly tease out these combinatorial relationships. Taken together, these results are in line with the expectation that different models are able to capture distinct aspects of the data. This prompted us to combine the results from the multiple models into a single, unified predictive score. Two different approaches were applied to this end: ridge regression (i.e. finding optimal weights to combine the outputs from the single classifiers) and an approach based on the weighted sum of the individual output-ranks. These strategies led to a significant improvement in the AUPRC as compared to RF (p-value = 0.03, one-tailed Wilcoxon signed-rank test, Fig 4B) and a smaller improvement when considering LASSO or linear SVM (p = 0.05, combined ridge model), but it was not significant as compared to radial SVM (p = 0.57, combined ridge model).
We then asked whether considering clusters of evolutionary conserved TF-binding sites could lead to a more consistent increase in the predictive power of the combined models. In order to limit the number of input features, we only considered binding motifs for TFs expressed in the developing limbs (see Methods, S3 and S4 Tables). The overall performances of the sequence features alone were markedly lower than those achieved by chromatin (Fig 4A and 4B, S3, S4 and S5 Tables). Even the application of gkm-SVM [35,36], a motif-agnostic machine learning approach which has been previously applied to enhancer prediction [37], lead to performances comparable to those achieved by our models (median AUPRC of 0.197, S11 Table and Discussion). However, combining our sequence- and chromatin-based predictions using ridge regression significantly outperformed the combined chromatin model (Figs 4B, S3 and S4; p = 0.019 in terms of AUPRC, one-tailed Wilcoxon signed-rank test; p = 0.08 when considering the combined SOR) and the best single chromatin model (namely the radial SVM, p = 0.024, one-tailed Wilcoxon signed-rank test). Model performances across the test sets are reported in S5 Table. Similar to the chromatin features, we combined the importance measurements from the LASSO and the RF, and short-listed the most relevant TFs (Fig 4C and S10 Table). Interestingly, these TFs are over-represented in publications in the field of limb development as compared to TFs whose motifs were selected less frequently or not selected at all (Fig 4D, p-value = 0.02, Mann-Whitney test, Top vs No). Among the most frequently selected motifs are those of Hoxa13 and Hoxd9, which are known regulators of digits and stylopod development, respectively [38]. Tp63 is a critical factor for epithelial development that, when mutated, can lead to severe developmental defects including complete absence of limbs [39]. Tfap2a has been previously associated with distal outgrowth of the developing limbs [40]. While Tp63 and Tfap2a have been associated to limb development, these results suggest they might exert their function by binding to enhancer elements. These findings underscore the importance of applying interpretable machine learning approaches to highlight relevant features, in turn helping to formulate new experimental hypotheses.
By combining the predictions from both the chromatin- and sequence-based models, our strategy outperformed the state-of-the-art approaches [11,12] both in terms of AUPRC and AUROC (Table 1). In line with this, the distributions of predicted values between the positive and negative regions in the training set showed a stronger separation in our combined models as compared to EnhancerFinder and EMERGE (Kolmogorov–Smirnov statistic, S5 Fig). We also compared the performance of our combined models to the predictive power of the strong enhancers chromatin states defined using two ChromHMM models [17], trained using eight histone modifications from two distinct biological replicates [41] from E11.5 limbs (see Extended Methods). This resulted in a recall of 0.162 and a precision of 0.447 for the enhancer calls from one of the two replicates. At the same level of recall, our combined models reached a much higher precision (0.787 and 0.741 for the ridge regression and the SOR approach, respectively). The second replicate led to comparable conclusions (recall of 0.256 and precision of 0.417 for the ChromHMM calls, as compared to a precision of 0.732 and 0.710, given that level of recall in our combined models). These results prompted us to train the models using the complete set of observations and to run them genome-wide. The mouse genome was tiled into overlapping windows of 2kb, which were assigned prediction values for tissue-specific enhancer activity in vivo using all models. The resulting predictions for each single model (either LASSO, SVM or RF) and type of feature considered (chromatin, sequence) as well as their combination were ranked, and the 20,000 highest scoring regions were binned into 10 groups (see Methods and S8, S13 and S14 Tables). These were used to evaluate the enrichment for proximity to genes involved in limb development and expressed in Theiler stages 19 and 20 (corresponding to the window from E11 to E13) using GREAT [42] (Figs 5A and S6). The enrichments from the combined predictions were higher than those of any single model trained on chromatin or clusters of conserved TF-binding sites alone, indicating that the combined models can identify thousands of bona fide, previously uncharacterized, enhancers. Interestingly, while showing lower performances on the test sets as compared to the combined ridge classifier (Fig 4B), the SOR showed the highest enrichment in terms of proximity to genes relevant to limb development (Fig 5A), especially for the highest-ranking elements (S6 Fig).
To further corroborate our predictions, we searched the literature for developmental limb enhancers that were robustly validated in vivo but are not part of the VISTA collection. We identified five elements, all of which overlapped the 10,000 highest scoring predictions (considering either the ridge regression or the sum of ranks, S12 Table). One of these regions is the ZRS (ZPA Regulatory Sequence), which is a well-known enhancer controlling the expression of Shh in the ZPA [43]. This element consistently ranked very high across both the sequence- and chromatin-based predictions, in line with the abundance of conserved TF-binding sites and the presence of a strong limb-specific DNase I signal (Fig 5B, left panel, and S12 Table). A further, independently identified element that drives the expression of Tfap2a in limbs and face [44] was correctly predicted within an intron of the Tfap2a gene itself (Fig 5B, right panel).
We then additionally verified the ability of the proposed approach to identify bona fide limb-enhancers by choosing four newly predicted elements close to the developmental regulator TF Hand2 and testing them in vivo through mouse transgenic enhancer-LacZ reporter assays (Fig 5C). Recently published promoter-Capture-C data [45] from developing limbs at E11.5 revealed that these elements are indeed located in a domain contacting the Hand2 promoter with high frequency (2/4 reaching statistical significance, S7 Fig) demonstrating their potential to act as enhancers for this gene. Hand2 displays critical developmental functions in various embryonic tissues such as the limb [23,24], the heart [46–48] and the craniofacial structures [49]. However, the Hand2 limb-specific enhancer landscape has been poorly characterized so far. Three out of four tested elements displayed reproducible LacZ reporter staining at E11.5, with patterns of activity specific to limbs and overlapping well-known subdomains of Hand2 expression [50]. Interestingly, the only element that tested negative also showed the lowest predicted combined score (Fig 5C and S15 Table).
Finally, we made the genome-wide predictions available at http://leg.lbl.gov/. These can be directly and systematically queried through a user-friendly interface. The website also provides two tutorials that leverage published datasets that were not used in the predictions.
In this work, we integrated >50 genome-wide chromatin datasets with sequence information and were able to improve our ability to recognize limb enhancers over previously published approaches. These include EnhancerFinder [11] and EMERGE [12] which represent computational state-of-the-art tools in the field (Table 1) and to our knowledge are the only two studies that employed the VISTA dataset in a way that is comparable to our approach. Combined with making the predictions readily available via a user-friendly interface, another advantage of the presented study is the extensive application of machine learning models (LASSO, RF) that are intrinsically designed to provide feature importance. In this way, we have shown that including a limb-specific DNase I enrichment score dramatically improves the prediction of developmental limb-enhancers in terms of both precision and recall over incorporating just the commonly used histone-mark H3K27ac (Fig 2). The availability of both DNase I accessibility and H3K27ac ChIP-seq data for midbrain, hindbrain, neural tube and facial prominence tissues at E11.5 (S17 Table) allowed us to reproduce this finding also in tissues other than limb at the same developmental time point. To this aim, we fit logistic regression models including p300, H3K27ac, DNase I and DNase I enrichment features consecutively, and found that inclusion of the DNase I enrichment scores on top of the other features considerably and significantly improved the performance of all the neuronal tissues and to a lesser extent of facial prominence (S8 Fig). Our ability to evaluate the performances for craniofacial enhancers is affected by the lower number of validated examples (74 versus 274, 310 and 196 for hindbrain, midbrain and neural tube, respectively) leading to greater variance in the overall performance. In addition to the DNase I enrichment score, the chromatin features H3K4me3, H3K9ac, H3K27me3 as well as the binding of CTCF and Smc1a were selected as informative (Fig 3). The inclusion of features other than H3K27ac and DNase I accessibility reflects the presence of VISTA elements showing combinations of these genomic features in the developing limbs, including many that failed to validate in vivo. These are very likely to be insulators, unannotated canonical promoters or poised enhancers, rather than active enhancers (S1 Fig). As many regions in VISTA were selected specifically because they showed canonical enhancer marks, these other classes are likely under-represented in our training data. Nevertheless, these types of elements could be misclassified as enhancers when only DNase I and H3K27ac are used for genome-wide scanning, and our approach proved effective at exploiting this information. When moving to the scale of genome-wide prediction, the small but significant improvements in performance observed when including all these features lead to important differences in the types of elements that are predicted. For example, models trained on H3K27ac and DNase I only are more enriched for insulator-like elements, as indicated by a larger overlap with CTCF-bound regions (S2 Fig).
Of note, LASSO also systematically selected with a positive coefficient the H3K27ac dataset specific for the AER sub region (Fig 3). This demonstrates the value of sub-regional-specific datasets. Nevertheless, there are two issues that need to be addressed in order to fully harness this kind of datasets in the future. First, the current number of tested VISTA enhancers (and more in general in the literature) showing activity for each different sub-region is still low. Second of all, there are very few high-quality genomic datasets generated from sub-regional dissected tissues available at present.
The performances of the sequence-based models were on the other hand lower than what we observed when incorporating experimentally derived, chromatin data (Fig 4). The performance of our models are seemingly lower than previously reported [37,51]. Previous publications mainly focused on the prediction of sequences with enhancer activity using randomly selected genomic regions (matched by GC- and repeat- content) as negative examples. In this study, we focused on a different problem, i.e. the identification of enhancers showing limb-specific activity, against regions that either show enhancer activity in a different tissue than limb, or anyway were selected based on partial experimental evidence of activity, but failed in vivo validation. In fact, when we applied gkm-SVM [35,36], a machine learning approach previously applied to enhancer prediction [37], we observed performances comparable to those achieved by our sequence-based models (S11 Table). More in general, the identification of transcription factor binding sites genome-wide suffers from a high false positive rate, a problem that was only partially mitigated by leveraging the information of TF-binding-sites-clustering and the evolutionary conservation of these sites. On top of this, developing tissues are complex, heterogeneous mixtures of lineages giving rise to multiple cell-types, each one of which depends on only partially overlapping gene regulatory networks. As such, the diversity of regulatory elements at the sequence level is expected to be much greater in tissues than in more homogeneous, in vitro cell populations. This factor is likely to have a major impact on the signal-to-noise ratio for the identification of sequence-encoded features. In line with this, the most important TFs identified by the sequence-based models are enriched for general regulators involved in enhancer function in the limb, like the Hox family genes, Tp63 or Tfap2a (Fig 4C and 4D). Despite these limitations, we found multiple evidences supporting the value of integrating both chromatin and sequence features in our predictive framework. These included functional analysis of the de novo predictions using GREAT [42] (Figs 5A and S4), our in vivo validation of three out of four newly predicted enhancers very likely involved in the transcriptional regulation of Hand2 (Fig 5C and S15 Table), and the recapture of previously validated limb-enhancers from a number of independent studies (Fig 5B and S12 Table). Of note, even though the incorporation of the sequence features significantly improved the predictions per se (Fig 4A and 4B), the use of evolutionary conserved TF-binding sites still led to a considerable number of false positive (S9 Fig). A more unbiased approach to mitigate all the issues highlighted in this paragraph would be the generation of high-quality ChIP-seq profiles for the cell-type-specific TFs involved in the development of the embryonic tissue under study.
Overall, our results will help instruct future strategies for the identification of enhancers. Our analysis strongly suggests that the use of a limited number of features relevant to the developing organ system under scrutiny (chromatin accessibility, high enrichment for H3K27ac and p300 binding and low to no enrichment for H3K27me3, H3K4me3 and CTCF, see Fig 3), as well as the integration of a previously unappreciated feature, the DNase I enrichment, will likely improve the prediction of enhancers active across development and showing diverse tissue and sub-regional specificity. We expect this to be the case in the near future, as soon as the relevant genome-wide datasets are generated. We envision that measuring relative chromatin accessibility across tissues by means of ATAC-seq [52] might provide the same information (and in turn the same boost in predicting bona fide enhancers) as the DNase I enrichment score proposed here. At the same time, while more sophisticated computational models could be applied [53], these are currently limited by the size of the training set. Data gathering remains the major limiting step (e.g. validation in transgenic mouse lines is still relatively low-throughput). Technological advancements to increase the throughput as well as to standardize the assays (e.g. by site-specific integration of the reporter transgene in the genome) will soon be required and extremely beneficial. Importantly, by providing the community with an easy access to significantly improved genome-wide prediction maps of the enhancers active during limb development, we anticipate these results to be of value for both developmental biologists and human geneticists. Our web-interface (http://leg.lbl.gov/) can be queried using human genomic regions. This will specifically help the functional contextualization of human non-coding variants, pinpointing their contribution to limb malformations. As an example, the LEG predictions overlapping published H3K27ac-enriched regions in embryonic human limbs [20] (S16 Table and Extended Methods) are provided.
All animal work was reviewed and approved by the Lawrence Berkeley National Laboratory Animal Welfare Committee. All mice used in this study were housed at The Animal Care Facility (ACF) at LBNL. Mice were monitored daily for food and water intake, and inspected weekly by the Chair of the Animal Welfare and Research Committee and the head of the animal facility in consultation with the veterinary staff. The LBNL ACF is accredited by the American Association for the Accreditation of Laboratory Animal Care International (AAALAC, IACUC-approved animal protocol #290008).
Human and murine validated elements were downloaded from the VISTA enhancer browser (http://enhancer.lbl.gov) [4] and mapped to mm10 coordinates using liftOver [25]. After filtering (see Extended Methods), 2,201 elements were used for machine learning (S1 Table).
ChIP-seq, DNase-seq, RNA-seq and CpG-methylation profiles collected for this study are listed in S2 Table. ChIP-seq and DNase I hypersensitivity reads were aligned to the mm10 release of the mouse genome (Dec. 2011, GRCm38) using bowtie2 [54]. ChIP-seq peaks were called using MACS v1.4.2 [55] for analysis regarding overlaps to enriched regions (not machine learning). RNA-seq datasets were aligned to the reference transcriptome (mm10, Ensembl 81 gene annotation release, [56]) using STAR v2.4.2a [57]. Transcripts were quantified with Stringtie v1.0.4 [58]. CpG-methylation bigWig tracks at base-pair resolution were downloaded from the ENCODE repository (http://www.encodeproject.org/) [3].
Log2-RPKM quantifications for ChIP-seq and DNase I samples for each one of the 2,201 mm10-mapped VISTA elements were performed after expanding them to a minimum size of 2kb around their center. For ChIP-seq samples, enrichments were computed relative to the corresponding control samples (input DNA) (see Extended Methods). Scaling of the input features was performed as z-scores. For CpG-methylation, the average fraction of methylated CpGs was determined for each region.
Position weight matrices (PWMs) [59] (S3 and S4 Tables) were limited to those representing binding preferences of TFs potentially expressed in the developing limb (see Extended Methods). Putative TF-binding sites were identified using FIMO v4.10.2 [60], with a p-value cutoff of 10−4 and using GC-content matched backgrounds (see Extended Methods). Clusters were identified using a sliding window (500bp); binding sites were weighted by evolutionary sequence conservation, as estimated by phastcons [61]. Either the mouse or the human sequence was scanned according to which version was tested in vivo (S1 Table). A complete table of the scores for each TF-gene across the 2,201 VISTA elements is provided in S7 Table.
The observations were split into ten equally sized groups. Each group was used as test set exactly once while the rest was used for training. Parameters were tuned by ten-fold cross-validation within each training set (see Extended Methods). For the chromatin data, four different classifiers were trained: 1) LASSO logistic regression [29]; 2) Support Vector Machines (SVM) [30] with linear kernel; 3) SVM with radial kernel and 4) Random Forests (RF) [31]. For the sequence data, radial SVMs were not fit. Fig 1C summarizes the modeling strategy. In order to combine the predictions from the separate models, two methods were applied (see Extended Methods for a detailed description): 1) ridge regression (i.e. finding optimal weights for the output from the single classifiers, or “model stacking”) and 2) the weighted sum of the individual output ranks. Predicted values for each one of the input observations are reported in S6 Table, overall performances in S5 Table.
For genome-wide predictions, models were fit using 10-fold cross-validation on the entire dataset. After extensive pre-processing of the values of the single features (see Extended Methods), the mouse genome was tiled into gap-less, overlapping 2kb tiles (with a step of 1kb). Tiles overlapping either gene promoters or elements in the training set were discarded. The top 20,000 disjoint elements predicted by each model (or combination) were obtained using an iterative merging strategy (see Extended Methods).
For the bootstrap LASSO, 1,000 bootstrap samples of the original data were extracted (see Extended Methods). Model parameters were estimated and selection probabilities for each feature were calculated by dividing the number of non-zero coefficients across bootstrap samples by the total number of bootstrap samples [62]. For the RF, the importance for each variable was defined as the average decrease in accuracy.
The Limb-Enhancer Genie (LEG) is an online tool (available at http://leg.lbl.gov/) aimed at facilitating the access to the genome-wide predictions generated in this study. Two separate analysis modes are available. The first one finds the overlap of a set of input regions with the top 10,000 predicted limb-enhancers. This can be used, for example, to scan large regions for potential limb-enhancers. The second one is conceived to assign scores to smaller regions (< = 10kb). For each input region, the highest scoring overlapping 2kb genomic tile is identified and returned along with its score and original coordinates. This also allows scoring of elements overlapping the training data or regions close to promoters, which were excluded from the top 10,000 reported predictions. This second mode of analysis accepts mouse (mm9 and mm10) as well as human (hg19 or hg38) regions. All the predictions along with tracks for the chromatin features and evolutionary conserved TF-binding sites (for the TF-features most correlated with activity in limb) are available on the UCSC Genome Browser [25] (for both mm10 and mm9) via the track hub available at http://portal.nersc.gov/dna/RD/ChIP-Seq/LEG_trackhub/hub.txt. The source code for training and combining the models is available for download at http://github.com/rmonti/limb_enhancer_genie/.
All the described data processing steps were performed in the statistical computing environment R v.3.2.1 (www.r-project.org). An overview of the packages used in this study along with references to them is given in the Extended Methods.
Newly tested elements were named according to the nomenclature current in use in the VISTA Enhancer Browser (http://enhancer.lbl.gov/; mm: mouse, hs: human). The elements were amplified from mouse genomic DNA and cloned into an hsp68-lacZ expression vector, as previously described [5]. Genomic coordinates are listed in S15 Table. Transgenic mouse assays were conducted as previously described [5,63]. Sample sizes were selected empirically based on past experience of performing transgenic mouse assays for >2,000 total putative enhancers [4]. Mouse embryos were excluded from further analysis if they did not encode the reporter transgene or if the developmental stage was not correct. All transgenic mice were treated with identical experimental conditions. Randomization and experimenter blinding were unnecessary and not performed.
Transgenic mouse assays were performed in Mus musculus FVB strain mice. The E11.5 developmental stage was considered. Animals of both sexes were used in the analysis. See the previous paragraph for details on sample size selection and randomization strategies.
|
10.1371/journal.pntd.0000862 | Targeting Antibiotics to Households for Trachoma Control | Mass drug administration (MDA) is part of the current trachoma control strategy, but it can be costly and results in many uninfected individuals receiving treatment. Here we explore whether alternative, targeted approaches are effective antibiotic-sparing strategies.
We analysed data on the prevalence of ocular infection with Chlamydia trachomatis and of active trachoma disease among 4,436 individuals from two communities in The Gambia (West Africa) and two communities in Tanzania (East Africa). An age- and household-structured mathematical model of transmission was fitted to these data using maximum likelihood. The presence of active inflammatory disease as a marker of infection in a household was, in general, significantly more sensitive (between 79% [95%CI: 60%–92%] and 86% [71%–95%] across the four communities) than as a marker of infection in an individual (24% [16%–33%]–66% [56%–76%]). Model simulations, under the best fit models for each community, showed that targeting treatment to households has the potential to be as effective as and significantly more cost-effective than mass treatment when antibiotics are not donated. The cost (2007US$) per incident infection averted ranged from 1.5 to 3.1 for MDA, from 1.0 to 1.7 for household-targeted treatment assuming equivalent coverage, and from 0.4 to 1.7 if household visits increased treatment coverage to 100% in selected households. Assuming antibiotics were donated, MDA was predicted to be more cost-effective unless opportunity costs incurred by individuals collecting antibiotics were included or household visits improved treatment uptake. Limiting MDA to children was not as effective in reducing infection as the other aforementioned distribution strategies.
Our model suggests that targeting antibiotics to households with active trachoma has the potential to be a cost-effective trachoma control measure, but further work is required to assess if costs can be reduced and to what extent the approach can increase the treatment coverage of infected individuals compared to MDA in different settings.
| Repeated ocular infection with the bacterium Chlamydia trachomatis leads to the development of trachoma, a major cause of infectious blindness worldwide. Mass distribution of antibiotics, a component of the current trachoma control strategy, has had success in reducing infection in some areas, but results in a large number of uninfected people receiving antibiotics. We have previously shown that transmission of the bacteria between people in the same household is very efficient. Here, we investigated the effectiveness and cost-effectiveness of targeting antibiotics to households with active trachoma (inflammatory disease) compared to mass distribution, using data from four trachoma-endemic populations and a mathematical model of transmission. We found a high correspondence between households with active trachoma and infected households. In all populations the household targeted approach was predicted to be as effective as mass distribution, but it reduced the number of uninfected individuals receiving antibiotics, making the targeted strategy more cost-effective when antibiotics are not donated. Assuming antibiotics are donated, we predicted the targeted strategy to be more cost effective if it increases the proportion of infected individuals receiving treatment. Further work to address the feasibility and the cost variability in implementing the targeted approach in different settings is now required.
| Trachoma, a ‘Neglected Tropical Disease’, is the leading infectious cause of blindness worldwide and there are currently an estimated 46 million people with the active stage of the disease [1]. The disease mostly affects impoverished populations where people cannot afford treatment and access to running water is scarce.
The World Health Organization (WHO) advocates the ‘SAFE’ strategy (Surgery for trichiasis, distribution of Antibiotics, Facial cleanliness and Environmental improvement) to work towards the Global Elimination of Trachoma as a public health problem by 2020. Annual mass drug administration (MDA) of antibiotics, to reduce the prevalence of the aetiological bacterium, Chlamydia trachomatis, is recommended for at least three years to members of communities in which the prevalence of Trachomatous Inflammation – Follicular (TF) in 1–9 year-olds is 10% or greater [2]. WHO recommends azithromycin as the first-line (oral) antibiotic for all, except infants under the age of 6 months who are given topical tetracycline [2]. MDA is advocated because screening individuals is not cost-effective and there is a poor correlation between active disease and infection of an individual [3], [4], [5], [6]. Field-ready and cost-effective diagnostic tests for infection with C. trachomatis are currently unavailable.
The ‘SAFE’ strategy has had success in reducing the prevalence of active trachoma in certain populations [7], [8], [9]. However, there are many costs associated with implementing MDA, particularly if the antibiotics are not donated, and many uninfected individuals receive treatment [10], [11], [12]. It has been estimated that there are fifty-seven countries endemic for trachoma [13]. Control programmes in eighteen of these countries currently receive azithromycin donated by the manufacturer, Pfizer, through the International Trachoma Initiative [14]. However a large disparity remains in certain countries between the number of individuals in the target population requiring treatment and the number of individuals receiving antibiotics [14].
If antibiotics can be successfully targeted to groups of infected individuals within a population rather than being administered to the whole population, the number of antibiotic doses required per population may be reduced. This saving of antibiotic resources could be utilised by other populations who require treatment to reduce trachoma. However the targeting method would only be justified if the method is as effective in reducing transmission as MDA and is also cost-effective.
Households with active trachoma are potential targets for antibiotic distribution. Trachoma clusters by household [15], [16], [17], [18] and we have previously shown that in most communities intra-household transmission is very efficient [19]. We found on average 71% of incident infections to be the result of household transmission (with the remainder due to transmission between households). An alternative approach to targeting treatment would be to limit treatment to children because they are the principal reservoir and source of infection in most communities. Children in some communities have been shown to have a relatively high prevalence of active disease [20], [21], [22] and a high burden of infection [23].
Here we investigate whether targeting antibiotics to households that have at least one member with active disease or to children alone is effective in the prevention of ocular chlamydial infection by analysing data on the prevalence of C. trachomatis and active disease from four endemic populations in West and East Africa (two in The Gambia, and two in Tanzania) with different baseline trachoma prevalence. We calculate the cost-effectiveness of targeted household treatment compared with MDA on the basis of a mathematical model and previously published data on the costs of these interventions [10], [11].
Conjunctival swabs were collected from a total of 4,436 individuals living in four endemic populations, which had not received prior interventions for trachoma control, in West and East Africa (Upper Saloum District and Jali village in The Gambia; Kahe Mpya sub-village and Maindi village in Tanzania) and the presence of infection was assessed using Polymerase Chain Reaction (PCR) amplification of a target sequence in the common cryptic plasmid of the bacterium C. trachomatis. Standard procedures current at the time of these surveys were followed to prevent contamination, described in [23] and [24]. In Maindi village, quantitative PCR amplification of the omp1 gene was used to indicate presence of infection. In all four studies clinical observations were made by experienced trained observers using a ×2.5 binocular loupe and pen torch or direct sunlight. In The Gambia the more detailed clinical diagnosis “FPC” system [25] was used but subsequently converted to the simplified WHO grading system [26] for this analysis. In Tanzania the simplified grading system was used. Active disease was defined as the presence of TF and / or Trachomatous Inflammation – Intense (TI). Detailed demographic information was collected including individual age, gender, and household membership. Full descriptions of the study populations and laboratory methods have been published elsewhere [17], [24], [27], [28] and details on community structure are summarized in [19]. Pre-control prevalences of infection in these populations (all ages) were 7.2%, 22.1%, 9.5% and 36.0% respectively. The age distribution of the prevalence of infection in these four communities is given in Table S1. The proportion of people present and consented to being screened for trachoma in the four data sets was 0.84 Upper Saloum district, 0.98 Kahe Mpya sub-village, 0.99 Jali village, and 0.86 for Maindi village. The work presented in this paper is based on further analyses of the data obtained in the original studies which had been granted ethical clearance [17], [24], [27], [28] and did not involve collecting further information. For this reason additional ethical approval was not sought.
The sensitivity and specificity of active disease (TF and TI) as a marker of infection were calculated among individuals in each community. We also calculated the sensitivity and specificity of active disease exhibited by at least one member of a household as a marker for infection of at least one household member (which we refer to as the household sensitivity and specificity).
Ocular chlamydial infection probably elicits only a limited protective immune response against re-infection and can be described by a simple Markov model where each individual may be either susceptible or infected. We have previously analyzed a susceptible→infected→susceptible (SIS) model where the population is structured into households [19], [29], [30]. Here we have extended this model to allow for different transmission parameters among ‘children’ (those aged less than ten years) and ‘adults’ (those individuals aged 10 years and older) (Text S1 ‘Model of Ocular Chlamydia Transmission’). We chose this classification of age because children under the age of ten are considered to be the principal reservoir of infection. Transmission parameters of each model for each dataset were estimated from the survey data using maximum likelihood, assuming endemic equilibrium. The most parsimonious yet adequate model for each dataset was selected using the Akaike Information Criterion (AIC) [31]. The transmission model was written in R (version 2.7.2). The rate of recovery from infection was taken as the reciprocal of the average duration of infection estimated from a Gambian cohort with frequent follow-up [4] (18.6 weeks for children, 7.1 weeks for adults and 17.2 weeks on average for the non-age-structured model).
The effectiveness of different treatment strategies was assessed using the most parsimonious model identified for each of the communities (Text S1 ‘Model Selection’, Table S2). With the exception of Upper Saloum district, the transmission models included a greater contribution of children to transmission than adults. Active disease at the household level was incorporated into the model at each round of treatment. At the time of treatment, each household was assigned a disease status by sampling from a Bernoulli distribution where the probability of a household having at least one individual with active disease was taken to be a function of the number of infected individuals within a household at the time of treatment. This probability function was calculated for each dataset on the basis of the observed distribution of infection and active disease in households of different sizes (Figure S1).
The outcome of three annual rounds of azithromycin treatment was investigated in all four populations as this is the number of treatment rounds recommended by the WHO prior to re-assessment of the prevalence of active disease when the baseline prevalence of TF in children is greater than 10%. For a transmission model parameterised to Maindi village, annual rounds were predicted to result in infection returning to almost baseline level in all strategies within one year after a treatment round suggesting that the treatment rounds need to be more frequent for this higher transmission setting. Therefore the effect of six bi-annual rounds was investigated for this setting.
Stochastic simulations of the model were used to examine four possible treatment scenarios:
A single treatment with azithromycin was assumed to be 95% efficacious in clearing infection [32]. We did not explicitly model treatment of infants aged <6 months with topical tetracycline instead of oral antibiotics. We assumed treatment coverage to be 80% in a), b) and d). One hundred simulations were run for each strategy to compare the effectiveness of each strategy. Further details of the stochastic model (written in R) are given in Text S1 ‘Stochastic Simulation Model’.
The cost-effectiveness of different antibiotic distribution strategies (compared with the ‘doing nothing’ option) from a government and societal perspective was assessed using previously published cost data from Mali and Nepal [10], [11] (summarised in Table S3) and the results from the stochastic simulations. The cost data were collected in 1998 and 2000 respectively for the studies in Mali and Nepal. Using the most recently available consumer price index for the two countries (2007) [33], [34], [35], the costs were converted to the value of US$ in 2007. Costs included the generic price of azithromycin per tablet, drug delivery costs per population size, and opportunity costs (the amount of money not earned per recipient whilst they attend the treatment campaign). Delivery costs in the Mali study consisted of governmental (salaries and vehicle investment) and distribution (dispatching, training of nurses and other health workers, per diems and fuel) costs specific to each strategy. Delivery costs in the study from Nepal were composed of salary and transportation costs and not the training of health workers. The delivery costs were higher for household-targeted treatment as they accounted for the extra training and salaries of nurses to diagnose trachoma in Mali and the increase in transport costs in Nepal (in this study two trips per community were assumed for this strategy: one for screening and one for treatment). We assumed that MDA was distributed via a central site. In agreement with the study in Mali, we assumed that opportunity costs equal to half a day or one hour's wages were incurred by individuals aged ≥10 years receiving treatment during MDA or HTT respectively. We assumed that individuals aged 10 years or older received an average of 3.43 azithromycin tablets and those under the age of 10 received an average of 1.02 tablets. (Text S1 ‘Cost Effectiveness Analysis’). One hundred stochastic simulations were performed for each strategy in each community and the costs were applied to the resulting simulations. The total cost of azithromycin was calculated by multiplying the number of individuals receiving treatment by the price per tablet and the mean number of tablets received in that age group. The delivery costs were scaled linearly to the size of the population in the four endemic areas under study and were assumed to occur at each round of treatment.
A discount rate of 3% per year was applied to all costs. Two estimates of total drug costs, delivery costs and opportunity costs were obtained using the two different sets of cost data and the mean cost of the two was calculated. Cost-effectiveness was calculated on the basis of the median effectiveness observed in the simulations, with lower and upper bounds based on the inter-quartile range of the simulations and the upper and lower costs from the two cost studies.
In all four communities (Upper Saloum district and Jali village in The Gambia, and Kahe Mpya sub-village and Maindi village in Tanzania) the sensitivity of active disease as a marker of infection was higher and specificity lower at the household level compared with the individual level (Table 1). Limiting clinical diagnosis in a household to children under the age of 10 years resulted in a similar household sensitivity and specificity compared to undertaking clinical diagnosis in all age groups (Table S4).
Targeting treatment to households, in which at least one resident has active disease, was predicted to result in post-treatment dynamics similar to MDA (Figure 1). The household-targeted approach had a slightly higher rate of return of infection and therefore the probability of eliminating infection five years after the last treatment round was predicted to be somewhat lower than the probability of eliminating infection after MDA (absolute difference between the probabilities in each setting was −0.22, −0.04, −0.25 and −0.12 for Upper Saloum district, Jali village, Kahe Mpya sub-village and Maindi village respectively). However if all individuals in targeted households were treated, then the probability of eliminating infection was predicted to greatly increase in each setting, being greater than MDA (absolute difference between the probability of eliminating infection after HTT with 100% coverage within the targeted households and the probability of eliminating infection after MDA was 0.26, 0.69, 0.07 and 0.44 respectively) (Figure 1). Limiting MDA to children under the age of 10 years resulted in an initial decrease in the prevalence of infection in the untreated older population (Figure S2) but the probability of eliminating infection in the whole community was greatly reduced compared to the other treatment scenarios investigated (Figure 1). There was a relatively smaller difference in effectiveness between the different treatment scenarios in the communities with relatively low baseline prevalence (Upper Saloum district and Kahe-Mpya sub-village) but HTT with 100% coverage, remained the most effective treatment scenario.
Modifying the model to account for variation in the efficiency of transmission among households resulted in faster return of infection for all treatment strategies in the simulations and the probability of eliminating infection was lower five years after the last treatment round (Figure S3). However, the relative impact of the different strategies remained robust to this additional complexity.
A household-targeted approach resulted in a similar number of infected individuals receiving treatment compared with MDA, but reduced the number of treatments given to uninfected individuals (Figure 2A). Assuming 80% therapeutic coverage and that azithromycin was not donated, HTT was predicted to be more cost-effective than MDA in all four communities when including the cost of generic azithromycin (Table 2). Assuming azithromycin was donated, HTT was predicted to be more cost effective when opportunity costs for individuals collecting drugs in the MDA approach were included (Table 2). Otherwise, MDA was estimated to be more cost effective. We did not calculate the cost-effectiveness of targeting treatment to children because the model simulations showed it to be the least effective of the four treatment scenarios at controlling infection.
If a visit to a household facilitates treatment of all members, then there was a large increase in the number of incident infections averted compared with either MDA or targeted approaches with 80% coverage in hyperendemic settings (Figure 2B). As a result, the household targeting strategy in which all members of diseased households are treated was predicted to be significantly more cost-effective in the areas with high baseline prevalence, even when azithromycin was assumed to be donated (Table 2).
A targeted approach for distributing azithromycin would result in fewer antibiotic doses distributed per head of population than MDA, thus saving medication for use by other trachoma endemic populations in need of treatment to reach ‘the Global Elimination of Trachoma as a public health problem by the year 2020’. However the approach would only be warranted if it is as effective in reducing ocular C. trachomatis prevalence in a population as MDA and as cost-effective.
Our results have indicated that targeting antibiotics to households with at least one member with active disease has a similar effect to MDA in the reduction of infection. Active disease was found not to be 100% sensitive as a marker of infection at the household level and this explains the small differences observed between the two strategies. However, we have shown that HTT results in a large reduction in the number of uninfected individuals receiving antibiotics compared to MDA (26%–51% reduction). When antibiotics were assumed to be donated, opportunity costs incurred by individuals taking time to collect tablets from the MDA program resulted in HTT being more cost-effective. Although the large majority of trachoma control programmes currently operate using donated azithromycin, we also estimated the cost-effectiveness of HTT assuming antibiotics were purchased at the generic price, to give a monetary value to the amount of antibiotic used in each strategy for the donor's perspective of the strategy and because some small scale programmes operating at village levels do purchase the drug [36], [37]. In this case the dominating cost was that of the antibiotics and so HTT was estimated to be more cost-effective.
If all members of visited households were assumed to be treated as a result of the visit by the treatment team, a much higher chance of eliminating infection from the community in all settings compared with MDA was predicted. The success of this approach will depend on the extent of household transmission and the degree to which household visits can boost treatment coverage. For example, in a community such as Kahe Mpya where household transmission was estimated to be limited [19], this approach can be hypothesised to be less effective. Baseline surveys of the prevalence of disease could be used as an indicator for the likely degree of household transmission, enabling the selection of communities that would benefit from a targeted approach. A large effort is typically required to achieve high coverage levels for MDA control programs [38]. In contrast, analogy can be drawn with other disease control programmes, such as vaccination for polio and measles, in which a house-to-house strategy of administering vaccination achieves much higher coverage than a fixed point campaign [39], [40]. Whether all household members can be reached with a single household visit remains to be investigated and further work is required to address whether coverage of infected individuals can be improved with HTT at what additional costs.
The cost per incident infection averted was greatly reduced when 100% of targeted-household members were assumed to be treated in areas with a relatively high prevalence of infection at baseline (Jali and Maindi villages), both when assuming azithromycin was and was not donated. In low prevalence settings the additional benefits of treating all household members were less apparent in our simulations because we investigated the effect of only three annual rounds of treatment, which, in these settings, were sufficient for any treatment scenario to have a greater than 50% chance at eliminating infection.
There are some caveats to our cost analysis: the cost data used in the study are a decade old and the linear scaling of delivery costs to the size of each community may not be appropriate for some costs (for example the time taken to perform a round of HTT may depend not only on the size but also on the geography of the population). However, the two cost studies referred to were the only published cost data at the time of our study that included the full cost of HTT. We assumed that individuals aged ≥10yr received a mean of 3.4 tablets whilst those aged<years received 1 tablet. This is a simplification and does not include azithromycin suspension given to younger children and topical tetracycline given to infants under 6 months. However, this would increase the total cost of antibiotics further, making targeted treatment more cost-effective.
We assumed that MDA occurred via a central site distribution. The WHO states that MDA can be carried out either via central site or by house to house distribution [2]. If we had assumed the latter for MDA there would have been a smaller difference in the distribution costs between MDA and HTT, (the only difference would be the cost of screening for active disease) and so HTT would have appeared more cost-effective in comparison to MDA. We took the assumption from the Mali cost study that MDA via a central site would result in adult antibiotic recipients having an opportunity cost of half a day's wages and HTT one hour's wages. However the WHO advises that MDA should be performed outside of the farming season [2] to try to minimise opportunity costs and improve the treatment coverage. In our analysis opportunity costs had a small impact on the cost-effective estimates but further studies could be performed to analyse what proportion of the recipient population's activities are interrupted by the different treatment campaigns.
The costs involved in treatment scenarios are likely to vary from country to country and by size of the community treated. Our work has investigated HTT in populations of approximately 1,000 people. If such an approach were to be implemented on a district or even country-wide scale, economies of scale will have to be considered e.g. a large number of nurses (or volunteers) will have to be trained for screening and there may be societal costs incurred as such personnel may stop working on other health programmes. Further studies are required to investigate these differences.
The delivery costs of targeting treatment to diseased households could be reduced in a number of ways which need to be researched further. We currently assume separate visits to households to assess disease and provide antibiotics. Assessment and treatment could be administered in a single visit, thereby reducing transport and salary costs. Furthermore, village volunteers could be trained to assess clinical disease to reduce the costs of ophthalmic nurses (a scheme which has been trialled with success in Ghana [41]). We also assumed that all residents would be screened for active disease at each round of HTT. Firstly, this could be limited to children under the age of ten: we have shown here that this approach has the same sensitivity as screening all ages but the difference in cost between the two approaches remains to be ascertained. Secondly, in practice, as soon as one person in a household is found to have active trachoma, the remainder of the household would not need to be screened. Therefore the cost of HTT in this work may be an overestimate in the higher prevalence settings where it is likely that in some households not all residents would be required to be screened. Further data are required to elucidate how the cost of screening for identifying target households will vary for different levels of prevalence and household clustering, including settings where WHO currently recommends HTT (active trachoma prevalence of 5%–9% in 1–9 year olds).
Data on active trachoma, analysed in this study, were collected in a scientific setting by experienced observers. The accuracy of trachoma grading may be more variable in a programmatic setting. A consequence of this would be that that sensitivity and specificity of active disease as a marker of infection at the individual level could worsen. However, this may be less significant at the household level, where diagnosis of just a single case of active disease is sufficient for treatment of that household. Further field studies would help understand the implications of trachoma grading error on HTT.
The original analyses of the cost data from Nepal and Mali differed from our work. The study in Nepal [11], [42] compared MDA of children to HTT of all ages. The study found the two strategies not to be significantly different from one another in the reduction of active disease and the costs involved (although this could be explained in part by the low power of the study). The original study in Mali [10] found HTT to be significantly less effective than MDA of the whole population with respect to the reduction of active disease prevalence one year after one round of treatment (although the age-adjusted odds ratio for prevalence active disease after HTT in relation to MDA was 1.56 with 95% confidence intervals of 1.00–2.43 indicating the strategies could have had the same outcome). The study found HTT to be more cost-effective except in low transmission settings. A difference between our work and the previous cost analyses is that here the cost was calculated as a cost per incident infection averted over five years rather than a change in point prevalence between baseline and one time point in the previous cost analysis. Measuring the number of infections avoided is not feasible in the field but measuring the cost-effectiveness in this way from model simulations gives a better insight into the impact of each treatment scenario on cumulative exposure to infection and therefore the ocular disease process.
Limiting treatment to children is another way to target treatment. Our models predicted that the prevalence in adults declines when children under the age of ten are treated, in agreement with House et al. [43], but this strategy is not as effective as MDA or HTT because the probability of eliminating infection is reduced in all four communities. Women could be included along with children in the target group as explored in the study in Mali [10]. However we did not investigate this strategy because the number of transmission parameters to be estimated would have been too large for the size of our dataset and the prevalence of infection did not differ largely between males and females in the study communities (excluding Maindi) [23]. Besides, there is considerable risk that specifically excluding adult males from treatment schedules would jeopardise community support for drug distribution.
Another method to target treatment would be to ‘graduate’ communities from MDA once the prevalence of ocular C. trachomatis infection is below a certain threshold, as suggested by Ray and colleagues [44]. Their study predicted graduating communities to be efficacious and drug-sparing (assuming a diagnostic test for infection becomes available in a field-ready format), by fitting a stochastic model allowing for heterogeneous transmission between communities, to the Upper Saloum district and Kahe Mpya sub-village data and a group of communities in Ethiopia.
Therefore two separate analyses of the Tanzanian and Gambian data sets have resulted in two different suggestions for targeting treatment. Here we fitted and simulated under a model of transmission which allows individuals to be infected by an infected member of their household or community at two different rates, specific to the setting. The Upper Saloum district contains 14 villages and this analysis grouped the villages together as one population. The Tanzanian sub-village contains balozis (groups of roughly 10 households that form an administrative unit) that we also grouped together. Additional analysis would be required to understand the relationship between within household transmission and heterogeneous community transmission where several of the communities constitute a larger population. This would then allow comparisons of the different targeting strategies to be made.
A recent study in Ethiopia [45] found that communities which had received MDA with azithromycin was associated with an odds ratio of 0.51 (0.29–0.90) for childhood (1–9 years) mortality one year after commencement of MDA compared to children in communities which did not receive the antibiotic. If this phenomenon extends to other settings then the impact of HTT with azithromycin on child mortality should be examined.
Caveats to our model of transmission have been described previously [19]. Infection status of individuals was characterised through PCR of ocular swabs. Standard precautions at the time of data collection were performed to prevent contamination of infection data (although the risk of contamination cannot fully be ruled out due to the absence of negative field controls). Sensitivity analysis of the assumption that each household is at equal risk of becoming infected found that increasing the level of heterogeneity in the household transmission parameters resulted in a faster rate of return of infection after treatment with a lower probability of eliminating infection for each treatment strategy. Further studies are needed to quantify differences in households' risk of becoming infected. Individuals were assumed not to move from one age group to the next but this is a reasonable simplification as the time spent in the lower age group (ten years) by each individual is far longer than the average duration of infection. We have assumed that the relationship between active disease and infection remains constant in a household after treatment. This requires further investigation but preliminary analyses of follow-up data from Upper Saloum District and Kahe Mpya sub-village indicates that households with at least one person with active disease at baseline can predict which households will contain individuals with ocular chlamydial infection at follow-up time points more accurately than households with active disease at follow-up.
The model did not include interventions to improve facial cleanliness (F) or the environment (E), the interventions advocated by WHO to accompany the distribution of antibiotics [2]. The exclusion of these interventions allowed the predicted effectiveness and cost-effectiveness of the different distribution strategies to be shown clearly. Inclusion of ‘F’ and ‘E’ would reduce the rate of return of infection and increase the probability of eliminating infection by an uncertain factor but is unlikely to alter the rank order of the impact of the different distribution strategies. If the cost of implementing ‘F’ and ‘E’ is independent of the antibiotic distribution strategy then the relative differences between the cost-effectiveness of implementing trachoma control for different antibiotic distribution strategies would remain unchanged. The exclusion of ‘F’ and ‘E’ from the model may explain why infection was observed to return relatively slowly in Maindi village following two rounds of treatment whereas our model predicts infection to rapidly return for an area with such a high baseline prevalence of infection. Changes in hygiene could have arisen in the village through residents receiving radio broadcasts by the National Trachoma Control Programme informing individuals to improve face washing and latrine usage [46] or alternatively, by simply the presence of the intervention itself, altering individuals' behaviour.
Our model suggests that targeting treatment to households that have at least one resident with active trachoma is as effective as MDA in a diverse variety of settings and can be more effective if the strategy increases the coverage of infected individuals. We also show that HTT is drug-sparing and has the potential to be more cost-effective but to have a better understanding of this in settings for which azithromycin is donated, more studies are required to evaluate whether HTT can improve antibiotic coverage levels of infected individuals and whether the cost can be further reduced compared with costs recorded in the studies in Mali and Nepal. The results of these studies will provide a better understanding of efficient and effective antibiotic distribution approaches for trachoma control programmes in countries with limited resources.
|
10.1371/journal.ppat.1003183 | Poxvirus Targeting of E3 Ligase β-TrCP by Molecular Mimicry: A Mechanism to Inhibit NF-κB Activation and Promote Immune Evasion and Virulence | The transcription factor NF-κB is essential for immune responses against pathogens and its activation requires the phosphorylation, ubiquitination and proteasomal degradation of IκBα. Here we describe an inhibitor of NF-κB from vaccinia virus that has a closely related counterpart in variola virus, the cause of smallpox, and mechanistic similarity with the HIV protein Vpu. Protein A49 blocks NF-κB activation by molecular mimicry and contains a motif conserved in IκBα which, in IκBα, is phosphorylated by IKKβ causing ubiquitination and degradation. Like IκBα, A49 binds the E3 ligase β-TrCP, thereby preventing ubiquitination and degradation of IκBα. Consequently, A49 stabilised phosphorylated IκBα (p-IκBα) and its interaction with p65, so preventing p65 nuclear translocation. Serine-to-alanine mutagenesis within the IκBα-like motif of A49 abolished β-TrCP binding, stabilisation of p-IκBα and inhibition of NF-κB activation. Remarkably, despite encoding nine other inhibitors of NF-κB, a VACV lacking A49 showed reduced virulence in vivo.
| The host response to infection provides a powerful means of restricting the replication and spread of viruses. Consequently, viruses have evolved mechanisms to reduce activation of host response to infection and this paper provides an example of this. Nuclear factor kappa B (NF-κB) is an important transcription factor that activates the host response to infection and is normally retained in an inactive form in the cytoplasm bound to an inhibitor called IκBα. However, upon stimulation by infection, IκBα is degraded and NF-κB moves to the nucleus to activate expression of genes mediating the host response. Here we describe how protein A49 from vaccinia virus, the vaccine used to eradicate smallpox, mimics IκBα to hijack the cellular degradation machinery and so stabilise IκBα and retain NF-κB in the cytoplasm. The importance of A49 is demonstrated by the fact that a virus lacking A49 was less virulent than control viruses, despite the expression of several other NF-κB inhibitors by vaccinia virus. Interestingly, HIV protein Vpu functions in a similar way to A49 and, given that A49 is highly conserved in variola virus, this work reveals a common strategy for suppression of host innate immunity by the viruses that cause smallpox and AIDS.
| Mammals respond to infection by activation of innate and adaptive immunity. In the past two decades, the discovery of pattern recognition receptors (PRRs) such as Toll-like receptors (TLR), intracellular nucleic acid sensors and inflammasomes has established the link between sensing pathogens and responding to them [1]. Following sensing of pathogen associated molecular patterns (PAMPs), signalling cascades lead to the activation of transcription factors that induce the expression of interferons (IFN), cytokines, chemokines and other pro-inflammatory molecules. Nuclear factor kappa B (NF-κB) is a transcription factor that plays a central role in switching on the immune system and proteins induced by NF-κB are responsible for the amplification of the innate response and for the recruitment of cells of the immune system, so linking innate and adaptive immunity [2], [3].
The signalling cascade leading to transcriptional activation of pro-inflammatory genes by NF-κB is well studied [4]. It can be initiated by TLR ligands, interleukin (IL)-1 or tumour necrosis factor (TNF)α and leads to the phosphorylation of the inhibitor of κB (IκB) by the IκB kinase (IKK) complex, the ubiquitination and degradation of phosphorylated IκB (p-IκB), and the translocation of the NF-κB heterodimer p65/p50 into the nucleus [5]. These steps are central in the NF-κB canonical signalling pathway and represent targets for pathogen evasion [6]. In particular, the ubiquitination and degradation of IκBα requires the recognition of its phosphorylated form [7] by the E3 ligase β-transducing repeat containing protein (β-TrCP) [8]. β-TrCP belongs to the Skp1, Cullin1, F-box protein (SCF) family and induces IκBα ubiquitination [9], [10]. β-TrCP was identified originally as the ubiquitin ligase targeted by human immunodeficiency virus (HIV)-1 viral protein U (Vpu) to cause CD4 degradation [11] and exists in 2 forms with very similar properties and specificity [12].
Poxviruses are large DNA viruses that replicate in the cytoplasm [13]. The prototypal poxvirus, vaccinia virus (VACV), was the live vaccine used to eradicate smallpox [14]. Poxviruses express many immunomodulatory proteins, encoded in the terminal regions of the genome [15], that can synthesise immunosuppressive steroids [16], [17] or block the production or action of cytokines, chemokines, IFNs and complement, for reviews see [18]–[20]. For instance, several VACV proteins inhibit activation of NF-κB: protein A52 binds TNF receptor associated factor 6 (TRAF6) and IL-1 receptor associated kinase 2 (IRAK2) and inhibits NF-κB activation downstream of TLRs and the IL-1 receptor [21], [22]. Protein A46 binds to different TLR adaptor molecules [21], [23]. Protein B14 binds to IKKβ and thereby reduces phosphorylation of IκBα [24]–[27]. Protein N1 inhibits NF-κB activation downstream of TRAF6 [28], [29] and protein M2 reduced p65 nuclear translocation [30]. Proteins K7 and K1 inhibit NF-κB activation by either inhibiting TLR-induced signalling (K7) [31] or by blocking IκBα degradation (K1) [32]. Protein E3 inhibits NF-κB activity [33] and antagonizes the RNA polymerase III-dsDNA-sensing pathway [34]. Lastly, protein C4 inhibits NF-κB at or downstream of the IKK complex [35]. However, genetic evidence predicts additional VACV inhibitor(s) of NF-κB activation that stabilise phosphorylated IκBα (p-IκBα) [36].
Here we show that VACV protein A49 is an inhibitor of NF-κB activation that contributes to virus virulence. Like HIV Vpu, A49 exploits molecular mimicry of p-IκBα to bind to the E3 ubiquitin ligase β-TrCP. Consequently, p-IκBα is not ubiquitinated or degraded and so remains in complex with the NF-κB (p65/p50) complex in the cytoplasm. A highly conserved counterpart of VACV A49 is encoded by variola virus, suggesting that the pathogens that cause smallpox and AIDS have evolved a common strategy to suppress innate immunity.
A screen of genes near the VACV genome termini for proteins that inhibited the induction of the IFNβ promoter led to the discovery of proteins C6 [37] and A49, the subject of this paper.
The A49R gene is near the right genome terminus between the genes encoding thymidylate kinase [38] and DNA ligase [39] and is predicted to encode an 18.8-kDa protein, which is conserved in other VACV strains and orthopoxviruses including variola virus [40] (Figure S1A) but not in monkeypox, camelpox and ectromelia viruses where the coding region is disrupted (www.poxvirus.org). However, outside poxviruses no clear counterparts were identified by bioinformatic searches. The A49R gene is transcribed both early and late during infection [41]. Consistent with this, an A49 antibody (Methods) detected the A49 protein in VACV strain Western Reserve (WR)-infected cells within 2 h of infection and in the presence of cytosine arabinoside (AraC), an inhibitor of DNA and late protein synthesis (Figure S1B). At later time points A49 expression was reduced by AraC indicating late expression also (Figure S1B), and this is consistent with a TAAAT motif upstream of the A49 open reading frame (ORF) that is a feature of VACV late promoters [42].
To study the role of the A49 protein in virus replication, a VACV WR strain in which the A49R ORF was deleted (vΔA49), and a revertant control with the A49R ORF reinserted (vA49rev) were constructed. These virus genomes were analysed by PCR and restriction enzyme digestion, and no differences were seen except at the A49R locus of vΔA49 (data not shown). These viruses had indistinguishable growth curves (Figure S2A and S2B) and ability to form plaques (Figure S2C) showing A49 is not essential for replication in vitro.
The contribution of A49 to virulence was tested by infecting groups of BALB/c mice intranasally and measuring weight loss and signs of illness [43]. Animals infected with vΔA49 lost less weight and recovered more quickly than controls (Figure 1A) and showed fewer signs of illness on days 4 to 10 post infection (pi) (Figure 1B). Measurement of infectious virus in lungs showed that all viruses replicated to similar titres by day 2 pi, but on days 5 and 7 pi mice infected with vΔA49 had significantly lower titres, showing more rapid clearance of virus (Figure 1C). Collectively, these data indicate that A49 is non-essential for replication, but is a virulence factor. The degree of attenuation seen by deletion of the A49 gene is similar to that deriving from deletion of many other VACV immunomodulators in this model [16], [17], [22], [23], [35], [37], [43]–[50].
Next, the mechanism by which A49 inhibited activation of the IFNβ promoter was studied. The A49R ORF was amplified from VACV WR genomic DNA and cloned into a mammalian expression vector with a C-terminal Flag or an N-terminal HA tag and tested for inhibition of the IFNβ pathway. HEK293 cells were co-transfected with an IFNβ promoter-firefly luciferase reporter and an A49 expression vector or the empty vector (EV), TLR3 (to allow IFNβ induction by poly(I∶C)) and a renilla luciferase transfection control. Cells were stimulated 24 h later with poly(I∶C), poly(dA∶dT) or infected with Sendai virus (SeV) (Figure 2). A49 blocked activation of the IFNβ promoter by poly(I∶C) (Figure 2A). The same effect was seen in RAW 264.7 cells stimulated with LPS and CpG, agonists of TLR4 and TLR9, respectively (Figure 2B). A49 also diminished transcription of IFNβ mRNA in poly(dA-dT)-stimulated HEK293T cells, as shown by quantitative PCR (Figure 2C), and inhibited production of the NF-κB responsive chemokine CCL5 in SeV-infected HEK293T cells (Figure 2D).
To understand how A49 inhibited IFNβ promoter activity, additional reporter gene assays were performed. HEK293 cells were transfected with an NF-κB luciferase reporter and a plasmid expressing A49. Upon stimulation with either IL-1α or TNFα, A49 reduced NF-κB activation (Figure 3A) in a dose-dependent manner (Figure 3B, C). Moreover, A49 blocked NF-κB activation mediated by TLR signalling in HEK293T cells transfected with TLR4 fused to the CD4 dimerisation domain (CD4-TLR4) (Figure 3D). To determine where A49 was acting, NF-κB was activated by overexpression of proteins operating at different stages in the signalling cascade. A49 blocked NF-κB activation after overexpression of TRIF (Figure 3D), TRAF2, TRAF6, TGFβ-activated kinase 1 (TAK1)-binding protein 3 (TAB3), and IKKβ (Figure 3E). However, when p65 was overexpressed, A49 was not inhibitory (Figure 3F), showing that A49 suppresses NF-κB activation downstream of IKKβ and upstream of p65.
To test if A49 blocked other transcription factors, such as IRF3, HEK293ET cells were transfected with a ISG56.1 promoter reporter (for IRF3) [51], along with plasmids expressing VACV A49, B14, which blocks NF-κB activation [29], or C6, which blocks IRF3 activation [37]. Under the conditions tested, after stimulation with poly(I∶C), C6 blocked ISG56.1 activation, whereas A49 and B14 did not (Figure S3A). Similarly, A49 and B14 did not inhibit induction of the canonical ISRE reporter after poly(I∶C) stimulation, whereas C6 did (Figure S3B), and A49 did not inhibit the ISRE promoter after stimulation with IFNα (Figure S3C). Collectively, these results indicate that A49 inhibits NF-κB.
Activation of NF-κB requires phosphorylation of IκBα on serines 32 and 36 within a short motif (DSGX2–3S) that is present in several proteins such as IκBα [8], Emi1 [52], β-catenin [12], HIV Vpu [11] and p105 [53]. Once this motif is phosphorylated, it is recognized by the E3 ligase β-TrCP [4]. Inspection of the A49 sequence identified the sequence SGNLES (aa 7–12) near the N terminus that matched the motif in β-TrCP substrates (Figure 4A). This suggested that A49 might bind β-TrCP and hence prevent β-TrCP from targeting its usual substrates. To test whether A49 interacted with β-TrCP, myc-tagged β-TrCP or TAK1 were co-expressed in HeLa cells with tandem-affinity purification (TAP)-tagged (streptavidin and FLAG) VACV proteins A49 or C6. After pull-down with streptavidin beads, an interaction between β-TrCP and A49 was seen by immunoblotting with anti-myc mAb (Figure 4B). Note that A49 did not interact with TAK1, nor did C6 with β-TrCP, confirming the specificity of the A49-β-TrCP interaction.
To test if this interaction occurred during virus infection, co-immunoprecipitation was done with extracts of HeLa cells transfected with TAP-tagged β-TrCP or retinoic acid induced gene I (RIG-I), and subsequently infected with vΔA49 or vA49rev. Both TAP-β-TrCP and TAP-RIG-I were pulled down with streptavidin beads and immunoblotting of the eluates showed that A49 associated only with TAP-β-TrCP (Figure 4C). Importantly, the β-TrCP-A49 interaction was demonstrated by immunoprecipitation of both proteins at endogenous levels after viral infection (Figure 4D).
The E3 ligase β-TrCP belongs to the F-box protein family and contains an N-terminal F-box domain and a C-terminal WD40 domain [9]. To study the A49-β-TrCP interaction, the F-box or the WD40 domains were deleted separately from β-TrCP (Figure 5A) and these truncated alleles were expressed in HeLa cells with TAP-tagged A49. After A49 pull-down, β-TrCP was co-purified only if it contained the WD40 domain (Figure 5B). This showed that A49, like IκBα, interacted with β-TrCP only if the WD40 domain was present, and that the F-box was dispensable.
To investigate if the IκBα-like motif within A49 was necessary for association with β-TrCP, A49 serines 7 and 12 were mutated to alanine (S7/12A) or glutamic acid (S7/12E). A third mutant was also made in which the glutamic acid at position 11 was changed to alanine in addition to serines 7 and 12 (7/11/12A). Expression of these TAP-tagged A49 alleles together with myc-tagged β-TrCP showed that the A49-β-TrCP interaction was lost with S7/12A or 7/11/12A, but was increased by the phospho-mimetic mutant S7/12E (Figure 5C). The association of A49 with another component of the β-TrCP SCF machinery that is found in complex with β-TrCP was also analysed. Both WT and S7/12E A49 co-immunoprecipitated with Skp1, but this was lost for the S7/12A A49 mutant (Figure S4). This is consistent with the SGLNES sequence near the A49 N terminus mediating binding to β-TrCP and, via β-TrCP, to other components of the SCF machinery.
Next, we addressed whether the IκBα-like motif within A49, which was required for β-TrCP interaction, was necessary for A49-mediated inhibition of NF-κB activation. HEK293T cells were transfected with an NF-κB responsive reporter plasmid together with TAP-tagged A49 alleles, and these cells were stimulated with TNFα. WT A49 inhibited NF-κB-luciferase expression compared to empty vector as expected, but the S7/12A mutant showed a statistically significant loss of function compared with WT (Figure 5D). It was also notable that the phospho-mimetic allele (S7/12E) inhibited NF-κB activation slightly more efficiently than WT A49 and this was consistent with its ability to bind β-TrCP slightly more strongly than WT (Figure 5D). Immunoblotting of these cell lysates revealed similar expression levels for each A49 allele (Figure 5E). Therefore, inhibition of NF-κB by A49 requires its β-TrCP recognition motif.
To compare the potency of A49 with another NF-κB inhibitor acting at the same stage of the pathway, a plasmid encoding the HIV Vpu protein fused at the N terminus with a TAP tag was transfected into HEK293 cells in parallel with TAP-tagged A49. Upon stimulation with TNFα or IL-1β, both A49 and Vpu inhibited NF-κB activation in a dose-dependent manner to a similar extent (Figure S5A, B). Immunoblotting of the lysates with a FLAG antibody demonstrated dose-dependent expression of each protein but HIV Vpu was expressed at higher levels than A49 suggesting that at equivalent levels of protein A49 was the more effective inhibitor under the conditions tested.
A49, like IκBα, contains a double serine motif that is needed to bind the C-terminal WD40 domain of β-TrCP. Therefore, it was possible that the ubiquitination and degradation of IκBα might be decreased in the presence of A49. As such, the IκBα/NF-κB complex would be stabilised and p65 would be retained in the cytoplasm. To address these possibilities, the effect of A49 expression on the level of total IκBα and p-IκBα (Ser 32/36) was examined by immunoblotting following TNFα stimulation (Figure 6A). In the absence of A49, p-IκBα was observed from 5 to 15 min after addition of TNFα and disappeared after 30 min, correlating with the reduction of IκBα levels at 30 mins. However, in the presence of A49, levels of both total IκBα and p-IκBα were higher when compared with cells transfected with the empty vector (Figure 6A). Conversely, A49 did not affect the phosphorylation of p65 on serine 536 by upstream kinases (Figure 6A), an event not involved in β-TrCP recognition. Furthermore, immunoprecipitation of p65 from HEK293T cells and immunoblotting for p-IκBα showed that the p-IκBα/p65 complex was de-stabilised by TNFα stimulation, but in the presence of A49 the complex remained intact (Figure 6B). To quantitate these effects, the intensity of the bands corresponding to p-IκBα and IκBα was analysed by densitometry. In all cases, the presence of A49 sustained both total and p-IκBα forms (Figure S6A, B).
Next the ubiquitination of IκBα was analysed in the presence and absence of A49. HEK293T cells expressing FLAG-tagged A49, or FLAG-tagged GFP, were treated with the proteasome inhibitor MG132 for 4 h and, following TNFα stimulation, IκBα was immunoprecipitated. After MG132 treatment, ubiquitinated forms of IκBα were observed upon TNFα activation, but these were reduced in the presence of A49 (Figure 6C). Densitometry analysis of the intensity of the higher molecular mass forms corresponding to ubiquitinated IκBα confirmed a reduction in IκBα ubiquitination in the presence of A49 (Figure S6C). To characterise the effects of A49 further, the levels of p27, a known target of the F-box protein Skp2, were measured in the presence of A49 and MG132. p27 accumulated after inhibition of the proteasome, but was unaffected by A49, suggesting that A49 did not interfere with the processing of targets of other F-box proteins or affected the proteasome non-specifically (Figure S7).
The translocation of p65 into the nucleus upon TNFα stimulation in the presence or absence of A49 was analysed. In HeLa cells A49 was present in both the nucleus and cytoplasm before and after treatment with TNFα and prevented the TNFα-induced translocation of p65 into the nucleus (Figure 6D). Quantitation showed highly significant differences (Figure 6E).
Collectively, these results show that A49 binds β-TrCP and thereby diminishes ubiquitination of p-IκBα. This stabilises p-IκBα and its interaction with NF-κB, so retaining p65 in the cytoplasm and preventing NF-κB-dependent gene expression.
Next, A49 function was tested during VACV infection. HeLa cells were infected with vA49rev or vΔA49 for 4 h, treated with MG132 for 1 h, and then stimulated with TNFα for 10 or 30 min. Infection by a VACV expressing A49 prevented IκBα degradation and stabilised p-IκBα, whereas infection with vΔA49 did not (Figure 7A). Remarkably, vA49rev induced accumulation of p-IκBα even without TNF stimulation, indicating that A49 blocked NF-κB activation triggered by viral infection. In addition, failure to accumulate p-IκBα could be reversed by MG132 (both without TNF activation or 30 min post-activation), suggesting that no other VACV protein interfered with the proteasomal degradation of p-IκBα downstream of A49. To obtain a more quantitative read-out, a similar experiment in which cells were infected with vA49rev or vΔA49 and treated with TNFα for 30 mins, was performed in triplicate. The amounts of p-IκBα and total IκBα were determined by quantitative fluorescence imaging of immunoblots, and plotted as a ratio compared to the amount of viral protein D8 to account for the efficiency of infection. Infection with vA49rev sustained levels of both p-IκBα and total IκBα 30 mins post-treatment compared to infection with vΔA49, and these differences were statistically significant (Figure 7B). In the absence of TNFα, accumulation of p-IκBα and IκBα during vA49rev infection was also detected (as observed by conventional immunoblotting), but with the sample sizes tested this was not significant.
Lastly, the stability of the IκBα/NF-κB complex was assessed during viral infection. HeLa cells were infected with vΔA49 or vA49rev and total IκBα was immunoprecipitated after TNFα stimulation. vA49rev infection stabilised both total IκBα and p-IκBα and these remained associated with p65 (Figure 7C). In contrast, vΔA49 infection failed to inhibit IκBα degradation and consequently no p65 co-precipitated with IκBα. So although there are other NF-κB inhibitors expressed by vΔA49, the A49 protein seems dominant in stabilising the IκBα/NF-κB complex.
The IRFs and NF-κB transcription factors are central to a coordinated immune response and their activation culminates in the expression of IFNβ, which induces an antiviral state in cells, and the production of inflammatory cytokines and chemokines which recruit lymphoid cells to the site of infection. In response to this host defence, viruses have evolved many mechanisms to suppress the host immune response and complete their life cycle. Understanding how viruses evade immune responses can aid understanding of virus pathogenesis and of the immune system itself.
Here, the VACV WR A49 protein is shown to be an inhibitor of innate immune signalling by blocking NF-κB activation. A49 was identified as an inhibitor of IFNβ expression in a screen of VACV proteins encoded near the genome termini, and dissection of the mechanism by which A49 inhibited IFNβ activation showed that A49 blocked NF-κB activation by binding to the WD40 domain of the E3 ubiquitin ligase β-TrCP (Figure 4 and 5). Thus, even though IκBα was phosphorylated by upstream kinases, β-TrCP-mediated ubiquitination of p-IκBα was reduced, p-IκBα was stabilised in complex with NF-κB and so NF-κB was retained in the cytosol (Figure 8).
A49 mediates its anti-NF-κB activity via molecular mimicry. Near the N terminus of A49 there is a SGLNES sequence that is closely related to motifs in IκBα, IκBβ, HIV Vpu and other β-TrCP substrates (Figure S1, 4A) suggesting that A49, like these other proteins, might bind β-TrCP. This interaction was shown by reciprocal immunoprecipitation of transfected tagged molecules and of endogenous β-TrCP and A49 made during VACV infection (Figure 4). Mutagenesis of the conserved serines within SGLNES to alanines prevented interaction with β-TrCP, stabilisation of p-IκBα, and inhibition of NF-κB activation (Figures 5 and S4). Conversely, mutation of these residues to glutamic acid increased the binding of A49 to β-TrCP and enhanced the inhibitory activity of A49 (Figure 5). This observation is in contrast to mutagenesis of IκBα where substitution of phospho-serines with glutamic or aspartic acid reduced IκBα recognition [8]. This indicates that in the A49 motif, charge-based interactions are sufficient to support binding. Notably, between the conserved serines A49 is one residue longer than in IκBα, like some β-TrCP substrates e.g. p105 [53]. This residue is after the glycine and before the hydrophobic amino acid known to insert in the β-TrCP hydrophobic groove [54].
The A49 protein is conserved in the majority of VACV strains and orthopoxviruses (OPVs) including variola virus, the causative agent of smallpox (Figure S1). But in ectromelia, camelpox and monkeypox viruses the A49 coding region is disrupted by mutation, showing A49 was non-essential for replication of those OPVs. The isolation of vΔA49 which replicated normally in cell culture and produced a normal sized plaque confirmed A49 is dispensable for VACV also (Figure S2). However, vΔA49 was less virulent compared to controls in vivo and this attenuation was characterised by lower weight loss, more rapid recovery and lower virus titres in infected lungs on days 5 and 7 pi (Figure 1). This attenuation, and the substantial increase of p-IκBα stability mediated by A49 during infection (Figure 7), was despite the fact that VACV expresses several other inhibitors of NF-κB activation (see introduction). The activity of A49 is clearly, therefore, not redundant. This might be explained by either these different proteins acting at different stages in the activation pathway, or them having multiple functions. In the former case, if an NF-κB inhibitor acted only downstream of the IL-1R or the TNFR before these pathways converge, it might give a different phenotype in vivo to an inhibitor that acted downstream of where these pathways converge. In addition, since there is crosstalk between the NF-κB pathway and other pathways, such as the MAP kinases for instance, an inhibitor might affect these other pathways too depending on its site of action. Concerning the possibility that the NF-κB inhibitors might have multiple functions, this has already been demonstrated in some cases. For instance, the N1 protein not only inhibits NF-κB activation [28], [47] but also inhibits apoptosis [55] and these functions are assigned to different binding surfaces of the protein [47]. With A49, there is a parallel with the HIV Vpu protein, which contains the conserved motif for interaction with β-TrCP, and has more than one function. Not only does Vpu interact with β-TrCP and so diminish degradation of p-IκBα, and hence block NF-κB activation [56], but Vpu also engages other proteins such as CD4 and tetherin and so brings these to β-TrCP for ubiquitination and degradation [57]. It is possible that A49 will also bind other cellular proteins and target these for β-TrCP-mediated ubiquitination and degradation in a manner advantageous for VACV. The identification of such targets may require their stabilisation by use of proteasomal inhibitors or mutants of A49, such as the S7/12A, which no longer bind to β-TrCP.
A comparison of the VACV-encoded inhibitors of NF-κB B14 and A49 is particularly interesting. The B14 protein binds IKKβ and reduces phosphorylation of IκBα, and thereby activation of NF-κB [24], [29]. In contrast, A49 reduces ubiquitination of p-IκBα and thus stabilises it. So B14 is acting at the step in the NF-κB activation pathway immediately upstream of the A49 protein. The benefit of multiple viral inhibitors of NF-κB was demonstrated here by the detection of more p-IκBα after infection with vA49rev than vΔA49 despite the presence of B14 in both viruses (Figure 7A, B). Therefore, these viral inhibitors work in combination to abrogate NF-κB activation during viral infection and each contributes to virus virulence. But, unexpectedly, the attenuation induced by loss of B14 or A49 is apparent in different in vivo models. In an intradermal model of infection [58], [59] a virus lacking B14 showed attenuation, but this virus had normal virulence in the intranasal model [25]. Conversely, vΔA49 was attenuated in the intranasal model (Figure 1) but not in the intradermal model (data not shown). This intriguing difference might be explained by either A49 or B14, or both proteins, having additional functions.
Although many inhibitors of NF-κB had been reported in VACV previously, genetic evidence had suggested the existence of additional inhibitor(s) because the VACV strain v811 stabilised p-IκBα despite lacking all known inhibitors of TNFα-mediated NF-κB activation [36]. The A49R gene is present in mutant v811 [60] and so the A49 protein probably represents such an inhibitor. A49 represents one of several virus proteins that target β-TrCP. In addition to A49 and HIV Vpu, rotavirus [61] and Epstein-Barr virus [62] also modulate β-TrCP activity. The widespread targeting of β-TrCP illustrates the importance of the SCFβ-TrCP complex for pathogen-induced responses.
In conclusion, the VACV A49 protein inhibits β-TrCP function by molecular mimicry and thereby blocks NF-κB activation, promotes immune evasion and enhances virus virulence. Given that a highly conserved version of A49 is encoded by all (∼50) strains of variola virus sequenced [63], it is probable that this strategy for increasing virulence by immune evasion is conserved in the pathogens that cause smallpox and AIDS.
This work was carried out in accordance with regulations of The Animals (Scientific Procedures) Act 1986. All procedures were approved by the UK Home Office and carried out under the Home Office project licence PPL 70/7116.
For mammalian expression, A49R was cloned into a pCI vector (Promega) with a Flag tag, into a pcDNA4/TO vector (Invitrogen) as an N-terminal TAP fusion containing 2 copies of the streptavidin binding sequence and 1 copy of the FLAG epitope [64], and into pCMV-HA (Clontech) with an N-terminal HA tag. Mutagenesis of A49 was performed by PCR amplification using forward primers containing the desired mutations. nTAP.C6 and FLAG-B14 have been described [29], [37]. To produce protein in bacteria (used for antibody generation), A49R was cloned into the pOPINE vector [65]. Myc-β-TrCP was obtained from Addgene (identified as β-TrCP2 after sequencing). The ORF was PCR amplified and cloned in pcDNA4/TO as TAP or HA fusions. PCR products covering residues 1–250 (F-box) or 251–569 (WD40) were cloned fused with HA. The fidelity of the PCR products was verified by DNA sequencing.
A polyclonal antibody against A49 was generated in rabbits by Eurogentec immunised with purified recombinant A49 protein. Monoclonal antibody against IκBα was a kind gift of Ron T. Hay (University of Dundee) and was used in ubiquitination assays. Other antibodies were: p65 (Santa Cruz), phospho-p65 (Ser 536) (Cell Signalling), IκBα and p-IκBα (Ser32/36) (Cell Signalling), β-TrCP (clone C-18, Santa Cruz), Skp1 (Santa Cruz), p27 (Cell Signalling), β-actin (Abcam), α-tubulin (Upstate Biotech), myc (Cell Signalling), HA (Covance) and FLAG (M2 clone, Sigma). The mouse monoclonal antibody AB1.1 against D8 was described [66] as well as the anti-N1 serum [45]. Poly(I∶C), poly(dA-dT) and MG132 were from Sigma, TNFα, IL-1β and IL-1α were from Peprotech, LPS and CpG were from Invitrogen. Sendai virus (strain Cantell) was grown in embryonated hen eggs [67] and was used at a single dose, at a dilution of 1∶200.
BSC-1, CV-1, HEK293T, HEK293ET and RAW 264.7 cells were cultured in Dulbecco's modified Eagle's medium (DMEM, Gibco) supplemented with 10% heat-treated foetal bovine serum (FBS, Harlan Sera-Lab), 50 IU/ml penicillin and 50 µg/ml streptomycin (Gibco) and 2 mM L-glutamine (Gibco). HeLa cells were maintained in Minimum Essential Medium (MEM - Gibco) supplemented with 1× non-essential amino acid solution (Sigma) and 10% heat-treated (56°C, 1 h) foetal bovine serum (FBS, Harlan Sera-Lab), 50 IU/ml penicillin and 50 µg/ml streptomycin (Gibco) and 2 mM L-glutamine (Gibco).
Alignment of the A49 amino acid sequence from poxviruses was performed using Clustal X and Genedoc [68]. Viruses and GenBank accession numbers are: VACV-Cop (vaccinia virus strain Copenhagen, acc. num. M35027), VACV-WR (vaccinia virus strain Western Reserve, AY243312), VACV-TT (vaccinia virus strain Tian-Tian, AF095689), CPXV (cowpox virus, NC_003663), HSPV (horsepox virus, DQ792504), VARV (variola virus strain India 1967, NC_001611).
A49R gene fragments were produced by PCR using VACV WR genomic DNA as template and were cloned into a pCI (Promega) derived plasmid. This plasmid contained E. coli guanylphosphoribosyl transferase (EcoGPT) fused in frame with the enhanced green fluorescent protein (EGFP), driven by a VACV promoter and enables transient dominant selection of recombinant viruses [69]. To produce an A49 deletion VACV, a DNA fragment containing the left and right flanking regions of A49R was produced by overlapping PCR. The 5′ fragment was generated with oligonucleotides 5′- CAGGGATCCAACAAAAGGTATTACAAGAAT – 3′ (LA), containing a BamHI restriction site (underlined), and 5′- ATATCGTTCGCGGATATAGTTTCTATCTTGGCAATAAC 3′ containing nucleotides from the 3′ fragment (italics) at the 5′ end. The 3′ fragment was generated with oligonucleotides 5′- CAAGATAGAAACTATATCCGCGAACGATATTTGTG -3′, with complementary sequence to the 5′ fragment (italics) and 5′- TGCAGCGGCCGCCGGATTTCTGTGTTCTCTTTGAAG -3′ (RA), containing a NotI restriction site. These two fragments were joined by PCR using the LA and RA oligonucleotides and cloned forming pΔA49. To make vΔA49, BSC-1 cells were infected with VACV WR and transfected with pΔA49. Recombinant viruses were collected 24 h later selected in the presence of mycophenolic acid, xanthine and hypoxanthine [69]. Intermediate EcoGPT+ viruses were resolved into WT or vΔA49 by plaquing on BSC-1 cells in the absence of drugs and their genotype confirmed by PCR. To generate pA49rev, a DNA fragment containing the entire A49R gene and flanking regions was generated with oligonucleotides LA and RA. vA49rev was generated in a similar manner by transfection of pA49rev in vΔA49-infected cells.
Growth kinetics of viruses was determined as described [37].
Virus plaque size was determined as described [70].
Virus virulence was analysed in a murine intranasal infection model [71]. Groups of 5 BALB/c mice 6–8 weeks old were inoculated with 5×103 PFU of the different recombinant virus in 20 µL PBS. Mice were weighed daily and signs of illness were recorded as described [43]. All experiments were conducted at least twice.
HEK293T cells in 96-well plates were transfected with 60 ng/well of firefly luciferase reporter plasmids, 10 ng/well of pTK-Renilla luciferase (pRL-TK, Promega) or 20 ng/well of pGL3-renilla luciferase [37] as transfection control, and the indicated amount of expression vectors with FugeneHD (Roche) or GeneJuice (Merck). A plasmid encoding HIV Vpu was a gift from Paul Lehner, and was amplified by PCR and cloned as N-terminal TAP fusion. IFNβ-promoter luciferase reporter was a gift from T. Taniguchi (University of Tokyo, Japan) and NF-κB-luciferase was from R. Hofmeister (University of Regensburg, Germany). ISRE-luciferase and pRL-TK (Renilla Luciferase) were purchased from Promega. ISG56.1 was a gift from Ganes Sen (Cleveland Clinic, USA). TLR3 was a gift from D.T. Golenbock (University of Massachusetts Medical School, USA). The concentration of the A49 expression vectors varied according to the vector used and the experiment. They are: for Figures 2A, B and Figures 3A, E and F, 60 ng/well; Figure 2D, 50 and 150 ng/well; Figures 3B, 3C and S5, 50, 100 or 150 ng/well; Figure 3D, 150 or 50 ng/well. DNA was kept constant during the transfections by the addition of empty vector control plasmid. Cells were stimulated as indicated in the figures and were harvested in passive lysis buffer (Promega). The relative stimulation of reporter-gene expression was calculated by normalizing firefly luciferase activity with renilla luciferase activity. In all cases, data shown are representative from at least three independent experiments. Data from experiments performed in triplicate are expressed as mean ± SD.
RNA from HEK293T cells in 6-well plates was extracted using the RNeasy kit (QIAGEN) and converted to cDNA using the Quantitect RT kit (QIAGEN). IFNβ mRNA was quantified by real-time PCR with the TaqMan gene expression assay Hs00277188_s1 and a β-actin endogenous control VIC-MGB probe (6-carboxyrhodamine–minor groove binder; Applied Biosystems). Cells were transfected with 2 µg/well of the A49 expression plasmid or the empty vector. Experiments were performed in triplicate.
HEK293T cells were stimulated as indicated in figure legends. Supernatants were analysed for CCL5 protein using Duoset reagents (R&D Biosystems).
For co-immunoprecipitation, HEK293T or HeLa cells were transfected using Fugene-6 (Roche). After 24 h, cells were washed once with ice-cold PBS and lysed with IP buffer (10% glycerol, 150 mM NaCl, 20 mM Tris-HCl [pH 7.4], 0.1% Triton-X100, and protease inhibitors [Roche]). Lysates were incubated with either streptavidin beads (Thermo Scientific) or protein G sepharose beads (GE Life Sciences) that were pre-incubated with the corresponding antibody for 2 h at 4°C. After 3 washes with ice-cold Tris-buffered saline (25 mM Tris-HCl, pH 7.4, 150 mM NaCl, 2 mM KCl), proteins were eluted and analysed by SDS-PAGE and immunoblotting.
Densitometry analysis was performed using ImageJ, an open-source image processing and analysis software provided by the National Institutes of Health (http://rps.info.nih.gov/ij). Films were transformed into digital pictures and intensities were calculated after removal of background signal. For quantitative immunoblotting, IRDye 800-conjugated donkey anti-mouse and goat anti-rabbit antibodies were used according to the manufacturer's instructions (LI-COR Biosciences). Membranes were then dried and scanned using the Odyssey infrared imaging system (LI-COR Biosciences). Quantitation was performed using the system software to determine total band intensities on the original scans.
HeLa cells on glass coverslips were transfected (Fugene 6) with 50 ng of plasmid expressing A49. After 24 h, cells were either treated with 50 ng/ml TNFα (Peprotech) (diluted in warm 2% FBS MEM) or were mock treated. After 30 min, the cells were stained with anti-p65 (1∶50) and anti-FLAG (1∶500) and prepared for imaging as described [35].
Data were analysed using unpaired Student's T test unless stated otherwise. Statistical significance is expressed as follows: * P-value<0.05, ** p-value<0.01, *** p-value<0.001.
|
10.1371/journal.ppat.1005600 | Distinct Roles of Type I and Type III Interferons in Intestinal Immunity to Homologous and Heterologous Rotavirus Infections | Type I (IFN-α/β) and type III (IFN-λ) interferons (IFNs) exert shared antiviral activities through distinct receptors. However, their relative importance for antiviral protection of different organ systems against specific viruses remains to be fully explored. We used mouse strains deficient in type-specific IFN signaling, STAT1 and Rag2 to dissect distinct and overlapping contributions of type I and type III IFNs to protection against homologous murine (EW-RV strain) and heterologous (non-murine) simian (RRV strain) rotavirus infections in suckling mice. Experiments demonstrated that murine EW-RV is insensitive to the action of both types of IFNs, and that timely viral clearance depends upon adaptive immune responses. In contrast, both type I and type III IFNs can control replication of the heterologous simian RRV in the gastrointestinal (GI) tract, and they cooperate to limit extra-intestinal simian RRV replication. Surprisingly, intestinal epithelial cells were sensitive to both IFN types in neonatal mice, although their responsiveness to type I, but not type III IFNs, diminished in adult mice, revealing an unexpected age-dependent change in specific contribution of type I versus type III IFNs to antiviral defenses in the GI tract. Transcriptional analysis revealed that intestinal antiviral responses to RV are triggered through either type of IFN receptor, and are greatly diminished when receptors for both IFN types are lacking. These results also demonstrate a murine host-specific resistance to IFN-mediated antiviral effects by murine EW-RV, but the retention of host efficacy through the cooperative action by type I and type III IFNs in restricting heterologous simian RRV growth and systemic replication in suckling mice. Collectively, our findings revealed a well-orchestrated spatial and temporal tuning of innate antiviral responses in the intestinal tract where two types of IFNs through distinct patterns of their expression and distinct but overlapping sets of target cells coordinately regulate antiviral defenses against heterologous or homologous rotaviruses with substantially different effectiveness.
| Two distinct families of interferons (IFNs), type I (IFN-α/β) and type III (IFN-λ) IFNs, are quickly produced in response to virus infection and engage distinct receptors to invoke shared rapid and broad-spectrum antiviral mechanisms against invading pathogens. However, the relative importance of type I and type III IFNs in protecting different organ systems against specific viruses or distinct strains of an individual virus remains to be fully explored. Here we demonstrated in suckling mice that neither type I nor type III IFNs are effective in blocking intestinal replication of murine rotavirus, rather, viral clearance is dependent upon adaptive immune responses. In contrast, both IFN types cooperate to control intestinal replication and extra-intestinal spread of simian rotavirus in neonatal mice. Unexpectedly, we found that although intestinal epithelial cells (IECs) respond to both types of IFNs in neonatal mice, responsiveness of IECs to type I IFNs, but not type III IFNs, is diminished in adult mice. Transcriptional analysis showed that both types of IFN receptors induced overlapping intestinal antiviral responses, which were abrogated only when both receptor types were deleted. Overall, these findings reveal a well-coordinated spatial and temporal regulation of antiviral defenses by type I and type III IFNs in the gastrointestinal tract that varies significantly depending on the viral strain examined.
| Mucosal surfaces of mammalian reproductive, respiratory and gastrointestinal (GI) tracts are functionally unique. Most pathogens enter the host through mucosal surfaces, and epithelial cells lining these tracts serve as a first line of defense against invading pathogens. Moreover, mucosal surfaces are constantly exposed to a variety of microbes and therefore have the unique task of distinguishing between harmful pathogens and commensal symbiotic microbes. This challenge is particularly important in the GI tract where tolerance to billions of commensal microbes must be established and maintained. At the same time, the GI tract provides protection against pathogenic bacteria and GI viruses such as rotaviruses (RVs). RV infection causes severe diarrhea in infants and young children and is a major cause of morbidity and mortality in the developing world. The overall incidence of RV infection and morbidity appears to be similar in all unvaccinated areas, however the majority of RV-related deaths occur in developing countries [1, 2]. Although RV replicates primarily in the mature intestinal epithelial cells (IECs) of the small bowel, it can breach intestinal barriers and spread to the circulation and extra-intestinal organs (e.g. mesenteric lymph node (MLN), central nervous system (CNS), liver and biliary tree) [1, 3–5].
Initial antiviral protection in mammalian hosts is mainly dependent on the coordinated action of type I and type III IFNs, which are quickly produced by virus-infected and bystander IECs, as well as by intestinal hematopoietic cells [6–8]. These IFNs invoke innate antiviral mechanisms within virus-infected and uninfected bystander tissues, and coordinately regulate the development of adaptive immune responses against viral pathogens including RV [9–11]. Both IFN types activate the same signal transduction pathway which culminates in the formation of a ternary transcription complex, composed of STAT1, STAT2 and IRF9, and designated as IFN-Stimulated Gene Factor 3 (ISGF3) [12–14]. Subsequently, type I and type III IFNs induce expression of the same sets of IFN-stimulated genes (ISGs) and have very similar biological activities in sensitive cells [13, 15–17]. However, type I and type III IFNs engage distinct receptor complexes for their signaling. Whereas all type I IFNs utilize a heterodimeric receptor complex composed of IFN-αR1 and IFN-αR2 subunits, type III IFNs, or IFN-λs, engage the IFN-λR1 and IL-10R2 receptor chains for signaling [8, 12, 14, 18]. A major difference between the type I and type III IFN-based antiviral systems resides in the distinct cell-type specific pattern of receptor expression. In contrast to the type I IFN receptor that is ubiquitously expressed, the IFN-λ receptor is expressed primarily by epithelial cells [19].
Recent studies identified type III IFNs as critical non-redundant antiviral mediators in the GI tract. Type I IFNs alone were unable to restrict reovirus replication in the IECs of mice deficient in the type III IFN receptor [20]. Efficient control of murine norovirus, which replicates in the IECs, dendritic cells and B cells of the mouse, also required a functional IFN-λ receptor [21]. Furthermore, type III IFNs have recently been identified as unique antiviral mediators that were indispensable for the protection of suckling mice against infection with murine RV strain EDIM [22, 23]. However, the latter result contradicted other studies demonstrating that the murine EW-RV strain (derived from the original EDIM strain) replicated to a similar extent in wild-type (WT) or STAT1-deficient suckling mice due to its ability to effectively antagonize IFN production and signaling [6, 24–27]. In contrast, heterologous simian RV (RRV strain) was found to replicate poorly in WT mice, but RRV infection of STAT1-deficient suckling mice resulted in substantially enhanced intestinal replication and efficient systemic virus replication and disease [25, 26]. It was also demonstrated that despite their contrasting IFN-sensitive replication phenotypes, infection of suckling mice with either EW-RV or RRV results in similar induction of several IFN-stimulated genes (ISGs) in the small intestine at 16 hours post infection (hpi), confirming prior observations and indicating that RRV replication is uniquely sensitive to one or more of these antiviral effectors [6]. In fact, the substantial restriction of non-homologous RV strain replication in heterologous host species likely underlies the attenuating principle of several live RV vaccines that were based upon restricted replication of bovine, lamb or simian RV in humans [28, 29]. To further investigate the relative importance of type I and type III IFNs in regulating antiviral defenses in the GI tract, we utilized two distinct strains of RV, the homologous murine EW-RV strain and the heterologous simian RRV strain, and mice deficient in STAT1, type I or type III IFN receptors, or both types of IFN receptors.
Experiments reveal that neither type I nor type III IFNs alone, or both IFN types together were able to efficiently suppress the intestinal replication or diarrheal disease of murine EW-RV, demonstrating that homologous RV have evolved highly effective measures to circumvent the innate responses of their murine host. In contrast, we now demonstrate that both type I and type III IFNs are important mediators of antiviral protection of the GI tract and work cooperatively to limit intestinal replication of the heterologous simian RRV in suckling mice. Transcriptional analysis in the suckling mouse of bulk intestinal tissues revealed that similar patterns of ISG induction occurred in RRV-infected WT mice and in mice lacking either type I or type III IFN receptors, and induction of most ISGs was completely abolished in mice deficient in receptors for both IFN types. Further specific analysis demonstrated that IECs of neonatal mice were responsive to both types of IFNs as determined by immunohistochemical (IHC) staining for IFN-induced tyrosine phosphorylation and nuclear translocation of STAT1. In addition, pretreatment of neonatal mice with either type of IFN resulted in suppressed intestinal RRV replication. We also observed that responsiveness of IECs of adult mice to type I IFNs was diminished, whereas lamina propria cells (LPCs) of both neonatal and adult mice were responsive to type I but not type III IFNs. Both type I and type III IFNs helped to limit extra-intestinal RRV spread, but only type I IFNs were essential for controlling RRV replication in MLN. Our results reveal a previously underappreciated contribution of type I IFNs to the protection of IECs against GI viruses in neonatal mice, and demonstrate that both type I and type III IFNs act as important mediators of antiviral defenses within the GI tract, acting cooperatively to suppress heterologous RV replication in IECs and restrict extra-intestinal spread.
To develop mice deficient in type III IFN signaling, exon 3 of the IFNLR1 gene was targeted for elimination (Fig 1 and S1 Fig). LoxP sites flanking exon 3 were introduced into corresponding introns away from the splice signals to ensure that normal splicing of the modified IFNLR1 gene is not disturbed (Fig 1A). The entire gene encoding IFN-λR1 consists of 7 exons. The deletion of exon 3 by Cre recombinase resulted in the generation of an abnormal IFN-λR1 transcript with exon 2 spliced to exon 4 leading to a reading frame shift and the premature termination of translation of the modified IFNLR1 transcript (Fig 1B and 1C). Mice with the deleted exon 3 in the IFNLR1 gene had intact type I IFN signaling, but were unresponsive to type III IFNs as demonstrated by the inability of IFN-λ to trigger STAT1 phosphorylation in various tissues (Fig 1D). In addition, freshly isolated kidney cells from these mice up-regulated MHC class I antigen expression only in response to type I IFN, whereas cells from WT mice responded to both types of IFNs (S1C Fig).
Previous studies have shown that RV strains differ in their ability to antagonize IFN responses, both in vitro and in vivo, in part dependent on the species origin of the virus and the host [6, 30–34]. RV strains are best able to circumvent innate immune responses in their natural, homologous species host. For example, although both heterologous simian RRV and homologous murine EW-RV induce similar levels of type I IFNs and several ISGs in the small intestine at 16 hpi, replication of RRV was highly sensitive to IFN-mediated antiviral defenses and occurred much more efficiently in the intestine of type I IFN receptor and STAT1 KO suckling mice than in WT mice [6, 25], whereas murine EW-RV strain was able to replicate comparably in the intestine of suckling mice in the presence or absence of IFNs; EW-RV shedding and clearance proceeded at similar rates in wild type and STAT1 KO mice, which are deficient in type I, II and III IFN signaling [6, 24, 26]. In contrast to these results with STAT1 and type I IFN receptor KO mice, two recent studies with IFN receptor-deficient animals indicated that type III IFNs, and not type I IFNs, could mediate very significant and biologically relevant innate antiviral protection of neonatal mice IECs during murine RV infection [22, 23]. In this studies, it was observed that the action of type III IFNs alone was sufficient to substantially restrict murine RV (EDIM-RV strain) intestinal replication and virus shedding while promoting suckling mouse weight gain and diminishing diarrheal disease. On the other hand, as previously reported by others [6, 24, 26], type I IFNs did not appear to play a substantial role in intestinal epithelial antiviral defenses, viral replication, or in protecting the suckling mice from murine RV associated disease [22, 23].
In order to better clarify the conflicting data in the reports that found no substantial changes on murine RV replication or disease in suckling Stat1-/- mice and the reports that found that murine RV replication and disease was substantially augmented in Ifnlr1-/- mice, eight-day-old WT, Ifnlr1-/-, Ifnar1-/ and Ifnar1-/-Ifnlr1-/- mice (on the C57BL/6J background) and WT, Stat1-/- and Rag2-/- mice (on 129S6/SvEv background) were orally inoculated with 104 diarrhea dose 50 (DD50) of the murine EW-RV strain derived from the original EDIM-RV isolate [33], and fecal EW-RV shedding was initially quantified by ELISA. We observed virtually identical fecal shedding of EW-RV in WT, Ifnlr1-/-, Ifnar1-/-, Ifnar1-/-Ifnlr1-/- or Stat1-/- mice during the first 7 days post infection (dpi) (Fig 2A and 2B). Slightly delayed viral clearance in suckling mice deficient in type I, type III, or both IFN receptors, together with small differences in virus shedding, was observed on 8 and 9 dpi (Fig 2A and 2B). In agreement with our previous studies [24–26], we saw little difference in the kinetics of EW-RV clearance between 129S6/SvEv WT and Stat1-/- suckling mice, with a slight increase in shedding only detected in Stat1-/- mice on 9 dpi (Fig 2A and 2B). Consistent with this observation, similar levels of EW-RV protein were detected in the small intestine of infected WT and Stat1-/- mice on 1 dpi (Fig 2C). Of note, unresolved shedding was observed only in Rag2-/- animals (Fig 2A and 2B), demonstrating that adaptive, rather than innate, immune responses are primarily responsible for resolving EW-RV infection.
We also detected similar patterns of viral replication in small intestines of EW-RV-infected C57BL/6J WT, Ifnlr1-/- and Ifnar1-/-Ifnlr1-/- mice, as measured by qRT-PCR (Fig 2D). Although IFN-λ transcripts were strongly up-regulated to similar levels in intestines of all mouse strains examined, IFN-β transcripts were induced less efficiently with considerably weaker IFN-β induction in Ifnar1-/-Ifnlr1-/- mice than in WT or Ifnlr1-/- mice (Fig 2E), possibly due to the absence of a positive feedback loop in these animals. ISG induction occurred with similar efficiency in WT and Ifnlr1-/- mice, but was completely abrogated when both IFN-λ and IFN-α receptors were lacking (Fig 2F). Thus, intestinal ISG expression following homologous RV infection can occur in the absence of IFN-λ receptor-mediated signaling. Nevertheless, the presence of IFN and ISG expression has minimal effect on EW-RV replication in small intestine of IFN receptor-sufficient or deficient mice (Fig 2A–2D and 2F), demonstrating that EW-RV efficiently antagonizes most IFN-mediated antiviral responses [27]. These data obtained in both Ifnlr1-/- and Ifnar1-/-Ifnlr1-/- mice and confirmed in Stat1-/- mice differ from the results of recently published studies, where suckling Ifnlr1-/- and Ifnar1-/-Ifnlr1-/- mice were found to be substantially more susceptible to the murine EDIM-RV strain, and type III IFNs were postulated to be the primary mediators of ISG expression in the intestinal epithelium [22, 23]. In these conflicting studies, mice reconstituted with a functional Mx1 gene were used. In our studies, all the mouse strains examined were on either C57BL/6J or 129S6/SvEv backgrounds and lacked the functional Mx1 gene. However, conventional C57BL/6J mice that are deficient in Mx1, and Mx1-reconstituted C57BL/6J mice showed no differences in either EW-RV replication or in their patterns of IFN and ISG induction (S2A–S2D Fig), ruling out the possibility that Mx1 was responsible for the observed differences between these studies. We also obtained the murine EDIM-RV strain that was used in the conflicting studies [22, 23] and compared it to our murine EW-RV strain that was also derived from the original EDIM strain. Both strains replicated similarly in WT or Stat1-/- 129S6/SvEv or C57BL/6J WT suckling mice (S3A and S3B Fig), indicating that differences in the replication phenotype of murine RV in the two studies were not likely due to viral strain variations. Overall, EW-RV replication in suckling mice was not significantly affected by the presence of either type I or type III IFNs, confirming the substantial insensitivity of murine RV to IFN-mediated antiviral effects on virus replication in the suckling murine host [6, 24, 26].
A previous study also revealed substantial growth retardation of EDIM-RV-infected Ifnlr1-/- pups compared to their WT counterparts, and correlated these differences with increased EDIM-RV replication in Ifnlr1-/- mice [23]. To determine whether lack of IFN signaling might affect pathophysiologic parameters other than RV replication, weight gain and diarrheal disease were also monitored in EW-RV-infected suckling WT, Stat1-/- and Rag2-/- mice (on 129S6/SvEv background). Diarrhea appeared on 2 dpi in all groups, affected virtually all inoculated pups, and resolved between 8 and 11 dpi (Fig 3A), with no difference in the numbers of animals affected, despite continuous virus shedding by Rag2-/- mice (Figs 2A and 3A). However, diarrhea was moderately prolonged in the EW-RV-infected Stat1-/- mice (Fig 3A), suggesting that IFN signaling may affect the duration of murine RV-associated diarrheal disease. Furthermore, similarly delayed resolution of diarrhea was observed in Stat1-/- mice infected with simian RRV (Fig 3A). Despite moderately prolonged diarrhea in Stat1-/- animals or the continued EW-RV shedding in Rag2-/- mice, body weight gain of WT, Stat1-/- and Rag2-/- mice in either EW-RV or RRV-infected groups remained similar (Fig 3B). These experiments suggest that RV-induced diarrhea and weight gain are not necessarily correlated with virus load, since the chronically infected Rag2-/- mice (Fig 2A) resolved diarrhea earlier than EW-RV or RRV-infected Stat1-/- pups and exhibited body weight gain comparable to WT mice (Fig 3A). Therefore, although IFNs may be involved in the timely resolution of murine RV-induced diarrhea, this is not directly correlated with either virus load or weight gain. Of interest, although the level of shedding and the severity of diarrheal disease have been directly correlated in children [35], this correlation is not invariable since the Rag2-/- mice resolved diarrhea while continuing to shed RV (Figs 2 and 3).
Because homologous murine RV is remarkably resistant to IFN-mediated innate responses in suckling mice (Fig 2) [6, 21–23] and because prior studies had indicated that heterologous RVs might be more responsive to IFN mediated suppression [5, 6, 25], we next examined the heterologous simian RRV to assess the relative contributions of type I and type III IFNs to the control of non-murine RV replication and clearance. Eight-day-old suckling WT mice or mice deficient in IFN type-specific signaling (Ifnlr1-/-, Ifnar1-/- and Ifnar1-/-Ifnlr1-/- mice, all on C57BL/6J background) were orally inoculated with 4 x 106 FFU RRV. Intestinal samples were collected and RV titers determined on 1, 3, 5 and 8 dpi. On 1 and 3 dpi, there were no significant differences in intestinal virus replication between Ifnlr1-/- and Ifnar1-/- mice, whereas either type I or type III IFN receptor-deficient (Ifnar1-/- or Ifnlr1-/-) animals supported significantly greater intestinal RRV replication (>100 fold) than did WT mice (Fig 4A and 4B). Importantly, RRV replicated to significantly higher titers in Ifnar1-/-Ifnlr1-/- and Stat1-/- mice than in mice lacking either IFN receptor alone (Fig 4A and 4B). Ifnar1-/-Ifnlr1-/- and Stat1-/- mice also showed delayed virus clearance, with virus still present in the small intestine on 8 dpi, a time point when virus could no longer be detected in WT, Ifnar1-/- and Ifnlr1-/- strains (Fig 4A and 4B). Low, but sustained, RRV levels were detected in the small intestine of Rag2-/- mice from 1 to 8 dpi (Fig 4A and 4B), which persisted through 15 dpi. Consistent with the virus titer results, infected IECs were rarely detected in the small intestines of RRV-infected Ifnar1-/- and Ifnlr1-/- mice by immunohistochemistry, with much more extensive antigen-staining present in the IECs of infected Ifnar1-/-Ifnlr1-/- animals (Fig 4C). Viral antigen was found primarily in IECs at the tips of the villi in type I or type III IFN receptor-deficient mice, and RRV-infected IECs were essentially absent in infected WT mice. These results indicate that both type I and type III IFNs independently restrict replication of the heterologous simian RRV in intestines of suckling mice, with resolution of infection mediated primarily by the adaptive immune response.
RV gene transcriptional analysis revealed robust RRV replication in Ifnar1-/-Ifnlr1-/- mice, with incremental decreases in replication occurring in the single IFN receptor KO and WT pups, respectively (Fig 4D). The induction of IFN-β transcripts by RRV in WT mice occurred primarily on 1 dpi. In comparison, in mice lacking receptors for either type I or type III IFNs, as well as in the double IFN receptor KO mice, IFN-β induction was more robust and occurred over a prolonged period of time following RRV infection, particularly in Ifnar1-/-Ifnlr1-/- mice and to a lesser extent in Ifnlr1-/- mice (Fig 4E). Similar expression patterns were observed for IFN-λ transcripts but with sustained up-regulation of IFN-λ expression on 2 and 3 dpi in all mouse strains (Fig 4E). Similar to IFN-β, expression levels of IFN-λ transcripts were elevated in all KO strains in comparison to WT mice, and sustained elevated expression of IFN-λ transcripts was mostly pronounced in either Ifnar1-/-Ifnlr1-/- or Ifnlr1-/- mice. Patterns of IFN expression (Fig 4E) mirrored the transcriptional RRV load (Fig 4D), suggesting that increased viral replication in the absence of the cognate IFN receptors and their effector pathways triggers prolonged and elevated expression of both IFN types in IFN receptor deficient mice. Because expression of IFN transcripts was similar in response to RRV infection of WT and single or double IFN receptor-deficient mice (Fig 4E), type I and type III IFNs appear to be induced independently during RV infection. To assess whether much more robust up-regulation of Ifnl transcription compared to Ifnb transcription correlates with higher levels of type III IFN protein expression, homogenates of small intestines from RRV-infected mice were collected on 1 dpi and used for IFN-λ ELISA and type I IFN bioassay (S4 Fig). Whereas IFN-λ proteins were detected at about 300 to 500 pg per 100 mg of tissue, levels of type I IFNs were below the detection level of the bioassay (<30 units/ml; ~300 pg per 100 mg). These results correlate with our transcriptional analyses and indicate that RRV infection predominantly triggers production of type III IFNs in the small intestine. Both the magnitude and kinetics of ISG expression were similar in WT and single IFN receptor-deficient mice (Fig 4F), demonstrating that either type I or type III IFN can up-regulate ISG expression in small intestine of RRV-infected mice independently. Moreover, the increased expression of these ISGs was abolished in double IFN receptor KO mice after RRV infection, despite the induction of type I and type III IFN transcripts (Fig 4E and 4F). The diminished ISG induction in Ifnar1-/-Ifnlr1-/- mice correlated with increased viral replication in these animals, reflecting the sensitivity of RRV to IFN-mediated innate antiviral defenses. The delayed RRV clearance in Ifnlr1-/- mice (Fig 4D) correlated with only a modest induction of Ifnb (Fig 4E) and the lack of transcriptional Ifna responses (S5A–S5C Fig). In contrast, levels of Ifnl2/3 transcription were substantially elevated in response to RRV infection (Fig 4E) and correlated with fast reduction of RRV by 2 dpi in Ifnar1-/- mice (Fig 4D), suggesting a predominant role of type III IFNs in the intestinal antiviral defense.
Because RRV replication was significantly increased in Ifnar1-/-Ifnlr1-/- and Stat1-/- animals when compared to single IFN receptor KO mice, it seemed likely that IECs in suckling mice can respond to either IFN-α or IFN-λ. Such a possibility is also supported by the abrogation of RRV-mediated intestinal ISG expression only in the absence of receptors for both IFN types. Signaling downstream of either the type I or type III IFN receptor leads to tyrosine phosphorylation of STAT1 (pSTAT1). To directly investigate responsiveness of IECs and cells within the lamina propria to IFNs, eight-day-old suckling C57BL/6J mice were subcutaneously injected with PBS, IFN-α, or IFN-λ, and levels and nuclear translocation of pSTAT1 in small intestine were assessed by immunohistochemical staining with pSTAT1 specific antibody (Fig 5A). Both IECs and LPCs of the small intestine of suckling mice were responsive to IFN-α, whereas only IECs were responsive to IFN-λ (Fig 5A). To exclude the possibility that lack of type III IFN signaling might alter responsiveness of IECs to type I IFNs, Ifnlr1-/- suckling mice were also treated with IFN-α and STAT1 phosphorylation was again examined in the small intestine. Type I IFN-induced pSTAT1 was detected in both IEC and LPC compartments while no pSTAT1 staining was found in response to IFN-λ treatment (Fig 5B). Therefore, in the suckling mouse, both type I and type III IFNs are capable of triggering STAT1 activation in IECs, whereas LPCs are only responsive to type I IFNs.
These data from suckling mice are different from previous observations wherein mouse IECs were found to be unresponsive to type I IFNs when adult mice were treated with plasmid-delivered IFNs [20, 22]. To investigate whether the IFN responsiveness of IECs might be age-dependent, IFN-mediated STAT1 activation was subsequently assessed in six to eight-week-old WT or Ifnlr1-/- mice. Only LPCs, but not IECs, were strongly responsive to type I IFNs in the older mice, whereas responsiveness of IECs to type III IFNs remained robust in adult animals (Fig 5C and 5D). Low levels of STAT1 phosphorylation were detected in PBS-treated IECs in adult WT mice, but not in mice deficient in type III IFN receptor (Fig 5C and 5D), suggesting that weak, constitutive IFN-λ signaling is likely to be maintained in IECs in adult mice. These results reveal an unexpected age-related change in the type I IFN responsiveness of IECs, which is robust in early post-natal life, and strongly diminished as the mouse matures.
To directly assess whether STAT1 activation in IECs correlates with antiviral protection, eight-day-old suckling mice were subcutaneously injected with either IFN-α or IFN-λ 6 h before RRV infection, and virus replication was analyzed on 1 dpi (Fig 5E). Pretreatment with either type of IFN resulted in reduced RRV levels on 1 dpi, demonstrating that both type I and type III IFNs inhibit intestinal RRV replication. In a previous study, unresponsiveness of IECs to systemic type I IFN treatment was explained by the polarized nature of IFN signaling in IECs [22]. In that study, polarized IECs were shown to respond to type I IFNs only when IFN-β was delivered apically, whereas type III IFNs were active on both basolateral and apical surfaces [22]. In contrast, our experiments with human SW-1116 colorectal carcinoma cells demonstrated that upon polarization, these cells strongly respond to either type I or type III IFNs only basolaterally (Fig 5F). Of note, sensitivity of SW-1116 cells to IFN-λ was enhanced upon polarization to higher degree than that to IFN-α, and weak responsiveness to IFN-λ at the apical surface was also detected (Fig 5F). Overall, these results demonstrate that type I and type III IFNs are capable of inducing antiviral protection in IECs of suckling mice in a redundant manner.
Previous mouse studies demonstrated that RRV can spread to and replicate in extra-intestinal sites as efficiently as murine RV including MLN, and that type I IFNs are important for restricting RRV replication and pathogenesis at these sites [4, 5, 25]. To further investigate the role of specific IFNs in controlling early extra-intestinal spread and replication, RRV titers in MLN of infected mice were assayed. There were no significant differences between virus titers in MLN of RRV-infected WT and IFN receptor-deficient animals on 1 dpi (Fig 6A). However, by 3 dpi, elevated virus titers on the order of 100-fold greater than WT were detected in MLN of RRV-infected Ifnar1-/- mice, and 1,000-fold above WT levels in Ifnar1-/-Ifnlr1-/- animals (Fig 6B). RRV replication was still detectable in MLN of Ifnar1-/- and Ifnar1-/-Ifnlr1-/- mice on 5 dpi, and MLN from several Ifnar1-/-Ifnlr1-/- mice were still RV positive on 8 dpi (Fig 6A and 6B). Levels of RRV were not elevated above WT controls in the Ifnlr1-/- mice. Therefore, type I IFNs mediate the primary control of extra-intestinal spread and replication of RRV in MLN, although a further deficiency in type III IFN signaling enhances virus replication in MLNs when combined with a type I IFN deficiency at early times post infection. Consistent with these data, elevated virus titers were also found in MLN of RRV-infected Stat1-/- mice on 3 dpi, and virus was still detectable in MLN of some Stat1-/- mice on 8 dpi (Fig 6A and 6B). RRV titers in the MLN of infected Rag2-/- were mainly unchanged from 1 dpi to the conclusion of the experiment on 8 dpi (Fig 6A and 6B), emphasizing the importance of adaptive immunity for virus clearance.
Murine and simian RV strains have also been shown to spread to the liver and replicate in the epithelial lining of the biliary tree [4, 5, 25]. To investigate involvement of type I and type III IFNs in limiting systemic simian RV infection in the liver, we determined hepatic virus titers in RRV-infected single and double IFN receptor-deficient mice. WT and single IFN receptor-deficient mice showed similar levels of virus replication in the liver on 1 dpi (Fig 6C and 6D). RRV titers had declined in WT controls by 3 dpi, but remained significantly elevated in infected Ifnar1-/-, Ifnlr1-/- and Ifnar1-/-Ifnlr1-/- animals (Fig 6C and 6D). Although virus was cleared from the liver of WT and single IFN receptor-deficient mice by 5 dpi, RRV persisted in the liver of double IFN receptor-deficient Ifnar1-/-Ifnlr1-/- suckling mice through 8 dpi (Fig 6C and 6D). These data indicate that type I and type III IFNs cooperate to limit RRV spread to and replication in the liver of infected mice. Furthermore, although liver virus titers were similar in 129S6/SvEv WT, Stat1-/- and Rag2-/- mice on 1, 3 and 5 dpi, both Stat1-/- and Rag2-/- mice had higher liver virus titers than WT controls on 8 dpi (Fig 6C and 6D). Of note, all mouse strains on the 129S6/SvEv background showed higher virus titers in the liver than any mouse strain on the C57BL/6J background (Fig 6C and 6D), indicating that levels of RRV replication in the liver are also affected by strain-specific genetic factors.
Recent studies concluded that intestinal antiviral responses are primarily mediated by type III, rather than type I, IFNs during infection with the homologous murine EDIM-RV [22]. In contrast, we observed that murine EW-RV replication was rather insensitive to the antiviral actions of both type I and type III IFNs in the homologous murine host (Fig 2). On the other hand, the replication of the heterologous simian RRV in suckling mice was substantially restricted by both IFN types (Figs 4–6). In addition, when either IFN was administered systemically, it was able to efficiently stimulate STAT1 activation in IECs of suckling mice (Fig 5A and 5B) and induce antiviral protection against RRV in IFN-pretreated suckling mice (Fig 5E). These findings provided a clear phenotype and biologically relevant RV strain to decipher the relative roles of these two types of IFNs in intestinal innate antiviral responses.
To perform transcriptional analysis, small intestines of RRV-infected WT mice, as well as pups lacking receptors for either type I or type III IFNs, or both receptors, were isolated 1, 2 and 3 dpi and used for microfluidic qRT-PCR analysis of selected antiviral response transcripts at the bulk whole intestinal level (Fig 7). In these experiments, we also included uninfected animals as well as murine EW-RV-infected WT, Ifnlr1-/- and Ifnar1-/-Ifnlr1-/- mice harvested at a single time point (2 dpi). Of note, we had previously shown that at 16 hpi, despite their substantially different replication capacity in vivo, both EW-RV and RRV infections result in comparable levels of ISG and type I IFN induction in bulk intestinal tissues of WT suckling mice [6]. Intestinal EW-RV replication was 1,000- to 10,000-fold greater than that of RRV in WT suckling mice (Fig 7A). Similar to earlier observations at 16 hpi [6], the overall transcriptional levels of EW-RV-induced antiviral cytokines such as IFN-λ and IFN-β, and several IFN-induced antiviral genes such as ISG15 and IFIT3, were similar to, or greater than those induced by RRV on 2 dpi (Figs 7B–7D). We found that infection with the IFN-sensitive RRV strain resulted in the robust induction of several well-defined ISGs, including those encoding IFIT1/2/3, ISG15, ISG20, RSAD2, and Mx2 (Fig 7C and 7D). Notably, transcription of such ISGs was also induced in the absence of either type I or type III IFN receptor in agreement with the ability of both IFN types to trigger STAT1 phosphorylation in IECs, but was almost completely abolished in Ifnar1-/-Ifnlr1-/- animals (Fig 7D). Thus, type I and type III IFNs drive a set of highly similar antiviral intestinal responses to both EW-RV and RRV, but effectively restrict the replication of only heterologous simian RRV. In the absence of type I and type III IFN signaling, the attenuated RRV-induced transcription of certain ISGs such as ISG20 and RSAD2 can be driven by interferon regulatory factors (IRFs) directly or mediated by other virus-induced mechanisms [36, 37]. The absence of types I and type III IFN receptors led to prolonged induction of CXCL10 and CCL5 chemokine genes, correlating with extended and increased viral replication. Of interest, genes encoding the anti-microbial proteins REG3B and REG3G (S6 Fig) were induced independently of IFNs by both EW-RV and RRV, with more consistent and higher levels of up-regulation in EW-RV-infected mice. Expression of these genes can be controlled by IL-22, which was recently implicated in host anti-RV restriction [23, 38]. Collectively, transcriptional analysis of bulk intestinal tissues revealed a surprising level of redundancy in the induction of intestinal antiviral responses in suckling mice by type I and type III IFNs.
Type I and type III IFNs are important mediators of innate antiviral defenses. Although these IFNs signal through distinct receptor complexes, the signaling cascades, sets of ISGs up-regulated and biological activities induced in response to these cytokines are almost indistinguishable [8, 9, 13]. For this reason, the relative contributions of type I and type III IFNs to overall antiviral protection of an entire organism can only be investigated with the use of animals deficient in individual and combined specific IFN receptors. Due to the cell-type specific pattern of type III IFN receptor expression that largely limits action of IFN-λs to epithelial cells [19], the target organs for type III IFNs are restricted, whereas type I IFN receptors are expressed ubiquitously, and therefore, expected to evoke antiviral defenses in all tissues and cell types. Nevertheless, there have been several reports demonstrating that mice deficient in STAT1, a transcriptional factor that is critical for signaling of all IFNs, are more susceptible to certain viruses than type I IFN receptor-deficient mice [39, 40]. For example, influenza virus replicated to much higher titers in STAT1 or STAT2 KO mice than in type I IFN receptor-deficient animals [39] suggesting that type III IFNs may also play a role in protecting mice against influenza virus infection. Indeed, it has been demonstrated that protection against influenza A virus replication in airway epithelium can be mediated by either type I or type III IFNs [41–43]. Similarly, it has been shown that RRV replicates better in Stat1-/- than in Ifnar1-/- suckling mice [6, 25, 26].
It was recently reported that the GI epithelium, and particularly IECs, are protected primarily by type III IFNs in suckling and adult mice, based on the observation that mice deficient in IFN-λ signaling had impaired control of murine RV infection when compared with strain-matched WT and Ifnar1-/- mice [22]. In our experiments, on the other hand, the level of murine RV shedding was high and almost indistinguishable in WT or IFN receptor-deficient or STAT1-deficient suckling mice with the exception of slightly delayed but significant clearance differences on 8–9 dpi in Ifnlr1-/- and Stat1-/- mice; with complete virus clearance from all mouse strains on 10 dpi (Fig 2). In addition, in the current study weight loss and the degree of diarrheal disease were not substantially enhanced in Ifnlr1-/- and/or Stat1-/- mice when compared to WT suckling mice. These data are inconsistent with the results of Pott et al. and Hernandez et al., which suggested that pathogenesis and susceptibility to murine RV was highly IFN-λ dependent [22, 23]. The basis of the different findings in the studies (Summarized in Table 1) is not readily apparent. Direct comparison of the two murine RV strains used in the two groups of studies indicates that they are highly related or identical in terms of replication capacity in WT and Stat1-/- suckling mice (S3 Fig). In addition, although mice, which were used in studies of Pott et al. and Hernandez et al. [22, 23], have a reconstituted functional Mx1 gene, whereas mice used in other studies possess a non-functional Mx1 gene, kinetics and magnitude of EW-RV replication were similar in WT suckling mice deficient or reconstituted with the functional Mx1 gene (S2A and S2B Fig) and profiles of transcriptional IFN and ISG induction were similar (S2C and S2D Fig). Of note, the Mx1-reconstituted mice were generated by breeding the A2G-Mx1+/+ mice onto the Mx1-/- C57BL/6 mice for several generation, however the purity of the genetic background of the resulting B6.A2G-Mx+/+ strain has not been characterized [44]. Moreover, B6.A2G-Mx+/- males, starting from F2 generation, were selected for backcrossing with C57BL/6 females based on their survival of the infection with lethal dose of influenza virus infection. This breeding strategy, in addition for maintaining the Mx+/- genotype in breeders, may also put a selective pressure skewing for genes, other than Mx1, which also enhance virus resistance. These B6.A2G-Mx+/+ mice were later crossed with the Ifnlr1-/- C57BL/6J mice [41, 45]. It should also be noticed that mice used in the current studies have only a small alteration within the Ifnlr1 gene, only exon 3 was deleted (Fig 1 and S1 Fig), whereas mice used in studies of Pott et al. and Hernandez et al. [22, 23], have the entire the Ifnlr1 gene (~20 kb) removed and replaced with the IRES-LacZ/MC1-Neo reporter gene/selection cassette (~5 kb) [45]. This substantial genomic alteration could potentially affect expression of other neighboring genes, particularly the Il22ra gene that is located downstream of the Ifnlr1 gene and encodes one of the IL-22 receptor chains. IL22RA (IL-22R1) shares epithelial cell specific expression pattern with IFN-λR1 and these two adjacent genes may share co-regulatory elements. Therefore, further studies are required to fully characterize and compare the two currently existing Ifnlr1-/- mice. In addition, animal diet, microbiota and persistent infections with murine norovirus or helicobacter, which are often present in pathogen-free animal facilities, have been shown to alter innate intestinal antiviral responses [21, 46–48], and therefore could potentially account for some of the observed differences in this and other studies.
Of note, we did observe declining responsiveness of IECs to type I IFNs in adult mice (Fig 5C and 5D), but saw robust IFN signaling in suckling mice IECs following treatment with either IFN-α or IFN-λ (Fig 5A and 5B). We also observed that low constitutive levels of STAT activation are present in IECs of WT adult mice, but not in Ifnlr1-/- mice (Fig 5D), suggesting that IFN-λ signaling seems to be maintained in IECs and may contribute to the well documented decreased ability of murine EDIM-RV to replicate as efficiently in adult as in suckling mice [49]. These findings are interesting but unlikely to fully account for the differences observed between the two sets of studies (Table 1) as both of these were carried out in suckling mice. Recent studies on intestinal antiviral immunity have shown compartmentalized effects of type I and type III IFNs, where LPCs were protected only by type I IFNs, whereas type III IFNs were indispensable for restricting reovirus or RV replication in IECs [20, 22]. Using a direct immunohistochemical assay of IFN-triggered STAT1 activation in IECs and LPCs, we also observed that LPCs responded only to type I IFNs by STAT1 activation (Fig 5). However, we observed that administration of either type I or type III IFNs could induce the restriction of RRV replication in the small intestine of suckling mice with one caveat: the responsiveness of IECs to type I IFNs was substantially more pronounced in neonatal mice, where RV disease is present, than in adults (Fig 5), where RV replication is restricted and RV associated disease is absent [50, 51].
RV infections are remarkably host specific. Homologous RVs replicate to significantly higher levels in the intestines of homologous hosts, require much lower doses to cause disease and spread more efficiently among non-immune susceptibles than heterologous RVs. As with several other virus infections, it has been shown that RV host-range restriction is, in large part, determined by the different efficiency of homologous versus heterologous RVs in antagonizing the host IFN response [6, 25, 26]. Indeed, we also observed that murine EW-RV replicated in mice much more efficiently than heterologous RRV (Fig 7A). However, the higher EW-RV load induced similar magnitude of IFN (Figs 2E, 4E and 7B) and ISG (Figs 2F, 4F, 7C and 7D) responses as the much lower load of RRV. Similar findings were reported when responses to EW-RV and RRV were compared at 16 hpi in suckling mouse intestines, and reinforce the notion that homologous murine RVs have evolved highly effective measures to circumvent host innate immune responses in order to replicate efficiently and cause diarrheal disease that promotes virus dissemination [6, 25, 26]. The robust murine RV replication was not appreciably augmented in IFN receptor or STAT1 deficient suckling mice (Fig 2 and [6, 24–27]). Therefore, in order to study the relevant importance of type I and type III IFNs in initiating and propagating intestinal innate immune responses, we used heterologous simian RRV that has been previously shown to be much more sensitive than homologous murine RV to innate antiviral defenses in suckling mice [6, 25, 26]. RRV replicates but only poorly in the WT suckling mouse intestine and is unable to spread from inoculated to susceptible litter mates while murine EW-RV is more virulent, replicates to much higher levels in the mouse intestine, and spreads very efficiently among litter mates.
Our results revealed that RRV replicated much more efficiently in either type I or type III IFN receptor-deficient suckling mice, and the complete lack of IFN responses in Ifnar1-/-Ifnlr1-/- or Stat1-/- mice allowed RRV replication to proceed to even higher titers with delayed clearance in comparison to single IFN receptor-deficient mice (Fig 4). Accordingly, pretreatment of suckling mice with either IFN-α or IFN-λ 6 h prior to RRV infection suppressed intestinal RRV replication to the similar extent and combined IFN-α or IFN-λ pretreatment provided a similar level of protection as pretreatment with either type of IFN alone (Fig 5E). On the other hand, pretreatment of suckling mice with type I or II interferon had no effect on homologous murine RV replication or diarrheal disease [52]. In addition, transcriptional analysis in the whole intestine of RV-infected suckling mice revealed that classical ISGs were induced to similar levels in either type I or type III IFN receptor-deficient animals (Fig 7), confirming independent and overlapping actions of type I and type III IFNs in the intestinal antiviral defense of suckling mice.
RRV was also able to replicate more efficiently in MLN of Ifnar1-/- or Ifnar1-/-Ifnlr1-/- mice than in WT or Ifnlr1-/- mice (Fig 6A and 6B), suggesting that at this site, type I and not type III IFNs were primarily responsible for controlling RV replication. Both single and double IFN receptor KO mice demonstrated impaired control of RRV replication in the liver (Fig 6C and 6D), but only Ifnar1-/-Ifnlr1-/- mice failed to clear RRV from the liver by 5 dpi (Fig 6C and 6D). Of note, RRV replication was better controlled by type III IFNs at earliest times post infection, because increased viral transcription was detected only on 1 dpi, and was quickly suppressed by 2 dpi in Ifnar1-/- mice, whereas viral transcripts were still elevated on 2 and 3 dpi in Ifnlr1-/- mice (Fig 4D), suggesting a somewhat more prominent role of type III IFNs in controlling intestinal RV replication and this correlated with more efficient Ifnl2/3 induction by RV than those of type I IFNs (Figs 4E and 7B, and S4 Fig). More prolonged intestinal RRV replication in Ifnlr1-/- mice might give virus more time to disseminate to and replicate in other organs. Nevertheless, we observed distinct patterns of RRV spread and replication in MLNs and liver, the former controlled primarily by type I IFNs (Fig 6A and 6B) and the latter by both IFN types (Fig 6C and 6D).
Collectively, these data demonstrate that neither IFN alone or together play a significant role in regulating the robust replication and disease phenotypes of the homologous murine RV in suckling mice. On the other hand, these studies clearly demonstrate that both type I and type III IFNs are required for optimal antiviral protection of the GI tract of suckling mice against the heterologous simian RRV infection, and that both IFN types independently contribute to innate antiviral defenses within the intestinal mucosal compartment (Figs 4, 5 and 7) and cooperate to restrict extra-intestinal RRV replication in other tissues (Fig 6). Our studies also identified a reduced sensitivity of IECs but not LPCs to the effects of type I but not type III IFNs as mice mature. Overall, our findings highlight a multi-faceted complexity of the virus-host interactions and reveal a well-orchestrated spatial and temporal tuning of innate antiviral responses in the intestinal tract where two types of IFNs through distinct patterns of their expression and distinct but overlapping sets of target cells coordinately regulate antiviral defenses. Our findings also highlight the fact that the antiviral capacity of the various IFNs can vary very significantly between strains of the same virus in a host dependent manner.
Conventional specific pathogen-free (SPF) WT C57BL/6J mice were purchased from Jackson Laboratory. Mice lacking functional IFN-λ receptor (Ifnlr1-/-) were generated in the laboratory. Recombineering techniques were used to create a KO targeting vector that contained exon 3 of the mouse IFNLR1 gene flanked with two LoxP sites and ~10 kb arms for homologous recombination (Fig 1A). A neo (G418-selection) cassette flanked with the FRT sites was introduced in front of the LoxP site in intron 4. The accuracy of all modified sequences within the targeting vector was verified by sequencing. Bruce4 mouse embryonic stem (ES) cells from C57BL/6J strain were transfected with the targeting vector, and G418-resistant ES clones were selected and screened by Southern blotting for correct integration of the targeting fragment (S1A and S1B Fig). Chromosomal DNA was obtained from G418-resistant ES clones, digested with EcoRV restriction endonuclease, subjected to Southern blotting with a hybridization probe corresponding to exons 1 and 2 of the mouse IFNLR1 gene that are positioned outside of the left arm for homologous recombination (S1A Fig). Twenty three clones were selected and their DNA was digested with AflIII restriction endonuclease, and Southern blotting was performed with a probe corresponding to exons 5, 6 and 7 that are outside of the right arm for homologous recombination (S1B Fig). One of the clones with the correct integration pattern at both 3' and 5’ ends was used for the generation of chimeric mice, and subsequently mice homozygous for the integration cassette. First, the neo cassette was eliminated by crossing the chimeric mice with C57BL/6J mice transgenic for the CMV promoter-driven flipase (Jackson Laboratory, Stock # 009086). Mice homozygous for the deletion of the neo cassette were selected, followed by the selection against the flipase gene. These mice were then crossed with C57BL/6J mice transgenic for the CMV promoter-driven Cre recombinase (Jackson Laboratory, Stock # 006054), and mice homozygous for the deletion of the IFN-λR1 exon 3 and lacking the Cre gene were selected. These IFN-λ receptor-deficient animals were crossed with C57BL/6J mice lacking functional type I IFN receptor (Ifnar1-/- mice) in the laboratory of Jörg Fritz at McGill University; and Ifnar1-/- and Ifnar1-/-Ifnlr1-/- mice were provided for these studies. Congenic B6.A2G-Mx1 mice carrying intact Mx1 alleles [44] and EDIM-RV isolate were provided by P. Staeheli. All mouse strains on C57BL/6J background were maintained at SPF barrier facility at NJMS, Rutgers. Mouse strains on 129S6/SvEv background were described previously [25] and maintained in the vivarium at the Veterinary Medical Unit of the Palo Alto VA Health Care System.
Eight-day-old suckling mice were orally inoculated with 104 DD50 of the murine EW-RV strain or 4x106 FFU of the simian RRV strain. The EW-RV strain was derived following serial suckling mouse passage from the original E. Kraft EDIM-RV isolate [33]. From 2 to 12 dpi for EW-RV infection or from 2 to 8 dpi for RRV infection, animals were examined daily for the occurrence of diarrheal disease. The percentage of diarrhea among inoculated littermates during the course of infection for each group was recorded. To measure the effects of RV infection and IFN deficiency on suckling mouse body weight gain, EW-RV or RRV-infected or non-infected WT 129S6/SvEv, Stat1-/- and Rag2-/- mice were weighed daily during the course of experiments. Daily mouse weight ratio was calculated for each infected mouse as weight of infected mouse (g) / mean weight of uninfected control mice (g) of the same age. Fecal specimens (approximately 10–20 μl) were collected from EW-RV-infected suckling mice into pre-weighed eppendorf tubes. Samples were stored at -80ºC prior to fecal EW-RV shedding detection by ELISA. At indicated day post EW-RV or RRV infection, a number of mice from each experimental group were sacrificed for tissue collection and histology.
All IFNs were injected intradermally in adult or suckling mice. Human hybrid IFN-αA/D and mouse IFN-λ2 were used at the concentrations indicated in the figure legends. Human IFN-αA/D was previously shown to be highly active on many mouse cell types in vitro and in vivo [53].
Human SW-1116 cells (ATCC CCL-233) were plated at confluency onto transwell filters and cultured for 56 days (media was changed every other day) until epithelial layer of well-polarized epithelial cells with high trans-epithelial resistance (TER > or = 2000 ohm/cm2) was established. In parallel, SW-1116 cells were also grown in continuously proliferating cultures on regular plates. The cells were left untreated or treated at the apical or basolateral surfaces with various amounts of IFN-α or IFN-λ as indicated. At 72 h, the cells were collected, and levels of MHC class I antigen expression were evaluated by flow cytometry.
At various time points post RRV infection, suckling mice were anesthetized, and tissue samples from liver, MLN, and small intestine were collected and stored at -80°C. Before assay, the thawed tissue samples were individually weighed and made to 10% (wt/vol) suspensions with serum free M199. Samples were homogenized in 5 ml polypropylene tubes and the homogenates were activated with trypsin (10 μg/ml) for 1 h at 37°C in a 5% CO2 incubator. Total homogenates were centrifuged at 1,500 rpm for 10 min and the supernatants were serially diluted in serum free M199. MA-104.1 cells (ATCC CRL-2378.1) were inoculated in a 24-well plate with 0.1 ml of diluted supernatant. After absorption for 1 h at 37°C in a 5% CO2 incubator, cells were re-fed with 500 μl 10% FBS M199 supplemented with 2 mM L-glutamine and penicillin/streptomycin (100 μg/ml / 100 I.U.) and cultured for 24 h. The cells were then fixed with 10% phosphate-buffered formalin for 30 min. Viral antigenic focus detection was accomplished by incubation with rabbit anti-rotaviral hyperimmune serum for 1 h, then alkaline phosphatase (AP)-conjugated goat anti rabbit IgG (Invitrogen) for 1 h, then the AP substrate BCIP/NBT (5-bromo-4-chloro-3-indolyl phosphate/nitro blue tetrazolium) (Sigma). Between each step, wells were washed twice with PBS-TA (PBS containing with 0.1% Triton X-100, 0.1% BSA, 0.1% sodium azide). The positive cells were enumerated and virus titers were expressed as focus forming unit (FFU) per gram of tissue.
Fecal EW-RV antigen shedding was measured as previously described [54]. Briefly, fecal samples were made to 10% wt/vol suspensions with PBS. Ninety-six-well polystyrene high binding plates (E&K Scientific) were coated with guinea pig anti-rotavirus hyperimmune serum. After washing and blocking with 5% BLOTTO (wt/vol fat free power milk in PBS) suspended stool samples were added to the plates for overnight incubation at 4°C. The plates were washed and rabbit anti-rotavirus hyperimmune serum was added to the plates. The plates were washed, and horseradish peroxidase (HRP)-conjugated goat anti-rabbit immunoglobulin G (IgG) antibody (γ chain specific, Thermo scientific) was added to the plates. TMB (3,3’,5,5’-Tetramethylbenzidine) substrate (Kirkegaard & Perry Laboratories) was used for the color reaction. A serial dilution of a standard RRV stock was used in each plate to control the level of color development. The absorption at A450 nm was measured with an ELISA reader (Bio-Tek Instruments). The fecal viral antigen shedding data were expressed as optical density (OD) values.
The small intestines were formalin-fixed and paraffin-embedded. Antigen retrieval was performed on deparaffinized 5 micron sections which were then incubated for 5 min with Super Block (ScyTek #AAA999), and 10 min in 3% H2O2 to block the endogenous peroxidase activity. Sections were then incubated at 4°C overnight with polyclonal goat anti-rotavirus antiserum (NCDV; Meridian, LS; 1:500) or monoclonal rabbit anti-phospho-STAT1 (Tyr701; 58D6) (Cell Signaling; 1:500). Slides were washed 2 times in PBST (0.05% Tween-20 in PBS), then incubated at room temperature for 30 min with UltraTek anti-Goat biotinylated antibody (Ready to Use) (ScyTek #AGL125) or UltraTek anti-rabbit biotinylated antibody (Ready to Use) (ScyTek #ABK125), followed by a 20 min room temperature incubation with UltraTek Streptavidin/HRP (Ready to Use). NovaRED substrate solution (Vector, SK-4000) was used as a substrate. After immunostaining, tissue sections were washed twice in water and counterstained with Mayer’s haematoxylin and Scott’s bluing buffer.
Mice were sacrificed and sections of the small intestines (all tissues (bulk) of the small intestine) were collected and lysed in Trizol (Life Technolgies) on ice. Total RNA was extracted following the manufacturer’s instructions and subjected to DNAse digestion before use in qRT-PCR. Synthesis of cDNA and subsequent microfluidics PCR on the Fluidigm platform was done as described earlier [6]. Serial 10-fold dilutions of mouse reference RNA (Agilent) were run in duplicate for each PCR run. Relative gene expression in infected and uninfected mouse intestinal samples was derived using the 2dCt method [6] with reference RNA serving as a calibrator and HPRT as housekeeping control.
Cell lysates were collected in lysis buffer containing protease and phosphatase inhibitors. Equal amounts of total protein was separated on 7.5% SDS-PAGE gels, transferred to Nitrocellulose 0.45 μm membrane (BIO-RAD), and subsequently probed with antibody against phosphorylated STAT1 (pY701; BD #612133) and β-actin (Sigma #A5441).
The supernatants of intestinal homogenates of RRV-infected mice were prepared as described above and assayed for IFN-λ protein using commercially available ELISA (R&D Systems), and for IFN-α/β protein by bioassay as previously described [55].
Sigmaplot 12.5 or GraphPad Prism software was used for data analysis. Virus levels in tissue were determined by either focus forming unit assay or real time quantitative RT-PCR, and was analyzed with one-way ANOVA and Bonferroni multiple comparison test with the log-transformed viral titers.
All animal studies were approved by the NJMS Institutional Animal Care and Use Committee (Protocol 13009C0316) and the VA Palo Alto Health Care System Institutional Animal Care and Use Committee (Protocol GRH0022/GRH1397) and carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.
|
10.1371/journal.pgen.1006414 | Budding Yeast Rif1 Controls Genome Integrity by Inhibiting rDNA Replication | The Rif1 protein is a negative regulator of DNA replication initiation in eukaryotes. Here we show that budding yeast Rif1 inhibits DNA replication initiation at the rDNA locus. Absence of Rif1, or disruption of its interaction with PP1/Glc7 phosphatase, leads to more intensive rDNA replication. The effect of Rif1-Glc7 on rDNA replication is similar to that of the Sir2 deacetylase, and the two would appear to act in the same pathway, since the rif1Δ sir2Δ double mutant shows no further increase in rDNA replication. Loss of Rif1-Glc7 activity is also accompanied by an increase in rDNA repeat instability that again is not additive with the effect of sir2Δ. We find, in addition, that the viability of rif1Δ cells is severely compromised in combination with disruption of the MRX or Ctf4-Mms22 complexes, both of which are implicated in stabilization of stalled replication forks. Significantly, we show that removal of the rDNA replication fork barrier (RFB) protein Fob1, alleviation of replisome pausing by deletion of the Tof1/Csm3 complex, or a large deletion of the rDNA repeat array all rescue this synthetic growth defect of rif1Δ cells lacking in addition either MRX or Ctf4-Mms22 activity. These data suggest that the repression of origin activation by Rif1-Glc7 is important to avoid the deleterious accumulation of stalled replication forks at the rDNA RFB, which become lethal when fork stability is compromised. Finally, we show that Rif1-Glc7, unlike Sir2, has an important effect on origin firing outside of the rDNA locus that serves to prevent activation of the DNA replication checkpoint. Our results thus provide insights into a mechanism of replication control within a large repetitive chromosomal domain and its importance for the maintenance of genome stability. These findings may have important implications for metazoans, where large blocks of repetitive sequences are much more common.
| Rif1 is a conserved eukaryotic protein implicated in regulation of both the temporal pattern of DNA replication initiation and the DNA damage response (DDR). We found that in budding yeast several of Rif1’s DDR-related phenotypes stem from its ability to interact with the Glc7/PP1 phosphatase and inhibit DNA replication initiation at the highly repetitive and highly transcribed rDNA locus. Each rDNA copy contains a potential replication origin flanked on one side by a proteinaceous replication fork barrier, a crucial player in rDNA array size maintenance. Additionally, the rDNA RFB ensures that replication proceeds in the same direction as transcription, thus presumably minimizing collisions between the replication and transcription machineries. Our results show that inhibition of rDNA origin firing by Rif1-Glc7/PP1 prevents the buildup of an excess of stalled forks within the rDNA locus, which can lead to genome instability and cell death. These findings highlight the challenges posed by the replication of repetitive loci, and in particular the need to limit DNA replication initiation events at such vulnerable regions. Our study may have important implications for metazoan genomes, which contain a much higher fraction of repetitive sequences than budding yeast. Finally, since tumor cells already exhibit elevated levels of replication stress, our results suggest that inhibition of systems that limit DNA replication initiation may jeopardize the viability of these cells and thus prove to be a useful therapeutic strategy.
| In eukaryotes DNA replication initiates from multiple sites (origins) in a characteristic sequential pattern referred to as ‘replication timing’ [1]. Replication timing is tightly regulated, with some origins being replicated early during S phase and others later [2, 3]. Mechanisms that determine replication timing are still unclear, though recent studies point to a model in which limiting factors (e.g. Sld3, Sld2, Dpb11, Dbf4) are sequentially re-distributed to origins with decreased levels of accessibility to these factors, thus generating a temporal program of origin firing [4, 5].
In the budding yeast Saccharomyces cerevisiae about one-third of all potential replication origins (autonomously replicating sequences, or ARSs) are located within the rDNA repeat array on chromosome XII [6]. The rDNA array comprises ~150–200 copies of a module containing 35S and 5S rRNA genes separated by two intergenic regions harboring an origin of replication (rARS) and a polar replication fork barrier (RFB) (Fig 1A). Interestingly, only ~20% of the rDNA origins fire during a given S phase in wild type cells [6, 7], and it has been shown that deregulation of rDNA replication leads to genomic instability [8, 9].
The rDNA repeat RFB, which is generated by sequence-specific binding of the Fob1 protein [10], is believed to prevent head-to-head collisions of the transcription and replication machineries and to mediate rDNA copy number homeostasis. Replication fork blockage caused by other proteins is observed elsewhere in the genome, for example at tRNA promoters, telomeres, silent mating type loci, dormant origins and centromeres [11–13]. The replisome protection (or pausing) complex, which consists of Tof1/Csm3 in S. cerevisiae, is essential for fork arrest at RFBs both within and outside of the rDNA [13, 14] and is proposed to act mainly by counteracting the Rrm3 helicase [14]. Consistent with this idea, genome-wide accentuation of RFBs in rrm3Δ cells leads to fork collapse and breakage, and to loss of viability in combination with mutation of DNA repair genes such as MRE11, SGS1, and SRS2 [15, 16]. Similarly, specific strengthening of the rDNA RFB by Fob1 overexpression decreases viability in mre11Δ mutants [17]. Apart from its conserved function in DNA double strand break (DSB) nucleolytic processing (resection) [18] the MRX complex (Mre11, Rad50 and Xrs2) also participates in replication fork stabilization under conditions of replication stress (e.g. dNTP depletion; [19]) and in replisome re-assembly after fork collapse [20].
Repair of broken replication forks at the yeast rDNA RFB leads to repeat array instability due to recombination-driven gain or loss of copies [10]. Accordingly, an increase in the efficiency of rDNA origin firing, such as that observed in cells mutated for the histone deacetylase Sir2 [6, 7], is associated with elevated rDNA instability [8, 9, 21], presumably due to an increase in the number of replication forks arrested at an RFB. One model suggests that, in addition to down-regulating rDNA origin firing, Sir2 also inhibits unequal sister chromatid exchange, by promoting cohesion binding within the rDNA intergenic spacers [22], thus defining a second mechanism by which Sir2 promotes rDNA stability.
Rif1, a budding yeast Rap1-interacting factor, was initially described as an inhibitor of telomerase-dependent lengthening of telomeres in yeast [23]. Rif1 is highly conserved in eukaryotes [24, 25] and has more recently been shown to be a regulator of DNA replication initiation in yeast, flies and mammals [26–33]. We and others found that Rif1, through its conserved RVxF/SILK motifs, interacts with protein phosphatase 1 (PP1; Glc7 in budding yeast), and that this interaction is crucial for inhibition of replication origin firing by counteracting the activity of the Dbf4-dependent kinase (DDK) [26, 27, 30]. Deletion of RIF1 in budding yeast leads to advancement of the replication timing of most late origins [28]. Importantly, loss of Rif1 in mouse cells leads to defects in S phase progression, hypersensitivity to the DNA polymerase inhibitor aphidicolin, and checkpoint kinase activation [31, 34]. Modulation of Rif1 activity may thus provide a valuable tool to study the molecular and cellular consequences of altering the replication-timing program.
Here we show that budding yeast Rif1 inhibits DNA replication initiation at the rDNA locus and thus promotes the stability of rDNA repeat array. Moreover, the increase of rDNA instability in rif1Δ accounts for the majority of its DNA damage response (DDR)-related phenotypes, suggesting that the rDNA is a key target of Rif1 action. These findings offer a new perspective on the relationship between replication timing, repeated DNA sequences and genome stability.
Driven by the hypothesis that disruption of RIF1 leads to an increase in the number of active replication forks during S phase [26–28, 30], we decided to determine where these forks are located. Chromatin immunoprecipitation (ChIP) of epitope-tagged DNA polymerase epsilon (Pol2) from cell cultures synchronously entering S phase provides a read-out of replication fork passage [26, 27, 35]. We thus measured Pol2 association at different loci, comparing wild type to rif1 mutants. Consistent with our previous observations [27], the timing of Pol2 recruitment to the early origins was not affected in cells deleted for RIF1 or cells mutated in its Glc7-binding RVxF/SILK motifs (see ARS305 at the Fig 1B, where in both WT and the rif1 mutants Pol2 is recruited at 45 minutes after the release into S phase). On the other hand, Pol2 was detected over a shorter time interval at the late replicating HMR locus in both rif1 mutants, which might reflect its earlier and/or faster replication. Given the fact that many dormant replication origins are located within the repetitive rDNA locus [6, 7], we also probed for the ARS element there (referred to as rARS). In wild type cells rARS displayed a peak of Pol2 binding in the middle of S phase (60 minutes following release from alpha-factor arrest). Interestingly, deletion of RIF1 or mutation of its Glc7 (PP1)-interacting RVxF/SILK motifs advanced Pol2 binding by ~15 minutes, to a time similar to that of the early ARS305 origin. Importantly, acute depletion of Rif1 from the nucleus in G1 phase by the anchor-away method (see Materials and Methods section for details) also led to advancement of Pol2 binding in the next S phase at rARS, HML and telomere sites, while replication timing of an early origin (ARS607) was not affected (S1A Fig, left and middle panels).
Origins that fire early in S phase, but not late-replicating regions, recruit the pre-replicative complex (pre-RC) component Sld3 in G1, prior to DNA replication initiation [36]. Significantly, we detected elevated Sld3 recruitment to the late replicating rARS in G1-arrested rif1Δ cells (Fig 1C), whereas the recruitment of Mcm4, part of the replicative helicase that is loaded synchronously on all origins, was not affected (S1A Fig, right panel). The decrease in Sld3 recruitment to a non-rDNA early origin (ARS607, Fig 1C) in rif1Δ cells might be due to re-localization of limiting amounts of this protein to the excess of activated late origins, both within the rDNA and elsewhere, due to the absence of Rif1. Taken together, these data indicate that Rif1’s interaction with the PP1 phosphatase (Glc7) is responsible for inhibition of rDNA replication, defining the rDNA locus as a novel Rif1-Glc7 target.
To further investigate the role of Rif1 in rDNA replication, we used bromo-deoxyuridine (BrdU) incorporation followed by anti-BrdU immunoprecipitation (IP) and quantitative PCR (qPCR) as a more direct method to measure newly synthesized DNA. We released G2/M arrested (nocodazole-treated) cells into S phase in the presence of 0.2 M hydroxyurea (HU) and BrdU (see FACS profiles in S1B Fig). HU slows fork progression and allows one to determine whether late origins fire, or are instead passively replicated by forks coming from nearby early origins. In accord with a recent genome-wide study [28], we detected higher levels of DNA synthesis in rif1Δ at late origins (ARS1212, ARS522, HMR locus and telomeres), whereas levels of BrdU incorporation at early origins (ARS305, ARS607) were not affected (Figs 1D and S1C). We also found higher BrdU incorporation in rif1Δ and rif1-RVxF/SILK mutant cells compared to wild type at and around rARS (Fig 1D). Importantly, the increased BrdU incorporation in rif1 mutants was specific to the rDNA and not a general feature of repetitive loci, since another repetitive locus (CUP1) incorporated BrdU very similarly in rif1Δ, rif1-RVxF/SILK and wild type cells. Analysis of the source data from the Peace et al. study [28] also revealed the same trend of higher BrdU incorporation at rARS in rif1Δ compared to wild type.
Next, we used 2D agarose gels [37] to observe directly the replication intermediates at the rDNA locus, again from cells released from a G2/M arrest into S phase in the presence of HU. Deletion of RIF1 led to a dramatic increase in bubble arc, Y arc, RFB spot, and X-shaped molecules signals at the rDNA in these conditions (Figs 1E and S1D), indicating a higher frequency of rARS firing. The effect of rif1Δ as seen in asynchronous cultures was less prominent (S1E and S1F Fig), presumably because the increase in fork density in the mutant also increases the rate at which blocked forks are resolved following arrival of a downstream fork moving in the opposite direction, which would convert the replication intermediates into linear molecules. This effect is nullified when synchronized cells are released into S phase in medium containing HU, which permits early origin firing but severely limits fork elongation. We thus conclude that the realm of Rif1-dependent inhibition of DNA replication initiation includes the rDNA locus.
As pointed out above, Sir2 plays an important role in several aspects of rDNA biology. We confirmed the previous observation [7] of an rDNA replication increase (as detected by BrdU incorporation) upon deletion of SIR2, and furthermore found that Sir2’s effect on rARS is quantitatively similar to that of Rif1 (Fig 2A). However, unlike rif1Δ cells, sir2Δ mutants do not display increased firing at either of the two late-replicating regions we examined, ARS522 or HMR (ARS317) (Fig 2A). To address the question of whether Rif1 acts independently of Sir2 to inhibit rARS firing, we examined a rif1Δ sir2Δ double mutant, but found no additive effect, suggesting that these two proteins act in a common pathway. In fact, and quite surprisingly, the rif1Δ sir2Δ double mutant displayed consistently lower BrdU incorporation at both rARS and a site 2 kb distant, compared to both single mutants. Nevertheless, replication at these sites was still increased at least 2-fold over that observed in wild type cells.
Using 2D gels we also observed more intensive replication and fork pausing in early S phase at the rDNA in sir2Δ cultures, similar to that in rif1Δ (Fig 2B), though with a marked difference in the relative intensity of the arc signals. Deletion of RIF1 mostly increased bubble arcs, whereas SIR2 deletion and the double deletion of SIR2 and RIF1 led to more Y arcs. This difference might be due to variations in the spatial pattern of origin activation, fork progression rates, or timing of origin activation in these mutants In conclusion, the above results indicate that Sir2 and Rif1 work in a common pathway to inhibit rARS firing, but suggest in addition that other players may be involved that create a more complex functional relationship between Sir2 and Rif1 (see Discussion).
As indicated above, replication of the rDNA repeats is highly polar in nature due to an orientation-dependent replication fork block (RFB; see Fig 1A). Replication proceeding rightwards from rARS is efficiently blocked at the RFB, which is thought to prevent potential collisions with an RNA polymerase I (RNAPI) complex transcribing the downstream copy of the 35S rRNA gene. Forks proceeding to the left from rARS, and in the direction of 35S rRNA gene transcription, are free to pass the RFB present at the upstream rDNA copy.
We hypothesized that rDNA locus stability might be sensitive to an increase in origin firing since this leads to a concomitant increase in the number of forks blocked at RFBs (Fig 1E, Fig 2B). Blocked forks can, with a certain probably, collapse, sometimes generating DNA breaks that will normally be repaired by homologous recombination (HR), non-homologous end joining (NHEJ) or alternative break-induced replication (BIR) pathways. Due to the repetitive nature of the rDNA, recombination between different repeats of the same or sister chromatids may lead to a change in the rDNA array size, which is usually referred as ‘rDNA instability’ [38]. The loss of repeats from the rDNA array can be conveniently measured when a single copy of the ADE2 gene is inserted in the array, in cells where the endogenous ADE2 gene is mutated. The ADE2 gene confers a white colony-color phenotype, whereas popping-out of this gene from the chromosome (together with adjacent repeats) leads to the accumulation of a red pigment when adenine in the medium is limiting, and the appearance of red sectors in colonies [39]. Indeed, using this colony-color marker-loss assay [39], we detected higher levels of rDNA instability in rif1Δ cells compared to wild type (Figs 3A and S2A), consistent with a recent report [40].
To further challenge the idea that the rDNA instability phenotype of rif1Δ is specifically linked to its effect on replication origin firing, we examined the rif1-RVxF/SILK mutant, which we showed previously [27] to result in a loss of the Rif1-Glc7 interaction and increased phosphorylation of two key DDK kinase targets at pre-RCs. As shown above, the rif1-RVxF/SILK mutant leads to increased and earlier rDNA origin firing (Fig 1B and 1D). We found that rif1-RVxF/SILK mutant cells also display a higher level of rDNA instability compared to wild type (Fig 3A), though smaller than the increase conferred by complete deletion of RIF1, perhaps because rif1-RVxF/SILK retains some residual binding to Glc7 [27].
We next hypothesized that strengthening of the RFB by deletion of RRM3 [15, 16], which encodes a helicase that promotes the passage of replication forks through RFBs [41], would lead to a further increase in rDNA instability. As predicted, we observed an additive increase in rDNA instability when combining rif1Δ or rif1-RVxF/SILK with rrm3Δ (Figs 3B, S2A and S2D). If the effect of Rif1 and Rrm3 on rDNA stability were linked to the RFB, deletion of the FOB1 gene, whose product is required to establish the fork block, would be expected to abolish the instability induced by rif1 mutants, rrm3Δ, or the double mutants rif1Δ rrm3Δ and rif1-RVxF/SILK rrm3Δ. This is indeed what we found (Figs 3B, S2A and S2D), strongly suggesting that Rif1, as well as Rrm3, act through the RFB. Surprisingly, neither single mutation (rif1Δ or rrm3Δ) nor the double mutation rif1Δ rrm3Δ affected cell growth, either under normal conditions or in the presence of DNA damaging agents (S2C Fig), suggesting that DNA repair pathways in these cells are sufficient to cope with the increased load of stalled forks [15, 16].
In accordance with its rDNA replication phenotype (Fig 2), sir2Δ also displays a large increase in rDNA instability that is fully rescued by FOB1 deletion (Figs 3C and S2B). The increase in rDNA instability caused by sir2Δ is larger than that of rif1Δ and is unaffected by rif1Δ, consistent with a previous report [40]. Increased instability of the rDNA locus leads to heterogeneity in the size of chromosome XII in a population of the cells [22]. As expected, then, pulse field gel electrophoresis revealed a heightened smearing (broader and less sharp band) of chromosome XII in rif1-RVxF/SILK and rif1Δ cells (Fig 3D), though not to the same extent as in sir2Δ (S3A Fig), consistent with their varying effect on rDNA stability measured by the sectoring assay. Again as expected, we found that the effect of rif1Δ on chromosome XII heterogeneity was reversed by the fob1Δ mutation (Fig 3D). Deletion of either RIF1 or RRM3 increases rDNA instability (Figs 3B and S3C), but only rrm3Δ leads to an increase in the ratio of Fob1-dependent blocked forks at the RFB to total forks at rDNA [41] (compare 2D gels in S1E and S3C Figs), since in rif1Δ the increase in RFB signal is paralleled by an increase in the number of forks at the rDNA (Figs 1E and 2B). These findings further support the argument that Rrm3 acts directly at RFBs, whereas Rif1 primarily acts through controlling DNA replication initiation.
Elevated blockage and collapse of replication forks at the rDNA may also lead to HR-dependent “popping-out” of rDNA repeats in the form of episomal circles [42], referred to as extrachromosomal rDNA circles (ERCs). Consistent with elevated rDNA array instability, we observed increased levels of ERCs in rif1Δ, rrm3Δ and sir2Δ cells (Figs 3E and S3B). It is not known whether the rARS is more or less active on the episomal ERCs, but it is conceivable that the change in ERC number in a cell may affect the apparent rDNA replication phenotype. Deletion of FOB1, which has been shown to significantly reduce ERC formation ([83]; Figs 3E and S3B) abolished ERC accumulation in rif1Δ and sir2Δ mutants (Fig 3E). However, we found that fob1Δ did not affect the rif1Δ-induced increase in rDNA replication, as detected by BrdU incorporation and 2D gels (Figs 3F and S3D), confirming that the loss of Rif1 influences chromosomal rDNA origin firing. Taken together, these results show that Rif1 and Sir2, but not Fob1, are involved in control of replication initiation at the rDNA locus.
Rif1 was originally identified as a telomere-binding protein involved in TG-tract length regulation. Deletion of FOB1 did not affect rif1Δ-dependent telomere elongation (S4A Fig), arguing that telomere- and rDNA-related functions of Rif1 are separable. However, early studies [45,47,84,85] indicated that Rif1 can compete with SIR proteins for binding to the Rap1 C-terminus at telomeres and that this competition can indirectly affect the availability of SIR proteins for binding elsewhere in the genome, in particular at silent mating type loci (where a Sir2/3/4 complex assembles) and within the rDNA, where Sir2 binds at two distinct sites. A more recent report thus suggested that rif1Δ increases rDNA instability indirectly by favoring the re-localization of Sir2 from its binding sites in rDNA to telomeres and silent mating type loci [40]. To determine whether Rif1 acts directly to affect rDNA stability, or instead works by modulating the distribution of Sir2 at its different target sites (rDNA, HM loci and telomeres), we first assessed rDNA instability in the rif1-RBM mutant, which, like rif1Δ, leads to an increase in telomeric silencing and telomere TG-tract length [43]. We found that rif1-RBM has no effect on rDNA stability (Fig 4A), suggesting that increased SIR-mediated telomeric silencing and telomeric TG tract length do not lead to rDNA instability. Furthermore, neither deletion of TEL1, which reduces telomere length in a rif1Δ background [44], nor deletion of RIF2, which further increases telomere length and telomeric silencing [45], had any effect on rif1Δ-promoted rDNA instability (Fig 4A). We also examined sir4Δ cells, where the Sir2 protein cannot be recruited to either telomeres or HM loci and is thus liberated for enhanced action within the rDNA [46]. However, sir4Δ had no significant effect on rif1Δ-induced rDNA instability (Fig 4B). Taken together, these findings do not support the notion that Rif1 affects rDNA stability by influencing Sir2 distributions in the nucleus, but are instead consistent with Rif1 having a direct effect on rDNA stability.
Next, we measured binding of Sir2 to chromatin by ChIP-qPCR in rif1Δ cells. We found that Sir2 binding was increased at the HMR silent mating-type locus (Fig 4C, left panel), in line with the idea that Rif1 competes with the SIR complex for Rap1 binding at HMR in wild type cells [47]. However, in contrast to a recent report [40] that found a small effect of rif1Δ on Sir2 binding at IGS1 (near the RFB) using a semi-quantitative ChIP assay, we found no difference in Sir2 binding there by ChIP-qPCR, nor at three other sites along the rDNA locus: at rARS (which is located in IGS2), at an adjacent region at the 35S rRNA gene promoter, and at a site within the 35S rRNA gene coding sequence (Fig 4C, left panel). Fob1 ChIP at rDNA was also unaffected by rif1Δ (Fig 4C, right panel). Taken together, these data suggest that any influence of Rif1 on SIR complex distribution is insufficient to account for its effects on rDNA instability, and instead argue that Rif1 has a direct effect on rDNA stability by maintaining the low level of rARS firing.
Martina et al. [48] recently proposed that Rif1 physically counteracts Rad9 binding to DSBs. We therefore asked whether rif1Δ-induced rDNA instability stems from unrestrained activity of Rad9. However, deletion of RAD9 alone had no effect on rDNA instability in the marker-loss assay and did not alleviate the increased instability caused by rif1Δ (S4B Fig). We therefore conclude that the effect of rif1Δ on rDNA stability is unrelated to the activity of Rad9. Since the rad9Δ mutation abolishes the DNA damage checkpoint (DDC) [49], these results also argue that rif1Δ-dependent elevation in rDNA instability is not a consequence of DDC activation.
Arrested replication forks need to be stabilized and/or restarted to avoid formation of DSBs and/or inappropriate recombination events [13]. Increased numbers of stalled replication forks might therefore compromise cell viability. Consistent with this idea, elevating the strength of RFBs, either by removal of the Rrm3 helicase or by overexpression of Fob1, leads to synthetic sickness in combination with disruption of the MRX complex [15–17], probably due to a role for MRX in fork repair [50], fork restart [20], or fork stabilization at RFBs [17].
As already reported, deletion of RIF1 also severely compromises growth of mre11Δ cells, both in untreated cells and upon exposure to phleomycin, which generates DSBs (Fig 5A; [48, 51]). We reasoned that this effect of rif1Δ might stem, at least in part, from an increased number of replication forks that are pausing at an rDNA RFB in rif1Δ cells, and thus prone to collapse and subsequent DSB formation. If this were the case, deletion of FOB1 or alleviation of the fork pause through removal of the replisome pausing complex (Tof1/Csm3) would be expected to rescue this synthetic sickness. Indeed, fob1Δ, tof1Δ, or csm3Δ deletions completely rescued rif1Δ mre11Δ synthetic sickness, both in normal conditions and upon treatment with genotoxic agents (Figs 5A, S5A and S5B).
We next examined the premise that the increased number of RFB-stalled forks in rif1Δ mre11Δ cells stems specifically from the effect of Rif1 on rARS firing. In support of this notion, we found that mutation of the Rif1 RVxF/SILK motifs alone conferred a synthetic sickness phenotype in combination with mre11Δ that was comparable to that of rif1Δ, whereas rif1-RBM had no such effect (Figs 5B and S5C). Given our finding that rif1-RVxF/SILK, but not rif1-RBM, increases rARS firing [27], these data point to a primary effect of Rif1 on rARS firing as the cause for synthetic sickness in combination with mre11Δ. To test this idea further, we introduced the temperature-sensitive cdc7-4 mutation, which compromises DDK kinase activity and thus decreases replication initiation rates genome-wide [52], into our rif1Δ mre11Δ strain. At 30°C, where compromised cdc7-4 activity begins to affect growth in a rif1Δ background, we note significant alleviation of rif1Δ mre11Δ synthetic sickness, both in untreated and phleomycin-treated cells (Fig 5C). As expected, at 37°C cdc7-4 is unable to support viability in either the rif1Δ or the rif1Δ mre11Δ background.
Finally, we reasoned that if elevated rDNA repeat replication coupled with fork blockage at the RFB were the source of toxicity in rif1Δ mre11Δ double mutants, then removing the rARS/RFB replication system from chromosome XII should improve growth of these cells. To test this idea we took advantage of a previously described rDNA array deletion strain, which leaves only 2 chromosomal rDNA repeats (rdn1Δ strain) [53]. Survival of this strain is maintained by a multi-copy plasmid harboring both the 35S and 5S rRNA genes but having a 2 μm replication origin instead of rARS. As predicted, this rDNA repeat- and rARS-deficient strain displayed no evidence of rif1Δ mre11Δ synthetic sickness, nor any effect of FOB1 deletion (Fig 5D).
In line with its strong additive effect in combination with rif1Δ (Fig 3B), rrm3Δ also displays strong synthetic sickness with mre11Δ. However, consistent with a more general role of Rrm3 at replisome barriers throughout the yeast genome [41], rrm3Δ synthetic sickness with mre11Δ was more severe than that of rif1Δ mre11Δ and was significantly but not completely suppressed by fob1Δ (S6A Fig). Moreover, double deletion of RRM3 and RIF1 was inviable in combination with mre11Δ, in line with the additive effects of Rrm3 and Rif1 on rDNA integrity (S6A Fig).
Consistent with the fact that sir2Δ, like rif1Δ, leads to elevated rDNA origin usage and rDNA instability (Fig 2, Fig 3C and 3D), we found that sir2Δ also shows fob1Δ-suppressible synthetic sickness with mre11Δ (Fig 5E). The triple mutant rif1Δ sir2Δ mre11Δ was slightly less sick than rif1Δ mre11Δ (S6B Fig), consistent with a partial decrease in replication at the rDNA when combining rif1Δ with sir2Δ (Fig 2A and 2B). In line with our conclusion that the Rif1 effect on rDNA is independent of Sir2 re-localization, sir4Δ mutation did not alleviate the synthetic sickness of rif1Δ with mre11Δ (S6C Fig) as it had no significant effect on rif1Δ-induced rDNA instability (Fig 4B).
We next reasoned that the burden of elevated replication in rif1Δ should lead to a synthetic growth defect in combination with other mutations affecting replisome integrity. In fact, combining rif1Δ with deletion of CTF4, which encodes a replisome component that couples CMG helicase and DNA polymerase alpha/primase [54], led to a strong synthetic sick phenotype (Fig 5F). The same is true for deletion of MMS22, whose product has been proposed to be recruited by Ctf4 to the replisome [55] as part of the Rtt101-Mms1-Mms22 ubiquitin-conjugating complex, essential for replisome maintenance at endogenous impediments and upon challenge with genotoxic agents [56] (Fig 5F; [57]). It worth noting that, together with histone acetyltransferase Rtt109, Mms22 also influences the downstream repair events at blocked forks [58–60] and participates in the maintenance of rDNA array size [61]. The synthetic interaction of rif1Δ with both ctf4Δ and mms22Δ was also alleviated by the deletion of FOB1, again consistent with the idea that the primary defect occurs at the rDNA RFBs (Fig 5F). All of the above results show that the function of Rif1 in rDNA stability becomes essential for survival when replisome maintenance and/or DNA break repair at rDNA RFBs is compromised.
Treatment of cells with the ribonucleotide reductase inhibitor HU leads to DNA replication checkpoint (DRC) activation due to accumulation of single-stranded DNA at stalled replication forks, and is accompanied by Rad53 phosphorylation (reviewed in ([62]). Interestingly, the strength of the DRC, as measured by Rad53 phosphorylation (Rad53-p), correlates with the number of arrested replication forks [63]. We thus reasoned that increased replication in early S phase in rif1Δ cells might lead to higher DRC activation (Fig 6A). Indeed, we observed reproducibly higher levels of Rad53-p following HU treatment in both rif1Δ cells (in the W303 background (Fig 6B–6D), and two other backgrounds, S288C and JC482 (S7A Fig, upper panels)), and in cells where Rif1 was rapidly depleted from the nucleus by anchor-away (S7A Fig, lower panel). Moreover, only the rif1-RVxF/SILK mutant, but not rif1-RBM, exhibited elevated DRC upon HU treatment (Fig 6B), emphasizing the connection with increased replication. This effect of Rif1 loss depended on the DRC adaptor Mrc1, but not the DDR adaptor Rad9 (S7B Fig), indicating that it is a bona fide DRC response [62]. Interestingly, the kinetics of Rad53 de-phosphorylation upon HU withdrawal (recovery from the DRC) were similar in wild type and rif1Δ cells (S7C Fig), suggesting that DRC deactivation is not altered by rif1Δ, consistent with previous observations [28]. Importantly, despite the elevated DRC, rif1Δ does not confer increased sensitivity to genotoxic agents such as HU, MMS, phleomycin or camptothecin (S2C and S3 Figs; [48]), suggesting that Rif1 has no essential role in the DDR, either in DSB repair, repair of damaged forks, or re-start of stalled replication forks.
Deletion of FOB1 had no effect on either the DRC (Fig 6C) or the replication phenotype (Fig 3F) of rif1Δ cells, suggesting that their elevated Rad53-p might result from an increase in rDNA replication per se and not from subsequent fork stalling and/or breakage at the rDNA RFBs. However, we found that removal of the majority of the rDNA repeats, leaving only 20 or 2 copies [53, 64], also did not alleviate the rif1Δ-induced increase of Rad53-p upon HU treatment (Fig 6D). This suggests that even when a large increase in rDNA replication initiation is not possible, the effect of rif1Δ on replication initiation elsewhere is sufficient to lead to elevated Rad53-p, a conclusion consistent with the observation that rif1Δ increases origin activation near telomeres and at other sites in the genome (this report and [26–28, 30]).
It is important to note that in contrast to rif1Δ, deletion of SIR2 did not increase the level of Rad53-p upon HU (Fig 6E). This might be due to the fact that in sir2Δ cells the increase in rARS usage is accompanied by a decreased efficiency of firing of genomic origins outside of the rDNA locus [7], whereas rif1Δ leads to elevation of the total replication load on the genome, due to its effects at both at rDNA and elsewhere. Thus, rDNA origins are only one example of a general Rif1 inhibitory role in DNA replication initiation. Nevertheless, due to the intrinsic vulnerability of the repetitive rDNA locus, caused at least in part by the RFB [64], Rif1’s role in genome integrity is manifested largely at this site.
Genome integrity is maintained by various mechanisms that either prevent damage to DNA or mediate its repair once the damage has occurred [65]. Inhibition of replication initiation at late-firing or normally dormant origins following DNA damage is one such preventive measure, referred to as the ‘replication checkpoint’ [62]. The exact benefit of late origin inhibition remains enigmatic. For example, it is still unclear whether or not a failure to down-regulate origin firing in the wake of exogenous damage leads to a decrease in genome integrity [66, 67].
Given the DDR-related genetic interactions and phenotypes of budding yeast Rif1 [48, 51] and the recently discovered role of mammalian Rif1 in the DSB repair pathway choice [68] we sought to determine whether yeast Rif1 participates directly in DSB repair, and if so, through what mechanism. To our surprise, we found instead that most of Rif1’s DDR-related phenotypes can be explained by its inhibitory effect on DNA replication within the repetitive rDNA locus and its consequent effect at the rDNA RFB (Fig 7). Our findings suggest that the down-regulation of rDNA replication by Rif1 is important to limit replication fork stalling at RFBs, and presumably a concomitant increase in fork collapse and DNA breakage. In the absence of Rif1, factors involved in either replication fork stabilization or repair (e.g. Mre11, Mms22, Ctf4 and perhaps others) become limiting for survival. This model places Rif1 at the top of a cascade of events leading to DNA damage and impaired growth.
Deletion of RIF1 or disruption of its interaction with PP1 (Glc7) phosphatase through mutation of the conserved RVxF/SILK motifs led to elevated instability of the rDNA array as detected by the frequency of marker loss from within the array, popping-out of the repeats in the form of ERCs and increased smearing of the chromosome XII on PFGE gels (Figs 3 and S3). Remarkably, the loss of Rif1 did not lead to either a specific increase or decrease in the size of the rDNA array compared to WT (Figs 3D and S3A) but rather to increased variation of the size of the rDNA array in the population of cells, which manifests itself as increased smearing of the chromosome XII band in the PFGE assay (Figs 3D and S3A). This variability may stem from ‘out of register’ repair of damaged rDNA repeats at the same or sister chromatid, from replication fork collapse and re-start on another repeat, or from re-insertion of ERCs into the chromosomal rDNA array. Furthermore, we argue that the rif1Δ-dependent increase in rDNA replication is upstream of elevated rDNA instability, since complete alleviation of the latter by removal of the essential causal rDNA fork-blocking factor Fob1 fails to reverse the enhanced rDNA replication in rif1Δ cells. Indeed, fob1Δ rif1Δ double mutant cells have a fixed rDNA array size (with a sharp chromosome XII band on PFGE), don’t accumulate ERCs, but still exhibit an increase in rDNA replication. These data show that chromosomal (rDNA array) instability and not exclusively that of episomal (ERC) rDNA repeats, are repressed by Rif1-Glc7 inhibition.
Our genetic data place Rif1 in the same pathway as Sir2 with respect to both rDNA replication control and rDNA stability, with the activities of both proteins being essential for inhibition of rARS firing. Based upon the known properties of both proteins, we envisage that Sir2 could restrict accessibility of the rDNA origins to replication initiation factors whilst Rif1-Glc7 decreases their activity by targeted dephosphorylation (e.g. Mcm4 and Sld3, [27]) (Fig 7A). Although both sir2Δ and rif1Δ deletions lead to more intensive rARS firing, we note an important difference in their effects on the firing of late origins outside of the rDNA. While sir2Δ leads to a decrease in the activity of non-rDNA origins, in accordance with models proposing the re-localization of limiting replication factors during S phase [4, 7], rif1Δ instead leads to increased firing of non-rDNA origins (Fig 2A and [27]). Our data do not contradict the limiting factors model, as we indeed see some evidence of Sld3 re-localization from an early origin in rif1Δ cells (Fig 1C). We imagine that the limiting factors, whose low abundance is unaffected by rif1Δ [27], become more active in origin firing in the rif1Δ setting due to hyper-activation of DDK and a concomitant increase in CMG DNA helicase activation [27].
Irrespective of the mechanisms of action of Rif1 and Sir2, their inhibition of rARS firing decreases the frequency of stalled forks at RFB and thus DSBs, thereby stabilizing the rDNA repeat locus (Fig 7A). When Rif1 or Sir2 are not present, this replication inhibition is at least partially lost, with a consequent increase of rDNA instability (Fig 7B and 7C). It is also worth noting that Sir2 is loaded onto two locations along the rDNA repeat: at the RFB and near the promoter of the 35S rRNA gene, through interactions with Net1 and its paralog Tof2 [69–71]. As deletion of FOB1 leads to the loss of Sir2 binding from the RFB site [69], we conclude that the 35S rRNA promoter-bound Sir2 complexes are sufficient to inhibit replication initiation events from the rARS (Figs 3F and 7A). Furthermore, we note that fob1Δ had no effect on the rif1Δ-dependent increase in BrdU incorporation at the late origins (Fig 3F) and on rDNA replication (Figs 3F and S3D).
At present we do not understand why the rif1Δ sir2Δ double mutant displays weaker BrdU incorporation at the rDNA than that of either single mutant. One potential mechanism may involve re-localization of DNA replication factors from rDNA to genomic regions de-repressed only in the rif1Δ sir2Δ double mutant and not in single deletions. Further whole-genome replication profiling of the respective mutants would be necessary to address this issue. Alternatively, other factors may come into play when both Rif1 and Sir2 are absent. We note that the Rpd3 deacetylase has been implicated in control of rARS firing [7] and may thus play an unanticipated role in the rif1Δ sir2Δ double mutant.
Although both rif1Δ and sir2Δ mutant cells have elevated numbers of replication intermediates at the rDNA and behave similarly in the BrdU incorporation assay, we noted a marked difference in the relative intensities of bubble and Y arcs in early S phase between rif1Δ and sir2Δ mutant cells. Namely, rif1Δ cells had more bubble arcs and X-shaped molecules, whereas sir2Δ and sir2Δ rif1Δ cells exhibited more prominent Y arcs. We speculate that these patterns may arise due to the repetitive nature of the rDNA array, where the spacing of the activated origins of replication and the rates of fork progression would affect the apparent replication patterns. For instance, one can imagine that the elevated bubble arcs in rif1Δ cells may be an indication of either more widely spaced activated origins (therefore the chance of merging of forks early in S phase would be low), or of slower fork progression speed. The increase in Y arcs in sir2Δ would fit with previously observed clustering of the activated origins in groups of ca. 2–3 adjacent repeats [6]. Deletion of SIR2 was shown to increase the sheer number of clusters without disrupting cluster formation. The clustering would lead to faster merging of the forks from adjacent repeats and loss of the bubble arc signal from the 2D gels. We imagine that the clustering of origins might be also beneficial for faster rescue of the forks arrested at the RFBs. Further studies using methods that give spatial resolution, such as DNA combing or electron microscopy would be necessary to investigate the effects of Rif1 and Sir2 on the spacing of origin firing and fork rates at the rDNA array.
Although both sir2Δ and rif1Δ show a similar increase in rDNA replication initiation, it is interesting to note that Sir2 has a quantitatively larger influence on rDNA stability compared to Rif1. This phenotype has been ascribed to the repressive effect of Sir2 on a bi-directional non-coding promoter (E-pro) located between the 5S rRNA gene and the RFB [22], and it will thus be interesting to determine whether Rif1 also regulates this promoter, with perhaps a weaker effect compared to Sir2. In contrast to its weaker effect on both rDNA silencing and rDNA stability, the Rif1 effect on rDNA replication is very similar in magnitude to that of Sir2, and presumably operates through a different mechanism (Glc7 recruitment versus deacetylation). It is conceivable that the failure to dephosphorylate the previously identified Rif1-Glc7 targets at the origins of replication (i.e. Mcm4 and Sld3) [27] is instrumental in the increased replication and instability of rDNA locus in rif1 mutants. However, our data do not exclude the possibility that other Rif1-Glc7 dephosphorylation targets exist at the rDNA locus, which could mediate the observed phenotypes. The deacetylation targets of Sir2 relevant for the biology of the rDNA locus also remain poorly explored. Further studies aimed at characterizing known and identifying additional Rif1-Glc7 and Sir2 target proteins should help to shed light on this important question. Another interpretation of our data is that Sir2 is required to recruit Rif1 to its site of action within the rDNA. At present, though, we have no evidence for Rif1 binding within the rDNA from ChIP assays, where we did not detect enrichment above background levels.
The detrimental effect of enhanced replication in a uni-directionally replicated locus may seem paradoxical. Indeed, one can imagine that the elevated number of unimpeded forks at the rDNA in rif1Δ and sir2Δ cells might be able to efficiently ‘‘rescue” the ones blocked at RFBs, thus neutralizing the effect of the latter, or perhaps even increasing rDNA stability. However, it may also be the case that the proximity of the RFB to the rARS (~ 1.2 kb, in the same rDNA repeat), compared to the nearest possible non-blocked fork (~ 8 kb from rARS in the adjacent rDNA repeat), will mean that forks blocked at RFBs will have to persist for an extended period of time before they can be rescued by a non-blocked fork approaching from the other side. Furthermore, it was shown that both elevated and decreased DNA replication initiation rates at rARS increase rDNA instability [8], and we imagine that some of the DNA replication factors recently shown to be necessary for rDNA stability [72] may act through regulating the replication initiation frequency at rARS, in addition to replisome integrity.
Since Rif1 is recruited to DNA DSBs generated by induction of the HO endonuclease [48] it is possible that Rif1 also participates directly in transactions that occur at accidental DSBs in general, or at DNA breaks that occur specifically at RFB sites. With respect to the former possibility, we note first that rif1Δ mutants themselves are not overtly sensitive to DNA damaging agents, and those mutations that display synthetic growth defects in combination with rif1Δ (mre11Δ, mms22Δ, or ctf4Δ) so far point to a specific role of Rif1 within the rDNA. However, our data do not rule out a subtle role for Rif1 in repair at all DSBs, and this is a subject worth further investigation. Regarding Rif1’s role within the rDNA, our data point to a specific role for Rif1 Glc7 (PP1) recruitment, and by inference in rDNA replication control. Nevertheless we cannot rule out an additional downstream role of Rif1 in processing breaks generated at the RFB. Finally, we note that the fob1Δ background, which would appear to bypass the role of Rif1 in the rDNA stability, may be a valuable tool to study rDNA independent functions of Rif1 in DNA replication and DNA damage response.
Our finding that the effect of budding yeast Rif1 on replication timing is of most consequence at the repetitive rDNA locus may have important implications in more complex eukaryotes where repetitive DNA sequences are much more prevalent. We imagine that there is a strong selection for replication origins within extensive repetitive sequences, and as a consequence mechanisms that help to assure that not all of these identical elements fire within any given cell cycle.
All yeast strains described in this study are listed in the S1 Table. General yeast manipulations were done according to standard methods [73]. For the growth assays, overnight cultures of the indicated genotypes were serially diluted 1:10 and spotted onto solid YPAD medium or YPAD medium supplemented with phleomycin (PHL), methyl methanesulfonate (MMS) or camptothecin (CPT). Plates were incubated at the indicated temperature for 2–4 days before being photographed.
The Rif1 anchor away strain (RIF1-FRB) was constructed on the basis of the starting strain HHY168, which contains RPL13A-2XFKB12, tor1-1, and fpr1Δ::NAT alleles [74], by transformation with a PCR amplified FRB tag substituting the stop codon of the RIF1 gene. The depletion of Rif1 from the nucleus was achieved by addition of rapamycin (1 μg/ml) to the yeast culture.
Cell cycle synchronization for DNA polymerase ChIP-qPCR experiments was achieved as described in [27].
BrdU incorporation was performed as described [75, 76] with minor modifications. Exponentially growing yeast culture arrested in G2/M phase (10 μg/ml nocodazole, US Biological) were pelleted, washed two times and released in fresh media containing 0.2 M hydroxyurea (US Biological) and 400 μg/ml BrdU (Sigma-Aldrich) for 2 hours. Genomic DNA extracted with phenol/chloroform and isopropanol precipitation was sonicated to 500–1000 bp fragments, purified with High Pure PCR Cleanup Kit (Roche) and denatured at 98°C for 5 minutes. Sonicated DNA was immunoprecipitated overnight with 1μg of anti-BrdU antibody (BD Pharmingen) pre-coupled with Dynabeads Protein G (Invitrogen) in IP genomic buffer (1X PBS + 0.0625% Triton X-100). The beads were washed 3 times with IP genomic buffer and once with TE (10 mM Tris-HCl pH8.0, 1 mM EDTA) and DNA was eluted with TE + 1% SDS and purified with High Pure PCR Cleanup Kit (Roche). Immunoprecipitated and input DNA were quantified by qPCR. The BrdU IP is shown as percentage of input or as a relative BrdU incorporation, which is percentage of input of the locus of interest divided by the percentage of input of the negative control (refereed as “ctrl” on Fig 1D) amplifying un-replicated region on chromosome V [7].
The neutral-neutral 2D agarose gel electrophoreses were performed according to [37] with minor modifications. Briefly, the total genomic DNA from asynchronous cell cultures or cultures released into S phase in the presence of 0.2M HU was isolated with Qiagen Genomic DNA Buffer Set and Genomic-tip. Genomic DNA was digested with NheI or BglII, then run in 1st dimension gels (1xTBE; 0.35% agarose) at 50V x 18 hrs. Lanes were excised and run on 2nd dimension gels (1xTBE; 0.9% agarose; 0.3 μg/mL ethidium bromide) at 175V for 13.5 hrs. The resolved DNA was transferred onto nylon membranes, UV cross-linked and hybridized with rDNA probes as described in [77]. The rDNA probe was PCR amplified from W303 yeast genomic DNA, gel-purified and radioactively labeled with Random Primed DNA Labeling Kit (Roche). The images were acquired with FX Personal Phosphorimager (Bio-Rad) and analyzed with Quantity One software. The intensity of the replication intermediates was normalized to the 1n spot and reported as ratio to corresponding replication intermediates in WT strain.
Undigested genomic DNA was resolved on 0.9% agarose gels, transferred onto nylon membranes and hybridized with rDNA probe.
Telomere southern blot to measure the length of the telomere terminal XhoI restriction fragments was performed essentially as described in [27].
Protein extraction by the TCA method and western blotting were performed essentially as described [27]. Rad53 protein was detected (as described in [78]) using Rad53-specific mouse monoclonal antibody raised against total Rad53 protein (Mab clone EL7), or against the active auto-phosphorylated state of Rad53 (Mab clone F9) [79] provided by A. Pellicioli (University of Milan).
Chromatin immunoprecipitation of TAP-tagged Fob1 and Sir2 [80] and quantitative PCR were performed as described [27] using Anti-TAP antibody (2 μl per IP, Thermo Scientific) for the immunoprecipitation. For the Sld3-13xMyc and Mcm4-13xMyc ChIP experiments, cells were arrested in G1 with alpha factor prior to the chromatin preparation. Antibody used: anti-Myc, 9E10 from culture supernatant.
PFGE was performed as previously described [81] using Bio-Rad DR II Contour-clamped Homogenous Electric Field (CHEF) apparatus. The running conditions were 68 hrs, 12°C, ramping from 300s to 900s switch time. DNA size standards (labeled M in Fig 2D) were purchased from Bio-Rad (H. wingei, 170–3667). The gel was stained with ethidium bromide, photographed under UV, transferred to a nylon membrane and hybridized with an rDNA probe.
rDNA instability was measured by the marker loss assay [39, 82]. Saturated yeast cultures where diluted and plated on complete YPD medium supplemented with 5 mg/ml adenine hemisulfate in order to obtain ca. 400 colonies/plate. Plates were sequentially incubated at 30°C (3 days), 4°C (2 days) and 25°C (1 day). The colonies were counted using ImageJ software Colony Counter plugin and the marker loss was calculated as the percentage of white colonies having red sectors. Completely red colonies, representing the progeny of the cells that lost the ADE2 marker, were excluded from the calculations.
The significance of the difference of the mean values obtained in BrdU IP-qPCR and rDNA instability assays was assessed with two-tailed paired Student’s t-test. The mean and standard error of the mean (SEM) are reported on the graphs.
|
10.1371/journal.ppat.1001066 | Direct Interaction between Two Viral Proteins, the Nonstructural Protein 2CATPase and the Capsid Protein VP3, Is Required for Enterovirus Morphogenesis | In spite of decades-long studies, the mechanism of morphogenesis of plus-stranded RNA viruses belonging to the genus Enterovirus of Picornaviridae, including poliovirus (PV), is not understood. Numerous attempts to identify an RNA encapsidation signal have failed. Genetic studies, however, have implicated a role of the non-structural protein 2CATPase in the formation of poliovirus particles. Here we report a novel mechanism in which protein-protein interaction is sufficient to explain the specificity in PV encapsidation. Making use of a novel “reporter virus”, we show that a quasi-infectious chimera consisting of the capsid precursor of C-cluster coxsackie virus 20 (C-CAV20) and the nonstructural proteins of the closely related PV translated and replicated its genome with wild type kinetics, whereas encapsidation was blocked. On blind passages, encapsidation of the chimera was rescued by a single mutation either in capsid protein VP3 of CAV20 or in 2CATPase of PV. Whereas each of the single-mutation variants expressed severe proliferation phenotypes, engineering both mutations into the chimera yielded a virus encapsidating with wild type kinetics. Biochemical analyses provided strong evidence for a direct interaction between 2CATPase and VP3 of PV and CAV20. Chimeras of other C-CAVs (CAV20/CAV21 or CAV18/CAV20) were blocked in encapsidation (no virus after blind passages) but could be rescued if the capsid and 2CATPase coding regions originated from the same virus. Our novel mechanism explains the specificity of encapsidation without apparent involvement of an RNA signal by considering that (i) genome replication is known to be stringently linked to translation, (ii) morphogenesis is known to be stringently linked to genome replication, (iii) newly synthesized 2CATPase is an essential component of the replication complex, and (iv) 2CATPase has specific affinity to capsid protein(s). These conditions lead to morphogenesis at the site where newly synthesized genomes emerge from the replication complex.
| Enteroviruses are single, plus-stranded RNA viruses that contain a large number of closely related pathogens. Human enteroviruses cause altogether >3 billion human infections per year, inflicting diseases ranging from benign (common cold) to very serious (poliomyelitis). Enterovirus replication has been studied for decades yet the mechanism of genome selection during encapsidation, a key step open for chemotherapeutic intervention, remains unknown. Attempts to identify a genomic “packaging signal”, instructing the genome to engage with capsid proteins in morphogenesis, have failed. We have used the similarities and dissimilarities of two closely related subspecies of enteroviruses, poliovirus and coxsackie A viruses (CAVs), and constructed chimeras in which the capsid coding region was interchanged. A chimera, with the CAV capsid domain and the poliovirus two non-structural domains of the polyprotein, synthesized its genome with wt kinetics yet was blocked in morphogenesis. Genetic and biochemical studies of chimeras led to the discovery that the non-structural protein 2CATPase (essential part of the replication complex) and capsid protein VP3 must directly interact for morphogenesis to proceed. This has led us to propose a novel mechanism by which the specificity of enterovirus morphogenesis is governed by protein-protein rather than RNA-protein interaction at the site of genome synthesis.
| Morphogenesis is a crucial step at the end of the virus' life cycle that provides newly synthesized genomes with a protective shell to survive in the extracellular environment yet assures attachment to and penetration into subsequent host cells. Morphogenesis of viral genomes must be specific because encapsidation of non-progeny nucleic acid is wasteful for the virus, for which reason elaborate mechanisms have evolved to discriminate against nucleic acids other than its own genome.
Here we describe our studies of the morphogenesis of a group of single, plus-stranded RNA viruses that belong to the genus Enterovirus of Picornaviridae, a family of viruses containing a large number of human and animal pathogens. Poliovirus (PV), the prototype enterovirus, has been extensively studied for a century and although much is known about its virion structure, uptake into host cell, genome structure and macromolecular events of replication, the mechanism of particle assembly is only poorly understood [1]. The key requirement of morphogenesis, namely the specific selection of viral genomes, has also remained obscure. We have discovered a novel mechanism for enteroviruses in which the specificity of encapsidation is facilitated by protein-protein interaction. It should be noted that this mechanism is different from the one used by some other RNA viruses such as hepatitis B virus and alphaviruses [2], [3]. The specificity of encapsidation with these viruses is dependent on an RNA encapsidation signal and RNA/protein interactions.
Enteroviruses synthesize only one protein, the polyprotein, which is cleaved by two virus-encoded proteinases, 2Apro and 3Cpro/3CDpro, into intermediates expressing specific functions (e.g. 3CDpro) and into mature proteins (Figure 1A). After its release from the polyprotein by 2Apro, the precursor of the structural proteins (P1) interacts with cellular chaperone Hsp90 [4], a requirement for its subsequent processing by 3CDpro into capsid proteins VP0, VP3 and VP1 (Fig. 1A) [5]. These cleavage products will spontaneously form a 5S protomer (VP0, VP3, VP1) that can oligomerize to the 14S pentamer (VP0, VP3, VP1)5; twelve pentamers, subsequently, assemble into a 75S empty capsid [(VP0, VP3, VP1)5]12, also called procapsid [6], [7]. It is not known at what stage progeny genomes interact with the capsid precursors. They may be inserted into the procapsid or, alternatively, pentamers may condense around RNA emerging from the replication complex. Either process will yield provirions {[(VP0, VP3, VP1)5]12RNA} [8], [9], [10] that mature to virions when VP0 is cleaved to VP4 and VP2 by a mechanism possibly involving an RNA-dependent autocatalytic process [6], [7].
The encapsidation process in PV morphogenesis is highly restricted to newly synthesized plus strand progeny RNA [11], [12]. Under normal conditions of replication in HeLa cells, cellular RNAs, PV mRNA lacking VPg or viral VPg-linked minus strand RNA are excluded from mature viral particles [6]. Numerous studies aimed at determining the specificity of encapsidation by searching for an RNA packaging signal have been unsuccessful. The very long 5′NTR of PV can be replaced with that of the distantly related coxsackie B3 virus (CVB3) [13] or CVB4 [14] yielding virions containing chimeric genomes that proliferate with PV wild type (wt) kinetics. Similarly, the cloverleaf of PV can be changed to that of HRV2 [15], or the PV IRES has been exchanged with IRESes from other picornaviruses [16], [17], [18] and even with that of HCV [19] without yielding impaired encapsidation phenotypes. The 3′NTR of PV, in turn, has been exchanged with that of HRV14, a single stem-loop structure with no apparent similarity in structure and sequence to that of PV. This chimera too proliferated with wt kinetics [20]. This makes it highly unlikely that the 5′- and 3′-NTRs of poliovirus contain packaging determinants. The genomic sequence encoding the capsid P1 precursor cannot harbor an encapsidation signal since the entire P1 encoding region can be deleted [21] or replaced by foreign genes [14], [22], [23]. Such PV replicons, all of which can replicate, can be efficiently encapsidated with PV capsid proteins in trans. Recent experiments from this laboratory, employing a “scrambled” sequence [24] of the P2 coding region (scramble of synonymous codons) have eliminated this region too from carrying an encapsidation signal (Song, Mueller, Ward, Skiena, Futcher, Paul and Wimmer, manuscript in preparation). Finally, genetic modification of the PV VPg coding sequence [25], [26] or engineering PVs carrying VPg sequences of other picornaviruses [25], [27], [28] have also eliminated the VPg coding sequence as providing an encapsidation signal. VPg, however, may still play a role in encapsidation (see below). Currently it seems unlikely that poliovirus or other enteroviruses (including the rhinoviruses that have recently been classified as enteroviruses) harbor an RNA signal that would instruct the capsid components to bind to and enclose the viral genome in a specific manner. Only one member of the extended family of Picornaviridae, Aichi virus (Kobuvirus genus), was reported to contain a 5′-terminal RNA stem loop with a role in particle assembly [29].
Among the nonstructural proteins of PV, 2CATPase and 3CDpro, have been reported to be involved in packaging although no mechanism(s) is known. Studies with an in vitro translation/RNA replication system, which produces viable viruses [11], have suggested that 3CDpro functions at a late step in the assembly process just before or during the maturation cleavage of VP0 to VP2 and VP4 [30]. Protein 2CATPase of PV has been implicated in virion capsid formation through genetic analysis of a cold-sensitive mutant [31] or by determining escape mutants from drug (hydantoin) inhibition [32]. The multifunctional 329 amino acids-long 2CATPase is the most complex and least understood nonstructural proteins of enteroviruses. The functions of this protein that are highly conserved among picornaviruses, include, in addition to encapsidation, host cell membrane rearrangements [33], [34], genome replication [35], [36] and even uncoating of viral particles [37]. Based on sequence analyses the protein has been classified as a member of the superfamily III helicases, which contain 3 conserved motifs (A–C), including two classical ATP binding motifs (A and B) (Fig. 1B) [38]. Purified 2CATPase possesses ATPase activity [39], [40], which is inhibited by guanidine HCl (GnHCl) [41], a known potent inhibitor of PV RNA replication [18]. In vitro the protein forms homo-oligomeric structures required for ATPase activity [42]. The N-terminal part of the protein contains a RNA binding domain and an amphipathic helix, which is involved in membrane binding and oligomerization [36], [42], [43]. Another amphipathic helix, a RNA binding domain and a cysteine rich domain that binds zinc are located near the C-terminus [44], [45]. In infected cells 2CATPase appears to be associated with viral RNA in the replication complexes on the surface of membranous vesicles [46].
Available evidence suggests that genome replication is a precondition of PV encapsidation [11], [12]. Electron-microscopic studies, which showed that RNA replication complexes co-localize with capsid precursors on membranous vesicles during infection [47], supported these observations. Nugent et al. [12] hypothesized that encapsidation specificity may be determined by the spatial arrangement of replication complexes with the capsid precursors. This intriguing hypothesis lacked an essential component: what brings the capsid precursors into the vicinity of the replication complexes since PV replicons lacking the P1 domain altogether can be efficiently encapsidated in trans?
Human enteroviruses have been divided into several clusters based on genotype relationships [48]. PV types 1–3, and eleven C-cluster coxsackie A virus serotypes share the C-cluster, also referred to as C-cluster human enteroviruses (C-HEVs). Their difference in affinity to cellular receptors, PVs using CD155 [49], [50] while C-CAVs using ICAM-1 [51], accounts for significant capsid dissimilarities between the member viruses of this species [52], [53]. In contrast, the differences between the non-structural proteins of PV vs C-CAVs are less pronounced. We have used the similarities and dissimilarities between PV and C-CAVs to separate RNA replication from encapsidation by constructing chimeric viruses in which capsid precursor P1 and/or 2CATPase have been exchanged. All of the chimeric viruses studied here replicated their genomes with wt kinetics in tissue culture cells but were blocked in encapsidation, a phenotype that we examined by genetic analyses. We present genetic evidence suggesting a specific interaction between 2CATPase and VP3, which is essential for genome encapsidation. The genetic evidence of the 2CATPase-VP3 interaction was further substantiated by biochemical assays. We propose that the primary determinant of encapsidation specificity in the enterovirus life cycle is protein-protein interaction.
HeLa H1 cells were maintained in DMEM (Life Technology), supplemented with 10% FCS, 100 units of penicillin, and 100 mg of streptomycin per milliliter.
The prototype strain of C-CAVs, CAV20, CAV21 (Kuykendall) and CAV18, propagated in HeLa H1 cells, were obtained from the American Type Culture Collection. Polioviruses (PVM) were derived from cDNA pT7PVM [54] by transfection.
Parental plasmids of PV: pT7PVM contain a full-length infectious cDNA of PVM.
Parental plasmids of C-CAVs: pT7CAV20 contains a full-length infectious cDNA of CAV20 [48]. pGEM-CAV21 and pT7CAV18, which contain a full length infectious cDNA of CAV21 and CAV18 (Kuykendall), respectively, were constructed by Elizabeth Rieder.
Chimeric genomes: Parental plasmids of PVs and C-CAVs were used as the backbone for cloning as described below. All plasmids contain the T7 promoter in front of the 5′ end of the full length genomic cDNA for in vitro RNA transcription by T7 RNA polymerase [54].
Parental plasmid cDNAs of CAV20, CAV21 and CAV18 were used as the backbone for cloning. Using a three-step overlapping PCR, chimeras between CAV20 and CAV21 (or CAV18) were generated by precise swapping of the genetic segment encoding the P1 region of the polyprotein [48]. The oligonucleotides and the templates for PCR are summarized in Supplementary Table 2 in Text S1. The overlapping PCR fragment and the vectors were digested with the same pair of restriction enzymes and ligated to produce the cDNA clone of the chimeric genome. The vectors and the restriction sites for construction of the chimeric genomes are listed in Supplementary Table 3 in Text S1. The DNA sequence of the final constructs was verified by sequencing analysis using BigDye Kit and ABI Prism DNA sequencer (model 310). Replacement of the original 2CATPase coding region in each of the chimeric genome with that of the same origin as the capsid coding region follows the same strategy described above (Supplementary Table 2 & 4 in Text S1).
Construction of the chimeric C20PP genome was described before [48]. For construction of the C20PP derivatives, listed in Supplementary Table 3 in Text S1 and Figure 4, a two-step overlapping PCR similar to the one described above was performed using C20PP as the template with the mutation(s) introduced into the internal primers (Supplementary Table 3 & 4 in Text S1).
To test the RNA replication efficiency of chimeric viruses, we used novel reporter viruses, which contain the Renilla luciferase gene fused to the N-terminal of the P1 coding region of the chimeric genomes. The same strategy (three-step overlapping PCR), described above, was used to introduce the Renilla luciferase gene into the N-terminal of P1 the coding region of the chimeric genomes (Supplementary Table 3 & 4 in Text S1). The luciferase protein is post-translationally cleaved from the remainder of the polyprotein by 3CDpro at a recombinant 3CDpro cleavage site.
Parental plasmid cDNAs pT7PVM, pT7CAV20, pGEM-CAV21, pT7CAV18 and the chimeric constructs were linearized at a unique restriction sites downstream the poly(A) tract (Supplementary Table 5 in Text S1) and used as templates for in vitro RNA synthesis using T7 RNA polymerase. RNA transcripts were transfected into HeLa H1 cell monolayers by the DEAE-Dextran method as described before [54]. Following transfection, virus was harvested from the transfected cells when 90–95% of the cells displayed cytopathic effect (CPE). Lysates of transfected cells from the chimeric genomes showing no CPE were inoculated into 35-mm-diameter HeLa H1 cell monolayers for 6–8 subsequent serial passages. The plaque phenotypes and virus titers (PFU/ml) of the parental and chimeric viruses were determined in triplicate by plaque assay [11] using 0.6% tragacanth gum. The identity of the chimeric viruses was confirmed by RT-PCR/sequencing analysis.
In vitro RNA translations were performed with HeLa cell S10 cytoplasmic extracts at 34 degree Celsius as described previously [11].
HeLa H1 cell monolayers (5×106) were infected with viruses that were purified from plaque assay. At 7-hr post infection, total cytoplasmic RNA was extracted with 1 ml Trizol reagent (Invitrogen) and amplified into DNA using Titan one tube RT-PCR system (Roche). The RT-PCR products were sequenced using the Bigdye Terminator Sequencing Kit (ABI, Applied Biosystems).
Dishes (35-mm diameter) of monolayered HeLa H1 cells were transfected with 5 µg of replicon RNAs and were incubated at 37 degree Celsius in standard tissue culture medium in the presence and absence of 2 mM GnHCl. Luciferase activities were determined in lysates of cells harvested 16 hrs after transfection. Cell lysates (10 µl) was mixed with 20 µl of luciferase assay reagent (Promega; luciferase assay system catalog no. E2810) and Renilla luciferase activity was measured in an Optocomp I luminometer (MGM Instruments, Inc.). Cell lysates from transfections were used to re-infect HeLa H1 cells in the presence and absence of 2 mM GnHCl and luciferase activities were determined in lysates of cells harvested 8 hrs after infection. Luciferase activity ratio (−GnHCL/+GnHCl) represents: luciferase activity without GnHCl divided by luciferase activity with GnHCl in either transfection or infection.
A PCR fragment containing full length PV VP3 was amplified and cloned into the pET21b vector (Novagen) with the restriction enzymes Sac I and Xho I.
GST-tagged 2CATPase and His-tagged VP3 recombinant proteins were expressed in E. coli. The GST-2CATPase proteins were expressed from pGEX-2C vector and purified by glutathione sepharose column (GE Healthcare) as described before [41]. The His-VP3 proteins were purified by nickel column chromatography (QIAGEN).
Briefly, 5 µg GST-2CATPase (or 2 µg GST as a control) were incubated with glutathione sepharose beads at 4°C for 3 hr in buffer containing 50 mM Tris-HCl pH7.5, 140 mM NaCl, 0.1% TritonX-100 with protease inhibitor cocktail tablets (Roche). The protein bound GSH beads were washed with PBS 3 times and then 5 µg His-VP3 was added. After 1 hr incubation at 4 degree Celsium, the glutathione beads were washed 3 times and were boiled in 1x SDS sample buffer for 5 min. The samples were analyzed by SDS-polyacrylamide gel (12.5% acrylamide) electrophoresis and followed by western blot analysis using antibodies against PV VP3 (polyclonal, kindly contributed by Dr. Delpeyroux, Pasteur Institute, France).
Plasmids used for in vitro translation of CAV20 structural protein VP3 (wt), VP3 (E180G) and CAV20 non-structural protein 2CATPase (wt), PV non-structural proteins 2CATPase (wt) and 2CATPase (N252S) were generated with A2 plasmid [55] and PCR fragments encoding wt and mutant proteins of VP3 and 2CATPase according to methods described previously [55]. 2 µg of each VP3 and 2C RNA transcripts generated in vitro by T7 RNA polymerase were co-translated in HeLa extract and labeled by 35S-labeled methionine. Using anti-PV 2C polyclonal antibody, Co-IP assay was performed with co-translated 35S labeled 2CATPase and VP3 proteins following standard protocols using protein A/G plus-agarose (Santa Cruz Biotechnology) and [56]. The radioactive signals from input proteins and Co-IP reaction products were quantified by a PhosphorImager (Molecular Dynamics, Storm 860) by measuring the amount of 35S incorporated into product. Interactions between 2CATPase and VP3 were represented by percentages of the levels observed in the Co-IP reaction with CAV 2CATPase and CAV VP3 after normalizing the amount of input 2CATPase and VP3. Numbers given for the extents of interactions represented the average of three independent experiments.
The lack of evidence for an RNA packaging signal in enterovirus proliferation has prompted us to study the specificity of encapsidation by searching for possible protein-protein interactions needed for this process. Previous studies have shown that chimeric constructs of the PV polyprotein with exchanges of varying coding regions of closely related picornaviruses can be utilized to analyze determinants of viral macromolecular interactions and replication [57], [58], [59], [60], [61], [62]. PV and C-CAVs share a high degree of amino acid identity in their nonstructural proteins but are not closely related in their capsid sequences probably because they evolved to use different cellular receptors [48]. In a chimera with the capsid of one C-HEV and the nonstructural proteins of another C-HEV, this difference may produce specific morphogenesis phenotypes due to incompatibility or poor interaction between capsid and nonstructural proteins. In our previous work, we have already observed that the replacement of the PV type 1 Mahoney (PVM) capsid with that of its closest relative, CAV20, resulted in a quasi-infectious virus, C20PP (Figure 2A) [48]. In C20PP, the first letter refers to the origin of the P1 region, the 2nd and 3rd letters refer to the origins of the P2 and P3 regions, respectively. The quasi-infectious phenotype means that a step in the life cycle of C20PP is so severely debilitated that only escape variants can be recovered from transfections with RNA transcripts [48]. The molecular basis of the defective phenotype of C20PP was not elucidated but could be proteolytic processing of the capsid precursor, genome replication or encapsidation. The observed quasi-infectious phenotype of C20PP made it possible to subject this chimera to a genetic analysis that may reveal a defect in encapsidation.
We first provided evidence that the quasi-infectious phenotype of C20PP was not due to abnormal translation or protein processing resulting from poor compatibility between the heterologous capsid and the 3CDpro polypeptide in C20PP. Translation of RNA transcripts of wt and chimeric constructs in a HeLa cell-free extract [11] showed normal translation and protein processing patterns (Figure 2B).
The search for the block of C20PP proliferation led us to develop a novel reporter virus in which the PV open reading frame (ORF) of the Renilla Luciferase (R-Luc) protein was fused to the N-terminus of the viral polyprotein. In the course of the infection the R-Luc is cleaved off from the viral polyprotein at an engineered 3CDpro proteinase cleavage site. Due to the small size of the inserted R-Luc gene this virus was stable for 1 passage after transfection and, thus, suitable for our experiments. This construct is similar to a previously described recombinant coxsackie B3 virus that stably expressed eGFP in tissue culture [63]. The advantage of using our reporter virus over conventional reporter replicons, in which the P1-coding sequence is replaced by the luciferase gene, is that it can distinguish between a defect in replication and encapsidation. A reporter viral genome (with the R-Luc sequence) that is unable to encapsidate itself will exhibit normal RNA replication levels as evidenced by a wt-like Renilla luciferase signal after RNA transfection. However, it would not generate infectious progeny and, consequently, passage to fresh cells will fail, leading to the loss of the luciferase signal.
We have made such reporter viruses from both the parental wt CAV20 and the chimeric C20PP (Figure 2C). RNA transcripts were transfected into HeLa H1 cells both in the absence and presence of 2 mM GnHCl, a potent inhibitor of PV [18] and CAV20 RNA replication [18], [48]. Luciferase activity was determined at 16-hr post transfection either in the presence of GnHCl (+GnHCl) throughout the incubation period that allows us to measure the translation of the transfecting RNA, or in the absence of GnHCl (−GnHCl) when the luciferase signal is increased because of RNA synthesis. The ratio of the luciferase signals –GnHCl/+GnHCl indicates the extent of genome replication. As shown in Figure 2D there was a 100 fold increase of the luciferase signal at −GnHCl compared to +GnHCl with both wt R-Luc-CAV20 and R-Luc-C20PP viruses, an observation indicating robust RNA synthesis under these conditions. Lysates of transfected cells were then inoculated to fresh HeLa H1 cells as 1st passage either in the presence or absence of GnHCl. Eight hours post infection the luciferase activity was measured (Figure 2D). The high level of luciferase activity obtained after the first passage of the R-Luc-CAV20 virus indicated the formation of virions that had encapsidated the wt genome in the course of transfection. In contrast, no luciferase signal could be detected after passage of the lysate harboring the R-Luc-C20PP chimera (Figure 2D). We conclude that the genome of the C20PP chimera, although competent in RNA replication, cannot form infectious progeny, e.g. it is defective in genome encapsidation. The reason for the defect in encapsidation, however, remains elusive.
As mentioned above, the C20PP chimera is quasi-infectious. Variants that escaped the block in encapsidation were found only after three blind passages following transfection. This indicated the emergence of mutation(s). To identify the rescuing mutation(s), we plaque purified two viruses that had emerged after three passages from two independent transfections with C20PP transcripts. RT-PCR and sequence analyses of the viral genomes revealed two independent single mutations that mapped to the coding region of either VP3 (E180G) or 2CATPase (N252S). By separately engineering these two mutations back into the cDNA of C20PP we obtained two viable viruses C20PP-VP3E180G and C20PP-2CN252S (Figure 3A). Their titers after transfection were as low as those observed with isolates after three passage of C20PP (Figure 2A) and 1000 fold lower than that observed with wt CAV20. Moreover, C20PP-VP3E180G and C20PP-2CN252S expressed a small plaque phenotype compared to that of wt CAV20 (Figures 2A and 3A).
The phenotypes of C20PP-VP3E180G and C20PP-2CN252S did not change after further passages, e.g. all attempts to isolate variants with improved proliferation phenotypes failed (data not shown). This prompted us to engineer both the VP3E180G and 2CN252S mutations into C20PP (C20PP-DM). Variant C20PP-DM expressed phenotypes (virus titer and plaque size) almost the same as that of wt CAV20 (compare Figures 2A and 3A). Currently, we cannot explain why in our experiments the C20PP-DM-like variant did not evolve during passaging of C20PP.
To confirm the rescue of the encapsidation defect of C20PP by the mutations in VP3 and 2CATPase, we constructed reporter viruses R-Luc-C20PP-VP3E180G, R-Luc-C20PP-2CN252S and R-Luc-C20PP-DM (Figure 3B). All three produced strong luciferase signals after transfection of their genomic RNAs into HeLa H1 cells, as expected (Figure 3C). After a passage into fresh HeLa H1 cells, the luciferase activity was highly impaired with C20PP but was found to be partially rescued if the virions carried the single 2CN252S or VP3E180G mutation or fully rescued by the double mutation (Figure 3C). Thus, although a single mutation can partially rescue the proliferation phenotype of C20PP, the double mutation 2CN252S/VP3E180G in the genome of the chimera is capable of producing a proliferation phenotype similar to that of wt CAV20. This observation, confirms the previous genetic data by Vance et al. [32] that the coding region of 2CATPase is linked to encapsidation. More importantly, the cooperative activity of VP3 with 2CATPase suggests that capsid protein VP3 functions through a direct interaction with 2CATPase in encapsidation. As indicated before, the coding sequence of 2CATPase, however, can be eliminated as carrying an encapsidation signal. The cre, the only known essential RNA structure in coding sequence of 2CATPase, are highly homologous in sequence between PV and CAV20. Moreover, scrambling of the P2 RNA sequence has no influence on PV proliferation if the essential cre is transplanted to the 5′NTR (Song, Mueller, Ward, Skiena, Futcher, Paul, and Wimmer, manuscript in preparation). It is, thus, likely that VP3 and 2CATPase or their precursors cooperate by protein-protein interaction, a novel mechanism for specific genome encapsidation of enteroviruses.
It should be noted that the mutation in 2CATPase, which rescues in part the encapsidation of C20PP-2CN252S, is an N/S change in PV 2CATPase at position 252 (Figure 4A). Amino acid sequence alignment demonstrated that CAV20 has a Gly at this position (Figure 4A), an observation indicating that a CAV-20 like, uncharged residue might be favorable at this position. Thus, since C20PP-2CN252S expressed a severe encapsidation phenotype, it was of interest to determine whether a C20PP-2CN252G mutant would yield a chimera equal or superior in encapsidation to C20PP-2CN252S. Similar to C20PP-2CN252S, the N252G mutation in the PV 2CATPase only partially rescued the encapsidation phenotype of C20PP (Figure 4B). The observed N/S mutation in a naturally selected escape mutant can be explained by the reasoning that the CAV20-like N/G change in codon N252 would have required two nucleotide changes (AAT/GGT) while the N/S mutation entails only a single nucleotide change (AAT/AGT). Overall, the CAV20 like single amino acid change at 252 of PV 2CATPase was favorable but not sufficient to fully rescue the VP3-related function of the PV 2CATPase protein in encapsidation.
Based on this observation, we reasoned that the replacement of the entire 2CATPase coding region in C20PP with its CAV20 counterpart should yield a chimera whose protein/protein interaction required for encapsidation would be sufficient and, thus, yield a virus with proliferation phenotypes similar to that of wt CAV20. Therefore, we generated a chimera designated as C20P(C202C)P that showed CPE after transfection with a virus titer comparable to that of wt CAV20 (compare Figures 2A and 4B). These results provide further support for the hypothesis that the 2CATPase protein is a partner required for encapsidation. It should be noted that the chimera C20P(C202C)P had a growth phenotype more similar to that of the wt CAV20 virus than the chimera C20PP-2CN252S (compare Figures 2A and 4B), an observation indicating that sequences besides residue 252 in 2CATPase are also important for function during viral encapsidation.
The observation that CAV20 VP3 needs its own 2CATPase to fully rescue the defect in encapsidation suggests that the cooperation or interaction between 2CATPase and capsid might be specific and generally true for C-HEVs. To test this hypothesis we extended our analyses to CAV18 and CAV21, two viruses that are phylogenetically related to CAV20 and PV [48]. Chimeras C20C21C21 and C18C20C20 (Figure 5A) displayed non-viable phenotypes as judged by the lack of virus in plaque assays even after 8 blind passages on fresh HeLa H1 monolayers (Figure 5A). In order to test whether the lethal phenotypes of the two chimeric viruses were due to the same encapsidation defect as that of C20PP, we constructed reporter viruses of CAV18 and of the chimeras C20C21C21 and C18C20C20 (Figure 5B) just as that of CAV20 (Figure 2C). After transfection of RNA transcripts into HeLa H1 cells, R-Luc activity at 16-hr post transfection showed that the parental and chimeric viruses replicated their genomes with nearly wt efficiency (Figure 5C). However, after the first passage on fresh HeLa H1 cells only wt CAV20 and wt CAV18 reporter viruses yielded normal luciferase signals (Figure 5C). The chimeric genomes R-Luc-C20C21C21 and R-Luc-C18C20C20 could not produce an infection, a result demonstrating an encapsidation defect.
To test whether the lethal proliferation phenotypes of the two chimeras (C20C21C21 and C18C20C20) are related to an incompatibility between the capsid and the 2CATPase proteins derived from different parental viruses, we constructed new chimeras [C20C21(C202C)C21, C18C20(C182C)C20] in which the capsid and 2CATPase proteins were derived from the same origin (Figure 6A). The resulting chimeras all showed CPE after transfection (Figure 6B) and the virus titers were comparable to that of the parental viruses (data not shown). The finding that the lethal growth phenotypes of the chimeras were fully rescued when the capsid and 2CATPase were derived from the same origin serves as further support of our hypothesis that 2CATPase and capsid proteins communicate with each other during the process of encapsidation. So far, we have not been able to determine, by continued passage, the necessary amino acid changes in the 2CATPase protein of CAV20 and CAV21 to allow encapsidation of the C20C21C21 and C18C20C20 chimeras. This observation might be explained by the fact that there are many amino acid differences either between CAV20 and CAV21 or between CAV20 and CAV18 flanking residue 252 of the 2CATPase protein (Figure 6C). This may make it difficult for the two chimeras (C20C21C21 and C18C20C20) to generate escape mutants simply by natural selection during passages.
The genetic evidence described above strongly suggested a direct interaction between the capsid proteins and 2CATPase, which is required for encapsidation. To confirm this interaction, we carried out a GST-pull down assay with purified PV proteins GST-2CATPase and His-VP3 (Figure 7A, lane 2). The same assay was performed with purified GST protein as a control (Figure 7A, lane 1). Our results clearly showed that PV GST-2CATPase interacts directly with the PV His-tagged VP3 protein. To provide further proof that direct interaction between VP3 and 2CATPas is required for the encapsidation process, co-immunoprecipitation (Co-IP) assays were performed with VP3 and 2CATPase proteins in three different combinations: CAV20 VP3 & CAV20 2CATPase, CAV20 VP3 & PV 2CATPase, CAV20 VP3 (E180G) & PV 2CATPase (N252S), which correspond to those observed in wt CAV20, nonviable chimera C20PP, and rescued C20PP-DM, respectively. In vitro transcribed RNA transcripts of 2C and VP3 coding sequences in different combinations were co-translated in HeLa cell extracts (Figure 7B, lanes 4–6) [11], [55]. Using PV 2C polyclonal antibody, which recognizes PV 2CATPase and CAV20 2CATPase with the same efficiency (data not shown) [56], CAV20 VP3 was co-immunoprecipitated readily by CAV20 2CATPase (100%, Figure 7B, lane 1) but only weakly by PV 2CATPase (32%, Figure 7B, lane 2). The extent of interaction between 2CATPase and VP3 was quantified using a PhosphorImager and is expressed as percentage of the level observed in the CAV20 2CATPase and CAV20 VP3 Co-IP reaction. These results indicated a strong, direct interaction between 2CATPase and VP3 of the same origin (CAV20) but not between PV 2CATPase and CAV20 VP3, a combination that yielded the nonviable C20PP chimera. In contrast, the interaction between PV 2CATPase (N252S) and CAV20 VP3 (E180G) proteins was restored (78%, Figure 7B, lane 3) when the two mutations were incorporated into PV 2CATPase and CAV20 VP3, respectively, This observation, which correlates with the rescue of the nonviable phenotype of C20PP by the two mutations, strongly support the notion that sufficient protein-protein interaction between 2CATPase and VP3 is essential for the encapsidation process. The same Co-IP assays were also performed with α-actin antibody and empty resin as controls to ensure that the interactions were not due to non-specific binding of the proteins to the antibody or to the resin (data not shown). It should be noted that there was an extra protein band shown below the band of 2CATPase in each of the input lanes (Figure 7B, lane 4–6, indicated by asterisk), which was possibly generated from the internal initiation or premature termination of the translation of RNA transcripts of 2C coding sequences. Apparently these incomplete translation products could not be recognized by the 2C antibody since protein bands disappeared after Co-IP (Figure 7B, lane 1–3). These results confirm the specificity of the Co-IP assay in which the detection of the VP3 protein was due to its co-immunoprecipatation with the 2CATPase protein, recognized by anti-2CATPase antibody.
The data support our hypothesis that 2CATPase is required for viral encapsidation through a direct interaction with capsid protein VP3 and also confirm that 2CATPase interacts with VP3 through protein–protein rather than RNA-protein interaction. Given that 2CATPase and capsid proteins are colocalized on the surface of membranous vesicles in the RNA replication complex [46], [47], it is likely that 2CATPase interacts with VP3 either in the form of the mature protein or in the context of one of the VP3-containing capsid precursors (5S, 14S, 75S and 150S).
The mechanism of picornavirus genome encapsidation has been a conundrum for many years. In previous studies on poliovirus morphogenesis, mostly trans encapsidation experiments were performed to determine the specificity of PV morphogenesis. In trans encapsidation experiments, the capsid proteins are offered from a different molecular entity to the parental genome either by coinfecting picornaviruses [14], by the vaccinia system [22], [23] or by expression from co-transfected cDNAs [64]. However, differences in the experimental design also affect the outcome of the experiments. Jia et al., [64] have reported trans-encapsidation of PV replicons into capsids of coxsackie B3, human rhinovirus 14, and coxsackie B24 viruses. These data, however, are in contrast to those of Porter et al., [23] and Barclay et al., [14], who failed to trans encapsidate the reporter PV replicon by super-infecting with CAV21.
In the current study, we have used a novel system to study encapsidation of closely related enteroviruses, the C-cluster enteroviruses (C-HEVs). They consist of two classes of virus serotypes: PV and C-CAVs. We have made use of differences between these viruses to investigate the effect of capsid exchanges on C-HEV morphogenesis. By studying a variety of C-HEV chimeric viruses in which the capsid P1 precursors were exchanged, we now present direct evidence for the involvement of 2CATPase in enterovirus morphogenesis via direct interaction with capsid protein VP3. Different from trans encapsidation assays, in which the capsid proteins are offered from a different molecular entity to the parental genome, our experiments were designed to measure cis encapsidation of genomes, in which the capsid is provided (generated) by the chimeric genome itself. It appears that in these chimeras the non-structural proteins are more discriminatory to capsid proteins because they are required to proteolytically process the heterologous capsid precursor. In addition, the quality and quantity of the heterologous capsid proteins produced might not be ideal for the chimeric genome. As we have shown here, favorable conditions for encapsidation are rarely met in the chimera, even if the viruses are as closely related as PV and CAV20.
In a previous study on genetic recombination between PV and C-CAVs, we observed that capsid chimera C20PP initially could not grow [48], an observation indicating that it harbors defect(s) debilitating the viral replication life cycle. Translation in HeLa cell extracts showed, however, that both translation and proteolytic processing of the CAV20 capsid precursor in C20PP was unimpaired. This is not surprising since the 3CDpro proteins of poliovirus and CAV20 are closely related in amino acid sequence [48] and because the 3CDpro cleavage sites within the P1 precursor of PVM and CAV20/21 are well conserved (Supplementary Table 1 in Text S1). These two facts reduce the likelihood of a processing defect of the foreign capsid precursor prior to packaging. In contrast, a chimera consisting of the PV capsid and the coxsackie virus B3 (CBV3) nonstructural domains was not only dead but it also revealed a processing defect of the PV capsid precursor [65]. CBV3 is an enterovirus belonging to the B-cluster, and its genetic kinship with PV is much more distant than that between PV and CAV20. Thus, CBV3 3CDpro proteinase was apparently unable to properly cleave the PV capsid precursor. Similar results were obtained with a chimera of PV in which the 3C-coding region was derived from HRV14. The foreign proteinase was not capable of recognizing the PV-specific processing sites within the capsid precursor [57].
The robust RNA replication phenotype of C20PP demonstrated here with the use of a new reporter virus construct suggested to us that this chimera might be quasi-infectious with respect to encapsidation. This was proven to be correct since viable viruses were found upon blind passages of C20PP but they harbored mutations. The virus isolates from two independent transfections contained a mutation either in 2CATPase (N252S) or in VP3 (E180G). Either of these single mutations was able to partially rescue the defective encapsidation and growth phenotype of C20PP. Introducing both mutations together into the C20PP genome fully rescued packaging and resulted in normal production of progeny virus. It is noteworthy that we never found both mutations in a single isolate even after eight passages. Perhaps the mutations conferred to the variants too little of an advantage to be selected under the conditions of the experiments. As discussed earlier, an involvement of essential RNA sequences in the 2CATPase and VP3 coding sequences during the process of encapsidation can be excluded. A direct interaction between the 2CATPase and VP3 proteins, suggested by the genetic experiments, was confirmed by biochemical assays using either purified or in vitro translated PV and CAV20 proteins. It should be noted that 2CATPase also functions in the viral life cycle in the form of its precursor 2BCATPase. A requirement of an interaction between 2BCATPase and VP3 for packaging is, however, unlikely since 2B is less homologous between PV and CAV20 in sequence than 2CATPase and the exchange of the mature 2CATPase protein alone is sufficient for full rescue. We are currently investigating whether the interaction between 2CATPase and VP3 involves the mature VP3 polypeptide or one of the capsid intermediates during viral assembly and/or maturation.
The encapsidation defect of C20PP could also be rescued by replacing the entire 2CATPase coding sequence of PV with that of CAV20. Additional experiments indicated that the lethal growth phenotypes of other CAV/CAV capsid chimeras could also be reversed by replacing their 2CATPase coding sequence with that of the capsid donor virus. It is noteworthy that these observations are not contradictory to the scenario of another chimera previously described. PC20C20, which, in contrast to C20PP, possesses a chimeric genome encoding PV capsid and CAV20 nonstructural protein sequences, grows as well as wt PV [48]. We have previously proposed that the evolutionary direction is from C-CAV to PV within C-HEVs resulting in a receptor switch from ICAM-1 to CD155 during the speciation [48]. If so, the newly emerged PV capsid may still be compatible with 2CATPase of the C-CAV ancestors and achieve sufficient interaction required for encapsidation.
Our data also clarify some unanswered questions about previous trans encapsidation experiments using poliovirus replicons containing a reporter gene in the capsid-coding region [14], [23]. Those studies showed that CAV21 was not able to encapsidate a PV replicon even though co-infection of cells with CAV21 resulted in high levels of replication of both CAV21 and the PV replicon. The most likely reason for those results is that the CAV21 capsid, similar to CAV20 capsid, fails to properly interact with PV 2CATPase. From their studies with hydantoin, Vance et al., have suggested that 2CATPase might have a role in encapsidation by an association of the progeny RNA with the capsid [32]. In other experiments with the same drug, however, Oh et al., [66] have recently proposed that hydantoin inhibits the release of the progeny RNA from the replication complex prior to encapsidation. Whether the interaction between 2CATPase and VP3, as we have observed in our studies, is required during or just before the union of the RNA with capsid proteins remains to be determined.
An intriguing phenomenon of poliovirus encapsidation is that only newly replicated RNA molecules are incorporated into virions [11], [12], an observation reported also for some other RNA viruses such as flock house virus [67], [68], [69], and brome mosaic virus [70]. Coupling encapsidation specifically with replication offers an efficient mechanism of discriminating against cellular RNAs or viral mRNA. Since genome replication is coupled with translation [71], [72] the link to encapsidation “can impose a form of late proofreading” for the progeny virus [12]. In Figure 8, we present a model of morphogenesis that is based, admittedly, on much speculation. We currently propose that in the context of the membrane-associated replication complex 2CATPase will directly interact with a 14S pentamer via VP3. The pentamer will then bind the newly emerging, VPg-linked genomic RNA [9] while the assembly of the virion proceeds in close contact with the membranous environment. This model offers a new mechanism for the specificity of enterovirus encapsidation: it is dependent on protein/protein interactions at the site of the active replication complex.
It is likely that our model will be applicable to most, if not all, picornaviruses as well as to other families of plus strand RNA viruses. For example, the requirement for an interaction between a capsid protein and nonstructural proteins for encapsidation has also been observed for members of the Flaviviridae. Murray et al., reported that several assembly-deficient core mutants of HCV genotype 2a (Hepacivirus genus) could be rescued by compensatory mutations in p7 or NS2 [73]. In addition, it was shown that HCV core and NS5A colocalize on the surface of lipid droplets, a process required for particle assembly [74]. Furthermore, recent reports indicate the importance of nonstructural proteins in the maturation of Kunjin virus and Yellow fever virus (Flavivirus genus) [75], [76] and of bovine diarrhea virus (Pestivirus genus) [77].
We have noted before that Aichi virus, a member of the Kobuvirus genus in the Picornaviridae, requires a 5′-terminal RNA element for encapsidation [29]. This signal by itself, however, is not sufficient to confer encapsidation specificity since it can be replaced by a similar stem loop from hepatitis A virus, a member of the genus Hepatovirus of Picornaviridae, and the resulting chimera expresses a severe proliferation phenotype [27]. On the other hand different genera of Picornaviridae may have evolved different strategies of encapsidation. This is not entirely unlikely considering the fundamentally different strategies that different genera of Flaviviridae are using to control genome translation (cap-dependent vs. IRES-dependent initiation of translation [78].
Our model does not explain, however, how VPg-linked minus-stranded poliovirus RNA is discriminated against in encapsidation. Previous studies have reported that several picornaviral 2CATPase proteins bind specifically to the 3′-end of minus strand RNA in vitro [79], [80]. However, the importance of this interaction, if any, for encapsidation is unlikely. Normally, plus strands are produced in great excess over minus strands [81], a phenomenon thought to lead to the depletion of free minus strands by forming the replicative form (RF) or replication intermediates (RI) [18]. If the balance between the plus and the minus strands is disturbed, free minus strands may emerge from the replication complex and they may then be encapsidated. Indeed, encapsidation of both plus and minus strand genomic RNAs was observed in non-cytopathogenic CBV3 that were isolated from persistently infected murine hearts and cardiac myocyte cultures [82]. It is possible that the non-cytopathogenic CBV3 produces minus stranded RNA in excess such that it will emerge from replication complexes where capsid precursors are waiting to encapsidate them. Whether the “cap” of a positively charged VPg on both plus or minus strands plays a role in this process remains to be seen.
|
10.1371/journal.pgen.1007326 | Functional diversification accompanies gene family expansion of MED2 homologs in Candida albicans | Gene duplication facilitates functional diversification and provides greater phenotypic flexibility to an organism. Expanded gene families arise through repeated gene duplication but the extent of functional divergence that accompanies each paralogous gene is generally unexplored because of the difficulty in isolating the effects of single family members. The telomere-associated (TLO) gene family is a remarkable example of gene family expansion, with 14 members in the more pathogenic Candida albicans relative to two TLO genes in the closely-related species C. dubliniensis. TLO genes encode interchangeable Med2 subunits of the major transcriptional regulatory complex Mediator. To identify biological functions associated with each C. albicans TLO, expression of individual family members was regulated using a Tet-ON system and the strains were assessed across a range of phenotypes involved in growth and virulence traits. All TLOs affected multiple phenotypes and a single phenotype was often affected by multiple TLOs, including simple phenotypes such as cell aggregation and complex phenotypes such as virulence in a Galleria mellonella model of infection. No phenotype was regulated by all TLOs, suggesting neofunctionalization or subfunctionalization of ancestral properties among different family members. Importantly, regulation of three phenotypes could be mapped to individual polymorphic sites among the TLO genes, including an indel correlated with two phenotypes, growth in sucrose and macrophage killing. Different selective pressures have operated on the TLO sequence, with the 5’ conserved Med2 domain experiencing purifying selection and the gene/clade-specific 3’ end undergoing extensive positive selection that may contribute to the impact of individual TLOs on phenotypic variability. Therefore, expansion of the TLO gene family has conferred unique regulatory properties to each paralog such that it influences a range of phenotypes. We posit that the genetic diversity associated with this expansion contributed to C. albicans success as a commensal and opportunistic pathogen.
| Gene duplication is a rapid mechanism to generate additional sequences for natural selection to act upon and confer greater organismal fitness. If additional copies of the gene are beneficial, this process may be repeated to produce an expanded gene family containing many copies of related sequences. Following duplication, individual gene family members may retain functions of the ancestral gene or acquire new functions through mutation. How functional diversification accompanies expansion into large gene families remains largely unexplored due to the difficulty in assessing individual genes in the presence of the remaining family members. Here, we addressed this question using an inducible promoter to regulate expression of individual genes of the TLO gene family in the commensal yeast and opportunistic pathogen Candida albicans, which encode components of a major transcriptional regulator. Induced expression of individual TLOs affected a wide range of phenotypes such that significant functional overlap occurred among TLO genes and most phenotypes were affected by more than one TLO. Induced expression of individual TLOs did not produce massive phenotypic effects in most cases, suggesting that functional overlap among TLO genes may buffer new mutations that arise. Specific sequence variants among the TLO genes correlated with certain phenotypes and these sequence variants did not necessarily correlate with sequence similarity across the entire gene. Therefore, individual TLO family members evolved specific functional roles following duplication that likely reflect a combination of inherited function and new mutation.
| Changes in gene copy number provide a rapid mechanism of adaptation to new or different environments by utilizing available functional sequences to cope with altered conditions. Gene duplication commonly arises through errors in DNA replication or sister chromatid recombination to produce a second identical gene copy [1–3]. The presence of functionally redundant genes loosens evolutionary constraints on the two paralogs and allows them to mutate through genetic drift [4]. As this process is repeated, gene duplication can lead to gene family expansion, which provides significant evolutionary fodder on which selection can act to promote adaptation.
Following gene duplication, the replicated sequence can either be lost or retained to serve a redundant or new function in the organism. In most cases, one of the paralogs is inactivated by deleterious mutations, thereby restricting further evolution of the other gene duplicate [5, 6]. However, if a mutation in a duplicated gene provides a selective advantage, both paralogs may be retained as they contribute separately to fitness of the organism [7–9]. Accumulated polymorphisms between gene duplicates can lead to subfunctionalization in which each gene performs a separate function that previously existed within the ancestral gene or neofunctionalization where one of the paralogs evolves a novel function and the other retains the ancestral function. Most studies of gene duplication and divergence rely on comparison of two paralogs to assess the selective pressures that operated following gene duplication because it provides a more simplified context for analysis [5, 10–13]. Such copy number variants may have arisen through small scale or whole genome duplication [14–17]. Although the evolutionary outcomes of gene duplication resulting from whole genome duplication have been studied extensively [18–22], small scale duplications are much more common, with copy number variation in some genes occurring at rates up to 1.7x10-4 duplications per cell division, far exceeding the basal point mutation rate [23]. The evolutionary fate of genes following small-scale duplication is driven largely by genomic context [24–26], gene dosage and protein complex formation [27–29], as well as by gene expression level [28, 30]. Yet, the evolutionary trajectories of gene families that encode many paralogous sequences remain largely unexplored.
Subtelomeres, or telomere-associated sequences, are genomic regions of linear chromosomes that separate the telomeric repeats from chromosome-specific sequences. These regions typically harbor a mixture of duplicated genes and repetitive sequences that often resemble fragments of mobile genetic elements [31, 32]. Subtelomeric regions evolve rapidly and are characterized by extensive genetic turnover due, in part, to the presence of these repetitive sequences [33, 34]. Frequent recombination, elevated mutation rates via acquisition of single nucleotide polymorphisms (SNPs) and insertions/deletions (indels), and the constant processes of gene duplication and gene disruption contribute to the rapid evolution of subtelomeric regions [25, 35–37]. Consequently, subtelomeres are often the most dynamic regions of the genome [25, 38, 39], with profound changes detectable over time scales readily achieved via experimental evolution studies [36].
Expanded gene families commonly reside within subtelomeric regions and are characterized by extensive copy number variation and a rapid accumulation of mutations that can alter their expression, structure, or function [40]. As a result, gene families that reside within the subtelomeres are typically under strong selection and are associated with species-specific lifestyles that promote organismal success [40–43]. For example, the MAL, MEL, and SUC genes in S. cerevisiae allow cells to utilize different carbon sources (maltose, melibiose, and sucrose, respectively), and fluctuate in copy number depending on the available growth substrate [40, 44, 45]. In this way, the subtelomeric genes contribute to phenotypic plasticity and rapid adaptation to nutrient availability across diverse environments.
The Candida clade of species includes mammalian commensals that are closely related to other Saccharomycotina but did not undergo a whole genome duplication event [46, 47]. Of these, C. albicans is the most clinically prevalent species for humans because it is a common commensal also capable of causing debilitating mucosal infections as well as life-threatening systemic infections [48, 49]. The success of C. albicans is due, in part, to its ability to occupy and persist in a range of commensal host niches including the gastrointestinal tract (pH 7.4–8, 37–40°C), the oral cavity (pH 6.3–7.4, 33–35°C), and the anaerobic colon [50, 51]. The organism often breaches these mucosal niches and becomes bloodstream-borne, especially in hosts with compromised immunity. Progression of disease is dependent upon host immunity as well as a battery of fungal virulence attributes including the ability to transition between different cell morphologies, to resist stresses within the host including oxidative and cell wall damage, and to evade immune system components [52–57].
The expansion of several gene families involved in virulence traits distinguishes C. albicans from other Candida clade species, and thus may have a role in elevated C. albicans virulence. Expansion of the ALS, SAP, and LIP gene families in C. albicans increases the functional capacity of adhesins, proteases, and lipases, respectively, which have known roles in pathogenesis [58–60]. The most dramatic gene expansion occurred within the telomere-associated (TLO) gene family, which has fourteen copies in C. albicans, two copies in the most closely-related C. dubliniensis species, and a single copy within all other Candida species [61, 62]. In C. albicans, these genes are typically the penultimate gene on each chromosome arm [63, 64]. The fourteen TLO genes were classified into three clades (α, β, and γ) based on sequence variation that clusters towards the 3’ end of the gene. TLO genes display ~97% nucleotide identity within a clade and 82% identity between clades (when excluding indels), yet the three Tlo clades differ in localization to different cellular compartments and in transcript abundance [63, 64].
All TLOs encode a conserved Med2 domain found in the Med2 component of the tail subunit of the Mediator complex. Accordingly, Tloα and Tloβ clade members are functional components of the C. albicans Mediator complex [65, 66]. Mediator functions as a major transcriptional regulator that recruits RNA polymerase II to specific promoters through interaction with transcription factors [67, 68]. It is unclear if TLO expansion has led to functional diversification in C. albicans and how continued evolution to produce diverse sequences affects functional specialization of the TLO genes. More broadly, it is not known how gene family expansion beyond a few members shapes the functional specificity of individual members within the amplified gene family.
Here we investigated the role of individual TLO genes across a breadth of biological functions relevant to virulence and to growth under different nutrient conditions. Induced expression of individual TLO genes using a Tet-ON approach altered a range of phenotypes including complex interactions such as virulence. In most cases, a phenotype was affected by more than one TLO gene, but this effect was not simply a function of TLO clade or phylogenetic relatedness. Two phenotypes were associated with specific changes to the Tlo protein sequence at the C-terminal end of the Med2 domain. Furthermore, different evolutionary pressures appear to be operating on the TLO gene family, with most polymorphisms encoding synonymous changes in the Med2 region and a vast excess of non-synonymous changes occurring within the gene/clade-specific 3’ end. Thus, expansion of the TLO gene family is associated with functional diversification, with significant evidence of selection operating on regions and specific sites within the genes.
Previous experiments have assessed aggregate information on TLO function as part of C. albicans Mediator [66] or for a select few TLOs under relatively isolated conditions [62]. Yet, retention of the recently expanded TLO gene family across multiple sequenced isolates of C. albicans, despite the high frequency with which it diverges [36], suggests that individual family members likely provide a selective advantage [69]. To test this hypothesis, we constructed strains in which the expression of individual TLO genes could be manipulated through a regulatable promoter, via the Tet-ON inducible expression system designed for use in C. albicans [70].
The Tet responsive promoter (pTET) was targeted to the endogenous locus of individual TLO genes where it replaced one of the native promoter alleles (Fig 1A). Integration of the targeting construct produced an in-line inducible expression system in which transcription is activated upon addition of doxycycline (+Dox) and repressed when no Dox (-Dox) is present. In the absence of Dox, only the TLO allele lacking the pTET promoter is expressed. Repeated transformations were performed to produce a series of strains with each strain containing a single Tet-inducible TLO gene (S1 Table). Ultimately, we isolated inducible strains for all TLOs with the exception of TLOα1 and TLOα10.
The strains harboring Tet-regulated TLOs were then tested for expression in the presence and absence of doxycycline. Primers unique to each TLO gene [64] were used to determine the total transcript abundance for individual TLOs. Addition of doxycycline to the parental SC5314 strain did not produce any consistent alteration on collective TLO gene expression (p = 0.371) (Fig 1B). Integration of the Tet-regulated promoter at TLO genes reduced the native expression levels of most targeted loci (Fig 1B, S2 Table), consistent with loss of expression of the regulated allele in the absence of Dox. Induction of the pTET-TLO allele by addition of Dox increased transcript abundance significantly for regulated TLO genes (p = 0.034) (Figs 1B and S1). Thus, integration of a Tet-regulatable promoter at individual TLO loci allows each TLO gene to be manipulated and assessed for phenotypic contributions individually.
Tlo proteins are incorporated into Mediator, which modulates the expression of a large proportion of the encoded genome [67, 68, 71]. Mediator has previously defined regulatory roles in carbon utilization during growth [72, 73], with MED2 playing a specific role in gluconeogenesis [74]. To identify alterations in growth that may result from changes in TLO expression, doubling times were calculated for all TLO inducible strains across a range of nutritional environments. When grown in in rich media conditions (peptides and carbohydrates) with dextrose as the primary carbon source, induction of five different TLO genes (TLOα12, TLOβ2, TLOγ8, TLOγ11, and TLOγ13) increased the observed doubling times, indicating a reduced growth rate relative to uninduced expression of the same strains (Fig 2A). Cells grown with sucrose as the primary carbon source displayed a wider range of doubling times for Tet-induced TLO genes, with all strains showing a similar trend towards slower growth (Fig 2B). Six TLOs increased doubling times when induced during growth in sucrose with two of these genes also having increased doubling times in dextrose. Little effect on growth rates was observed when cells were cultured in fructose-containing media (Fig 2C). Inclusion of maltose as the primary carbon source had the opposite effect (Fig 2D): most strains grew more rapidly (lower doubling times). Three strains (TLOα9, TLOγ11, and TLOγ13) had significantly faster growth on maltose under inducing conditions. Importantly, addition of Dox to the parental strain had no significant effect on doubling time across the assayed growth condition. These data suggest that there is a complex interplay between growth rates, carbon sources, and the expression of different constellations of TLO genes.
Although Tet-induced TLO expression affected growth rates across a range of carbon sources in rich media, regulated expression had little effect on growth rates in nutrient-poor media (Spider, YP, or sorbitol) with the notable exception of growth on YP media (0.3% yeast extract, 0.5% peptone), in which growth rates increased for strains with induced expression for six of eight TLOs (S2 Fig). The six Tet-regulated TLOs that influenced growth in YP included genes that had no effect in YP media supplemented with different sugars, suggesting that the nutrients other than carbon source, such as those in yeast extract, had a different and perhaps stronger effect than did the different carbon sources in rich medium.
To determine if altered expression of the TLO genes played a role in response to other stress conditions, we tested growth in the presence of a variety of physiological stresses using spot dilution assays in which the doxycycline used to regulate TLO expression had no effect on growth. Similarly, induced TLO expression had little effect on growth under several physiological stresses, including growth on synthetic complete defined (SCD) medium at 30°C, 37°C, pH 4.0, pH 8.0, or in the presence of 100μg/mL Calcofluor White (S3 Fig). However, induced expression of TLOα3 and TLOα9 provided a growth advantage relative to –Dox in the presence of 1M NaCl, suggesting that these two alpha-clade TLOs confer some resistance to high salt conditions (Fig 2E). By contrast, under oxidative stress conditions, TLOs from the gamma-clade provided an advantage in 2mM H2O2 (Fig 2F). All strains failed to grow well at higher oxidative concentration (6mM H2O2) and induction of any single TLO did not rescue growth (S3C Fig). Induction of TLOα3 expression revealed a growth advantage in the presence of hydroxyurea (HU), a DNA damaging agent (S4 Fig). Conversely, TLO induction had more prominent effects in response to methylmalonyl sulfonate (MMS), a different DNA damaging agent. Induction of TLO genes both increased resistance to MMS, as was seen with TLOα9, and reduced resistance with induction of TLOγ4 and TLOγ11 (Fig 2G). Thus, TLO genes may influence survival under a range of stress conditions but they appear to play a more prominent role in carbon utilization.
Preliminary observations of prepared overnight cultures indicated that cells expressing inducible TLOs (supplemented with Dox) were more flocculant because they settled more rapidly when left undisturbed, compared to SC5314 +Dox. A more quantitative analysis of flocculation, in which optical density (OD600) of vortexed cells was monitored at 15 min intervals, found that induction of all Tet-regulated TLO strains flocculated faster than SC5314 +Dox (Fig 3A). Furthermore, induction of TLO expression resulted in faster flocculation relative to the–Dox condition for half of the assayed TLO genes (Fig 3B).
Increased flocculation may result from changes in cell size and/or cell aggregation. Neither introduction of the pTET promoter nor induction of TLO expression for any strain caused a noticeable change in cell size (S5 Fig). Conversely, induced TLO expression significantly altered cell aggregation. Whereas SC5314 formed aggregates composed of roughly equal numbers of cells in the presence or absence of Dox, induction of TLO expression significantly decreased aggregate size for seven of twelve TLOs (Fig 3C and 3D). Reduced aggregate size would be expected to decrease the degree of flocculation, because larger aggregates should settle more rapidly. This suggests that additional factors likely contribute to the enhanced cell settling phenotype in Tet-induced TLO strains.
Filamentous growth can contribute to flocculation by both increasing cell size and altering the surface properties of C. albicans cells, such that they adhere to one another more readily [75, 76]. To test the degree to which filamentous growth affected flocculation, we performed solid agar adhesion-invasion assays for all Tet-regulated TLO strains. No tight adhesion to solid YPD or Spider media at 30°C was detected for any of the strains under any condition (S6 Fig). However, Tet-induced TLO expression did influence the degree of agar invasion as measured by observable hyphal density and/or prevalence. Induction of TLOγ7 and TLOγ8 decreased and increased, respectively, the extent of agar invasion on YPD at 30°C (Fig 4A and 4D). On solid Spider media at 30°C, increased agar invasion occurred for strains containing four regulated TLO genes (TLOα9, TLOγ8, TLOγ13, and TLOγ16 (Fig 4A and 4D).
An alternative approach to assess filamentous growth is to measure a modified M score [69], which quantifies the relative abundance of filamentous growth within a colony’s mass. After 7 days of growth on either YPD or Spider media in the presence of absence of Dox, colonies were imaged and the degree of filamentous growth was measured. Custom scripts assisted in these measurements that differentiate radial filamentous regions of the colony (green) from the central colony body (red) (Fig 4B). This script also accounts for colonies that fail to produce any significant filamentation (blue) in the overall filamentous growth score. As with agar invasion, addition of Dox to SC5314 parental cells did not induce a change in filamentous growth. In contrast, TLOα9 +Dox increased and TLOγ4 +Dox decreased filamentous growth (Figs 4C, 4D and S7A). Of note, TLOγ4 was scored as ‘hypofilamentous’ upon Tet induction relative to the uninduced condition because it was hyper-filamentous in the–Dox conditions relative to wildtype levels of filamentous growth in the presence of Dox. Induction of TLO expression on Spider media did not alter filamentous growth for any of the assayed strains (S7B Fig). Thus, Tet-regulated expression of TLO genes affected filamentous growth in a condition-dependent manner influenced by nutrient, carbon source, stress, and potentially other environmental conditions.
In C. albicans, biofilms require both the adhesion of yeast cells to the substrate at the base of the biofilm and subsequent filamentous growth to form an interwoven hyphal mat that accounts for much of the biofilm biomass [77]. Biofilm formation on silicone implanted devices is clinically relevant because it can seed disseminated infection and complicate patient treatment [49, 78, 79]. To assess biofilm formation, we used a simplified in vitro system in which cells were incubated with silicone elastomer squares and allowed to form communities for approximately 3 days (Fig 5A). TLO expression was induced overnight, prior to incubation on the silicone substrate, and was discontinued during the process of biofilm formation. Tet-regulated induction of two TLO genes, TLOα34 and TLOα9, reduced biofilm mass in the absence of induction compared to the parental SC5314 strain (S8A and S8B Fig). When induced in the presence of Dox, TLOα3 and TLOα34 increased biofilm mass while TLOγ16 decreased biofilm mass significantly relative to SC5314 (S8A and S8C Fig). Three genes affected biofilm biomass when induced. TLOα3 and TLOα34 increased biofilm mass and TLOγ16 reduced biofilm mass when induced (Fig 5B). Transcript levels of both TLOα3 and TLOγ16 in the Tet-regulated strain increased dramatically during biofilm production compared to growth in liquid YPD (Fig 5C). Dox-induction of TLO expression yielded a small increase in TLOα3 transcript abundance and a sharp decrease in TLOγ16 transcript abundance relative to their–Dox levels (p = 0.013), which mirrors the change in biofilm production following induction. Thus, integration of the Tet-inducible expression system at specific TLOs altered phenotypes independent of induction of TLO expression with Dox. Additionally, biofilm formation is a complex phenotype involving multiple processes; accordingly, TLOs involved in biofilm production did not completely overlap with those contributing to filamentation or cell-cell adhesion.
Recent work has highlighted a role for the Mediator tail subunit in resistance to azole class antifungal drugs [80–82], but involvement by TLO (Med2 in Mediator) was not specifically addressed. Tet-regulated TLO strains incubated overnight with or without Dox were plated onto solid agar and allowed to grow in the presence of a 25 μg fluconazole disc. After two days of growth, the size of the zone of inhibition appeared similar across most strains and induction conditions. The susceptibility phenotype (size of the zone of inhibition (ZOI)) of two induced TLOs, TLOα3 and TLOα34, decreased and increased, respectively, when induced compared to the uninduced state (Fig 6). Changes in resistance due to regulated TLO expression were relatively minor, typically altering the size of the ZOI by no more than 15%. No alterations to fluconazole tolerance (measured by the fraction of growth inside the ZOI [83]) were apparent for any strain in the presence of absence of Dox and the rate of change in growth (slope) differed for only a single TLO, TLOγ11, in the presence of Dox (S9 Fig). Thus, expression of a few TLO genes, one telomeric and one located far from the telomeres had some effects on azole drug resistance, although this effect was neither broadly conserved among TLOs nor profound.
To more directly test the role of TLOs in virulence, Tet-regulated TLO strains were co-incubated with RAW 264.7 macrophages in vitro at an MOI of two following logarithmic phase growth in the presence or absence of Dox. After 16 hours, LDH release from infected cultures was measured to quantify macrophage survival (Fig 7A). C. albicans cells with induced expression of TLOα34, TLOα9, TLOα12, or TLOγ11 resulted in more macrophage death compared to the uninduced cells of the isogenic strain (Fig 7B).
To test virulence with an in vivo model, we infected Galleria mellonella, a model for disseminated candidiasis, with C. albicans [69]. Larvae were infected with overnight cultures of C. albicans cells that had been induced or not induced with Dox and larval survival was monitored during the infection. Induction of three TLOs, TLOα34, TLOγ4 and TLOγ7, altered the morbidity of infected Galleria worms. Of these, Tet-induced expression of two genes, TLOα34 or TLOγ4, significantly increased lethality (Fig 7C), while induction of TLOγ7 reduced virulence compared to the uninduced state (Fig 7D). Thus, individual TLO genes, when induced, have different effects on virulence attributes such as macrophage lysis and G. mellonella viability.
Taken together, the above results reveal that TLO genes evolved varying degrees of influence on different virulence traits of C. albicans. A heat map displaying all significant associations of individual TLO expression (+Dox vs.–Dox) with each assayed phenotype reveals that there are few conserved functions shared by most of the TLO genes (Fig 8). Induced TLO expression promoted unidirectional changes in a number of phenotypes such as cell aggregation, growth at 30°C, and macrophage killing. Yet, a number of phenotypes can change in either direction upon induction of specific TLO genes. Thus, TLO gene family members have shared and unique modes of transcriptional regulation. This suggests a complex pattern of genotype-phenotype associations due to evolution and inheritance of TLO genes in C. albicans.
To better visualize the relationship between TLO genes in controlling phenotypic traits, hierarchical clustering was performed using the phenotypic data for all Tet-induced loci. Comparison of phenotypic scores (Fig 8) in all pairwise combinations for the TLO genes served as the basis for calculated relative distance (S10 Fig). One major branch-point separated the TLOs into two main clusters, which were each composed of a mixture of TLOα and TLOγ genes (Fig 9A). This suggests that the functions acquired by different Tlo proteins are not clade-specific. Yet, replotting the data for each TLO using principal components analysis (PCA) assigned 25.5% and 19.6% of the variation among the data set to PC1 and PC2, respectively (Fig 9B). Interestingly, this approach separated the TLOγ genes into two clusters on either side of the main TLOα genes cluster and TLOβ2. The two TLOγ groups separated primarily along PC1 with TLOγ4 and TLOγ11 being lower on PC1. A single gene, TLOα9, remained an outlier. This suggests that a mixture of clade-associated and TLO-specific features produce the functional variation observed among Tet-regulated TLO strains.
The sequence of TLO genes can be separated into roughly two halves, an N-terminal Med2 domain and a C-terminal gene/clade-specific region [64]. While the Med2 domain is responsible for the association of Tlo with the Mediator complex, the function of the C-terminal region is less clear and may interact with specific transcription factors to recruit RNA polymerase II through Mediator [65, 84]. To map individual phenotypes to specific polymorphisms that differ between members of the TLO gene family, we focused only on the Med2 domain, as the gene/clade specific region sequence diverged too much to allow individual substitutions to be analyzed across all TLO clades. Within the first 315 nucleotides (nt) of the genes, encompassing the Med2 domain, seven polymorphisms could be correlated relative to 15 phenotypes that were altered upon TLO induction. Three phenotypes mapped to specific sites within the Med2 domain. Doubling time in YPD rich media associated specifically with a synonymous polymorphism (A or G) at nucleotide 201 (p = 0.034) (Fig 10A). Two traits, the ability to lyse macrophages and the growth rate in YPS mapped to polymorphisms at nucleotide positions 303 to 306 near the end of the Med2 domain (p = 0.025). This polymorphic site includes synonymous G to A transition at position 303 together with a three nucleotide CGT indel beginning at position 304, which alters the coding sequence by introducing an arginine. Many other positions in the clade/gene specific region of TLOs may affect phenotypic properties of C. albicans, but the high prevalence of indels following the Med2 domain precludes a systematic analysis.
Comparing variants in TLO sequences to the phylogenetic tree allows a reconstruction of the mutational history of the TLO gene family during evolution. To identify mutations that arose during gene family expansion, two closely-related TLOs (i.e., TLOγ5 and TLOγ13) were compared to build a common ancestral sequence that occupied the node connecting those two genes (S11A Fig). This process was repeated until all nodes were connected through reconstructed sequences. This reconstruction identified 146 unique mutations that arose during TLO expansion, with most polymorphisms clustered towards the 3’ end of the gene in the gene/clade-specific domain (S10B Fig). Importantly, the ratio of non-synonymous to synonymous mutations was highest in the gene/clade-specific region; the Med2 domain harbored a significantly higher frequency of synonymous than non-synonymous SNPs. This suggests that different evolutionary pressures are operating on the TLO sequence: purifying selection acting on infrequently maintained SNPs has promoted sequence identity of the Med2 domain, while positive selection has diversified the gene/clade-specific domain. This implies that all of the Tlo proteins continue to function through their interaction with Mediator.
Sequencing of TLOα34, the one TLO gene not located at a telomere, identified multiple polymorphisms relative to the genome reference sequence, including nine SNPs (four of them non-synonymous substitutions that produced significant amino acid changes (A502D, V509D, V511D, and L535S) and two insertions/deletions (indels) within the 3’ gene/clade-specific domain of TLOα34 compared to the Assembly 21 (A21) sequence (S12 Fig., S3 Table). Together with an eighteen nucleotide insertion and three nucleotide deletion, these mutations suggest that rapid TLO evolution is not limited to those genes found within subtelomeres.
TLO sequences within a clade had relatively neutral selection coefficients (mean Ka/Ks of 0.76 and 1.25 for TLOα and TLOγ intra-clade diversity, respectively), which increased dramatically to Ka/Ks = 5.33 between TLOβ2 and TLOα-clade ancestral sequences and to Ka/Ks = 2.45 between TLOα/β and the TLOγ ancestral sequences (S11A Fig). The average selection coefficient for all TLOs is higher than for other Candida expanded gene families, including the serine aspartyl proteases (SAPs) in C. albicans (Ka/Ks = 1.70), the EPA adhesins in C. glabrata [85] (Ka/Ks = 1.41), and other expanded gene families in C. albicans (S11B Fig). Thus, it appears that selection has propelled divergence within the TLO gene family that appears to have had phenotypic consequences on TLO function.
Evolutionary studies of functional divergence following gene duplication commonly analyze variation between two paralogous sequences, to facilitate direct comparison. The degree to which extensive gene family expansion associates with continued functional diversification remains largely unexplored due to the complex nature of assessing individual family members for specific phenotypes. Here, we performed functional analysis of 12 of the 14 C. albicans TLO genes and found that different TLOs regulate distinct phenotypic properties to different degrees. Each TLO affected multiple phenotypes and most phenotypes were affected by multiple TLO genes when individual TLO genes were induced for expression. The TLO gene family has undergone extensive genotypic evolution with a significant proportion of variation occurring within the gene/clade-specific 3’ end, which is experiencing significant positive selection for acquired mutations. Furthermore, phenotypic variation in three traits could be mapped to specific polymorphic sites in the TLO gene family, suggesting specific mutational events following gene duplication lead to diverse functions.
The TLO gene family encodes a highly similar set of interchangeable protein subunits, yet individual genes affect distinct sets of biological functions. For example, induction of TLOα9 altered outcomes in seven phenotypic assays ranging across cell growth, filamentous growth, stress responses, and interactions with macrophages. Additionally, most TLO genes caused a mixture of phenotypic outcomes when induced, suggesting that a single TLO likely affects the differential expression of a significant number of downstream genes to produce the observed phenotype. Indeed, the two C. dubliniensis TLO homologs each regulate a large combination of unique and overlapping gene sets that promote exclusive and overlapping phenotypes [86]. We assume that incorporation of a particular Tlo protein into the Mediator complex may shift the relative expression of a distinct set of genes and thereby modulate a particular phenotypic response. The resulting phenotypic plasticity has the potential to confer a repertoire of available Mediator ‘types’ that could operate as a primary driver of TLO expansion. Thus, retention of divergent TLOs would act as a bet-hedging mechanism by which shifts in the incorporation of certain Tlo proteins would provide greater adaptability during changes in growth conditions or new host niches.
Induced expression of individual TLOs provided the most direct route to assess paralog function. Regulated transcription of candidate genes can overcome difficulties in phenotypic expression due to compensation and redundancy but introduces its own caveats such as toxicity, pathway overload, stoichiometric imbalance, and promiscuous interactions with non-physiological targets when a gene is overexpressed [87]. Thus, overexpression can produce phenotypes that are not directly attributable to the gene of interest but other affected cellular processes [88–90]. Indeed, genetic analysis using an inducible deletion or overexpression system in C. albicans found disagreements between the two approaches that may reflect these effects [91]. Furthermore, the strength of induced expression in C. albicans can alter observed phenotypes [92, 93]. Aberrant phenotypes produced by induced TLO expression were mitigated, in part, by a lack of noticeable toxicity and association with known pathways. Previous studies demonstrated that Tlo proteins exist in excess of other Mediator components as a free Tlo pool [65, 94], which suggests inherent stoichiometric imbalance with regards to Mediator. The Tlo incorporated into Mediator also appears quite plastic as multiple Tlos have been biochemically purified from the complex [65]. It is possible that induced TLO expression leads to target promiscuity either through Mediator’s role in transcriptional regulation or the Mediator tail’s role in chromatin remodeling [95]. However, the excess of Tlo protein found in C. albicans favors a model in which induced expression alters the relative availability of the regulated Tlo to be incorporated into Mediator and regulate expression of gene sets through interaction with different transcription factors although target promiscuity may occur.
Most tested phenotypes were affected by multiple different TLO genes. Nearly all TLOs had a similar effect on cell aggregation; in contrast, regulation of growth in YPS and macrophage killing fell primarily on specific TLO clades. In most cases, at least two separate TLO genes affected each phenotype altered by TLO induction and different subsets of TLO genes modulated most of the phenotypes (Fig 8). Furthermore, different genes that regulated the same phenotype displayed both enhancing and suppressing effects, indicating that the evolution of individual TLO genes as well as TLO clades likely influences the phenotypic consequences of induced expression.
15 of 22 phenotype assays detected a phenotype associated with induction of at least one or two of the 12 TLO genes tested. Six of the seven phenotypes that were not affected corresponded to different environmental stresses such as pH and high temperature, and other stresses typically were significantly affected by only one or two Tet-induced TLO genes, suggesting that the TLO genes may not have a prominent role in stress responses. Previous expression profiling of SC5314 grown in a range of stress conditions that overlap with those tested here (pH 4.0, pH 8.0, Calcofluor white, etc.) supports this hypothesis: TLO genes did not display significant expression changes in a range of stress conditions [96]. This contrasts sharply with Med2 and other tail components of Mediator in S. cerevisiae that have integral roles in the regulation of general stress response pathways [97, 98]. On the other hand, overlapping contributions to a given phenotype among TLO genes may reduce the phenotypic effect of inducing a single gene or result in minor effects that are difficult to distinguish. Therefore, contributions of individual TLOs to some phenotypes may be underscored or missed entirely.
A notable exception is exposure to MMS, which methylates DNA and leads to DNA replication fork stalling. Mutants of the single copy of MED2 in S. cerevisiae display defects in viability following DNA damage analogous to the results here seen with induction of certain TLOs (TLOα9, TLOγ4, and TLOγ11) [99, 100]. We suggest that this may reflect a broader role for MED2 and Mediator in DNA regulation and repair rather than transcriptional regulation in response to extracellular stress.
In contrast to the role in stress response, induction of a diverse set of TLO genes affected growth rates in a range of carbon sources but did not significantly alter growth in minimal media. In some cases, the same gene (i.e., TLOα9) produced opposite effects when grown in media supplemented with two different disaccharide sugars. The C. albicans genome encodes 20 different hexose transporters that are regulated through complex signaling networks that remain to be fully elucidated [101, 102]. Modulation of these pathways by TLO genes could have important phenotypic consequences on carbon source utilization, which in turn could have major effects on C. albicans biology and host interactions [103–105]. Of note, induction of TLO gene expression altered growth rates either by unidirectionally increasing or decreasing doubling times for any single carbon source. Similarity in the metabolic response to induced expression of disparate TLO genes across clades suggests a conserved role for Med2 subunits that may have existed prior to TLO expansion.
Induction of TLO expression also had a profound impact on filamentous growth and biofilm formation in different media contexts. Different TLO genes promoted or suppressed filamentous growth and biofilm formation although the TLOs involved varied depending on the conditions used. Thus, TLOs whose induction increased filamentous growth did not promote biofilm formation and vice versa. The critical step of cell-cell adhesion decreased when most TLOs were induced, although correspondence with decreased biofilm mass was only seen for Tet-induced expression of TLOγ16. Interestingly, expression of TLOγ16 decreased following induction and biofilm formation whereas TLOα3 increased in expression and biofilm formation after induction, suggesting individual TLOs may affect different components of the regulatory circuits controlling biofilm formation that can, in turn, affect their own expression [106]. Thus, the cumulative steps of cell adhesion and filamentous growth are not necessarily additive for biofilm formation. Since biofilm formation can proceed without induction of the filamentous growth transcriptional circuit and filamentous growth does not necessarily yield biofilms, the two processes are not entirely codependent [77, 107, 108]. Modulation via expression of different Tlo proteins may play a direct role in promoting different aspects of biofilm formation, including adhesion, filamentous growth, and substrate invasion. Indeed, Mediator components have fundamental roles in cell state transitions that are analogous to yeast-hyphal differentiation in other organisms [109–111], suggesting a conserved role for the complex in defining cell type-specific expression. In C. albicans, Mediator and the tail module, which includes Med3 and Med2/Tlo, function within the phenotypic switch between the ‘sterile’ white and mating-competent opaque states [66]. Thus, TLOs have the potential to control cell state transitions broadly across the breadth of cell types described for C. albicans [55, 112–114] and, more specifically, for adhesion and the yeast-hyphal transition important for flocculation and biofilm formation.
Induced expression of single TLO genes had surprising effects on complex phenotypes such as immune cell survival and virulence. Unexpectedly, Tet-induced expression of a single member of the large paralogous TLO gene family increased the ability of C. albicans to kill macrophages and G. mellonella hosts. While many C. albicans mutants have defects in virulence (reviewed in [53]), attributing a given mutation to virulence traits can be complicated by general fitness defects. In contrast, TLOs that regulated pathogenicity had equivalent or slightly reduced fitness when induced despite also displaying increased virulence. This suggests that TLOs modulate genes specific to virulence properties that do not significantly influence growth and filamentous growth processes under the assayed conditions in vitro.
The subtelomeric position of the TLO genes exposes them to significantly elevated frequencies of expression variability and genome change relative to other regions of the genome. Variability in TLO expression is observed in these strains consistent with previous reports although this expression plasticity is somewhat dampened [115], presumably due loss in variation of the regulated allele. The observed variation in expression may act to promote variation in available Mediator ‘types’ by altering the available Tlo pool over time [115]. This could explain some of the phenotypic variability in these assays, as the relative abundance of the regulated Tlo in the total cellular pool is being altered but is still competing with other Tlos for incorporation into Mediator to produce an observable effect. The subtelomeric context of TLOs may also account for, in part, the large number of polymorphisms that distinguish each TLO gene among its paralogous sequences. Yet TLOα34, the only non-telomeric TLO gene, underwent significant sequence evolution as well, indicating an underlying chromosomal or genetic feature that contributes to this process. Additionally, it is reasonable to assume selection has acted on the TLO gene family during expansion. Extensive sequence variation exists among TLO family members and there is strong evidence of positive selection during their evolution, especially at the major nodes that separate the three TLO clades. Thus, a large number of SNPs differentiate the TLO clades and these SNPs often produces a change to the protein sequence. Interestingly, different selective pressures appear to be operating across the TLO gene sequence. Purifying selection across the Med2 domain likely reflects the continued requirement for integration within Mediator [116], whereas positive selection operates on the variable 3’ end of the gene that has evidence of encoding a transcriptional activation domain (TAD) [94, 116]. Variation within the TAD could provide a mechanism for recruitment of different transcription factors. Emergence of SNPs within TLO sequences to produce allelic variants may further differentiate function within a single gene although we have identified few heterozygous SNPs within chromosomal homologs of single TLO genes. This may be a consequence of frequent recombination between chromosome homologs in the subtelomeres that rapidly fixes heterozygous positions through gene conversion or break-induced replication [33, 36].
Importantly, an indel position at the end of the Med2 domain was associated with growth in sucrose and macrophage interaction, demonstrating that variant positions may be under selection for specific phenotypes. Numerous indels within the gene/clade-specific region complicates further analysis of variant positions in the 3’ end of the TLOs. Yet, it is tempting to speculate that divergent evolution of the TLO sequence, especially within the TAD, affects phenotypic plasticity among the Tet-regulated strains by affecting the expression of different sets of target genes. Thus, expansion of the Med2-domain containing TLOs in C. albicans led to sequence variation that results in phenotypic variation to promote a highly adaptive lifestyle.
Strains of Candida albicans used in this study are listed in S1 Table. Strains were grown on YPD agar at 30°C unless otherwise noted. For induction of the tetracycline-inducible system, cultures were grown overnight in 3 mL of YPD liquid media with constant agitation in the presence (induced) or absence (uninduced) of 50 μg/mL of doxycycline. Saturated cultures were then prepared for individual experiments using their respective protocols.
Strains were transformed by standard lithium acetate transformation procedures as described previously through multiple rounds of transformation [117]. For integration of the tetracycline-inducible system at the endogenous TLO locus, the tetracycline-responsive promoter, the reverse tetracycline transactivator (rtTA), and the nourseothricin resistance marker (SAT1) were amplified from plasmid pNIM6 [70] using primers ALO110 and ALO111. Primer sequences are listed in S2 Table. These primers target this amplicon to the native TLO locus corresponding to a direct integration at the ATG start codon. The integration site was determined by polymerase chain reaction (PCR) using primers ALO108 and ALO109, corresponding to the pTET promoter and downstream in the TLO coding sequence, respectively. In some cases, additional sequencing was required to specify the TLO targeted by integration. Amplification of the integration site with ALO108 and primers ALO225, ALO226, and ALO227, which bind farther downstream within the clade-specific region, were used identify integration at specific TLO genes for clades α, γ, and β, respectively.
RNA was collected from 2x106 cells grown for four hours in liquid YPD medium in the presence or absence of 50 μg/mL of doxycycline. Cells were removed from the medium and RNA isolated using the MasterPure Yeast RNA Purification Kit (EpiCentre, Madison, WI) according to the manufacturer’s instructions. Subsequently, 1 μg of RNA was used to synthesize cDNA using oligo-(dT)18 and Superscript III reverse transcriptase (Thermo Scientific, Waltham, MA). cDNA was assayed for genomic DNA contamination using intron-spanning primers, ALO30 and ALO31, for ribosomal protein large subunit 6 (RPL6) and only cDNA lacking genomic contamination was used for qRT-PCR (S4 Table).
qRT-PCR was performed with PowerUp SYBR Green (Applied Biosystems, Foster City, CA) using an Applied Biosystems QuantStudio 3 qPCR machine and analyzed with the QuantStudio Design and Analysis Software package version 1.4.2. Primers used are listed in S4 Table. Quantification of individual TLO genes was assessed relative to ACT1. The comparative Rq method was used measure expression levels. Experiments for each gene were performed a minimum of three biological replicates in technical duplicates.
Overnight cultures were grown in 300 μL YPD liquid medium with or without 50 μg/mL doxycycline. Cultures were diluted 1:2000 into the appropriate growth medium with continued application or lack thereof of doxycycline. Optical density was measured every 15 minutes for 18–48 hours at 30°C shaking at 250rpm using an AccuSkan FC plate reader (Fisher Scientific, Hampton, NH). The polynomial measurement of the curve was used to derive doubling times. These experiments were completed with a minimum of three biological replicates with two technical replicates each.
Overnight cultures were grown in YPD liquid medium in the presence or absence of 50 μg/mL of doxycycline. Cultures were vortexed and diluted to an initial optical density (OD600) of 2.0 in a channeled cuvette. OD600 readings were taken at the start of the assay and every 15 minutes on a ThermoFisher NanoDrop One (Fisher Scientific, Hampton, NH) for a total of 210 minutes to plot cell settling. These experiments were completed with six biological replicates.
Overnight cultures were grown in YPD liquid medium in the presence or absence of 50 μg/mL of doxycycline. Cultures were diluted 1:2 in a total volume of 100 μL YPD liquid. An aliquot was visualized across a minimum of 10 random fields of view using a Leica DM750 with an attached Leica MC170HD digital camera (Leica, Wetzlar, Germany). The number of cells per aggregate was tallied across all fields of view and plotted as average for each induced TLO gene with standard error. Two biological replicates at a minimum were performed per strain.
Overnight cultures were grown in YPD liquid medium in the presence or absence of 50 μg/mL of doxycycline. Cells from each overnight culture were counted by hemocytometer and plated at a concentration of 100 cells per plate onto solid YPD and Spider medium. These plates were grown at 30°C for 7 days and imaged using a BioRad ChemiDoc XRS+ imaging system (BioRad, Hercules, CA). Images were processed by the visual analysis tool MIPAR v1.4.1 (MIPAR, Worthington, OH) and scored using the following formula: Filamentation score = 100 * (Cf) * (0.8 (Rh/Ry) + 0.2 (Sw)). Cf is the proportion of filamenting cells, Rh is the radius of the hyphal halo, Ry is the radius of the yeast colony, and Sw is the score for colony wrinkling. Three biological replicates were performed at a minimum per strain.
Overnight cultures were grown in YPD liquid medium in the presence or absence of 50 μg/mL of doxycycline. Cells were counted with a hemocytometer and plated at a concentration of 100 cells per plate onto solid YPD and Spider medium. These plates were grown at 30°C for 5 days and imaged prior to rinsing as described above for filamentation. A steady stream of water was run over the plate to remove non-adherent colonies and imaged. Remaining colonies were then rubbed off with a gloved finger and imaged to assess agar invasion. Three biological replicates were performed at a minimum per strain.
Biofilm production and measurement was performed as outlined in (Nobile et al, Cell, 2012). Briefly, silicone squares were pre-treated overnight in 12 well tissue culture plates with 2 mL of adult bovine serum. Wells were washed with PBS and 2 mL of Spider medium was added to each well. Overnight cultures were grown in YPD liquid medium in the presence or absence of 50 μg/mL of doxycycline. Cells were introduced at an OD600 of 0.5 to each well and incubated for 90 minutes at 37°C, shaking at 120 rpm. Silicone squares were then removed with sterile forceps, rinsed in a separate PBS wash well, and transferred to a new well with 2 mL of Spider media. Cultures, now adhered to the silicone squares, were incubated for 60–65 hours at 37°C, shaking at 120 rpm. After incubation, media was gently pipetted from the wells and plate was left to dry on benchtop, slightly ajar, for 24 hours. Produced biofilm was then scraped off and weighed. Four biological replicates were performed at a minimum per strain.
Overnight cultures were grown in YPD liquid medium at 30°C in the presence or absence of 50 μg/mL of doxycycline. Cell density was determined using OD600 and cultures were adjusted to an OD600 of 1.0 in 1 mL ddH20. These dilutions were used as a base for five sequential ten-fold dilution done in a 96 well plate. Or each stress condition, 5 μl of each dilution was spotted to the appropriate prewarmed agar plates including a synthetic complete defined (SCD) medium plate absent any stressor as a control for growth. Plates were then incubated at 30°C unless otherwise indicated and imaged at 24 hours and 48 hours.
C. albicans macrophage killing was assessed by using the CytoTox96 nonradioactive cytotoxicity assay (Promega, Madison, WI). RAW 264.7 macrophages were seeded at 2.5 x 104 cells per well in a 96 well plate in RPMI supplemented with 10% fetal bovine serum (FBS) and incubated overnight at 37°C and 5% CO2. Overnight C. albicans cultures were grown in YPD liquid medium at 30°C in the presence or absence of 50 μg/mL of doxycycline. These overnight cultures were then diluted 1:20 and grown for 3 hours into logarithmic phase growth in YPD medium with or without DOX. Log phase C. albicans cultures were then washed with PBS three times, inoculated into macrophages at a multiplicity of infection (MOI) of 2, and incubated overnight at 37°C and 5% CO2. To assess macrophage killing, plates containing C. albicans infected macrophage were centrifuged at 250 x g for 5 minutes and 10 μL from each well was transferred to a new plate. The transferred solution was diluted 1:5 with 40 μL of PBS and assayed using the Promega CytoTox Assay Kit according to the manufacturer’s instructions. The abundance of lactate dehydrogenase (LDH) release was calculated according to the manufacturer’s protocol.
Galleria mellonella infections were carried out using previously described protocols (Fuchs et al. 2010). In brief, overnight cultures were grown in YPD liquid medium at 30°C in the presence or absence of 50 μg/mL of doxycycline. Cultures were washed 3 times in 5 ml sterile PBS. Cell density was quantified through hemocytometer. Cell suspensions (~2.5x10^5 CFUs) in a 10 μl volume of sterile PBS were injected into the terminal pro-leg of G. mellonella larvae (Vanderhorst Wholesale, www.snackworms.com) using a 26 G, 10 μl syringe (Hamilton, No.80300) (n = 30 larvae per TLO). Dilutions of cell suspensions were plated onto YPD agar and CFUs counted to confirm inoculum. After infection, G. mellonella larvae were incubated at 37°C for 7 days. G. mellonella larvae were scored daily for signs of death (immobility and darkened pigmentation). The Log-rank (Mantel-Cox) test was used for statistical analysis of survival curves.
Overnight cultures were grown in YPD liquid medium at 30°C in the presence or absence of 50 μg/mL of doxycycline. Cells for each strain were cultured overnight in YPD at 30°C in the presence or absence of 50 μg/mL of doxycycline. Optical density measurements were used to dilute the cultures to 0.04 OD/ml (800,000 cells/ml) and 70 μL plated onto solid YPD agar. Inoculated plates were left for one hour to dry and a single 25 μg fluconazole disc (Liofilchem, TE, Italy) was placed in the center of the plate. Cells were allowed to grow for 48 hours at 30°C and images taken using a BioRad ChemiDoc XRS+ imaging system (BioRad, Hercules, CA). Drug resistance was quantified using the diskImageR program which allows for analysis of drug response parameters [83].
Alignment of TLO sequences was performed using the Multiple Sequence Comparison by Log Expectations (MUSCLE) [118]. A phylogenetic reconstruction was produced using maximum likelihood in MEGA7.
Phenotypic correlations between TLOs were produced by converting significant phenotypic changes across all assays into either a 1, 0, or -1, indicating an increased, unchanged, or decreased phenotype, respectively. A dendrogram was constructed from this matrix using Euclidean distances in R (v3.4.2) [119]. Principal components were constructed and visualized using the pca3D package.
Statistics were performed using Microsoft Excel or R (v3.4.2) developed by the R Development Team [119]. Statistics were performed with a Student’s t-test unless otherwise annotated.
|
10.1371/journal.pntd.0000147 | Loop-Mediated Isothermal Amplification (LAMP) Method for Rapid Detection of Trypanosoma brucei rhodesiense | Loop-mediated isothermal amplification (LAMP) of DNA is a novel technique that rapidly amplifies target DNA under isothermal conditions. In the present study, a LAMP test was designed from the serum resistance-associated (SRA) gene of Trypanosoma brucei rhodesiense, the cause of the acute form of African sleeping sickness, and used to detect parasite DNA from processed and heat-treated infected blood samples. The SRA gene is specific to T. b. rhodesiense and has been shown to confer resistance to lysis by normal human serum. The assay was performed at 62°C for 1 h, using six primers that recognised eight targets. The template was varying concentrations of trypanosome DNA and supernatant from heat-treated infected blood samples. The resulting amplicons were detected using SYTO-9 fluorescence dye in a real-time thermocycler, visual observation after the addition of SYBR Green I, and gel electrophoresis. DNA amplification was detected within 35 min. The SRA LAMP test had an unequivocal detection limit of one pg of purified DNA (equivalent to 10 trypanosomes/ml) and 0.1 pg (1 trypanosome/ml) using heat-treated buffy coat, while the detection limit for conventional SRA PCR was ∼1,000 trypanosomes/ml. The expected LAMP amplicon was confirmed through restriction enzyme RsaI digestion, identical melt curves, and sequence analysis. The reproducibility of the SRA LAMP assay using water bath and heat-processed template, and the ease in results readout show great potential for the diagnosis of T. b. rhodesiense in endemic regions.
| Control of human African trypanosomiasis (HAT) or sleeping sickness relies on diagnosis and treatment of infected patients. However, the diagnostic tests in routine use have limited sensitivity, due to a characteristically low parasitaemia in infected individuals. Differentiation of infections by Trypanosoma brucei rhodesiense (causes acute disease) and T. b. gambiense (causes chronic disease) is essential, as the two forms of disease have different treatment regimens. In the present work, loop-mediated isothermal amplification (LAMP) of DNA was successfully used to detect T. b. rhodesiense, with a sensitivity of up to one trypanosome/ml of blood. The LAMP test was efficient and robust, and results were obtained within 35 min. Amplification was possible when a water bath was used to maintain the temperature at isothermal conditions (60–65°C), and results could be read by visual observation of colour change. These findings have increased the prospects for developing a simple molecular test for HAT that can be used with limited equipment at point of care in endemic rural areas.
| Human African trypanosomiasis is endemic in tropical Africa. In eastern and southern Africa the disease is caused by Trypanosoma brucei rhodesiense, while T. b. gambiense infections are common in central and West Africa. T. b. rhodesiense causes an acute form of disease, whereas T. b. gambiense causes a more chronic form. Moreover, the treatment regimen for the two infections is different, expressing the need for a specific diagnostic test for each trypanosome. The geographical demarcation of T. b. rhodesiense and T. b. gambiense to a large extent forms the basis of trypanosome identification and treatment. In East Africa the introduction of T. b. rhodesiense into the T. b. gambiense region is certain to occur due to the closeness of the two disease foci and continuous movement of the livestock-reservoir host for T. b. rhodesiense. This prospect further obligates the development of test kits that can differentiate the two parasites. The serum resistance-associated (SRA) gene [1],[2] is conserved and specific to T. b. rhodesiense [3]–[5] and therefore provides unequivocal identification of this parasite. It is a low-copy gene, therefore the polymerase chain reaction (PCR) test is inadequate to amplify this target reliably in clinical samples without recourse to parasite multiplication in mice. Besides, available molecular methods of parasite detection require elaborate precision instruments [3]–[7], which make their use under field conditions unfeasible. There is therefore a need for a simplified method of amplification and product detection that would compliment the available tests and make feasible molecular diagnosis for case detection and confirmation of cure in the regions that are endemic for sleeping sickness.
Recently, a technique called loop-mediated isothermal amplification (LAMP) of DNA has been developed [8]. The technique uses four to six primers that recognise six to eight regions of the target DNA, respectively, in conjunction with the enzyme Bst polymerase, which has strand displacement activity. The simultaneous initiation of DNA synthesis by multiple primers makes the technique highly specific. The LAMP test is carried out under isothermal conditions (60–65°C) and produces large amount of DNA [8]. The reaction shows high tolerance to biological products [9], meaning that DNA extraction may not be necessary [10], and the product can be inspected visually by the addition of SYBR Green I [11],[12]. Briefly, LAMP proceeds when the forward inner primer (FIP) anneals to the complementary region (F2c) in the target DNA and initiates the first strand synthesis, and then the outer forward primer (F3) hybridises and displaces the first strand, forming a loop structure at one end [8]. This single-stranded DNA serves as template for backward inner primer (BIP)-initiated DNA synthesis and subsequent outer backward (B3)-primed strand displacement DNA synthesis, leading to the formation of dumbbell-shaped DNA structures [8]. The stem-loop thus formed acts as a template, and subsequently one inner primer hybridises to the loop on the product and initiates the displacement DNA synthesis, forming the original stem loop and a new stem loop that is twice as long [13]. The final products are stem-loop DNAs with several inverted repeats of the target DNA, and cauliflower-like structures bearing multiple loops [8].
A number of LAMP tests to detect parasitic protozoa have been designed and used successfully [14]–[16]. The rapidity, specificity, and simplicity of the technique make it appealing for use in trypanosomiasis-endemic regions. The purpose of the present study was to develop a LAMP test for detection of T. b. rhodesiense based on the SRA gene and compare it with PCR test that is specific for T. b. rhodesiense. Our results indicate that the SRA LAMP is sensitive and specific and has the potential to be developed into a field-friendly diagnostic test.
Institutional Ethical Clearance for the collection of human samples had been obtained from the Livestock Health Research Institute (LIRI), Tororo, Uganda, and the Uganda National Council of Science and Technology (UNCST), Kampala, Uganda, which records and regulates all research activities in the country. At Murdoch University, Perth, Australia, the use of mice was approved by Murdoch University Animal Ethics Committee (AEC).
The trypanosome DNA samples used in this study are shown in Table 1. The samples which most had been passaged in mice were chosen to ensure a wide geographical representation, different times of isolation, and hosts (Table 1). Six samples designated as JE (three each from blood and cerebrospinal fluid [CSF]) were direct isolates from human hosts. The DNA had been prepared using several methods (see footnotes in Table 1). The samples for studying analytical sensitivity and tolerance of LAMP were obtained from the blood of mice infected with T. b. rhodesiense and divided into two portions. The first portion was centrifuged at 3,000 rpm for 10 min and the buffy coat was collected, and the second portion was divided into aliquots of 10 µl. Then each of the two portions was mixed with 40 µl of ultrapure water, boiled for 3 min, and centrifuged at 14,000 g for 5 min. Samples of 10–15 µl of supernatant were recovered and stored at −20°C for later use.
Trypanosomes belonging to the subgenus Trypanozoon were analysed using TBR1 and 2 primers [7]. Furthermore T. b. rhodesiense was detected by a PCR specific for the SRA gene [3]), whereas T. b. gambiense was detected using a PCR for the T. b. gambiense-specific glycoprotein (TgsGP) gene [17].
LAMP reactions of 25 µl were standardised for optimal reagent concentrations, temperature, and time conditions using T. b. rhodesiense isolate LVH 56 and following the Taguchi design [18]. Briefly, the FIP and BIP were varied from 0.8 µM to 2.4 µM, dNTPs from 100 µM to 400 µM, betaine from 0.2 M to 0.8 M, and MgSO4 from 0 to 4 mM. The FIP, BIP, F3, and B3 primers were designed using the PrimerExplorer v3 software (http:/primerexplorer.jp/lamp) based on the SRA gene sequence (GenBank accession number Z37159) (Table 2). Loop primers [loop forward (LF) and loop backward (LB)] were designed manually. The reactions were optimised at 2.0 µM for FIP and BIP primers, 0.8 µM for loop primer (LF and LB), 0.2 µM for F3 and B3 outer primers, 200 µM for each dNTP, 0.8 M betaine (Sigma), 20 mM Tris-HCl (pH 8.8), 10 mM KCl, 10 mM (NH4)2SO4, 2 mM MgSO4, 0.1% Triton X-100, and 8 U of Bst DNA polymerase large fragment (New England Biolabs). For real-time reactions 3.34 µM SYTO-9 fluorescence dye (Molecular Probes) was added. The template was ∼100 pg for trypanosome lysate DNA samples and 2 µl of buffy coat and supernatant prepared from boiled blood. To find the optimum temperature for the LAMP test, the reactions were carried out for 1 h at 58, 60, 62, and 64°C using the Rotor-Gene 3000 thermocycler (Corbett Research) or in a water bath at the same temperature settings. The reaction was terminated by increasing the temperature to 80°C for 4 min.
Three methods were used to analyse DNA amplification, and included electrophoresis in 1.5% agarose gels stained with ethidium bromide, direct visual inspection of the LAMP product after addition of 1 µl of 1/10 dilution of SYBR Green I (Invitrogen), and by monitoring fluorescence of the double-stranded DNA (dsDNA)-specific dye SYTO-9 [19] in a Rotor-Gene 3000 thermocycler. Real-time fluorescence data was obtained on the FAM channel (excitation at 470 nm and detection at 510 nm) [19]. Three approaches were used to confirm that the SRA LAMP test amplified the correct target: (1) the product was digested with restriction enzyme RsaI (New England Biolabs) at 37°C for 3 h, followed by electrophoresis in 3% agarose gel; (2) following amplification, the DNA melting curves were acquired on the FAM channel using 1°C steps, with a hold of 30 s, from 62 to 96°C [19]; and (3) some of the LAMP amplicon bands were excised from an agarose gel and cloned into a TOPO-TA vector (Invitrogen), transformed in E. coli and inserts sequenced using an automated DNA 3730 analyser (Applied Biosystems). The resulting sequences were aligned with the target sequence using the DNAman computer software version 5.0 (Lynnon Biosoft).
10-fold dilutions were made from infected mouse blood containing 1.0×106 trypanosomes/ml and from 100 ng of purified T. b. rhodesiense DNA, and used to determine the analytical sensitivity of SRA LAMP and PCR tests. The reactions were done in triplicates and repeated after 2 wks. The LAMP test was carried out using both cold and heated templates. The specificity of the tests were assessed with DNA from human, tsetse fly, bovine, Plasmodium falciparum, and trypanosomes belonging to other species (Trypanosoma brucei brucei, T. b. gambiense, T. b. evansi, Trypanosoma congolense savannah, T. c. kilifi, T. c. forest, Trypanosoma simiae, T. s. tsavo, Trypanosoma godfreyi, Trypanosoma vivax, and Trypanosoma lewisi).
The results of the SRA LAMP assay are shown in Figures 1–4 and Table 1. When the test was carried out without loop primers a product was detected after 50 min. The inclusion of loop primers reduced the reaction time from an average of 50 min down to between 20 and 25 min and increased the sensitivity 100-fold. The best results were obtained when the reaction temperature was maintained at 62°C. All the positive LAMP reactions produced a characteristic ladder of multiple bands on an agarose gel (Figure 1A and 1B), indicating that stem-loop DNA with inverted repeats of the target sequence was produced. Positive reactions turned green on addition of SYBR Green I, while the negative ones remained orange (Figure 3). RsaI restriction enzyme digestion and electrophoresis gave the predicted sizes of 90 bp and 114 bp (Figure 1B). The SRA LAMP amplicons showed reproducible melt curves with a Tm of ∼87.5°C, suggesting amplicons of the same sequence (Figure 4). The cloned sequence showed 100% identity with the target sequence, and revealed that the length varied with sequence repeats of primers and there complementary sequences. The analytical sensitivity of SRA LAMP assay improved from a dilution of 10−4 to 10−6 when a template (DNA or supernatant) was preheated before being added to a reaction (Figure 2), with the best detection limit of dilution 10−7 recorded with supernatant prepared from the buffy coat. The classical PCR based on the same gene [3] showed a detection limit of dilution 10−4. The SRA LAMP detected all the 49 (100%) T. b. rhodesiense (including the six samples isolated directly from HAT patients), while TBR1 and 2 primers detected 39 out of 46 (84.8%) and SRA PCR detected 31 out of 46 (67.4%) samples (Table 1). The SRA LAMP test was specific and no cross-reaction was recorded with nontarget DNA.
In the present study we were able to demonstrate the successful amplification of T. b. rhodesiense DNA within 20–25 min at 62°C using the SRA LAMP assay. However, we set the optimal time at 35 min to amplify DNA at low concentrations. The results of the SRA LAMP assay were identical when either a water bath or a thermocycler was used to maintain the temperature at 62°C, demonstrating its robustness. Preheating of the template increased the efficiency of the assay by shortening the duration (Figure 2) and increasing sensitivity of the test. DNA amplification is preceded by strand separation under isothermal conditions using betaine, which destabilises the DNA helix [8]. It would appear that preheating of the sample produced a faster and/or a greater amount of strand separation, which translated into a far more rapid assay. All positive samples detected by gel electrophoresis or in real-time using SYTO-9 fluorescence dye could also be detected visually by addition of SYBR Green I to the product. This ability highlights another advantage of LAMP technique: the results of amplification can visually be observed through addition of a DNA intercalating dye (Figure 3), eliminating the need for gel electrophoresis and greatly reducing the time taken for result analysis.
When pure trypanosome DNA was used, the detection limit of the SRA LAMP test without loop primers was an equivalent of 1,000 trypanosomes/ml. This limit was improved to an equivalent of one trypanosome/ml with the inclusion of loop primers. Increased sensitivity and reduction in LAMP reaction time with the addition of loop primers is well documented [20] and has been demonstrated in detection of Mycobacterium [11], periodontal pathogens [12], and Plasmodium falciparum malaria [10]. Loop primers accelerate the LAMP reaction by hybridising to the stem-loop region, initiating further DNA amplification [20]. When different templates were used, heat-treated buffy coat from mice blood performed better than the supernatant obtained after boiling blood samples. The higher sensitivity recorded could be the effect of concentrating the parasites in the buffy coat through centrifugation; therefore, buffy coat seems a superior template for SRA LAMP test.
The robustness of the LAMP test is demonstrated by the ability to amplify target DNA from various templates without the expensive and time-consuming process of DNA purification. We observed no inhibitory effects in using 2–5 µl of supernatant in a 25 µl reaction or an increase in sensitivity beyond 2 µl, indicating that this volume was the optimal for our samples. The possibility of using heat-processed samples without compromising sensitivity eliminates the need for DNA extraction and further shortens the LAMP reaction. Other studies have shown superior tolerance of LAMP tests for biological substances [9],[13] and heat processed blood has been used successfully in detection Malaria [10]. The method of template preparation for use in LAMP tests, however, needs to be further developed.
The potential usefulness of SRA LAMP is confirmed by its ability to detect T. b. rhodesiense directly from parasitaemic and apparently aparasitaemic clinical samples (human blood and CSF). The human blood (JE2 and JE3) and CSF samples JE8–JE10 used in the present study were negative by microscopy at the time of sampling. Parasites were demonstrated only following inoculation of the samples in mice. When the samples were tested, they were positive by SRA LAMP assay while only JE4, JE9, and JE10 were positive using TBR PCR (Table 1) [7]. Detection of aparasitaemic samples demonstrates one of the practical values of SRA LAMP in sleeping sickness diagnosis-time-consuming parasite multiplication assays in mice are unnecessary, and early diagnosis increases the chances of cure after treatment.
In the present study, amplification of the target sequence was confirmed by restriction enzyme digestion using RsaI, melting curve analysis, and sequence analysis. It is important to distinguish T b. rhodesiense and T. b. gambiense since the two parasitic infections have different treatments. In recent years the T. b. rhodesiense region in Southern Uganda has been expanding towards the T. b. gambiense focus as a result of livestock movement [21],[22]. There is therefore a need to continue development of rapid and sensitive techniques to differentiate the two parasites and to compliment the available PCR tests, and to this end the SRA LAMP assay has shown great potential for this application.
The LAMP test should theoretically not amplify nontarget sequences, since the specificity is enhanced by using a set of six primers. However there is a high risk of amplicon contamination since the tubes have to be opened to add the dye. Analysis of any false positive reactions through sequencing and restriction enzyme analysis would easily distinguish between false positive and contamination. To reduce the chances of contamination, similar protocols to those followed for PCR are required. However, the great potential for LAMP is that reactions can be performed and results read without opening tubes [23]. On this end, more work is needed to develop such a closed reaction system for diagnosing sleeping sickness.
This study has shown that the SRA LAMP assay could be developed into an assay for T. b. rhodesiense that is simple to use at point of care. The detection of the equivalent of one trypanosome/ml in the buffy coat (with the possibility of reducing this further to 0.1 trypanosomes/ml) compares well with the normal parasitaemia in humans. Since DNA amplification and reading of results require minimum equipment, the technique has great potential for use in the HAT-endemic countries as back-up test for other HAT tests currently in use. |
10.1371/journal.ppat.1004010 | Lack of Detectable HIV-1 Molecular Evolution during Suppressive Antiretroviral Therapy | A better understanding of changes in HIV-1 population genetics with combination antiretroviral therapy (cART) is critical for designing eradication strategies. We therefore analyzed HIV-1 genetic variation and divergence in patients' plasma before cART, during suppression on cART, and after viral rebound. Single-genome sequences of plasma HIV-1 RNA were obtained from HIV-1 infected patients prior to cART (N = 14), during suppression on cART (N = 14) and/or after viral rebound following interruption of cART (N = 5). Intra-patient population diversity was measured by average pairwise difference (APD). Population structure was assessed by phylogenetic analyses and a test for panmixia. Measurements of intra-population diversity revealed no significant loss of overall genetic variation in patients treated for up to 15 years with cART. A test for panmixia, however, showed significant changes in population structure in 2/10 patients after short-term cART (<1 year) and in 7/10 patients after long-term cART (1–15 years). The changes consisted of diverse sets of viral variants prior to cART shifting to populations containing one or more genetically uniform subpopulations during cART. Despite these significant changes in population structure, rebound virus after long-term cART had little divergence from pretherapy virus, implicating long-lived cells infected before cART as the source for rebound virus. The appearance of genetically uniform virus populations and the lack of divergence after prolonged cART and cART interruption provide strong evidence that HIV-1 persists in long-lived cells infected before cART was initiated, that some of these infected cells may be capable of proliferation, and that on-going cycles of viral replication are not evident.
| Anti-HIV compounds are highly effective for preventing the onset of AIDS but they do not cure infected individuals. Very low levels of virus remain detectable in the blood of most patients despite antiviral treatment and levels surge if treatment is stopped. It is crucial to understand why current treatments are not equipped to cure HIV infection so that new therapies addressing these shortcomings can be developed. By characterizing genetic sequences of HIV in patients before and during antiviral treatment, we found that the low levels of virus detected in the blood of treated patients did not result from newly infected cells but originated from cells, or the daughters of cells, that were already infected when treatment was initiated. This finding demonstrates that HIV present in blood after prolonged antiviral treatment is derived from cells infected prior to treatment which likely expanded over time through cell division. Such long lived, infected cells are likely the critical target for developing strategies to cure HIV infection.
| The HIV-1 lifecycle includes rapid and error prone nucleic acid replication that results in large and genetically diverse virus populations in vivo. The consequences of broad HIV-1 genetic diversity include the presence of viral variants containing mutations that escape immune responses or confer resistance to individual antiretroviral agents. The use of antiretroviral agents in combination results in potent suppression of HIV-1 replication and reverses immune deficiency, at least in part. Despite the ability of cART to inhibit HIV-1 replication, treatment does not eradicate infection and plasma viremia persists at low levels in the majority of patients [1], [2]. If cART is discontinued, viremia rapidly rebounds to pre-therapy levels [3], [4]. Determining the sources and mechanisms for viral persistence during cART and rebound after interruption is essential for designing strategies to eradicate infection.
The dynamics of HIV-1 decay after initiating cART can be divided into four phases [1], [2], [5]. The first phase, reflecting rapid clearance of ca 90% of productively infected cells with half-life of 1–2 days, is followed by a more gradual clearance of infected cells with a half-life of 2–3 weeks. A study by Palmer, et al. described a third phase consisting of long-lived, perhaps latently-infected, cells with a half-life of 6–44 months as well as a fourth phase having a slope not significantly different from zero [1]. The plateau in the fourth phase suggests that long-term cART fully inhibits HIV-1 replication and that the source of persistent viremia is either long-lived virus-expressing cells or activation of virus expression from latently-infected cells. In this regard, studies by Dinoso et al., McMahon et al., and Gandhi et al. showed no decrease in the level of persistent viremia in patients on long term suppressive therapy before, during, or after intensification with an additional antiretroviral suggesting the absence of ongoing new rounds of replication during suppressive cART [6], [7], [8]. Bailey et al. investigated plasma viral sequences after long-term cART and found that HIV-1 populations often contain sets of identical sequences, referred to as “predominant plasma clones,” suggesting that viral subpopulations are lost over the course of treatment [9]. Wagner, et al. found an increasing frequency of identical sequences in blood cells during cART suggesting proliferation of infected cells [10], and Joos, et al. showed that homogeneous populations rebound after cART interruption [11]. These findings suggest that a reservoir of long lived infected cells, perhaps capable of expansion, may be responsible for persistent viremia and its rebound following interruption of cART.
In contrast to these findings, other studies have indicated that low-level virus replication may occur in specific anatomical compartments despite suppression of plasma HIV-1 RNA by cART [12], [13], [14], [15], [16], [17], [18], [19], [20]. For example, in 2008, Chun, et al. suggested that phylogenetic clustering of sequences obtained from different cellular compartments after long-term cART demonstrated cross-infection between reservoirs, consistent with full cycles of replication as a source of persistent viremia [13]. Although such phylogenetic clustering may be indicative of on-going replication, it may also result from compartmental mixing of infected cells before or subsequent to initiating therapy. Demonstrating the emergence of new viral variants during cART without corresponding increases in total HIV-1 RNA would provide clear evidence of virus replication. Previous studies that demonstrated genetic change during therapy were in the context of drug resistance, rebound viremia, or stimulation following vaccination, each occurring in subsets of study patients in conjunction with increases in plasma HIV-1 RNA levels, likely reflecting ineffective therapy [12], [14], [21], [22]. Several studies using integrase inhibitors to intensify cART have detected transient increases in 2-LTR circles in peripheral blood lymphocytes, especially in individuals undergoing protease inhibitor-based cART suggesting that some cells may be newly infected during treatment [23] [24]. However, changes in 2LTR circles were not associated with decreases in viral RNA levels and genetic analyses did not show divergence during the intensification period [25]. Notably, all of these clinical studies have been conducted with patients already undergoing cART for prolonged periods. No studies have investigated HIV-1 populations prior to and following initiation of cART. Comparing pre- and post-therapy populations can shed new light on HIV-1 reservoirs, the sources of persistent viremia, and changes in HIV-1 populations at each phase of viral decay after introducing cART.
To investigate further the effect of cART on virus replication, we examined HIV-1 populations in patients prior to cART, during each phase of viral decay including long-term cART (fourth phase), and during viral rebound after interruption of cART. By investigating the genetics of HIV-1 in all phases of viral decay and comparing on-therapy populations to pre-therapy virus we were able to directly assess HIV-1 replication and molecular evolution during long-term suppressive cART. We found that both short and long lived cellular compartments were seeded with the same diverse virus populations and that new viral populations rarely emerged after up to 15 years of cART.
Participants were enrolled in prospective studies aimed at determining the role of antiretroviral therapy on HIV-1 infection (protocols 97-I-0082, 08-I-0221) or on HIV-1 population genetics in infected individuals (00-I-0110) conducted at the NIH Clinical Center in Bethesda MD [26] [27]. All participants were ≥18 years of age at study entry, with chronic HIV-1 infection (Fiebig Stage VI) and reported no prior antiretroviral therapy (Table 1). Study participants were enrolled from 1997–2002; Patients 2–4 and 6–13 initiated therapy with 2 NRTIs + nevirapine + indinavir as part of a study of HIV-1 decay kinetics [26] and Patients 1, 5, and 14 initiated therapy with 2 NRTIs + efavirenz as part of a study of HIV-1 population genetics [27] (Table 1). Frequent plasma samples were obtained prior to and following introduction of cART (Supplemental Table S1). Patients are described in Table 1 and samples analyzed in Supplemental Table S1. Patients were categorized into three partially overlapping groups according to their sample collection and treatment history (Table S1). Blood samples were collected prior to initiating cART in all patients (N = 14). In 10/14 patients (group 1) frequent samples were collected during short-term treatment (up to one year on cART). In 5 patients (group 2), samples were collected after long-term therapy (average 9 yrs on cART), and in 5 patients (group 3), samples were collected after a patient-initiated treatment interruption as well as after re-suppression in 3/5 (Table S1). Results from the sequence analysis from all groups were compared to data obtained using the same methods from a cohort of elite controllers (data previously published) [28]. The elite controllers served as untreated controls since they have similar levels of viremia (mean 0.8 copies/ml) without cART.
All participants in this study were enrolled in clinical protocols (00-I-0110, 97-I-0082, 08-I-0221) approved by the NIAID Institutional Review Board (FWA00005897) administered at the NIH Clinical Center in Bethesda, Maryland. Individuals underwent an informed consent process and provided written consent for participation.
HIV-1 RNA levels were determined using bDNA Versant version 3.0 (Bayer, Inc) as previously described [29]. Single-genome sequencing (SGS) of a portion of HIV-1 gag-pro-pol amplified from plasma HIV-1 RNA was performed as previously described [30], [31], [32]. Sequences were aligned using ClustalW. Population genetic diversity and divergence were calculated as average pairwise difference (APD) using MEGA5 [33] (http://www.megasoftware.net) and an in-house program [32]. Shifts in population structure were calculated using a subdivision test for panmixia with a significance cut off level of p<10−3 as described by the original report to account for the high number of comparisons between sequences and nucleotide sites [34], [35] [27]. The probability of 10−3 for assigning a significant change in intra-patient HIV populations was derived statistically taking into consideration that every nucleotide position is compared in every two possible sets of sequences. This approach results in more than 1012 comparisons between populations of only 10 sequences. The test was derived from a geographic population structure test proposed by Hudson et al. [36]. It compares the APD in single-genome sequences obtained from samples taken at different times (or places) to distances calculated from imaginary populations containing the same sequences randomly reassigned to two groups. Random mixing of the populations to be compared, reassignment, and distance comparisons are performed 10,000 times, generating a p-value for the probability that the randomized populations' structures are the same between sets of sequences. Neighbor-joining phylogenetic analyses were performed using MEGA5 [33]. Trees were rooted on the subtype B consensus sequence (http://www.HIV-1.lanl.gov). Tests for molecular evolution were done with BEAST [37] (http://beast.bio.ed.ac.uk) using the HYK+G model with a relaxed clock, uncorrelated log normal and constant size, followed by estimating the root to tip distances with TreeStat1.2 (www.tree.bio.ed.ac.uk/software/treestat). Linear regression was used to determine the slopes for the root-to-tip analyses.
To investigate if genetic bottlenecks occurred after initiating cART, we evaluated changes in the number of heterozygous sites over time [38]. A genetic bottleneck was present if the number of heterozygous sites in equal numbers of sequences in post-therapy samples were in excess (chi-square probability <0.05) compared to those in pre-therapy. We also investigated if CTL escape mutations were enriched or depleted during cART by calculating the allele frequencies at each amino acid position in pre-therapy and post-therapy data in patients in group 1 and 2 with 7 or more sequences at distal time points (N = 8). Positions with amino acids undergoing significant change in frequency after cART (Fisher exact test 0.05) were identified and mapped onto predicted CTL epitope maps [39]. Changes within the 9 amino acid peptide or in +1 and -1 amino acids flanking the peptide were considered to be part of the CTL epitope. The predicted binding affinities of the pre- and post-therapy peptides were compared to determine if amino acid changes occurring after initiating cART resulted in decreased binding affinity; ≥10 fold decreases in affinity were considered escape; ≥10 fold increases were considered return to wild-type allele.
To investigate the effect of cART on plasma HIV-1 diversity, we assessed HIV-1 genetics by single-genome sequencing of plasma HIV-1 RNA in individuals undergoing cART. Plasma samples were obtained prior to and following introduction of cART; and, for some patients, after planned patient-initiated treatment interruptions. Single-genome sequences were obtained at time points throughout the study period, and population genetics parameters were measured. Genetic diversity was measured by APD of virus populations in patients' plasma prior to treatment, during each phase of viral decay, and during viral rebound (Figures 1, 2). Group 1 patients were sampled during the first and second phases of HIV-1 decay on cART (up to 200 days) to investigate the effect of declining viremia on virus diversity (Figure 1a, 2a). Group 2 patients were sampled on long term cART (c. 4–12 years) without treatment interruption during the third and/or fourth phases of viral decay (Figure 1b, 2b). Group 3 patients with long-term suppression of HIV-1 underwent brief planned treatment interruptions and were sampled before and after treatment initiation and after virus rebound (Figure 1c, 2c).
Most patients (13 of 14) showed no significant difference in APD of HIV-1 populations during any phase of viral decay, after long-term therapy, or after viral rebound, compared to pre-therapy virus populations (Figure 1, 2). This finding shows, in most cases, that HIV-1 plasma diversity is not associated with the level of viremia (Figure 2), with the duration of cART, or with viral rebound after stopping cART. Figure 1 shows the diversity of plasma HIV-1 populations in each patient before and during or after interruption of cART (the value above the bar in Figure 1 shows the number of years the sample was collected after initiating cART). Of 14 patients, only one (PID 8) showed a significant reduction in viral diversity after treatment with cART (Figure 1b, 2b). The mean virus diversity across patients in each group and as a whole did not change after initiation of cART or during cART (Figure 1d), indicating that plasma virus diversity is sustained during each phase of viral decay despite the large decreases in the replicating population size. This result suggests that the cellular reservoir of persistent viremia in most patients is seeded with the same highly diverse replicating population of virus that exists prior to therapy. This observation is in contrast to elite controllers who have significantly lower levels of diversity than noncontrollers (p = 0.005) [28] correlating with their lower levels of viremia. The contrasting results suggest that the infected cell population in patients treated with cART is large while the reservoir of infected cells in elite controllers is likely to be significantly smaller.
Although HIV-1 populations revealed no significant change in APD with cART, genetic bottlenecks may occur in large, diverse populations without producing a detectable change in the overall diversity. During a bottleneck, low frequency alleles, which do not contribute substantially to overall diversity or to phylogenetic signal, are lost [38]. As a result, the total number of alleles is decreased while the diversity is maintained. Because bottlenecks will have substantial effects on the occurrence of low frequency alleles, we specifically investigated the total number of alleles prior to and following introduction of cART (Table 2) in patients with sampling during viral RNA decay on therapy (N = 9). We found a significant decrease in the numbers of alleles in only a single patient (PID 1), suggesting that a genetic bottleneck occurred in this patient alone. In two patients (PID 2 and 7), a modest but detectable increase in alleles occurred suggesting genetic shifts but not population contraction. The remaining patients had no changes in the numbers of alleles (Table 2), indicating that, for the majority of individuals, no genetic bottleneck accompanies the profound decrease in HIV RNA after initiation of cART.
To specifically investigate whether prolonged HIV-1 suppression resulted in changes in amino acid sequences, we investigated nonsynonymous changes alone in patients from groups 1 and 2 for which there were more than 7 sequences at time points with <50 copies/ml (N = 8) (Table S2). We found that amino acid frequencies were remarkably stable during cART. In fact, virus populations in 4/8 patients had no significant change at any of the PR or RT loci. As all enrolled patients underwent HLA testing, we were able to investigate, using in silico techniques, the predicted positions of all the CTL epitopes in the HIV-1 sequence as well as the estimated binding affinity of all the HIV-1 peptides at each epitope site [39]. As shown in Table S2, there was no consistent trend to enrich or deplete CTL escape mutations after prolonged cART suppression, including in those patients who underwent a significant population shift (e.g., PID 1). Taken together, these data suggest that the population of virus-producing cells present after prolonged suppression is not shaped in a substantial way by new CTL selection following introduction of cART. This finding is in stark contrast to the strong selection at CTL epitopes in elite controllers ranging from 11–66% of epitopes carrying escape mutations [28].
Divergence of HIV-1 populations during cART could result either from on-going cycles of replication leading to the emergence of new variants or as a consequence of shifts in the viral variants present in the plasma during suppression, indicating a dynamic reservoir. To investigate the possibility of population shift (divergence) during cART, we used a test for panmixia to detect changes in the population structure during therapy compared to pretherapy virus. The panmixia test compares populations of single-genome sequences obtained from longitudinal samples and provides a p-value for the probability that the populations are the same [34]. Probabilities of <10−3 were considered to indicate significantly different populations, taking into account the large numbers of comparisons. Figure 3 and Table 3 show the panmixia results for single-genome sequences from group 1 (Figure 3a, Table 3), group 2, (Figure 3b, Table 3), and group 3 (Figure 3c, Table 3) compared to pretherapy sequences. Panmixia probabilities of virus populations in samples collected from patients on cART compared to pre-therapy populations did not achieve significance (Figure 3a) in 8/10 patients from group 1. These results indicate that there is typically no significant shift in the plasma virus population during the first and second phases of decay after initiating cART despite up to 10,000-fold declines in levels of viremia. Two patients in group 1 (PID 6, 7), however, did show a significant change in population structure after 173 and 193 days on therapy. Additional analyses describing the nature of these changes are presented below. Three of 5 patients in group 2 (long-term cART) showed a significant change in population structure during cART for 4–12 years with no treatment interruptions, suggesting either that new variants emerged during therapy or that the reservoir for persistent viremia is dynamic. Four of 5 patients in group 3 (long-term cART but with brief treatment interruptions) showed a significant shift in population structure using the panmixia test. The results from group 2 and 3 show that, although plasma HIV-1 populations do not typically change in the early phases of viral decay, shifts in virus populations (without a change in overall diversity) are readily detectable after long-term therapy and in rebound viremia. They imply that either a compartment allowing on-going cycles of replication exists during cART or subsets of infected cells expressing virus particles shift over the course of treatment (through proliferation and/or death).
To further determine if the population shifts detected in the plasma of some patients during and after long-term cART were the result of on-going cycles of virus replication or were due to a shift in the population of cells that express virus particles during therapy, we performed phylogenetic analyses and tests for molecular evolution. Such tests can detect with high sensitivity the emergence of new viral variants indicative of full cycles of replication during cART. We used neighbor-joining trees to first evaluate the direct relationship of the sequences obtained prior to, during, and after therapy and we subsequently used tests for molecular evolution and calculations of root-to-tip distances to detect the emergence of new virus populations during cART.
Figure 4a shows two examples of the population structure in patients in group 1 who had no detectable shift in the virus population using the test for panmixia or the divergence analysis. Consistent with the panmixia results, the structure of sequences obtained during viral decline (gray triangles) and early suppression on cART (black triangles) showed no change from pre-therapy virus (open circles). Figure 4b shows the neighbor-joining trees for the two additional patients in group 1 (PID 6, 7) whose virus had a detectable shift in the population during short-term treatment with cART using the test for panmixia. It is evident from the trees that the shift in population and significant panmixia resulted from clusters of identical sequences that were revealed when levels of viremia were <50 copies/ml (circled black triangles). To confirm that the identical sequences found in PID 6 and 7 resulted in the population shift measured by the test for panmixia, we collapsed the alignment to include only one of each identical sequence and repeated the test. The collapsed alignments resulted in p values of 0.044 and 0.011, respectively, for panmixia (not significant), rejecting the null hypothesis. The revealing of populations of identical sequences during therapy suggests that either a single infected cell is proliferating and releasing virus resulting in a dominant variant appearing in the plasma or that a single variant is expanding through full cycles of replication despite cART. Additional analyses to investigate this question are presented later.
Phylogenetic trees of virus sequences from 6 patients on long-term cART are shown in Figure 4c–d. Trees from representative patients in group 2 (long-term suppression - Figure 4c) and group 3 (re-suppression after brief treatment interruption - Figure 4d) show that the population shift detected by the test for panmixia in these groups resulted from clusters of identical sequences in the plasma (black triangles), and not from additional accumulation of mutations. As noted above, patient 8 was the only one who also showed a significant change in the diversity of the virus population during therapy. The phylogenetic analysis shows that the loss of diversity of the virus population in this patient also resulted from over-representation of identical sequences in the plasma, possibly masking the presence of other viral variants. The presence of identical sequences after long-term cART suggests a proliferating infected cell population as a major source of persistent viremia during therapy. These data also suggest that the virus-producing reservoir of HIV-1 infection may contract during prolonged cART. In addition to the identical sequences, there were also some unique sequences detected in patients in groups 2 and 3 after long-term suppression. The presence of unique variants in the plasma during long-term treatment in PID 1 (Figure 4c) may indicate that on-going replication is another source of residual viremia during therapy in this patient. However, unique variants present in PID 11 (Figure 4c) are more likely due to replication that occurred during the brief treatment interruption in this patient.
The genetics of rebound viremia are shown in two patients from group 3 (PID 2, 9) in Figure 5; and demonstrate that rebounding virus is primarily due to populations of identical sequences, as seen during cART, suggesting a stable, non-evolving reservoir as a likely source of rebound viremia. The presence of multiple populations of rebounding virus argues against the identical sequences persisting during suppression being the source of viral rebound since, in most patients, we detected only a single population of identical variants during suppression. Rebound viremia in these two patients also includes unique variants, some of which may be recombinants between the rebounding rakes of identical sequences and accumulation of new mutations that occurred after interrupting cART.
Neighbor-joining analyses allowed us to visualize the plasma virus populations present during cART compared to those in pretherapy, but cannot be used to determine if the variants present during treatment are newly emergent resulting from full cycles of replication or if they are merely the expression of variants from cells infected prior to treatment. For this purpose, we applied a test for molecular evolution using Bayesian analysis as implemented in BEAST (http://beast.bio.ed.ac.uk) to determine if the plasma virus populations present during therapy were newly emergent variants or were pre-existing. The molecular evolution test was performed by measuring the distances from the root of the tree (rooted on consensus B) to the tip of each branch (Figure 6a–d). If the population structure results from the emergence of new variants, then those sequences will be on branches that are more distant from the root of the tree than variants present in pre-therapy, resulting in positive slopes in Figure 6 as shown in Table 3. The molecular evolution test revealed slopes that were close to 0 (median = 1×10−5±4.5×10−5 nt/day) with no significant differences between groups (t-test between groups 1 and 2 had p value = 0.72, between groups 2 and 3 p = 0.67, and between groups 1 and 3 p = 0.74), showing that the variants present during the second and third phases of decay and after prolonged therapy were not more distant from the root of the tree than variants present prior to initiating therapy. In a few cases the sequences were actually slightly closer to the root (consensus B) resulting in a negative slope. By contrast, the slopes in untreated elite controllers with similar levels of viremia have significantly positive slopes (median = 15 nt/day) (p = 0.009) when measured over similar intervals [28]. These findings indicate that the viruses with identical sequences that are revealed during cART are not the result of full cycles of replication, but are likely being released from a proliferating cell population that was infected prior to therapy. Although all patients had root-to-tip slopes close to 0, one had a slightly but significantly positive slope after long-term treatment (PID 1, Figure 6a,b, Table 3) suggesting that there is a subset of patients for whom treatment (for some period) is not fully suppressive. The remaining 13/14 patients had slopes not significantly different from zero consistent with complete suppression of viral replication.
The source(s) of persistent viremia during suppressive antiretroviral therapy remains uncertain, and there have been a number of studies to investigate whether repeated full cycles of virus replication occur during adherence to cART or if low-level viremia present in the plasma of successfully treated patients is the result of viral expression from long-lived cells infected prior to treatment. Population genetics and phylogenetic approaches represent powerful techniques to detect genetic change in temporally spaced samples, but in the setting of relatively high genetic diversity it is often difficult to determine whether observed change represents molecular evolution from ongoing replication or a shift in the population of reservoir cells producing virus. One way to resolve this issue is to compare HIV-1 populations prior to and following initiation of cART and to compare temporal changes in viral populations in treated patients to untreated elite controllers with similar levels of viremia and duration of control. In this study, we investigated HIV-1 gag-pro-pol populations in infected individuals before, during, and after cART by analyzing the effect of cART on viral genetic diversity and population structure and compared the results to similar data set from a cohort of elite controllers [28]. We previously showed that viral replication and molecular evolution occur in spontaneous HIV-1 elite controllers at levels that are not significantly different from non-controllers [28]. This finding demonstrates that our analytical methods are sensitive enough to detect the emergence of new viral variants despite very low levels of viremia. In fact, with these methods, we are able detect the emergence of new variants even if evolutionary rates are only 10% of those measured in elite controllers (15 nt changes/day). To address the question of ongoing replication during cART, we applied the same analytical approach used in the elite controller cohort [28] to HIV-1 populations in non-controllers on cART for evidence of molecular evolution during treatment. In contrast to our findings in elite controllers, we found clear evidence for virus molecular evolution in only one patient on long-term cART (without treatment interruptions) while we found no evidence for the appearance of new variants in any of the other suppressed patients.
First, we investigated virus populations in samples collected within the first 6 months of initiating cART and compared these populations to the viral sequences obtained from pretherapy samples. We found no change in the diversity, divergence, or phylogenetic structure in populations obtained before and after 6 months of ART, and no evidence of any genetic bottleneck in 8/9 patients on study. These results indicate that both short- and long-lived cellular compartments are seeded with the same viruses and that these compartments are sufficiently large to support highly diverse populations of HIV-1 genomes. The sustained diversity of HIV-1 populations over months of suppressive therapy without a genetic bottleneck or loss of low frequency alleles also implies that pre-existing low-level drug resistance mutations are not likely to be lost during antiretroviral therapy.
To investigate the genetics of persistent HIV-1 during long-term cART, we also sequenced plasma virus populations during 4–15 years of suppressive therapy; and again, compared these populations to those obtained from pretherapy samples. In contrast to earlier samples on cART, we found clusters of identical sequences in plasma samples collected after long-term treatment. However, using phylogenetic tests for molecular evolution (root-to-tip distance analyses), we found no evidence for the appearance of new variants during long-term cART and the clusters fit within the phylogenies of virus populations present before therapy (with the exception of one patient). The presence of identical sequences during ART suggests that virus particles are being produced by an HIV-1 infected clonal cell population, such as stem cell-like CD4+ memory T-cells or other proliferating cell types. These conclusions are consistent with previous studies that indicate that persistent plasma viremia during cART is derived from viral expression in long-lived cells [9], [11] [1] [6], [7], [8]. The specific cell populations giving rise to plasma viremia during cART have not yet been determined but one study nicely demonstrated significantly different populations structures between residual viremia and resting CD4+ cells in 11/13 patients [40] suggesting alternative sources for persistent viremia.
Given that long term therapy reduces the average level of viremia from about 30,000 to about 1–3 copies of RNA per ml on average [1] and that a minority (up to about 1/3) of the sequences in patients on long term therapy are clonal, we can estimate that the cells that produce such virus represent about 1 in 100,000 of the total virus-producing cell population in an untreated individual. Our findings also suggest that these cells are neither expanded nor depleted during therapy as a result of CTL selection. Our observations and those of others (14) that rebound viremia after long-term cART contains homogeneous populations suggests that rebound viremia results from the expansion of identical sequences present during suppression or from a small number of founder viruses (as seen in acute infection). Further experiments are required to determine the relationship of virus populations that persist during therapy to those that rebound after treatment interruption.
The conclusion that cART effectively and completely halts HIV-1 replication in those infected cells that are responsible for viremia is consistent with prior studies by us and others showing that low levels of viremia on therapy are independent of the therapeutic regimen used and they cannot be further suppressed by additional drugs [6], [7], [8]. Our conclusions are also consistent with the initial observations of Persaud and coworkers who demonstrated that drug resistant mutations do not emerge in patients with suppressed viremia [41]. Several observations, including transient increases in 2LTR circles in some cART treated patients undergoing raltegravir intensification, and studies measuring relative levels of HIV-1 RNA in specific compartments [42] have suggested the presence of localized, limited HIV-1 replication. However, the relationship between the 2-LTR circles and low level viremia has not been firmly established. It is likely that a very small fraction of the virions released during suppressive cART give rise to the 2 LTR circles and that these represent dead-end events, not continuous replication, most likely related to the use of a specific antiviral treatment regimen [42]. Our findings here suggest that low level viremia persisting during cART results primarily from expression of virus in expanding cell populations infected prior to initiating therapy. Cure of HIV-1 infection will require strategies that either eliminate the extremely rare cell population that can chronically produce infectious virus or prevent regrowth of virus from these reservoirs following cessation of cART.
|
10.1371/journal.pgen.1003380 | Yap- and Cdc42-Dependent Nephrogenesis and Morphogenesis during Mouse Kidney Development | Yap is a transcriptional co-activator that regulates cell proliferation and apoptosis downstream of the Hippo kinase pathway. We investigated Yap function during mouse kidney development using a conditional knockout strategy that specifically inactivated Yap within the nephrogenic lineage. We found that Yap is essential for nephron induction and morphogenesis, surprisingly, in a manner independent of regulation of cell proliferation and apoptosis. We used microarray analysis to identify a suite of novel Yap-dependent genes that function during nephron formation and have been implicated in morphogenesis. Previous in vitro studies have indicated that Yap can respond to mechanical stresses in cultured cells downstream of the small GTPases RhoA. We find that tissue-specific inactivation of the Rho GTPase Cdc42 causes a severe defect in nephrogenesis that strikingly phenocopies loss of Yap. Ablation of Cdc42 decreases nuclear localization of Yap, leading to a reduction of Yap-dependent gene expression. We propose that Yap responds to Cdc42-dependent signals in nephron progenitor cells to activate a genetic program required to shape the functioning nephron.
| The mammalian kidney undergoes reiterative and stereotypical morphogenetic changes to create the elaborately convoluted adult nephron, the functional filtration unit of the kidney. How these sequential morphological events are controlled remains poorly understood. Here we show that the transcriptional activator Yap is essential in the developing murine kidney. Yap mutants have reduced nephrogenesis and defective morphogenesis. Yap function in nephrogenesis is independent of its previously described role in regulation of cell proliferation and apoptosis. Instead, Yap activity is needed for proper expression of a suite of genes that control cell signaling and cell structure. Remarkably, we find that ablation of Cdc42 phenocopies loss of Yap. We show that Cdc42 is essential for nuclear access of Yap, both in vivo and in tissue culture studies. Taken together, our work shows that Yap and Cdc42 are essential for the cell fate and morphogenesis decisions necessary to shape functioning nephrons, and suggests that Yap functions downstream of Cdc42 during kidney development.
| Nephrons are the functional units of the kidney. Variability in nephron number (300,000 to 1 million in each kidney [1]) in human depends on both environmental and genetic factors. Low nephron number at birth correlates with increased incidence of renal failure later in life [2]. Thus, it is critical to understand the molecular mechanisms underlying nephron induction and patterning.
Kidney organogenesis is a remarkably orchestrated, reiterated process that depends on reciprocal signaling between the epithelial ureteric bud (UB) and the surrounding metanephric mesenchyme [3]–[6]. Signaling from the mesenchyme induces successive rounds of UB branching, generating the collecting duct system of the kidney. Surrounding the UB are self-renewing mesenchymal progenitor cells called the cap mesenchyme (CM). A subset of CM cells is reciprocally induced by the UB to form a pretubular aggregate (PA), which subsequently undergoes a mesenchyme-to-epithelial transition (MET) to form a renal vesicle (RV). The RV then undergoes morphogenesis, first changing into a comma-shaped body (CSB) that then elongates and folds back on itself to form a S-shaped body (SSB) (Figure 1A). Finally, the SSB further elongates and undergoes patterned differentiation to give rise to the various segments of the nephron (Figure 1A′). This process is repeated thousands of times, resulting in the stereotypical structure of the mature nephron which includes the distal tubules, proximal tubules, Henle's loops and glomeruli. How this intricate morphogenetic process is regulated is not fully understood.
The Hippo pathway is a highly conserved kinase cassette that regulates tissue growth in metazoans by controlling the activity of Yap and Taz (reviewed in [7]–[9]). Yap and Taz are closely related transcriptional co-activators that control expression of genes that promote cell proliferation and inhibit apoptosis. When the Hippo kinases Mst and Lats are active, Yap and Taz are phosphorylated and excluded from the nucleus. Loss of Hippo signaling leads to unrestricted proliferation in flies and mammals, and has been linked to a variety of cancers (reviewed in [10], [11]). Yap knockout (Yap−/−) embryos die at embryonic day 8.5 (E8.5) [12], and Taz−/− mice have polycystic kidney disease [13], [14]. Yap−/−;Taz−/− embryos die prior to the morula stage with defects in trophectoderm specification, indicating redundant roles in early embryonic development [15].
Blocking the inhibitory effects of Mst/Lats signaling on Yap, either through disruption of Salvador (Sav1/WW45) or by forced expression of a constitutively active form of Yap, leads to hyperproliferation of cells in the gut and skin [16], [17]. Hippo signaling has also been shown to restrict heart size in mice [18]. Upstream of Hippo kinases lie a number of cell surface regulators, which include cadherins, cell polarity complexes and GPCRs [19]. These and other data in flies, fish and mice (reviewed in [7], [8], [20]) have led to a model in which Yap and Taz primarily function to regulate tissue growth.
Surprisingly, recent studies in tissue culture demonstrate that Yap and Taz also respond to mechanical stresses [21], [22]. Plating cells on a rigid substrate induces the nuclear localization of Yap and Taz, promoting transcription of Yap/Taz targets that enhance cell proliferation and inhibit apoptosis. Disruption of the cytoskeleton has also been shown to regulate cell proliferation and Hippo pathway activity in Drosophila (reviewed in [20]). These studies specifically highlight the importance of the establishment of cytoskeletal tension for Yap/Taz dependent-mechanotransduction and implicated RhoA activity and the integrity of stress fibers in mediating Yap/Taz responses to mechanical cues. The in vivo developmental relevance of these findings has not been addressed before. In addition, no study has examined to date whether the activity of other Rho GTPases can influence Yap function.
We have investigated the role of Yap signaling in the context of nephrogenesis in the murine kidney. To bypass the early lethality of Yap−/− mutants, we used tissue-specific deletions to study Yap function during nephron formation. We found that Yap conditional mutants display early defects in nephron induction, and that the stereotypical morphogenesis that remodels the RV into SSB is disrupted. Importantly, these early defects occur independently of major alterations in proliferation or apoptosis. Using microarrays, we identified a suite of genes whose expression depends on Yap during kidney morphogenesis. Intriguingly, we find that loss of the small Rho GTPase Cdc42 leads to a reduction of nuclear Yap in the CM and cytoplasmic accumulation of Yap in cultured fibroblasts. Importantly, nephrogenic inactivation of Cdc42 leads to loss of Yap-dependent gene expression. Moreover, Cdc42 and Yap removal in the CM leads to remarkably similar morphological defects and abnormalities in nephron gene expression. Together these data support a model in which Cdc42 acts upstream of Yap in nephron progenitor cells, to promote gene expression required to establish and shape nephrons.
To investigate a potential function of Yap in nephrogenesis, we first stained developing kidneys with antibodies to Yap, and found that Yap was dynamically expressed throughout nephrogenesis. As a transcriptional co-activator the function of Yap is primarily regulated at the level of access to the nucleus [23]. Yap is expressed in the ureteric compartment and cortical stromal cells, with lower levels of expression in the CM (Figure 1B, 1E). Strikingly, we noted that the distribution of Yap is regulated spatially and temporally during nephrogenesis. In early nephrogenic structures, Yap is strongly expressed in proximal cells of the RV (Figure 1B, 1E) and in most distal and proximal cells of the SSB (Figure 1E, 1F–1F″). This dynamic expression pattern was seen using two different Yap antibodies, and was lost upon deletion of Yap from the CM using Six2:Cre (Figure 1D and Figure S1).
Yap localization in the nucleus is often regulated by phosphorylation. We stained embryonic kidneys with antibodies that recognize Yap phosphorylated at S127, a site that is phosphorylated by Lats in response to Hippo activation [24], and found that phospho-Yap staining is detectable throughout kidney development (Figure 1C and Figure S2). However, we found no correlation between phospho-Yap staining and Yap localization in the RV or SSB stages.
To directly assess the function of Yap during nephron formation, we removed Yap from the CM with Six2:CreTGC/+[25]. Since all components of the nephron, from the glomerulus to the distal tubule derive from Six2-expressing CM cells, this system removes Yap from the CM and all of its epithelial derivatives (i.e. podocyte, Bowman's capsule, proximal tubule, Henle's Loop and distal tubule). We found that Six2:CreTGC/+ Yapflox/flox (termed YapCM−/−) newborns were obtained at Mendelian ratios. However, despite successful feeding, YapCM−/− animals died within 48 hours of birth. Gross anatomical examination revealed that neonatal (P0) YapCM−/− animals had hypoplastic kidneys and an empty bladder suggesting a failure to produce urine (Figure 1G, 1H). Histological examination of E18.5 kidney sections revealed a smaller papilla and a reduced nephrogenic zone. Convoluted renal tubules and glomeruli were not distinguishable in the inner cortex of the mutant, and the medulla was mainly composed of collecting ducts, suggesting a dramatic reduction in Henle's loop formation (Figure 1I–1L). YapCM−/− mutant kidneys had few detectable glomeruli and proximal tubules (Figure 1K–1N). Strikingly, the rare glomeruli observed in YapCM−/− mutants were ultrastructurally abnormal characterized by simplified capillary tufts ensheathed with podocytes having effaced foot processes (Figure S3).
Labeling with Dolichos Biflorus Agglutinin (DBA) lectin or Calbindin (Figure S4 and data not shown) confirmed that a branched collecting duct system was present as expected, as this structure does not derive from the CM. To determine which nephron compartment was affected by Yap inactivation, we used markers of distinct CM-derived nephron segments. Podocin staining labeled numerous glomeruli in wild-type kidneys, however, considerably fewer podocin-positive structures were detected in YapCM−/− at birth, consistent with the reduced number of glomeruli seen in histological analysis (glomeruli number per section at P0: control:50±5; YapCM−/−:6±2; ***p<0.001. Figure 2A, 2B). Furthermore, the few glomeruli observed in YapCM−/− mice had abnormal structures as seen by triple staining with podocin, WT1 and tomato-lectin (Figure 2C–2D′ and Figure S3). Yap and phospho-Yap antibodies failed to stain the Six2-positive compartments in YapCM−/− (Figure 1D, low and high magnification image panels are shown in Figure S1 and S2), suggesting that Yap excision was efficient, and that the rare nephron derivatives that form in mutants are likely not due to incomplete inactivation of Yap. Examination of markers at E18.5 revealed a dramatic loss of Lotus tetragonolobus lectin (LTL)-positive proximal tubule structures (Figure 2E, 2F). Interestingly, the morphogenesis of the remaining LTL-positive tubules was severely affected as they have barely discernable lumens at E18.5 (Figure 1M, 1N and Figure 2E, 2F). Staining for Ezrin, LTL and Par3 was normal in the residual tubules, indicating that cell polarity was retained (Figure S5). The reduced lumens may reflect an absence of filtration due to the dramatically reduced number of glomeruli. Strikingly, YapCM−/− kidneys also have defects in Henle's loop (Slc12a1) and distal tubule (Slc12a3) formation (Figure 2G–2J). Thus Yap is necessary in CM cells for normal nephron development.
The CM plays an essential role in supporting branching morphogenesis of the developing kidney. To determine if loss of Yap from the CM alters branching, we analyzed the number of ureteric tips at different time-points using immunofluorescent staining with antibodies to Calbindin, which marks both the CD and the UB tips (Figure S4). While similar branching is observed in wild-type and YapCM−/− kidneys at E14.5, the number of UB tips slightly decreases at E16.5 in Yap mutants with a significant reduction of tip number at P0. Thus loss of Yap in the CM does not affect early branching but has late-onset impairment of branching.
To determine the developmental origin of the defective nephrogenesis in YapCM−/−kidneys, we examined kidney development from E13.5 to E18.5. Nephrogenesis occurs in a repetitive manner, with new nephrons being formed throughout development at the outer cortex of the kidney. This process is highly regulated, involving both inductive and repressive signals (reviewed in [4]). Six2-expressing precursor cells residing in the cortex self-renew to replenish a pool of mesenchymal cells that are then transformed into nascent nephrons [25]. Maintenance of the progenitor population requires Six2, as deletion of Six2 results in premature differentiation of the CM cells [23]. Histological analysis revealed that YapCM−/− kidneys have limited nephrogenesis (Figure 2K, 2L) with abnormal morphogenesis of SSB (Figure 2M–2N′). CM cells in both YapCM−/− and wild-type kidneys were detected by histological analysis at E14.5 (Figure 2K–2N′) and by Six2, Gdnf and Sall1 expression (Figure 3A, 3B, 3F–3K), indicating that nephrogenic precursors cells are present in Yap mutant kidneys. Clear Six2 staining is obvious even at P0 (Figure 9W), although there is a mild reduction in the total number of Six2-positive cells in Yap mutant kidneys compared to wild-type kidneys (Figure S6). In contrast, however, the number of nascent nephrons (PA, RV, CSB and SSB) was clearly and dramatically reduced in Yap mutant kidneys early in development, as revealed by histological analysis (Figure 2K, 2L), NCAM staining (Figure 3C, 3D) and WT1 staining (Figure 3H, 3I). Quantification of NCAM-positive nephrogenic structures at E15.5 further validated a significant decrease in total nephrogenesis due to Yap deletion (Figure 3E). Since no change in branching morphogenesis could be detected at this stage (Figure S4), the limited nephrogenesis in Yap mutants is not secondary to impaired ureteric branching. Few PA could be detected in Yap mutant kidneys (Figure 3E), further showing that nephron induction is severely disrupted. In addition, the number of CM derived structures that reached the SSB stage in the mutant was dramatically reduced when compared to controls (Figure 3E). Thus, while the self-renewing capacity of CM cells is largely Yap independent, Yap-depleted CM cells are less potent to undergo nephrogenesis, and are unable to execute regulated morphogenesis to form regular SSB.
Formation of a functioning nephron requires polarization of the emerging epithelium along a proximodistal axis to specify diverse cell types. Polarization and segmentation is detectable as early as the RV stage. Segmentation becomes clearly apparent at the SSB stage with distal-, medial- and proximal-specific gene expression. We examined nephron segmentation at both the RV and SSB stages. Yap deletion did not impair RV polarization as proximal (WT1) and distal (E-cadherin, Hnf1ß, Sox9, Jag1) markers showed similar expression patterns (Figure 4A–4H). Similarly, in later nephrogenic structures, no segmentation defect could be seen in distal and medial SSB (Figure 4A–4H, Distal:E-cadherin, Hnf1ß, Sox9; Medial:Jag1, Hnf1ß, Sox9; Proximal:WT1 [26]–[28]). However, Yap-null SSB have a reduced WT1 positive proximal segment (in particular compare Figure 4C′ versus 4D′; 4E′ versus 4F′) consistent with defects in proximal fate seen in P0 kidneys.
Formation of a functional nephron also requires fusion to the ureteric bud, a process that occurs at the late RV stage [29]. Staining with Laminin to mark the basement membrane (BM) and Cytokeratin to mark the ureteric epithelium (UE) of an early RV shows that the RV is surrounded by its own BM and separated from the adjacent UE by the ureteric epithelial BM in both controls and YapCM−/− mutants (Figure 4I, 4J). At the comma stage (Figure 4I, 4J) fusion of the early nephron to the UE is complete in both genotypes as seen by a continuous BM. However, we note that YapCM−/− mutants consistently display aberrant morphology at the connecting segment, where the SSB connects to the UE (asterisk, see also Figure 3K, Figure 5F, Figure S2B). In particular, the distal segment of the SSB does not correctly merge with the outermost edge of the UB.
Yap, downstream of the Hippo pathway, has been extensively shown to regulate organ size by promoting cell proliferation and inhibiting apoptosis. We therefore analyzed cell proliferation throughout nephrogenesis to ascertain if altered proliferation or apoptosis could explain the morphological defects in Yap mutants. Quantification of BrdU incorporation in nephron progenitors cells (Six2 positive cells) did not reveal any significant changes in CM proliferation (n = 1,000 Six2 positive cells from 4 kidneys of each genotype, Figure 5A, 5B and 5G). Interestingly, while no significant changes could be detected in overall RV proliferation, or in distal RV proliferation, a slight reduction in proliferation was detectable in the proximal part of Yap-null RV (Hnf1ß negative, n = 16 RV per genotype - Figure 5C, 5D and 5G). Finally, we investigated proliferation in the distal (cells located between the UE and Jag1 expressing domain), medial (Jag1 positive cells) and proximal segments of the SSB (n = 12 SSB per genotype). Similarly to the RV stage, no significant change in proliferation could be detected in the overall SSB, however segment-specific analysis revealed slightly decreased proliferation in the distal segment of Yap mutant (Figure 5E, 5F and 5G). TUNEL staining in control and YapCM−/− kidneys (E18.5) did not reveal any changes in apoptosis in mutants relative to controls (Figure 5H, 5I). Our data indicates that early defects in nephron formation in Yap mutants are not due to death of the nephrogenic cell population, nor to a dramatic failure to proliferate.
Recent studies have revealed functional interactions between Yap and ß-catenin, [18], [30]. While activation of the canonical ß-catenin signaling pathway is necessary for nephron formation, its repression is required for epithelialization to occur [31]. Wnt9b secreted from the UB induces mesenchymal condensation via canonical ß-catenin signaling, activating a molecular cascade involving Fgf8, Wnt4, Pax8 and Lim1 [32]. Expression of Wnt9b was unchanged in Yap mutant kidneys (Figure S7A, S7B). To see if ß-catenin signaling was affected in Yap mutant kidneys, we examined expression of ß-catenin signaling targets. Significantly, expression of the established Wnt target genes Pla2g7, C1qdc2 and Lef1 were unchanged in YapCM−/− kidneys (Figure S7C–S7H). Moreover, Fgf8, Wnt4, Pax8, and Lim1 expression levels were also unchanged in YapCM−/− kidneys (Figure S7I–S7P). Finally, removing one allele of ß-catenin in YapCM−/− mice (by generating Six2:CreTGC/+ Yapflox/flox ß-catenin(KO)flox/+ embryos) does not alter the YapCM−/− phenotype (Figure S8). Taken together, these data indicate that Yap functions in nephron formation independently of major changes in the Wnt/ß-catenin signaling pathway.
The defects in glomeruli and proximal tubules that occur in YapCM−/− kidneys are reminiscent of defects in Notch signaling [27]. We therefore assayed different members of the Notch pathway by in situ hybridization (ISH). In particular, no changes were detected in the expression levels of Notch1, Notch2, the ligand Jagged1, or the Notch targets Hes1 or Hes5 (Figure S7Q–S7X and data not shown). These data indicate that the loss of Yap does not lead to nephrogenesis defects via loss of Notch signaling.
Since no defects were observed in ß-catenin or Notch signaling, and proliferation and apoptosis were largely unaffected, we sought an unbiased approach to determine the molecular basis of the defects seen in Yap mutants. We used whole-genome transcript profiling (Mouse Whole Genome-6 v2.0 BeadChip) at E13.75 to determine gene expression changes in Yap mutant kidneys. Of the ∼45,000 transcripts represented on the array, 334 genes were found to be differentially expressed in YapCM−/− kidneys (fold change>1.27, p-value<0.05). We used both Genepaint (www.genepaint.org) and Gudmap (www.gudmap.org) databases to examine candidate expression in the developing kidney. This approach allowed us to concentrate on 24 candidates (Table S1). To confirm changes in YapCM−/− mutants, we performed ISH and antibody staining in E14.5 control and YapCM−/− kidneys. Our analysis confirmed that expression of Cited1, Meox2, Traf1 and Capn6 was lost in Yap-null CM cells (Figure 6D–6O). Similarly, expression of Pax2, Uncx4.1 and Sostdc1 were significantly reduced in Yap mutants (Figure 6A–6C, and Figure S9A–S9F). While Fgf10 expression was barely detectable in wild-type CM cells, strong mesenchymal expression of Fgf10 was observed in YapCM−/− kidneys (Figure 6P–6R). Surprisingly, expression of both Ret and Raldh3 was greatly increased respectively in UB tips and collecting ducts of Yap knockout kidneys, indicating that loss of Yap in the CM non-autonomously affects expression of these genes (Figure S9G–S9L). This work identifies a set of genes that depend on Yap expression during nephron development that function in differentiation and morphogenesis rather than proliferation and apoptosis.
Staining with antibodies to phospho-Yap did not indicate any obvious spatial or temporal regulation by Hippo kinases that could explain the regulation of Yap localization or activity during nephrogenesis (Figure 1B, 1E, 1F and Figure S1). We therefore searched for other potential regulators of Yap activity. Recent studies in cultured mammalian cells have demonstrated that Yap can be regulated in a Hippo kinase independent manner by mechanical signals exerted by extracellular matrix rigidity and cell shape [21]. Mechanical signals regulate Yap localization via small GTPase activity and the actin cytoskeleton. Cdc42 is a conserved and critical regulator of the actin cytoskeleton, acting through Arp2/3 and N-Wasp [33]. To examine the role of Cdc42 in nephrogenesis, we used Six2:Cre to delete Cdc42 from the CM population (Cdc42CM−/−). Loss of Cdc42 from the CM resulted in a severe defect in kidney development that was strikingly similar to YapCM−/−, with hypoplastic kidneys with empty bladders indicating lack of functional nephrons (compare Figure 7A, 7B to Figure 1G, 1H). The histology of E18.5 Cdc42CM−/− kidneys strikingly resembles that of YapCM−/− with a distinctively reduced nephrogenic zone and a smaller papilla (Figure 7C, 7D). Convoluted renal epithelia and glomeruli were absent in the cortex of the mutant (Figure 7E, 7F). Staining with Podocin and quantification of glomeruli demonstrated a significant reduction in glomerular number in Cdc42CM−/− (glomeruli number per section at P0: control:51±2; Cdc42CM−/−:4±2; ***p<0.001), similar to that seen in YapCM−/− kidneys (Figure 7G, 7H). Cdc42CM−/− kidneys also have fewer and truncated proximal tubules with barely discernable lumens (Figure 7I, 7J). Similar to YapCM−/−, nephrogenic precursors are present in Cdc42CM−/− (seen by PAS staining, Six2, Sall1 and WT1 expression; Figure S10A–S10H), but the capacity of these cells to undergo nephrogenesis is dramatically reduced (NCAM staining - Figure 7K, 7L and Figure S10A, S10B). Together these data show a remarkable similarity between the effects of loss of Yap and the loss of Cdc42 in the CM, suggesting they might function together in kidney development.
The primary mechanism of regulating Yap activity is controlling Yap nuclear localization. We therefore tested if loss of Cdc42 affected Yap nuclear localization in developing kidneys. Detailed examination of Cdc42CM−/− kidneys revealed reduced nuclear Yap in Six2 positive CM cells at E12.5 (Figure 8A–8B′″). Quantification using ImageJ software further confirmed a significant decrease of nuclear Yap in mutant CM cells compared to wild-type, while no change in the levels of Yap were observed in the nuclei of adjacent UB cells (Figure 8E).
The small GTPase RhoA has been shown to regulate Yap nuclear localization in mammalian tissue culture [21], [22], however no studies to date have examined the effects of Cdc42 on Yap localization. To better visualize changes in Yap localization upon removal of Cdc42, we examined cultured mouse embryonic fibroblasts (MEFs) isolated from E13.5 Cdc42flox/flox embryos. Cdc42flox/flox MEFs were infected with an adenovirus expressing Cre. Yap staining is predominantly in the nucleus in isolated control MEFs (Figure 8C–8C′″), while removal of Cdc42 in MEFs results in more diffuse Yap staining, with reduced nuclear accumulation (Figure 8D–8D′″, lower magnification in Figure S11). Thus, loss of Cdc42 in MEFs, as in embryonic kidneys, leads to a decrease in nuclear Yap, indicating that Cdc42 function is necessary for Yap to be normally localized in the nucleus.
The remarkable phenotypic similarities of YapCM−/− and Cdc42CM−/−, coupled with the observation that loss of Cdc42 leads to reduced levels of nuclear Yap, suggested the hypothesis that Cdc42 is necessary for Yap-dependent gene expression. We tested this hypothesis by examining expression of Yap-dependent genes in Cdc42CM−/−, by immunofluorescence and ISH. Staining of E14.5 Cdc42CM−/− kidneys revealed dramatic loss of Cited1, Capn6, and Traf1, and a clear reduction of Pax2, Uncx4.1 and Meox2 (Figure 8F–8M and Figure S10I–S10L). In addition, there were marked increases in Fgf10 expression in Cdc42CM−/− mutants (Figure 8N–8O). All these changes mimicked the changes seen in Yap mutants (Figure 6). Thus, defective nuclear Yap localization in Cdc42 mutants results in a loss of Yap-dependent gene expression. These data support a model where Cdc42 function is necessary for proper Yap localization and activity to control early nephron formation.
As described above, YapCM−/− mice have dysplastic kidneys with minimal nephrogenesis. The Yap paralogue Taz is required for proper kidney development since Taz−/− mice have cystic kidneys [13],[14].
To investigate the function of Taz in the Six2 progenitor cells, we generated Six2:CreTGC/+ Tazflox/flox mice (TazCM−/−). Macroscopic analysis of TazCM−/− urogenital systems shows functional kidneys (bladder filled with urine) similar in size to controls, with spotty hemorrhages at P0 (Figure 9B). Histology at P0 reveals highly cystic tubules in the cortex of TazCM−/− kidneys (Figure 9F, 9J), similar to Taz−/− mutants. Importantly, neither the number of Six2-positive progenitor cells present at birth (Figure 9V) nor the mesenchymal-epithelial transition (NCAM - Figure 9R) were changed in TazCM−/− mutants. Furthermore, the number of glomeruli (data not shown) remained unaltered in the TazCM−/− mutants when compared to littermate controls indicating that Yap and Taz have distinct functions during the process of nephron formation.
To determine if the residual glomeruli and proximal tubules that form in YapCM−/−kidneys are due to a low level redundancy by Taz, we generated Six2:CreTGC/+ Tazflox/flox Yapflox/flox mice. Significantly, TazCM−/−;YapCM−/− double mutants showed no exacerbation of glomeruli or proximal tubules deficits relative to Yap mutant kidneys (Figure 9H–9P and data not shown). However, some of the few proximal tubules that formed were cystic (Figure 9L, 9P), similar to Taz single mutants. Together, these data indicate that Yap and Taz play distinct roles during nephrogenesis.
Yap and Taz have well-described roles in the regulation of cell proliferation and apoptosis. Here we show that loss of Yap leads to defects in nephron formation and morphogenesis during renal development. We demonstrate that these defects occur independent of major changes in apoptosis or proliferation, and identify a novel set of Yap-dependent genes implicated in morphogenesis. We further show that Cdc42 function is necessary for Yap to be correctly localized in developing CM cells and in cultured MEFs, and that loss of Cdc42 disrupts nephron formation and abolishes Yap-dependent gene expression. We propose a model in which Yap localization occurs in response to Cdc42-dependent signals, leading to expression of Yap-dependent genes during nephron development.
Yap and Taz are closely related transcriptional co-activators that have been shown in many systems to have similar, and at times partially redundant roles in control of cell proliferation and apoptosis. A striking finding of our in vivo analysis is that loss of Yap leads primarily to misregulation of genes involved in cell fate and morphogenesis. Another surprising finding in our study was the discovery that Yap and Taz play distinct roles during kidney development. While loss of Yap leads to reduced nephrogenesis, with clear morphological defects at the SSB stage, loss of Taz leads to normal sized kidneys, with functioning nephrons, as indicated by a full bladder at birth. Proximal tubules in the Taz mutants are cystic, while the proximal tubules in Yap mutants have barely discernable lumens. Moreover, TazCM−/−;YapCM−/− double mutants show both loss of glomeruli and proximal tubules, with some tubules becoming dramatically dilated - underscoring the independence of the Taz and Yap phenotypes.
Surprisingly, we did not detect any significant changes in proliferation or apoptosis in Yap mutants indicating that in nephrogenesis Yap is functioning independently of previously described roles. We did not detect any spatial regulation of Hippo-dependent Yap phosphorylation. Moreover, we found that both Sav1 and Mst1/2 (Pax3:Cretg/+ Mst1−/− Mst2flox/flox) knockout kidneys were superficially normal (Figure S12). Further studies are needed to fully ascertain the contribution of Mst and Lats kinases to nephron development.
Although Yap is not essential for proliferation or apoptosis in early nephron development, we cannot exclude a later role for Yap in cell proliferation in the tubules. In fact the extremely short tubule segments seen in Yap mutants may reflect a role for Yap in later proliferation. Consistent with this possibility, we found that forced overexpression of Yap in adult kidneys leads to increased proliferation (data not shown). This later role of Yap during nephrogenesis may be controlled by the Hippo pathway, but further examination is required to examine this possibility.
Using microarray analysis, we identified a number of novel Yap-dependent genes that function in nephrogenesis. Unexpectedly, these genes were not involved in control of apoptosis and proliferation, but instead involved in cell fate and morphogenesis. A subset of these genes are involved in controlling cell shape and the cytoskeleton. Capn6, for example, bundles and stabilizes microtubules [34]. Other genes, such as Sostdc1 and Fgf10 are involved in cell-cell signaling, whereas others such as Meox2 and Cited1 are markers of the early steps of nephrogenesis, and linked to stem cell renewal. Yap has been previously linked in multiple systems with stem cell proliferation and stem cell pluripotency ([16], [35], reviewed in [23]). Since Six2 expression is largely unaltered in Yap mutants, the observed defects lie downstream of commitment to a progenitor fate, but upstream of nephron formation.
Interestingly, of the genes we identified as Yap-dependent in nephrogenesis, only Pax2 is absolutely required for kidney development [36]. Thus, there is likely redundancy in the morphogenesis mediated by Yap-dependent gene expression, and more complex approaches, such as double or triple knockout may be needed to uncover the critical gene programs needed for each step of nephron formation.
While removal of Yap leads to a dramatic loss of Cited1, Meox2, Capn6 and Traf1, and reduction of Pax2 and NCAM, removal of Yap has little effect on Six2, Sall1, WT1 or Gdnf expression in the CM. This indicates that Yap deletion does not result in a loss of the CM cell population. Instead Yap is needed for proper CM differentiation. Interestingly, expression of the stromal markers Raldh2 and Foxd1 was unaltered in Yap mutants, suggesting that observed changes in the CM of Yap mutants are not due to a loss of the CM cells to the stromal lineage (Figure S13). Moreover, loss of Yap does not primarily affect survival or proliferation of nephrogenic precursors, as no change in apoptosis and BrdU incorporation could be seen in Six2-positive cells. Thus this analysis indicates that genes required for progenitors cell survival and self-renewal are separable from those involved in their differentiation.
Some MET occurs even in the absence of Yap, but functional nephrons failed to form, with defects clearly visible at the SSB stages. We found that there is a dynamic pattern of nuclear Yap during early nephrogenesis. Yap localizes mainly to the nucleus in the proximal RV and the most distal and proximal cells of the SSB. Additional analysis using early nephron-specific Cre lines will be required to distinguish if changes in gene expression of progenitors cells are responsible for defective nephron formation or if Yap is needed for proper gene expression in RV and SSB.
In parallel to nephron formation, signals from the CM also promote branching morphogenesis. Our analysis of branching morphogenesis revealed that branching in Yap mutants slowed at E16.5 resulting in a 30% decrease in UB tips at birth. The increased Fgf10 and Ret may partially compensate for the loss of CM genes seen in Yap mutants to maintain branching morphogenesis.
We found that Cdc42 function is necessary for Yap to be localized normally both in vivo and in vitro, and that loss of Cdc42 phenocopies loss of Yap, and abolishes Yap-dependent gene expression. How could Cdc42 affect Yap localization and activity? Cdc42 has conserved roles in the regulation of cell polarity and in the organization of the actin cytoskeleton. We found that neither loss of Yap nor Cdc42 alters cell polarity in the developing nephron, suggesting that Cdc42-dependent cell polarity is not the primary mechanisms to control Yap localization. Instead, we favor a model in which the loss of Cdc42 affects Yap localization via alterations in the cytoskeleton. Yap relocalization by mechanical stresses has been shown to be dependent on an intact actin cytoskeleton. Cdc42 regulates the actin cytoskeleton, in part via Neuronal-Wiskcott Aldrich Syndrome protein (N-WASP) [37]. Interestingly, we found that N-WaspCM−/− kidneys (Figure S14) were hypoplastic, with significant loss of glomeruli and proximal tubules reminiscent of the defects of Yap and Cdc42 deficient kidneys.
Studies in cultured mammalian cells have demonstrated that Yap nuclear localization is regulated by mechanical signals exerted by extracellular matrix rigidity and cell shape [21]. Notably, recent studies have shown that Cdc42 is essential for matrix contraction in 3D tissue culture assays. Loss of Cdc42 may result in defects in nephrogenesis due to loss of cytoskeletal tensions between the matrix and the tubules as they form, twist and bend during nephrogenesis. We propose that the dramatic changes in cell shape as cells aggregate, epithelialize and contort during formation of nephrons, generates mechanical stresses that are sensed via the cytoskeleton, leading to changes in the nuclear localization of Yap. Once in the nucleus, Yap then promotes expression of genes that are necessary for subsequent steps in nephron formation. While this idea is appealing, clearly much more work is needed to understand how loss of Cdc42 leads to disruption of Yap localization, and changes in Yap-dependent gene expression.
We have shown here that loss of Cdc42 leads to loss of Yap-dependent gene expression and loss of nuclear Yap localization. Taken together, our data suggest a model in which Cdc42 function is necessary for Yap localization and activity during development to shape functioning nephrons.
Cdc42flox [38], Mst1flox and Mst2−/− [39], N-Waspflox [40], Pax3:Cretg/+ [41], Sav1−/− [17] and Six2:CreTGC/+ [25] mouse strains have been described. Yapflox and Tazflox alleles were generated by inserting LoxP sites for Cre-mediated excision flanking exons 2 as described in Figure S15. All mice were maintained on a mixed genetic background. Husbandry and ethical handling of mice were conducted according to guidelines approved by the Canadian Council on Animal Care. Embryos were genotyped by standard PCR protocol. Genotyping was done by PCR using genomic DNA prepared from mouse ear punches.
Embryonic samples from timed matings (day of vaginal plug = E0.5) were collected, fixed in 4% paraformadehyde overnight at 4C, serially dehydrated and then embedded in paraffin. Microtome sections of 7 µm thickness were examined histologically via periodic acid-Schiff staining.
For immunofluorescent analysis, paraffin sections were dewaxed and re-hydrated via ethanol series. Antigen retrieval was performed by boiling the sections for 20 minutes in Antigen Unmasking Solution (H-3300, Vector). Sections were incubated for 1 hour in blocking solution (3% BSA, 10% goat serum, 0.1% Tween20 in PBS) at room temperature. Blocking solution was replaced by a solution of primary antibodies diluted in 3% BSA, 3% goat serum, 0.1% Tween20 in PBS. The following primary antibodies were used in this study: Calbindin (PC253C, Calbiochem), Cited1 (RB-9219-P0, Neomarkers), Cytokeratin (F3418, Sigma), E-cadherin (Mouse, 610181, BD Transduction Laboratories), E-cadherin (Rabbit, #3195, Cell Signaling Technology), Ezrin (sc-58758, Santa Cruz Biotechnology), FoxD1 (gift from Andrew P. McMahon (Harvard University, Cambridge), Hnf1ß (sc-2280, Santa Cruz Biotechnology), Jag1 (#2620, Cell Signaling Technology), Laminin (L9393, Sigma), LTL (FL-1321, Vector Laboratories), NCAM (C9672, Sigma), Par3 (07-330, Millipore), Pax2 (PRB-276P, Covance), Phospho-Yap (#4911, Cell Signaling Technology), PKC (sc-216, Santa Cruz Biotechnology), Podocin (P0372, Sigma), Six2 (11562-1-AP, Proteintech), Sox9 (AB5535, Chemicon), Tomato-lectin (TL-1176, Vector Laboratories), WT1 (C-19, Santa Cruz Biotechnology), Yap (sc-101199, SantaCruz Biotechnology), Yap/Taz (#8418, Cell Signaling Technology). Relevant Cy3- or FITC-conjugated secondary antibodies (Jackson Laboratories) were used for primary antibody detection. Slides were mounted using Vectashield with or without DAPI (Vector Labs). Fluorescent images were taken with a Nikon C1 plus Digital Eclipse confocal microscope.
For immunohistochemistry, the same procedure was used, with the addition of one step after the re-hydration. Slides were immersed in 3% H2O2 in PBS for 20 minutes to block endogenous peroxidases. The Yap/Taz antibody was incubated for 48 hours at 4 degrees. Then undiluted secondary antibody (EnvisionPlus from Dako) was applied to the sections for 1 hour at room temperature. Samples were washed, developed with DAB, counterstained with hematoxylin and mounted in pertex.
P0 kidneys were dissected in PBS, fixed in 4% paraformaldehyde, embedded in paraffin and sectioned. Immunostaining against Podocin was performed on P0 median kidney sections and glomeruli were counted. Means were calculated per kidney and genotype. An unpaired two-tailed t-test was used to determine the statistical significance among genotypes.
Quantification of Yap nuclear staining was performed using Image J software. Images were imported into Image J, and, by using DAPI staining to mark the cell nuclei, nuclear Yap signal was measured. The mean signal was calculated from 100 cells for each genotype. An unpaired two-tailed t test was used to determine the statistical significance.
BrdU solution containing 5-Bromo-2′-deoxyuridine (10 mg/ml) was injected intraperitoneally in pregnant mice (50 mg BrdU/kg of mice) 2 to 3 hours before embryonic dissection. The samples were prepared and sectioned as described above before being incubated overnight with anti-mouse BrdU antibody (Clone Bu20a, Dako).
Terminal deoxynucleotidyl transferase, mediated digoxigenin-deoxyuridine nick end labeling (TUNEL) was performed using the Roche Cell Death Detection Kit on E18.5 kidney sections.
Embryos were fixed in 4% paraformaldehyde in PBS overnight at 4C, and then paraffin embedded. Further processing of the embryos and ISH were carried out as described [42]. Riboprobes for Capn6 [34], Raldh3 [43], Fgf10 [44], Ret (gift from Frank Costantini), Slc12a1 and Slc12a3 (gift from A. Brandli), Meox2, Traf1 and Uncx4.1 (from the SLRI Open Freezer) were used [45].
We used IlluminaMouseWG-6 v2.0 Expression BeadChips for whole-genome expression profiling. Pregnant mice were sacrificed at E13.75, embryos collected in ice-cold PBS and immediately decapitated. Kidneys were quickly removed and flash frozen in liquid nitrogen and stored at −80°C until RNA extraction. Both kidneys from one embryo were pooled to make one mutant or control sample. RNA was extracted using RNeasy Mini Kit (Qiagen) according to the manufacturer's protocol. The quality and quantity of collected RNA samples was checked (Agilent, Bioanalyzer) prior to submitting samples for microarray analysis. Expression profiling of three mutants and three controls samples and data analysis were done at the UHN Microarray Centre in Toronto.
Dissected kidneys were fixed in 0.1 M cacodylate buffer containing 4% paraformaldehyde and 2% glutaraldehyde. Subsequently, P0 kidneys were postfixed in 1% OsO4, dehydrated, and embedded in Quetol-spurr resin. Ultrathin resin sections stained with uranyl acetate and lead citrate were viewed using an FEI CM100 transmission electron microscope (FEI, Hillsboro, OR).
Mouse embryonic fibroblasts (MEFs) were derived from E13.5 embryos homozygous for the floxed Cdc42 allele and maintained in 10% fetal bovine serum supplemented DMEM. MEFs were seeded into 8-well glass culture slides (BD Falcon, Bedford, MA) that were precoated with 200 µg/mL polethyleneimine to promote cell adhesion. To establish Cdc42 null MEFs, cells were infected with Cre-expressing adenovirus (Vector Biolabs) at MOI of 100. Control (uninfected) and Cdc42 null MEFs were serum-starved overnight and stimulated with serum-containing medium for 2 h and subsequently fixed in chilled 4% PFA in PBS for 10 min. Fixed cells were permeabilized for 2 min with 0.3% Triton X-100 in PBS and stained with an anti-Yap monoclonal antibody (Santa Cruz Biotechnology, Santa Cruz, CA) followed by an Alexa-Fluor488 anti-mouse secondary antibody conjugate. Cells were doubly counterstained with Texas-Red conjugated phalloidin and the DNA-binding dye Hoechst 33258. The efficiency of Cdc42 excision was assessed by western blot of MEF lysates probed with a Cdc42 antibody (sc-8401, Santa Cruz Biotechnology) and an α-tubulin (DM1A, Sigma Aldrich).
|
10.1371/journal.pbio.1000107 | Aberrant Herpesvirus-Induced Polyadenylation Correlates With Cellular Messenger RNA Destruction | Regulation of messenger RNA (mRNA) stability plays critical roles in controlling gene expression, ensuring transcript fidelity, and allowing cells to respond to environmental cues. Unregulated enhancement of mRNA turnover could therefore dampen cellular responses to such signals. Indeed, several herpesviruses instigate widespread destruction of cellular mRNAs to block host gene expression and evade immune detection. Kaposi's sarcoma-associated herpesvirus (KSHV) promotes this phenotype via the activity of its viral SOX protein, although the mechanism of SOX-induced mRNA turnover has remained unknown, given its apparent lack of intrinsic ribonuclease activity. Here, we report that KSHV SOX stimulates cellular transcriptome turnover via a unique mechanism involving aberrant polyadenylation. Transcripts in SOX-expressing cells exhibit extended poly(A) polymerase II-generated poly(A) tails and polyadenylation-linked mRNA turnover. SOX-induced polyadenylation changes correlate with its RNA turnover function, and inhibition of poly(A) tail formation blocks SOX activity. Both nuclear and cytoplasmic poly(A) binding proteins are critical cellular cofactors for SOX function, the latter of which undergoes striking nuclear relocalization by SOX. SOX-induced mRNA turnover therefore represents both a novel mechanism of host shutoff as well as a new model system to probe the regulation of poly(A) tail-stimulated mRNA turnover in mammalian cells.
| During viral infection, many essential cellular functions are targets for viral manipulation, yet aside from RNA interference, surprisingly few examples of viruses disrupting RNA turnover have been documented. Kaposi's sarcoma-associated herpesvirus (KSHV) is an oncogenic virus that induces widespread cellular messenger RNA destabilization during lytic infection. The viral protein SOX is a critical effector of this phenotype, yet it lacks ribonuclease activity, so presumably it targets cellular factors governing RNA stability. Here, we show that SOX stimulates host mRNA destruction via a unique mechanism involving polyadenylation. During SOX expression, newly formed messages have longer than normal poly(A) tails, leading to their retention in the nucleus. Coincident with this hyperadenylation, poly(A) binding protein (PABPC) is relocalized from the cytoplasm to the nucleus. PABPC has prominent roles in translation, messenger RNA stabilization, and quality control in the cytoplasm; we find its nuclear relocalization by SOX correlates with enhanced mRNA turnover in the cytoplasm. Thus, KSHV appears to have evolved distinct polyadenylation-linked mechanisms to target both new messages in the nucleus and preexisting cytoplasmic messages for destruction, thereby effectively inhibiting cellular gene expression.
| Kaposi's sarcoma-associated herpesvirus (KSHV) is the most recently discovered human herpesvirus and the etiologic agent of several neoplasms, the most prominent of which is Kaposi's sarcoma (KS) [1]. Originally described as a rare tumor found predominantly in elderly Mediterranean or African men, with the onset of the AIDS pandemic, KS became the most common neoplasm associated with untreated human immunodeficiency virus (HIV) infection. KSHV is a large double-stranded DNA virus that undergoes both latency and lytic replication. Although the majority of infected cells in vitro and in vivo harbor the virus in a latent state, the lytic cycle is required both for viral replication and KS development [2],[3]. One striking feature of lytic KSHV infection is that it destroys the host transcriptome by promoting global messenger RNA (mRNA) degradation via unknown mechanisms [4],[5]. The magnitude of cellular transcript loss is significant; nearly 75% of all messages are massively down-regulated, with another 20% undergoing a more modest decrease [6],[7]. This phenotype, termed host shutoff, is mediated by the viral factor SOX (shutoff and exonuclease) which has homologs across the entire herpesvirus family [5]. In other herpesviruses, this protein has DNA exonuclease (DNase) and recombinase activities that contribute to processing and packaging the newly replicated viral genomes in the nucleus, but has no role in mRNA turnover [8]–[10]. By contrast, in KSHV and its closest viral relatives within the lymphotrophic γ-herpesviral subfamily—including the human cancer-associated Epstein-Barr virus—SOX retains these conserved functions but has evolutionarily acquired a novel and distinct role in global mRNA decay [11],[12]. The host shutoff and DNA processing functions of SOX are genetically separable, as single-function point mutants can dissociate the two activities [4]. Despite its ability to induce widespread mRNA destruction, KSHV SOX has neither homology to known ribonucleases nor predicted RNA recognition motifs, and thus far no intrinsic ribonuclease (RNase) activity has been detected with the purified protein. SOX is therefore presumed to function by modulating one or more cellular RNA turnover pathways.
Control of message stability obviously represents a powerful means of regulating gene expression both on an individual and a global scale. Nearly all eukaryotic mRNAs are protected from exonucleolytic attack by a 5′ cap structure and a 3′ poly(A) tail. Cleavage and polyadenylation are cotranscriptional events, and their successful completion is required to signal formation of an export competent message. Poly(A) site recognition is mediated by specific sequence elements bound by the cleavage factors CPSF, CtsF, and CFIm [13]–[15]. Poly(A) polymerase (PAP) is recruited to the complex during the cleavage reaction and initiates polymerization of the adenosine tract in a biphasic manner; initial slow distributive adenosine addition proceeds until a sufficiently long tail has been formed to allow binding of nuclear poly(A) binding protein (PABPN), then rapid polymerization of the remaining 200–250 nucleotides (nt) ensues, whereupon PAP reverts to a distributive mode and dissociates from the transcript [16]. Upon nuclear export, PABPN is replaced by the cytoplasmic poly(A) binding protein (PABPC), which enhances mRNA stability and translation efficiency, in part through its interactions with the eIF4G translation initiation factor [17],[18].
In eukaryotes, polyadenylation generally serves to stabilize mRNAs, whereas in bacteria and some organelles, it facilitates RNA turnover [19]. However, it is now becoming clear that polyadenylation can be a facilitator of eukaryotic mRNA degradation as well. In particular, yeast possess a nuclear polyadenylation complex (TRAMP) that marks aberrantly processed RNAs for quality control-mediated turnover via the addition of short poly(A) tails [20]–[22]. Additionally, many yeast mutants defective in RNA processing or export accumulate hyperadenylated transcripts, suggesting a link between the polyadenylation process and RNA surveillance and turnover [23]–[25]. Analogous pathways may function in higher eukaryotes, as polyadenylated precursors to RNA turnover have also been detected in mammalian cells [26]–[28].
Here, we reveal that KSHV SOX-induced host shutoff is intimately linked to mRNA polyadenylation. SOX promotes aberrant polyadenylation of cellular transcripts in a manner dependent on its RNA turnover activity. Transcript degradation requires both nuclear and cytoplasmic poly(A) binding proteins, the latter of which undergoes striking nuclear relocalization in a host shutoff-dependent manner. In the absence of cellular 3′ end processing and polyadenylation, SOX can no longer target mRNAs for destruction, although addition of a templated poly(A) tail reinstates SOX-induced turnover. These findings suggest SOX is directing a novel polyadenylation-dependent mechanism of host shutoff, and demonstrate a link between polyadenylation and mRNA destruction in higher eukaryotes.
The KSHV SOX protein is found both in the nucleus and in the cytoplasm of cells, whereas its herpesviral homologs lacking mRNA turnover activity are restricted to the nucleus [4]. We considered that this distinct localization pattern could play a role in its mRNA degradation function, and therefore evaluated whether blocking CRM1-dependent nuclear protein export using the drug leptomycin B (LMB) could restrict SOX to the nucleus and alter its function. Although LMB treatment significantly increased the population of SOX in the nucleus (Figure S1), it did not completely eliminate the cytoplasmic fraction and did not abrogate the mRNA turnover activity of SOX (Figure 1A–1C). However, upon LMB treatment, a slower-migrating population of the reporter green fluorescent protein (GFP) mRNA appeared specifically in the presence of SOX, indicative of some form of SOX-induced RNA modification (Figure 1A). To gain insight into the specificity of this SOX activity, we tested a panel of SOX mutants lacking only the conserved DNase activity associated with viral genome processing or lacking only the mRNA turnover activity responsible for host shutoff [4],[5]. The production of these slower-migrating species correlated very strongly with the host shutoff function of SOX; they were not observed in cells expressing SOX mutants defective for mRNA degradation (T24I, P176S, L20/23A), but they were produced in cells expressing a SOX mutant (Q129H) lacking only the conserved DNase activity (Figure 1B). Furthermore, expression of the SOX homolog from herpes simplex virus (HSV AE) that exhibits DNase activity [8],[29], but has no role in host shutoff [5], also has no effect on the reporter mRNA mobility (Figure 1B). Thus, we conclude that this RNA modification correlates with the RNA turnover activity of SOX responsible for host shutoff.
One obvious mRNA modification that could significantly alter message size is polyadenylation. To determine whether the altered mRNA mobility was due to extended poly(A) tails (hyperadenylation), we investigated whether deadenylation of the messages by oligo(dT) hybridization followed by RNaseH digestion would eliminate their size differences. Indeed, northern blotting revealed that poly(A) tail removal caused the high MW mRNA from SOX-expressing cells to shift down in size, such that it precisely co-migrated with the ‘normal’ mRNA (Figure 1C).
Although LMB treatment may somehow stabilize the hyperadenylated mRNA species thereby facilitating their detection by northern blotting, it was important to confirm both that this modification also occurs in untreated cells and on endogenous cellular transcripts. To this end, total endogenous poly(A) RNA accumulation was measured by in situ hybridization of HEK 293T cells with a fluorescently labeled oligo(dT) probe (Figure 1D). Significantly, all wild-type (WT) SOX-expressing cells contained elevated levels of endogenous nuclear poly(A) RNA, as visualized by enhanced oligo(dT) staining. Accumulation of the poly(A) RNA specifically in the nucleus can be seen most clearly in the higher magnification images (Figure 1D, far right). Biochemical fractionation studies also show that the hyperadenylated mRNA is absent from the cytoplasmic fraction of cells (Figure S3). SOX single-function mutants lacking host shutoff activity, such as P176S and the HSV SOX homolog (AE) that possesses only DNase activity, fail to increase cellular poly(A) RNA levels (Figure 1D and unpublished data). These observations indicate that hyperadenylation is widespread on endogenous messages in SOX-expressing cells. Thus, although polyadenylation has traditionally been viewed as a stabilizer of eukaryotic transcripts, our data indicate that it is associated with mRNA destruction in the presence of SOX.
Three poly(A) polymerase proteins with molecular masses of 90, 100, and 106 kDa have been identified in HeLa cell nuclear extracts [30]. The 106-kDa isoform is likely a phosphorylated version of the 100-kDa isoform, and collectively, they are referred to as PAPII, whereas the 90-kDa protein, termed PAP-γ, is the product of a distinct locus [31]. A testes-specific PAP has also been identified [32],[33], but was not examined here due to its tissue-restricted expression. The contribution of PAPII and PAPγ towards SOX-induced hyperadenylation was assessed using small interfering RNA (siRNA)-based knockdown of each protein and measuring the resulting effects on endogenous poly(A) RNA accumulation by oligo(dT) in situ hybridization (Figure 2). Western blotting confirmed efficient knockdown of PAPII and PAPγ upon transfection of two independent siRNA oligo pairs (Figure 2A). We consistently observed that SOX-induced hyperadenylation was diminished in the absence of PAPII, whereas no decrease in the oligo(dT) signal in SOX-expressing cells was detected upon PAPγ knockdown (Figure 2B). Thus, hyperadenylation in SOX-expressing cells is mediated by the canonical PAP responsible for the majority of mRNA poly(A) tail synthesis. If hyperadenylation by PAPII participates in mRNA turnover by SOX, we would predict that inhibition of PAPII might stabilize mRNAs in SOX-expressing cells. We therefore monitored GFP mRNA turnover by SOX upon siRNA-mediated knockdown of PAPII or PAPγ (Figure 2C). Indeed, PAPII knockdown increased GFP mRNA levels in the presence of SOX, as well as increased the mobility of the GFP mRNA in the presence of SOX (Figure 2D, compare lane 4 to lanes 2 and 6; quantification shown in Figure S2). By contrast, GFP mRNA was still efficiently degraded and hyperadenylated upon knockdown of PAPγ. These data therefore suggest that PAPII-induced hyperadenylation plays an important role in SOX-induced host shutoff.
While PAPII mediates poly(A) tail formation, its activity is greatly stimulated by the nuclear poly(A) binding protein PABPN, which has also been proposed to help mediate poly(A) tail-length control [16]. Once synthesized, poly(A) tails are immediately coated with PABPN, which remains bound to the mRNAs until their transport into the cytoplasm, whereupon PABPN is replaced by cytoplasmic PABP (PABPC). PABPC effectively circularizes mRNAs by virtue of its interaction with eIF4G, an event that both protects the mRNA ends from exonucleolytic attack and enhances message translation via the closed loop model [17],[18]. Although both bind poly(A) tails, PABPN and PABPC have distinct functions and subcellular localizations and in fact do not share significant sequence homology.
Given the strong associations between the PABPs and poly(A) tail formation, length control, and mRNA stability, we hypothesized that one or both of these proteins could be involved in the poly(A)-dependent mRNA turnover by SOX. We began by monitoring the localization of these proteins in cells with or without SOX. Remarkably, immunofluorescence experiments revealed that whereas PABPN localization was unchanged by SOX (unpublished data), there was a striking redistribution of endogenous PABPC from the cytoplasm to the nucleus (Figure 3A). This phenotype was confirmed using two independent PABPC antibodies (Figure 3A and 3C). Interestingly, we have not observed an interaction between SOX and PABPC in co-immunoprecipitation experiments from cells transfected with SOX or lytically infected with KSHV (unpublished data), suggesting that PABPC relocalization is not due to direct binding and recruitment by SOX.
Although PABPC is a nuclear–cytoplasmic shuttling protein [34] its steady-state localization is almost exclusively cytoplasmic and possible roles for PABPC in nuclear events such as mRNA 3′ end formation and quality control have not been elucidated. To link PABPC nuclear import mechanistically to SOX-induced mRNA turnover, we examined a panel of SOX mutants for their ability to redistribute PABPC. Single-function SOX mutants such as P176S and the HSV SOX homolog AE lacking the mRNA turnover and hyperadenylation functions failed to alter PABPC localization (Figure 3B and 3C). However, the Q129H SOX mutant that lacks the conserved DNase activity but retains the ability to promote host shutoff and hyperadenylation induced PABPC nuclear recruitment to the same extent as WT SOX (Figure 3B). Thus, nuclear accumulation of PABPC requires the host shutoff activity of SOX.
Removal of PABPC from the cytoplasm would be predicted to destabilize mRNAs in that locale. To test this hypothesis, we performed fractionation experiments to monitor the half-life of GFP mRNA specifically in the cytoplasm of cells with and without SOX. Indeed, we observed that cytoplasmic mRNAs were more rapidly turned over in SOX-expressing cells compared with cells lacking SOX (7.5 h versus >30 h) (Figure 3D; gels shown in Figure S3). The shortened half-life of cytoplasmic GFP mRNA was comparable to that of total GFP mRNA extracted from unfractionated cells expressing SOX (6.5 h) (Figure 3D).
SOX expression initiates 12 h into the KSHV lytic cycle, but cellular mRNA destruction becomes most prominent at 18–24 h and is maintained throughout the lytic cycle [5]. To monitor PABPC localization during infection, telomerase-immortalized microvascular endothelial (TIME) cells were either mock infected, latently infected with KSHV, or lytically infected with KSHV for a time course of 8–24 h. PABPC staining was predominantly cytoplasmic in mock-infected cells, as well as during latent infection and at 8 h post lytic infection when SOX is not expressed and host shutoff does not occur (Figure 4). However, beginning at the onset of host shutoff at 12 h post lytic infection, PABPC concentration in the nucleus began to increase, and by 24 h, the majority of infected cells showed prominent nuclear PABPC staining. These results confirm that PABPC relocalization into the nucleus is similarly induced during KSHV infection and is temporally coincident with SOX-induced host shutoff.
The host shutoff-dependent nuclear redistribution of PABPC suggested that this factor could play a prominent role in mRNA turnover in SOX-expressing cells. Additionally, the fact that PABPN is functionally linked to poly(A) tail formation and length control prompted us to examine possible roles for this protein in SOX-induced hyperadenylation and mRNA turnover as well. To this end, we monitored SOX activity by northern blotting and quantitative real-time PCR (qPCR) upon siRNA-mediated knockdown of either PABPC or PABPN. Indeed, northern blotting showed there was a significant decrease in the ability of SOX to promote GFP mRNA degradation upon knockdown of either PABPN or PABPC (Figure 5A and 5B). In contrast, SOX-expressing cells transfected with a control nonspecific siRNA showed robust turnover of the GFP reporter mRNA. These effects are highly specific; we have performed siRNA-mediated knockdowns of approximately 10 other cellular proteins involved in mRNA stability with no resulting decrease in SOX function (Figure S4, unpublished data). Additionally, the siRNA treatment did not affect the levels of SOX protein expression or the mRNA levels of the GFP reporter in the absence of SOX (Figure 5A and 5B). These results were confirmed by qPCR analysis of GFP mRNA levels from these samples, which showed a strong inhibition of SOX host shutoff activity upon PABPN or PABPC knockdown (Figure 5C). Of note, we have consistently observed that knockdown of PABPN, but not PABPC, blocks hyperadenylation detected by northern blotting (compare lanes 3 and 4 in Figure 5B). In addition, we monitored hyperadenylation of endogenous messages in SOX-expressing cells upon PABPC knockdown using oligo(dT) in situ hybridization (Figure S5). In agreement with our northern blots, siRNA-mediated depletion of PABPC also failed to inhibit SOX-induced hyperadenylation in these experiments. Although similar experiments were also performed upon PABPN siRNA treatment, these were more difficult to interpret because a fraction of the vector-transfected control cells exhibited enhanced nuclear dT staining (possibly due to mRNA export defects). However, it appeared as though a reduced number of SOX-expressing cells lacking PABPN exhibited hyperadenylation (Figure S5). Collectively, these data suggest that hyperadenylation may be necessary, but not sufficient, for SOX-mediated RNA turnover, and that the contributions of PABPC to SOX function may be downstream of those of PABPN. That PABPC is required for the mRNA turnover activity of SOX suggests its host shutoff-dependent nuclear import is not simply a byproduct of hyperadenylation, but rather plays an integral role in directing turnover of cellular transcripts in the presence of SOX. Indeed, siRNA-induced knockdown of PAPII (which inhibits hyperadenylation; see Figure 2) produces no defect in SOX-induced PABPC relocalization (Figure S6), indicating that hyperadenylation is not a prerequisite for nuclear import of PABPC.
Our data indicate that polyadenylation plays a key role in the host shutoff function of SOX. We therefore sought to more directly evaluate the contribution of a poly(A) tail towards SOX-induced mRNA turnover by preventing polyadenylation of the GFP reporter message. This was accomplished by replacing the portion of the GFP 3′ UTR containing the AAUAAA polyadenylation signal sequence with a self-cleaving hammerhead ribozyme element (GFP-HR; Figure 6A). The 3′ end cleavage of this mRNA is mediated by the ribozyme rather than cellular machinery, and it is not polyadenylated at steady state and should not associate with the PABPs. Notably, although SOX promoted turnover of the polyadenylated GFP message, it failed to degrade the GFP-HR RNA (Figure 6C). To determine whether absence of polyadenylation was the primary cause for the inability of SOX to promote GFP-HR mRNA turnover, we next generated GFP constructs lacking the polyadenylation signal sequence but containing a templated stretch of either 60 adenosine residues (GFP-A60-HR) or, as a control, 60 uridine residues (GFP-U60-HR) immediately upstream of the ribozyme cleavage site (Figure 6A). The minimum poly(A) interaction site size for PABPN is 10 nt [35] and for PABPC is 12 nt [36], although when coated along a poly(A) tail, each PABPC protein covers approximately 25 nt [38]. Thus, the 60-nt templated poly(A) tail is of sufficient length to bind multiple copies of PABPC and/or PABPN. As would be predicted, the GFP-HR mRNA fails to be translated, whereas addition of the A60 tail partially rescues this defect, and very weak protein expression is observed with a U60 tail (Figure 6B). Significantly, the presence of a templated poly(A) tail, but not a poly(U) tail, was sufficient to reinstate SOX-induced degradation of the GFP reporter (Figure 6D), indicating that a poly(A) tail specifically participates in targeting mRNAs for turnover by SOX. We observed similar results with the SOX homolog from a related γ-herpesvirus, MHV68 (Figure S7).
To examine the requirement for PABPC and PABPN in SOX-induced turnover of the GFP-A60-HR mRNA, we performed siRNA-mediated knockdowns of these factors and monitored the resulting ability of SOX to degrade the GFP message (Figure 6E and 6F). The cells were also treated with LMB to determine whether SOX-induced hyperadenylation occurs on this ribozyme-terminated mRNA. Cells transfected with a scramble siRNA showed SOX-induced turnover of the GFP-A60-HR transcript, but no hyperadenylation (Figure 6E), as anticipated given our observation that hyperadenylation requires the machinery involved in cellular 3′ end formation, which does not participate in processing of the HR constructs. Knockdown of PABPN, which is required for poly(A) tail formation and hyperadenylation, did not prevent SOX-mediated turnover of GFP-A60-HR. However, knockdown of PABPC effectively prevented SOX-mediated turnover of this mRNA. Thus, PABPN is necessary for SOX-induced hyperadenylation and destruction of mRNAs processed by the cellular 3′ end machinery but is not required if a poly(A) tail is already in place, whereas PABPC is a critical cofactor for SOX-induced destruction of already polyadenylated mRNAs. Reproducibility of all northern blotting results shown in Figure 6 was demonstrated by quantification of multiple replicates (n≥3; Figure S8).
Finally, histone mRNAs are the only known cellular mRNAs lacking poly(A) tails, as they instead terminate in a 3′ stem loop (SL) structure that recruits a number of the same processing and degradation factors as poly(A) mRNAs [38],[39]. We examined whether a GFP construct containing the histone SL and the downstream element required for termination would be sensitive to SOX-mediated turnover (Figures 6C and S7). Interestingly, histone SL-terminating mRNA was degraded by SOX, suggesting that this unique message may recruit one or more factors in a poly(A)-independent manner that facilitate SOX targeting. This is in contrast to other non-polyadenylated RNAs (e.g., ribosomal RNAs), which are not subject to SOX-mediated degradation [5]. An important future direction will be to define the specific elements or factors that render RNAs like the histone mRNA susceptible to SOX-mediated RNA turnover, as they are anticipated to identify areas of convergence between polyadenylation-dependent and -independent pathways of mRNA degradation.
The ability to regulate cellular gene expression is a key aspect of the lifecycles of a diverse array of viruses. Global inhibition of cellular protein synthesis serves not only to ensure maximal viral gene expression by diverting the cellular resources towards the virus, but also assists in evasion of host immune responses detrimental to viral replicative success. Although the outcome of host shutoff may be similar for some pathogens, the mechanisms they use to achieve this endpoint are quite distinct. For example, poliovirus prevents cap-dependent translation by cleavage of eIF4G and PABP [40],[41], vesicular stomatitis virus blocks nuclear mRNA export via disruption of Rae1 function [42], and herpes simplex virus (HSV) both inhibits splicing and encodes a ribonuclease that degrades cytoplasmic mRNA [43]–[46]. Although KSHV infection elicits global mRNA turnover via the activity of SOX, the mechanisms driving this phenotype remained enigmatic.
Here, we demonstrate that SOX engages in a novel mechanism of host shutoff involving aberrant mRNA polyadenylation (Figure 7). To our knowledge, this is the first example of enhanced RNA turnover coupled to hyperadenylation by PAPII in metazoans. We further show that degradation of these cellular transcripts requires nuclear and cytoplasmic PABPs, the latter of which undergoes striking nuclear relocalization during KSHV infection. Manipulation of this cellular RNA 3′ processing event and PABPC nuclear recruitment require the host shutoff activity of SOX, and are therefore intimately linked to KSHV-induced transcriptome turnover. Although we do not yet know the identity of the ribonuclease ultimately responsible for mRNA destruction in the presence of SOX, we propose that SOX-induced alterations in mRNA processing events may render these nascent RNAs targets of cellular quality control machinery.
Several lines of evidence suggest that polyadenylation plays an integral role in SOX-mediated mRNA turnover. First, SOX mutants selectively defective for mRNA turnover, but retaining the conserved DNase activity, fail to promote hyperadenylation. Second, siRNA-mediated knockdown of PABPN, which stimulates poly(A) polymerase activity and has important roles in poly(A) tail formation and length control [16], reduces SOX-induced hyperadenylation and mRNA turnover. Finally, ribozyme-terminating transcripts that bypass cellular 3′ end cleavage and polyadenylation cannot be targeted by SOX for destruction, whereas turnover is reinstated upon addition of a templated poly(A) but not poly(U) tail. Collectively, these data suggest that the poly(A) tail serves as a signal for degradation in SOX-expressing cells and/or participates in the recruitment of decay factors, perhaps via the PABPs (Figure 7).
Polyadenylation-triggered mRNA decay is well established in prokaryotes such as Escherichia coli [47]. Both mature and fragmented bacterial mRNAs can be polyadenylated, but rather than stabilizing the messages (as generally occurs in eukaryotes), here polyadenylation facilitates RNA degradation [19]. In this regard, prokaryotic poly(A) tails are thought to serve as unstructured “landing pads” for exoribonucleases, thereby assisting their progression through structured regions of the RNA. In contrast, the eukaryotic nuclear mRNA polyadenylation reaction that is coupled to 3′ end cleavage and processing generally helps protect messages from exonucleolytic attack, and mutation of poly(A) polymerase results in rapid depletion of mRNAs [48],[49]. However, it is likely that polyadenylation can also serve as an important signaling mechanism for the cell to monitor the fidelity of RNA processing in eukaryotes. Compelling evidence linking polyadenylation directly to destruction of eukaryotic RNAs emerged in yeast with the discovery of the TRAMP nuclear polyadenylation complex that tags aberrant messages with short poly(A) tails to stimulate turnover [20]–[22]. Indeed, accumulation of polyadenylated forms of RNAs that do not normally have poly(A) tails, such as small nucleolar RNA (snoRNA) [50],[51], ribosomal RNA (rRNA) precursors [52]–[54], and intergenic transcriptional events [22],[55] can be detected in yeast lacking the exosome component Rrp6; these are presumably TRAMP-labeled degradation intermediates stabilized in the absence of surveillance-mediated decay. Although a number of homologs of the TRAMP polymerase Trf4 exist in humans, none have yet been shown to function in an analogous manner. However, short poly(A) tails reminiscent of TRAMP products have been detected on cotranscriptionally cleaved β-globin pre-mRNAs in mammalian cells [28], as well as on human mitochondrial transcripts and rRNAs [26],[27].
Proper 3′ end formation and polyadenylation are required for mRNA export into the cytoplasm, and defects in these processes trigger nuclear retention and RNA destruction by quality control pathways [56]. Interestingly, yeast nuclear export and 3′ end processing mutants can lead to hyperadenylated transcripts that accumulate at the site of transcription [23]–[25]. We therefore predict that hyperadenylated messages produced in SOX-expressing cells would be regarded as aberrantly processed and retained in the nucleus for eventual destruction. Of note, we have not observed significant defects in SOX activity upon depletion of the human exosome components (Figure S4), suggesting that this quality control complex does not play a major role in SOX-mediated host shutoff. This is perhaps not unexpected, as exosome depletion would be anticipated to only affect turnover of the nascent hyperadenylated nuclear transcripts, rather than the bulk of transcripts in the cytoplasm that become destabilized coincident with PABPC import. This idea is further supported by our observation that inhibiting hyperadenylation by depleting PAPII results in partial reduction rather than complete inhibition of SOX-induced mRNA turnover. It should be noted, however, that the nuclear exosome is apparently more refractory to siRNA-mediated turnover than the cytoplasmic exosome [57]. It is therefore possible that in our knockdown experiments, enough exosomal proteins remain in the nuclear fraction to promote turnover of the hyperadenylated messages. An alternative possibility is that other enzymes are involved in polyadenylation-triggered nuclear mRNA turnover during KSHV infection.
Given our evidence linking polyadenylation to host shutoff, the degradation of a reporter bearing the histone mRNA termination signals by SOX was unexpected. We envision at least two possible scenarios explaining degradation of this non-polyadenylated transcript. First, despite lacking the canonical mRNA cleavage and termination signals or a poly(A) tail, the histone mRNA 3′ end is nonetheless able to recruit a significant number of the same factors involved in nuclear RNA processing as well as cytoplasmic RNA turnover as polyadenylated mRNAs [39],[58],[59]. The histone mRNA 3′ end SL structure may therefore be able to bind proteins normally associated with a poly(A) tail that are necessary for SOX targeting. PABPC, for example, has been shown to associate with a non-poly(A) element within the MKK-2 mRNA 3′ UTR to control its stability [60], although to our knowledge, there is currently no evidence that it binds histone mRNA. Secondly, it is formally possible that in contrast to uninfected cells, upon SOX expression, this mRNA becomes polyadenylated and thus subject to destruction. An important future goal will be to determine whether either of these possibilities is correct.
Whereas PABPC is known to shuttle between the nucleus and cytoplasm, its steady-state localization is cytoplasmic, and possible roles for it in the nucleus remain largely unknown. Recent observations, however, suggest that PABPC is likely to have functions in the nucleus, because it interacts with polyadenylated nuclear pre-mRNAs as well as with PAP, suggesting that its association with the pre-mRNAs occurs during tail formation in the nucleus [61]. However, our observation that siRNA-mediated knockdown of PABPC prevents mRNA turnover, but not hyperadenylation by SOX, indicates that PABPC functions downstream of PABPN, perhaps in the recruitment of RNA decay factors. The fact that PABPC but not PABPN depletion blocks SOX-mediated degradation of the ribozyme-terminating GFP with a 60-nt templated poly(A) tail further bolsters this conclusion.
It should be noted that four cytoplasmic PABPs have been identified in human cells [62]. In addition to PABPC1, HEK 293T cells also likely express PABPC4 (iPAPB) [63] and PABPC5 (X-linked PABP) [64], whereas PABPC3 is reported to be testes specific [65]. Our knockdowns selectively targeted PABPC1, thought to be the predominant PABPC in these cells, and removal of this protein clearly has detrimental effects on SOX activity. However, there is extensive sequence homology between the PABPCs [62], and the PABPC1 antibody is predicted to cross-react with at least PABPC4 as well; it is therefore possible that SOX induces nuclear relocalization of multiple lineages of PABPC during host shutoff. We hypothesize that the presence of these additional PABPCs is what prevents general mRNA destabilization upon siRNA-mediated depletion of PABPC1 in the absence of SOX. In this regard, an additional consideration is that while PABPC1 depletion did not abrogate the hyperadenylation phenotype, it remains to be determined whether other PABPC proteins are similarly dispensable for this function.
Surveillance-mediated destruction of aberrantly processed pre-mRNAs could explain how cellular messages are depleted from the nucleus during lytic KSHV infection, but what about the abundance of cytoplasmic cellular transcripts? Given that the average half-lives of mammalian mRNAs are relatively long [66] and we have shown that SOX-expressing cells exhibit enhanced turnover of cytoplasmic messages, we predict that SOX activity must stimulate both nuclear and cytoplasmic mRNA decay either by distinct or overlapping mechanisms. In this regard, it is significant that PABPC has established roles in preserving cytoplasmic mRNA integrity by promoting mRNA circularization via interactions with eIF4G to enhance stability and translation via the closed loop model [16]. Thus, removal of PABPC from the cytoplasm could render these messages less translatable, unprotected at their 3′ termini, and more susceptible to nucleolytic attack, in agreement with our observation that the cytoplasmic GFP mRNA half-life is reduced in SOX-expressing cells. Indeed, multiple RNA viruses with unique translational strategies have evolved means to disrupt PABPC activity presumably to facilitate selective translation of viral messages and promote host translational shutoff; enteroviruses, caliciviruses, and HIV encode proteases that cleave PABPC [67]–[69], and the rotavirus NSP3 protein competes with PABPC for eIF4G binding and promotes PABPC relocalization [70]–[72]. Thus, multiple diverse groups of viruses have all evolved strategies to target this cellular factor, presumably to divert resources away from cellular gene expression. We hypothesize that SOX host shutoff activity consists of a nucleus-specific component requiring PABPC and PABPN-stimulated aberrant mRNA hyperadenylation and turnover, and a cytoplasmic component involving inhibition of mRNA translation followed by destabilization coincident with PABPC depletion. KSHV may have additional roles for PABPC during infection, as a recent report showed a limited amount of the K10/K10.1 viral protein associates with PABPC in the nucleus during the lytic cycle, although the functional significance of this observation remains unknown [73].
Finally, one particularly intriguing issue is how viral messages manage to evade host shutoff. KSHV mRNAs closely resemble cellular transcripts, in that they are 5′ capped and polyadenylated, some are spliced, and they are transcribed, processed, and translated using cellular machinery; yet these messages must escape the mRNA destruction fate suffered by cellular transcripts. Our findings suggest that successful viral gene expression during host shutoff requires navigating at least two obstacles: first, evading aberrant 3′ end mRNA processing and destruction in the nucleus and, second, keeping the messages stable and efficiently translated in the cytoplasm in the face of significantly reduced PABPC levels. One clever mechanism of evading nuclear degradation has been delineated for the highly abundant KSHV noncoding PAN RNA; this polyadenylated nuclear transcript contains a 79-nt RNA element (termed the ENE) near its 3′ end which acts post-transcriptionally to stabilize and enhance the nuclear levels of PAN or other reporter RNAs [74]. Modeling experiments predict the ENE folds into a secondary structure reminiscent of box H/ACA snoRNAs, and indeed, this element can form intermolecular interactions with the PAN poly(A) tail in a manner that blocks deadenylation, thereby stabilizing the RNA [75]. An important future challenge will be to delineate the mechanisms by which the bulk of the remaining viral mRNAs achieve these tasks, as such information may provide clues as to how cellular quality control checkpoints could be bypassed during viral infection or other human disease.
GFP-HR was created by replacing the AAUAAA signal sequence of pd2EGFP-N1 (Clontech) with a hammerhead ribozyme (CCTGTCACCGGATGTGTTTTCCGGTCTGATGAGTCCGTGAGGACGAAACAGG) by deleting the NotI/AflII-flanked poly-A signal and cloning in annealed hammerhead ribozyme oligos with NotI and AflII overhangs. GFP-A60-HR and GFP-U60-HR were generated by ligation of an A60 oligo into the Not1 site of GFP-HR in the forward or reverse orientation, respectively. GFP-hisSL was created by replacing the AAUAAA signal sequence of pd2EGFP-N1 (Clontech) with a histone stem and downstream element (HDE) sequence (ATGTAAGTCTAGAGGATGGGGAGCAAAAGGCTCTTTTCAGAGCCACCCACTGAATCAGATAAAGAGTTGTGTCACGGTAGCCA) by deleting the NotI/AflII flanked poly-A signal and cloning in annealed histone stem and downstream element oligos with NotI/AflII overhangs. The cloning of pCDEF3-SOX, pCDEF3-HA-HSV AE [5], and the pCDEF3-SOX single-function mutants (Q129H, T241, and P176S) [4] were described previously. Mutant L20/23A was generated by overlapping PCR and then cloned into the EcoR1/Not1 sites of pCDEF3.
HEK 293T cells (American Type Culture Collection) were maintained in DMEM supplemented with 10% FBS. Telomerase-immortalized microvascular endothelial (TIME) cells [76] were maintained using EBM-2 medium bullet kits (Clonetics). TIME cells were infected with KSHV and lytically reactivated with an adenoviral vector expressing the lytic switch protein RTA as described previously [77].
siRNA duplex oligos (Bioneer; Dharmacon) were generated against the following target sequences: PAPII (accession number NM_032632) siRNA #1: CTGCGTACTTACACAGAAA, PAPII siRNA #2: GATTAGGAGTGCATACAAA; PAPγ (GenBank accession number NM_022894) siRNA #1: CAACAGAATTCTACGTATA, PAPγ siRNA #2: GGAGAAACAGAAAGGAATA; PABPC1 (GenBank accession number NM_002568) siRNA #1: GAAAGGAGCTCAATGGAAA; PABPC1 siRNA #2: GGACAAATCCATTGATAAT; and PABPN (GenBank accession number NM_004643) siRNA #1: GTAGAGAAGCAGATGAATA; PABPN siRNA #2: CTATTTAGAGGAAGGCAAA. Nonspecific control siRNA duplexes #1 and #2 were purchased from Ambion. HEK 293T cells were transfected with siRNA oligos at a final concentration of 200 nM using Lipofectamine 2000 (Invitrogen), both at 48 h and 24 h prior to DNA transfections, and harvested 24 h after the DNA transfections for total RNA and protein or processed for in situ hybridization.
Cells were harvested 24 h post DNA transfection for both oligo(dT) in situ hybridization and immunofluorescence analyses. In situ samples were processed as described (http://www.singerlab.org/protocols) using 2 ng/µL of AlexaFluor 546-labeled oligo-dT(15) (Molecular Probes). After oligo hybridization, samples were incubated with either α-SOX J5803 or α-HA (Abcam) primary antibodies at a 1∶500 dilution in 2× SSC, 0.1% triton X-100 for 3 h at 37°C, then subsequently with Alexa fluor 488-labeled goat α-rabbit secondary antibodies (Molecular Probes) and mounted with DAPI-containing Vectashield mounting medium (Vector Labs). IFA not performed in conjunction with in situ hybridization was done as described previously [77] using either SOX J5803 polyclonal antibodies (1∶500 dilution), 10E10 monoclonal PABPC antibodies (generously provided by Dr. G. Dreyfuss) (1∶1,000 dilution), rabbit polyclonal PABPC #39473 antibodies (generously provided by Dr. R. Andino) (1∶100 dilution), rabbit polyclonal PABPN antibodies (generously provided by Dr. E. Wahle) (1∶250 dilution), monoclonal HA 12CA5 antibodies (Abcam) (1∶500 dilution), and Alexa Fluor 488- or 546-labeled goat α-rabbit or α-mouse secondary antibodies (1∶1,500 dilution) (Molecular Probes).
Lysates were prepared in RIPA buffer (50 mM Tris-HCl [pH 8.0], 150 mM Nacl, 1% [v/v] Nonidet P-40, 0.5% [w/v] sodium deoxycholate, 0.1% [w/v] sodium dodecyl sulfate [SDS]) containing protease inhibitors (Roche). Equivalent amounts of each sample were subjected to immunoblotting with the following antibodies: PAPII (1∶1,000 dilution) (generously provided by Dr. J. Manley), PAPγ (1∶1,000 dilution) (generously provided by Dr. A. Virtanen), HA (Abcam) (1∶5,000 dilution), PABPC #39473 (1∶2,500), or SOX J5803 (1∶5,000 dilution) (see below), and either HRP-conjugated goat-α-rabbit or goat-α-mouse secondary antibodies (Southern Biotechnology Assoc.). Rabbit polyclonal antisera were raised against a maltose binding protein (MBP)-tagged full-length SOX by standard methods [78].
Where indicated, the transfected HEK 293T cells were treated with 5 ng/ml leptomycin B (LMB) (Sigma) for 6–12 h prior to RNA isolation. Total cellular RNA was isolated using RNA-BEE (Tel-Test), resolved on 1.2% agarose-formaldehyde gels, and probed with a 32P-labeled GFP DNA probe generated using the RediPrime II random prime labeling kit (Amersham). Membranes were subsequently incubated with an 18S probe as a loading control. For half-life analysis, cells were treated with LMB for 12 h, then washed with PBS and transferred to medium containing 1 µg/ml Actinomycin D (ActD) minus LMB for the indicated times. For fractionation studies, NE-PER Nuclear and Cytoplasmic Extraction Reagents Kit (Pierce) or Paris Kit (Ambion) was used according to the manufacturer's instructions. The level of GFP mRNA was divided by the corresponding level of 18S rRNA to correct for errors in sample loading. The log of normalized data was then plotted versus the time of treatment of ActD. The reported data are the means of a minimum of three independent experiments. Northern blots were analyzed using a Typhoon 8600 phosphorimager (Molecular Dynamics).
RNaseH digestions were performed by combining 10 µg of RNA with 500 pmol of oligo(dT) primer in a 25 µl reaction, incubating at 65°C for 8 min, then adding 1 U of RNaseH (New England Biolabs), RNaseH buffer to 1×, and 40 U of RNasin (Promega). Reactions were incubated at 37°C for 30 min, then terminated by adding 1 µl of 0.5 M EDTA (pH 8.0) and ethanol precipitating the RNA prior to gel electrophoresis.
cDNAs were synthesized from 1 µg of total RNA using AMV reverse transcriptase (Promega), diluted 1∶5, and used directly for qPCR analysis. GFP cDNA was amplified using the 5′ primer 5′CAACAGCCACAACGTCTATATCATG and 3′ primer 5′ATGTTGTGGCGGATCTTGAAG, along with a Taqman probe 5′FAM-CAAGCAGAAGAACGGCATCAAGGTGA-BHQ1. Taqman Ribosomal RNA Control Reagent (Applied Biosystems) with VIC-labeled probe and forward and reverse primers for human 18S rRNA was used as a loading control. Standard curves were prepared for each primer/probe set using 10-fold serial dilutions of either the 97-nt GFP fragment or the 55-nt 18S fragment derived from a pGem-T-easy vector (Promega). The qPCR reaction was performed using Taqman Gene Expression Mix (Applied Biosystems) in the presence of 100 nM GFP primers, 200 nM GFP probe, 50 nM 18S rRNA primers, 200 nM 18S rRNA probe, and 9 mM MgCl2. The level of GFP mRNA was calculated using a mathematical model of relative expression in qPCR [79] to quantify the relative level of GFP mRNA in comparison to the 18S rRNA.
|
10.1371/journal.pgen.1006541 | Bdf1 Bromodomains Are Essential for Meiosis and the Expression of Meiotic-Specific Genes | Bromodomain and Extra-terminal motif (BET) proteins play a central role in transcription regulation and chromatin signalling pathways. They are present in unicellular eukaryotes and in this study, the role of the BET protein Bdf1 has been explored in Saccharomyces cerevisiae. Mutation of Bdf1 bromodomains revealed defects on both the formation of spores and the meiotic progression, blocking cells at the exit from prophase, before the first meiotic division. This phenotype is associated with a massive deregulation of the transcription of meiotic genes and Bdf1 bromodomains are required for appropriate expression of the key meiotic transcription factor NDT80 and almost all the Ndt80-inducible genes, including APC complex components. Bdf1 notably accumulates on the promoter of Ndt80 and its recruitment is dependent on Bdf1 bromodomains. In addition, the ectopic expression of NDT80 during meiosis partially bypasses this dependency. Finally, purification of Bdf1 partners identified two independent complexes with Bdf2 or the SWR complex, neither of which was required to complete sporulation. Taken together, our results unveil a new role for Bdf1 –working independently from its predominant protein partners Bdf2 and the SWR1 complex–as a regulator of meiosis-specific genes.
| Chromatin modifying proteins play a central role in transcription regulation and chromatin signalling. In this study we investigated the functional role of the bromodomains of the chromatin protein Bdf1 during yeast gametogenesis. Our results show that the bromodomains of Bdf1 are essential for meiotic progression and the formation of mature spores. Bdf1 bromodomains are required for the expression of key meiotic genes and the master regulator NDT80. Forced expression of NDT80 can partially rescue the formation of spores when Bdf1 bromodomains are mutated. The results presented here indicate that Bdf1 forms two exclusive complexes, with Bdf2 or with the SWR complex. However, none of these complexes are required for sporulation progression. To conclude, our findings suggest that Bdf1 is a new regulator of the meiotic transcription program and of the expression of the master regulator NDT80.
| Protein members of the BET family share a conserved modular architecture, with two bromodomains in their N-terminal part, an extra-terminal recruitment (ET) domain and other conserved motifs. Bromodomain modules bind acetylated lysines in histones and other proteins, and BET bromodomains specifically recognise the acetylated lysines of core histones, in particular H3 and H4 [1].
BET proteins are present in unicellular eukaryotes such as the model organism Saccharomyces cerevisiae, where two homologous genes, Bdf1 and Bdf2, are expressed. Bdf1 has a typical BET protein structure, with two bromodomains, one ET domain and other conserved motifs. Multiple functional roles have been proposed for this protein, which was first described as a regulator of snRNA [2] and then identified as part of the yeast’s general transcription factor, TFIID [3]. In yeast, unlike its human homologue TAF(II)250, the TAF(II)145 protein lacks a module with bromodomains. Bdf1 could thus represent the missing piece of this yeast TAF complex [3,4]. Moreover, Bdf1 plays a role in regulating gene expression in response to various stresses [5–8].
Bdf1 has been shown to be part of the SWR complex, which is responsible for the incorporation of the histone variant H2A.Z into chromatin [9,10]. Bdf1 is not required for the enzymatic activity of Swr1 but facilitates H2A.Z incorporation in vivo [5,11,12]. Two studies demonstrated that Bdf1 bromodomains are functional and bind multi-acetylated histones, where they mediate an anti-silencing function and prevent the spreading of Sir proteins at heterochromatin boundaries [13,14]. Unexpectedly, a high throughput study also implicated Bdf1 in pre-mRNA splicing [15]. Thus, this protein could potentially link chromatin remodelling, transcription initiation and pre-mRNA splicing.
Most of the data describing the functions of Bdf1 were obtained using vegetative S. cerevisiae cells. However, Bdf1 was described to play an essential role for sporulation when its gene was first identified in yeast [16]. In this process, nutrient starvation triggers a specific differentiation program in diploid S. cerevisiae, starting with meiosis and ending with the formation of four quiescent spores [17,18]. In their seminal work, Chua et al. [16] identified Bdf1 as a chromosomal protein with two bromodomains which was required for meiosis progression. We also found that spore chromatin is highly compacted, hyperacetylated, and enriched in Bdf1 [19]. Altogether, this information suggests that Bdf1 plays a functional role during the meiotic and post-meiotic stages of sporulation.
In mammals, Brd2, Brd3 and Brdt play specific roles in cellular differentiation in a range of tissues [20–22]. Thus, Brd2 is involved in neuronal development and differentiation [23,24], while Brd3 is essential to erythropoiesis [25]. Brdt is only expressed in the testis and regulates the expression program of meiotic and post-meiotic genes, driving the transcription program throughout sperm differentiation [26]. In addition, it is a key player in the final chromatin reorganisation step found in the final spermatic structure [27]. Brdt preferentially binds to hyperacetylated histones and promotes a drastic chromatin reorganisation during sperm differentiation [28,29]. Brd4 is also involved in regulating transcription during post-meiotic sperm differentiation and final chromatin reorganisation [30].
Altogether, BET proteins are important for gametogenesis. The role of Bdf1 during sporulation was the focus of the study presented here. Our results indicate that, although Bdf1 bromodomains are not required for growth on non-fermentable carbon sources, they are essential for sporulation to complete. Transcriptomic analysis during sporulation revealed that Bdf1 bromodomains play an essential coordinating role in the transcriptional program of meiosis leading to the formation of spores. In particular, Bdf1 was found to accumulate on the promoter of the master regulator NDT80 before meiosis. This recruitment requires functional Bdf1 bromodomains and is essential for a normal expression of NDT80. Thus the ectopic expression of Ndt80 can partially overcome the defects observed when Bdf1 bromodomains are mutated. Finally, Bdf1 forms two exclusive complexes, neither of which was required for meiotic gene regulation or to complete sporulation.
Deregulation of some master regulators of sporulation could be an obvious explanation for failure to induce this differentiation program. The expression of early genes IME1 and IME2, which is required for progression through meiotic S and G2 phases, was not significantly affected by mutation of the Bdf1 bromodomains (S2A Fig). Middle and late sporulation genes are essential for meiotic divisions and post-meiotic spore differentiation. The expression of a selection of these genes was tested by RT-qPCR and found to be defective when Bdf1 bromodomains are mutated (S2B and S2C Fig). This result is consistent with the sporulation phenotype of the strain, which is blocked at the transition between the pachytene phase and meiotic divisions.
RNA-seq experiments provided a general view of the transcriptional defects caused by mutation of both Bdf1 bromodomains. The expression of BDF1 was monitored in triplicate experiments performed on the WT and in bdf1-Y187F-Y354F strains at three time-points (before sporulation induction (0 h), 4 and 8 hours after induction, Fig 3A). The bioinformatics pipeline used in the analysis of the results of these experiments is presented in S3A Fig. Increased BDF1 transcript and protein levels were detected when its bromodomains were mutated, possibly as part of a compensatory response by the cell (S3B Fig). Furthermore, our results confirmed that Bdf1 regulates BDF2 mRNA levels, particularly during sporulation when the mutation of Bdf1 bromodomains may induce a compensatory overexpression of BDF2 (S3B Fig, [34]).
The distance between each replicate was analysed and, as expected, biological replicates appeared clustered by strain and sporulation time-point (Fig 3B). Thus, before sporulation was induced, the transcription program was similar in the WT and BDF1 bromodomain mutant strains. Defects started to accumulate from 4 h and culminated at 8 h when the BDF1 bromodomain mutant strain stopped progressing through the sporulation program (Fig 3B and 3C). The same observation was made based on an independent assessment using principal component analysis (S3C Fig).
A total of 833 genes were differentially downregulated when combined between the 4h and 8h time points. The examination of their promoters revealed that binding sequences of the transcription factors NDT80, CUP9, CUP2, FKH1 and ROX1 were statistically enriched (p-value <0.001, Fig 4A). Of these, only Ndt80 has been described as essential for sporulation and is known to be a key meiotic regulator [38–41]. Indeed, Ndt80 is induced during meiosis where it is expressed at much higher levels than Cup9, Cup2, Fkh1 and Rox1 (Fig 4A). Defects in Ndt80 expression also explain the non-induction of middle sporulation genes, such as SMK1 and SSP1, as expression of these genes is controlled by Ndt80; both were downregulated in the bromodomain mutant strain (S2B Fig, [35,42]).
Ndt80 is the key transcription factor that activates the promoters of middle sporulation genes and is expressed just before the middle genes are induced [41]. The examination of the promoters of the 833 differentially downregulated genes identified 531 genes with putative NDT80 binding sites (Fig 4B, S5 Table). In addition, the probability that an Ndt80 binding sequence is present within the promoter of Bdf1-regulated genes was calculated based on the ROC-AUC value produced by the MORPHEUS webtool [43] (Fig 4C). The ROC-AUC value of 0.63, which is above 0.5, indicates an enrichment of the Ndt80 binding sequence in the genes regulated by Bdf1 bromodomains when compared to a list of unrelated genes. Finally, a list of genes potentially regulated by Ndt80 was obtained from Chu et al [35]. These genes were statistically over-represented in the list of genes downregulated in the bdf1-Y187F-Y354F mutant (Fig 4D, p-value = 10−48 with a hypergeometric test). In conclusion, a significant proportion of the genes which fail to induce when Bdf1 bromodomains are mutated are regulated by Ndt80.
An oestradiol-inducible form of NDT80 was used to test whether a reduction in Ndt80 levels could explain the phenotype observed when Bdf1 bromodomains are mutated [44,45]. This system was introduced into the bdf1-Y187F-Y354F mutant and oestradiol was added 6 hours after sporulation induction. Induction of NDT80 expression at this stage partially alleviated the defects observed when Bdf1 bromodomains are mutated, with 40% of spores formed in this mutant strain (Fig 4E). This result indicates that mutation of Bdf1 bromodomains affects the expression of the master regulator NDT80 alters the expression of middle sporulation genes and that these sporulation defects can be partly rescued by ectopic overexpression of NDT80.
We next wondered whether Bdf1 interactants could contribute to the regulation of the meiotic transcription program and mediate the transcriptional activation of NDT80. The interactome of Bdf1 during vegetative growth was determined using mass spectrometry-based proteomic analysis of affinity purified samples (Fig 5). The relative abundances of Bdf1 partners were determined using the mass spectrometry-based iBAQ metrics [49,50]. Interestingly, the most abundant Bdf1 partner was Bdf2. The 12 subunits of the SWR complex, involved in the incorporation of the yeast histone variant H2A.Z, were also identified [9]. Both of these interactions with Bdf1 were confirmed by western blot (S4B Fig). The relative stoichiometry of the SWR subunits found to interact with Bdf1-TAP was compatible with the recently published structure of the complex, where the subunits Rvb1/2 are 2.5 times more abundant than the other subunits (Fig 5A bottom, [51]). Finally, the kinase CK2 was also detected (Fig 5A and S6 Table). This protein had previously been associated with Bdf1 [9].
Additional purifications were then performed to refine our knowledge of how the Bdf1 interactome is organised. Bdf2-TAP purifications were performed and identified Bdf1 as the major partner (Fig 5B). None of the SWR subunits was associated with Bdf2.
The Swr1 protein was then TAP-tagged and its complex purified. This analysis identified all the subunits of the SWR complex (Fig 5C). Bdf1 was identified as a minor interacting partner of Swr1 compared to the well-established SWR complex subunits. In these experiments, a 9 fold difference in relative abundance was detected for Rvb1/2 and the other SWR subunits. This difference with the data obtained in the Bdf1 complex and the published structure of the SWR complex could be explained if the SWR complex is present in vivo in several forms. This hypothesis is strengthened by results from structural studies using cosslinking reagents to obtain a homogeneous population of the complex [51]. Interestingly, Bdf2 was never identified in SWR complex. This result suggests that Bdf1 could form two exclusive complexes, one with the SWR complex and another one with Bdf2.
Finally, Bdf1-TAP was purified from a strain lacking YAF9, the subunit of the SWR complex responsible for recruiting Bdf1 [52]. In the yaf9Δ strain, Bdf1 still associated with Bdf2, but its interaction with the SWR complex was completely abolished (Fig 5D). Taken together, these results suggest that Bdf1 can be part of two exclusive complexes: one with the SWR complex or a Bdf1 / Bdf2 complex whose function remains to be characterised (Fig 5E).
The sporulation phenotypes induced by the mutations of Bdf1 bromodomains could be functionally explained by its interactions with the SWR complex or with Bdf2. To investigate this hypothesis, YAF9 and SWR1 were mutated in the SK1 genetic background and compared to bdf1 mutant strains. Deletion of YAF9 or SWR1 had no effect on growth in any of these conditions (Fig 6A), nor did YAF9 and SWR1 deletions affect sporulation efficiency (Fig 6B). Therefore, the sporulation defects observed when Bdf1 bromodomains are mutated are not related to the function of the SWR complex.
In the absence of a role for the SWR complex in sporulation, we hypothesised that the Bdf1 / Bdf2 complex could play a role during sporulation which may be related to the defects observed when Bdf1 bromodomains are mutated. As previously published [2,8,16,32], strains deleted for BDF1 and BDF2 exhibit a slow growth phenotype on fermentable carbon sources, and fail to grow on non-fermentable carbon sources such as acetate or glycerol (Fig 6A). Different regions of Bdf2 were deleted to map the regions required for its interaction with Bdf1 (Fig 6C). This analysis indicated that Bdf2 bromodomains are essential for its interaction with Bdf1 (Fig 6D). Interestingly, deletion of the two Bdf2 bromodomains had no phenotypic effects either on non-fermentable carbon sources or during sporulation (Fig 6E, left). Finally, we studied the functional role of the Extra-Terminal domain (ET), which is a conserved protein–protein interacting domain that regulates transcriptional activity [53,54]. Deleting this domain in BDF2 disrupted its interaction with Bdf1 (Fig 6D) but had no effect on sporulation (Fig 6E, left). Deletion of the same domain in Bdf1 did not affect growth on non-fermentable carbon sources at 30°C but revealed some defects at 37°C (S5B Fig). In contrast to Bdf2, the ET deletion in Bdf1 affects sporulation to a similar extent to the mutation of its bromodomains (Fig 6E, right).
These last results underline the fact that even though Bdf1 and Bdf2 are highly similar in their modular organisation, they play different roles during sporulation. Thus, only Bdf1 bromodomains, and not Bdf2 bromodomains, are essential for the formation of spores. Moreover, the sporulation defects observed in strains with mutated Bdf1 bromodomains are not mediated by any of its partners. Indeed, the disruption of the interaction of Bdf1 with Bdf2 or with the SWR complex generated no sporulation defects.
Finally, we examined whether deletion of BDF2 impacted Bdf1 recruitment to chromatin. ChIP experiments suggest that Bdf1 does not require Bdf2 to bind to its chromatin targets (Fig 6F). In addition, we were unable to detect any recruitment of Bdf2 to these genes, which raises the issue of the recruitment of Bdf2 to chromatin [55]. The fact that Bdf1 loading onto chromatin does not depend on Bdf2 reinforces the idea that Bdf2 does not contribute to the transcriptional regulation of middle sporulation genes.
In this study, the functional role of Bdf1, and more particularly its bromodomains, was analysed during sporulation. Our results showed that Bdf1, but not its bromodomains, is essential for growth on non-fermentable carbon sources. The bromodomains are, however, required for meiosis to progress and for mature spores to form. The gene expression program during sporulation when Bdf1 bromodomains are mutated is blocked in the transcriptional cascade at the point when Ndt80 intervenes, culminating in meiotic arrest and the non-expression of middle genes. Detailed characterisation of Bdf1 interactants revealed that it forms two exclusive complexes, one with Bdf2 and the other with the SWR complex. However, neither of these complexes is involved in the meiotic block observed when Bdf1 bromodomains are mutated. Therefore, we hypothesise that Bdf1 transcriptionally regulates meiotic genes without input from either of its major partners (Fig 5E).
Here, and in previous studies, deletion of BDF1 or BDF2 was found to lead to severe growth defects in the absence of a fermentable carbon source [2,16,32]. Indeed, Bdf1 was shown to preserve mitochondrial activity in various metabolic conditions [6,56]. However, a putative role for Bdf2 has yet to be investigated. Our results indicated that this phenotype is not controlled by Bdf1 or Bdf2 bromodomains, thus possibly revealing a new function of BET proteins in cell metabolism, which is independent of their bromodomains and remains to be characterised.
Bdf1 was first linked to sporulation when its gene was cloned, its bromodomains were identified and the growth and sporulation defects induced by its deletion were analysed [16]. Furthermore, a transposon-based screen indicated that the second Bdf1 bromodomain, but not the first one, was essential for the completion of meiosis and spore formation [16]. Here, we found that both bromodomains of Bdf1 have to be mutated to impair the formation of spores and cause an arrest during meiosis. These apparent discrepancies compared to the results presented in Chua et al. [16] could be explained by the facts that (i) they used a transposon mutational screen which did not precisely disrupt Bdf1 bromodomains and (ii) a transposon insertion in the second bromodomain may have also disrupted the subsequent ET domain, which we also found to be essential for sporulation (Fig 6E, right).
When Bdf1 bromodomains are mutated, early inducers of the transcription program for sporulation are correctly expressed but transcriptional defects accumulate as sporulation progresses (Fig 3D, S2A Fig). In particular, expression of the middle sporulation genes and of the master regulator NDT80 failed when Bdf1 bromodomains were not functional (Figs 4B and S2B). Interestingly, 65% of the genes downregulated in the bdf1 bromodomain mutant were also predicted targets of Ndt80 (Fig 4B). Several studies identified Bdf1 as important for gene splicing [15,34]. This functional role appears to be independent of its bromodomains as no splicing defects were observed in the bdf1-Y187F-Y354F mutant (S3D and S3E Fig).
Quantitative proteomics analysis of the Bdf1 interactome revealed that Bdf1’s main partners are Bdf2 and the SWR complex. These interactions occur as part of two different, mutually exclusive, complexes (Fig 5E). It is interesting to note that Bdf1, Bdf2 and Yaf9, the subunit of Swr1 which interacts with Bdf1, share similarities: first, they all contain domains which recognise acetylated lysines: bromodomains for Bdf1 and Bdf2 and a YEATS domain for Yaf9 [57]. The Yaf9 YEATS domain was recently described as a reader of lysine crotonylation and to promote active transcription [58]. Second, they all possess a coiled-coil domain, well known to mediate protein dimerization (S6 Fig, [59]). The B motif present in all the members of the BET family is also a coiled-coil domain and could mediate the dimerization of BET proteins [60]. Therefore, we hypothesised that coiled-coil domains could contribute to the formation of the two complexes the Bdf1 protein is involved in; Bdf1’s coiled-coil domain could interact exclusively either with the coiled-coil domain of Bdf2 or of Yaf9. This would explain how Bdf1 forms exclusive complexes with these two proteins.
In the SWR complex, Bdf1 promotes the incorporation of H2A.Z into acetylated regions [10,61,62]. The function of the interaction between Bdf1 and Bdf2 remains to be characterised. This interaction seems to be tight because deletion of any of the Bdf2 domains -bromodomains or ET- disrupts its interaction with Bdf1. However, none of these BDF2 mutations induces any sporulation defects.
Although Bdf1 and Bdf2 are functionally redundant—as only one is needed to maintain cell viability—they could play different roles in controlling chromatin dynamics. Indeed, they occupy distinct locations in the genome and larger amounts of Bdf1 bind to chromatin compared to Bdf2 [55]. Deletion of BDF2 in vegetative cells only affects a few transcripts, whereas the deletion of BDF1 provokes major transcriptomic changes [14]. The situation is similar during sporulation: Bdf1 is loaded on various loci, and especially middle sporulation genes, whereas Bdf2 does not seem to be strongly attached to chromatin (Fig 4F and ref [19]). The timing of the recruitment of Bdf1 onto chromatin is particularly intriguing, with Bdf1 detected on the promoters of middle sporulation genes before their transcriptional activation. This recruitment is essential for the activation of these genes and involves Bdf1 bromodomains. When these domains are mutated, Bdf1 is no longer recruited to these promoters and their genes fail to induce. The non-induction of NDT80 causes the major defects observed when Bdf1 bromodomains are mutated. Indeed, this Bdf1 mutation phenocopies deletion of NDT80, which also induces stalling in the pachytene phase and failure to accumulate middle-meiotic mRNA. Moreover, the ectopic expression of NDT80 can partially rescue the sporulation defects caused by mutation of Bdf1 bromodomains. However, the mechanisms through which Bdf1 specifically regulates the expression of NDT80 and middle sporulation genes remain unclear. Bdf1 was originally described as a missing piece of the TFIID complex, although as a loosely interacting protein [63], which could partially explain why no TFIID proteins were identified in our purifications. Perhaps a minor fraction of Bdf1 interacts with the transcription machinery. Bdf1 alone could promote a chromatin state favourable for the expression of middle sporulation genes. Thus, Bdf1 could promote the initial low-level pool of Ndt80 in meiotic G2 and also promote the feed-forward autoregulatory loop required to trigger high expression levels of middle genes [64]. Bdf1 is a member of the BET family, which also includes the mammalian proteins Brd2, Brd3, Brd4 and Brdt. Brd4 has been shown to facilitate transcriptional activation and was recently found to have an intrinsic HAT activity [48]. However, the residues controlling this activity are not conserved in Bdf1. Finally, H4 acetylation and H4K5 butyrylation have been shown to regulate the binding of Brdt to chromatin, resulting in highly active transcription [65]. Similarly, different acyl marks could regulate Bdf1 binding to chromatin and may explain why the expression of middle genes specifically requires Bdf1 bromodomains.
This novel role for Bdf1 during yeast meiosis could be shared by mammalian BET proteins during spermatogenesis. Indeed, Brd2 and Brd4 are known to associate with mitotic chromosomes and Brd4 binds post-mitotic genes to promote their transcriptional reactivation after the M phase [66–68]. During male meiosis in mouse, Brdt is essential for the gene expression program, and its deletion affects the expression of testis-specific cyclin genes such as Ccna1, which are required for meiotic division [26]. In yeast, the results presented here revealed that mutation of Bdf1 bromodomains abolished the expression of almost all the subunits of the APC complex, which is required for meiotic progression [69]. Interestingly, in mice, the absence of Brdt leads to a loss of FZR1 expression—an essential activator of the APC complex during meiosis—and subsequent deregulation of the APC complex (S7 Fig and reference [70]).
In conclusion, Bdf1 appears to be a key chromatin protein for the meiotic transcription program, regulating the expression of middle sporulation genes and promoting their rapid activation. Its role during meiosis appears to be evolutionarily conserved from yeast to mouse, and BET proteins are essential for the functionality of the APC complex and its regulation of meiotic progression. Our results suggest that the control of gene activation by Bdf1 does not rely on any of its major protein partners. Further studies will be required to explore its precise mode of action and whether these molecular mechanisms are conserved during the mammalian cell cycle in somatic cells and during meiosis and gamete differentiation.
GST, TAP and HA antibodies were obtained from Dutscher (Ref. 27-4577-01), Fisher (Ref. 10506450) and Roche (Ref. 11867423001), respectively. Anti-Bdf1 antibody was developed in-house by injecting recombinant protein into rabbits (Covalab). The use of bdf1Δ and Bdf1-TAP strains validated its specificity (S4A Fig).
Yeast Bdf1 bromodomain 1 (residues 132–263) or 2 (residues 317–430) and human Brd4 bromodomain 1 (residues 22–204) were cloned into pGEX-4T1 as GST-tagged proteins. Expression was induced in E. coli strain BL21 (DE3) grown in LB medium with kanamycin (50 μg/mL) at 37°C by adding 0.5 mM IPTG at OD600 of 1 and then incubated for 16 h at 16°C. Cells were lysed by sonication in Tris-HCl 50 mM pH 7.5, NaCl 150 mM and protease inhibitors. Clarified lysate was incubated with glutathione sepharose (Fisher, ref. W7349W) and washed in Tris-HCl 50 mM pH 7.5, NaCl 500 mM, NP-40 1%. Bound proteins were eluted in glutathione 10 mM.
For peptide pull-downs, 0.5 nmol of biotinylated H4 (Millipore, ref. 12–405) and H4K5ac K8ac K12ac K16ac peptides (H4ac4, H4 (Millipore, ref. 12–379)) were bound to Streptavidin magnetic beads (Thermo Dynabeads MyOne Streptavidin, ref. 65001) and used to pull-down 1.25 μg of GST-tagged Bdf1 bromodomain 1 and 2 (wild-type and mutants, bdf1-Y187F and bdf1-Y354F) and GST-tagged Brd4 bromodomain 1 of Brd4 in Tris-HCl pH 7.5 50 mM, NaCl 150 mM, NP-40 0.1%, Glycerol 10% and DTT 1 mM. Washes were optimized using 500 mM of NaCl. Bound proteins were eluted by boiling in Leammli buffer and proteins analysed by western blot using an anti-GST antibody.
Histone peptide array were purchased from Active Motif (ref. 13005) and used in accordance with the supplier’s recommendations. Briefly, arrays were blocked and then incubated with 1 μM of purified GST-tagged Bdf1 bromodomain 1 and 2. Incubation was allowed to proceed for 2 h at 4°C in binding buffer (Tris-HCl 50 mM pH 7.4, NP-40 0.1%, NaCl 150 mM, glycerol 10%). The membrane was washed in PBS Tween 0.1% and detected with an anti-GST antibody mixed with an anti-myc antibody used as a positive control and for normalisation. The membrane was washed and incubated with horseradish-peroxidase-conjugated anti-Goat antibody (Jackson, ref. 705.035.147). The membrane was submerged in Clarity Western ECL Substrate (Bio-Rad, ref. 170–5060), imaged (Bio-Rad ChemiDoc XRS Imaging) and data were quantified using an array analyser software provided by Active Motif.
Signal intensities are presented in S1 Table.
Yeast strains, plasmids and primers are listed in S2, S3 and S4 Tables respectively. The genotypes/expression profiles of all deletion mutants and tagged strains were confirmed by PCR analysis, sequencing and/or western blot analysis. Frogging assays on agar plates used 10-fold serial dilutions.
Diploid yeast in the SK1 background was grown to an OD600 of 0.5 in YPD. Cells were washed and resuspended in YPA at an OD600 of 0.03. After 12 h cells were washed and transferred into sporulation media (K acetate 2%) supplemented with auxotrophic amino acids. Six hours after sporulation induction, GAL-NDT80 GAL4.ER strain was released from the meiotic arrest by the addition of 1 μM oestradiol (5 mM stock in ethanol, Sigma E2758-1G). Sporulation efficiency was assessed after 24 h of induction and is defined by the number of cells which formed tetrads.
Meiotic progression was monitored by fixing 500 μl of cells in EtOH 70% and staining the nuclei with 0.5 μg/ml of DAPI. Fluorescence microscopy was used to determine the proportion of mono-, bi-, tri- and tetra-nucleated cells.
The frequency of meiotic recombination between the heteroalleles his4-N/his4-G was quantified by plating sporulated yeast cultures on YPD and SC-HIS plates. The number of yeasts growing on SC-HIS plates was counted and normalised to the total number of cells growing on YPD.
Samples were collected from sporulating cells (wild-type and Bdf1 double point mutant bdf1-Y187F-Y354F) at 0, 4 and 8 h. Cell pellets were washed and stored at -80°C. Total RNA was extracted with phenol:chloroform (Sigma Ref. 2190191) from three independent biological replicates. DNaseI treatment (Thermo, ref. AM2222) was performed using 1 μg of total RNA subsequently purified by phenol:chloroform extraction. Intact poly(A)+ RNA was isolated by using the NEB Next Poly(A)+ mRNA Magnetic Isolation kit (E7490S). cDNA libraries were constructed from 10 ng of total RNA using the NEB Next Ultra RNA Library Prep Kit for Illumina (New England Biolabs, ref. E7530S). Fragments were enriched by 15 cycles of PCR amplification. Agilent 2100 Bioanalyzer (2100, Agilent Technologies, CA) was used to assess the quality and quantity of each library. Eighteen barcoded cDNA libraries were pooled and sequenced on an Illumina HiSeqTM 2000, generating 5–8 million single reads (SR50) for each sample. The bioinformatic analysis pipeline, presented S3A Fig can be summarized as follows: reads were quality-checked using FastQC before and after cleaning by Trimmomatic applying default parameters [71]. Reads were then aligned to S. cerevisiae S288c genome build R64-1-1.82 using Bowtie2 [72]. The attribute file was downloaded from Ensembl website and modified to include yeast intronic sequences, obtained from Yeastmine. Raw read counts for each gene were calculated using HTSeqCount using defaults parameters [73]. Read count data was normalised using the DESeq2 R-package (S5 Table, [74]). DESeq2 normalised read counts were used to identify differentially expressed genes with an adjusted p-value < 0.05 and a fold change below -2 or above 2 (S5 Table). Gene ontology analysis was performed using the generic GO-term mapper developed by the Max Planck Institute (http://cpdb.molgen.mpg.de/YCPDB).
Pscan tool was used to explore the promoter of the genes which are downregulated in bdf1-Y187F-Y354F 4 and 8h after sporulation induction (http://159.149.160.88/pscan/, [75]). Binding motifs for yeast transcription factors were obtained from the JASPAR 2016 database [76] and researched in the promoter of each gene (-500 to 0 bp upstream each transcription start side). A binding motif was considered significantly over-represented when the p-value was inferior to 0.001 (Fig 4A).
The probability that an Ndt80 binding sequence was present in the promoter of bdf1-Y187F-Y354F deregulated genes was calculated using the Morpheus ROC tool [43]. This tool computes the ROC-AUC value as a measure of the enrichment of the binding sequence of a transcription factor in a list of candidate genes compared to a mock list of genes (Fig 4C).
Venn diagram was generated using Venny 2.1–0 (http://bioinfogp.cnb.csic.es/tools/venny/index.html). A list of genes potentially regulated by Ndt80 was obtained from Chu et al [35]. The statistic representation of these genes in the list of genes downregulated in the bdf1-Y187F-Y354F mutant was tested using a hypergeometric test (Fig 4D).
ChIP analyses were performed as described [19] with minor changes. Crosslinking was done with 1% EGS for 15 min and then 1% formaldehyde for 10 min before quenching for 5 min with Glycine 125 mM. Cells were lysed in lysis buffer (Hepes 50 mM pH 7.5, NaCl 140 mM, EDTA 1mM, Triton X-100 0.1%, PMSF 0.5 mM, cOmplete, TSA 100 mg/l, phosphatase inhibitor cocktail (Sigma, ref. P2850)) in a Fastprep (MP Biologicals) for three periods of 45 s. Extracts were sonicated in cycles of 20 s with intermediate incubation for 40 s over a total of 30 min (EpiShear, Active Motif). Clarified extracts were immunoprecipitated by incubating with Pan-Mouse IgG Dynabeads (LifeTechnologies, ref. 11041) or with Protein G Dynabeads (Thermo, ref. 10004D) conjugated with anti-HA antibody (Roche, ref. 11867423001) for TAP or HA tagged strains, respectively. ChIP data are presented as percent of input. In all cases, at least three biological replicates were performed to determine the standard errors in each experiment.
TAP purifications were performed as described previously [77,78]. Calmodulin eluates from the TAP-purified complexes were analysed by SDS-PAGE by using Novex 4–12% gradient gels (Invitrogen) and visualised by staining with SilverQuest Silver Staining Kit (Thermo, LC6070). Western blot analysis of the calmodulin eluates was performed using the corresponding antibodies. Protein preparation and mass spectrometry-based proteomic analyses were carried out as described in [79]. Briefly, eluted proteins were stacked as a single band in a SDS-PAGE gel (NuPAGE 4–12%, Invitrogen) and submitted to in-gel digestion using trypsin (Promega, sequencing grade). Resulting peptides were analysed by online nanoLC-MS/MS (UltiMate 3000 and Q-Exactive Plus or LTQ-Orbitrap Velos Pro, Thermo Scientific) using a 120-min gradient. Peptides and proteins were identified and quantified using MaxQuant (version 1.5.3.30, [80]) and the SGD database (November 2015 release). Proteins were quantified based on the iBAQ value [81] calculated by MaxQuant. Only proteins identified with a minimum of two unique + razor peptides were taken into account. For statistical analysis of results obtained with Bdf1-TAP triplicates, we used ProStaR [82]: proteins identified in the reverse and contaminant databases and proteins exhibiting fewer than 6 iBAQ values in a single condition (2 conditions and 2 analytical replicates per biological replicate) were discarded from the list. After log2 transformation, iBAQ values were normalised by condition-wise median centring before imputing missing values using the QRILC algorithm (missing not at random-devoted imputation method); statistical testing was conducted using limma. For a protein to be considered as a potential binding partner of Bdf1-TAP, it had to exhibit a p-value < 0.001 and a log2 (fold change) ≥ 7.
Coimmunoprecipitation was performed using 50 mL of cells grown on YPD to an OD600 of 0.5. Cells were harvested, washed with water, and resuspended in lysis buffer (HEPES 50 mM pH 7.5, NaCl 140 mM, EDTA 1 mM, glycerol 10%, NP-40 0.5%, PMSF 1 mM and protease inhibitors). An equal volume of glass beads was added. Cells were lysed in a Fastprep homogenizer (MP Biomedicals) for 45 s. The clarified extracts were immunoprecipitated during 2 h at 4°C using Dynabeads Pan Mouse IgG (LifeTechnologies). The immunoprecipitates were washed and resuspended in 50 μl of SDS-PAGE sample buffer for subsequent western blot analysis.
Reverse transcription was performed by following manufacturer instructions using kits iScript RT Supermix and Universal SYBR Green Supermix (Bio-Rad) on a CFX384 Touch qPCR machine (Bio-Rad). Primers are presented (S4 Table). At least three biological replicates were performed to determine the standard errors for each experiment. RT-qPCR data were then normalised relative to the reference gene, NUP85 [19].
RNA-seq data obtained in this study are available from the GEO repository (GSE89530). Mouse transcriptomic data was derived from RNA-seq dataset GEO GSE39909 as described in (S7 Fig, [26]). Proteomic data produced in this study have been deposited on ProteomeXchange under identifier PXD005227.
|
10.1371/journal.pntd.0004339 | Novel Characteristics of Trypanosoma brucei Guanosine 5'-monophosphate Reductase Distinct from Host Animals | The metabolic pathway of purine nucleotides in parasitic protozoa is a potent drug target for treatment of parasitemia. Guanosine 5’-monophosphate reductase (GMPR), which catalyzes the deamination of guanosine 5’-monophosphate (GMP) to inosine 5’-monophosphate (IMP), plays an important role in the interconversion of purine nucleotides to maintain the intracellular balance of their concentration. However, only a few studies on protozoan GMPR have been reported at present. Herein, we identified the GMPR in Trypanosoma brucei, a causative protozoan parasite of African trypanosomiasis, and found that the GMPR proteins were consistently localized to glycosomes in T. brucei bloodstream forms. We characterized its recombinant protein to investigate the enzymatic differences between GMPRs of T. brucei and its host animals. T. brucei GMPR was distinct in having an insertion of a tandem repeat of the cystathionine β-synthase (CBS) domain, which was absent in mammalian and bacterial GMPRs. The recombinant protein of T. brucei GMPR catalyzed the conversion of GMP to IMP in the presence of NADPH, and showed apparent affinities for both GMP and NADPH different from those of its mammalian counterparts. Interestingly, the addition of monovalent cations such as K+ and NH4+ to the enzymatic reaction increased the GMPR activity of T. brucei, whereas none of the mammalian GMPR’s was affected by these cations. The monophosphate form of the purine nucleoside analog ribavirin inhibited T. brucei GMPR activity, though mammalian GMPRs showed no or only a little inhibition by it. These results suggest that the mechanism of the GMPR reaction in T. brucei is distinct from that in the host organisms. Finally, we demonstrated the inhibitory effect of ribavirin on the proliferation of trypanosomes in a dose-dependent manner, suggesting the availability of ribavirin to develop a new therapeutic agent against African trypanosomiasis.
| Only a limited number of therapeutics for human African trypanosomiasis also known as African sleeping sickness is available today, and it narrows the choice of the drugs to escape from the side effects and the emergence of drug-resistant pathogens. The parasitic protozoa Trypanosoma brucei is the causative reagent of African trypanosomiasis, and is infective to various mammalian species. T. brucei and its mammalian hosts share almost the same metabolic machinery, and therefore it is important to understand the differences in biochemical properties of the metabolic enzymes between T. brucei and its hosts. Here we report that guanosine 5’-monophosphate reductase (GMPR) of T. brucei showed apparent differences in its primary structure and biochemical properties from those of its host counterparts, and was more sensitive to purine nucleotide analogs such as monophosphate forms of ribavirin and mizoribine than were the host GMPRs. Furthermore, ribavirin prevented the proliferation of trypanosomes in vitro. Our present findings may imply the availability of ribavirin and/or its derivatives in a treatment of African trypanosomiasis.
| In general, purine nucleotides are synthesized de novo from their precursors such as amino acids and ribose 5-phosphate, and are also produced from purine bases and ribose 5-phosphate through a salvage pathway. Guanosine 5’-monophosphate reductase (GMPR) catalyzes the reductive deamination of guanosine 5’-monophosphate (GMP) to inosine 5’-monophosphate (IMP) in the presence of NADPH, a route to recycle guanine nucleotides into adenine nucleotides [1]. GMPR has been identified in various species from bacteria to mammals including parasitic protozoa [2], and has been structurally characterized by X-ray crystallography, which indicated that GMPR belongs to the family of (β/α)8 barrel proteins also known as TIM barrel proteins. It is also known that GMPR shows high similarities in amino acid sequence and structure to inosine 5’-monophosphate dehydrogenase (IMPDH), the enzyme catalyzing the NAD+-dependent oxidation of IMP to xanthosine 5’-monophosphate (XMP); nevertheless, GMPR and IMPDH are generally distinguished by the cystathionine β-synthase (CBS) domain, which is well conserved in IMPDHs but absent in GMPRs [1]. Recent studies have demonstrated that the catalytic mechanism of E. coli GMPR follows an ordered bi-bi kinetic mechanism [3], and that the GMPR reaction uses the same intermediate E-XMP* as IMPDH, but in this reaction the intermediate reacts with ammonia instead of water [4]. However, detailed studies on GMPRs have been performed only on human and bacterial enzymes, and so the GMPRs in other organisms including protozoa are still poorly defined.
Trypanosoma brucei is a protozoan parasite and the causative agent of African trypanosomiasis, a vector-borne parasitic zoonosis known as African sleeping sickness in humans and as nagana disease in cattle. Nearly all the protozoa are incapable of de novo purine biosynthesis and depend on the purine salvage pathway, which has been regarded as an attractive chemotherapeutic target of parasitemia [5]. Indeed, T. brucei lacks the enzymatic machinery for the de novo synthesis of purine nucleotides, and therefore it solely depends on salvaging purines acquired from the extracellular environment for survival [6]. Recently, several groups have investigated the genomic information of T. brucei, and a gene registered as Tb927.5.2080 was annotated as a putative T. brucei GMPR (TbGMPR) in TriTrypDB and GeneDB [7,8]; however, the molecular identification and characterization of TbGMPR still remain to be made.
In this study, we examined the GMPR activity of the recombinant protein of the Tb927.5.2080 gene, and identified the subcellular localization of TbGMPR in T. brucei bloodstream forms. Furthermore, we compared the characteristics of TbGMPR with those of GMPRs of host animals in terms of their enzymatic kinetics and structures and found that ribavirin 5'-monophosphate, a purine nucleotide analog, was a inhibitor of T. brucei but not of its host GMPRs.
All chemicals were purchased from Wako Pure Chemical Industries, Ltd. (Osaka, Japan) or Nacalai Tesque Inc. (Kyoto, Japan), except for GMP (Sigma-Aldrich Japan K.K., Tokyo, Japan) and RMP (Toronto Research Chemicals, Ontario, Canada). MZP was synthesized as described previously [9].
T. brucei brucei ILTat 1.4, a strain with a monomorphic bloodstream form in vitro, was cultured in HMI-11 medium [10], a modified IMDM medium (Sigma-Aldrich Japan), which was supplemented with the following (mM): hypoxanthine (1.0), thymidine (0.16), cysteine (1.5), pyruvate (1.0), bathocuproine sulfonate (0.05), 2-mercaptoethanol (0.2), and 10% fetal bovine serum. Cultures were maintained in a humidified atmosphere of 5% CO2 and 95% air at 37°C. A pleomorphic GUTat 3.1 strain was cultured and maintained as described above.
Total RNA from T. brucei was prepared with Sepasol-RNA I Super (Nacalai Tesque). The cDNA was synthesized from the total RNA by using ReverTraAce and oligo-d(T)20 primers (TOYOBO, Osaka, Japan). The sequences of the primers were as follow: TbGMPR primers, 5’-GTTCCTATCGTTGGGCAAAA-3’ and 5’-ATACAAGGTACACCCCTGCG-3” α-tubulin primers, 5’-TCAAGTGCGGTATCAACTAC-3’ and 5’-AGTGCTGCAAGGTCTTCAC-3’. PCR reactions were performed with a thermal cycler PC-300 (ASTEC Co., Ltd., Fukuoka, Japan) operated under the following conditions: 95°C for 3 min; 30 cycles of 95°C for 10 s, 55°C for 10 s, and 72°C for 30 s, with an additional extension at 72°C for 5 min. The reaction products were analyzed by electrophoresis on agarose gels, and subsequently by PCR-direct sequencing.
The coding region of the TbGMPR gene was amplified by using the genomic DNA of T. brucei as a template, with primers 5’-CGGAATTCATGTCCTTCAATGAATCGGCA-3’ and 5’-CGGTCGACTTAAAGTTTGGCAACACCGTGA-3’. The PCR product was cloned between Eco RI and Sal I sites of the expression vector pGEX 6P-1 (GE Healthcare Japan Co., Tokyo). The recombinant TbGMPR was expressed as an N-terminal glutathione S-transferase fusion protein in E. coli BL21 (DE3). The water-soluble fraction of the cell extract was applied onto an Affi-Gel Blue Gel 100–200 mesh (Bio-Rad) equilibrated with 50 mM Tris-HCl (pH 8.0) containing 150 mM NaCl, 1 mM EDTA and 1 mM DTT. The column was washed with the same buffer containing 1 M KCl, and the recombinant protein was eluted with the buffer containing 3 M KCl. The eluted protein was further purified with glutathione-Sepharose 4B and PreScission protease (GE Healthcare Japan) according to manufacturer’s instructions. The purified protein was dialyzed against 50 mM Sodium phosphate buffer (pH 7.0) containing 100 mM KCl, 3 mM EDTA, 1 mM DTT, and 5% glycerol. The concentration of the purified recombinant protein was determined with BCA Protein Assay Reagent (Thermo Fisher Scientific) by using bovine serum albumin as a standard, and the purity was analyzed by SDS-PAGE.
The mixture for the GMPR reaction consisted of 50 mM Tris-HCl (pH 8.0), 100 mM KCl, 3 mM EDTA, 500 μM GMP, and 500 μM NADPH. Following pre-incubation at 37°C, the reaction was initiated by adding recombinant enzyme at 0.5 ng/ml. The reaction was terminated by filtering out the enzyme from the reaction mixture with VIVASPIN 500 (Sartorius AG, Goettingen, Germany), and 10 μl of the filtrate was subsequently subjected to analysis with an HPLC system (LaChrom Elite L2000, Hitachi High-Technologies Co., Tokyo, Japan) equipped with a Cadenza CD-C18 column (2.0 mm x 150 mm x 3 μm, Imtakt Co., Kyoto, Japan). The following mobile phases were prepared separately: solvent A contained PIC-A reagent (Waters) diluted according to the manufacturer's manual. Solvent B consisted of acetonitrile containing 0.1% trifluoroacetic acid. Both solvents were pumped at a flow rate of 190 μl/min under the following gradient protocol: The concentration of solvent B was gradually increased from 2.5 to 20% for the first 15 min, from 20 to 50% for the next 6 min, and finally kept at 70% for 10 min. The eluents were monitored through a UV detector (Waters) with a wavelength at 254 nm. Commercial IMP, GMP, NADPH, and NADP+ were used as standards to identify and quantify the corresponding compounds in the reaction mixture.
Immunization of a rabbit with the purified recombinant TbGMPR as an antigen, and preparation of the antiserum were performed by Medical and Biological Laboratories Co., Ltd. (Nagoya, Japan). Whole IgG was fractionated from the antiserum by ammonium sulfate precipitation followed by Ab-Rapid SPiN EX column chromatography (ProteNova Co. Ltd., Kagawa, Japan). Crude lysates of T. brucei were prepared in RIPA buffer and subsequent centrifugation. The lysates were separated on a 10% SDS–PAGE gel and transferred to a PVDF membrane (Bio-Rad). The membrane was then incubated overnight at 4°C with 5 μg/ml of the anti-TbGMPR polyclonal IgG in Tris-buffered saline with 0.05% Tween-20 (TBST) containing 5% skim milk (Wako). After having been washed with TBST, the blots were incubated with anti-rabbit IgG conjugated with horseradish peroxidase (Cell Signaling Technologies, Danvers, MA). The immunoreactivities were visualized by using ImmunoStar Zeta (Wako) and LAS4000 imaging device (GE Healthcare Japan).
T. brucei (5.0 x 105 cells) were washed with PBS and fixed with 5% formalin in PBS for 15 min on ice, and then adhered to PLATINUM-coated glass slides (Matsunami Glass Ind., Ltd., Osaka, Japan). Cells were permeabilized for 20 min with 0.3% (v/v) Triton X-100 in PBS, followed by blocking for 10 min with 50% (v/v) fetal bovine serum in PBS. They were then incubated with polyclonal rabbit anti-TbGMPR (5 μg/ml) for 45 min and, after a serial washing with PBS, were incubated for 45 min with secondary antibody conjugated with Alexa Fluor 488 (goat anti-rabbit IgG; Life Technologies) at a 1: 200 dilution. Subsequently, the cells were washed with PBS and mounted in glycerol containing DAPI (1 μg/ml). As a negative control, staining was performed as described above but in the absence of primary antibody. Differential interference contrast and fluorescence optical images were captured under nonsaturating conditions by using a confocal laser scanning microscope LSM-700 (Zeiss).
For immunoelectron microscopy, T. brucei (1.3 x 108 cells) were fixed with 4% paraformaldehyde and 0.1% glutaraldehyde in 0.1 M phosphate buffer (PB, pH 7.6) at 4°C for 30 min. They were then washed with 0.1 M PB and embedded in 1% agarose. Next, the samples were sequentially treated with 20% sucrose in PB for cryoprotection, embedded in TISSU MOUNT compound (Shiraimatsu co., LTD., Osaka, Japan), and sectioned at a 10-μm thickness on a cryostat. After having been dried on glass slides, the sections were rinsed in PBS containing 0.1% Triton X-100 and treated with 10% normal goat serum in PBS for 10 min. Then they were incubated with anti-TbGMPR antibody (10 μg/ml) at 4°C overnight. After washing in PBS, the sections were incubated at 4°C overnight with nanogold-conjugated anti-rabbit IgG (Nanoprobes, NY) at a 1: 200 dilution. After a washing with PB, the gold particles were enhanced by using an HQ Silver enhancement kit (Nanoprobes) and then post fixed with 1% osmium tetraoxide in 0.1M PB at 4°C for 1 h. After the samples had been dehydrated by passage through a graded series of ethanol, they were embedded in epoxy resin. Ultrathin sections were cut and examined by electron microscopy (Hitachi High-Technologies Co.).
GMPR activity was measured at 35°C in a buffer containing 50 mM sodium phosphate buffer (pH 7.0) including 100 mM KCl, 3 mM EDTA, and 1 mM DTT. The concentrations of GMP (0 to 1,000 μM) or NADPH (0 to 150 μM) were varied to determine the kinetic parameters for each compound. The activities were measured as NADPH consumption as described above. The initial velocity data (n = 3) were fitted to the Michaelis-Menten equation by using Origin 6.0 (Microcal, Northampton, MA) or IGOR Pro 6.3 (WaveMetrics, Inc., OR) [11].
The activity of recombinant TbGMPR was examined at different temperatures in 50 mM sodium phosphate buffer (pH 7.0), with the ionic strengths adjusted to 0.2 by adding NaCl. Reactions were initiated by adding 100 nM enzyme to each buffer containing 1 mM GMP and 200 μM NADPH. Measurements of TbGMPR activity at different pHs were performed in 50 mM sodium phosphate buffer (pH 6.5 to 8.0) or 50 mM sodium borate buffer (pH 8.0 to 9.0). GMPR activity was defined as NADPH consumption, which was monitored at 340 nm (ɛ340 = 6220 M-1·cm-1) with a DU-800 spectrophotometer (Beckman Coulter). All measurements were performed in triplicate.
The effects of monovalent cations on the activity of recombinant TbGMPR were examined at 35°C in 50 mM sodium phosphate buffer (pH 7.0) containing 1 mM GMP, 200 μM NADPH, and 150 mM monovalent chloride (NaCl, KCl or NH4Cl). The reaction was initiated by the addition of 100 nM enzyme, and the NADPH concentrations in the reaction mixture were monitored as described above. The GMPR activities of human and bovine enzymes in the presence of 150 mM NaCl or KCl were examined by using the same procedure, except that the NADPH concentration was 800 μM.
Inhibitory activities of RMP and MZP toward GMPRs were tested by the same procedure applied for the determination of kinetic parameters, except that the concentrations of GMP (1,000 μM) and NADPH (200 μM for TbGMPR, and 800 μM for mammalian GMPRs) were fixed unless otherwise specified. Various concentrations of RMP or MZP were added 5 min prior to the reaction initiation, and subsequent NADPH consumption was monitored with a spectrophotometer. The initial velocity data (n = 3) at the various concentrations of the compounds were fitted to the Hill equation to calculate the IC50 values. For RMP, the data were plotted on Lineweaver-Burk plots, and fitted to a competitive inhibition equation to calculate the Ki values [11].
T. brucei GUTat 3.1 cultured in HMI-11 medium was inoculated to the fresh medium and placed on 96-well plates at a density of 2 x 103 cells/mL in the presence of ribavirin. PBS was used as a control. The final concentrations of the drug were ranged from 1 to 1,000 μM. The parasites were incubated for 72 h in 5% CO2 and 95% air at 37°C. Four hours prior to finish the incubation, 1/10 vol. of 1.25 mg/mL resazurin was added. The number of the parasites in each well after treatment was measured by colorimetric method on a plate reader Benchmark Plus (BIO-RAD), by subtracting the absorbance at 600 nm from those at 570 nm. The cell densities of the drug treatment groups were obtained relative to the control group (n = 4).
In the study on the GMPR activities in the presence of the monovalent cations, the significance of differences in the enzymatic activities between the NaCl group and others was analyzed by using one-way ANOVA followed by Dunnett’s test. The differences were considered as significant when the p values were below 0.05 (n = 3).
The molecular characterization of GMPR in trypanosomatids had remained unreported until now. Our sequence alignment analysis showed that all the putative GMPRs of trypanosomatids, including the Tb927.5.2080 gene of T. brucei, shared a unique insertion of about 130 amino acid residues; and a domain search on PROSITE [12] revealed that this insertion possibly corresponded to a tandem repeat of CBS domains (S1 and S2 Figs), which are absent in human, bovine, and E. coli GMPRs.
We performed RT-PCR analysis to clarify endogenous mRNA expression of the Tb927.5.2080 gene in T. brucei. Total RNA was extracted from bloodstream forms of T. brucei in culture, and used as a template. RT-PCR analysis with primers specific for the Tb927.5.2080 gene produced a single-sized product (Fig 1A), and subsequent sequencing analysis of this PCR product revealed that it was certainly derived from the mRNAs of the Tb927.5.2080 gene.
We then generated a recombinant TbGMPR by use of an E. coli expression system and found that the purified recombinant protein exhibited a single band on an SDS-PAGE gel with the expected molecular weight of approximately 53 kDa (Fig 1B). The purified protein was used as an antigen to raise a polyclonal antibody against TbGMPR. By Western blotting against the whole cell extracts of cultured T. brucei with prepared antibody, a single band was detected at a position of approximately 53 kDa which corresponded to the size of the recombinant TbGMPR (Fig 1C).
To examine the GMPR activity of the Tb927.5.2080 gene product, we incubated the recombinant protein with GMP and NADPH, a substrate and a co-enzyme of GMPRs, and subsequently analyzed the reaction mixture by HPLC at various time points. The absorbance at retention times of 15.0 and 27.2 min (corresponding to those of authentic GMP and NADPH, respectively) decreased in a time-dependent manner; and conversely, the absorbance at 15.6 and 24.4 min (corresponding to those of authentic IMP and NADP+, respectively) increased with the same time course (Fig 2). These results indicate that the recombinant protein of Tb927.5.2080 gene catalyzed the conversion of GMP into IMP in the presence of NADPH, meaning that Tb927.5.2080 actually encoded the GMPR of T. brucei. Based on the above results, we designated the Tb927.5.2080 gene and its encoded protein as TbGMPR.
Immunofluorescent staining with the anti-TbGMPR polyclonal antibody showed that the fluorescent signals were distributed throughout the cells of T. brucei bloodstream forms in culture and seemed to be located in vesicle-like organelles (Fig 3A). Further anatomical study was performed by immunoelectron microscopy, and the majority of the immunoreactivities were found in glycosomes (Fig 3B). All GMPRs in trypanosomatids possess a peroxisomal-targeting signal (Ala/Ser-Lys-Leu) on their C-terminus (S2 Fig), which is commonly found in glycosomal proteins of T. brucei [13,14]. Taken together, all our results clearly demonstrate that GMPR was consistent to be localized in the glycosomes of T. brucei bloodstream forms, though we did not use antibodies against bona fide glycosomal markers.
In order to characterize the kinetic properties of TbGMPR, we calculated kinetic constants for the recombinant enzyme. The initial velocities at a various concentrations of GMP or NADPH were collected and processed according to Michaelis-Menten equation (Fig 4). The steady-state kinetic parameters of TbGMPR were determined to be the following: kcat value, 0.519 ± 0.012 s-1, and Km values for GMP and NADPH, 89.3 ± 9.0 μM and 12.3 ± 0.8 μM, respectively (Table 1). We also determined the steady-state kinetic parameters of recombinant GMPRs of human and bovine enzymes, as summarized in Table 1. Based on these data, TbGMPR had lower affinity for GMP than either human or bovine GMPRs. On the other hand, NADPH bound to TbGMPR with higher affinity than did the mammalian GMPRs. Consequently, the differences in the affinities for the substrate and co-enzyme of GMPRs among species raised the kcat value of TbGMPR about 2-fold in comparison to mammalian values.
The enzymatic properties of the recombinant TbGMPR at various temperatures and pHs were assessed to determine the optimal reaction conditions for analysis of steady-state kinetics of the protein. In terms of temperature dependency, TbGMPR exhibited the maximum enzymatic activity of 26.5 ± 0.8 (nM·s-1) at 38°C (Fig 5A). The enzymatic activity of the recombinant TbGMPR increased as the pH was reduced from pH 9.0 down to 6.5 (Fig 5B). The isoelectric point of TbGMPR was expected to be pI 9.7 from its amino acid sequence, and this value fell within those of common glycosomal proteins ranging from pI 8.8 to 10.2 [15]. It has been reported that the glycosomal pH of T. brucei procyclic forms is around 7.4 [16]. Our results with the previous findings suggest that TbGMPR is able to exhibit its enzymatic activity in T. brucei glycosomes.
We examined the effects of various monovalent cation chlorides such as NaCl, KCl, and NH4Cl on the activity of the recombinant TbGMPR. The activity of TbGMPR in the presence of NaCl was determined as 21.7 ± 1.2 nM·s-1 under the conditions employed here (Fig 6). Interestingly, TbGMPR activities were increased to 43.8 ± 0.2 and 30.6 ± 0.5 nM·s-1 when NaCl was replaced with KCl and NH4Cl, respectively. Furthermore, neither human nor bovine GMPRs were activated in the presence of KCl instead of NaCl. As opposed to the activity of TbGMPR, the activities of the mammalian GMPRs in NH4Cl were significantly decreased, being 61.1 to 81.6% of those in NaCl. TbGMPR showed significantly higher homologies to mammalian IMPDHs than to other GMPRs on BLAST analysis (S1 Table). The sequence alignment of IMPDHs and GMPRs revealed that the amino acid residues that interact with the potassium ion were conserved in TbGMPR and in all IMPDHs but not in other GMPRs examined in this study (Fig 7), supporting our findings of TbGMPR activation by potassium ions. This was the first observation of the activation of GMPR in the presence of monovalent cations.
Next we evaluated the inhibitory effects on GMPRs of ribavirin 5’-monophosphate (RMP) and mizoribine 5’-monophosphate (MZP), which are purine nucleotide analogs known as potent inhibitors of IMPDHs that share a similar conformation with GMPRs [4,11]. Inhibitory kinetic analysis revealed that RMP inhibited TbGMPR with an IC50 value of 101.8 ± 0.9 μM; however, no or little inhibition of human and bovine GMPRs was observed in the presence of RMP (Fig 8A). Lineweaver-Burk plots showed that RMP inhibited TbGMPR in a competitive manner, and the Ki value was determined to be 4.46 ± 0.46 μM (Fig 8B). MZP also inhibited TbGMPR (IC50 = 23.7 ± 1.9 μM), though it also exhibited moderate inhibition of HsGMPR1 (119.9 ± 16.1 μM) and HsGMPR2 (69.3 ± 7.8 μM).
To investigate the inhibitory effect of ribavirin on trypanosome proliferation, trypanosomes were cultured with various concentrations of ribavirin for 72 h and the numbers of the living cells relative to the vehicle treatment were examined by using resazurin. The proliferation of trypanosomes was inhibited by the addition of ribavirin in a dose-dependent manner (Fig 9). The curve fitting to Hill equation revealed that the IC50 value of ribavirin against trypanosomes was estimated to be 25.4 ± 3.9 μM.
In this study, we identified the GMPR in T. brucei for the first time and demonstrated its unique properties, which were distinct from those of GMPRs in host animals in terms of protein structure, enzymatic property, and cytological distribution.
So far, 2 genes in T. brucei, Tb10.61.0150 and Tb927.5.2080, have been annotated as IMPDH in the NCBI database, whereas none have been identified as GMPR. Previously we showed that the former gene encodes IMPDH of T. brucei [11]; however, the product of the latter gene has not been characterized at the molecular level. Nonetheless, some databases such as TritrypDB and GeneDB annotate this gene as a putative GMPR inferred from its sequence orthology [7,8]. We observed the mRNA expression of Tb927.5.2080 gene in cultured T. brucei, and this was well corresponding to the previous findings obtained from transcriptome analyses [19–23]. Our present study clearly demonstrated the GMPR activity of the recombinant protein of the Tb927.5.2080 gene, and so this is the first molecular identification of GMPR in T. brucei. On the other hand, TbGMPR showed no enzymatic activity to catalyze IMP to XMP, despite TbGMPR had higher homology to IMPDHs than to GMPRs of mammals on our BLAST search (S1 Table). All IMPDHs annotated to date share 5 amino acid residues for IMP-binding [18], and yet Tyr411 and Gly415 (HsIMPDH2 numbering) of those residues are substituted in TbGMPR to Ile398 and Ala402, respectively. These differences may account for the lack of IMPDH activity in TbGMPR.
We also examined the expression and localization of TbGMPR and found that TbGMPR was dominantly localized in the glycosomes of T. brucei bloodstream forms in culture. Glycosomes are peroxisome-like vesicles that contain the machineries for glycolysis and some parts of purine nucleotide metabolism in trypanosomatids. TbGMPR possesses a signal peptide sequence to allow translocation into glycosomes. TbGMPR has been found in the glycosomal extract prepared from the procyclic forms of T. brucei [24]; however, recent proteome analysis failed to detect the enzyme in the glycosomes of bloodstream forms [25]. Therefore, our present data are the first observation that TbGMPR is consistently localized in the glycosomes of T. brucei bloodstream forms.
In this study, we showed that TbGMPR possessed a unique insertion of a single pair of putative CBS domains that were absent in GMPRs of species besides trypanosomatids. The CBS domain pairs have been found in proteins such as cystathionine β-synthase, AMP-activated protein kinase, and IMPDH; and they are known to bind adenine nucleotides to function as sensors for intracellular metabolites [26]. AMP-activated protein kinase is involved in switching ATP generation and consumption by sensing AMP concentrations through its CBS domains, and mutations in these domains have been shown to cause severe heart defects known as Wolff-Parkinson-White syndrome [27,28]. CBS domains of IMPDHs of various species including T. brucei [29] are able to bind adenine nucleotides [28,30], and a mutation in the CBS domain is associated with retinitis pigmentosa in humans [31,32]. These previous findings together with our results suggest that the CBS pair of GMPR, as that of IMPDH, functions as a sensor of adenosine nucleotides in T. brucei, which could be a novel function specific to trypanosomal GMPRs.
We showed here that TbGMPR was activated by the addition of monovalent cations such as K+ and NH4+. No GMPRs so far except IMPDHs have been found to be activated in the presence of K+ and NH4+, though E. coli GMPR was reported to catalyze the opposite reaction in the presence of NH4+ to form GMP from IMP [4]. Our study of the recombinant GMPR enzyme of T. brucei is the first to show the enhancement of GMPR activity in the presence of monovalent cations. Both GMPR and IMPDH in mammals are known as homotetramers; and especially in IMPDH, the Cys loop existing between the β6 strand and the α6 helix forms a half of the monovalent cation-binding site, whereas the another half is contributed by the end of a C-terminal α-helix from an adjacent monomer [1,18]. Our sequence alignment of the ion-binding sites of mammalian IMPDHs with the corresponding regions of GMPRs showed that the TbGMPR had higher similarity to IMPDHs than to mammalian GMPRs. This observation is consistent with our findings showing that TbGMPR, but neither human nor bovine GMPRs, was activated by monovalent cations, suggesting that either K+ or NH4+ functions as a molecular “lubricant” to facilitate a conformational change in TbGMPR, as suggested in the case of IMPDHs of various species [18]. These structural and functional similarities between GMPRs and IMPDHs in trypanosomatids might have occurred during the evolution from their ancestors, and further investigation might provide a new knowledge regarding the molecular phylogenetics of protozoa.
Our studies with purine nucleotide analogs and recombinant GMPRs clearly showed that RMP was an competitive inhibitor of TbGMPR, with a Ki value similar to that observed for the inhibition of T. brucei IMPDH (Ki = 3.2 ± 0.16 μM) [11]. It has been found that GMPR of E. coli is inhibited by RMP with a Ki value of 98 ± 25 μM [4], though this value is about 20-fold higher than that for TbGMPR (4.46 ± 0.46 μM). Interestingly, RMP showed no or little inhibition of human and bovine GMPRs, whereas MZP, another purine nucleotide analog tested, inhibited the activities of both T. brucei and human GMPRs. These compounds are known to inhibit IMPDHs in a competitive manner by interfering with the IMP binding site [18]; and therefore, it is of considerable interest that these compounds also share the same binding site on TbGMPR. Our findings that RMP preferentially inhibited TbGMPR rather than mammalian GMPRs provide the knowledge to aid in the design of novel species-specific inhibitors of GMPRs. In addition, TbGMPR had higher Km and kcat values for GMP as compared with human and bovine GMPRs or with E. coli GMPR [4]. These findings suggest that the reaction center of trypanosomal GMPR has a unique structure among species, and further structural analysis will be necessary to examine this possibility.
In the present study, we found the inhibitory effect of ribavirin on the proliferation of T. brucei. Ribavirin is widely used in the treatment of hepatitis C, and is believed to act as its nucleotide forms through the phosphorylation by intracellular kinase activities. We have previously reported that RMP is a potent inhibitor of TbIMPDH [11], in addition, we demonstrated here that RMP also inhibited TbGMPR more potently than its mammalian counterparts. Trypanosomes are known to have purine nucleoside transporters [33] and adenosine kinase [34,35], which participate in the RMP production from the extracellular ribavirin in the cells [36]. These findings suggest that ribavirin taken up to the parasite cells is converted to RMP and exerts its anti-trypaosomal activity through the inhibition of TbGMPR and/or TbIMPDH; however, another possibility still remains that ribavirin interferes the post-transcriptional modification via inhibiting the capping guanylation of mRNA [37]. Further investigation is required to ascertain whether these enzymes are essential to trypanosomes, and if proven so TbGMPR and/or TbIMPDH can be possible targets of ribavirin.
In conclusion, we clearly demonstrated in this study that GMPR of T. brucei apparently had characteristics distinct from those of its orthologs in the host animals and that the purine nucleotide analog RMP inhibited TbGMPR but not the host enzymes. Further studies to improve the potency and specificity of such inhibitors might lead to new therapeutics against African trypanosomiasis.
|
10.1371/journal.pntd.0005570 | External quality assessment study for ebolavirus PCR-diagnostic promotes international preparedness during the 2014 – 2016 Ebola outbreak in West Africa | During the recent Ebola outbreak in West Africa several international mobile laboratories were deployed to the mainly affected countries Guinea, Sierra Leone and Liberia to provide ebolavirus diagnostic capacity. Additionally, imported cases and small outbreaks in other countries required global preparedness for Ebola diagnostics. Detection of viral RNA by reverse transcription polymerase chain reaction has proven effective for diagnosis of ebolavirus disease and several assays are available. However, reliability of these assays is largely unknown and requires serious evaluation. Therefore, a proficiency test panel of 11 samples was generated and distributed on a global scale. Panels were analyzed by 83 expert laboratories and 106 data sets were returned. From these 78 results were rated optimal and 3 acceptable, 25 indicated need for improvement. While performance of the laboratories deployed to West Africa was superior to the overall performance there was no significant difference between the different assays applied.
| For the highly infectious and deadly ebolavirus disease (EVD) to date neither specific treatment nor vaccines are available. Rapid and adequate isolation of patients is the only option to contain and to combat spreading of the disease. Reliable and sensitive diagnosis that allows efficient identification of infected individuals is a pre-requisite for outbreak management. External Quality Assurance (EQA) studies are a vital tool to assess individual diagnostic laboratory performance particularly important during the outbreak of novel emerging infections. Therefore, a panel of inactivated ebolavirus samples was generated in order to perform an EQA for ebolavirus diagnostic during the recent outbreak in West Africa to assess performance of mobile laboratories sent to the outbreak countries from different parts of the world. Further, the panel was provided to laboratories in other parts of the world to improve global preparedness in case EVD would spread through international travel or evacuation of infected international staff members deployed to West Africa to fight the disease. While 73.6% of all results reported during this study were rated optimal the performance of the laboratories from the outbreak countries was even better with 82.1% of the results rated optimal.
| The Ebola outbreak in West Africa that started in December 2013 in the southeast of Guinea [1] has developed into the largest yet documented outbreak. While the human infection initiating this outbreak most likely was a zoonotic bat to human transmission of a Zaire ebolavirus variant [1, 2] subsequent spreading of ebolavirus disease (EVD) occurred via infected bodily fluids through close human-to-human contact. The core area of the outbreak was limited to the three most affected countries Guinea, Liberia, and Sierra Leone. The virus variant meanwhile has been named Makona (EBOV/Mak) after the Makona River in the Guinea/Liberia/Sierra Leone border region [3]. Close to 29 000 individuals were infected with EBOV/Mak and more than 11 300 died from EVD as of March 27, 2016. But imported cases and small outbreaks were also reported in Mali, Nigeria, Senegal, Europe and the United States [4]. EVD symptoms which can comprise high fever, nausea, vomiting, diarrhea, exanthema, coughing, and hemorrhage are highly unspecific, in particular in a region where malaria is highly endemic and where also other infectious diseases occur that might interfere with a clinical diagnosis of EVD [5]. Therefore, a quick and reliable diagnostic of suspected patients is of high priority to identify, isolate and treat infectious patients. Detection of viral RNA by reverse transcription polymerase chain reaction (RT-PCR) has proven effective for diagnosis of ebolavirus infection from acute cases since serology is only useful in the later stage of illness.
Several ebolavirus-specific RT-PCR assays have been published and commercial assays are available as well [6]. However the quality of these assays in particular regarding sensitivity and specificity are largely unknown and require a serious evaluation.
External Quality Assurance (EQA) studies are a vital tool to assess individual diagnostic laboratory performance and became especially important to assess technical capacities during outbreaks of novel emerging infections. Participants could benchmark the quality of their diagnostic performance, identify possible weaknesses and improve their diagnostic capabilities accordingly, allowing the most accurate EVD diagnostic in order to rapidly identify and isolate new cases.
Participating laboratories from Africa were nominated by the WHO Geneva office with a strong focus on the outbreak countries. Ebola-PCR EQA panels were also distributed among the members of the European Network for Diagnostic of Imported Viral Diseases (ENIVD), the German National Laboratory Network for Diagnostic of Biothreat Agents (NaLaDiBa) and the Global Health Security Action Group Laboratory Network (GHSAG-LN). Participation was free of charge. Laboratories were coded and after evaluation of the results participants received a table with all data sets but only their own laboratory was identified.
Viruses were grown on Vero E6 cells. Supernatant was inactivated by heat treatment (1h, 56°C), subsequent gamma irradiation on dry ice at 25–30 kgray (Synergy Health Radeberg GmbH, Radeberg, Germany), and tested for inactivation by cultivation in tissue culture. Cultures were passaged three times on Vero E6 cells. In supernatants no replication of virus was detected by specific real-time RT-PCR, thus confirming absence of infectivity. Inactivated virus stocks were stored at -80°C until further use.
To determine sensitivity of the ebolavirus diagnostic performed by the participating laboratories a 10-fold serial dilution of the Zaire ebolavirus (EBOV Gabon 2003, Genbank Acc. No. EF490230) preparation in distilled water and lyophylisation reagent (OPS Diagnostics, Lebanon, USA) was generated. Dilutions from 10−2 to 10−6 as well as 10−3 and 10−4 dilution steps of an early field isolate from the outbreak region Mak-C05 (GIN/2014/Makona-Guéckédou-C05, GenBank Acc. No. KJ660348), were included. To test for reproducibility the 10−4 sample of this isolate was included in duplicate. Marburgvirus isolate Popp (Genbank Acc.No. Z29337) was included as a 10−3 dilution of the virus stock. Two negative controls contained human plasma from blood donors. Aliquots of 100 μl of virus were freeze-dried together with 100 μl of 2x Lyophilization reagent (Ops diagnostics, NJ, USA) in 0.5ml glass vials with plugs (SP Industries, USA) in a freeze dryer (Epsilon 2-6D, Martin Christ Gefriertrocknungsanlagen GmbH, Germany). The samples of the panel were encoded with randomly distributed numbers from 1 to 11 (Fig 1) and stored at 4°C in the dark.
Sets of freeze-dried samples were pre-tested by three expert laboratories. Stability of samples was verified after 3 months at 4°C and an additional 4 weeks at room temperature (approx. 22°C) and reconfirmed after 6 months at 4°C. No obvious loss of genome copies was detected.
Due to the continuing demand for the Ebola EQA panel preparation of a second set of panels became necessary. Starting from the inactivated virus stocks used for panel-1 pre-dilutions identical to panel-1 were made. To allow joint evaluation of results great care was taken to prepare panel-2 according to the same specifications as panel-1. Therefore, both panels were pre-tested by real-time RT PCR side by side. While most corresponding samples of the 2 panels showed little variation of Cq values the final dilution of Zaire ebolavirus (10−6) for panel-2 was very close to the detection limit of the PCR assay. Therefore, from this dilution step only 1 out of 4 aliquots was positive (Cq 35) while 3 were negative (Cq 45). Similar results were obtained from the three pre-test sites. In addition, sample numbering between the final two dilution steps (10−5 and 10−6) were exchanged. Panel-2 was also stored at 4°C and stability was confirmed by PCR after one month and repeatedly up to 14 months.
Ebolavirus EQA panels were shipped with appropriate documentation in small zip-lock bags with desiccation bags at ambient temperature either with regular mail or by courier service.
Performance of molecular diagnosis of virus infection is based on specificity, sensitivity, and reproducibility of the applied assays and reliability largely depends on the prevention of cross-contamination. To test for these parameters a panel of samples was established from inactivated stocks of filoviruses (Fig 1). In order to speed up the preparation process a 10-fold serial dilution was established of a Zaire ebolavirus isolate (EBOV Gabon 2003) that previously had been inactivated and tested as an Ebola standard. This dilution series was used in the panel to test for sensitivity and to establish the limit of detection for the individual participant. At the time of conception of the EQA, only limited sequence information on the outbreak strain and potential sequence divergence from former ebolaviruses were available. Therefore, one first human isolate Mak-C05 from the current outbreak originating in Guéckédou, Guinea was included. Laboratories from Africa, and in particular field laboratories in the outbreak region were identified by WHO for participation in the EQA. The EQA panel was also offered to members of ENIVD, NaLaDiBA, and GHSAG-LN. In addition, the panel was distributed globally to interested laboratories responsible for ebolavirus diagnostic in their respective countries. Panels were always sent without any refrigeration, participants were asked to resolve the lyophilized material in 100 μl of sterile bidest. water prior to extraction and to handle the resolved samples as regular serum samples that potentially might contain ebolavirus. They were instructed to extract the entire sample and to analyze all samples for the presence of ebolavirus RNA genome according to their established protocol, to report their results directly to RKI, and to include information on the extraction method and PCR assay used.
A total of 106 data sets with PCR results for either Ebola PCR EQA panel-1 or panel-2 were returned by 83 labs in 42 countries (see acknowledgments). Of these 28 data sets were from 21 laboratories working in Sierra Leone and Guinea during the outbreak. While for 12 data sets samples were reported only as PCR-positive or PCR-negative all other 94 results were given with Ct values (Cq according to MIQE guidelines) [7] for samples tested positive. Participants were anonymized with numbers in the order of the arrival of results. Results for both panels were compiled in Fig 2. Several laboratories applied more than one assay to the panel and reported 2 or more results. These results were coded with the participant’s number discriminated by a, b, c and d, as required. While approx. half of the laboratories performed a diagnostic that also allowed identification of Marburgvirus the remaining participants reported this sample as “ebolavirus negative”. Since this study was specifically designed for the Ebola outbreak in West Africa participants were only asked to analyze the samples of the EQA panel for ebolavirus. Consequently a result for the Marburgvirus sample was also assessed “correct” if just reported as “ebolavirus negative”.
From the 106 data sets reported 55 (51.9%) were obtained with commercial assays (Table 1). From the remaining 51 data sets 26 results were obtained with 9 different in-house assays. Further 15 results were obtained with in-house assays but no reference was given. For 10 results no or too little information was available to determine if they had been obtained with commercial or with in-house assays.
For the evaluation of this ebolavirus PCR EQA inaccurate results were rated with weighted inaccuracy points (Fig 2), 1 point for the highest dilution of the Zaire ebolavirus dilution series (marked orange), 2 points for the second highest dilution and for every other result not correctly analyzed 3 points (all marked red). The final dilution step (10−6) for Zaire ebolavirus in panel-2 (#4) was slightly more diluted than the corresponding sample in panel-1 (#5) and in the hands of the pre-test sites this sample was not reliably tested positive in RT qPCR. Therefore, a negative result for this sample was not rated as false-negative (marked in grey) and consequently was not attributed inaccuracy points. However, RT qPCR results for the second highest dilution (10−5) of panel-2 (#5) matched results for the second highest dilution from panel-1 (#4) in pre-tests. Therefore a false-negative result for this sample in panel-2 was also attributed 2 inaccuracy points. Results were ranked according to points. Identical scores were ranked according to arrival date.
78 data sets (73.6%) reported by 67 participants were rated optimal since all samples were identified correctly. Three additional results were rated acceptable since only the highest dilution of the panel had been analyzed false-negative (1 inaccuracy point) suggesting slightly reduced sensitivity for the assay performed. The remaining 25 results (23.6%) reported by 23 participants showed a clear need for improvement (2 to 9 inaccuracy points). However, out of these 23 participants 8 had reported additional results with an alternative assay that was rated optimal. This reduces the number of laboratories with urgent need for an improved assay to 15. Out of the 25 results not rated optimal or acceptable 10 were obtained with commercial, 13 with in-house assays. For the remaining 2 results there was no information available on the assay used.
All participants either identified the Marburgvirus sample correctly as PCR positive for Marburgvirus or reported it as “ebolavirus negative”. Only one participant reported a false-positive result for one of the two negative controls, all other participants identified the negative samples correctly.
Most participants used real-time RT-PCR assays for their analyses. While 102 data sets came with Cq values for positive samples only for 10 results copy numbers had been calculated and consequently quantification could not be considered for evaluation. Therefore, Cq values were taken as a semi-quantitative indicator and used for statistical analysis. For this purpose all negative results were translated into a Cq value of 45, since in many laboratories real-time PCR is routinely performed with a maximum of 45 cycles. For all positive samples Cq values were taken as reported but reduced to 1 position after decimal point if required. Since not all participants reported details on extraction procedure (e.g. extraction kit, elution volume) and volume of purified RNA applied to the PCR reaction this approach introduces an inaccuracy. Further, a few laboratories did not follow all the instructions provided with the panel and they only extracted part of the sample or pre-diluted the sample prior to testing. However, most participants (as far as indicated in the reports) used the entire sample for extraction and took 1/10 to 1/12 of the extracted RNA for PCR analysis. Additional variations might have been introduced since several different PCR assays and different real-time PCR instruments were used. These parameters might to some extend affect amplification efficiency and Cq value obtained for a given sample. Despite these ambiguities box-plot analysis showed a good over-all correlation for the panel between dilution and the mean for Cq values while overall variation of Cq values for a sample in the panel increased with the degree of dilution seen by an increase in the size of the boxes representing the 25 to 75 percentile (Fig 3). Reproducibility of results was tested with the identical samples #2 and #7. Good reproducibility was seen for 93 data sets. For these the difference between Cq values was not more than 1. For the remaining 13 results only for 4 data sets the difference was more than 2.
This study was initiated in late summer of 2014 in the light of the rapidly developing ebolavirus outbreak in West Africa with already high numbers of cases, disastrous predictions for the future development, and under the threat of global spreading through international air travel [17]. At present identification can be achieved best with sensitive, reliable, and rapid diagnostic PCR assays [18]. While WHO aimed at evaluating various laboratories that had been deployed to the outbreak countries in West Africa in the combined international effort to fight the disease, an additional motivation of RKI, ENIVD, NaLaDiBa, GHSAG-LN and WHO was to promote preparedness and evaluate quality of ebolavirus diagnostic beyond the outbreak region and on a global scale by supporting the development of diagnostic capacities and improving capabilities to allow rapid diagnosis of potentially emerging suspect cases.
In infected individuals ebolavirus can be detected early in the course of the disease and rapidly reaches high viral loads. In the absence of reliable rapid tests the method of choice for diagnosis of an acute infection is molecular detection of the viral genome via RT-PCR. In particular detection of pre-symptomatic patients would be beneficial for epidemic control [19]. Even at the end of the outbreak molecular detection of ebolavirus is still needed. It has been shown that survivors of the disease still can shed virus over long periods of time through seminal fluids, vaginal secretions, and breast milk, as well as several other body fluids [20, 21, 22] which have the potential of triggering new infections and flare-ups in countries that previously had already been declared Ebola-free [23].
This EQA was designed for laboratories prepared to inactivate and handle suspect samples under BSL3 conditions to perform PCR diagnostic on EVD suspect individuals [6]. This type of diagnostic does not require infectious virus and can be performed on inactivated clinical specimens. While preparation of purified viral genomic RNA, e.g. with the Qiagen viral RNA kit, has the advantage of efficiently eliminating infectivity [24, 25] distribution of purified RNA to participants also has several disadvantages: i) important pre-analytical steps are omitted and purification of nucleic acids prior to RT-PCR analysis cannot be evaluated, ii) stability of shipped RNA is critical [26] and shipment on dry ice would considerably increase costs for the study. In contrast analysis of carefully inactivated viruses also controls the pre-analytical steps prior to RT-PCR. Indeed, for some results indicating a need for improvement of diagnostic performance the information sent by the participating laboratories clearly pointed to problems at the level of sample preparation and not to the PCR analysis itself. While heat-treatment and gamma-irradiation reliably inactivate infectious viruses [27] the genomic RNA is still protected from degradation through RNases within viral particles although the Ebolavirus morphology of the viral particles has been altered through the inactivation process (M. Laue, personal communication). This combined inactivation was used as a safe inactivation method that has been used for different viruses including hemorrhagic fever viruses in other EQA studies [28]. Lyophylization of the inactivated material allows long-term storage and shipment at ambient temperature. Our most recent test of an Ebola-EQA panel set after 14 months of storage at 4°C confirmed unchanged quality of the material even for samples with low copy numbers. The samples from these sets could still be used to spike human blood and semen for validation studies [29].
This EQA study for Ebola PCR-diagnostic was developed during the largest hemorrhagic fever outbreak reported to date [4] and which—with the imminent threat of global spreading—was an enormous challenge for international public health. This had multiple consequences for the development of this particular EQA. The principle of former EQAs organized by ENIVD always had been to keep individual performance of participants confidential. Performance was only revealed to the individual participant after final evaluation of all results. This principle of confidentiality was kept for the majority of participants. However, the laboratories operating in the outbreak countries have been coordinated and supported as well as nominated for participation in the EQA by the Emerging and Dangerous Pathogens Laboratory Network at WHO headquarters in Geneva. In order to allow inclusion into the global evaluation of the public health situation performance of all those laboratories working in the outbreak countries was also revealed to WHO which shared results with the Ministries of Health for the respective outbreak countries. Participants were alerted to this fact by including this information into the documents provided with the EQA panel.
Since one of the main intentions of this study was to evaluate the laboratories deployed to the outbreak countries it is interesting to note that from 28 data sets received from 21 laboratories operating in Guinea or Sierra Leone during the outbreak 23 (82.1%) were optimal while 5 (17.9%) results were lacking sensitivity. Since the initial virus load for most of suspect EVD patients diagnosed in the outbreak countries was quite high [30], sensitivity of detection methods is most likely not the most important issue in an outbreak situation for the majority of samples. However, alternative sampling methods required for collection of blood samples from small children or because of religious or cultural reservations might affect sensitivity of PCR diagnostic [31] and therefore might require the most sensitive performance of the analytical method for certain patients also in an outbreak situation. Compared to the overall performance in this study with 73.6% of optimal results performance of laboratories from the outbreak countries was better. This may not surprise considering that the majority of mobile laboratories had technical expertise with EVD and / or BSL3 practice. Also, these laboratories had built a solid working routine during the outbreak. And although laboratory staff was exchanged on a regular base established and improved protocols were passed on from group to group and were even further refined if the necessity occurred (H. Ellerbrok, personal communication). Also, initial protocols and working routines were established from experts in the field and staff scientists volunteering to serve in these laboratories were also skilled and highly motivated bringing a solid working routine from their home institutions. Therefore, the results from this study revealing such an elevated performance for the deployed laboratories also show that international cooperation is a role model how to handle an emergency in future outbreaks. The panel was useful to provide confidence to the Ministries of Health in the outbreak countries that the international deployed laboratories were technically accurate. Further, the panel allowed increase of testing capacity beyond the limited number of laboratories with prior experience in working with EVD.
|
10.1371/journal.ppat.1006603 | Anthrax edema toxin disrupts distinct steps in Rab11-dependent junctional transport | Various bacterial toxins circumvent host defenses through overproduction of cAMP. In a previous study, we showed that edema factor (EF), an adenylate cyclase from Bacillus anthracis, disrupts endocytic recycling mediated by the small GTPase Rab11. As a result, cargo proteins such as cadherins fail to reach inter-cellular junctions. In the present study, we provide further mechanistic dissection of Rab11 inhibition by EF using a combination of Drosophila and mammalian systems. EF blocks Rab11 trafficking after the GTP-loading step, preventing a constitutively active form of Rab11 from delivering cargo vesicles to the plasma membrane. Both of the primary cAMP effector pathways -PKA and Epac/Rap1- contribute to inhibition of Rab11-mediated trafficking, but act at distinct steps of the delivery process. PKA acts early, preventing Rab11 from associating with its effectors Rip11 and Sec15. In contrast, Epac functions subsequently via the small GTPase Rap1 to block fusion of recycling endosomes with the plasma membrane, and appears to be the primary effector of EF toxicity in this process. Similarly, experiments conducted in mammalian systems reveal that Epac, but not PKA, mediates the activity of EF both in cell culture and in vivo. The small GTPase Arf6, which initiates endocytic retrieval of cell adhesion components, also contributes to junctional homeostasis by counteracting Rab11-dependent delivery of cargo proteins at sites of cell-cell contact. These studies have potentially significant practical implications, since chemical inhibition of either Arf6 or Epac blocks the effect of EF in cell culture and in vivo, opening new potential therapeutic avenues for treating symptoms caused by cAMP-inducing toxins or related barrier-disrupting pathologies.
| Recent anthrax outbreaks in Zambia and northern Russia and biodefense preparedness highlight the need for new therapies to counteract fatal late-stage pathologies in patients infected with Bacillus anthracis. Indeed, two toxins secreted by this pathogen—edema toxin (ET) and lethal toxin (LT)—can cause death in face of effective antibiotic treatment. ET, a potent adenylate cyclase, severely impacts host cells and tissues through an overproduction of the ubiquitous second messenger cAMP. Previously, we identified Rab11 as a key host factor inhibited by ET. Blockade of Rab11-dependent endocytic recycling resulted in the disruption of intercellular junctions, likely contributing to life threatening vascular effusion observed in anthrax patients. Here we present a multi-system analysis of the mechanism by which EF inhibits Rab11 and exocyst-dependent trafficking. Epistasis experiments in Drosophila reveal that over-activation of the cAMP effectors PKA and Epac/Rap1 interferes with Rab11-mediated trafficking at two distinct steps. We further describe conserved roles of Epac and the small GTPase Arf6 in ET-mediated disruption of vesicular trafficking and show how chemical inhibition of either pathway greatly alleviates ET-induced edema. Thus, our study defines Epac and Arf6 as promising drug targets for the treatment of infectious diseases and other pathologies involving cAMP overload or related barrier disruption.
| Bacterial pathogens enhance infectivity by secreting toxins that deregulate immune signaling pathways or disrupt host cellular barriers. One class of toxins produced by diverse bacterial species dramatically increases intracellular concentrations of cAMP. This striking evolutionary convergence suggests that over-production of this second messenger represents a successful strategy to promote growth and dissemination of infectious agents and associated disease symptoms [1]. These toxins include adenylate cyclases (AC), such as edema factor (EF) from Bacillus anthracis (B. a.), CyaA from Bordetella pertussis, and ExoY from Pseudomonas aeruginosa. Other toxins modify host proteins to induce cAMP production by endogenous cellular machineries. For example, cholera toxin (Ctx) from Vibrio cholerae, and the related heat-labile toxin from enterotoxigenic Escherichia coli, both ADP-ribosylate the α subunit of trimeric G proteins to stimulate cAMP synthesis by host AC, while pertussis toxin (Ptx) from Bordetella pertussis ADP-ribosylates and inactivates Gi subunits that normally inhibit endogenous ACs (reviewed in [2]).
B.a., the etiological agent of anthrax, produces two A-subunit toxins, edema factor (EF) and lethal factor (LF), which are secreted together with a shared B-subunit, protective antigen (PA), and then assemble to form edema toxin (ET) and lethal toxin (LT), respectively [3,4]. ET and LT can enter a wide array of mammalian cells expressing either of two related surface receptors, CMG2 or TEM8, where upon the toxins are internalized, leading to the release of the enzymatic A-subunits into the cytoplasm [5]. LF is a zinc metalloprotease that cleaves and inactivates mitogen-activated protein kinase kinases (MAPKKs or MEKs) to block MAPK signaling pathways [6] and, in some hosts, also cleaves NLRP1 to activate the inflammasome [7]. EF is a calmodulin-dependent AC, estimated to be more than a hundred times more potent than its mammalian counterparts in raising intracellular cAMP concentrations [8]. During the early stages of anthrax infection, LT and ET inhibit the innate immune response, reducing cell viability, disrupting chemotaxis and phagocytosis and deregulating cytokine production by macrophages, dendritic cells, and lymphocytes. These combined toxic effects promote bacterial growth and dissemination throughout the host [9,10]. In late fulminant stages of the disease, increasing amounts of ET [11] are released into the bloodstream, and in combination with LT cause edema, bleeding and hemorrhagic lesions (ET), and atypical collapse of the cardiovascular system (LT), often culminating in cardiac arrest and death [12,13].
Molecular pathways altered by the concerted effects of EF and LF were analyzed in transgenic Drosophila models by tissue-specific and conditional expression of the A-toxin subunit using the GAL4/UAS system [14]. Expression in the developing wing revealed that EF caused a phenotype very similar to that of a dominant-negative form of Rab11, a small GTPase of the Rab subfamily essential for endocytic recycling [15,16]. Consistent with EF blocking Rab11-dependent trafficking, two known cargo proteins, Delta (a transmembrane ligand activating the Notch receptor) and the homophylic adhesion protein E-cadherin[17,18] failed to reach their normal destination at apical adherens junctions (AJs). In addition, Rab11 levels were severely reduced in response to EF expression in the wing imaginal disc. This newly recognized activity of EF was also observed in mammalian cells, where ET caused a clear disruption of AJs and Notch signaling in several endothelial cell lines, and was essential for B. a.-induced vascular effusion in vivo [19]. To promote cargo vesicle fusion with the plasma membrane at proper apical sites, Rab11 relies on its effector Sec15, which physically binds to the GTP-bound/active form of Rab11[13,20,21]. Sec15 is a key component of the exocyst, an octameric protein complex that triggers docking and SNARE-mediated fusion of cargo vesicles with the plasma membrane [22]. When over-expressed in various cell types, Sec15 promotes the assembly of large punctate structures[20] that also contain Rab11, Sec15, and other exocyst components. Consistent with previous observations, we found that EF prevented the formation of such Sec15-rich punctae. Interestingly, LF led to a similar inhibition of Sec15 punctae assembly, although via a Rab11-independent mechanism, indicating that Sec15 acts as a convergence point that integrates the effects of both anthrax toxins to block exocyst-mediated trafficking and disrupt integrity of the endothelial barrier [19].
Subsequent studies revealed that cholera toxin also blocks Rab11-mediated trafficking, an activity expected to increase intestinal epithelial permeability, paracellular water loss and diarrhea [23]. These similar cellular effects of ET and Ctx are likely to contribute to the hallmark pathological features and symptoms associated with anthrax and cholera respectively [24].
In the present studies, we delve deeper into the molecular pathways connecting ET-induced cAMP overload to inhibition of Rab11. We apply a combination of approaches involving GTPase isoform-specific transgenes and antibodies, different Drosophila epithelial tissues, human cell lines, and in vivo experiments in mice. Our results indicate that EF disrupts Rab11-dependent processes after the GTP loading step. In flies, both cAMP effectors PKA and Epac disrupt Rab11-mediated junctional transport when artificially activated, but disable early versus late steps of the trafficking process, respectively. However, the Epac/Rap1 pathway seems to serve as the primary mediator of EF-induced toxemia in mammalian systems as well as in the Drosophila wing epithelium. Constitutive activation of Arf6, a small GTPases involved in endocytic retrieval of junctional proteins [25], causes phenotypes nearly identical to that of EF, and similarly alters Rab11 levels and distribution. These findings have potentially important practical implications, since chemical inhibition of Epac (using the selective cAMP analog ESI-09[26]) or Arf6 (using SecinH3[27] or Slit[28]) can reverse the effect of EF in a mouse footpad edema assay and in human cells. Such small molecule interventions open new potential therapeutic avenues for alleviating pathological effects of cAMP toxins and potentially other barrier disruptive agents.
In an effort better understand how EF blocks Rab11-dependent trafficking, we initially examined the behaviors of three YFP-tagged forms of Rab11: wild-type (wt), activated (*), and dominant-negative (DN) [29]. These variants were first expressed in the wing primordium in which inhibition of Rab11 by EF was initially discovered and analyzed [19,23]. The sub-cellular distribution of Rab11wtYFP detected by immuno-fluorescence appears as a grainy stain restricted primarily to the apical pole of epithelial cells (Fig 1A). In addition to this wt pattern, activated Rab11 (Rab11*YFP), a mutant that cannot hydrolyze GTP to GDP, displayed an additional staining component that accumulates at or near apical adherens junctions (AJs) (Fig 1B). This latter staining is in line with the known role of Rab11 in junctional delivery (see S1 Fig for co-localization of Rab11* and Drosophila E-cadherin, D-Ecad). We conclude that active GTP-bound Rab11 is selectively directed to cell-cell contacts at AJs. Consistent with this hypothesis, a dominant-negative Rab11 (Rab11DNYFP), locked in its inactive GDP-bound conformation, did not display a preferential junctional distribution nor apical accumulation (Fig 1C). We then turned our analysis to larval salivary glands, which are comprised of large polyploid secretory cells[30], where the junctional-specific distribution of Rab11*YFP appears more pronounced (Fig 1E). In these cells, over-expressed Rab11wtYFP was distributed throughout the cytoplasm, albeit excluded from densely packed secretory granules, with higher levels detected in the vicinity of intercellular junctions (Fig 1D). Rab11*YFP behaved similarly but, in addition, exhibited a strong junctional staining component (Fig 1E and 1H). In contrast, Rab11DNYFP did not concentrate at junctions, but altered the size and shape of secretory granules (Fig 1G, thin arrows), suggesting that Rab11 normally plays a role in the formation or trafficking of these granules. These findings in the salivary gland confirm our results in wing discs suggesting that the activated GTP-bound form of Rab11 is selectively directed to AJs.
According to the hypothesis that only the activated form of Rab11 traffics to junctions, factors blocking Rab11 upstream of the GTP-loading step should have no effect on Rab11* distribution, whereas inhibitory factors acting downstream of Rab11 should prevent Rab11* from accumulating at AJs. To test this model, we employed two RNAi constructs, one for knocking-down expression of Crag, which is the only known GEF specifically dedicated to activating Rab11 [31], and the other for knocking-down Sec15, an important Rab11 effector required for junctional delivery [18]. Specific inhibitory activities of these RNAi lines were confirmed using epitope-tagged forms of Crag and Sec15 (See S2 Fig). When co-expressed with Rab11*YFP, Sec15-RNAi clearly prevented Rab11*YFP from reaching the AJs (S3 Fig), consistent with Sec15 acting downstream of Rab11 activation. In contrast, Crag-RNAi had no effect on Rab11*YFP distribution, consistent with Crag acting upstream of Rab11 (S3 Fig). Next, we examined whether EF blocks activation (GTP loading) of Rab11 or a subsequent step, by testing the effect of EF on Rab11*YFP localization. Expression of EF blocked all Rab11* junctional accumulation (Fig 1F, compare panels 1H and 1I showing higher magnifications, see S4 Fig for quantifications of junctional Rab11* in response to EF expression). As the constitutively activated mutant Rab11*YFP remains sensitive to EF, we conclude that this toxin acts after the GTP-loading step.
We next examined the behavior of endogenous Rab11 in salivary glands and its response to EF challenge using an antibody that detects all forms of Rab11 (α-Rab11). In wt glands, Rab11 shows a granular distribution with a higher concentration in the vicinity of cell junctions (Fig 1J and 1L), which may represent an enrichment in activated Rab11. In EF-expressing glands, this juxta-junctional staining was clearly reduced: Rab11 dots were detected at similar levels as in wt glands, but very few accumulated around the junctions (Fig 1K and 1M). These findings are consistent with the hypothesis that EF prevents activated Rab11 from reaching the AJs. To test this model further we employed an antibody that specifically detects the activated Rab11 pool (α-Rab11*). Consistent with the observations described above, we found that endogenous activated Rab11 localized predominantly to AJs (Fig 1N and 1P). In EF-expressing glands, the overall levels of activated Rab11* were not obviously altered, however, less activated Rab11 accumulated at the AJs (Fig 1O and 1Q, compare with 1N and 1P). Similarly, in EF-expressing discs, activated Rab11 levels remained comparable to wt levels, while junctional accumulation was severely reduced by EF (Fig 1R and 1S). We conclude that EF does not interfere with Rab11 activation (GTP loading), but instead blocks Rab11 function at a subsequent step(s) to prevent the activated form of Rab11 from trafficking to AJs.
Next, we tested whether the association between Rab11 and its known cargo protein D-Ecad was affected by EF. As expected, co-labeling of Rab11 and D-Ecad in wt salivary glands revealed strong co-localization at cell junctions (Fig 1T and 1U -wide arrow-) and in punctate structures near the junctions (Fig 1U -thin arrow-). In glands expressing EF, however, D-Ecad approached the cell surface (Fig 1V and 1W), but failed to fully localize to AJs, as revealed by gaps in staining between cells (Fig 1W, arrows). Similarly, expression of Rab11DN led to an accumulation of D-Ecad just under the junctions, while many gaps were visible between cells (S5 Fig). These observations suggest that in salivary glands, Rab11 is not required for trafficking D-Ecad to the proximity of junctions, but is critical for the final delivery at the plasma membrane through vesicular fusion. Importantly, in EF-expressing glands, Rab11-DEcad co-localization was abrogated (Fig 1V and 1W, wide arrows). We conclude that EF blocks the association between Rab11 and trafficking vesicles containing cargo proteins such as D-Ecad, leading to a failure in final step of junctional delivery with the consequence of weakened AJs.
cAMP stimulates two main effectors: PKA and Epac, a GEF that activates the small GTPase Rap1 [32] [33]. We activated each branch of the cAMP pathway separately, using either a constitutively active form of PKA (PKA*, consisting of the catalytic domain only [34]), or an activated form of Rap1 (Rap1*, which is locked in its GTP-bound form [35]). We previously reported that both PKA* and Rap1* expressed in the wing primordium caused a reduction in Rab11 levels, blocked apical accumulation of Delta, and prevented the formation of Sec15 structures in the wing primordium [23], suggesting that over-stimulation of each branch of the cAMP pathway can inhibit Rab11. We thought to resolve the respective activities of PKA and Rap1 on Rab11 function further, by co-expressing the activated form of Rab11 (Rab11*YFP) with either PKA* or Rap1*. PKA* profoundly altered Rab11* distribution, both by eliminating accumulation of Rab11* at AJs (Fig 2B and 2E, compare with 2A and 2D) in a similar, albeit stronger, fashion to EF (Fig 1F and 1I), and also by preventing the formation of secretory granules (or dramatically reducing their size). These combined effects of PKA* result in Rab11* being ubiquitously distributed throughout the cytoplasm (Fig 2B and 2E). A similar pattern was observed when staining for total endogenous Rab11, which lost its tendency to concentrate around the junctions in response to PKA* expression (Fig 2H, compare with 2G). Surprisingly, PKA* induced a strong increase in overall Rab11 levels in salivary glands, which is opposite to its effect in wing imaginal discs[19,23]. Consistent with Rab11-dependent trafficking being disrupted by PKA*, adherens junctions appeared weakened in PKA*-expressing glands, with more D-Ecad accumulating in the cytoplasm and around the AJs (Fig 2K) than in the wt glands (Fig 2J). In contrast to PKA*, Rap1* expression in salivary glands did not prevent Rab11*YFP from accumulating near cell boundaries (Fig 2C and 2F). However, instead of the typical single sharp line coinciding with cell junctions observed with Rab11*YFP alone (Fig 2D), co-expression with Rap1* resulted in a double row of Rab11* staining, revealing a narrow gap between adjacent cells (Fig 2F, arrows). This phenotype suggests a failure of the final fusion event between cargo vesicles and the plasma membrane. Consistent with these observations, endogenous total Rab11 staining was also concentrated in a sub-junctional zone in response to Rap1* expression (Fig 2I,arrows), revealing narrow intercellular gaps. These Rap1*-expressing glands also showed an accumulation of small D-Ecad-rich vesicles near inter-cellular boundaries, while normal AJs failed to form (Fig 2L). These results confirm the view that over-activation of each branch of the cAMP pathway can block Rab11-dependent trafficking, but that PKA* does so at an early step when vesicle loading takes place, while Rap1* acts later during the final vesicle delivery process.
In adult Drosophila wings, both PKA* and Rap1* cause phenotypes similar to that of EF (compare Fig 2N and 2O with 2Q) consisting of smaller wings with blisters and thicker veins. The PKA* phenotype, however, is predominantly restricted to the center of the wing (Fig 2N), while Rap1*, like EF, affects the entire wing blade (Fig 2O and 2Q). Consistent with PKA* and Rap1* intersecting a common pathway, we found that co-expression of Rap1* and PKA* led to a drastically enhanced synergistic phenotype (Fig 2P). While these gain-of-function studies reveal that both PKA and Rap1 signaling can interfere with Rab11 trafficking when artificially stimulated, we also tested which cAMP pathway might be required to mediate the effects of EF. We selectively blocked the Epac/Rap1 branch by expressing different EpacRNAi transgenes, which did not produce any notable phenotype on their own (S6 Fig). When combined with EF, however, EpacRNAi significantly reduced the EF phenotype (Fig 2R, compare with 2Q, see S6 Fig for quantifications). In contrast, reducing the levels of PKA-C1, the major PKA catalytic subunit in Drosophila (by two heterozygous loss-of-function PKA-C1 alleles), had little if any effect on the EF phenotype (S6 Fig). We conclude that the Epac/Rap1 pathway is the predominant mediator of EF in the wing epithelium.
In order to direct cargo vesicles to the AJs and promote their fusion with the plasma membrane, Rab11 must interact with several known effectors, including Rab11-FIPs (Rab11 Family-Interacting Proteins [36]) and Sec15, a component of the exocyst complex that is critical for its assembly [20]. Drosophila has a single ortholog of Rab11-FIP (dRip11 [37]), as well as unique representatives of all core exocyst components [18]. We first tested the effect of EF on Rab11 effectors by expressing a full length GFP-tagged dRip11 UAS transgene [37] in the salivary glands. When expressed alone, this fusion protein was strongly concentrated at cell junctions (Fig 3A). Co-expression of EF with dRip11 reduced, but did not eliminate junctional accumulation of dRip11 and also resulted in forked and irregular cell borders (Fig 3B). Because Rip11 is a Rab11-binding protein, we also examined association between these two proteins, which we visualized by expressing Rab11*YFP (detected with a rat anti-GFP antibody) and staining for the endogenous dRip11. This particular double stain revealed frequent co-localization of the two proteins in bright dots in the vicinity of intercellular junctions (Fig 3C, arrows in lower panel). Co-expression of EF with Rab11*YFP severely reduced its co-localization with Rip11 (Fig 3D and 3G), supporting the hypothesis that high levels of cAMP trigger a dissociation of Rab11 and Rip11, or prevent their initial association. Similarly, co-expression of PKA* with Rab11* also largely eliminated co-localization of Rab11* and Rip11 (Fig 3E and 3G). Interestingly, Rap1* also affected this association: Rab11*YFP and Rip11 proteins remained present in adjacent but non-overlapping vesicles (Fig 3F, quantifications in 3G), suggesting that both Rap1* and PKA* have effects on the Rab11*-dRip11 interaction albeit through distinct mechanisms.
In contrast to full-length dRip11, a truncated dominant-negative form of dRip11 (dRip11DN) retaining only the C-terminal Rab11-binding domain[37], did not accumulate at cell-junctions in salivary glands, consistent with its N-terminal cholesterol-binding domain being essential for associated cargo vesicles to traffic to AJs. Instead, dRip11DN was distributed in a reticulated pattern throughout the cytoplasm, although it did show higher juxta-junctional levels (S7 Fig). Small cytoplasmic Rab11 staining punctae strongly co-localized with dRip11DN-GFP (S7 Fig), consistent with dRip11DN retaining its Rab11-binding domain. Interestingly, when co-expressed with EF, this punctate co-localization was not reduced, but rather transformed into rings that encircled secretory granules (S7 Fig). Thus, EF does not abrogate association between Rab11 and dRip11DN. Because deletion of the first 700 aa of Rip11 (a region containing a verified PKA phosphorylation site in humans [38] and several such predicted sites in Drosophila) results in an EF-resistant association between Rab11* and dRip11DN, it is possible that PKA phosphorylation may contribute to this dissociation.
We next examined the relationship between Rip11 and Rab11 in mammalian Madin-Darby canine kidney (MDCK) cells, in which the role of Rab11 in cadherin trafficking has been well established [39]. Co-expression of human Rab11-DsRed and EGFP-Rip11 constructs in these cells revealed strong co-localization throughout the cytoplasm, and a tendency for both proteins to accumulate at cell margins (Fig 3H). Upon treatment with ET, however, we observed a significant reduction in Rab11 and Rip11 co-localization, and a reduction in Rab11 localization at the plasma membrane (Fig 3I, see S8 for Pearson’s coefficient quantifications). Mirroring our observations in Drosophila salivary glands, EGFP-Rip11 accumulation at cell boundaries was reduced by ET-treatment (Fig 3I). Interaction between endogenous activated Rab11 and its effectors was also tested in human brain microvascular cells (HBMECs) transfected with a mammalian Sec15-GFP construct. High-level expression of Sec15-GFP led to formation of punctate fluorescent structures (S9 Fig), the formation of which depends on Rab11 [19]. Consistent with Sec15 associating with the active form of Rab11, we detected, using an anti-Rab11* antibody, a high degree of co-localization between Sec15-GFP fluorescence and Rab11*. In this context of Sec15 over-expression, we also visualized co-localization of Rab11* with endogenous Rip11 (S9 Fig). When these cells were treated with ET, Sec15-GFP punctae were significantly reduced after 6 hours, and the remaining punctae no longer co-localized with Rab11* or Rip11 (S9 Fig). Cumulatively, these experiments suggest that EF-induced dissociation of Rab11* from its effectors Rip11 and Sec15 is a well-conserved process across species.
Junctional homeostasis is also established by a balance of Rab11-mediated delivery of junctional cargo and retrieval of proteins via endocytic processes. Arf6, a small GTPase of the Arf subfamily (ADP-ribosylation factors) is involved in early steps of endocytosis from the plasma membrane, exocytosis, and endosomal recycling, and is predominantly localized to the plasma membrane and endosomes [25]. Arf6 activation contributes to sepsis by promoting vascular leakage through excessive internalization of VE-cadherins [40] and additionally interacts directly with exocyst components [41]. We tested whether Arf6 also exerted a role in mediating the phenotypes caused by cAMP-producing toxins in our system by expressing an activated form of this small GTPase (Arf6*). Strikingly, Arf6* caused a wing phenotype nearly identical to that induced by EF (Fig 4A and 4B) or Rab11DN [19], consisting of small narrowed wings with thickened veins and blisters. In contrast, the wild-type form of Arf6 (Arf6wt) when expressed alone did not cause any detectable phenotype (Fig 4D). However, both activated and wild-type forms of Arf6 strongly enhanced the EF wing phenotype (Fig 4C and 4E). Further analysis revealed that Arf6* reduced the levels and apical restriction of Rab11 in the wing discs (Fig 4H and 4I), diminished the formation of Sec15-rich structures, and reduced total Sec15 levels (Fig 4J and 4K), in a manner similar to what we observed with EF [19]. Arf6* expression also reduced the levels of junctional and total D-Ecad (Fig 4M, compare to 4L), as would be expected from its wing phenotype and effects on Rab11 and Sec15. Given the striking similarities between Arf6* and EF phenotypes, we tested whether Arf6 contributes to mediating the effect of EF in the developing wing, making use of an Arf6-RNAi construct that is highly effective in suppressing Arf6 expression (S2 Fig). Arf6-RNAi did not produce any noticeable phenotype on its own (Fig 4F), but did exert a significant suppression of the EF wing phenotype (Fig 4G, compare with 4A). Arf6-RNAi suppression of the EF phenotype was yet more pronounced at the level of junctional E-Cadherin expression (Fig 4O, compare to 4N). We also tested whether Arf6* altered the distribution of Rab11*YFP in salivary glands. As observed with EF (Fig 1F and 1I), Arf6* reduced the concentration of Rab11*YFP at the AJs (Fig 4Q, compare to 4P), revealing that Arf6* similarly inhibits Rab11 at a step subsequent to GTP loading. In contrast to EF, however, Arf6* induced an intracellular accumulation of Rab11, while also reducing Rab11 levels near the junctions (Fig 4S, compare to 4R), and caused striking accumulations of D-Ecad below the apical plasma membrane (Fig 4U, compare to 4T). In aggregate, these observations suggest that the Arf6 pathway inhibits Rab11 activity, but does so through a mechanism distinct from that of EF.
As described above, activation of PKA*, Rap1*, and Arf6* mimic features of the EF phenotype in Drosophila. We wondered whether the same might be true in mammalian systems and thus examined the relative contributions of each of these pathways in EF-induced toxemia in various experimental models relevant to B.a. infection in mammals. In HBMECs, ET treatment reduced total cadherin levels and weakened AJs as indicated by staining with an anti-pan-cadherin (p-Cad) antibody (Fig 5B, compare with 5A), as shown previously [19]. Western-blot analysis confirmed a drastic decrease in p-Cad and Rab11 levels in response to treatment with ET or dcAMP (S11 Fig). Similar reductions in Rab11 levels in response to EF have been documented histochemically in Drosophila wing imaginal discs [19]. To inhibit the Arf6 pathway, we treated HBMECs with Slit2, a secreted peptide that activates the Robo4 receptor to promote vascular stability via stimulation of the ArfGAP GIT[28]. Cells co-treated with ET and Slit2 appeared resistant to ET, as clearly illustrated by the robust rescue of junctional pan-cadherin accumulation (Fig 5C). These findings suggest that Arf6 contributes to EF-induced inhibition of the Rab11/exocyst complex and weakening of AJs. Next, to determine the relative contribution of each branch of the cAMP pathway, we co-treated ET-intoxicated cells with ESI-09, an inhibitor specific for Epac[26], or with H89, a well-characterized inhibitor of PKA[42]. We found that only ESI-09 could partially restore cadherin expression at AJs (Fig 5D), although junctions did not appear as regular as in untreated cells. In contrast, H89 provided no obvious rescue to the ET-induced junctional phenotype (Fig 5E). We conclude that Epac/Rap1 is the predominant pathway mediating the effects of ET on exocyst-dependent junctional cadherin trafficking in HBMECs.
We next examined the relative contributions of the PKA and EPAC pathways as well as Arf6 in a quantitative in vivo footpad-swelling assay, in which intra-dermal injection of ET results in a robust and quantifiable edema (Fig 5F) [43]. In mice pretreated with SecinH3, a compound that inhibits the Arf6-GEF ARNO thereby lowering Arf6 activity [28], ET-induced swelling was strongly reduced (Fig 5G). Indeed, in animals in which systemic pre-treatment with the drug induced observable symptoms of malaise (presumably indicative of potent systemic pharmacological action), ET-induced edema was virtually abolished. We then examined contributions of the cAMP effector PKA and Epac to ET-induced edema, by comparing the relative abilities of H89 and ESI-09 to block ET-induced footpad swelling (Fig 5F and 5H). Reinforcing the results of our experiments in flies and with HBMECs, we found that while ESI-09 virtually abolished ET-induced edema, H89 had little or no effect (Fig 5H). We conclude that the Epac/Rap1 pathway is the primary mediator of EF-induced edema. In addition, we tested the effect of AG1024 [44], an inhibitor of insulin-like growth factor receptor (IGF-1R) [45]. Because IGF-1R has been shown to indirectly stimulate Rap1 [46], we hypothesized that inhibition of IGF-1R by AG1024 might conversely result in Rap1 inhibition. Indeed, pre-treatment of mice with AG1024 also led to significant reduction of edema, which was particularly strong at early time points (Fig 5I). These findings, together with results described above provide a framework for how effector pathways contribute to cAMP-mediated disruption of Rab11-dependent membrane trafficking (See Fig 6 for summary diagram). In addition to a prominent role of the Epac/Rap1 branch in mediating the effect of ET, our study reveals a previously unappreciated form of negative cross-regulation between the machineries responsible for the delivery versus retrieval of membrane bound cargo. Importantly, small molecule inhibitors such as SecinH3, ESI-09, and AG1024 offer potential for new therapeutic avenues for treating a range of diseases involving compromised barrier integrity of epithelial or endothelial sheets.
In previous studies, we established that two cAMP toxins, EF from Bacillus anthracis [19] and Ctx from Vibrio cholerae [23], block Rab11-mediated endocytic recycling of cargo such as signaling ligands and adhesion proteins (reviewed in [15,16]), ultimately leading to inhibition of Notch signaling and loss of barrier integrity. However, the precise mechanisms by which cAMP overproduction interfered with Rab11-dependent trafficking remained to be explored. Here, we examined how cAMP effector pathways converge on discrete nodes of the trafficking process subsequent to the GTP loading step to efficiently interrupt endocytic recycling.
As is typical of small GTPases, Rab11 cycles between active (GTP-bound) and inactive (GDP-bound) conformations, the former permitting interaction with effector proteins to carry out downstream functions. Two types of regulators, activating GEFs and inactivating GAPs provide control for this essential cycle. In the particular case of Rab11, Crag (the Drosophila homolog of human DENND4A) is the only known Rab11-dedicated GEF [31]. Similarly, only one Rab11-specific GAP has been identified: EVI5 [47–49]. Neither of these regulators contains an identified cAMP-binding domain that could provide a direct link between cAMP and upstream regulation of Rab11. Consistent with this inference, we found that EF acted on Rab11 at a step subsequent to GTP loading. Indeed, transport of vesicles carrying the constitutively activated mutant Rab11*YFP were blocked by EF, while total endogenous levels of Rab11-GTP did not appear to be greatly altered.
Association between Rab11 and its effectors Rip11 and Sec15 was abrogated by EF in several settings, including Drosophila salivary glands and human cells. The Rab11 effector Rip11 is an attractive candidate for mediating some of EF effects, as it contains a verified PKA phosphorylation site located in the central portion of the protein [38]. Indeed, PKA-dependent phosphorylation of Rip11 is required for cAMP-potentiated insulin secretion in pancreatic β-cells [38]. In addition, Ser/Thr phosphorylation is responsible for Rip11 transition from the insoluble to cytosolic fraction in intestinal CACO-2 cells [50]. Although it was not determined whether the latter modification was specifically PKA-dependent, this study proposed a model in which phosphorylation of Rip11 is essential for cycling to a free state following interaction with Rab11 and specific membrane compartments prior to its re-associating with Rab11. Our data show that the association between Rab11 and Rip11 can be disrupted by EF in Drosophila and mammalian endothelial or embryonic kidney cells. It is possible that unrelenting phosphorylation of Rip11 by PKA may cause the premature dissociation of Rab11 and its effectors, potentially leading to a failure to reach the AJs. While this PKA-dependent phosphorylation of Rip11 has been demonstrated in human pancreatic cells, it is not known whether it occurs in Drosophila. As dRip11 contains 19 candidate PKA phosphorylation sites, further investigation will be necessary to determine whether phosphorylation of one or more of these sites occurs and promotes the dissociation between dRip11 and Rab11. Intriguingly, Drosophila Sec15 also harbors several putative PKA phosphorylation sites, although such predicted sites are missing in its human counterpart. Importantly, we found that artificial stimulation of Rap1 also causes a loss in Rab11*/Rip11 co-localization resulting in correlated but separated staining foci of these two proteins, suggesting that the later acting Epac/Rap1 pathway may feedback on this process (see below).
The second branch of the cAMP pathway mediated by the cAMP-regulated GEF Epac and its partner Rap1 [32] contributes significantly to the effect of EF in flies, and surprisingly appears to play the predominant role in the mammalian systems we examined. In flies, activated Rap1 (Rap1*) causes a wing phenotype more similar to that of EF and Rab11DN than that of PKA*. We previously reported that Rap1* reduces the levels of Rab11 and prevents formation of Sec15 punctae [23]. In the present study, we find that blocking expression of Epac significantly reduces the intensity of the EF phenotype. In addition, Rap1* alters the distribution of Rab11* and inhibits Rab11*/Rip11 co-localization. We hypothesize that the final exocyst- and SNARE-dependent fusion event with the apical plasma membrane is subjected to inhibition by exuberant Rap1* activity, leading to accumulation of non-functional Rab11* just beneath the plasma membrane. Consistent with this hypothesis, Rap1 has been implicated by many studies in regulating of both cadherin and integrin-mediated cell-cell adhesion (reviewed in [51] [52] [53]). Further indicating a functional connection between Rap1 signaling and Rab11-dependent trafficking, Rap1 and Rab11 over-expressed in human cells co-localize in a recent study [54]. Additional experiments will be required to elucidate the molecular interactions connecting the activities of these two GTPases. The small GTPase RalA is a possible candidate for mediating the activity of Rap1, through activation of the Rap1 effector Rgl1, a positive regulator (GEF) of RalA. Because lowering the dose of Rgl1, or expressing a dominant-negative form of RalA, can suppress Rap1*-induced phenotypes in Drosophila, it has been proposed that RalA may act downstream of Rap1 [35]. Also, RalA is known to directly bind to exocyst components Sec5 [55] [56] [57] and Exo84 [58] and plays a central role in regulating exocyst-mediated processes in several settings, including the release of Von-Willebrand Factor from endothelial cells, or insulin secretion in pancreatic β-cells (reviewed in [51] and [59]). In addition, a recent study identified Arf6 as a key component acting downstream of RalA, mediating its effect on exocyst-dependent delivery of raft micro-domains to the plasma membrane [60]. Thus, RalA over-activationmay contribute to mediating the effect of cAMP toxins on exocyst inhibition downstream of Rap1, although this hypothesis needs to be tested in future experiments.
We previously showed that EF caused a drastic reduction in total Rab11 levels in wing epithelial cells[19]. Here, we find that this effect is also evident in HBMECs treated with ET, but is dependent on cell context, since inhibition of Rab11 function can be uncoupled from reduction in total Rab11 levels in Drosophila salivary glands. This reduction in Rab11 levels is unlikely to derive from transcriptional inhibition, as infection of HBMECs with B. a Sterne did not result in any change in levels of Rab11 transcripts (Nina Van Sorge, personal communication). Similarly, in Drosophila wings, where EF also triggers great reduction in Rab11 protein levels, mRNA transcript levels again were not greatly affected (Valentino Gantz, personal communication). In HBMECs, where Rab11 levels are reduced by ET treatment, we observed that total levels of cadherins were also severely reduced in ET-treated cells. Although the precise mechanism responsible for the loss of these proteins following ET treatment remains to be explored, it is worth noting that degradation of VE-cadherins has been observed following silencing of Rab11 in human endothelial cells [61], in which Rab11 is important for stabilizing cadherins at the AJs. Thus, it is possible that following EF intoxication, Rab11 and cadherins are routed to the lysosomal pathway and degraded, further impairing endocytic recycling and junctional integrity. Such an attractive hypothesis could explain the catastrophic loss of cadherins observed in ET-treated cells.
Numerous studies have demonstrated the positive role of physiological induction of cAMP in junction establishment and stabilization, through stimulation of both PKA and Epac [38,62]. It may therefore seem counterintuitive that cAMP produced by EF or other toxins may exert an opposing effect and jeopardize junctional integrity. In principle, high versus low concentrations, sustained versus transient production, and perinuclear vs cortical subcellular distribution of toxin-delivered cAMP could elicit such opposite outcomes. In the particular case of Rab11-dependent trafficking, low physiological levels of cAMP may exert their positive effects by promoting the release of Rip11 from Rab11, as necessary to allow the final fusion event between recycling endosomes and the plasma membrane. In contrast, pathologically elevated cAMP concentrations may cause premature dissociation of the Rab11-Rip11 complex and permanently block that cycle. Similarly, uncontrolled stimulation of Rap1 by Epac could also have a negative impact on junctional transport: titration of critical partners, failure to return to complete the necessary GTP/GDP cycle, or negative feedback interference with other important steps, could explain the occurrence of this apparent paradox. Another molecule potentially at play during the response to cAMP is the small GTPase RhoA. RhoA can be phosphorylated by PKA, which inhibits its activation and prevents increased endothelial permeability during inflammation [63], the potential interplay between RhoA and the exocyst downstream of cAMP signaling in EF-intoxicated cells also merits further examination.
The small GTPase Arf6 initiates retrieval of membrane proteins from cell junctions in a wide variety of cells types [25]. Arf6, a member of the ADP-ribosylation factor subfamily, is located at the plasma membrane and some endosomal compartments, and is involved in endocytosis from the plasma membrane, vesicular recycling, and exocytosis [64]. Importantly, Arf6 plays a role during sepsis to mediate acute VEGF-induced vascular permeability [40,65]. Whether linchpin regulators of opposing vesicular trafficking pathways such as Arf6 and Rab11 interact had not yet been extensively explored. In this study, we present evidence that these trafficking systems do in fact engage in cross-inhibitory interactions. Consistent with the published role of Arf6 in promoting VE-cadherin endocytosis [66], the activated form of Arf6 (Arf6*) caused phenotypes similar to those of EF. Our findings suggest that the activity of Arf6 negatively feeds back on vesicular transport to the plasma membrane by inhibiting Rab11 function. Previous studies showed that Arf6 physically interacts with the exocyst component Sec10 [41], defining a possible avenue for our observed effects of Arf6 on Rab11 levels and distribution. Given the negative regulation of Rab11 by Arf6 in flies and its known role in compromising barrier function in the mammalian vasculature during sepsis [28,40], we tested whether inhibitors of this pathway might antagonize the effects of EF. In human endothelial cells, we indeed found that treatment with Slit2, a secreted peptide indirectly blocking Arf6 function, could reverse the effects of EF, restoring junctional integrity. Similarly, pharmacological inhibition of Arf6 by SecinH3, a compound that inhibits the ArfGEF ARNO, potently blocked EF-induced edema in a mouse footpad assay.
An emerging lesson from the current and prior studies is that blocking multiple steps of branching pathways that converge on critical nodes in endocytic recycling may allow pathogens to weaken host protective mechanisms that rely on junctional integrity [24]. For example, LF, the other toxic factor secreted by B.a, blocked exocyst-mediated vesicular docking downstream of Rab11 via inhibition of MAPK signaling. It will be interesting to explore how the various effects of EF and LF are integrated to achieve an efficient inhibition of junctional delivery, and if any compound identified in this study can also block some of the downstream effects of LF. Altogether, our study suggests that a broad range of barrier disruptive diseases ranging from cAMP related toxemia to inflammatory autoimmune diseases that involve positive feedback loops between immune activation and barrier disruption, could potentially be treated with compounds that inhibit Arf6 or Epac/Rap1, or by yet undiscovered compounds that may boost Rab11 activity.
All experiments were performed in strict accordance with guidelines from the National Institute of Health and the Animal Welfare Act, approved by the Animal Care and Use Committee of University of California, San Diego and the National Institute of Allergy and Infectious Diseases, National Institutes of Health (approved protocols s00227m and LPD-8E). Anesthesia and euthanasia were performed using Isoflurane and CO2, respectively. All efforts were made to minimize suffering of animals employed in this study.
UAS-EF construct and line were described previously[19,23]. UAS-Rab11wtYFP/TM3 (#9790), UAS-Rab11*YFP (3rd chr. # 9791), UAS-Rab11DNYFP (#23261), UAS-CragRNAi (2nd chr. #53261), UAS-CragHA (3rd chr. #58463), UAS-Arf6RNAi (3rd chr. #51417), UAS-PkaDN/CyO (#5282), and Pka-C1B10 (#32018) lines were obtained from Bloomington Drosophila Stock Center (BDSC). UAS-EpacRNAiv50272 and UAS-EpacRNAiv50273 were obtained from Vienna Drosophila Resource Center (VDRC). UAS-Rap1*/TM6 and UAS-PKA*/CyO were generated by I. Hariharan (UCB), and D. Kalderon (Columbia University), respectively. UAS-Rip11GFP and UAS-Rip11DNGFP were kindly provided by Don Ready (Purdue University). UAS-Arf6* was generated in the Olson laboratory (UT Southwestern).
Imaginal discs were dissected, fixed and stained using standard procedures. Salivary glands were dissected similarly, fixed for 30 minutes, and left attached to carcasses until ready to mount in SlowFade (LifeTechnologies #S36936), using double sided tape as a spacer to prevent tissue squashing. Antibodies: rabbit anti-GFP antibody (1/500, ThermoFisher #A6455), rat-anti GFP antibody (1/500, SCBT #sc-101536), mouse anti-Rab11 (1/200, BD Biosciences #610657), mouse anti Rab11-GTP (1/100, NewEast Biosciences #26919), D-Ecad (1/500, DSHB #DCAD2). The rabbit anti-Rip11 (1/1000) was a gift from D. Ready (Purdue University) and A. Satoh (Hiroshima University, Japan), and guinea pig anti-Sec15 (1/1000) was kindly provided by Hugo Bellen (Baylor College of Medicine). Images were collected by confocal microscopy on a Leica TCS SP5. All images were acquired using a 40X or 63X objective, and all higher magnifications were obtained using a 4X digital zoom. Co-localization quantifications in Fig 3 used the coloc2 tool in ImageJ.
MDCK cells (ATCC CCL-34) were maintained in DMEM (Corning; Manassas, VA) containing 10% FBS, 1% Penicillin/streptomycin, 2 mM L-glutamine and were incubated in 37°C, 5% CO2 atmosphere. Cells were gently dislodged with 0.05% trypsin (Mediatech) and were electroporated with cDNA expressing DsRed-Rab11A (Addgene) and Rip11-EGFP (Kind gift from Dr. Rytis Prekeris, Univ. of Colorado, Denver) using Neon Transfection system (Life Technologies) according to manufacturer’s protocol. Briefly, cells were rinsed once with PBS and resuspended at a density of 107 cells/ml. cDNA expressing DsRed-Rab11A and Rip11-EGFP were added to the suspension, and cells were electroporated with a 10 μl Neon tip at 1650 V, 20 ms width and 1 pulse. Cells were transferred to 600 μl pre-warmed medium of which 300 μl cell suspension was plated on each well of 8 chamber tissue culture treated glass slide (BD Falcon, Bedford, MA). Cells were treated with 10 μg/ml EF +20 μg/ml PA for 4 h before fixation with 4% para-formaldehyde in PBS for 30 min at 37°C and processed for imaging. Fluorescence images were collected using a Delta Vision RT microscope. Colocalization between Rab11 and Rip11 was determined by measuring the Pearson's correlation coefficient (PCC) using the Velocity 6.3 imaging and analysis software (PerkinElmer). Costes automatic thresholding method [67] was applied for background discrimination.
HBMEC cultures were maintained in DMEM (Corning) containing 10% FBS, 1% Penicillin/streptomycin, 2 mM L-glutamine, and were incubated in 37°C, 5% CO2 atmosphere. Cells were gently dislodged with 0.05% trypsin (Mediatech Inc.) and cultured on glass poly-D-lysine coated chamber slides (BD Falcon #354108). At about 80% confluence, EF and PA (0.2 μg/ml and 0.4μg/ml, respectively) were added to cells. Drug co-treatments included: Slit2 (10μg/ml, R&D systems # 8616), ESI-09 (TOCRIS #4773, 100μM), and H89 (TOCRIS #2910, 10μM). After 24 h (Fig 5), cells were fixed for 10 min at -20°C in 100% Methanol, then washed with 0.1% Triton in PBS. Cells were stained with a mouse anti pan-Cadherin antibody (Abcam, clone CH-19, 1/100). For S6 Fig, transfection of the Sec15-GFP was performed with the FuGENE 9 transfection reagent (Roche) according to manufacturer recommendations. Cells were treated with ET (2 μg/ml EF and 4 μg/ml PA), and fixed after 6hrs of treatment for 30 mins in 4% paraformaldehyde in PBS. Cells were stained with rabbit anti Rip11 (Novusbio #NBP1-81855, 1/500) and mouse anti Rab11-GTP (NewEast Bioscience #26919, 1/100) antibodies overnight at 4°C. Coverslips were washed, and incubated with secondary antibodies before mounting with Prolong Gold with DAPI mounting media (ThermoFisher).
BALB/cJ mice (8–10 weeks old, female; Jackson Laboratories) were intraperitoneally injected with drugs. SecinH3 (TOCRIS #2849), AG1024 (Selleckchem, S1234), ESI-09 (TOCRIS #4773) or H89 (TOCRIS #2910), or with vehicle (70% DMSO in isotonic glucose for SecinH3, 30% DMSO in isotonic glucose for other drugs) 2–3 h prior to injection of ET (0.15 μg/20 μl, right footpad) or PBS (20 μl, left footpad), and in the case of SecinH3, also 2 h post toxin injection. ESI-09, AG1024 and H89 were administered at 10 mg/kg and SecinH3 was administered as 250 μl of 2.5 mM solution. Edema was assessed at 8–10 h, and 18–24 h by dorsal/plantar measurements using digital calipers.
Untreated or ET intoxicated (24h) HBMEC cells were lysed in RIPA buffer (Cell Signaling Technology) supplemented with mammalian protease inhibitor cocktail (Sigma Aldrich). The lysates were clarified by centrifugation at 1000 g for 10 min at 4°C and LDS sample buffer (NuPAGE) was added. Samples were boiled at 95°C, run on a 4–12% SDS polyacrylamide gel (Life Technologies) and transferred onto PVDF membrane (Bio-RAD). After incubation with primary antibodies against Rab11 (#71–5300, Thermo Scientific), Cadherins (CH-19, Abcam #ab6528), and actin (sc69879, Santa Cruz Biotechnology), blots were probed with respective HRP conjugated secondary antibodies and developed using SuperSignal West Pico chemiluminescent substrate (Thermo Scientific).
|
10.1371/journal.pcbi.1000389 | Molecular Mechanics of the α-Actinin Rod Domain: Bending, Torsional, and Extensional Behavior | α-Actinin is an actin crosslinking molecule that can serve as a scaffold and maintain dynamic actin filament networks. As a crosslinker in the stressed cytoskeleton, α-actinin can retain conformation, function, and strength. α-Actinin has an actin binding domain and a calmodulin homology domain separated by a long rod domain. Using molecular dynamics and normal mode analysis, we suggest that the α-actinin rod domain has flexible terminal regions which can twist and extend under mechanical stress, yet has a highly rigid interior region stabilized by aromatic packing within each spectrin repeat, by electrostatic interactions between the spectrin repeats, and by strong salt bridges between its two anti-parallel monomers. By exploring the natural vibrations of the α-actinin rod domain and by conducting bending molecular dynamics simulations we also predict that bending of the rod domain is possible with minimal force. We introduce computational methods for analyzing the torsional strain of molecules using rotating constraints. Molecular dynamics extension of the α-actinin rod is also performed, demonstrating transduction of the unfolding forces across salt bridges to the associated monomer of the α-actinin rod domain.
| The cell interacts with its environment in both biochemical and mechanical ways. In this study we explore one of the ways in which the cell interacts mechanically with its environment. α-Actinin is a cytoskeletal crosslinker: it functions to scaffold the cytoskeletal actin filaments that provide mechanical reinforcement to the cell. In its functional environment α-actinin is exposed to a multitude of mechanical stresses as it attaches itself to a dynamic network of actin filaments. The actin filaments extend, rotate, and bend the α-actinin crosslinkers. In this study we employ molecular dynamics techniques to understand the structural characteristics of α-actinin that underlie its ability to provide a scaffold in such a stressed environment. We analyzed the natural frequencies of α-actinin and simulated force-induced bending, extension, and twisting. Our results suggest that α-actinin has structural flexibility facilitating crosslinking in a dynamic environment and also structural rigidity stabilizing the linkage in the stressed environment. We have discovered novel natural bending movements of the rod domain that enhance its function as a crosslinker. We have also demonstrated the specific structural characteristics of α-actinin that give it the previously suggested property of having partial flexibility. Our results enhance the understanding of structural mechanics in the cytoskeletal molecules.
| Cytoskeletal microfilament networks contribute to the mechanical stability of the cell by dynamically arranging and rearranging actin filaments for reinforcement. The dynamic arrangement of actin filament requires actin filament crosslinking molecules such as α-actinin. α-Actinin is a 200 kDa homodimer with three major structural motifs: the actin binding domain (ABD), the calmodulin homology domain (Cam), and the central rod domain [1]. Each monomer contains all three structural domains but the two monomers are arranged anti-parallel so that the two ABDs are at opposite ends of α-actinin. The arrangement of the two ABDs at opposite ends allows for α-actinin to crosslink parallel actin filaments [2]. Actin filaments in the parallel arrangement are very dynamic; the actin filaments move laterally and horizontally in relationship to each other, and continuously bind and unbind α-actinin crosslinking molecules [3]. Several cellular processes involving actin filament dynamic rearrangement and scaffolding by α-actinin include: focal adhesion formation near membrane bound integrin molecules [4], cytokinesis and cytoplasmic dumping in the final stages of mitosis [5],[6], and z-disk formation and stabilization in muscle cells [7]. In order for α-actinin to maintain its function as an actin filament scaffold in such a dynamic environment, the α-actinin molecule must be partially flexible, meaning it must simultaneously be rigid and stable at some regions to resist external stress and be flexible at other regions to maintain binding in a dynamic environment [8]–[10].
Structure of the α-actinin rod domain underlies the function of α-actinin as a partially flexible actin filament crosslinker. Each central rod domain monomer is 240 Å long and made up of 4 spectrin (R1–R4) repeats connected by helical linkers (see Figure 1) [11],[12]. Other molecules with spectrin repeats include dystophin and utrophin. The α-actinin rod domain differs from the other spectrin family molecules by its shorter length, its more rigid helical linkers, and its dimerization [13]. The spectrin repeats structure of the rod domain contributes several vital characteristics to the α-actinin rod domain: aromatic packing and hydrophobic residues within each repeat stabilize secondary structure [8]; acidic and basic surfaces on R1 and R4 confer strong dimerization interactions [1], Kd of 10 pM between monomers [14]; interaction of hydrophobic residues between R2 and R3 on both monomers and electrostatic interactions produce a coiled-coil homodimer conformation with a 12 degree bend and a 90 degree left handed twist [15]. Together these characteristics account for the rod domain maintaining both structural rigidity and flexibility.
The goal of this investigation is to understand the structural mechanisms of the partial flexibility of the α-actinin rod domain. The coiled-coil nature of the rod domain is an essential component of the rod domain structure. Coiled-coils are the dominant conformation for fibrous proteins [16]. Most coiled-coils have a heptad conformation, with hydrophobic residues every seventh residue [17],[18]. The heptad conformation allows for hydrophobic insertion of one linker region into that of the other monomers by a knobs-into-holes mechanism [19]. The presence of heptad hydrophobic residues is common in coiled-coil structure but neither necessary nor sufficient [17],[18]. Coiled-coils with antiparallel dimers like the α-actinin rod domain are stabilized mainly by electrostatic interactions between the monomers, and within the monomers [20]. In general the knobs-into-holes mechanism of coiled-coil conformation exists only when stabilized by electrostatic interactions [21]. The tendency of electrostatic interactions to play a key role in stabilizing coiled-coil dimers like α-actinin is in contrast to globular proteins, where hydrophobic, VDW, and electrostatic interactions are equally significant to molecular stability [22]. The coiled-coil conformation of the rod domain is a significant structural feature, and the significance of electrostatic interactions to coiled-coil structure stability suggests a significant role of electrostatic interactions in mechanical properties of the α-actinin rod domain.
Several studies have examined the mechanical properties of other molecules with rod-like coiled-coil conformations. These studies on DNA [23]–[26], myosin [27]–[29], and keratin [30],[31] together suggest the coiled-coil rod like structure contributes extensible rigidity and torsional and bending flexibility. The tertiary structure of DNA is referred to as coiled-coil, and more commonly as a double-helix, because it consists of two intertwined α-helices. In contrast, α-actinin and other fibrous proteins are referred to as coiled-coil due to intertwining in their quaternary structures. The difference between the DNA coiled-coil conformation and the protein coiled-coil conformation is significant, but the mechanical properties can still be compared. DNA is the most studied of the coiled-coil conformations and has been described as an elastic rod [32]. Its global mechanical behavior has been described as like a thin isotropic homogeneous rod, but its local mechanical behavior has been described as like an anisotropic heterogeneous rod with bending and torsional flexibility [23]–[26]. Myosin has an S2 region that functions as a lever arm in muscle sarcomeres. Using a single molecule assay in a total internal reflection microscopy experiment [27], it has been shown that the S2 region has significant torsional flexibility underlying its lever arm function. Keratin, the first coiled-coil structure to be discovered [33], is the major molecule in hair fibers, and investigation of its mechanical behavior with molecular dynamics has shown that it has strong stretching rigidity, over 1 nN of force is needed to stretch keratin 90% [30]. Removing the electrostatic interactions underlying the coiled-coil conformation of keratin significantly reduces its rigidity [31]. These studies suggest that the coiled-coil conformation in α-actinin contributes extension rigidity but torsional and bending flexibility.
Studies of α-actinin and other spectrin repeat molecules have similarly demonstrated extension rigidity of the coiled-coil rod domain [34]–[37]. Experimental investigation using atomic force microscopy (AFM) of spectrin unfolding demonstrated that spectrin repeats unfold in a cooperative mechanism [35]. Several molecular dynamics investigations further characterize the extension rigidity of the α-actinin rod domain as resulting from the strength of the helical linker between the spectrin repeats, and electrostatic and hydrogen bonding within each repeat [31]–[34]. There has been no investigation of the bending or torsional flexibility of the α-actinin rod domain or other spectrin repeats, but investigation of α-actinin structure using cryoelectron microscopy has shown that there must be some structural flexibility since α-actinin molecules form stable actin filaments crosslinks in a range of crosslinking angles [38]. Is the flexibility of α-actinin in crosslinking actin filaments due to torsional and bending flexibility of the rod domain? What features of the coiled-coil structure of α-actinin underlie its partial flexibility?
Using molecular dynamics and normal mode analysis this study investigates the mechanical partial flexibility of the α-actinin rod domain. Bending, torsion and extension simulations demonstrate that, as with other coiled-coil molecules, the α-actinin rod domain has bending and torsional flexibility and extensional rigidity. Normal mode analysis shows that the rod-like structure of α-actinin contributes towards its bending and torsional flexibility. Our simulations suggest that aromatic packing interactions determine the trajectory of torsion on the rod domain, and that electrostatic interaction between the monomers contributes extension rigidity to the rod domain.
Structural properties of α-actinin can be inferred from its natural vibrations, therefore, to reveal the naturally rigid and flexible regions of the α-actinin rod domain, we carried out normal mode analysis (NMA). Results from NMA convey properties inherent in the structure of α-actinin regardless of what intermolecular interactions are present [39]. The purpose of our NMA is to determine the contributions of the α-actinin rod-like structure to its mechanical behavior. Later sections will investigate contributions of intermolecular interactions to its mechanical behavior. NMA was carried out on both the monomer conformation of the rod domain containing only four spectrin repeats and on the dimer conformation containing eight total spectrin repeats arranged in an anti-parallel coiled-coil conformation. The NMA results suggest the α-actinin rod domain to have natural bending and torsional flexibility.
NMA from a single monomer of α-actinin suggested that the monomer has significant bending flexibility and some torsional flexibility. The six lowest frequencies of α-actinin are rotational and translational modes. Mode 7 and mode 8, the two lowest frequency vibrational modes, show bending movement mainly at the termini with a single hinge at the central linker (Figure 2 A,B, and E). Of the other lowest frequency vibrational modes, three modes: 9, 10, and 12, exhibit bending modes with three hinge regions, each at the three linker regions between the spectrin repeats (Figure 2 A, C, and F). A three-hinge bending movement refers to bending with three different hinges, resulting in the molecule being divided into four sections, each undergoing movement in different directions. The other lowest frequency normal mode, mode 11, exhibits torsional movement (Figure 2 A, D, G, and Video S1). The torsional motion in mode 11 is localized to the regions near the termini. The α-actinin monomer NMA results suggest bending and torsion movements to be natural movements, movements that are exhibited by the natural vibrational frequencies, and reveal the residues near the termini to be more flexible and the linker residues to be more rigid.
NMA of the rod domain dimer suggested that the bending flexibility is retained in dimerization, while a new natural movement involving torsion and bending is present (Figure 3 C and F). The lowest vibrational frequency for the dimerized α-actinin rod domain (mode 7) is a single-hinge bending mode with hinge action at the central repeat as with the monomer (Figure 3 A, B, and E). The other bending modes, modes 8, 10, and 11, in the dimer also show torsional motion along with the bending motion (Figure 3 A, C, and F). Mode 8 showed the most pronounced bending and torsion, while modes 10 and 11 showed similar characteristics with more subtle movements. Two torsional modes exist for the dimer conformation, modes 9 and 12 (Figure 3 A, D, G, and Video S2). The torsional movement is localized to the termini as with the NMA of the α-actinin monomer. Dimerization maintained natural vibrations of bending and torsion, and still exhibited flexibility near the termini of the α-actinin rod domain. Vibrational movement in several additional normal modes for both the monomer and dimer are listed in Table S1.
The above results (Figures 2 and 3) indicate bending to be the most natural movement for the α-actinin rod domain since the lowest natural frequency vibration of the α-actinin rod in both monomer and dimer conformations was a single-hinge bending motion. To test normal mode analysis findings and to determine the mechanical consequence of having bending be the natural normal mode for α-actinin, we simulated bending using constant force molecular dynamics simulations (Figure 4). The simulations suggested that while the α-actinin rod domain has bending flexibility, dimerization of the α-actinin rod domain enhances the bending rigidity.
Simulation of bending of the monomer suggested bending flexibility. Interestingly, the bent monomer arranged itself into a coiled-coil structure. Total force as low as 24 pN was able to achieve complete bending of the α-actinin rod domain monomer (Figure 4A). The bending simulation showed initial movement by repeats (R1 and R4) followed by movement by the other two repeats and swift collapse of the two ends together (Figure 4B). Once bending of the α-actinin rod domain monomer to zero degrees between the termini was completed, the molecule proceeded to adopt a coiled-coil conformation similar to the dimer conformation, only two spectrin repeats in length not four (Figure 4F). The C-terminus spectrin repeat moves from being in plane with the central linker and the N-terminus to being in plane with the N-terminus but rotated 90 degrees relative to the central linker. The final conformation of the bent monomer is similar to the conformation of half of the full dimer with R1 and R4 in surface contact, and R2 and R3 in surface contact.
The molecular dynamics simulations under bending forces with the α-actinin dimer showed more resistance to bending than the α-actinin rod domain monomer (Figure 4C). The minimum force required to completely bend the dimer was 100 pN compared to the 24 pN to bend the monomer (Figure 4 A and C). Different force levels also showed different rates of bending (as with the simulations of monomer bending). The trajectory of bending shows non-localized movement of the entire molecule during bending (Figure 4D). Repeats 1 and 4 moved before repeats 2 and 3, but the bending force was later transduced and the entire dimer showed movement together. A side view of the bent dimer (Figure 4E) shows that the two bent halves of the molecule fail to collapse on top of each other because of the 90 degree coiled-coil conformation in the dimer conformation; the molecule halves sit next to each other. Further supercoiling is not seen upon bending as in the monomer simulations since R1 and R4 are already in surface contact at both ends of the dimer, and there are no exposed complementary surfaces or unsatisfied salt bridges to interact and coil.
Using rotating constraints, we implemented a molecular dynamics study of torsion induced conformational changes in the α-actinin rod domain. Torque was applied first to a single monomer of the rod domain at both the N-terminus and C-terminus, and also to a single monomer in dimer conformation. For comparison between torsional simulations of the monomer and dimer we define terminus A as the terminus with residue 1 (N-terminus in the monomer) and terminus B as the terminus with residue 475 (C-terminus in the monomer) (Figure 5). Torque studies of the dimer involved rotation of terminus A and terminus B of one monomer in dimer conformation. Direction of rotation is defined as either clockwise or counterclockwise with respect to the viewing angle looking along the axis of rotation at the site of torsion. Clockwise rotation at terminus B refers to clockwise rotation if viewed along the axis of rotation at terminus B, and would therefore be seen as counterclockwise if viewed along the axis of rotation from terminus A. To avoid ambiguity, all rotation directions referred to are viewed from the terminus at which the rotation is taking place.
The trajectory of the α-actinin monomer under torsional simulation was largely determined by interaction between terminal aromatic residues near the torsional stress. Rotation at both terminus A and terminus B required different levels of torque for the clockwise and the counterclockwise rotation (Figure 6 A and B). Clockwise rotation beyond 140 degrees required significantly more torque than counterclockwise rotation at both terminus A and terminus B. The increasing demand for torque to continue rotation in the clockwise direction is explained by the existence of aromatic packing in the α-actinin rod domain. Aromatic packing describes the arrangement of nearby aromatic residues in either an orientation with the aromatic rings stacked on top of each other or with the edge of one aromatic ring stacked against the face of another [40]. These aromatic packing arrangements are highly stable due to a combination of van der Waals (VDW) interactions, hydrophobic interactions, hydrogen bonding, and electrostatic interactions [41]. Aromatic stacking interactions can be further stabilized by π electron sharing between two stacked aromatic rings [42]. Near terminus B there is interaction between aromatic residues W381, Y417, and W453 (Figure 7A). Clockwise rotation at terminus B disrupted the aromatic packing between these residues after 140 degrees of rotation (Figure 7B and Video S3) and further rotation is thereafter increasingly energetically unfavorable. Rotation of terminus B in the counterclockwise direction failed to disrupt the aromatic packing (Figure 7C) even after 140 degrees of rotation. Similarly, at terminus A there are two sets of aromatic packing interactions. One set immediately near to terminus A is made up of Y15, F74, and F89 (Figure 7D). The other set further back from terminus A but still in repeat R1 is made up of amino acid residues W104, Y25, Y55, W32, F52, Y114, and W117. Although rotation in either clockwise (Figure 7E) or counterclockwise (Figure 7F), direction at terminus A fails to completely disrupt the aromatic packing at these two sites, clockwise rotation does reduce the extent of aromatic packing in the first set nearest the terminus. The decrease in favorable aromatic interactions increases the energetic cost of further rotation and thus requires more torque. Rotation in the counterclockwise direction at both terminus A and terminus B showed a slight decrease in torque required for rotation beyond 140 degrees. The reduction can be explained by free rotation about bonds in adjacent residues and the fact that aromatic packing interactions are not disrupted. Rotation at terminus B shows less sensitivity to the direction of rotation as compared to rotation at terminus A, suggesting a more significant impact of aromatic interactions on stability near terminus A than near terminus B.
The trajectory of the α-actinin rod domain dimer under torsional simulation was largely a result of steric interactions between the rod domain monomers. The topology of R1 near terminus A is different from the topology of R4 near terminus B, and the steric interactions with each of their respective complementary repeats after rotation is different (Figure 8). Rotation of terminus A in the clockwise direction lacks steric interactions with the complementary monomer and thus requires less torque for rotation than does the counterclockwise rotation at terminus A (Figure 6 C). Counterclockwise rotation of terminus A results in steric interactions with the complementary monomer after only 60 degrees of rotation, and results in an increase in torque required for rotation after only 60 degrees of rotation. At terminus B both rotation in the clockwise and the counterclockwise direction result in steric interactions (Figure 8A and Video S4). The steric interactions occur when α-actinin is rotated beyond 150 degrees (Figure 6D). There is a greater resistance to clockwise rotation than counterclockwise rotation at terminus B because clockwise rotation involves both steric interactions and disruption of the aromatic packing.
The results from torsional molecular dynamics simulations suggest that the dimerization of the α-actinin rod domain prevents rotation by introducing steric interactions that increase the torque needed to achieve rotation. Simulations of rotation of the monomer needed maximum 330 pN*nm of torque to rotate, while simulation of rotation of the dimer needed maximum 450 pN*nm (Figure S1). Aside from the steric interactions, the rotation of each monomer was resisted by aromatic packing interactions. The aromatic interaction had the additional effect of localizing most of the rotation to terminal regions (Figure 9). Normal mode analysis (Figure 2 and 3) also showed localization of the torsional movements to the terminal residues. Because of the strength of the aromatic interactions, it is likely more favorable for the α-actinin rod domain to further rotate terminal residues than to disrupt the aromatic packing and rotate other regions of the molecule. It is conceivable that the rotation may be localized to the terminal regions because of the speed of the rotation simulation, i.e. due to insufficient time for rotation to propagate to central regions as torque is applied. To test the effects of angular velocity on rotation we rotated the C-terminus of the two central repeats of the rod domain in the clockwise direction using two different rotational velocities: 0.5 degrees/ps and 0.05 degrees/ps (Figure S2). Rotation of the C-terminus of two central repeats in the clockwise direction at a slower rate showed a decrease in localization of conformational change to the terminus. Equilibration after rotation at terminus A did not, however, result in increased propagation of rotation (Figure S3). Both rotation at a slower rotational velocity and equilibration after rotation allow time for the torque to propagate to nearby residues. The aromatic packing interactions in repeat 4 however prevent propagation of rotation when torque is applied at terminus A, even after adequate equilibration after application of the torque (Figure S3). Comparison of the angular velocity used in the molecular dynamics simulations to the angular velocity calculated for torsional normal modes (∼0.005 degrees/ps) suggests rotation during molecular vibration is 100× slower than in our molecular dynamic simulation.
Using constant-force molecular dynamics we explored the mechanical properties of α-actinin. Constant-force molecular dynamics is simulation of the conformational changes in a molecule resulting from application of a constant external force to specific residues. In the α-actinin simulations here, we apply a constant force to the C-terminus (terminus B) of one monomer and hold the other terminus fixed. Specific conformational changes are indicative of structural properties. Our simulations suggest that specific electrostatic interactions within each monomer and between the monomers contribute significantly to the stability of the rod domain dimer under extensional simulation.
The α-actinin rod domain showed significant extensional rigidity. Forces were applied ranging from 100 pN to 200 pN to a monomer alone and to a monomer in dimer conformation, and only the monomer conformation with 150 pN or more of force was completely extended after simulation (Figure 10). Interactions within a single α-actinin rod domain stabilize the individual monomers under extension. Three key interactions dictate the extension trajectory of a single α-actinin monomer (Figure 11): T41-E129, E278-K440, and E159-R321. Studying the extension trajectory of an α-actinin rod domain monomer under a 100 pN force shows that every break in one of these three interactions corresponds to an increase in the extension rate of the monomer. Simulations with larger forces induce conformational changes at too rapid a rate to capture specific interactions. The simulation at 100 pN shows that before the T41-E129 interaction breaks, the extension of the monomer is due to helical regions near the terminus and linker helical regions. Once T41-E129 breaks (Figure 11B), the extension increases from helical unraveling in R1 and R2. The next big increase in extension occurs when the E278-K440 interaction breaks (Figure 11C). The extension continues and is once again extended with the extension and eventual breaking of the third key interaction E159-R321 (Figure 11D). The precise correlation of extension profile to salt bridge breakdown indicates the role these internal interactions play in stabilizing the individual monomers in the α-actinin rod.
Dimerization increases the structural rigidity of α-actinin by introducing 22 specific charged interactions between the two monomers: E71-R937, R56-E929, R56-E922, E115-R925, R57-D897, E900-R925, R53-E908, K134-D745, K138-D743, K138-E741, R186-E741, E266-R661, E266-K613, D270-K613, D270-K609, E433-R528, E447-R531, E425-R450, E454-R531, E447-R450, D422-R532, and R462-E546. These interactions anchor one monomer under forced extension to the other monomer, increasing rigidity and reducing the length of extension. The extension profile of the α-actinin rod dimer (Figure 12 and Video S5) demonstrates that the salt bridges between the two monomers do not break as a result of the extensional force being applied. Furthermore, the extensional force is being applied to one monomer but in all four repeats (Figure 12 C–F) the extension is not limited to only the stressed monomer; the salt bridges cause extension in the associated monomers as well. The associated monomer extends as well because it is more energetically favorable to extend its helices than to break the salt bridges holding the two monomers together.
The actin filament cytoskeletal network is highly dynamic and stressed. Actin filaments are in continuous movement and rearrangements, and are also continuously exposed to external stresses. As an actin filament crosslinker, it is therefore functionally necessary for α-actinin to be both rigid and be able to withstand the external stress exposure, and to flexibly and dynamically scaffold the actin filaments. The structural mechanisms determining the necessary partial flexibility of α-actinin are as yet not fully understood. By applying methods of computational simulation and normal mode analysis we suggest several possible explanations for the molecular basis of the partial flexibility. Our results suggest that the α-actinin rod domain is flexible and dynamic near its termini while its central helical linkers are rigid, and its dimerized surface is highly stable. Our results also suggest that α-actinin has bending flexibility. Our molecular dynamics simulations suggest aromatic packing interactions play a role in resisting torsion, while electrostatic interactions play a role in resisting extension.
Our NMA study suggested that the most natural vibrational mode of the α-actinin rod domain, both for monomer and dimer conformations, is bending (Figure 2 and 3). This result was then further tested by our molecular dynamics simulations, which suggested bending forces as low as 24 pN could bend the α-actinin rod monomer. The NMA results suggested that extensional natural frequencies are at physiologically irrelevant higher frequencies and not in the low frequency normal modes (up to mode 18 at 21 Hz). Again, our molecular dynamics simulations suggested that extension of the α-actinin dimer requires extensional forces of over 200 pN. Physiologically, α-actinin needs to stabilize the actin filaments even if they are moving closer together, and the bending flexibility suggested by our study could contribute to α-actinin's ability to do so. These results are consistent with previous studies on DNA, myosin S2 region, and keratin, which also suggested that the coiled-coil structure allows for bending and torsional flexibility but has extensional rigidity [21]–[28].
Previous studies have shown through electron microscopy that α-actinin crosslinks actin filaments at numerous angles and lengths [35]. Their work suggests that although the actin filaments are continuously in lateral and horizontal movement, the α-actinin crosslinker is able to maintain its crosslink of the actin filaments. Our molecular dynamics simulations suggest that the α-actinin rod domain has bending flexibility, and that the rod domain is likely to bend while the crosslinker is under compressive stress. The exact direction of any physiological compressive stress is likely different from the bending forces used in our simulations, but the bending flexibility suggested by our simulations is likely the same.
The NMA results as well as the torsional molecular dynamics simulations suggested that the torsion of the molecule during its natural vibrations occurred mostly near its termini (Figure 9). Flexibility near the termini suggests that the flexible region of the α-actinin structure is its neck region, which connects the regions near the termini of the rod domain to the ABD and Cam domains (Figure 1). Flexibility in the neck region has been suggested several times, most recently by Sjoblom [43], and our results further suggest that the flexibility in that region is structurally facilitated by flexibility at the rod domain near its termini.
An interesting result from the bending simulations was the super coiling that occurred when bending a single α-actinin rod domain monomer (see Figure 4F). The monomer was bent in half and immediately arranged itself into a coiled-coil conformation. The bent monomer had several similarities to the dimer molecule; it was twisted 90 degrees, it had R1 interacting with R4, and R2 interacting with R3. The phenomenon suggests that the coiled-coil conformation is a result of the complementary surfaces on R1 and R4 and on R2 and R3.
The torsional studies pointed out the role the aromatic residues in each spectrin repeat play in stabilizing the α-actinin molecule. The aromatic packing play a critical role in resisting torsion and more importantly in restricting rotation to regions near the termini while the molecule is subjected torsion. The importance of the aromatic residues as suggested by our results is consistent with a previous study which suggested that aromatic residues in the spectrin repeats are conserved [8], and with previous studies on DNA which showed that aromatic packing plays a critical role in DNA structural characteristics [21]–[24].
Our extension results provide interesting insight into the mechanisms of dimer stabilization. The α-actinin rod domain was stabilized by both charged interaction within each monomer and by electrostatic interactions at the surface between the two monomers. Furthermore, the charged surface interactions relay extension on only one monomer into extension in the other monomer. Several previous studies are consistent with this phenomenon and our results strengthen these past studies. Law et al. [44] studied the interaction of α-spectrin and ß-spectrin in dimerization while external forces were applied. They suggested that the regions with stronger intermonomer interactions had unfolding at both monomers, whereas regions with weaker intermonomer interactions had only unfolding in one monomer. Ortiz et al. [45] studied the effects of hydration on linker stability. They found that once the molecular regions blocking off the linker from hydration were broken, the linker lost its conformation. We find in our study that the interaction preventing hydration of the linker region in α-actinin is the electrostatic interactions between charged residues in front of the linker regions (see Figure 11). Other studies [13],[46] have suggested that the mechanisms for stabilization of the α-actinin dimer are electrostatic interactions on the surface and acidic residues on the surface. Our results suggest that the charged interactions on the surface between the monomers play a vital role in providing rigidity.
The normal mode analysis results were determined using the WEBnm@ [44] method which shows global conformational movements of the rod domain. In contrast, the molecular dynamics simulations of rotation, bending, and extension, all showed local conformational changes determined by interactions at the atomic level. The results from both sets of investigations are similar and therefore suggest that the global conformational changes seen in the normal modes are correlated to local atomic interactions seen in the molecular dynamics simulations.
Results from our study suggest the α-actinin rod domain is stabilized by electrostatic interactions between the monomers, aromatic interactions within each monomer, and steric interactions between the monomers. The results also suggest that the rod domain has bending flexibility, torsional flexibility near its termini, and extensional rigidity. These mechanical properties could play a role in facilitating α-actinin's role as an actin filament crosslinker. Experimental investigations can further test some of the results presented here. Our results suggest that mutation of aromatic residues would reduce resistance to torsion, and mutation of electrostatic residues would reduce extensional rigidity. Mutagenesis of such residues, therefore, could reduce the mechanical stability of α-actinin and decrease its effectiveness as an actin filament crosslinker.
Natural frequencies of both the monomer of the α-actinin rod and its dimer form were determined using WEBnm@ [47]. WEBnm@ is an online normal mode analysis tool developed using the computational methods of Hinsen [48]. Normal modes are calculated by determining the eigenvectors of the matrix of second derivatives of energy with respect to displacement of the Cα atoms of each residue. Because NMA represents movements resulting from overall structure, the use of Cα force fields are sufficient for NMA calculations [49]. WEBnm@ uses MMTK [50] toolkits internally and provides a web graphical user interface for implementing MMTK scripts. MMTK is an open source library of molecular modeling scripts developed in the PYTHON programming language. The 6 lowest frequency modes of α-actinin represent the translational and rotational normal modes and are ignored since these modes involve whole protein movement and demonstrate no conformational dynamics [51]. The next 6 lowest frequency normal modes (modes 7–12) exhibit the lowest energy natural vibrational frequencies and conformational movement. The natural frequencies are a property of the molecular structure and can be used to differentiate rigid molecular regions from flexible molecular regions. The natural frequencies also reveal movements (such as bending, torsion, or extension) that are natural to the molecular structure of α-actinin.
The crystal structure for the α-actinin rod domain (PDB ID = 1HCI) [15] was retrieved from the Protein Data Bank (PDB) and used for NMA by WEBnm@ after 1000 steps of the Adopted Basis Newton-Raphson (ABNR) method for minimization in CHARMm [51]. The NMA simulation with WEBnm@ uses only Cα force fields and calculates normal modes without a solvent environment. Resulting vibrational frequencies were visualized using the molecular visualization software VMD [52]. Vector fields of the vibrational movements were produced by WEBnm@ and used in VMD along with a new ribbon representation to illustrate the modal movements. WEBnm@ also calculated the individual residue vibrational movements using the RMSD methods of Schulz [53]. NMA analysis was carried out separately for both the monomer (spectrin repeats 1–4) and dimer (8 spectrin repeats).
Molecular dynamics simulations were carried out using the commercially available software CHARMm [54]. α-actinin structure (PDB ID = 1HCI) was retrieved from PDB and minimized using the Adopted Basis Newton-Raphson (ABNR) method for 1000 steps. For equilibration and simulation the implicit water solvation model ACE [55] was used. After minimization, each simulation was heated to 310 K and equilibrated using the VERLET loop function in CHARMm. Throughout the equilibration and bending simulation the temperature was controlled using the Hoover temperature control method [56]. Charmm22 force field definitions [57] were used along with the SHAKE method [58] for applying harmonic constraints on the bond lengths to hydrogen atoms. All molecular dynamics simulations were carried out with 1 fs timesteps and simulations were run for 500,000 timesteps (500 ps). Results were visualized using VMD [52].
Bending simulations were carried out on both dimer conformations (8 spectrin repeats) of α-actinin and monomer conformations (4 spectrin repeats). For the simulations on the monomer conformations, constraints were placed on the α-carbon of I320, S162, and L240, the three central residues suggested to be hinges in the NMA studies. Bending forces were applied to S1 and N85 at terminus A and to G398 and D475 at terminus B. Vector direction of the bending forces was taken directly from the bending normal modes. Total bending force applied to the α-actinin monomer ranged from 8 pN to 200 pN.
Bending simulations on the dimer conformation were carried out similarly. Constraints were placed on the α-carbon of S161, L240, I320, S637, L715, and S798. Bending forces were applied to S1, N85, H873, and D950 at terminus A and G398, D475, S476, and N560 on terminus B. Total bending force applied to the α-actinin dimer ranged from 24 pN to 200 pN.
Α-actinin crystal structure with PDB ID = 1HCI [15] was solvated in a water box with a 10 Å radius of solvation. For the dimer conformation with 950 residues about 100,000 water molecules were added. For the monomer conformation with 475 residues about 65,000 water molecules were added. Molecular dynamics were carried out with NAMD [59]. Each structure was minimized for 1000 steps and equilibrated for 400,000 steps (400 ps). Equilibration and simulations were run with 1 fs timesteps. Results were visualized using VMD [52].
Torsion was implemented using the rotating constraints function of NAMD using the CHARMM22 force field definition [57]. The target atom to be rotated was attached to a reference atom with a spring of known stiffness K. The reference atom was rotated at a known angular velocity about the axis of rotation and the resulting conformational changes in α-actinin were determined by the molecular dynamics. If ν is the unit vector of the axis of rotation, M the rotation matrix, P the coordinates of the pivot point, R the coordinates of the reference atom, and R0 the original coordinates of the reference atom, then the location of the reference atom can be determined by R = M(R0−P )+P. If X is the location of the target atom, then the normal is defined as N = (P+((X−P) .ν)ν)−X. The force applied to the target atom by the spring attaching it to the rotating reference atom can be calculated as F = 2 K(R−X). The torque is then calculated as torque = F×N.
The torsion simulations were carried out with rotating constraints applied to terminus B or terminus A, and rotated the target atoms in either the clockwise or the counterclockwise direction. For simulations on both the α-actinin monomer and dimer rotated at terminus B, fixation constraints were placed on residues 1, 2, 3, 4, 5, 84, 85, and 86 at terminus A, and rotating constraints were placed on residues 396, 397, 398, 399, 400, 401, 469, 470, 471, 472, 473, 474, and 475 at terminus B (Figure 5). For simulations on the α-actinin monomer and dimer rotated at terminus A, fixation constraints were placed on residues 396, 397, 398, 399, 400, 401, 469, 470, 471, 472, 473, 474, and 475 at terminus B, and rotating constraints were placed on residues 1, 2, 3, 4, 5, 84, 85, and 86 at terminus A (Figure 5). Residue 1 and 475 are also part of the axis of rotation, and application of rotation constraints to these residues has the purpose of inducing rotation to these residues once they move off the axis of rotation due to conformational changes induced by rotation of nearby residues. For simulations with rotation targeted at terminus B, the pivot for the axis of rotation was at the α-carbon of residue 1 and the axis was defined in the direction from residue 1 to residue 475. For simulations with rotation targeted at terminus A, the pivot for the axis of rotation was placed at the α-carbon of residue 475 and the axis of rotation defined in the direction from residue 475 to residue 1. The specific axis of rotation used in the molecular dynamics simulations was chosen to be consistent with rotation as seen in normal mode analysis. Normal mode analysis of the monomer (Figure 2) suggested that the residues near the termini rotate around the axis running through the center of the molecule. The axis from residue 1 to residue 475 is set up to also run through the center of the molecule. For all simulations a spring stiffness of K = 10 Kcal/molÅ2 was used. The reference atoms were rotated at a constant angular velocity of 0.005 degrees per timestep (0.5 degrees/ps) in the positive direction for counterclockwise simulations, and in the negative direction for clockwise simulations. Simulations were run for 400 ps to achieve near 180-degree rotation. To illustrate the effects of rotation speed, the C-terminus of the two central repeats of α-actinin were rotated clockwise at two different rotational speeds: 0.005 degrees per timestep (0.5 degrees/ps) and 0.0005 degrees per timestep (0.05 degrees/ps) (Figure S2). Quantitative results of applied torque and angle of rotation were calculated using MATLAB (2007a, The Mathworks, Natick, MA). The average torque applied on each residue during simulation is reported below. Torque values have been interpolated using a fourth-order polynomial fit to reveal trends from the oscillations in applied torque values.
Extension studies were carried out first in implicit solvent using CHARMm [54] and later verified using explicit simulations in NAMD [59]. α-Actinin atomic coordinates were taken from the crystal structure PDB ID = 1HCI [15]. For implicit solvent simulation of the α-actinin molecule the ACE [55] implicit solvent method was used. For the explicit solvent simulations a water box with a solvation radius of 15 angstroms was elongated to 48.8 nm×17.2 nm×8.5 nm to ensure solvation of an extended α-actinin. The resulting water box had over 230,000 water molecules. The cut-off length of non-bonded interactions was set to 12 Å. Explicit solvent simulation was run for 100 ps with 100 pN of force, and a timestep of 1 fs. Forced extension simulations were carried out on a monomer conformation with only the four spectrin repeats, a dimer conformation with 8 total spectrin repeats, and single spectrin repeat (R1) created based on the repeat definitions given by Gilmore et al. [60]. The results were visualized using VMD [52].
For accurate and efficient simulation the hydrogen atom bond length was constrained using SHAKE [58]. The SHAKE method fixes bond lengths between large atoms and hydrogen atoms preventing unnecessary calculation of irrelevant interactions. One fs timesteps were used in both implicit and explicit simulations, and simulations were run at 310 K. The α-actinin molecule was minimized using ABNR for 1000 steps and equilibrated using the VERLET loop function of CHARMm. Charmm 22 force field [57] definitions were used. Constant forces ranging from 100–200 pN were applied to the terminus B α-carbon at residue 475, and fixation constraints were applied to the α-carbon of residue 1 at terminus A in both the monomer and dimer simulations. Forces were applied in the vector direction from terminus A to terminus B. The implicit solvation simulations were run for 500 ps, and the explicit solvation simulations were run for 100 ps for verification.
|
10.1371/journal.pgen.1003563 | Multi-organ Abnormalities and mTORC1 Activation in Zebrafish Model of Multiple Acyl-CoA Dehydrogenase Deficiency | Multiple Acyl-CoA Dehydrogenase Deficiency (MADD) is a severe mitochondrial disorder featuring multi-organ dysfunction. Mutations in either the ETFA, ETFB, and ETFDH genes can cause MADD but very little is known about disease specific mechanisms due to a paucity of animal models. We report a novel zebrafish mutant dark xavier (dxavu463) that has an inactivating mutation in the etfa gene. dxavu463 recapitulates numerous pathological and biochemical features seen in patients with MADD including brain, liver, and kidney disease. Similar to children with MADD, homozygote mutant dxavu463 zebrafish have a spectrum of phenotypes ranging from moderate to severe. Interestingly, excessive maternal feeding significantly exacerbated the phenotype. Homozygous mutant dxavu463 zebrafish have swollen and hyperplastic neural progenitor cells, hepatocytes and kidney tubule cells as well as elevations in triacylglycerol, cerebroside sulfate and cholesterol levels. Their mitochondria were also greatly enlarged, lacked normal cristae, and were dysfunctional. We also found increased signaling of the mechanistic target of rapamycin complex 1 (mTORC1) with enlarged cell size and proliferation. Treatment with rapamycin partially reversed these abnormalities. Our results indicate that etfa gene function is remarkably conserved in zebrafish as compared to humans with highly similar pathological, biochemical abnormalities to those reported in children with MADD. Altered mTORC1 signaling and maternal nutritional status may play critical roles in MADD disease progression and suggest novel treatment approaches that may ameliorate disease severity.
| Mitochondrial disorders have multiple genetic causes and are usually associated with severe, multi-organ disease. We report a novel zebrafish model of mitochondrial disease by inactivating the etfa gene. Loss of this gene in humans causes multiple acyl-Co dehydrogenase deficiency (MADD) that manifests with brain, liver, heart, and kidney disease. While presentations are variable, many children with MADD have a severe form of the disease that rapidly leads to death. We report that etfa gene function is highly conserved in zebrafish as compared to humans. In addition we uncovered potential disease mechanisms that were previously unknown. These include the impact of maternal nutrition on disease severity in their offspring as well as the role mTOR kinase signaling. Inhibition of this kinase with the drug rapamycin partially reversed some of the symptoms suggesting this may be a new approach to treat mitochondrial disorders.
| Multiple acyl-CoA dehydrogenase deficiency (MADD), also known as glutaric aciduria type II (GA-II, OMIM #231680), is a rare autosomal recessive inherited metabolic disorder first described in 1976 [1]. The precise incidence and prevalence are unknown but are likely underreported given the variability in clinical presentation. MADD is caused by mutations in electron transfer flavoprotein genes A (ETFA), B (ETFB) or the ETF dehydrogenase (ETFDH) [2]. The ETFA and ETFB gene products, ETFα and ETFβ respectively, form an ETF heterodimer located in the mitochondria matrix [3]. This complex receives electrons from at least nine distinct dehydrogenases that are involved in fatty acid β-oxidation, amino acid and choline metabolism [4], [5], [6], [7]. Patients with MADD are classified by disease severity with type 1 having severe neonatal-onset with congenital anomalies, rapid deterioration and death [8]. Type 2 patients with MADD do not have congenital anomalies but still have a severe course with death usually during the few years of life [9]. Finally, type 3 patients have later onset and an overall milder course. However they still have hypoglycemia, metabolic acidosis, cardiomyopathy, hepatomegaly, kidney defects and neurological manifestations such as encephalopathy and leukodystrophy [10], [11].
Current treatments are mainly aimed at relieving symptoms though anecdotal reports of improvement after administration of riboflavin or Coenzyme Q have been reported [11]. While all types of MADD can be caused by ETFA, ETFB or ETFDH mutations, it is not understood why there is such variability in disease severity. Several reports indicate a marked buildup of fatty acids, amino acid or toxic compounds in multiple organs in patients with MADD. However, comprehensive cellular and molecular analyses have not been possible as there are no animal models available that recapitulates the spectrum of abnormalities seen in patients with MADD. The first animal model of MADD was created by inactivating the zebrafish etfdh gene [12]. This mutant zebrafish was named xavier (xav) with conserved metabolic abnormalities also observed in MADD patients including increased levels of acylcarnitines and glutaric acid. However xav mutant zebrafish did not recapitulate morphological defects observed in MADD patients. This may be due to early lethality seen in this model prior to later stages of organogenesis.
Using forward genetic screening for mutants with abnormal livers, we identified a mutant zebrafish called dark xavier (dxavu463, termed hereafter as dxa) due to its phenotype of a dark fatty liver and hepatomegaly. Dxa mutant zebrafish have a nonsense mutation in the etfa gene resulting in widespread abnormalities broadly similar to those observed in MADD patients. We found large increases of acylcarnitines and glutaric acid in dxa mutants associated with multiple abnormalities of various organs including brain, liver, kidneys and heart. Marked accumulation of neutral lipid drops including cerebroside sulfate and free cholesterol in multiple organs was also observed. Analyses by mass spectrometry [13] found a large increase in triacylglycerides in dxa mutants but also a significant decrease of phosphatidylserine species which was also observed in human tissue derived from a patient with MADD [14]. The multiple defects seen in dxa mutant zebrafish closely recapitulate many core abnormalities observed in human patients with MADD. Interestingly, dxavu463 mutant developed hyperplasia with increased cell size in multiple organs including brain, liver and kidney suggesting activation of mTORC1 signaling. Excessive maternal feeding also exacerbated the phenotype in dxa mutants. We confirmed that mTORC1 signaling is highly elevated in dxa with increased phosphorylation of S6 and 4E-BP1. Treatment of dxa zebrafish with rapamycin alleviated a subset of signaling and cellular proliferation abnormalities suggesting that targeting mTORC1 signaling could be a rational therapeutic approach for patients with MADD.
We identified dxa mutants during a forward genetic screen using ENU mutagenized zebrafish. Homozygous dxa mutants had a large and dark colored liver at 7 days post fertilization (dpf) (Figure 1A). However, when more closely examined, dxa mutants had a broad spectrum of defects during development and post developmental stages (Figure 1A). About 20% of mutants had severe congenital defects (type I) that included a small head and cardiac edema, these larvae died by 5–6 dpf. Approximately 18% of mutants were type II with moderate defects including an abnormal head and dark liver, intestine and brain. These died by 7–8 dpf. The remainder of the mutants (approximately 62%) classified as type III had mild defects that were morphologically close to wild type zebrafish except for a darker appearing liver, intestine and brain (n = 218). Type III mutants lived for 10 dpf in the unfed state whereas control siblings live for 10–12 dpf. Overall, type I, II and III mutants accounted for approximately 25% of the total zebrafish in each cross suggesting the dxa phenotype was due to a defect in an autosomal recessive gene. This was later confirmed (see below) as a mutation in the etfa gene known to be involved in mitochondrial function. Given the potential for metabolic influences on mitochondrial disease, we studied whether maternal overfeeding prior to egg laying could influence the phenotype. One week of extra feeding caused a dramatic shift in severity with 57% of dxa mutant zebrafish now classified as type I, 32% type II and 11% type III (n = 151) (Figure 1B). This result suggests that the maternal nutritional state dramatically affects the severity of dxa zebrafish and may also explain similar phenotypic variability reported in patients with MADD.
To identify the mutant gene in dxa zebrafish, we performed conventional linkage mapping and were able to map the likely gene to approximately 0.18 cM from the galk2 gene, located on zebrafish chromosome 25 (1/547 recombination, data not shown). Whole genome sequencing of dxavu463 and control zebrafish ultimately identified a mutation in the etfa gene approximately 360 kb from galk2. This G to T mutation introduces a premature stop codon (G290X) in etfa (Figure 1C). Zebrafish etfa has 80% homology to the human ETFA gene suggesting a highly conserved function (sequence alignment not shown). Whole mount in situ hybridization of etfa mRNA shows maternal and ubiquitous expression during early development with subsequent high expression levels maintained in the midbrain and blood vessels at 30 hours post fertilization (hpf) as well as liver and pectoral fins at 2 dpf (Figure S1A). Immunofluorescent staining with an anti-Etfa antibody also revealed high expression of Etfa protein in neural progenitor cells located adjacent to the ventricles of the brain of wild type larvae (Figure 1D, head, inset). We also saw strong expression in neuromast hair cells as well as kidney, liver and skeletal muscle of the pectoral fin of wild type at 9 dpf (Figure 1D, trunk, inset, n = 9/9). However, negligible Etfa protein was detected in dxa zebrafish (Figure 1D, bottom panel, n = 9/9. There is some residual signal in the dxa zebrafish that we interpret as non-specific binding of the secondary antibody to the outer pial membranes of the brain and outer eye (Figure 1D). Immunoblot analyses also detected a very minimal amount of Etfa protein in the dxa mutant (Figure S1C). This also supports a loss of function mutation due to non-sense mediated decay of etfa mRNA given the location of the premature stop codon in exon 10 and the in situ expression data (Figure 1D lower panel, Figure S1B. 10/10). This expression pattern of etfa further supports an important role in high energy demanding cell types such as neural progenitors within the brain, hepatocytes and kidney tubule cells.
While genetic testing of affected patients is ideal, MADD can be strongly suspected in symptomatic children who exhibit increased serum acylcarnitines and glutaric aciduria [7]. Using tandem mass spectroscopy to determine acylcarnitine levels, we found significantly higher level of multiple long-, medium- and short-chain acyl-CoA species and isovalerylcarnitine in the dxa mutant larvae compared to control siblings (Figure S2A). This suggests that dysregulation of mitochondrial β-oxidation is highly similar in dxa to that observed in patients with MADD. Further analysis of organic acids using gas chromatography-mass spectrometry found approximately 6.5 µg of glutaric acid per dxa larvae (Figure S2C), but no detectable amount seen in control siblings (Figure S2B). This pattern is highly reminiscent of that seen in patients with MADD, also known as glutaric aciduria Type II (GA-II).
Hepatic steatosis is a central sign in MADD, likely resulting from defective fatty acid β-oxidation that may be exacerbated during episodes of hypoglycemia. Dxa mutants exhibit progressive accumulation of dark colored granules in multiple organs including brain, liver and intestine after 6 dpf (see Figure 1A). Oil Red O (ORO) staining in whole mounts of type II dxa mutant larvae and coronal sections showed massive accumulations of neutral lipid in the brain, liver and intestine as well as blood vessels at 8 dpf (Figure 2A, B). Interestingly, toluidine blue staining of thick sections used for Transmission Electron Microscopy (TEM) revealed many heterogeneous sized vacuoles in the liver with brown colored drops in the cytosol of hepatocytes (Figure 2C). As glycosphingolipid can include glucose or galactose and sulfate groups (cerebroside sulfate) that can be stained with toluidine blue, the brown drops within hepatocytes may contain cerebroside sulfate. Intensive Periodic Acid Schiff (PAS) staining was also seen in dxa liver (Figure 2C, middle) suggesting that these lipid drops are comprised of cerebroside sulfate as liver glycogen is normally undetectable in unfed 8 dpf zebrafish. Additional support that the drops do not contain glycogen is supported by TEM analyses where we did not observe any glycogen containing granules at 6 or 8 dpf (data not shown). We also found high levels of free cholesterol in the cytosol of dxa hepatocytes using filipin staining (Figure 2C, right). However, we did not see lipid and free cholesterol accumulation in dxa liver at 6 dpf although the mutants already exhibit hepatomegaly and enlarged hepatocytes (Figure S3A–D)
Enlarged cell size was a consistent phenotype at later stages (Figure S3E, F). Dxa hepatocytes were approximately three times larger those seen in control siblings (Figure S3G). These results suggest that intrinsic abnormalities of hepatocytes led to both lipid and cholesterol accumulation in dxa mutant zebrafish. We then analyzed cellular ultrastructure using TEM to investigate possible organelle defects. The internal mitochondrial cristae density was markedly decreased in type II dxa hepatocytes at 6 dpf although total mitochondrial size was not changed (Figure 2D). Strikingly, we found extremely large mitochondria with minimal cristae in type II mutants just 2 days later at 8 dpf (Figure 2E). From these TEM results, we conclude that the “vacuoles” we saw in the mutant liver are actually grossly swollen mitochondria (Figure 2C, E). This suggests that Etfa is required for mitochondrial maintenance as well as energy metabolism. We assessed mitochondrial function in dxa mutants by measuring oxygen consumption over time. We found significantly decreased oxygen flux in dxa mutant zebrafish compared to sibling controls (Figure S4). This strongly supports an impairment of mitochondria function in etfa mutant cells.
It has been reported that many patients with severe neonatal onset MADD have polycystic kidney disease though Bohn et al. and Harkin et al. reported that these kidneys were pathologically distinct from typical polycystic kidney disease [15], [16]. We did note high Etfa expression within pronephric tubules of wild type kidneys (Figure 1D, trunk). Histological analysis of toluidine blue stained sections of dxa zebrafish kidney showed clear abnormalities possibly resulting from hypertrophy of the pronephric tubular epithelium. We also found a large number of prominent vacuoles in both type II and III dxa kidney epithelium compared to wild type (Figure 3A). TEM analysis of type II dxa mutants showed similar to hepatocytes, they are very large mitochondria with minimal cristae (Figure 3C, 2E). This finding implies that the “cystic kidney” pathology ascribed to patients with MADD may be due to massively swollen mitochondria in kidney tubule cells. We further found lipid and free cholesterol accumulation in the cytosol of dxa mutant kidney cells (Figure 3B, D). However, we did not see significant increases of lipid and free cholesterol in dxa mutants that have only mild defects at earlier stages, though hypertrophic kidney tubule cells with swollen mitochondria are already present.
ORO staining also revealed extensive lipid drops in the brain (Figure 2A). We then analyzed brain sections to more precisely determine the location and cell types that contain lipid. We found large lipid accumulation in the ventricular zone (VZ) of the brain where the neural progenitor cells are found (Figure 4A). Within the same region of dxa mutant brain, we also found cerebroside sulfate containing lipid drops at 9 dpf (Figure 4B). Interestingly, type II mutants with more severe defects have VZ cells with very large nuclei compared to other neurons, but these large cells had pale intracellular staining (Figure 4B). TEM revealed that these cells do not have a discernible subcellular structure other than swollen nucleus and mitochondria (Figure 4C, D). These findings are suggestive of ongoing necrosis although we could not identify ruptured cell membranes. Of note we did not see enlarged nuclei at 6 dpf but swollen mitochondria were still observed in the type II mutant brain (data not shown). These results suggest progressive and rapid brain damage after 6 dpf in dxa zebrafish.
To analyze whether those abnormal cells are neural progenitors, we performed immunostaining for Sox2 in both type II and III dxa larvae at 8 dpf. We found increased numbers of Sox2 positive cells in both type II and III mutants (n = 9/9). Statistical analysis of type II mutants showed a 75% increase of Sox2 positive cells in the dorsal part of the VZ at 8 dpf (Figure S5). Finally, we found that brain lipid-binding protein (BLBP) positive neural progenitor cells were increased and distorted in their morphology though had decreased processes within the white matter of dxa mutants (Figure 4H, asterisk and yellow magnified inset, n = 6/6). Altogether, these results suggest that neural progenitor cells in the dxa mutant are hyperplastic and hyperproliferative. They may also be unable to properly generate neurons given the abnormal appearing grey matter observed in dxa mutant brain (Figure 4).
Given the overt increases in lipids seen by Oil Red O staining, we performed lipid profiling with mass spectrophotometry (MS) to identify differences of lipid molecular species between control and dxa mutant larvae at 8 dpf. We found moderately decreased monoacylglycerol (MAG) and diacylglycerol (DAG) in the dxa mutant though only the MAG decrease was statistically significant. However a large increase of triacylglycerol (TAG) was observed in the dxavu463 mutants (Figure S6).
Using MS analyses of glycerophospholipids, we also found significant decreases in phosphatidylserine (PS) species (Figure S6). By contrast, the two most abundant phospholipid species phosphatidylcholine (PC) and phosphatidylethanolamine (PE) did not show any statistically significant differences between control and dxa mutant zebrafish. These results provide a rationale for future lipid modifying therapies in patients with MADD and will help focus future experiments on lipid abnormalities seen in etfa mutant zebrafish.
High Etfa expression was found within neuromast cells that are zebrafish sensory organs (Figure 1D). Interestingly, type I and II dxa mutants had a decreased response to touch stimulation (data not shown) that was correlated with the severity of defects and increased age. Neuromast cells had short or absent kinocilia in type II dxa mutant and using ORO staining we found the dark granules seen in dxa neuromasts are comprised of lipids (Figure 5A, seen in 10/10 mutant larvae examined). Given these widespread lipid abnormalities in sensory structures, we also looked for alterations in sensory nerves tracts. Using acetylated tubulin, we noted decreased staining with a disorganized appearing, “kinked” axonal track in type II dxa mutants (Figure 5B, n = 8/9). We then examined expression of myelin basic protein (MBP), the most abundant myelin associated protein in the brain and spinal cord. We found decreased MBP staining in dxa type II mutants (Figure 5C, n = 6/6) but did not see significant changes in type III mutants (data not shown, n = 5/6). Using TEM to examine myelination in the spinal cord, we again found abnormally increased size and morphology of mitochondria in the Mauthner axon track (sensory pathway mediating escape responses) as well as decreased myelination (Figure 5D). These hypomyelination findings are reminiscent of the leukodystrophy reported in patients with MADD [17].
The markedly enlarged cells in dxa mutant brain, liver and kidney suggests that mTORC1 could be involved as signaling through this kinase is a key controller of cell size [18]. We previously showed activation of this pathway in zebrafish causes increased cell size that can be reversed with rapamycin, a potent mTORC1 inhibitor [19]. We then evaluated the phosphorylation status of the mTORC1 downstream effectors S6 ribosomal protein and 4E-BP1 by immunofluorescence. We found markedly elevated phospho-S6 and phospho-4E-BP1 in dxa mutants especially in neural progenitor cells (n = 4/6 (phospho-S6), 6/6 (phospho-4E-BP1)) and pial cells (n = 6/6 (phospho-S6), 6/6 (phospho-4E-BP1)) on the midbrain and central canal of the hindbrain (Figure 6A, B). Increased level of phospho-4E-BP1 was seen more broadly than phospho-S6 in dxa mutants notably within midline cells and the central canal of the hindbrain as well as the intestine (Figure 6B). Immunoblots of dxa mutants also revealed increased mTORC1 signaling compared to control larvae at 6 dpf (Figure S7A–C). We also found increased phospho-S6 and phospho-4E-BP1 in the kidney and liver at 8 dpf (Figure 6B). Even at earlier time points before the pathology was overt, we still found increased mTORC1 signaling in the liver (Figure S7D). Given these findings we hypothesized that mTORC1 inhibition could potentially rescue dxa mutants. However, rapamycin treatment from 3–9 dpf was not able to suppress the dxa mutant phenotype (data not shown). To address whether activated mTORC1 observed in dxa mutants is rapamycin sensitive, we treated with rapamycin daily from 5 dpf to 8 dpf at a concentration of 300 nM. Rapamycin freely crosses the blood brain barrier in zebrafish and typically inhibits mTORC1 downstream completely [19]. In contrast to results we have obtained with other zebrafish models of human disease, we found that levels of phospho-S6 and phospho-4E-BP1 were not fully suppressed in the brains of treated larvae (Figure 6C). In the liver rapamycin did in fact suppress phospho-S6 levels but phospho-4E-BP1 actually appeared to be increased (Figure 6D) [20]. This intriguing finding suggests a novel regulation of mTORC1 downstream effectors in dxa mutant zebrafish.
We previously found that tsc2 mutant zebrafish with prominent mTORC1 activation had both increased cell size and increased proliferation [19]. To assess proliferation in dxa mutants, we analyzed the proportion of kidney and liver cells expressing proliferating cell nuclear antigen (PCNA) at 8 dpf. Very few PCNA positive cells were detected in control siblings but a highly increased proportion of cells express PCNA in type III dxa mutant at 8 dpf (Figure 7A). We quantified these differences in the liver and found 1/246 PCNA positive cells in control larvae versus 20/245 in type III dxa mutant zebrafish, this difference was statistically significant, p<0.006 (Figure 7B). We then treated with rapamycin from 5 to 8 dpf to verify if suppression of mTORC1 signaling could rescue aspects of the dxa phenotype. While mutants treated with rapamycin still developed a fatty liver, there was a clear decrease in cellular proliferation (Figure 7B). This result suggests that mTORC1 may be considered as a potential therapeutic target for some of the pathological features seen in patients with MADD.
MADD is a complex genetic disease with multi-organ involvement and widespread biochemical abnormalities. These features likely reflect the impairment of multiple acyl-CoA dehydrogenases with each enzyme normally handling different substrates. Additional MADD complexity may be due to distinct mutations in ETFA, ETFB or ETFDH. While patients with ETFDH mutations predominate in the literature, this is possibly due to a bias of genetic testing for relatively milder forms of MADD that are compatible with longer survival. Patients with ETFA mutations in contrast may have a more severe course and rapidly succumb to this disease prior to an accurate clinical, biochemical and genetic assessment. MADD is now screened in newborns in many countries and the true prevalence of all genotypes should eventually emerge from prospective analysis of confirmed positive cases.
Comprehensive analysis of MADD features and pathological mechanisms including genotype/phenotype relationships has been severely hampered by the lack of genetic animal models that recapitulates key features of MADD. In this study we analyzed a novel zebrafish model with a loss of function mutation of the etfa gene. Remarkably, dxa zebrafish recapitulates many key MADD features including biochemical abnormalities, a phenotypic spectrum from severe (type I and II) to moderate (type III) and multi-organ defects of the brain, liver and kidney. The shared phenotype of hepatic steatosis and dysmorphic kidneys seen in patients with MADD and dxa mutant zebrafish are likely due to defects of fatty acid β-oxidation as well as disruptions of amino acid and choline metabolism. C4 (butyryl) and C5 (isovaleryl) acylcarnitines and glutaric acid were highly elevated, this confirms the remarkable conservation of zebrafish and human mitochondrial function. However we did note species-specific differences. For example, patients with MADD have elevations of C14:1 but we also observed increases of fully saturated C16 and C18 in zebrafish. This suggests that the substrate availability for very long chain acyl-CoA dehydrogenase (VLCAD) in the zebrafish diet differs from the human fatty acid pool and that zebrafish primarily oxidize saturated fatty acids.
In contrast to MADD/GAII, GA Type I is caused by mutation in glutaryl-CoA dehydrogenase (GCDH), however this is one of dehydrogenases coupled to the ETF complex. Gcdh knockout mice did not have any obvious brain defects, but on a high lysine diet, these mice had neuronal loss, defective myelination and swollen mitochondria [21], [22]. Though abnormalities of neural progenitor cells were not reported, their results suggest that accumulation of glutaric acid may be sufficient to cause defective myelination and mitochondrial abnormalities although the clinical differences between GA-I and GA-II support distinct pathological mechanisms for each disease. Strikingly, we found the severity may be caused by the nutritional state of the parents as extra feedings prior to egg fertilization produced a much higher proportion of type I and II mutants in each cross. Ongoing studies in our laboratory will investigate whether this mechanism is due to additional metabolic “stress” or from alterations of key maternal proteins, lipids or mRNA in the yolk. However, our findings indicate that a better understanding of nutrition and overfeeding may positively impact fetuses with MADD and could reduce severe congenital anomalies. The low frequency of this disease and the lack of prenatal diagnosis in families without a previously diagnosed proband makes this scenario unlikely but given the trend towards precise genetic diagnoses for all aspects of medicine, maternal diet potentially exacerbating the MADD phenotype may be a crucial finding.
mTORC1 signaling is a key mediator of cell size control and differentiation. Using other zebrafish models of human disease, rapamycin treatment reversed abnormalities of cell size and mTORC1 signaling in the brain [19]. In marked contrast, brain abnormalities and other aspects of mTORC1 signaling in dxa zebrafish appeared to be rapamycin resistant. Phospho-S6 was entirely inhibited in the liver of dxa zebrafish but levels of phospho-4E-BP1 were actually elevated by rapamycin. It was previously shown that rapamycin inhibits phosphorylation of S6 and 4E-BP1 differentially [20]. This group reported that S6K and S6 phosphorylation were readily abolished throughout the duration of rapamycin treatment but phosphorylation of 4E-BP1 can recover despite initial inhibition and repeated application of rapamycin. We do not understand the mechanism leading to rapamycin resistant mTORC1 signaling in the brain and liver but speculate it may be related to increased amino acids in dxa zebrafish that could activate Rag proteins [23]. Increased leucine for example is sufficient to cause Rag GTPase dependent translocation of mTORC1 to lysosomes [24]. Isovaleric Co-A dehydrogenase requires the ETF complex and loss of function mutations in the ISOVALERYL-CoA DEHYDROGENASE (IVD) gene are known to cause accumulation of isovaleric acid, a metabolite of leucine [25], [26]. Leucine accumulation in dxa may then be activating mTORC1. We found markedly increased leucine levels in dxa larvae supporting this potential mechanism (Figure S8). We also found markedly increased p62/sequestosome 1 in dxa mutant zebrafish (data not shown) that was recently shown to be essential to activate mTORC1 [27]. These findings suggest that restricting intake of leucine and other branched amino acids may be important in MADD to suppress symptoms due to mTORC1 activation. However, increased aerobic glycolysis was observed in etfdh mutant zebrafish, this may compensate for a failure of mitochondrial beta oxidation [12]. Increased glycolysis may provide key intermediates for cell proliferation [28] and elevated mTORC1 signaling could further increase glycolysis by modulating transcription of genes required for this metabolic process [29]. This may represent a compensatory mechanisms and inhibition of mTORC1 with rapamycin could exacerbate the MADD phenotype or precipitate a metabolic crisis. We have seen no evidence for this in our animal model but caution against the use of mTORC1 inhibitors in patients with MADD outside of well regulated clinical trials.
The myelination defects in etfa mutant zebrafish are notable given the severe neurologic deficits including encephalopathy that is usually seen in patients with MADD. Lysosomal disorders such as metachromatic leukodystrophy (MLD) [30], Krabbe disease [31] and Gaucher disease [32] all have accumulation of cerebroside sulfate that appears to cause myelination defects in nerve system as well as hepatomegaly. We also see accumulation of cerebroside sulfate in radial glia and hepatocytes in dxa mutant larvae. It is possible that inhibition of autophagy by mTORC1 activation might contribute to symptoms in MADD. The markedly increased p62 levels in dxa mutants supports such a mechanism.
In conclusion, we report the first animal model of MADD due to mutations of the etfa gene. Dxa mutant zebrafish larvae have an array of biochemical and pathological features that strongly indicates this is a relevant model for MADD. Dxa zebrafish can be effectively employed to generate and test further hypotheses about disease specific mechanisms. In addition, dxa mutant zebrafish will be invaluable for future in vivo chemical screens to identify therapeutic compounds that may ameliorate disease aspects of MADD and potentially other mitochondrial disorders.
Zebrafish strains used in this study included AB* and dxavu463. Embryos were obtained from natural matings and raised at 28.5°C in egg water (0.3 g of sea salts/L). For overfeeding experiments, we gave an extra meal of TetraMin Tropical Flakes daily for one week prior to fertilization of eggs. The normal diet is twice a day meal of brine shrimp and Tropical Flakes Monday through Friday and once day on Saturday and Sunday of each week. Short term extra feeding does not cause any obvious phenotypes. 5 pairs of heterozygous siblings were used for this experiment. We fed normally one week and each pair of each was mated. Then we gave the same zebrafish extra food for the subsequent week and mated again. This cycle was repeated three times.
Antisense digoxigenin-labeled RNA probe for etfa was produced using a DIG-RNA labeling kit (Ambion). Embryos were fixed in 4% paraformaldehyde overnight, and dehydrated in 100% methanol at −20°C. Whole mount hybridization was performed using standard protocols [33]. BCIP/NBT (Vector laboratories) mixture was used as a chromogenic substrate. In situ images were acquired using a Zeiss Axioscope and Nikon Coolpix 4500 digital camera.
To avoid staining variation, 3 control and 3 dxa mutant larvae were processed together in the same slide glass. Slides were processed in a Sequenza Slide Rack. Embryos were fixed in 4% paraformaldehyde from overnight to two days at 4°C. Fixed embryos were embedded in 1.2% agarose/5% sucrose and saturated in 30% sucrose at 4°C for 1 to 2 days. Tissue blocks were frozen in 2-methyl butane. 10 µm sections were collected on microscope slides using a Leica cryostat. Sections were kept in −80°C before use. Sections were rehydrated in 1× PBS and blocked in 5% sheep serum in PBS for 2 hours, they were then incubated with primary antibodies to Etfa (Genetex, #GTX124324, dilution 1∶300), Sox2 (abcam, #97959, dilution 1∶500), PCNA (Sigma, #P8825, dilution 1∶3000), BLBP (abcam, #ab32423, dilution 1∶500), phospho-S6 ribosomal protein (Cell Signaling #2215 Ser235/236, dilution 1∶300), and phospho-4E-BP1 (Cell Signaling #2855 Thr37/46, dilution 1∶300) overnight at 4°C, washed 10 minutes×3 times with 1× PBS and then incubated for 2 hours with Alexa Fluor conjugated goat anti-rabbit secondary antibodies. Sections were then washed with 1× PBS for 30 minutes and mounted in Vectashield with DAPI (Vector laboratories). Antigen retrieval for PCNA staining was performed for 30 minutes of boiling in 10 mM sodium citrate before blocking. Images were acquired using Zeiss Axiovert 200M microscope with Zeiss AxioCam MRm and Hamamathu digital camera. Digital images were processed using Adobe Photoshop CS5 and Adobe illustrator CS5. All images received only minor modifications with control and mutant sections always processed in parallel.
Fixed samples were rinsed with PBS-DT (1× PBS, 0.5% Triton X-100, 2% DMSO) and both control and mutant were incubated with blocking solution (PBS-DT, 5% goat serum) for 2 hours at room temperature in a single tube. The antibody against acetylated-tubulin (Sigma, #T7451, dilution 1∶500) was used overnight at 4°C. Larvae were rinsed with PBS-DT 3 times (10 minutes each). Secondary was a goat Cy3-anti-mouse for overnight at 4°C. Specimens were rinsed with 700 µL of PBS-Tween for 10 minutes and repeated 5 times. Zebrafish were fixed in 4% PFA and mounted in glycerol before being imaged.
For whole mount staining at larvae stage, larvae were fixed in 4% PFA overnight. Control and dxa mutant larvae were rinsed three times (5 minutes each) with 1× PBS/0.5% Tween-20 (PBS-Tween). After removing PBS-Tween, larvae were stained with mixture of 300 µL of 0.5% ORO in 100% isopropyl alcohol and 200 µL of distilled water for 15 minutes. Larvae were then rinsed with 1× PBS-Tween for three times. Larvae were rinsed twice in 60% isopropyl alcohol for 5 minutes each. They were briefly rinsed in PBS-Tween and fixed in 4% PFA for 10 minutes. Larvae were mounted in glycerol prior to imaging. For high resolution ORO staining on transversely sectioned larvae, 10 µm sections were dried at room temperature for 5 minutes. 150 µL of working ORO solution was added to slides and stained for 30 seconds. They were then washed with tap water and mounted using Vectashield with DAPI.
For free cholesterol staining on transversely sectioned larvae, slides were soaked with 1× PBS for 5 minutes, then Filipin complex diluted 1∶500 (Sigma, F-976) was added directly to slides and stained for 1 minute in the dark. Slides were washed with PBS and mounted with 75% glycerol. Images were taken using the DAPI channel of a fluorescent microscope.
Frozen sections were used for PAS staining. The PAS stain was conducted in the Translational Pathology Core laboratory at Vanderbilt University using a DAKO Artisan Link Staining System.
Glycerophospholipids from zebrafish larvae were extracted using a modified Bligh and Dyer procedure [34]. Forty of 8 dpf larvae of each genotype, either mutant or sibling control were homogenized in 800 µl of ice-cold 0.1 N HCl∶CH3OH(1∶1) using a tight-fit glass homogenizer (Kimble/Kontes Glass Co, Vineland, NJ) for about 1 minute on ice. The suspension was then transferred to cold 1.5 mL Eppendorf tubes and vortexed with 400 µl of cold CHCl3 for 1 min. Centrifugation (5 minutes at 4°C, 18,000× g) to separate the two phases. Lower organic layer was collected, an odd carbon internal standard was added and solvent evaporated. The resulting lipid film was dissolved in 100 µl of isopropanol∶hexane∶100 mM NH4COOH(aqueous) 58∶40∶2 (mobile phase A). Quantification of glycerophospholipids was achieved by the use of an LC-MS technique employing synthetic odd-carbon diacyl and lysophospholipid standards. Typically, 200 ng of each odd-carbon standard was added per sample. Glycerophospholipids were analyzed on an Applied Biosystems/MDS SCIEX 4000 Q TRAP hybrid triple quadrupole/linear ion trap mass spectrometer (Applied Biosystems) and a Shimadzu high pressure liquid chromatography system with a Phenomenex Luna Silica column (5-µm particle size) using a gradient elution as previously described [35], [36]. Individual species were identified based on their chromatographic and mass spectral characteristics. This analysis allows identification of the two fatty acid moieties but does not determine their position on the glycerol backbone (sn-1 versus sn-2). Neutral lipids from zebrafish (forty of 8 dpf larvae/sample) were extracted by homogenization in the presence of internal standards (500 ng 14∶0 monoacylglycerol and 24∶0 diacylglycerol and 1 µg 42∶0 triacylglycerol) in 2 ml 1× PBS and extracting with 2 mL ethyl acetate∶trimethylpentane (25∶75). A dried lipid film was dissolved in 1 mL hexan∶sopropanol (4∶1) and passed through a bed of Silica gel 60 Å to remove remaining polar phospholipids. Solvent from the collected fractions was evaporated and lipid film was redissolved in 90 µl 9∶1 CH3OH∶CHCl3, containing 10 µl of 100 mM CH3COONa for MS analysis essentially as previously described [36], [37]. Samples were analyzed in triplicates and p-values determined using Student's t-test.
Forty 9 dpf control and dxa mutant larvae were lysed using pellet pestles (Sigma, #Z359947) and passed through a 25 gauge syringe in 150 µL of PBS. For acylcarnitine analysis, the total lysate was placed into a 96 well plate containing stable isotope labeled internal standards (Cambridge Isotope Laboratories, Andover, MA) and acylcarnitine analysis performed according to the published methods [38]. Briefly, the lysate was dried under nitrogen, reconstituted with fifty µL of acetonitrile and one µL was injected into a Xevo-TQS tandem mass spectrometer (Waters Corp. Waltham, MA). Acylcarnitines were quantified against an isotope–labeled internal standard of the nearest chain-length using the parent ions of the carnitine-specific fragment of m/z 85. For organic acid analysis, the total lysate was made up to a final volume of 2.5 mL using deionized water, acidified to pH 2.0 and the acid fraction extracted three times into equal volumes of ethyl acetate. The pooled organic phases were dried down under a stream of nitrogen at room temperature and trimethylsilyl derivatives analyzed by gas chromatography-mass spectrometry using an Agilent 7890A gas chromatograph fitted with a 5975C Mass Selective Detector (Agilent Technologies, Santa Clara, CA) using a method initially developed for urine and vitreous humor analysis [39]. The acylcarnitine and organic acid assays are validated and in routine clinical use and were also previously used for analyses of etfdh mutant zebrafish [12].
In brief, samples were fixed in 2.5% gluteraldehyde for 1 hour then transferred to 4°C overnight. Samples were washed 3 times in 0.1 M cacodylate buffer then incubated for 1 hour in 1% osmium tetraoxide and washed with cacodylate buffer. Samples were dehydrated through a graded series of ethanol, then incubated in ethanol and propylene oxide (PO). Samples were infiltrated with 25% Epon 812 resin and 75% PO for 35 minutes, then 50% Epon 812 resin and 50% PO for 1 hour then exchanged with new 50% Epon 812 resin and 50% PO and incubated overnight. Samples were exchanged with 75%: 25% (resin: PO), then pure epoxy resin for 3–4 hours, then overnight. Finally, the resin was exchanged with epoxy resin for 3 hours, embedded in epoxy resin and polymerized at 60°C for 48 hours.
Sectioning and Imaging: 500 nm to 1 µm thick sections were collected using a Leica Ultracut microtome. Thick sections were stained with 1% toluidine blue and. 70–80 nm ultra-thin sections were cut from this block and collected on 300-mesh copper grids and stained with 2% uranyl acetate (aqueous) for 16 minutes and then with lead citrate for 12 minutes. Samples were imaged on the Philips/FEI Tecnai T12 electron microscope at various magnifications.
Average basal oxygen flux was quantified by high resolution respirometry using the Oroboros O2k Oxygraph (Oroboros Instruments, Innsbruck, Austria). Ten larvae (8 dpf) per chamber were maintained in Instant Ocean at 28°C and initially equilibrated to room air. Measurements of oxygen concentration were recorded every 2 seconds with no stirring. When the measured O2 concentration stabilized, chambers were stirred at 100 rpm for as short a time as possible to permit recording of a new stable O2 concentration, which was reflective of the true O2 concentration in solution. Stirrers were then turned off. This process was repeated until no further decrement in O2 concentration was measured and the fish were no longer motile. Average oxygen flux was calculated from the change in O2 concentration over time from the beginning of the experiment to the end. Data are from 4 measurements made with 10 larvae of each genotype.
Forty control siblings and homozygote dxa mutant larvae from 3 to 5 clutches were homogenized in 100–750 µL of 0.1 M TCA, containing 10 mM sodium acetate, 100 µM EDTA, 5 ng/ml isoproterenol as an internal standard and 10.5% methanol at pH 3.8. Samples were centrifuged at 10,000× g for 20 minutes. Supernatant was removed and stored at −80 degrees. Samples of the supernatant were then analyzed for biogenic monoamines and/or amino acids. Leucine was quantified with a Waters AccQ-Tag system with a Waters 474 Scanning Fluorescence Detector. Ten µL samples of the supernatant are diluted with 70 µL of borate buffer to which 20 µL aliquots of 6-Aminoquinol-N-Hydroxysuccinimidyl Carbamate and 10 µL 250 pmol/µL α-aminobutyric acid (as internal standard) are added to form fluorescent derivatives. 10 µL of sample was then injected into the HPLC system, and separation of the amino acids accomplished by means of a Waters amino acid column and supplied buffers using a specific gradient profile.
Error bars in Figure S3 and S6 represent standard error of the mean (SEM), error bars in Figure 7 and S4 represent standard deviations (SD). Student's t-test was used to determine statistical significance.
All animal experiments were done with the approval of the Vanderbilt University IACUC.
|
10.1371/journal.pmed.1002907 | Vitamin D status and risk of incident tuberculosis disease: A nested case-control study, systematic review, and individual-participant data meta-analysis | Few studies have evaluated the association between preexisting vitamin D deficiency and incident tuberculosis (TB). We assessed the impact of baseline vitamins D levels on TB disease risk.
We assessed the association between baseline vitamin D and incident TB in a prospective cohort of 6,751 HIV-negative household contacts of TB patients enrolled between September 1, 2009, and August 29, 2012, in Lima, Peru. We screened for TB disease at 2, 6, and 12 months after enrollment. We defined cases as household contacts who developed TB disease at least 15 days after enrollment of the index patient. For each case, we randomly selected four controls from among contacts who did not develop TB disease, matching on gender and year of age. We also conducted a one-stage individual-participant data (IPD) meta-analysis searching PubMed and Embase to identify prospective studies of vitamin D and TB disease until June 8, 2019. We included studies that assessed vitamin D before TB diagnosis. In the primary analysis, we defined vitamin D deficiency as 25–(OH)D < 50 nmol/L, insufficiency as 50–75 nmol/L, and sufficiency as >75nmol/L. We estimated the association between baseline vitamin D status and incident TB using conditional logistic regression in the Lima cohort and generalized linear mixed models in the meta-analysis. We further defined severe vitamin D deficiency as 25–(OH)D < 25 nmol/L and performed stratified analyses by HIV status in the IPD meta-analysis. In the Lima cohort, we analyzed 180 cases and 709 matched controls. The adjusted odds ratio (aOR) for TB risk among participants with baseline vitamin D deficiency compared to sufficient vitamin D was 1.63 (95% CI 0.75–3.52; p = 0.22). We included seven published studies in the meta-analysis and analyzed 3,544 participants. In the pooled analysis, the aOR was 1.48 (95% CI 1.04–2.10; p = 0.03). The aOR for severe vitamin D deficiency was 2.05 (95% CI 0.87–4.87; p trend for decreasing 25–(OH)D levels from sufficient vitamin D to severe deficiency = 0.02). Among 1,576 HIV-positive patients, vitamin D deficiency conferred a 2-fold (aOR 2.18, 95% CI 1.22–3.90; p = 0.01) increased risk of TB, and the aOR for severe vitamin D deficiency compared to sufficient vitamin D was 4.28 (95% CI 0.85–21.45; p = 0.08). Our Lima cohort study is limited by the short duration of follow-up, and the IPD meta-analysis is limited by the number of possible confounding covariates available across all studies.
Our findings suggest vitamin D predicts TB disease risk in a dose-dependent manner and that the risk of TB disease is highest among HIV-positive individuals with severe vitamin D deficiency. Randomized control trials are needed to evaluate the possible role of vitamin D supplementation on reducing TB disease risk.
| Although multiple lines of evidence suggest that vitamin D may play a role in host susceptibility to tuberculosis (TB) disease, the impact of low vitamin D on the risk of developing TB disease has not yet been firmly established.
We conducted a study in Lima, Peru, in which we measured serum 25–(OH)D levels in individuals at high risk for TB disease and followed them for development of TB over 1 year.
We pooled individual-level data from seven previously published prospective studies conducted worldwide and from our Lima cohort.
We found that individuals with low levels of vitamin D were at higher risk of future progression to TB disease.
These findings suggest that vitamin D deficiency is a risk factor for developing TB disease.
Randomized control trials are needed to determine whether vitamin D supplementation reduces risk of TB disease.
| The global burden of tuberculosis (TB) remains high, with approximately one-fourth to one-third of the world’s population infected with Mycobacterium tuberculosis (MTB), and the World Health Organization (WHO) estimates 10 million people developed TB disease in 2017 [1]. Concurrently, vitamin D deficiency (VDD) is a widespread problem globally, with a high degree of geographic variability; reported adult prevalence of VDD ranges from 10% in North America to >80% in parts of Asia [2,3]. Vitamin D is an important regulator of the immune system [4], and in vitro studies have elucidated multiple mechanisms by which vitamin D influences the pathogenesis of TB infection or disease [5–8]. At the level of the macrophage, vitamin D is involved in cathelicidin- and interferon gamma (IFN-γ)-mediated activity against mycobacteria, [5,6], induction of oxidative species [7], and the promotion of phagolysosome fusion, which leads to degradation of mycobacteria [8]. Discoveries about the various ways vitamin D modulates specific host immune responses to TB infection have focused attention on the possibility that low vitamin D levels may contribute to TB disease progression.
Numerous observational studies have also documented lower serum vitamin D levels among TB patients compared to healthy controls, and prior meta-analyses investigating the association between vitamin D and TB have concluded that low vitamin D increases TB disease risk [9–12]. However, most studies were cross-sectional studies and assessed vitamin D status after the diagnosis of active TB disease, rather than the impact of preexisting vitamin D levels on the risk of progression to TB disease. Given TB disease can induce profound metabolic abnormalities, it is unclear whether VDD increases TB disease risk or whether underlying TB infection or disease leads to decreased serum 25–(OH)D levels. Furthermore, prior studies evaluating the association between vitamin D and TB disease have used different cutoffs to categorize vitamin D levels or define VDD [9–12]. Hence, it is challenging to determine whether there is a vitamin D threshold below which individuals are at the greatest risk of TB disease.
Here, we address the association of vitamin D status on the risk of TB progression in two ways. We first report results of a case-control analysis nested in a prospective cohort study of household contacts (HHCs) of TB patients that we conducted in Lima, Peru. We next pool these data with those from other published prospective studies of vitamin D status and TB risk to conduct an individual-participant data (IPD) meta-analysis synthesizing available evidence on the association between vitamin D status and incident TB disease.
Upon completing our nested case-control study analysis, we conducted a systematic review of prior prospective studies investigating the association between vitamin D status and incident TB disease. After identifying fewer than 10 studies, most of which were conducted in small cohorts and used different thresholds for categorizing vitamin D, we undertook an IPD meta-analysis to harmonize the definition of VDD and increase the power to detect differences in risk of TB by vitamin D status.
Among 6,751 HIV-negative HHCs with baseline blood samples, 258 developed TB disease, 66 within 15 days of enrollment and 192 thereafter. Among these 192 secondary TB cases, 152 (79.1%) were microbiologically confirmed, and viable blood samples were available for analysis for 180 (93.8%) at the end of follow-up (Fig 1). Monthly incident of TB cases is shown in S1 Fig. Among the 180 TB cases analyzed, 147 (81.7%%) were microbiologically confirmed. Table 1 lists baseline characteristics of the incident cases and their matched controls. The median levels of 25–(OH)D at baseline were similar among cases (53.9 nmol/L; interquartile range [IQR] 42.7–64.0 nmol/L) and controls (54.7 nmol/L; IQR 44.5–67.1 nmol/L; p = 0.32) (Table 2). Median serum 25–(OH)D levels during spring/summer (57.2 nmol/L; IQR 46.1–69.0 nmol/L) were higher than during winter (49.6 nmol/L; IQR 40.5–59.9 nmol/L; p < 0.001) (S2 Fig).
In the univariate analysis, the differences in the risk of incident TB disease among HHCs with baseline VDD and baseline VDI compared to those with sufficient levels were not statistically significant (odds ratio [OR] VDD 1.54; 95% CI 0.88–2.71; p = 0.13 and OR VDI 1.23; 95% CI 0.72–2.08; p = 0.45) (Table 3). After we adjusted for BMI categories, SES, heavy alcohol consumption, tobacco use, IPT, TB infection status, comorbid disease, self-reported DM, index patient smear status, and season of sample collection, compared to those with sufficient vitamin D levels, the ORs for HHCs with baseline VDD and VDI were 1.63 (95% CI 0.75–3.52; p = 0.22) and 1.12 (95% CI 0.57–2.23; p = 0.75), respectively (Table 3). When we further adjusted for VAD, the ORs for HHCs with VDD and VDI were 1.59 (95% CI 0.71–3.55; p = 0.26) and 1.12 (95% CI 0.55–2.27; p = 0.75), respectively. A test for interaction between VAD and VDD was not statistically significant (p for interaction = 0.07) (S1 Table).
Our conclusions did not differ from results of the main analysis when we restricted our analyses to cases (and matched controls) diagnosed at least 60 days after index patient enrollment or to microbiologically confirmed TB cases (Table 4). We did not find a statistically significant association between VDD and risk of TB diagnosed less than 60 days after index case enrollment (OR 0.98; 95% CI 0.20–4.72; p = 0.98).
We identified 2,689 citations from the initial PubMed and Embase searches through December 31, 2017. After screening titles and abstracts, we excluded 2,678 articles because they were reviews, meta-analyses, letters, editorials or protocols (n = 1,212), case reports (n = 515), studies of other diseases or other outcomes (n = 331), animal or in vitro studies (n = 247), case-control or cross-sectional studies that assessed vitamin D status after TB disease diagnosis (n = 159), studies that did not measure vitamin D (n = 144), studies of TB treatment outcomes (n = 63), and studies of TB infection (n = 7) (Fig 2). We reviewed full texts of the remaining 11 articles [28–38] and further excluded three studies that assessed outcomes of TB-IRIS [35–37] and one study of TB infection with seasonality of TB [38]. Table 5 provides information about the seven eligible published studies [28–34] identified from the systematic review through December 31, 2017. All seven studies included in the IPD meta-analysis attained at least seven points on the NOS scale and were categorized as “good quality” studies (S2 Table).
In the updated search during peer review, we identified one additional eligible study published between January 1, 2018, and June 8, 2019, that had not been included in the IPD meta-analysis [39]. Details are provided as supporting information (S3 Fig, S3 Table).
We obtained IPD from all eligible seven studies published through December 31, 2017. One study provided patient data from a multisite evaluation conducted in nine countries [34]. Six of the seven studies were prospective cohort or case-cohort studies [28, 29, 31–34], whereas one study was a nested case-control study [30]. The final combined dataset with our Lima cohort study included 3,544 participants from 13 countries: Brazil, The Gambia, Haiti, India, Malawi, Pakistan, Peru, South Africa, Spain, Tanzania, Thailand, US, and Zimbabwe. We analyzed a total of 456 TB cases. The median time to TB diagnosis from enrollment was 151.0 days (IQR 44.0–342.0 days). Table 6 lists the baseline characteristics of all patients analyzed. The majority of the participants (86.5%) were over 15 years of age. HIV status was unknown for 629 (17.7%) patients, whereas 1,711 (48.3%) were HIV positive. One study assessed serum 25–(OH)D levels using HPLC [32], whereas others used immunoassay [28–30,34] and ELISA [31,33]. The median baseline level of 25–(OH)D was 65.0 nmol/L (IQR 48.8–83.5 nmol/L). The prevalence of VDD at baseline was 26.2% and of severe VDD was 4.8%. Most of the participants with severe VDD were from studies conducted in Pakistan (35.7%) [33], Spain (25.2%) [28], and India (17.5%) [30]. Median serum 25–(OH)D levels were higher among HIV-positive participants (74.3 nmol/L; IQR 58.0–90.0 nmol/L) compared to HIV-negative individuals (56.5 nmol/L; IQR 44.6–72.5 nmol/L; p < 0.0001). The studies in the IPD meta-analysis did not collect similar covariates; therefore, we did not compare additional baseline variables by HIV status.
In the univariate analysis, baseline VDD was associated with a 49% increased risk of progression to TB disease (OR 1.49; 95% CI 1.07–2.07; p = 0.02), and the OR for VDI compared to vitamin D sufficiency was 1.26 (95% CI 0.95–1.66; p = 0.11) [Table 7]. Both VDD and VDI remained associated with an increased risk of TB disease after we adjusted for age, gender, BMI, and HIV status (adjusted OR [aOR] for VDD: 1.48; 95% CI 1.04–2.10; p = 0.03; R2 = 0.97 and aOR for VDI: 1.33; 95% CI 1.00–1.78; p = 0.05; R2 = 0.98).
When we stratified by HIV status, HIV-positive individuals with VDD were twice as likely to develop TB disease compared to those with normal levels (aOR 2.18; 95% CI 1.22–3.90; p = 0.01; R2 = 0.97), whereas the aOR for TB disease among HIV-negative participants with VDD was 1.20 (95% CI 0.74–1.93; p = 0.46; R2 = 0.99) (Table 8, p for interaction = 0.17).
In the entire IPD cohort, the aOR for incident TB among those with severe VDD was 2.05 (95% CI 0.87–4.87; p trend for stepwise decrease in serum 25–(OH)D levels from sufficient to severe deficiency = 0.02; R2 = 0.91) (Table 9). Among HIV-positive individuals, the OR for severe VDD compared to sufficient vitamin D levels was 4.28 (95% CI 0.85–21.45; p = 0.08; R2 = 0.90). In contrast, among HIV-negative individuals, the OR for severe VDD was 1.55 (95% CI 0.55–4.34; p = 0.41; R2 = 0.97) (Table 9, p for interaction 0.17).
When we separately considered incident cases diagnosed at least 60 days after enrollment, VDI remained associated with increased risk of TB disease (aOR 1.40; 95% CI 1.02–1.92; p = 0.04; R2 = 0.97), whereas VDD was no longer associated with incident TB (aOR 1.02; 95% CI 0.67–1.57; p = 0.92; R2 = 0.95). Only 51.1% of incident TB cases in the IPD meta-analysis cohort had data on microbiologic confirmation, and we therefore did not conduct a sensitivity analysis.
In the one publication we identified that appeared after the target dates for the IPD meta-analysis, Maceda and colleagues [39] reported on a nested case-control study of 72 male prisoners in Brazil. Mean 25–(OH)D levels did not differ significantly among cases (92.5 ± 37.0 nmol/L) and controls (93.8 ± 27.5 nmol/L), and there was no association between serum 25–(OH)D < 75 nmol/L and risk of incident TB disease during 1 year of follow-up (aOR 0.59; 95% CI 0.13–2.62) (S3 Table).
Our IPD meta-analysis provides consistent support for a modest dose-dependent effect of vitamin D on future progression of TB disease across multiple studies conducted in diverse contexts. In the IPD, the association of low serum 25–(OH)D levels with increased TB disease risk was most pronounced among HIV-positive individuals with severe VDD.
Although some previous studies have documented lower vitamin D levels among patients with active TB compared to healthy controls [9,10,12,40], others have not confirmed the association between VDD and increased TB disease risk [11]. Furthermore, even among studies that found lower vitamin D levels in TB patients compared to healthy controls, the causal direction of this association is difficult to infer, since TB disease can lead to reduced dietary intake and micronutrient deficiencies, which resolve with successful treatment [10]. Although a number of intervention studies have found little evidence of an impact of vitamin D supplementation on TB treatment outcomes [10,41–46], a recent study demonstrated improved outcomes in a subgroup of multidrug-resistant patients who received supplementation [47]. In contrast to studies of micronutrients and TB treatment outcomes [10,41–46], relatively few studies have prospectively investigated the role of preexisting vitamin D status in the development of TB disease.
Our findings in these human studies support the role of vitamin D in TB infection and disease that has been inferred from more fundamental research that has enumerated multiple mechanisms by which VDD modulates host immune response to MTB. In macrophages, vitamin D is implicated in the activation of cathelicidin-mediated killing of ingested mycobacteria [5,48] induction of IFN-γ-mediated activity in macrophages [6], induction of reactive oxygen and nitrogen species [7], stimulation of phagolysosome fusion in infected macrophages [8], and inhibition of matrix metalloproteinases involved in the pathogenesis of cavitary pulmonary TB [49]. Recent evidence has also demonstrated vitamin D is involved in reduced dendritic cell–mediated priming of the adaptive immune response [50].
In addition to its impact on immunity, vitamin D status has also been linked to human metabolic phenotypes that may be involved in the pathogenesis of TB. In vitro studies have demonstrated various ways by which vitamin D promotes insulin sensitivity [51], and animal models have shown VDD impairs insulin secretion in pancreatic beta cells [51,52]. Numerous observational studies have also found an inverse association between vitamin D levels and incident type 2 diabetes mellitus (DM) [51]. Given DM is a well-described risk factor for TB disease [53,54], VDD may also contribute to increased TB risk through its role in modifying risk of diabetes.
Two other lines of evidence point to a possible association between vitamin D and TB. First, multiple studies have reported an association between specific polymorphisms of vitamin D receptor (VDR) and vitamin D binding protein (VDBP) and increased TB risk [10,55–57]. Although it is not clear that the functional effect of these polymorphisms recapitulates the impact of low vitamin D levels, several studies show that the impact of VDR variants is stronger in the presence of VDD [29,55,58]. Secondly, TB incidence varies with season and peaks in spring and summer months when vitamin D levels are highest. Some observers have postulated that low levels of sunshine, and hence vitamin D, in winter contribute to an increase in TB infection followed by a rise in TB disease incidence after a 6-month lag [38,59–61].
We also note previous studies have reported that low vitamin A is a strong predictor of incident TB disease [34, 62]. Vitamins A and D mediate changes at the cellular level by binding to nuclear hormone receptors, retinoic acid receptor, and VDR, respectively; and both receptors bind to retinoid X receptor (RXR) [63]. Some in vitro evidence further suggests vitamins A and D have synergistic activity in restricting MTB entry and reducing survival within macrophages [64]. Although we did not find evidence of a statistically significant interaction between vitamins A and D deficiencies on risk of TB, our study may not have been powered to detect this interaction. In a previous analysis of the Lima cohort, we found VAD conferred a 10-fold increase in TB disease risk [13], and here, we show that adjustment for vitamin A modestly attenuates the impact of vitamin D. Similarly, Tenforde and colleagues also reported that adjusting for retinol levels attenuated the effect of vitamin D on TB disease risk [34]. This raises the possibility that vitamin D levels correlate with other micronutrients implicated in the pathogenesis of TB, and these micronutrients may be potent mediators of increased TB risk.
Although we did not detect a statistically significant interaction between vitamin D and HIV status in the IPD, our findings raise the possibility that the effect of low serum 25–(OH)D on TB risk may be more pronounced among HIV-positive patients. Studies have shown that among HIV-infected individuals, VDD is associated with deleterious immune activation [65], lower CD4 counts [65,66], higher viral loads [65], and accelerated HIV disease progression [65,67]. Thus, VDD may exacerbate existing immune dysregulation in HIV infection to further increase TB risk, or low vitamin D levels may reflect severity of HIV-related immunosuppression. We also note that vitamin D status fluctuates with season, with declines in serum 25–(OH)D in the winter when UVB exposure is lower [68,69], and in vitro evidence suggests there is winter-associated increase in HIV replication [68]. Vitamin D also restricts mycobacterial growth in the presence of HIV infection [70]. A clinical trial is currently underway to evaluate the efficacy of vitamin D supplementation in preventing incident TB among adults with HIV in Tanzania [71]; the results may help clarify role of vitamin D in HIV-associated TB disease. We also plan to measure inflammatory markers in the Peru cohort to explore association between VDD and immune dysregulation.
We considered possible explanations for why we did not detect a significant association between VDD and TB risk in the Peru cohort, despite its relatively large size. First, since TB incidence is highest in summer [60,61] and HHCs were recruited when the index case was diagnosed, they are more likely to have been recruited and assessed when their vitamin D levels were highest. If levels later fell and this fall precipitated TB progression, this would not have been detected. Secondly, VDR variants are heterogeneously distributed in different populations and may modify the effect of vitamin D on TB risk. We did not measure VDR variants in Peru and are therefore unable to assess the prevalence of different VDR genotypes in this cohort. The Peru study is also limited by the relatively short (1-year) period of follow-up and the fact that it was only powered to detect a 3-fold or greater difference in TB incidence among people with VDD.
Our IPD meta-analysis also has some important limitations. Firstly, many possible confounding covariates were not measured across all studies. Therefore, we were unable to account for other important confounders such as baseline infection status, other micronutrient levels, and comorbidities, including DM, that might be associated with both VDD and TB risk. Secondly, although we only examined prospective studies of incident TB disease, the included studies were all observational, and we cannot exclude the possibility that participants had early, undiagnosed TB at baseline that lowered vitamin D levels. Although we addressed this by conducting a sensitivity analysis excluding incident TB cases diagnosed less than 60 days after enrollment, the smaller number of incident cases diagnosed after 60 days reduced the power to detect a statistically significant association. Thirdly, we also cannot exclude the possibility of publication bias if studies with nonsignificant findings on the link between vitamin D and incident TB have remained unpublished. We did not construct a funnel plot to assess publication bias, because we analyzed seven studies, and guidelines suggest tests for funnel plot asymmetry are not sufficiently powered to distinguish real asymmetry from chance with fewer than 10 studies [72]. These studies further used different methods to categorize vitamin D levels and therefore provided effect estimates that are not directly comparable on a funnel plot. During our systematic review, we attempted to address this by considering data reported from meeting abstracts, and none met our inclusion criteria. Of note, in the publication identified after the initial meta-analysis target dates, although Maceda and colleagues found no association between serum 25–(OH)D < 75 nmol/L and increased TB risk, our effect estimate for VDD in the IPD (OR 1.48) falls within the confidence intervals of this small cohort study [39].
We also note that the different 25–(OH)D assays employed in the meta-analysis studies vary in their sensitivity and precision. However, it is unlikely such variability introduced a bias in one direction, since within any given study, 25–(OH)D levels were analyzed using the same assay in individuals that progressed to TB and nonprogressors. The effect of any imprecision would be more likely to increase the noise:signal ratio, which would bias results of the analysis toward the null.
Despite the aforementioned limitations, we present findings from one of the largest cohorts to date evaluating the role of vitamin D on incident TB disease with comprehensive adjustment for possible confounders. The concurrent IPD meta-analysis further increased the sample size, providing the statistical power to detect more modest associations and enabling the evaluation of this relationship across different locations and by HIV status.
In conclusion, in our meta-analysis of prospective studies, we found low serum 25–(OH)D levels were associated with increased risk of future progression to TB disease in a dose-dependent manner. Randomized control trials are needed to determine whether vitamin D supplementation among individuals at high risk can mitigate the risk of developing TB disease.
|
10.1371/journal.pntd.0004056 | Financial and Economic Costs of the Elimination and Eradication of Onchocerciasis (River Blindness) in Africa | Onchocerciasis (river blindness) is a parasitic disease transmitted by blackflies. Symptoms include severe itching, skin lesions, and vision impairment including blindness. More than 99% of all cases are concentrated in sub-Saharan Africa. Fortunately, vector control and community-directed treatment with ivermectin have significantly decreased morbidity, and the treatment goal is shifting from control to elimination in Africa.
We estimated financial resources and societal opportunity costs associated with scaling up community-directed treatment with ivermectin and implementing surveillance and response systems in endemic African regions for alternative treatment goals—control, elimination, and eradication. We used a micro-costing approach that allows adjustment for time-variant resource utilization and for the heterogeneity in the demographic, epidemiological, and political situation.
The elimination and eradication scenarios, which include scaling up treatments to hypo-endemic and operationally challenging areas at the latest by 2021 and implementing intensive surveillance, would allow savings of $1.5 billion and $1.6 billion over 2013–2045 as compared to the control scenario. Although the elimination and eradication scenarios would require higher surveillance costs ($215 million and $242 million) than the control scenario ($47 million), intensive surveillance would enable treatments to be safely stopped earlier, thereby saving unnecessary costs for prolonged treatments as in the control scenario lacking such surveillance and response systems.
The elimination and eradication of onchocerciasis are predicted to allow substantial cost-savings in the long run. To realize cost-savings, policymakers should keep empowering community volunteers, and pharmaceutical companies would need to continue drug donation. To sustain high surveillance costs required for elimination and eradication, endemic countries would need to enhance their domestic funding capacity. Societal and political will would be critical to sustaining all of these efforts in the long term.
| River blindness (onchocerciasis) is a parasitic disease transmitted by blackflies. Symptoms include severe itching, skin lesions, and vision impairment including blindness. More than 99% of all cases are concentrated in sub-Saharan Africa. Fortunately, vector control and community-directed treatment with ivermectin have significantly decreased morbidity, and the treatment goal is shifting from control to elimination in Africa. To inform policymakers’ and donors’ decisions, we estimated financial resources and societal opportunity costs associated with alternative treatment goals—control, elimination, and eradication. We found that rapid scale-up of ivermectin treatment for elimination and eradication would result in substantial cost-savings in the long term as compared to staying in a control mode, because regular active surveillance would allow treatments to end earlier, thereby saving the economic costs of community volunteers and donated ivermectin. To realize cost-savings, policymakers should keep empowering community volunteers, and pharmaceutical companies would need to continue drug donation. To sustain high surveillance costs required for elimination and eradication, endemic countries would need to enhance their domestic funding capacity. Societal and political will would be critical to sustaining all of these efforts.
| The treatment goal for onchocerciasis (river blindness) has shifted from control to elimination as shown by the World Health Organization’s (WHO’s) roadmap for neglected tropical diseases (NTDs) and the London Declaration on NTDs in 2012 [1,2]. Onchocerciasis is a parasitic disease transmitted by blackflies, and notable symptoms include severe itching, skin lesions, and vision impairment including blindness. Those affected by onchocerciasis suffer negative socioeconomic consequences as a result of their symptoms [3]. The disease is endemic in parts of Africa, Latin America, and Yemen, and more than 99% of all cases are concentrated in sub-Saharan Africa [4]. In Africa, morbidity caused by onchocerciasis was significantly reduced by the vector control activities of the Onchocerciasis Control Programme (OCP) in West Africa (1975–2002) and by the community-directed treatment with ivermectin (CDTi) under the African Programme for Onchocerciasis Control (APOC) in sub-Saharan Africa and parts of West Africa (1995–present) [4]. Studies of foci in Mali, Senegal, and Uganda have proved that eliminating onchocerciasis through ivermectin administration is feasible for amenable epidemiological settings under effective treatments and surveillance [5,6].
Onchocerciasis elimination and subsequent eradication will generate health benefits by reducing the incidence of infection to zero, first in a defined area and then globally. These benefits would be higher than those of staying in a control mode that keeps disease prevalence at a locally acceptable level. In addition to epidemiological evidence, national and global policymakers must consider economic, social, and political aspects when deciding whether to invest in elimination in settings with limited resources and competing health priorities. To assess these broad aspects, a working group at the Ernst Strüngmann Forum suggested developing and analyzing eradication/elimination investment cases [7]. Tediosi and colleagues examined the suggested approach focusing on three NTDs including onchocerciasis [8]. Referring to this study, Kim and colleagues defined investment options for onchocerciasis as scenarios, and compared the respective timelines and needs for treatment in endemic African countries [9]. Each scenario consists of strategies of treatments and surveillance—epidemiological surveillance to track the infection levels in human and/or entomological surveillance to track the infectivity rates of blackflies.
To reduce disease prevalence to a locally acceptable level (i.e., microfilaria prevalence≤40% or community microfilarial load≤5mf/s [3]), all endemic African countries implement annual CDTi in hyper- and meso-endemic areas, and after at least 25-years of CDTi, conduct epidemiological surveillance to confirm that CDTi can be safely stopped (former OCP projects having implemented regular surveillance continue their surveillance strategies).
To reduce the incidence of infection to zero in a defined area, all endemic African countries except those with epidemiological and political challenges implement annual or biannual CDTi, and conduct regular active epidemiological and entomological surveillance to evaluate epidemiological trends, to decide a proper time to stop CDTi, and to detect and respond to possible recrudescence.
To reduce the incidence of infection to zero in Africa, which would lead to global eradication, all endemic African countries implement not only annual or biannual CDTi but also locally tailored treatment strategies to deliver sustainable treatments to areas with operational challenges, and implement regular active epidemiological and entomological surveillance to evaluate epidemiological trends, to decide a proper time to stop CDTi, and to detect and respond to possible recrudescence.
We estimated financial resources and societal opportunity costs for endemic African countries (Table 1) associated with the control, elimination, and eradication scenarios to support policymakers’ and donors’ informed decisions and provide a basis for further economic evaluation of the elimination and eradication of onchocerciasis.
We estimated financial costs to predict how much the governments of endemic countries and donors would have to pay for implementing the required interventions for alternative treatment goals of control, elimination, and eradication, and economic costs to assess societal opportunity costs of donated services and goods. The time horizon of the analysis was from 2013 to 2045, based on the predicted timeline for reaching the post-elimination phase in endemic African regions [9].
There are different methods for estimating health intervention costs, ranging from micro-costing (bottom-up approach) to gross-costing (top-down approach) [14]. We used a micro-costing method to more precisely estimate time-variant resource utilization depending on epidemiological trends and to incorporate the heterogeneity in the demographic, epidemiological, and political situation at project level. Fig 1 shows the six steps of the micro-costing approach calculating from the cost of a single cost item to the total financial and economic cost for a project.
We defined the key activities and resources required for onchocerciasis elimination and eradication with reference to an APOC report of the technical consultative committee [15], an APOC protocol for epidemiological surveillance, and a guide for post-treatment epidemiological and entomological surveillance (developed for the Onchocerciasis Elimination Program for the America) [16]. Based on the identified activities and resources, we defined cost items under five categories—CDTi, surveillance, capital costs, overhead and administrative costs, financial support for (post) conflict endemic areas—and their characteristics which include the type (financial or economic), the unit cost, and the time-variant unit quantity (depending on relevant phases among the three phases of treatment, confirmation of elimination, and post-elimination). The details about each step of the micro-costing approach and the characteristics of cost items are described in S1 Text.
We obtained 2012 budgets from APOC, approved for onchocerciasis CDTi, that cover 67 of all 112 ongoing projects (as of November 2013) in sub-Saharan Africa. All budget documents include information on the unit cost and the unit quantity of each resource, demography, available human resources (community health workers, community volunteers), and funding from the ministry of health, APOC, and non-governmental organizations. These data were used as the main sources to estimate financial costs. To estimate economic costs, agriculture value added per worker was used as an opportunity cost of community volunteers’ unpaid time [17], considering most volunteers are farmers in remote rural areas [18]. The opportunity cost of donated ivermectin was $1.5054 per treatment (three 3mg-tablets), based on Merck’s suggested drug price of $1.5 per treatment before the donation was decided [19] and on the insurance and freight cost of $0.0018 per tablet [20].
For projects with missing unit costs, we used the national average if relevant unit costs were available; otherwise, the regional average (Table 2) across available national averages for endemic African countries. For the countries that did not have agriculture value added per worker, we used the regional average for sub-Saharan Africa [17]. For projects with missing data for the determinants of unit quantities (e.g., the ratio of health workers over population, the ratio of volunteers over population), we used the national average if relevant data were available; otherwise, the regional average across available national averages for endemic African countries. Unit costs and the determinants of unit quantities at the country and regional levels are included in S1 Text.
We conducted sensitivity analysis to assess the robustness of results to parametric uncertainties. The parameters included either cost items for which unit costs were missing for more than one third of total projects or total countries with available budgets. Also the parameters included the time-variant determinants of unit quantities: population living in endemic areas, the number of required treatments (determined by population, treatment coverage linked to required treatment duration, and possible delay in starting and ending treatments), the number of required community volunteers (determined by population and the ratio of community volunteers over population), and the number of required community health workers (determined by population and the ratio of community health workers over population). We conducted one-way sensitivity analysis to examine the impact of parameters related to CDTi performance, the cost items with high uncertainty, and discount rates on total costs. We conducted multivariate probabilistic sensitivity analysis (PSA) to examine the joint effects of uncertainties about all selected variables on total costs. For PSA, we attached statistical distributions to the selected cost items and the determinants of unit quantities, and fitted to relevant data. S1 Text describes the methodological details of the sensitivity analysis.
Total financial and economic costs would be concentrated in the early stage during which treatments are scaled up to remaining endemic areas, and decrease as the treatment phase nears the end (Fig 2). In endemic African regions, total financial and economic costs over the period 2013–2045 would be $4.3 billion (95% central range from multivariate PSA: $3.9 billion[bn]–$5.0bn) for the control scenario, $2.9 billion ($2.6bn–$3.4bn) for the elimination scenario, and $2.7 billion ($2.4bn–$3.2bn) for the eradication scenario. That is, switching from control to elimination and eradication would lead to cost-savings of $1.5 billion ($1.0bn–$1.9bn) and $1.6 billion ($1.2bn–$2.1bn), respectively (S1 Fig). The eradication scenario would lead to cost-savings of $144 million (-$25 million[M]–$462M) as compared to the elimination scenario.
Unit financial and economic cost per treatment for the control scenario would decrease from $2.5 to $0.9 over 2013–2045. For the elimination scenario, it would decrease from $2.5 to $1.3 until 2035, and increase to $1.6 afterwards. For the eradication scenario, it would decrease from $2.5 to $1.5 over 2013–2030, and increase to $3.9 afterwards until the end of the treatment phase in endemic African regions (Fig 3).
Total financial costs over the period 2013–2045 would be $640 million ($572M–$711M) for the control scenario, $650 million ($574M–$751M) for the elimination scenario, and $649 million ($566M –$745M) for the eradication scenario (Fig 4). That is, the total financial costs associated with the elimination and eradication scenarios are slightly lower than those associated with the control scenario; however, these cost differences are not robust to sensitivity analysis (S2 Fig). The main difference between scenarios is the proportion of surveillance costs in total costs. Total surveillance costs over 2013–2045 would increase from 7% ($47M) of total financial costs under the control scenario to 33% ($215M) and 37% ($242M) under the elimination and eradication scenarios, respectively (Fig 4).
Unit financial cost per treatment for the control scenario would decrease from $0.4 to $0.1 over 2013–2045. For the elimination scenario, it would stay between $0.4 and $0.5 until 2035, and increase to $0.9 afterwards. For the eradication scenario, it would stay between $0.4 and $0.5 until 2030, and increase to $3.1 as the treatment phase nears the end in endemic African regions (Fig 3).
Economic costs would be six times higher than financial costs under the control scenario and three times higher under the elimination and eradiation scenarios. Total economic costs over 2013–2045 would be $3.7 billion ($3.3bn–$4.3bn) for the control scenario, $2.2 billion ($2.0bn–$2.7bn) for the elimination scenario, and $2.1 billion ($1.8bn–$2.5bn) for the eradication scenario (Fig 5). That is, the total economic costs associated with the elimination and eradication scenarios are lower than those associated with the control scenario by $1.5 billion ($1.1bn–$1.9bn) and $1.6 billion ($1.2bn–$2.1bn), respectively (S3 Fig). Donated ivermectin and community volunteers would account for 75% and 25% of the total economic costs over 2013–2045 in all scenarios.
One-way sensitivity analysis (Fig 6) shows that, among the parameters related to CDTi performance, the delay in ending CDTi (after the infection levels reach the threshold for stopping CDTi) is the most influential parameter, leading total costs to increase by $2 billion (undiscounted) over 2013–2045 in all scenarios. Among the cost items with high uncertainty (based on the number of missing data), the most influential one is the salary top-ups for stabilizing new projects in the elimination and eradication scenarios, leading total costs (undiscounted) to range from $3.807 billion to $3.847 billion, and from $3.460 billion to $3.498 billion, respectively. Increasing the discount rate from 0% to 6% would decrease total costs over 2013–2045 by 46% from $6.1 billion to $3.3 billion for the control scenario, by 39% from $3.8 billion to $2.3 billion for the elimination scenario, and by 35% from $3.5 billion to $2.2 billion for the eradication scenario.
The elimination and eradication scenarios are predicted to generate substantial cost-savings in the long run compared to the control scenario. The main factors contributing to cost-savings are the reduction in economic costs of community volunteers and donated ivermectin due to a shorter treatment phase as a result of regular active surveillance. This finding implies that the saved volunteers’ time and ivermectin can be used for other health programs. Willing volunteers and their well-established networks, which have enabled successful implementation of CDTi in Africa, could contribute to improving access to primary health care in remote rural areas with insufficient human resources. In addition, the saved ivermectin drugs could be used for other disease programs, for example, anti-LF mass drug administration. To realize these possibilities, policymakers would need to keep empowering community volunteers through training and societal or economic appreciation. Also, pharmaceutical companies’ continuous commitment to donating drugs would be needed.
The main operational difference between the elimination/eradication scenarios and the control scenario is regular active surveillance. Our analysis shows that the cumulative financial costs for surveillance over 2013–2045 in the elimination and eradication scenarios would be five times higher those in the control scenario. This implies that endemic countries would need to improve their domestic funding capacity to sustain high surveillance costs to achieve elimination, as the post-treatment surveillance period could last beyond 2045 [9] and external funding would be temporary. The development and operationalization of new affordable and effective diagnostic tools, for example, OV-16 (ELISA and Rapid Test) and the DEC patch test under development [27,28], might lead to the savings of surveillance costs.
The financial unit cost per treatment in the elimination and eradication scenarios would increase by factors of respective two and eight as the regional intervention phase nears the end. This increase is driven by the reduction in the number of people in need of treatment and steady or increasing costs for surveillance and capital goods. Additionally, in the last stage, the majority of people in need of treatment are expected to live in areas with epidemiological and political challenges [9]. This implies that, in the last mile towards elimination and eradication, political, financial, and societal commitment across a whole spectrum of stakeholders will be essential to meet high unit costs and to deliver treatments in challenging areas [29]. Studies based on social choice theory and game theory [30–33] show that the elimination and eradication of infectious diseases are public goods that can only be achieved through the coordinated efforts of multiple countries. These studies suggest that high benefit-cost ratios associated with elimination and/or eradication could incentivize endemic countries to pursue elimination and/or eradication and global donors to finance endemic countries lacking the financial capacity. Equity and social justice arguments for elimination and eradication [34,35] could also complement and strengthen the economic rationality. The role of global stakeholders can play a decisive role to overcome national challenges. A study by Shaffer suggests that, to prevent potential holdout problems caused by unwilling or unable countries, which could hinder elimination and eradication, the centralized efforts led by international organizations would be necessary [36]. In line with this, it has been argued that the explicit inclusion of NTDs elimination in the Sustainable Development Goals (SDGs) of the United Nations (UN) [37,38] would further motivate the commitment of national and global policymakers and donors. Societal commitment at local level will be also essential, because delivering treatments to operationally challenging areas would require successful drug administration by community volunteers and communities’ compliance to treatments. To promote such commitment by communities, endemic countries’ continuous investments in enhancing the operational capacity of community volunteers and in mobilizing communities will be needed.
The uncertainty analysis showed that the delay in ending CDTi would have the highest impact among those related to CDTi performance on total costs. Thus, planning to move towards the post-treatment phase, along with regular monitoring and evaluation to decide the proper time of stopping treatments, would be important to avoid the delay in ending CDTi.
The uncertainty analysis also showed that the salary top-ups for stabilizing new projects would have the most influence of all cost items on total costs. Many new projects are in potential hypo-endemic areas where parasitological surveys are still needed to confirm endemicity [39]. This suggests that complete epidemiological mapping should be a priority to choose areas to start new projects and to predict required human resources for those projects.
The results presented in this study should be interpreted considering the limitations of the approach and data used. To calculate financial costs for projects without available budgets, we relied on national or regional average unit costs which might only approximately represent the actual costs in those projects. For economic costs, we assumed agriculture value added per worker as an opportunity cost of community volunteers’ unpaid time. However, other studies used different proxies such as national minimum wage and GNI per capita [24,40]. We did not use national minimum wage, as it was unavailable for 11 of 28 endemic countries [41]. We did not use GNI per capita, as it does not represent the income level in remote rural areas. In the opportunity cost of donated ivermectin, we did not include tax deduction provided to donating manufacturers [19], as the relevant detailed information is proprietary and unavailable.
There were some other factors that could affect resource utilization, but were not included in the analysis. We assumed no recrudescence, because it was difficult to predict when recrudescence would happen. If that were to happen, costs would increase because the treatment phase would have to be restarted. We did not consider the potential impact of new diagnostic and treatment tools, because it was difficult to predict when they would be developed and operationalized. If new effective and affordable tools are operationalized, the strategies of treatment and surveillance could change, thereby influencing costs. We assumed no unexpected political unrest that could interrupt interventions and would increase costs to restart the interventions.
Despite these limitations, to our knowledge and based on literature review (see S1 Text), our study is the most up-to-date cost analysis of potential regional elimination strategies in Africa. National and global policymakers and donors could use our cost analysis to make informed policy decisions and to predict the funding needs for implementing elimination programs in Africa. Our cost estimates could also be used by policymakers and researchers to compare costs and potential benefits associated with potential elimination strategies in Africa.
|
10.1371/journal.pbio.2004979 | Unconventional function of an Achaete-Scute homolog as a terminal selector of nociceptive neuron identity | Proneural genes are among the most early-acting genes in nervous system development, instructing blast cells to commit to a neuronal fate. Drosophila Atonal and Achaete-Scute complex (AS-C) genes, as well as their vertebrate orthologs, are basic helix-loop-helix (bHLH) transcription factors with such proneural activity. We show here that a C. elegans AS-C homolog, hlh-4, functions in a fundamentally different manner. In the embryonic, larval, and adult nervous systems, hlh-4 is expressed exclusively in a single nociceptive neuron class, ADL, and its expression in ADL is maintained via transcriptional autoregulation throughout the life of the animal. However, in hlh-4 null mutants, the ADL neuron is generated and still appears neuronal in overall morphology and expression of panneuronal and pansensory features. Rather than acting as a proneural gene, we find that hlh-4 is required for the ADL neuron to function properly, to adopt its correct morphology, to express its unusually large repertoire of olfactory receptor–encoding genes, and to express other known features of terminal ADL identity, including neurotransmitter phenotype, neuropeptides, ion channels, and electrical synapse proteins. hlh-4 is sufficient to induce ADL identity features upon ectopic expression in other neuron types. The expression of ADL terminal identity features is directly controlled by HLH-4 via a phylogenetically conserved E-box motif, which, through bioinformatic analysis, we find to constitute a predictive feature of ADL-expressed terminal identity markers. The lineage that produces the ADL neuron was previously shown to require the conventional, transient proneural activity of another AS-C homolog, hlh-14, demonstrating sequential activities of distinct AS-C-type bHLH genes in neuronal specification. Taken together, we have defined here an unconventional function of an AS-C-type bHLH gene as a terminal selector of neuronal identity and we speculate that such function could be reflective of an ancestral function of an “ur-” bHLH gene.
| Across the animal kingdom, transcription factors of the basic helix-loop-helix (bHLH) family act during embryonic nervous system patterning as proneural genes to promote neuroblast identity. We describe here a distinct function for a specific member of this family, hlh-4, in the nematode Caenorhabditis elegans. hlh-4 is exclusively expressed in a nociceptive neuron class and is not required for this neuron class to be generated but is rather required for the execution of its terminal differentiation program. hlh-4 directly controls the expression of scores of terminal identity features of this neuron class, including its large battery of chemoreceptor-encoding genes. We propose that a role of bHLH genes in controlling terminal differentiation may be the ancestral function of members of this gene family.
| Nervous system development proceeds through sequential steps, starting with the early commitment to a neuronal fate, followed by the progressive restriction of fates, to finally reaching a terminal, differentiated end state. Proneural genes of the basic helix-loop-helix (bHLH) family play a key role in the initial stages of this process [1]. Mutant analysis in Drosophila revealed that loss of members of the Achaete-Scute complex (AS-C), as well as the related Atonal gene, resulted in the loss of the ability to generate neuroblasts in the peripheral nervous system [2–5]. Vertebrate orthologs of proneural AS-C and Atonal genes (the Mash and Math genes) also provide critical proneural function in vertebrate nervous system development [1,6–8]. Thus, the proneural function of AS-C-type and Atonal bHLH genes is broadly conserved throughout evolution.
The C. elegans genome encodes a canonical complement of homologs of proneural bHLH genes, including seven AS-C-like genes (hlh-4, hlh-3, hlh-14, hlh-19/hnd-1, hlh-12, hlh-6, hlh-16) and one Atonal ortholog (lin-32) [9]. The function of many of these C. elegans bHLH genes in the nervous system has not been as extensively studied as their fly and vertebrate orthologs, but it is nevertheless clear that a number of these bHLH genes also provide proneural activities [10–12]. Like in flies and vertebrates, C. elegans proneural bHLH genes operate in a lineage-specific manner. For example, the C. elegans AS-C ortholog hlh-14 and the C. elegans Atonal ortholog, lin-32, provide proneural activity in several distinct sensory neuron lineages of the peripheral and central nervous system (CNS) of the worm [10–12]. In both cases, the proneural activity of hlh-14 and lin-32 is exemplified by a transformation of neuroblasts into cells with a hypodermal identity in the respective mutant backgrounds.
One question that has been studied extensively over the years is whether AS-C/Atonal-type bHLH genes have functions in the nervous system that go beyond their proneural activity. In both vertebrates and flies, nonproneural functions of AS-C and Atonal-like genes have indeed been described in the context of later neuronal differentiation events (reviewed in [1,6,13]). Similarly, C. elegans lin-32/Ato has functions beyond its proneural activity in male ray lineages in which lin-32 also allocates fates in subsequently developing ray sublineages [14]. However, in all these cases, the respective bHLH gene is either transiently expressed; acts through downstream, intermediary regulatory factors; or only affects selected aspects of the differentiated state of the respective neuron.
In this study, we describe a novel, nonproneural, and noncanonical function of an AS-C-type bHLH gene. We find that the AS-C homolog hlh-4 displays a spatial and temporal specificity of expression that is unprecedented for any bHLH gene. hlh-4 is exclusively and continuously expressed in a single postmitotic nociceptive sensory neuron class in which it initiates and maintains the terminal identity of this neuron via direct binding to scores of terminal effector genes that are expressed in a neuron class–specific manner and that define the differentiated state of this neuron. Among its many functions in ADL, hlh-4 directly regulates the expression of the unusually large repertoire of olfactory receptor proteins in ADL. We hypothesize that the direct control of “neuron function genes” may have been an ancestral function of bHLH genes.
Strains were maintained by standard methods [15]. A list of all strains used is listed in S3 Table.
Green fluorescent protein (GFP) reporters for rescue and ectopic expression were generated using RF-cloning [16]. For making G-protein coupled receptor (GPCR) transgenic reporters (listed in S3 Table), a PCR fusion approach was used [17]. Genomic fragments were fused to the GFP coding sequence, which was followed by the unc-54 3′ untranslated region. All transgenic lines created in this study were injected at 50 ng/μL with the unc-122::gfp into wild-type animals or with the pha-1 rescuing plasmid (pBX) as a coinjection marker (50 ng/μL) into pha-1 mutant animals. For each construct, two independent lines were scored.
Fosmid-based reporters for hlh-2, hlh-3, and hlh-4 were generated by insertion of yfp at the 5′ end of the hlh-2 locus [18], 3′ end of hlh-4 (this paper), and gfp at the 3′ end of hlh-3 [19] using standard fosmid recombineering approaches [19,20].
The arrd-4 promoter (1,587 bp) was cloned together with hlh-4 genomic sequences and unc-54 3′UTR into a pPD95.75 backbone and injected (50 ng/μL) into OH14884 as a simple array, with unc-122::gfp (50 ng/μL) as a coinjection marker. The unc-3 promoter fusion was generated by amplification of 558 bp of unc-3 promoter, fused to hlh-4 genomic (including its own 3′UTR), using the PCR fusion approach [17]. Fifty nanograms per milliliter of this construct were injected into OH14884, with ttx-3::mcherry as a coinjection marker.
The eat-4 reporter constructs were generated by PCR and subcloning into pPD95.75 vector. eat-4prom6-1 contains 4,450 bp of the upstream region of the ATG and eat-4prom2 contains 1,150 bp of the genomic region just upstream of the ATG. The E-Box and homeodomain motif are found at positions -693 and -726 relative to the ATG start codon, respectively. The specific sequences deleted are, for the E-Box, AACAGGTGTT, and for the homeodomain site, ATTAGATAAT. The deletions were generated by mutagenesis with the QuickChange Site-Directed Mutagenesis kit (Stratagene). The plasmids were injected into OH13645 [otIs518;him-5(e1490)] at 50 ng/μL, using unc-122::gfp (50 ng/μL) as a coinjection marker.
Worms were anesthetized using 50 mM sodium azide (NaN3) and mounted on 5% agarose on glass slides. Images were acquired using an automated fluorescence microscope (Zeiss, AXIO Imager Z.2) or LCS-8 laser point scanning confocal. Representative images are shown following maximum projection of Z-stacks using the maximum intensity projection type. Image reconstruction was performed using Fiji software [21].
ADL neurons were identified by labeling subsets of sensory neurons with DiD or DiO (Thermo Fisher Scientific). For dye filling, worms were washed with M9 and incubated at room temperature with DiD (1:500) in M9 for 1 hour for Adults or (1:250) for 2 hours for L1 stage animals. After incubation, worms were washed three times with M9 and plated on agar plates coated with food (OP50 bacteria) for 1–3 hours before imaging.
The expression of bHLH fosmid reporters was manually lineaged using SIMI BioCell program, as previously described [22]. Briefly, the gravid adults of hlh-4Fosmid::yfp (otIs683) and hlh-3fosmid::gfp (otIs648) were dissected and single two-cell embryos were mounted and visualized on a Zeiss Imager Z1 compound microscope using the 4D microscopy software, Steuerprg (Caenotec). Nomarski stacks were taken every 30 seconds and embryos were illuminated with LED fluorescence light (470 nm) at predetermined time points during development.
Avoidance assay was performed as previously described [23,24]. L4 stage animals were picked onto OP50 seeded plates before a day of assay. We used 100 nM or 500 nM ascr#3 or 1M glycerol diluted in M13 buffer. In the assay, M13 buffer was firstly dropped in front of animals’ heads. When the animals didn’t respond to M13 buffer, we then dropped ascr#3/glycerol and checked avoidance to the stimulus. Long reversals were counted as avoidance [25]. The tests were done at least 5 times with 10 animals each.
Motif discovery was carried out using information-theoretic analysis as implemented in the Finding Informative Regulatory Elements (FIRE) algorithm [26]. De novo motifs were discovered by running FIRE in discrete mode, with all the genes in the C. elegans genome labeled as either belonging to class 1: the neuron-specific expression class (e.g., 117 ADL-expressed genes) or class 2: the complementary set of all other remaining genes. The starting k-mer seed length was set to k = 6 and the sequence search space was confined to 2-kb upstream regions. The discovered CACCTG motif had a robustness score of 10/10 with a significance z-score of 18.3.
We used TargetOrtho [27] to find whole genome CACCTG motif matches in five nematode genomes searching 2 kb upstream of each gene plus introns. ADL-expressed genes and all C. elegans genes, excluding noncoding RNAs, were compared using the Wilcoxon rank sums test to assess alignment independent species conservation scores, motif match position relative to the start codon, and motif match frequency per gene. Only genes with at least one CACCTG match were analyzed.
As a first step toward a systematic analysis of the neurogenic function of C. elegans bHLH genes, we undertook a nervous system–wide expression pattern analysis of all C. elegans AS-C-like genes. Using fosmid-based reporter transgenes, we found that many bHLH genes are expressed during embryonic development within and outside neuronal lineages, but we noticed that one AS-C-like bHLH gene, hlh-4, displays an unusual expression pattern, both in terms of spatial and temporal specificity (Fig 1). hlh-4 expression is not observed in any blast cells during embryonic or postembryonic development but rather is first expressed in two pairs of postmitotic cells in the precomma stage embryo, shortly after their birth (Fig 1A). One pair is the ADL neurons and the other pair is the sisters of ADL, which die shortly after their birth by programmed cell death [28]. Expression of hlh-4 in ADL is observed for the remainder of embryogenesis, continues during larval and adult stages, and is never observed in any other cell throughout the entire organism (Fig 1A). The fosmid on which the yfp reporter construct is based is able to fully rescue the hlh-4 mutant phenotype that we describe below (rescue data are shown in Table 1). The ADL-specific fosmid-based reporter expression pattern is recapitulated by a 700-bp 5′ promoter fusion reporter (Fig 1C).
With the exception of hlh-3, which is expressed in a subclass of postmitotic motor neurons of the ventral nerve cord [31], none of the other C. elegans AS-C-like bHLH genes (hlh-6, hlh-12, hlh-14, hlh-16, hlh-19/hnd-1) share the postmitotic, post-developmental neuronal expression feature of hlh-4 [12,32–34]. We note that while our fosmid-based hlh-3 reporter showed extensive expression in blast cells during embryogenesis, it does not recapitulate the postembryonic ADL expression previously reported using a reporter that only contained 1.5 kb of 5′ sequences upstream of the gene [35].
The only other bHLH reporter expressed in postmitotic neurons throughout embryonic, larval, and adult stages is the Daughterless homolog hlh-2/Da [29], a binding partner of many C. elegans AS-C-related bHLH genes [30]. Expression of HLH-2/DA protein in a specific subset of postmitotic neurons, including the nociceptive neurons ADL and ASH, has been previously reported using anti-HLH-2 antibody staining [29], but it was not reported whether expression persisted into later larval and/or adult stage. Using a fosmid-based reporter of hlh-2/Da expression, we found that ADL and ASH expression of hlh-2/Da, as well as expression in a few other head and tail neurons, is maintained throughout all larval stages into adulthood (Fig 1B). We conclude that hlh-4/AS-C and its heterodimerization partner hlh-2/Da are continuously coexpressed specifically in the nociceptive ADL neuron class.
One well-documented mechanism by which transcription factors ensure their continuous expression throughout the life of a neuron is through transcriptional autoregulation (e.g., [36–39]). To assess whether continuous expression of hlh-4 throughout the life of the ADL neuron is also ensured by autoregulation, we used a 5′ promoter fusion of the hlh-4 locus, which recapitulated the continuous expression of hlh-4 in ADL (Fig 1C). We crossed this reporter into an hlh-4 mutant allele, tm604, a putative null allele generated by the C. elegans knockout consortium in Tokyo [40] in which the bHLH domain is largely deleted (Fig 1A). We found that hlh-4 reporter expression in the ADL neuron pair is initiated normally in hlh-4 mutant embryos, but expression fails to be maintained beyond the first larval stage (Fig 1C). As yet unknown factors may initiate hlh-4 expression in the embryo and, after its initiation, hlh-4 takes over to regulate its own expression.
We furthermore tested whether continuous expression hlh-2/Da in ADL requires hlh-4 activity. Crossing the hlh-2 fosmid reporter into the hlh-4 mutant background, we indeed found this to be the case (Fig 1B). We conclude that the continuous expression of both hlh-4 and its putative cofactor hlh-2/Da is based on transcriptional autoregulation.
In most if not all organisms examined, AS-C genes have proneural function, characterized by a loss of neuroblast identity in the absence of the AS-C gene and ensuing conversion into an ectodermal identity [1,3,6,13]. Previous work has demonstrated that in the lineage that produces ADL, as well as other sensory neurons, the transiently and early-expressed AS-C gene hlh-14 acts as a proneural gene, such that loss of hlh-14 results in a neuroblast to hypodermal fate conversion [12]. In striking contrast, we find that the later-expressed hlh-4 gene does not act as a proneural gene. Specifically, in hlh-4 null mutants, the ADL neuron pair is still generated and differentiates as a neuron, as assessed by (a) intact expression of a panneuronal reporter, rab-3, (b) intact filling of the ADL neuron with the dye DiI (which is taken up by the dendritic endings of several sensory neurons, including ADL [41]), and (c) presence and intact speckled appearance of the ADL neuronal nucleus by Nomarski optics (Fig 2A). Corroborating this notion, we find that the two genes that are expressed by all ciliated sensory neurons, osm-6 and ift-20 [42,43], are still normally expressed in the ADL neurons of hlh-4 mutants (Fig 2B). Even though we could not confirm the previously reported expression of hlh-3 in ADL (Fig 1A), we nevertheless generated hlh-3; hlh-4 double null mutants and found that in these animals the ADL neurons are also still generated normally, as assessed by intact DiI filling and characteristic neuronal nuclear speckles (Fig 2A).
The expression of the hlh-4 promoter fusion in hlh-4 mutants until the first larval stage permitted us to visualize the anatomy of the ADL neurons in the absence of hlh-4 gene function. While the cell body of ADL is normally positioned, we find that ADL axons and dendrites display severe morphological defects (Fig 2C). The sensory dendrites of ADL are often detached from the nose. Even when attached, the cilia of ADL often do not display their characteristic bifurcated ciliated endings. The axons of ADL, which in wild-type animals display a highly stereotyped extension and branching pattern, show pathfinding and branching defects (Fig 2C).
To examine whether and to what extent hlh-4 is required to specify ADL neuron identity, we examined the differentiation program of the ADL neurons in detail. The ADL nociceptive neuron pair coexpresses an unusually large number of olfactory-type GPCRs [44–46]. Reporter genes generated for about one fifth of the approximately 1,300 GPCR encoding reveal the expression of more than 60 GPCR genes from diverse families in ADL [46]. Extrapolating to the complete set of GPCRs encoded in the C. elegans genome, about 300 GPCR-encoding genes may be expressed in ADL. We asked whether hlh-4 is required for the expression of 12 GPCR-encoding genes. We chose these genes to cover the diverse set of GPCR gene families expressed in ADL (sra, sre, sri, srz, srh, srxa, and srx families). We found that expression of all of the tested 12 GPCR reporters is abrogated in hlh-4 mutants (Fig 3A). While all defects were routinely scored at the adult stage, we note that these defects are already apparent at the first larval stage. Consistent with the absence of expression of the hlh-4 paralog hlh-3 in postmitotic ADL neurons, we find that hlh-3 does not affect srh-127 expression in ADL.
To test whether hlh-4 does not only affect expression of chemoreceptor proteins but also affects the chemorepulsive function mediated by the ADL neurons, we considered its chemorepulsive function toward a specific nematode pheromone, the ascaroside ascr#3 (asc-ΔC9, C9)[24]. While wild-type hermaphrodites are repelled by ascr#3, this repulsion is significantly reduced in hlh-4 hermaphrodites (Fig 3B). This is not a reflection of an overall failure to engage in a nociceptive response because another chemorepulsive behavior, mediated by the ASH neurons (glycerol avoidance) [47], is not affected in hlh-4 mutants (Fig 3B).
We tested whether hlh-4 function is restricted to controlling olfactory receptor expression and function in the ADL neurons or whether other identity features of ADL are disrupted as well. A TRP channel protein encoded by the osm-9 gene, expressed in a restricted set of sensory neurons, is required in ADL to signal the response to distinct chemorepulsive sensory inputs [24,48,49]. We find that osm-9 expression is selectively lost in the ADL neurons of hlh-4 mutant animals (Fig 4).
Going beyond signal perception and transmission, we asked whether ADL requires hlh-4 to communicate with its synaptically connected neurons [50]. Based on the expression of the vesicular glutamate transporter eat-4/VGLUT, the key defining feature of all glutamatergic neurons, ADL neurons have previously inferred to be glutamatergic [51]. We find that the glutamatergic identity of ADL, as assessed by eat-4 fosmid reporter gene expression, is defective in hlh-4 mutant animals (Fig 4). Apart from using glutamate as a likely fast neurotransmitter, the expression patterns of various neuropeptide-encoding genes indicate that ADL also utilizes distinct peptides for neurotransmission [52,53]. We find that the expression of four neuropeptides, previously known to be expressed in ADL, as well as other neurons (FMRFamides flp-4 and flp-21 and neuropeptides nlp-7 and nlp-10) [52,53] specifically fail to be expressed in the ADL neurons of hlh-4 mutants, while expression in other neurons is unaffected (Fig 4).
Apart from peptidergic and chemical synaptic transmission, electrical synaptic transmission is likely also affected in hlh-4 mutants. ADL forms electrical synapses with a select number of neighboring neurons [50]. Electrical synapses are formed by transmembrane innexin proteins [54], and 3 of the 24 C. elegans innexin genes, unc-7, inx-18, and che-7, are expressed in ADL, as well as a specific set of other neuron types [55]. The expression of all three innexin genes is lost specifically in the ADL neurons of hlh-4 mutants (Fig 4).
Transmembrane ion channel expression is also affected in hlh-4 mutants. Na+/Ca2+-K+ exchangers are important regulators of intracellular calcium homeostasis in the nervous system, and members of this family show remarkably specific gene expression profiles in the C. elegans nervous system [56]. Two Na+/Ca2+-K+ exchangers, ncx-6 and ncx-7, are each exclusively expressed in the ADL neurons of wild-type animals [56]. The expression of both genes in ADL is abrogated in hlh-4 mutants (Fig 4).
To examine whether these defects are a consequence of the failure of solely maintaining the differentiated state versus failure of initiation of the differentiated state, we examined the expression of several ADL markers right after hlh-4 mutant embryos had hatched. Testing four specific markers (srh-127, sre-43, srt-47, and ncx-6), we found that expression is already affected at this early stage of development.
In conclusion, we find that several distinct identity features that define functional features of the ADL neuron are coregulated by the same transcription factor. The affected identity features share the common theme of providing the ADL with a unique molecular signature and identity. In contrast, hlh-4 does not affect generic neuronal features (i.e., pansensory or panneuronal features).
hlh-4 is not only required for the expression of ADL identity genes, but ectopic expression of hlh-4 is also sufficient to induce ADL identity features. We drew this conclusion by driving expression of hlh-4 in many other ciliated sensory neurons, using the arrd-4 promoter [57] (S1 Fig). The arrd-4prom::hlh-4 construct is not only able to rescue the loss of srh-127::gfp expression in ADL in hlh-4 mutants (Table 1), but these transgenic animals display ectopic expression of the normally ADL-expressed srh-127::gfp reporter in many ciliated sensory neurons (Fig 5A). Similarly, the TRP channel osm-9, the neuropeptide-encoding flp-4 gene and the vesicular glutamate transporter eat-4 also are ectopically expressed in other sensory neurons in these transgenic animals (Fig 5A).
To further probe the ability of hlh-4 to induce ADL identity features in other neurons, we misexpressed hlh-4 under control of a promoter fragment from the unc-3 locus, which is expressed in ventral cord motor neurons and a small set of head neurons (S1B Fig). Transgenic animals expressing a unc-3prom::hlh-4 construct show ectopic expression of the ADL marker srh-127::gfp in head neurons but not in ventral cord motor neurons (Fig 5B). The apparent cellular context dependency of hlh-4 function mimics the context dependence of other master regulators of cellular identity, such as Eyeless/Pax6 [58].
Because gene expression is usually examined in C. elegans via reporter gene constructs, a large library of reporter transgenes that monitors the expression of thousands of genes has been amassed by the C. elegans community over the past few decades. In many cases, expression patterns of these reporter transgenes have been defined on a single neuron level. Almost 200 reporter transgenes have been found to be expressed in the ADL neurons (www.wormbase.org, S2 Table). The genes tested above for their dependence on hlh-4 belong to this dataset. We took a subset of these genes (117) and asked whether 5′ upstream regulatory regions of genes whose expression is monitored by these reporter transgenes are enriched for the presence of a specific sequence motif using the FIRE motif analysis platform [26] (see Materials and methods). We restricted the search space to the first 2 kb upstream of these genes. As a control, we also considered several other neuron classes that Wormbase associated with a large number of reporter genes (AIY, ASE, ALM, HSN, ASI, ASK, ASH, PHA; www.wormbase.org) and interrogated the upstream regulatory control regions of those genes. In the ADL dataset, we indeed identified a motif found in 75% of the ADL-expressed reporter genes (Table 2, S1 Table; S2 Table). The motif, shown in Fig 6A, has a completely invariant 6-nucleotide core, CACCTG, and no striking sequence features outside this core. There is no orientation preference for this motif on the plus versus minus strand. This motif is not enriched in the control datasets (AIY, ASE, ALM, HSN, ASI, ASK, ASH, or PHA expressed reporter genes).
The CACCTG motif matches experimentally determined bHLH binding sites (CANNTG) [59] and specifically matches the in vitro binding site of the C. elegans HLH-4/HLH-2 heterodimer, CA(G/C)CTG [30]. Probabilistic segmentation analysis of upstream regulatory sequences of ADL neuron-expressed GPCR genes had previously also identified a similar CA(G/C)CTG motif [45].
All the 23 terminal effector genes that we described above as depending on hlh-4 in their expression in ADL (Fig 3; Fig 4) contain at least one copy of this motif within 2 kb upstream of the 5′ start of the gene (Table 2, S1 Table). The one hlh-4-dependent GPCR reporter (srh-79) that does not contain a perfect match to the E-box motif contains a 1-nucleotide-mismatched copy of the motif (CACGTG versus CACCTG).
The hlh-4 locus itself and, specifically, the 700-bp 5′ upstream regulatory region that shows hlh-4 autoregulation (Fig 1C) contains two copies of the perfectly matched CACCGT motif (both motifs are located in the 245-bp-long intergenic region). Moreover, the upstream region of the hlh-2/Da gene, the putative cofactor of hlh-4, which is also continuously expressed in ADL, also contains three copies of this motif in its 5′ upstream intergenic region. The regulation of hlh-2/Da expression by hlh-4 (demonstrated above) is therefore also likely a reflection of direct autoregulation of the hlh-2 locus by the HLH-4/HLH-2 heterodimer.
Three lines of evidence further validate the importance of the CACCGT E-box motif for ADL expression:
We used phylogenetic footprinting in the TargetOrtho pipeline [27] to assess the extent of conservation of the CACCTG motif among five Caenorhabditis species, C. elegans, C. briggsae, C. remanei, C. brenneri, and C. japonica (S2 Table). This analysis provided a genome-wide assessment of the location of the CACCTG motif in these five different species and allowed us to define a number of features of the CACCTG motif:
Moreover, we find that two of the ADL-expressed genes that do not contain a perfect match to the CACCTG motif (srh-79 and srh-186, one of which, srh-79, we confirmed to be hlh-4-dependent) contain a motif with a single mismatch to the CACCTG motif (CACGTG), yet all Caenorhabditis species that have orthologues of these two genes contain perfect CACCTG motif matches (Table 2, S1 Table).
In conclusion, a CACCTG motif defines a signature for ADL-expressed genes. Given that this motif is a known in vitro binding site for a HLH-4/HLH-2 dimer [30], hlh-4 appears the most likely candidate to directly activate the expression of scores of genes that uniquely and combinatorially define the terminally differentiated state of the ADL neuron pair.
The partially penetrant effect of hlh-4 on eat-4/VGLUT expression suggested that hlh-4 partly relies on additional factors to control eat-4/VGLUT expression. This notion is further corroborated through an examination of the cis-Regulatory control regions of the eat-4/VGLUT locus. We find that 4.5 kb of sequence upstream of the eat-4/VGLUT locus directs reporter gene expression to many glutamatergic neurons, including ADL (prom6-1; Fig 7A). This 4.5-kb region contains a phylogenetically conserved CACCTG motif 691 bp upstream of the ATG. Deletion of this motif results in loss of expression in ADL (Fig 7A). However, while this motif is required for ADL expression, it is apparently not sufficient: deleting 3.2 kb from the 4.5-kb 5′ reporter fusion leaves the E-box unaffected but abolishes expression in ADL (prom2; Fig 7A), suggesting that these deleted sequences contain binding site(s) for a transcription factor that cooperates with hlh-4 to activate eat-4/VGLUT expression.
The LIM homeobox gene lin-11 was previously shown to be expressed in postmitotic ADL neurons throughout their lifetime [61]. We find that lin-11 expression in ADL is not affected in hlh-4 mutants (Fig 7B). Corroborating a role of lin-11 in parallel to hlh-4, we find that lin-11 null mutants are defective in the ADL-mediated chemorepulsive response to C9 ascaroside (Fig 7C). Consistent with this behavioral defect, we observed that lin-11 null mutants display defects in the expression of several of hlh-4-dependent and E-box-containing genes, including ncx-6, srh-234, and flp-21 (Fig 7D). However, lin-11 does not affect the hlh-4-dependent flp-4 gene, nor does it affect eat-4/VGLUT fosmid reporter expression (Fig 7D).
We tested whether a function for lin-11 on eat-4/VGLUT expression could be revealed in the context of an hlh-4 mutant background, in which eat-4/VGLUT fosmid reporter expression is only partially affected. lin-11; hlh-4 double mutants still normally express pansensory markers in ADL, but they display a dye filling defect that neither mutant alone displays, corroborating the parallel nature by which hlh-4 and lin-11 affect ADL differentiation (Fig 7E). Surprisingly, in hlh-4; lin-11 double null mutants, the partially penetrant loss of eat-4/VGLUT expression observed in hlh-4 single mutants was not enhanced but instead completely suppressed (Fig 7D). The same effect is observed on the flp-4 gene. Its completely penetrant loss in hlh-4 mutants is suppressed in hlh-4; lin-11 double mutants (Fig 7D).
The reinstatement of eat-4/VGLUT fosmid expression even in the absence of hlh-4 is mirrored by a mutation in the cis-Regulatory control region of eat-4/VGLUT. The 1.2-kb upstream region of eat-4/VGLUT, which contains an hlh-4 binding site but is not expressed in ADL, becomes expressed in ADL upon deletion of a predicted homeodomain binding site, a potential recognition motif for LIN-11 (Fig 7A). This result suggests that eat-4/VGLUT expression is controlled via a collaboration of hlh-4 with an as yet unknown transcription factor X whose activating effect is normally antagonized by LIN-11. If all activators (hlh-4 and X) are present, lin-11 cannot prevent activation of eat-4/VGLUT (eat-4prom6-1delta12); hence, eat-4/VGLUT is expressed in ADL. If, however, the system is partially destabilized by hlh-4 removal (or by removal of the E-box sequence in the reporter construct), lin-11 can counteract the ability of factor X to activate eat-4/VGLUT expression (eat-4prom2delta 12) (as assessed by the restoration of eat-4 expression upon removal of lin-11). The effect of lin-11 on ADL-expressed genes is, however, clearly target gene dependent. While in the case of one target gene, eat-4/VGLUT, lin-11 appears to antagonize hlh-4 function, it may positively cooperate with hlh-4 on those other target genes whose expression is either completely or partially lost in hlh-4 and/or lin-11 mutants. We conclude that hlh-4 is a central regulator of ADL identity that may interact in a target gene–dependent manner with distinct collaborating factors.
The identification of proneural genes that act very early in neuronal development to allocate neuroblast identity to distinct neuronal lineages via classic genetic loss of function analysis in Drosophila represents one of the classic landmark achievements of developmental neurogenetics [2,3]. The subsequent cloning of vertebrate AS-C and Atonal homologs has revealed the deep conservation of this fundamental neural patterning mechanism [1,6–8]. We have described here a novel functional property of an AS-C gene, demonstrating that C. elegans hlh-4 joins the rank of terminal selector-type transcription factors that act in postmitotic neuron classes to initiate and maintain the differentiated state of a specific, postmitotic neuron class. hlh-4 displays all the hallmarks of a terminal selector [62,63]: it is required for initiation of the terminal differentiation program of the ADL neuron pair, it is continually expressed throughout the life of the neuron (suggesting that it also maintains neuronal identity), this continuous expression is mediated by direct autoregulation via HLH-2/HLH-4 binding sites in the hlh-2 and hlh-4 loci, and, most importantly, hlh-4 controls the vast majority of neuron class–specific genes whose combinatorial coexpression defines ADL identity, yet it does not control generic neuronal features (panneuronal and pansensory features). Hence, exactly like other terminal selectors [62,63], hlh-4 separates the adoption of neuron type–specific features (hlh-4-dependent) from the acquisition of an overall, panneuronal/pansensory identity (hlh-4-independent) (Fig 8A). It is important to precisely appreciate this fundamental dichotomy in neuronal gene expression programs, repeatedly observed in many different neuron classes and corroborated here by the hlh-4 mutant phenotype: as schematized in Fig 8A, genes that are expressed in specific subsets of neuron classes are terminal selector dependent, while genes that are expressed in a non-neuron-class–specific manner are regulated by independent means [60].
The terminal selector function of hlh-4 is likely exerted in collaboration with the canonical AS-C cofactor, hlh-2/Da, which shares with hlh-4 the unusual feature of postmitotic expression throughout the life of the ADL neuron class. hlh-2 is also continuously expressed in a small number of additional neuron classes, but its function in these neurons remains unknown. In yeast one-hybrid assays, HLH-4/HLH-2 has been shown to bind to the CACCTG sequence that we describe here [30]. While the HLH-4/HLH2 complex and its cognate binding site is essential—and at least in some context also sufficient—for gene expression in ADL, it is unlikely to act on its own. With its 6-bp length, the recognition element of the HLH-4/HLH-2 heterodimer occurs too frequently in the genome to direct HLH-2/HLH-4 exclusively to ADL-expressed genes. We find that the LIM homeobox gene lin-11 assists hlh-4 in the regulation of some but not all hlh-4-dependent target genes. As no DNA cis-Regulatory motif was found to be significantly enriched in ADL-expressed genes by our bioinformatic analysis in addition to the E-box, we propose that hlh-4 is a central core inducer of all ADL-specific genes but may be assisted in its function, i.e., provided the proper specificity, by interaction with a suite of distinct, target gene–dependent collaborating factors, such as lin-11 and perhaps other, as yet to be discovered factors (Fig 8B).
Previous work on AS-C genes in worms has revealed that the AS-C-type hlh-14 gene acts as a conventional proneural gene during early embryonic patterning to specify the neuronal identity of an AB-blastomere-derived lineage branch that produces several sensory neurons, including ADL [12]. In the absence of hlh-14, cells in this lineage branch convert to a hypodermal identity [12] (Fig 8C). Hence, the ADL neuron depends on the successive activity of two distinct AS-C-type genes, one acting as a conventional proneural gene (hlh-14), followed by hlh-4, which acts in a subbranch of this lineage, to specify terminal ADL identity (Fig 8C). Whether hlh-14 directly activates hlh-4 expression is presently unclear. Notably, though, the E-box motif in the hlh-4 locus that is required for maintaining hlh-4 expression is not required for initiation of hlh-4 expression in the embryo.
Even though a proneural function of AS-C-type genes is clearly a deeply conserved function of bHLH genes, our findings prompt the intriguing question as to whether a function of bHLH genes in directly controlling the differentiated state of a neuron may have been an even more ancestral function of AS-C-type bHLH genes. In support of such notion, the AS-C ortholog in the cnidarian Hydra magnipapillata, Cnash, was previously reported to not be expressed in neuronal precursors but rather in differentiating and mature neurons, leading the authors of that report to postulate a role of hydra Cnash in initiating and maintaining the neuronal phenotype [64], exactly as we propose here for C. elegans hlh-4. Loss of function studies of the AS-C orthology NvashA of the sea anemone Nematostella vectensis cannot distinguish between a proneural versus terminal differentiation role [65].
Subsequent to such terminal differentiation role, an “ur-” bHLH may then have become co-opted into more upstream regulatory events in proliferating blast cells. A somewhat similar trajectory has been proposed for the Pax6/Eyeless gene, originating with a function in regulating lens protein to subsequent recruitment to earlier steps of eye development [66]. Of course, it is also conceivable that the terminal selector function of hlh-4 may be a derived feature, one that perhaps came into existence via the acquisition of an E-box motif in the hlh-4 locus that lead to hlh-4 expression being “locked” into a terminal and continuous function. More detailed expression pattern analysis of AS-C and Ato-like genes in the adult nervous system of other species will provide hints whether hlh-4-like, terminal selector functions may also be carried by AS-C/Atonal genes in other organisms. In fact, such function may be conceivable in an already previously reported case. Drosophila Atonal is expressed in mature dorsal cluster neurons in the dorsolateral CNS of the flies [67]. In these neurons, Ato has no proneural function but instead serves to control arborization patterns. However, whether Ato has an impact as broad as hlh-4 on controlling the differentiated state of these neurons is not yet known.
C. elegans sox-2/SoxB1 is another gene whose orthologs in other organisms (SoxB factors) act in early neuronal patterning [68] but that has become employed as a terminal selector in C. elegans [69,70]. Here again, the question is whether such late role is a reflection of an ancestral or derived function of this gene. It is important to keep in mind that the existence of such late functions (in addition to the well-characterized early functions) may have very easily escaped detection in other organisms, because straight knockout approaches will only reveal the early function of a gene in the lineage. Only if an early function is not existent, as apparently is the case for sox-2 and hlh-4, will a late function be revealed with relative ease using standard genetic loss of function, i.e., straight knockout approaches (this paper) [69,70].
Defining hlh-4 as a terminal selector of ADL identity sheds additional mechanistic context on previous studies about the feeding state–dependent regulation of a sensory-type GPCR gene, srh-234, in the ADL neuron [35,71]. Focusing on this specific gene, the authors found that the MEF-2 transcription factor, a well-known mediator of neuron activity–dependent processes in many different organisms [72], down-regulates hlh-4-dependent srh-234 expression under starvation conditions. This effect is mediated via a MEF-2 binding site in the srh-234 locus that is located next to the HLH-4/HLH-2 binding E-box [35]. Together with our description of a broad role of hlh-4 in controlling the differentiated state of ADL, an intersectional strategy of a “genetically hardwired” identity factor with a condition-dependent factor becomes apparent. Such an intersectional strategy could perhaps be a general strategy to explain the cellular specificity of broadly acting signals that convey environmental or physiological information.
One of the remarkable features of the chemosensory system of C. elegans is the coexpression of multiple sensory receptors of the GPCR family in individual neuron types [44–46]. Even though the expression of only about one fifth of C. elegans chemosensory-type GPCRs has been examined so far [46], there are several chemosensory neurons that coexpress several dozens of GPCRs. This tremendous extent of coexpression only applies to a select set of chemosensory neurons, with the most prominent set being the nociceptive ADL, ASH, PHA, and PHB neurons [46]. One could have imagined several scenarios by which such coexpression is controlled. A previous bioinformatic analysis already strongly hinted toward coregulation of coexpressed GPCRs via a common cis-Regulatory motif [45]. However, it is only through the present analysis that we can conclude that a single trans-acting factor instructs, apparently via direct binding to a cis-Regulatory element shared by most if not all coexpressed GPCRs, the enormously broad spectrum of chemosensory capacities of one of these nociceptive neurons, ADL.
|
10.1371/journal.ppat.1003987 | The Effect of Cell Growth Phase on the Regulatory Cross-Talk between Flagellar and Spi1 Virulence Gene Expression | The flagellar regulon controls Salmonella biofilm formation, virulence gene expression and the production of the major surface antigen present on the cell surface: flagellin. At the top of a flagellar regulatory hierarchy is the master operon, flhDC, which encodes the FlhD4C2 transcriptional complex required for the expression of flagellar, chemotaxis and Salmonella pathogenicity island 1 (Spi1) genes. Of six potential transcriptional start-sites within the flhDC promoter region, only two, P1flhDC and P5flhDC, were functional in a wild-type background, while P6flhDC was functional in the absence of CRP. These promoters are transcribed differentially to control either flagellar or Spi1 virulent gene expression at different stages of cell growth. Transcription from P1flhDC initiates flagellar assembly and a negative autoregulatory loop through FlhD4C2-dependent transcription of the rflM gene, which encodes a repressor of flhDC transcription. Transcription from P1flhDC also initiates transcription of the Spi1 regulatory gene, hilD, whose product, in addition to activating Spi1 genes, also activates transcription of the flhDC P5 promoter later in the cell growth phase. The regulators of flhDC transcription (RcsB, LrhA, RflM, HilD, SlyA and RtsB) also exert their control at different stages of the cell growth phase and are also subjected to cell growth phase control. This dynamic of flhDC transcription separates the roles of FlhD4C2 transcriptional activation into an early cell growth phase role for flagellar production from a late cell growth phase role in virulence gene expression.
| Flagellar-mediated motility is fundamental to Salmonella pathogenesis, which takes the lives of hundreds of thousands of people each year. The genes of the Salmonella pathogenicity island 1 and those of the flagellar regulon are part of the same transcriptional hierarchy. We report the novel finding where the key control of this network takes place at the flhDC promoter region. We followed the transcription from the two “live” flhDC promoters as a function of the cell growth phase. P1 comes on early in the cell cycle, while P5 comes on late. Transcription of P5 is HilD dependent, which represents a totally new finding and Salmonella specific: there is no HilD in E. coli flhDC control, no P5 transcription. P1 & P5 can express flhDC to equivalent levels, yet only P1- dependent expression produces motility UNLESS we artificially induce P5 EARLY in the cell cycle. This work is the foundation for the cell cycle stages a Salmonella bacterium experiences during host infection. This is a significant conceptual advance in Salmonella pathogenesis: one can no longer consider gene regulation at 37°C and OD 0.6 as a reflection of the Salmonella infection cycle; the whole cell growth cycle must be considered in understanding this complex biological processes.
| Tens of millions of human cases of Salmonellosis, a foodborne gastroenteritis caused by Salmonella enterica, occur worldwide every year killing more than a hundred thousand people annually (World Health Organization Fact sheet N°139, August 2013). Typhoid fever caused by Salmonella Typhi kills an equivalent number of people each year. A prominent player in Salmonella pathogenesis is the bacterial flagellum. The bacterial flagellum is an ion-powered, complex motor organelle that endows bacterial cells, such as Escherichia coli and Salmonella enterica, with the ability to propel themselves through liquid medium and across hydrated surfaces [1]. Motility also plays an important role in biofilm formation and in the ability of many pathogens to reach their sites of infection and establish disease [2], [3].
Early work on the discovery of Salmonella virulence genes identified a transposon insertion in the flagellar filament cap gene, fliD, as defective for survival of cells in macrophages [4]. However, fliD is in an operon with the fliT gene whose product is a regulator of the flagellar and Spi1 virulence genes master regulatory complex FlhD4C2 [5], [6]. The transposon insertion in fliD was polar on fliT gene expression and thus identified regulation of FlhD4C2 activity as critical for Salmonella virulence. The two proteins that make up the FlhD4C2 transcriptional regulatory complex are co-expressed from the flhDC operon, class 1 promoter, which is at the top of a complex transcriptional hierarchy for both flagellar and Spi1 virulence genes expression. The decision whether or not to produce flagella is regulated at the levels of flhDC transcription, translation, FlhD4C2 assembly and stability [7]. Positive regulators of flhDC operon transcription include cAMP-CRP, Fis, Fur, H-NS and QseB [8]–[14]. A large number of regulatory factors are also reported to inhibit flhDC transcription. These factors include, LrhA, RcsB, RtsB, SlyA, DskA, PefI-SrgD, FimZ, HdfR, OmpR and RflM [15]–[20]. The FlhD4C2 activity generates an auto-regulatory loop by activating transcription of the rflM gene encoding a LysR-type DNA binding protein RflM, which in turn inhibits the transcription of flhDC [21]. The post-transcriptional factors regulating flhDC include, CsrA [22], [23], Hsp70 chaperone DnaK [24] and ClpXP protease [25]. Recently an FlhD4C2 repressed gene, ydiV [26], was shown to code for a protein (YdiV) that will bind to FlhD4C2, in its free or DNA-bound form, remove FlhD4C2 from DNA and serves as an adapter that targets FlhD4C2 for ClpXP-dependent degradation [27], [28].
In Salmonella, an initial characterization of the flhDC promoter region identified six transcriptional start sites (TSSs) [13]. In a recent study, only four of the original six TSSs were detected [29]. The presence of six TSSs in the Salmonella flhDC regulatory region combined with the presence of DNA binding sites of CRP, LrhA, RtsB, HilD, RcsB, HNS and others indicated a complex level of the flhDC transcriptional regulation.
Salmonella enterica is an intracellular facultative pathogen causing a range of diseases in a variety of hosts [30]. Important virulence factors required for Salmonella invasion of epithelial cells and development of Salmonellosis are encoded within the Salmonella pathogenicity island 1 (Spi1) genes. Spi1 encodes a virulence-associated type III secretion system (T3SS) as part of an injectisome structure required for the secretion and injection of multiple effector proteins into the cytoplasm of host cells [31]–[36]. Expression of Spi1 genes is controlled in response to specific combinations of environmental signals in a complex hierarchical process with multiple transcriptional regulators. These include, HilA, a member of the OmpR/ToxR family of transcriptional regulators, which promotes transcription of genes encoding the necessary components for a functional Spi1 injectisome system [32], [35], [37], [38]. Also included are the hilC and hilD genes whose products are members of the Ara/XylS family of transcriptional regulators that control hilA gene transcription. HilD is at the top of the regulatory network controlling Spi1 expression because most regulators controlling hilA transcription appears to be HilD-dependent [39], [40].
It is noteworthy to mention that many protein components of the Spi1 and flagella T3SS exhibit a significant degree of amino acid identity, leading to the production of remarkably similar T3SS structures [16], [33], [34], [41], [42]. Furthermore, many of the transcriptional and posttranslational regulatory factors of flhDC also target the main transcriptional regulators of Spi1, such as HilA and HilD [11], [43]–[52]. In addition, the ATP-dependent Lon protease was shown to degrade both FlhD4C2 and HilD [24], [25]. Coordinated expression of Spi1 and flagellar genes has been recently demonstrated [53]. In Salmonella, expression of Spi1 genes is activated by FliZ [54]–[57], which is encoded within the flagellar fliAZY operon. FilZ activates the hilD gene expression at the posttranslational level and HilD in turn promotes transcription of the rtsAB operon, which encodes a pathogenesis-related DNA-binding regulatory proteins. RtsA and RtsB reciprocally regulate both the Spi1 and flagellar genes [17]. The direct binding of RtsB to the flhDC promoter region inhibits flhDC transcription and motility [17].
We decided to investigate how input from different regulatory factors might integrate multiple environmental or cell cycle signals into the control of flhDC expression in Salmonella enterica. We explored how and when positive and negative regulators affect flhDC expression throughout the cell growth cycle. We measured the effect of RcsB, LrhA, RflM, SlyA, RtsB and HilD regulatory factors on flhDC operon transcription at different cell growth phases. We characterized the specific TSSs within the flhDC promoter region and their involvement in the positive and negative control of flhDC cell-cycle dependent transcription. Finally, we examined how the individual TSSs and protein regulatory factors controlled the interconnection between the flagellar and Spi1 regulons.
To investigate flhDC operon transcription at different phases of the cell growth, we constructed a transcriptional fusion of the flhDC promoter region to the luciferase operon of Photorhabdus luminescence (luxCDBAE operon). Because the flhDC operon is autoregulated negatively by RflM and positively by HilD, we designed strains harboring an intact copy of the flhDC operon under the control of its native promoter (PflhDC) and an in-frame fusion of a second copy of the promoter region of flhDC (through the first 272 nucleotides of flhD coding sequence) to the luciferase operon: DUP[(PwtflhDC-luxCDBAE)*Km*(PwtflhDC-flhD+C+)] (Figure 1A). Thus, individual PflhDC promoter regions transcribe both the luminescence operon reporter and the flhDC operon. This results in a strain with luminescence readout for the level of transcriptional activation of flhDC under conditions that also preserves the wild-type expression of the flagellar regulon including flhDC autoregulation through FlhD4C2-dependent expression of rflM and hilD genes. For simplicity, we will refer to the DUP[(PwtflhDC-luxCDBAE)*Km*(PwtflhDC-flhD+C+)] construct as PwtflhDC.
Following batch inoculation of an overnight culture of the PwtflhDC strain into fresh media with shaking at 30°C, transcription of the flhDC genes declined 4-fold during the initial lag phase transition to log phase growth to a minimal value (Figure 1B). This observation is consistent with that reported in an earlier study [11]. After the transition to log phase growth, transcription of flhDC increased more than 10-fold between OD 0.3 and 1.2, followed by a decline in flhDC transcription as cells enter late log and stationary phase growth (Figure 1B).
In Salmonella enterica, flagellar regulon transcription is highest during the exponential phase of growth and decays in late stationary phase [58]. Transcription of the flagellar master regulatory operon, flhDC, is under both negative and positive control by multiple regulatory factors. Null mutations in any one of the rcsB, rflM, lrhA, slyA, and rtsB genes result in increased transcription of the flhDC operon, which is consistent with an inhibitory activity on flhDC expression. HilD is an activator of flhDC transcription such that over-expression of the hilD gene increases flhDC expression (Singer et al. submitted). The diversity of transcription factors controlling expression of flhDC reflects the complexity of flhDC transcriptional regulation and suggests that flhDC transcription is controlled when Salmonella cells are experiencing different metabolic or environmental states, or different growth conditions under which these transcriptional factors are active. We examined both the timing and magnitude of individual regulatory proteins on flhDC transcriptional control throughout the cell's growth phase. We tested flhDC transcriptional levels as a function of the cell's growth phase in strains missing the individual negative regulators RcsB, LrhA, RflM, RtsB, SlyA and the positive regulator HilD (Figures 1C & D). As was presented above for the wild-type strain, this was done by growing PwtflhDC cells in liquid culture at 30°C using luciferase as the reporter for flhDC transcription levels. Luciferase levels were determined at specific optical densities shown in Figure 1. As expected, removal of individual inhibitors resulted in an increase in flhDC transcription levels while removal of HilD decreased flhDC transcription. Importantly, our assay revealed a growth phase-dependent hierarchy of the effect of these regulators. At OD 0.3, basal flhDC transcription was elevated in the absence of LrhA and RcsB, while removal of RflM, RtsB, SlyA or HilD exhibited the same basal level of transcription as wild type (Figures 1C & 1D). This suggests that RcsB and LrhA act earlier, during lag phase, to inhibit flhDC transcription. This effect could also represent a carry-over of repression from stationary phase that keep flhDC transcription low during the transition to log growth. In the absence of RflM we observed an earlier transition to flhDC activation than in the other mutant strains. Since FlhD4C2 transcribes the rflM gene and RflM protein inhibits flhDC transcription (flhDC auto-inhibition), this result suggests that flhDC auto-inhibition through RflM occurs during early exponential phase to control when full FlhD4C2-dependent gene expression occurs at log phase. The negative effect of RtsB and SlyA on flhDC transcription was detected as cells enter early stationary phase. We also observed that the maximum flhDC transcription level peaked earlier for both the hilD and rflM mutants at OD 1, while the wild type and mutants in rcsB, lrhA, slyA and rtsB peaked around OD 1.2.
The data presented in Figure 1C demonstrate that initial flhDC transcription is kept low by a combination of repressors including at least RcsB and LrhA. Initial FlhD4C2 expression during the stationary to log phase transition produces enough RflM to maintain a low level of flhDC transcription until an OD of ∼0.3 is reached. After OD 0.3, flhDC transcription increased significantly, but RflM, RcsB and LrhA reduce the overall level. Interestingly, the wild-type level is balanced by the presence of the HilD activator of flhDC transcription, the hilD-activated inhibitor of flhDC transcription RtsB and by the virulence associated factor SlyA (Figure 1D).
In order to obtain more detailed information relating the effect of specific regulatory proteins on flhDC transcription as a function of the cell's growth phase, we determined luciferase levels for the PwtflhDC grown in liquid culture at 30°C in 96 well plates with a microplate reader. Using this assay system, we could measure the activity of flhDC transcription at shorter times intervals (6 min) with continuous shaking at 150 rpm. We observed the same trend of regulation of the flhDC operon as seen in batch cultures for lrhA, rcsB (Figure 2A), rflM (Figure 2B), slyA and rtsB (Figure 2C), and hilD mutants (Figure 2D). However, the pattern observed in 96 well plates was somewhat different compared to the batch growth. We observed that activation of flhDC transcription took place earlier at OD∼0.2 rather than OD∼0.3. Consistent with this observation, the differences between the activity of flhDC in wild-type versus mutant strains also occurred at an earlier OD measurement in microtiter plate growth compared to growth in batch culture. The cells in 96 well plates reached maximum expression at OD∼0.6 compared to OD∼1.2 in the batch culture. We attribute these differences to the mode of growth in 96 well plates (150 rpm) where bacterial cells are grown in much lower volumes and likely to be subjected to different oxygen levels in the medium compared to batch cultures. It has been shown that activation of flgA, a gene under the control of flhDC, under static conditions (no shaking of 96 well plates) occurred immediately after dilution of an overnight culture into LB-1% Salt [53]. When we tested the activation of flhDC operon in standing batch culture in LB, we observed that flhDC transcription increased at OD∼0.12 (Figure S1), which is earlier compared to what we observed either in batch shaking (OD∼0.3) or 96 well grown cultures (OD∼0.2). Moreover, the shutdown of flhDC transcription observed in standing cultures took place after cells reach an OD∼0.6 compared to shaking batch culture where the shutdown started at an OD∼1.2.
Because flhDC transcription is differentially regulated by different transcription factors in a growth phase dependent manner, we hypothesized that the effect of each of these regulators is observed at the time when they are produced during the cell growth cycle. To investigate this possibility we placed the luxCDBAE operon reporter under control of the promoters of the six regulatory genes lrhA, rcsB, rflM, slyA, rtsB and hilD, whose products have been demonstrated to bind directly to the flhDC promoter region and monitored their expression profile at different optical densities (binding of RflM or SlyA to the flhDC promoter region has not been reported). We monitored the activities of these constructs in 96 well plates over time. We observed that the transcription of the autoregulated gene lrhA is immediately activated following dilution from an overnight culture, and before the activation of flhDC (Figure S2.A). Transcription of rcsD (which is the first gene of the rcsDB operon transcribed from the rcsD promoter) also initiated before flhDC (Figure S2.B), whereas transcription of rflM overlapped with that of flhDC (Figure S2.B). Since rflM transcription is dependent on FlhD4C2, these results suggest that low basal levels of FlhD4C2 are sufficient to promote rflM gene transcription. In addition, transcription of rflM reached a maximum at OD∼0.35 and decayed very quickly (Figure S2.B) compared to the rest of the regulators tested in this study. The transcription of hilD gene is under positive autoregulatory control by HilD itself [59] and by HilD-activated RtsA [17]. In addition, the product of an flhDC activated gene, FliZ controls HilD at a posttranslational level [54], [57]. We observed that transcription of hilD increased at OD of ∼0.4 (Figure S2.C), at the same time when HilD promoted transcription of flhDC (Figure 2D). The expression of the HilD-activated rtsA gene (the first gene of the rtsAB operon) appeared to be activated at the same time as hilD (Figure S2.C). Transcription of the slyA gene was activated just after flhDC transcription started and before initiation of hilD and rtsA transcription, with a peak of expression at entry into stationary phase (Figure S2.D). These results suggest that there is a hierarchy of transcription of the factors regulating flhDC transcription that mirrors their effect on the transcriptional regulation of the flhDC operon.
We next asked if the protein levels of the regulatory factors controlling flhDC transcription were also growth phase dependent. We performed Western blot analysis of whole cell lysates of Salmonella strains (LrhA-HA, RcsB-3×Flag, RflM-HA, SlyA-HA, RtsB-HA and HilD-Flag) at different optical densities (Figure 3A). We established that LrhA is present at an early time point during cell growth (OD∼0.2) with maximum expression at OD∼0.6 followed by a decay at late stationary phase (note that both the N-terminal and C-terminal HA-tag fusion to LrhA are made but not functional and therefore there is no positive feedback regulation of lrhA transcription by LrhA protein [18]). The level of RcsB protein, the transcriptional regulator of the phosphorelay system RcsDBC, also appeared to be growth phase dependent because RcsB protein was detected early in the growth phase (OD∼0.2) and increased at the stationary phase of cell growth. The FlhD4C2 activated RflM, was produced early in the growth phase (OD∼0.2), and increased at OD∼0.4 followed by a quick decay during the rest of the cell's growth phase. HilD protein, the positive activator of flhDC transcription, was detected at OD∼0.4 and increased at stationary phase (Figure 3A). RtsB, whose gene is under the transcriptional control by HilD, was not detected early in the growth phase and was present at OD∼1.3. The absence of RtsB at an early time point in the blot might be due to the detection limits for low protein levels in our experiment (See CHIP, Figure 3B, where RtsB was already associated with the promoter of flhDC at OD∼1). In contrast, the negative regulator SlyA was produced during all the phases of cell growth, with a sharp increase at OD∼1. These results demonstrate a hierarchy at the level of expression of flhDC regulators that specifically mimics the differential dynamics of flhDC operon transcription.
We examined the in vivo binding dynamics within the flhDC promoter region by these regulatory factors. At different optical densities (0.4 to 1.4), chromatin immunoprecipitations (ChIP) were conducted using strains with individually tagged transcriptional factors, RcsB, RflM, HilD, RtsB, LrhA and SlyA (Figure 3B). Expression of RcsB and binding of RscB to its target DNA at the flhDC promoter was detected throughout the entire growth phase. However, RcsB bound levels increased as cells progressed to exponential phase (OD 0.4 to 0.6) followed by decreased binding at latter stages of growth. The transcriptional regulator RflM binding to DNA was detected at OD∼0.4 with maximal binding at OD∼0.6, but was no longer bound the flhDC promoter region beyond OD∼0.6. HilD, a transcriptional activator of flhDC, was bound to the flhDC promoter region at OD∼0.4 increasing to a maximum bound level at OD∼0.8 and followed by absence of bound HilD at OD∼1. SlyA was not physically associated with the flhDC promoter at OD∼0.4 and ∼0.6, but was bound to the flhDC promoter region at OD∼0.8. There was no binding of RtsB to the flhDC promoter at an early time point of cell growth OD∼0.4 to 0.6. Binding by RtsB had initiated by OD 0.8 and increased through OD 1.4. We were unable to immunoprecipitate LrhA tagged protein because C-terminal or N-terminal tagged LrhA behaved like lrhA null mutant (Figure S5; ). These results highlight the binding dynamics of different regulators to the flhDC promoter region resulting in a dynamic of flhDC operon transcription.
Six transcriptional start sites, designated P1flhDC, P2flhDC, P3flhDC, P4flhDC, P5flhDC and P6flhDC, within Salmonella flhDC promoter region were obtained by primer extension [13]. However only P1flhDC, P3flhDC, P4flhDC and P5flhDC were detected by RNA-Seq based approach [29]. Each of these TSSs was preceded by a hexamer motif (−10 box) with the consensus invariant residues adenine at position 2 (A2) and thymine at position 6 (T6), except for P4 (Figure 4A). To investigate the authenticity of these TSSs, we made alterations of the −10 sequences targeting the conserved residues A2 and T6 by changing them to a cytosine residue (C) and also by totally changing the −10 box to a GTTGGT sequence (Figure 4B). As controls, additional mutations were made in each −10 box, in a nucleotide other than A2 or T6 (Figures 4B & S3A) that supposedly should not alter significantly the effect of RNAP on transcription [60]. Because flhDC is subjected to negative and positive transcriptional feedback, mutations of the promoters responsible for transcription of flhDC operon in the wild-type strain might affect the positive and negative auto-regulation of flhDC transcription. We thus monitored the activities of the promoters mutants fused to luciferase operon in an flhD+C+ background (described above). Mutations of wild-type sequence P1flhDC (TATAGT) to GTTGGT (P1−.1flhDC); TCTAGC (P1−.2flhDC) or TCTAGT (P1−.3flhDC) but not TACAGT (P1−.4flhDC) were associated with a significant reduction of flhDC transcription (Figure 4C). Mutations of the wild-type P5flhDC (TATGCT) to TCTGCC (P5−.2flhDC) or TCTGCT (P5−.3flhDC) but not to TACGCT (P5−.4flhDC) reduced significantly the transcription of flhDC to the same extent as the mutation of −10 to GTTGGT (P5−.1flhDC) (Figure 4D). These results indicated that P1flhDC and P5flhDC are bona-fide promoters.
Analysis of mutations of −10 sequences of the P2flhDC and P6flhDC (overlapping with the CRP binding site which is required for the transcription of flhDC from P1flhDC promoter) and P3flhDC (overlapping with the LrhA binding site) and P4flhDC were not conclusive (Supplementary Text S1 & Figure S3).
We further investigated the authenticity of the six putative TSSs of the flhDC operon, by engineering strains with combined mutations in the promoter region of flhDC leaving only one wild-type −10 sequence from the six described promoters. Thus, P1+ designates a strain that has only a functional P1 promoter, etc. We also constructed a control strain with combined mutations in all the six promoters, AP−flhDC (All Promoters mutated). We established that P1+flhDC and P5+flhDC were able to promote flhDC operon transcription but to a lesser extent to what is observed in the wild-type strain (Figure 4E). The transcription of flhDC was totally abolished in strains harboring P2+flhDC, P3+flhDC, P6+flhDC and APs−flhDC, while P4+flhDC mutants showed very low level of flhDC transcription (1.8% relative to the wild-type strain) (Figure 4E). These results suggested that in the wild-type background P1flhDC and P5flhDC are the main promoters driving flhDC operon transcription with a marginal activity from the P4flhDC promoter. Yanagihara et al., 1999; have demonstrated that P6flhDC is only active in the absence of CRP, we confirmed that P6+flhDC (only P6 is functional) is inhibited by CRP, because in a crp null mutant there was an increase of transcription of P6+flhDC compared to wild-type (Figure S3H).
Since only mutations in P1flhDC and P5flhDC promoters significantly affected the expression of flhDC, we would expect the level of transcription of flhDC operon in the absence of both P1 and P5 promoters to be similar to the level of transcription of flhDC operon in the absence of all flhDC promoters (P1 through P6). To investigate this hypothesis, we measured the luciferase activity in a strain with combined mutations in P1flhDC and P5flhDC promoters (P1−P5−flhDC) and compared it to the luciferase activity of a wild-type strain and to a strain with all six promoters mutated (strain AP's). We observed that transcription of flhDC operon in strain P1−P5−flhDC was totally abolished to the same levels observed in a strain with all flhDC promoters mutated (Figure 4F). These results demonstrated that in a wild-type background P1flhDC and P5flhDC are the major promoters driving transcription of flhDC operon. We concluded that transcription of the flhDC operon in strain P1−flhDC (harboring mutations of the −10 box of P1flhDC) is driven from the P5flhDC promoter and that transcription of the flhDC operon in strain P5−flhDC (harboring mutations of −10 box of P5flhDC) is driven from P1flhDC.
Once we established that P1flhDC and P5flhDC are the main promoters driving transcription of the flhDC operon, we monitored the expression of the P1flhDC and P5flhDC promoters at different optical densities using PwtflhDC, P1−flhDC and P5−flhDC constructs (Figure 5A). The transcription profile of flhDC operon in strains P1−flhDC (P5-expressed) and P5−flhDC (P1-expressed) demonstrated that both promoters are required for transcription of flhDC because the expression of flhDC operon in constructs P1−flhDC (P5-expressed) and P5−flhDC (P1-expressed) did not reach the expression levels of the wild-type strain, PwtflhDC (both P1 and P5 are expressed) (Figure 5A). Moreover, transcription of flhDC operon from P1flhDC is activated earlier than P5flhDC because (i) the transcription profile of the flhDC operon in construct P5−flhDC (P1-expressed) overlapped with that of the wild-type strain from OD 0.1 to OD 0.4, (Figure 5A) and (ii) there was a delay in the transcription of flhDC operon in construct P1−flhDC (P5-expressed) where the transcription started taking place at OD∼0.35 (Figure 5A) compared to the wild-type PwtflhDC and P5−flhDC (P1-expressed) strains (OD∼0.2). The same hierarchy of expression of P1 and P5 was observed in batch culture (Figure 6A & B). The transcription of flhDC operon in P5−flhDC (P1-expressed) started declining at OD∼0.4–0.5, meanwhile, transcription of flhDC operon in strain P1−flhDC (P5-expressed) was more pronounced at a later growth stage accounting for ∼60% relative to the wild-type at OD∼0.6 (Figure 5A). It is apparent from the dynamic profile of flhDC operon transcription, that P5flhDC promoter transcription occurs concomitantly with a cessation or decline in the transcription from P1flhDC (Figure 5A). These results indicate that P1flhDC is an early promoter, whose activation drives the transcription of flhDC operon synthesis at early growth phase followed by a cessation or decline once P5flhDC promoter is activated.
We have demonstrated that HilD is a positive regulator of flhDC transcription (Figure 1D & 2D). As shown (Figure 2D), when cells are grown in the 96 well plate format, the effect of HilD on the transcription of flhDC takes place starting at OD∼0.4. In order to determine which of the two promoters, P1flhDC or P5flhDC, is the target of the positive regulation by HilD, we compared the dynamic profile of flhDC transcription in PwtflhDC, P1−flhDC and P5−flhDC constructs in a wild-type and its isogenic strain hilD null mutant (Figures 5B & C). We established that, relative to the wild-type strain background, a deletion of hilD (i) reduced PwtflhDC promoter transcription; (ii) abolished the transcription of flhDC operon in construct P1−flhDC (P5-expressed) (Figure 5B) and (iii) did not affect the transcription of flhDC operon in construct P5−flhDC (P1-expressed) (Figure 5C). These results indicate that HilD promotes transcription from P5flhDC and has no apparent effect on P1flhDC promoter transcription.
Transcription of the flhDC operon is subjected to negative feedback by RflM, which is activated at the transcriptional level by FlhD4C2 [21]. To further study the effect of the negative autoregulation on flhDC operon transcription kinetics, we monitored the transcription profile, over time, in the three strains PwtflhDC, P1−flhDC and P5−flhDC in the absence and presence of RflM. We established that there was an increase in the transcription from PwtflhDC in the absence of RflM (Figure 2B). We demonstrated that the P1flhDC promoter is under negative autoregulation by RflM because the expression of flhDC operon in strain P1−flhDC (P5-expressed) was similar between the wild-type and its isogenic rflM null mutant (Figure 5D). Additionally, we found that RflM did not appear to regulate P5flhDC because flhDC transcription in strain P5−flhDC (P1-expressed) increased in the absence of RflM (Figure 5E). These results demonstrated that in the wild-type background the P1flhDC promoter is subjected to negative autoregulation through RflM, while transcription from P5flhDC appeared to be RflM independent.
We employed an alternative approach to confirm which of the flhDC promoters is specifically inhibited by the transcriptional factor RflM. We monitored the transcription of flhDC in a strain that overproduces RflM under control of the arabinose promoter, ParaBAD::rflM+. In the presence of arabinose, used to induce overexpression of rflM, we observed an inhibition of transcription of flhDC operon in the three strains tested, PwtflhDC, P1−flhDC and P5−flhDC (Figure 5F). These results suggest that RflM is able to inhibit transcription of flhDC operon from both promoters, P1 and P5, which is in contradiction to the specific inhibition of the P1flhDC but not the P5flhDC promoter by RflM observed in Figures 5D & E. RflM protein production or stability appears to decline in function of cell growth cycle (Figure 3A), suggested that continuous production of RflM might affect indirectly the expression of P5flhDC. Because HilD is an activator of the P5flhDC promoter, we hypothesized that overexpression of RflM inhibits transcription of hilD gene. In order to test this hypothesis, we monitored the activity of a luciferase transcriptional fusion of the hilD promoter, PhilD, in two genetic backgrounds: (i) ParaBAD::FCF (ii) ParaBAD::rflM+. We observed that under conditions that overproduce RflM, presence of arabinose, there was an inhibition of transcription of the autoregulated gene hilD (Figure 6, compare column 1 to column 2). Note that the strains used to determine luciferase activity are all flhD+C+, and overexpression of RflM inhibits flhDC transcription required for production of the posttranslational regulator of HilD. Thus, the effect of RflM, on hilD transcription could be indirect through inhibiting flhDC. To test if the effect of RflM on hilD, is direct or indirect we used two additional strains (i) ParaBAD::FCF PflhDC::T-POP and (ii) ParaBAD::rflM+ PflhDC::T-POP. For the PflhDC::T-POP backgrounds the flhDC operon is transcribed from the tetracycline(Tc)-inducible tetA promoter, and as such are flhDC− in the absence of tetracycline and flhDC+ in the presence of tetracycline. First, we observed that flhDC controlled transcription of the hilD gene, because in the absence of Tc, there was a 2-fold decrease in the PhilD transcription level in the PflhDC::TPOP strain background (Figure 6D, compare column 3 to column 5). Moreover, we demonstrated that under condition of RflM overexpression, there was a higher level of inhibition of hilD transcription compared to the reduction observed in the PflhDC::T-POP background (Figure 6D, compare column 5 to column 6). The overproduction effect of RflM was not rescued by addition of Tc to induce flhDC transcription, an activator of hilD transcription (Figure 6D, compare column 6 to column 8). These results demonstrated that RflM could inhibit transcription of the hilD gene in an flhDC independent manner. Thus flhDC and rflM have opposite effects on the transcription of hilD, where flhDC is an indirect positive regulator of HilD, yet high levels of RflM inhibit hilD transcription. Since HilD is an activator of P5flhDC transcription, we conclude that the negative effect of RflM overproduction on transcription of P5flhDC is indirect and through inhibition of hilD gene transcription,
The presence of two principal TSSs within the flhDC operon promoter region combined with the hierarchical regulation by different transcriptional factors, suggests that there is differential regulation at the promoter by different transcriptional regulators at different cell growth phases. We investigated which of the specific regulators: RcsB, LrhA, SlyA and RtsB control transcription of flhDC through the P1flhDC and P5flhDC promoters start-sites.
There appears to be five stages of flhDC transcription that are controlled by three clusters of response regulators. Deletion of either, rcsB, lrhA or rflM resulted in increased motility compared to the wild-type strain [18], [19], [62]. We observed that null mutations in any of the late regulators: hilD, rtsB or slyA, did not affect motility (Figure 7A). Based on the expression profiles of the flhDC operon in these mutant strains, these results establish that Salmonella wild-type motility will only need to reach a threshold of flhDC expression for motility, while increased flhDC expression later in the growth phase has no further effect on motility. It is noteworthy to mention that factors that affected the early transcription of the P1flhDC promoter: LrhA, RcsB (Figure 6A) and RflM (Figure 5E) affected motility while transcriptional factors, HilD and SlyA, that regulate P5flhDC promoter late in the growth phase (Figure 5B & 6B) did not affect motility (Figure 7A). Moreover, RtsB, by inhibiting transcription from P1flhDC at later stages of growth (Figure 6B), did not inhibit motility suggesting that the growth phase combined with activation of flhDC promoters is important for motility (Figure 7A). It is noteworthy to mention that the factors that affected transcription of P5 flhDC but not motility are bona fide virulence factors.
We decided to study the effect of the flhDC promoter mutations on the motility of Salmonella. We constructed strains harboring single mutation in each of the promoters separately. Thus P1− refers to a strain that has a mutation in P1 promoter, etc. Note that these strains in contrast to strains harboring the luciferase constructs do not harbor a duplication of the flhDC operon. We demonstrated that strains defective in P1flhDC start-site transcription (only P1 is mutated) were non-motile while P5flhDC defective strains (only P5 is mutated) exhibited no apparent reduction of motility (Figure 7B). There was a motility defect of the strains P2− and P6− that is related to the effect of CRP (as discussed earlier and in Supplementary material). The motility of P3− and P4− were not significantly different from the wild-type strain. These results confirmed that in the wild-type background transcription from P1flhDC is a prerequisite for motility while P5flhDC is not required for motility. These results also suggested that the right timing of expression of flhDC is essential for motility. If this hypothesis is correct, we could expect that if flhDC is expressed from P5flhDC promoter at an early time point it should confer a motility phenotype. To test this hypothesis we used the non-motile strain P5+ (only P5 is functional and the other promoters are mutated) (Figure 7C) to isolate suppressors of motility inhibition. This strain was used in order to limit isolating mutations in the other promoters of flhDC that would otherwise suppress motility [16]. We isolated a spontaneous suppressor that restores motility to the P5+ strain (Figure 7C) and mapped the mutation to the promoter region of hilD gene (addition of a thymine residue at position −51 from the start codon of HilD and resulting in higher expression of hilD (labeled hilDup)). The isolation of this mutation confirmed that HilD regulates the P5flhDC promoter. If the hypothesis that the timing of expression of flhDC as a prerequisite for motility is correct, then a hilD-up mutation should promote transcription of flhDC operon from P5 promoter at early growth phase. To test this hypothesis we used a transcriptional lac fusion to fliL, a class 2 promoter that is positively regulated by FlhD4C2, as readout to determine the expression of the P5 promoter transcription. The transcription of fliL indicates the presence of FlhD4C2-dependent transcription. Transcription of fliL in the P5+ strain was very low during early growth phases and increased when cells reached an OD of 1.4 (Figure 7D). These results suggest that P5+ cells are able to express flagellar genes at later stage of cell's growth phase yet they are not motile. Interestingly, overexpression of hilD, hilDup mutant resulted in a premature activation of P5flhDC, leading to the transcription of fliL at early growth phase and similar to the timing and levels of the wild-type strain (Figure 7D). These results suggested that the timing of FlhD4C2 production during an early growth phase is critical for motility.
The complex networks and the number of factors necessary for the production of functional flagella and the resulting motility, though beneficial for the bacteria, represent a significant requirement on the cell's resources [63], [64]. At the top of this cascade sits the flhDC operon [7]. We established now that Salmonella flhDC operon is primarily transcribed from two promoters, P1flhDC and P5flhDC. The activities of these two promoters are coupled to five stages controlling flhDC transcription and each stage is differentially controlled by a set of transcriptional factors: (1) repression of transcription of flhDC during the initial growth phase by LrhA and RcsB (2) repression by RflM at early exponential phase (3) activation through the action of HilD at mid exponential phase (4) repression by SlyA and RtsB at the onset of stationary phase, and finally (5) shut down at late stationary phase.
The pre-log steady state transcription of flhDC regulation is controlled by two transcription factors, RcsB and LrhA. Null mutation in any of these transcriptional regulators, promoted flhDC transcription early in the growth phase and this inhibition was maintained throughout the rest of the growth phase (exponential and stationary). We found that the effect of LrhA and RcsB was coincident with activation of transcription of their respective genes. As cell densities reached an OD of 0.2–0.3, transcription of flhDC increased. The increased flhDC transcription resulted in transcription of rflM, which in turn resulted in the feedback inhibition of flhDC transcription. This effect was consistent with the concurrent transcriptional activation of flhDC and rflM, where a surge of transcription of rflM mimicked that of flhDC and decayed quickly compared to the rest of the regulators controlling flhDC transcription. At the protein level, RflM appeared to follow the same early production and a quick decay as observed at the transcriptional level. We conclude that RflM limits flhDC transcription perhaps to efficiently control the kinetic expression of the middle and late flagellar class genes to facilitate flagellum assembly. Class 2 promoters respond differently to FlhD4C2 levels allowing the cell to control the timing of an individual class 2 operon transcription with respect to the other class 2 operons. Auto-repression at the transcriptional level has been shown to reduce relative variance and duration of fluctuations, and consequently limits noise in downstream expression [65], [66]. Expression of fliC, encoding the filament component of the flagellum, has been demonstrated to be bistable [67], [68]. We suggest that RflM would fulfill the noise reduction of flagellar class 2 and class 3 promoters transcription during exponential growth phase, by controlling class 1 flhDC operon transcription. In support of this hypothesis, a null mutation of rflM gene has been shown to increase heterogeneity of fliC expression in a cell population when compared to wild-type [21].
Once bacteria reach mid-exponential phase growth, there is a second layer of control on flhDC operon transcription. This control is positive, and is brought on by the effect of a virulence-associated transcription factor, HilD. There was a delay in the positive effect of HilD compared to the negative control exerted by RcsB, LrhA and RflM. This delayed HilD effect on flhDC operon transcription was due to the time required to activate HilD expression through FlhD4C2-dependent FliZ production. FlhD4C2 activates fliZ gene transcription from a flagellar class 2 promoter and FliZ, in turn, activates hilD expression at the post-translational level [57]. Finally, a third layer of flhDC transcription takes place and, unexpectedly, is also controlled by HilD. HilD activates the transcription of two regulatory factor genes, rtsB [17] and slyA (data not shown). RtsB and SlyA are two DNA binding regulators, which then act to inhibit flhDC transcription.
There is no doubt that flagellar motility provides a significant survival advantage over non-motile bacteria in many environmental situations. Furthermore, the link between production of flagella and other regulatory networks [69]–[72] would be affected if an unchecked production of flagella occurs. The overexpression of the flagellar regulon also attenuates Salmonella virulence [73]. These observations could explain the array of negative regulators controlling transcription of flhDC operon and keeping a check on the flagellar synthesis as well as FlhD4C2 production.
While the literature reports the presence of either four or six transcription start-sites in the flhDC promoter region [13], [29], our work suggests that only the P1flhDC and P5flhDC promoters are functional in a wild-type strain under laboratory growth conditions. First, we demonstrated that there was a reduction in flhDC operon transcription in the absence of P1flhDC or P5flhDC compared to the wild-type strain (Figure 4C & D). Second, we showed that flhDC operon transcription was totally abolished in P1−P5−flhDC double mutant (Figure 4F). We confirmed that the P6flhDC promoter is active only in the absence of CRP [13]. Moreover, there was no apparent effect of P4flhDC, P3flhDC and P2flhDC promoters on flhDC transcription. In E. coli, CsrA, a carbon storage global regulator, activates flhDC expression in an RNaseE-dependent manner through protection of 5′end cleavage [23]. The 5′-UTR of the P5flhDC start-site transcript is 534 bases in length. We suspect that the presumed P3flhDC and P2flhDC start-sites resulted from RNAseE-dependent RNA-processing and/or degradation of the P5flhDC transcript. The P4flhDC start-site might also result from RNA processing; however, the isolation of mutants in the −10 region of P4flhDC that result in increased flhDC transcription suggests there might be unknown conditions where transcription from P4flhDC occurs [16].
Genes with multiple transcription start-sites combined with auto-regulatory networks have been described in other systems. These include, Salmonella phoP, Bordetela pertussis bvgA, E.coli rrnA, and Salmonella fliAZ operon [27], [74]–[78]. These four cases bear similarity with flhDC operon transcription from P1flhDC and P5flhDC promoters. However, the case of flhDC is more elaborate, where two disparate pathways are used as feedback control. First, we demonstrated a sequential activation of P1flhDC and P5flhD transcripts that are growth phase dependent (Figure 5A). The P1flhDC promoter activating two regulatory pathways resulting in both a negative and a positive regulatory loop and each of these loops has a specific effect on the flhDC operon promoters. The negative loop starts with P1flhDC, leading to the production of FlhD4C2 that activates rflM, which in turn feedback inhibits the P1flhDC promoter (Figure 5E). The positive feedback loop is also generated from P1flhDC, where transcription of flhDC operon from P1flhDC leads to fliZ gene transcription followed by FliZ activation of hilD. HilD then activates the second flhDC transcriptional cycle from P5flhDC (Figure 5B). Paradoxically, HilD controls transcription of rtsB and slyA genes, whose products binds to the flhDC promoter region (Figure 3B) and inhibit transcription, from P1flhDC and P5flhDC, respectively (Figures 6A & B).
The three promoter classes of the flagellar regulon, class 1, class 2 and class 3; are expressed in a temporal cascade that coincides with flagellum assembly [79]. The control of flagella production is ultimately determined through the production of FlhD4C2. However, when flhDC is highly over-expressed the cells are not motile for reasons that are not understood. Thus, an intricate temporal control of gene expression and specific quantities of a functional FlhD4C2 master regulator are essential for motility. For example, the activator of type I fimbriae gene expression, FimZ, represses flhDC transcription suggesting that adherence is impeded in the presence of functional flagella. Neither deletion of flhDC nor over-expression of flhDC affect type I fimbriae gene expression suggesting that the presence of fimbriae (at wild-type levels) does not impede swimming. FlhD4C2 activity is also required in other cell processes such as Spi1 gene expression and other genes less characterized such as the srfABC operon [80], which is implicated in surfactin production and the modABC operon [80], which is involved an anaerobic respiration. This leads us to speculate that P1flhDC is required for flagella production and P5flhDC is required for growth in various environmental conditions such as in biofilms or in host cells. One possibility is that the activation of flhDC transcription from P5flhDC might represent a mechanism of protein amplification by a surge of transcription, when it is necessary to turn on the Spi1 regulatory network, even under conditions where flagella synthesis is inhibited at the level of fliA and fliC. This scenario can be very useful after infection when the bacteria requires expression of virulence factors to survive the physical and immune clearance of the eukaryotic host.
Flagella appear to be required for reaching and selecting point of entry of bacteria into host cells [81]. The low pH of the stomach will cause flagella already present to depolymerize [82]. In the intestine, the early transcription of flhDC operon from the P1 promoter provides the transcription factor, FlhD4C2 for expression of functional flagellar machinery to reassemble filaments and allow bacterial cells to swim to selected points of entry into epithelia cells. At the time of invasion, expression of both T3SS1 and flagella has been shown to be required. Thus, in the second step, the already expressed flhDC from P1flhDC promoter activates transcription of fliZ, the posttranslational regulator of HilD. In turn, HilD promotes transcription of Spi1 genes, leading to invasion. Thus P1-expressed flhDC fulfills two functions: driving the cells near the point of entry and also boosting the expression of Spi1, necessary for invasion, through its effect on HilD. It is noteworthy to mention that invasion of epithelial cells is a rapid process occurring within 10 to 15 minutes after introduction of S. typhimurium into the intestinal lumen [83]. Translocation of bacteria across the epithelial barrier and into the underlying tissue is observed within 2 hours after infection of ligated ileal loops [83], [84]. Interestingly Salmonella can replicate within two distinct intracellular environments: intravacuolar and cytosolic [85]. Once inside the host, the expression of both flagella and Spi1 appear to be downregulated but not abolished with most of the cytosolic population expressing both flagella and Spi1 at latter stage of infection. In addition, only a subset of T3SS1-induced cytosolic bacteria was motile [85]. We speculate that once bacteria invade epithelial cells, HilD activates P5flhDC and down-regulates the transcription of P1flhDC in an RtsB-dependent manner. The transcription from P5flhDC is bistable leading to two populations of cells, one is flagellated and the other is not (∼10% of cells being flagellated). This bistable expression of P5flhDC is reminiscent with the bistable expression of Spi1. We suggest that the presence of two populations inside epithelial cells could be explained by the bistability from P5flhDC promoter and the consequent downregulation of P1flhDC might represent a mechanism to limit the number of flagellated cells. The cytosolic growth of Salmonella leads to the extrusion of epithelial cells as a host defense mechanism [85]. The consequent release of the invasion-prone flagellated cells bacteria back into the mucus rich and inflamed gut endows Salmonella with a fitness advantage to use the energy-taxis mechanism to benefit from inflammation [86]. We speculate that the different timing of expression of flagellar promoters P1 and P5 and the bistable expression of P5flhDC represent a mechanism by which bacteria can disseminate and increase transmission by fecal shedding. These hypotheses are under investigation.
An additional scenario is that the transcription from P5flhDC has no effect on the synthesis of flagella but rather leads to the production of single subunits of the active transcriptional complex FlhD4C2. It has been shown that the inhibition of FlhD4C2-dependent transcription inside host cells is due to the effect of YdiV-mediated ClpXP degradation of the FlhD4C2 complex. The expression from P5flhDC late during cell growth will not allow for motility because the activation of the ClpXP leads to the degradation of the complex. However, ClpXP in addition to degrading the FlhD4C2complex also degrades the FlhC single subunit but not FlhD. This leads to the hypothesis that single FlhD or FlhC subunits might activate transcription of other genes required for virulence [87]
Our finding can be rationalized in terms of a model (Figure 8). Two regulatory factors, LrhA and RcsB regulate flhDC by inhibiting transcription from P1flhDC and P5flhDC. The effect of RcsB is more dominant on P1flhDC then on P5flhDC, whereas LrhA represses more strongly P5flhDC than P1flhDC. Transcription activation of P1flhDC by CRP leads to a rapid transcription of rflM, which in turn reduces transcription of P1flhDC, and limits a rapid class 2 and class 3 genes expression. The FlhD4C2 complex, already produced, allows motility to proceed and also promotes activation of HilD at the posttranslational level through FliZ, ultimately leading to activation of transcription from the P5flhDC promoter. This positive autoregulation also generates a subsequent inhibition of flhDC operon transcription, of both P1flhDC and P5flhDC promoters, by two HilD-induced regulatory factors SlyA and RtsB, themselves regulated by different environmental cues. The activation of transcription from P5flhDC would lead to higher expression of FlhD4C2. Though not necessary for motility, it could affect expression of HilD. Because, HilD is required for Salmonella survival inside host cells, this positive circle of activation might be well suited for virulence.
Bacterial strains and primers used in this study are listed in Table S1 and Table S2, respectively (Supplementary Information). Bacterial cells were routinely grown in Luria-Bertani (LB) broth and, when necessary, supplemented with appropriate antibiotics at the following concentrations: Kanamycin (5–10 µg/ml), tetracycline (15 µg/ml) in agar plates and for induction of T-POP 3.5 µg/ml). L-arabinose was used at 0.2% (w/v) when needed. Motility agar plates were prepared as described earlier [62]. The generalized transducing phage of S. typhimurium P22HT105/1 int-201 was used in all transductional crosses [88]
For the construction of strain TH18684 DUP[(PwtflhDC8093-luxCDBAE)*Km*(PwtflhDC-flhD+C+)] primers 5104 and 5103 [designed to delete the replication origin and tetracycline resistance (TcR) cassette of the plasmid pRG38 [89]] were used to amplify the kanamycin cassette of pKD3. The PCR product was electroporated into TH18710 (LT2/pKD46/pRG38) followed by selection for kanamycin resistance (KmR). KmR colonies were pooled and infected with P22 to produce a transducing lysate. This lysate was used to transduce LT2 selecting KmR. The KmR transductants were replica-plated in LB+Km and LB+Tc. Tc-sensitive (TcS) and KmR colonies should have resulted from integration of PflhDC-luxCDBAE into the chromosome generating a duplication of the promoter region of the flhDC operon. To check the integration of a single copy of PflhDC-luxCDBAE-Km and to screen for the presence of any duplication of the luxCDBAE upon integration, a set of primers [1401 (reverse for luxC)- 3091 (forward in upstream of PwtflhDC promoter region not present in the plasmid pRG38)] demonstrated the correct integration of the plasmid at the flhDC promoter region. A second PCR reaction using [Primers 267 (Km) and 1403 (luxE)] demonstrated the correct placement of Km cassette after the luciferase operon. Amplification with primers, 1403 and 1401, indicated a single copy integration of the plasmid without its origin of replication. Five candidates were obtained having a single integration of PwtflhDC-luciferase into the chromosome. One of the five candidates was sequenced and used in this study (TH18684). The Duplication of PflhDC was maintained in the presence of 5–10 µg/ml Km.
Mutations in the promoter region of PflhDC-lux were constructed using the λ-Red recombinase system, as reported previously [90], using the primers listed in Table S2. All transcriptional fusion constructs using the luciferase operon reporter used the strain TH18727: (DUP[(PflhDC8093::tetRA-luxCDBAE*Km*(PflhDCflhD+flhC+)]/pKD46) as the electroporation recipient. Individual fusion constructs with specific promoter regions were designed as follows: the rcsB promoter region included 400 bp upstream of the start codon through 230 bp of coding region, the rcsD promoter region included 466 bp upstream of the start codon through 260 bp of coding region, the slyA promoter region included 258 bp upstream of the start codon and 290 bp of the coding region, the hilD promoter region included 300 bp upstream of the start codon through 240 bp of coding region, the rtsA promoter region included 264 bp upstream of the start codon through 290 bp of coding region, the lrhA promoter region included 880 bp upstream of the start codon through 200 bp of coding region and the rflM promoter region included 460 bp upstream of the start codon through 284 bp of coding region. The promoter regions defined above were amplified by PCR using the respective primers listed in Table S2, and electroporated into strain TH18727, using the Lambda-Red recombinase system selecting for replacement of tetRA element with a PCR-amplified DNA fragment [90].
Chromosomal FLAG-tagged HilD, RcsB and chromosomal HA-tagged RtsB, SlyA, RflM, HilD and LrhA were generated by the Lambda-Red recombinase system, as described previously [91] using gene-specific primer pairs, as shown in Table S2. All strains were verified by PCR amplification and DNA sequence analysis.
LB+Km medium containing 1% tryptone, 0.5% yeast extract, and 0.5% NaCl was used for growth of all bacterial cultures to determine the transcription activities of luciferase. Overnight cultures in LB+Km cultures were adjusted to the same OD 595 nm, then, 8-ml glass tubes containing 2 ml of LB+Km were inoculated with a 500-fold dilution of the bacterial suspensions and incubated at 30°C in a water bath with shaking at 250 rpm. For determination of luciferase activity in batch cultures, samples (200 µl) were taken at different time point and the light production along with the OD595 were measured in 96 well plates in a microplate reader (PolarStar Optima). For the determination of luciferase activity in 96 well plates, adjusted OD595 of overnight bacterial cultures at 37°C were diluted 500-fold in LB+Km and 200 µl of diluted bacteria were added to 96 well dark plates (Greiner). The plates were sealed with breathe easy membrane (to minimize evaporation and to allow growth in semi-aerobic conditions) and incubated in a chamber/shaker of a PolarStar Optima microplate reader (BMG labtech) set at 30°C. The conditions of the plate reader to determine the light production and OD 595 nm were as follow: orbital Shaking for 300 s at 150 rpm, 5 s stop and 95 s for luciferase light reading of the wells. For normalization of results a 0.1 s integration time was used. The OD 595 nm and light production (luciferase) was measured over time using a PolarStar Optima microplate reader (BMG labtech). For the background, we took the average measurements of the strain (TH18402) harboring mutations in all the promoters of flhDC. After background correction, relative light units (Arbitrary Units) were calculated by dividing the lights reading with its corresponding OD 595 nm. The OD 595 nm in our setting of the PolarStar Optima reader corresponds to ∼1.69 factor of the OD 595 nm read with 1 ml spectrophotometer.
Whole-cell extracts were prepared from samples of cultures grown in LB. 500-ml flasks containing 100 ml of LB were inoculated with a 500-fold dilution of the bacterial suspensions and incubated at 30°C in an orbital shaker at 150 rpm. Cells were collected at different optical densities (0.25, 0.4, 0.6, 0.8, 1 and 1.3) and washed twice with ice cold PBS. Pellets were lysed, at room temperature for 15 minutes, using B-PER reagent (Fisher, product #78243) with freshly added lysozyme (1 mg/ml) and protease inhibitors (Roche). The lysates were clarified by centrifugation at 4°C for 10 minutes. Supernatants were transferred to new eppendorfs and the extracted proteins were quantified using the BSA assay (BioRad). Samples, containing 50 µg of total protein per lane, were electrophoresed onto 12% to 14% Tris/SDS gels. To detect RtsA-HA a 15% Tricine-SDS gel was used as described [92]. Following transfer onto a 0.45 µm pore size polyvinylidene difluoride (PDVF) membrane (Immobilon P, Millipore) using a semidry transfer apparatus (Bio-Rad), membrane were blocked for 1 hour at room temperature with freshly prepared non-fat dry milk (5% w/v) in PBS. For detection of HA-tagged or Flag-tagged proteins, membrane blots were incubated overnight at 4°C with anti-HA (Covance) or anti-Flag M2 (Sigma) mouse monoclonal antibodies at 1∶1,000 and 1∶2,000 dilutions respectively. DnaK was detected using Anti-DnaK (Covance) diluted 1∶10,000. The blots were washed three times with PBS-T (PBS+0.1% tween) and incubated protected from light with green or red infrared dye-conjugated secondary antibody in non-fat dry milk (3% w/v) in PBS-T for 45 minutes at room temperature. Following three washes in PBS-T and one wash in PBS. Labeled proteins bands were detected using the Odyssey Infrared Imaging System (Li-COR Biosciences, Lincoln, NE, USA).
CHIP was performed as in [93] with modifications. Bacterial batch cultures were grown at 30°C to different ODs, at which point formaldehyde (final concentration of 1%) was added to cells. After 20 min at room temperature in an orbital shaker, cross-linking was quenched by the addition of glycine (500 mM) for 10 minutes. Samples were then placed on ice for an additional 10 minutes to complete quenching. Cells were collected by centrifugation, and washed twice with cold phosphate-buffer saline (pH 7.5). Cells pellets were resuspended in 1 ml of lysis buffer (10 mM Tris, pH 8.0, 20% sucrose, 50 mM NaCl, 10 mM EDTA, 10 µg/ml of lysozyme) and incubated at 37°C for 30 min. Following lysis, 1 ml of immunoprecipitation buffer (50 mM HEPES-KOH, pH 7.5, 150 mM NaCl, 10 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate, 0.1% sodium dodecyl sulfate) and phenylmethylsulfonyl fluoride (final concentration of 1 mM) were added. To shear cellular DNA to an average size of 500 to 1,000 bp, the cell extracts were sonicated on ice using Misonix Sonicator 3000 with a microtip at power 2 for three 10 s pulses, with 30 s rests on ice between pulses. The lysates were clarified by centrifugation and the supernatant were treated with 5 µl RNaseA (10 µg/ml) at 37°C for 30 minutes. The treated supernatant was retained for use as the input sample in the immunoprecipitation experiments. Aliquots of sheared samples were uncross-linked by incubation for 2 h at 42°C and 6 h at 65°C in 0.5× elution buffer containing freshly added 0.8 mg/ml of Proteinase K. DNA was purified using a PCR purification Kit (Bioline). An aliquot of purified DNA was run in a 1.25% agarose gel to confirm the shearing of DNA to 500–1000 bp fragments and DNA was quantified using Nanodrop spectrophotometer. An Aliquot of the input sample (2 µg) was used for each immunoprecipitation experiment. The sample was incubated with 50 µl of proteinPlus A/G beads (Santa Cruz) and 4 µl of HA monoclonal antibody (Covance) or Flag M2 antibody (Sigma) for 90 min at room temperature on a rotating wheel. An immunoprecipitation experiment without antibody was also set up as a negative control. The beads were collected by centrifugation and subsequently washed three time with immunoprecipitation buffer and once with immunoprecipitation buffer plus 300 mM NaCl, once with wash buffer (10 mM Tris-HCl, pH 8.0, 250 mM LiCl, 1 mM EDTA, 0.5% Nonidet-P40, 0.5% sodium deoxycholate) and finally with PBS buffer (pH 7.5). Immunoprecipitated complexes were then removed from the beads by treatment with elution buffer (50 mM Tris-HCl [pH 7.5], 10 mM EDTA, 1% SDS). Crosslinking of immunoprecipitated samples was reversed by incubation for 2 h at 42°C and 6 h at 65°C in 0.5× elution buffer with 0.8 mg/ml of Pronase (Roche). Prior to analysis, DNA was purified from the immunoprecipitate by using a PCR purification kit (Bioline) and resuspended in 30 µl of TE and quantified using a Nanodrop spectrophotometer. Two micrograms of the fragmented DNA, isolated from DNA-protein complexes, was used as the input in all ChIP assays. Following purification, Real-time PCRs were run on a C1000 thermal cycler (BioRad) to analyze immunoprecipitated DNA. DNA samples were used in a 20 µl reaction mix containing a 1 µM concentration of each oligonucleotide and 10 µl of 2× SYBR-Green Reaction mix. Two pairs of primers, 3569-3477 and 3753-3090 covering the promoter region of flhDC were used (Table S2). PCR conditions were as follow: Initial denaturation at 95°C for 3 min, and 40 cycles of 95°C for 15 s and 60°C for 1 min, followed by the default melting curve program of the PCR machine. Fold-enrichments were determined by the 2−ΔCT method described in SA Biosciences User manual. To account for chromatin sample preparation differences, CHIP DNA fractions Ct values (Mean threshold cycles) were normalized (ΔCt(normalized ChIP) to the Input DNA fraction Ct values by substracting the Ct-values of the sample from the corresponding no antibody control. The percentage input of each ChIP fraction was calculated using 2(−ΔCt(normalized ChIP) and adjusted to the normalized background (No antibody) using the following formula: ΔΔCt(Chip) = ΔCt(normalized ChIP)−ΔCt(normalized NoAb). The IP fold enrichment was then calculated using 2(−ΔΔCt(ChIP/NAC)) to evaluate the fold amount of starting material of the sample applied in the real-time PCR.
|
10.1371/journal.pcbi.1005513 | lumpGEM: Systematic generation of subnetworks and elementally balanced lumped reactions for the biosynthesis of target metabolites | In the post-genomic era, Genome-scale metabolic networks (GEMs) have emerged as invaluable tools to understand metabolic capabilities of organisms. Different parts of these metabolic networks are defined as subsystems/pathways, which are sets of functional roles to implement a specific biological process or structural complex, such as glycolysis and TCA cycle. Subsystem/pathway definition is also employed to delineate the biosynthetic routes that produce biomass building blocks. In databases, such as MetaCyc and SEED, these representations are composed of linear routes from precursors to target biomass building blocks. However, this approach cannot capture the nested, complex nature of GEMs. Here we implemented an algorithm, lumpGEM, which generates biosynthetic subnetworks composed of reactions that can synthesize a target metabolite from a set of defined core precursor metabolites. lumpGEM captures balanced subnetworks, which account for the fate of all metabolites along the synthesis routes, thus encapsulating reactions from various subsystems/pathways to balance these metabolites in the metabolic network. Moreover, lumpGEM collapses these subnetworks into elementally balanced lumped reactions that specify the cost of all precursor metabolites and cofactors. It also generates alternative subnetworks and lumped reactions for the same metabolite, accounting for the flexibility of organisms. lumpGEM is applicable to any GEM and any target metabolite defined in the network. Lumped reactions generated by lumpGEM can be also used to generate properly balanced reduced core metabolic models.
| Stoichiometric models have been used in the area of metabolic engineering and systems biology for many decades. The early examples of these models include simplified ad hoc built metabolic pathways, and biomass compositions. The development of genome scale models (GEMs) brought a standard to metabolic network modeling. However, the vast amount of detailed biochemistry in GEMs makes it necessary to develop methods to manage the complexity in them. In this study, we developed lumpGEM, a tool that can systematically identify subnetworks from metabolic networks that can perform certain tasks, such as biosynthesis of a biomass building block and any other target metabolite. By generating alternative subnetworks, lumpGEM also accounts for the redundancy in metabolic networks. We applied lumpGEM on latest E. coli GEM iJO1366 and identified subnetworks/lumped reactions for every biomass building block defined in its biomass formulation. We also compared the results from lumpGEM with existing knowledge in the literature. The lumped reactions generated by lumpGEM can be used to generate consistently reduced metabolic network models.
| Stoichiometric models have been extensively used since 1980s [1–3] and prediction capabilities of these networks have been proven to be very useful. The size and structure of these models varied among different studies. One of the pioneering studies on E. coli through a small stoichiometric model is performed by Varma et al. [4,5] in where the authors described the model as composed of core carbon metabolism pathways namely, glycolysis, pentose phosphate pathway, TCA cycle and formation of some by-product formations accompanied by a part of the Electron Transport Chain (ETC). This stoichiometric definition is further extended by the integration of a biomass composition formulation that is provided in the classic text published by Neidhardt [6]. In this textbook, E. coli metabolism has been explained extensively, and all the components (amino acids, lipids, DNA, RNA etc.) that constitute 1gDW of cell were reported based on previous experiments [7]. Moreover, the amounts of 12 precursor metabolites from the core carbon metabolism (erythrose-4-phosphate, ribose-5-phosphate, pyruvate, alpha-ketoglutarate, phosphoenolpyruvate etc.) along with the requirement of cofactors (ATP, NADH, NADPH) and inorganic compounds (S, NH4) to synthesize these biomass building blocks (BBB) were estimated. Such representation has been used in different studies to understand the core carbon metabolism and its relation with biomass accumulation [8,9]. This information for E. coli allowed the authors to develop a model that can describe the growth requirements and limitations of the organism without including the complex biosynthesis routes for each individual biomass building block. With such a small stoichiometric representation of the core metabolism (~50 reactions), authors were able to predict many aspects of E. coli physiology. Similar metabolic models with a reduced representation of the biosynthesis of the building blocks have been used in many other studies for E. coli and other organisms [10–15].
In the following years, with the development of sequencing and high-throughput technologies, the gene-protein-reaction (GPR) associations [16] have been improved and the number of sequenced genomes has sparked off [17,18]. This accumulation of knowledge eventually has led to the development of Genome scale metabolic networks (GEMs) [19,20], which encapsulate all the known biochemistry of organisms. These comprehensive representations of metabolism are accompanied by biomass formulations that account for the cell composition [21]. The contribution of each biomass building block is either determined empirically, or approximated from phylogenetically close species [22].
Many metabolic models for various organisms [20,23–27] and their strains [28,29] have been constructed and they proved to be extremely useful for many different purposes ranging from strain design for biosynthesis of industrial chemicals to drug discovery [30]. Although GEMs are widely used and have provided important insight and guidance, small and yet predictive models are still widely in use in different areas, such as metabolic flux analysis (MFA) and kinetic modelling. And while the biomass formulation of Neidhardt is still in use and has proven to be valid in the last 25 years, as Pramanik et al. [12] have shown, the changes in the biomass composition have significant effects on the internal fluxes, thus should be considered very carefully. In this respect, the extended and curated biochemistry in GEMs can be used to validate and to improve the approximations made by Neidhardt, and to extend it for any organism and for every biomass building block defined in biomass compositions of GEMs and to account for alternative synthesis routes.
Towards this, we developed lumpGEM, a mixed-integer linear programming algorithm that identifies all the alternative reaction subnetworks that should be used to produce a cellular metabolite or biomass building block from a defined set of metabolites. For each subnetwork, lumpGEM derives a lumped reaction that can capture the overall stoichiometry of the subnetwork, while preserving the elemental balance. The concept of identifying minimum number of reaction or enzymes to perform a certain function is not new, throughout the last two decades, different studies have been performed around this idea. Hatzimanikatis et al. developed a mixed-integer linear programming (MILP) formulation to determine the minimum number of regulatory structures to manipulate to optimize a bioprocess, along with possible alternatives [31]. In another pioneering study, Burgard et al. [32] have determined the minimum number of reactions that can sustain growth in E. coli contemporary GEM [19]. Later, the concept of finding all possible optimal solutions in metabolic networks has been introduced [33]. Elementary flux modes analysis (EFMA) has been another popular tool to analyze the metabolic networks in terms of minimal functional units [34] and used before to study nucleotide production [35]. The main limitation of EFMA is the requirement of generation of all the EFMs, which is computationally too costly to be applied to genome-scale networks. Due to this limitation, Figueiredo et al. [36] discussed the concept of K-shortest EFM, an MILP approach, and identified 10 minimal reaction sets that can produce lysine in E. coli and C.glutamicum. In lumpGEM, we follow a similar but an efficient approach compared to the previous studies and we have merged it with thermodynamics-based flux analysis (TFA) to account for bioenergetics constraints. With lumpGEM we can generate thousands of thermodynamically feasible minimal subnetworks that can produce a target metabolite from a set of selected precursor metabolites.
In this study, we focused on the biosynthesis pathways of biomass building blocks of E. coli iJO1366 [29], and S. cerevisiae iMM904 [37] and with lumpGEM (Fig 1), we have identified known and possible other alternative synthesis routes/subnetworks for all BBBs from core carbon metabolism as defined in [4,5]. We demonstrated that lumpGEM is capable of building lumped reactions in where the contribution of each core carbon metabolite to synthesize a biomass building block is identified and properly accounted.
Then, we performed a comparison between the values reported by Neidhardt and those generated by lumpGEM and found that lumpGEM recovered Neidhardt tables and extended them by including alternative biosynthetic routes/lumped reactions. We also performed a comparison between E. coli and S. cerevisiae for their metabolic capabilities to produce BBBs and revealed their similarities and differences. Such studies will help us to understand the capabilities of E. coli and S. cerevisiae ‘per’ biomass building block and identify the flexibility of the organism to survive by activating different parts of the metabolism to accumulate biomass. The generality of the method makes it applicable to any GEM that has a well-defined biomass composition. In addition, lumpGEM can generate lumped reactions from any part of the metabolism for any target metabolite, either a biomass building block or a biochemical and chemical compound or sets of compounds, which makes it a versatile tool to be used for different purposes.
The genome-scale models (GEMs) can simulate the growth characteristics of organism since they include the necessary biosynthetic paths to biomass building blocks (BBBs), which make up 1 gDW of cell. However, when the GEM is optimized for maximum specific growth using Flux Balance Analysis, the contribution of Mcore (the core metabolites, for definitions, see Material and methods) to synthesize a biomass building block is not evident from the flux distribution due to the degrees of freedom that the system has, and the alternative routes that a BBB can be synthesized. In order to overcome this limitation, lumpGEM utilizes a Mixed-Integer Linear Programming (MILP) formulation (See Material and methods) to reveal the contribution of Mcore and Rcore (core reactions) while preserving the elemental balance. MILP formulations have been often used on biochemical networks for many purposes [31,36,38,39] since they allow the control of reactions with an on/off manner. We made use of this binary decision in order to control the flux through the reactions of RncGEM (the reactions defined in GEM network other than Rcore, (also see Postulate 1, Material and methods), and lumpGEM allowed us to build minimal subnetworks that can synthesize BBBs from any selected part of the metabolism in GEMs, in this specific case, the core carbon metabolism.
The biomass formulation defined in E. coli IOJ1366 is very well characterized and detailed and contains 102 biomass building blocks. It is mainly composed of amino acids, lipids, nucleoside triphosphates (NTPs), deoxy-NTPs and inorganic compounds (Nickel, Zinc, Iron, etc.) along with cofactors such as NAD+/NADH, NADP+/NADPH, CoA/AcCoA, and FAD. Experimental estimates of the growth associated ATP maintenance and production of diphosphate are also included in the biomass composition.
The main difference between our approach for generating synthesis routes for the BBBs and the database-based analysis is that the subnetworks that our method generates may include branches from linear synthesis pathways. The difference emerges from the mass conservation constraint that we force during our analysis. For instance, the smallest subnetwork that lumpGEM generated for the synthesis of histidine is composed of 21 reactions and the precursors are ribose-5-phosphate (R5P) and oxaloacetate. In the databases such as EcoCyc [40], the pathway for histidine synthesis is linear and composed of 10 steps, specifying ribose-5-phosphate (R5P) as the only precursor. When we analyse the 21-reaction subnetwork (Fig 2), we see branching points in the linear route from R5P to histidine. Due to the mass balance constraint, three metabolites, 1-(5-Phosphoribosyl)-5-amino-4-imidazolecarboxamide, L-Glutamine and diphosphate cannot be balanced in a network that is composed of core reactions and the linear pathway from ribose-5-phophate to histidine. Hence, the generated sets of reactions are not only the linear routes from precursor metabolites to biomass building blocks, but balanced subnetworks with stoichiometrically proportional branches (Postulate 3, Materials and methods). lumpGEM captured reactions from various subsystems that are parts of histidine subnetwork, namely alanine and aspartate metabolism, anaplerotic reactions, folate metabolism, glutamate metabolism, histidine metabolism, nucleotide salvage pathway, purine and pyrimidine biosynthesis. In addition, the lumped reaction that is generated from this subnetwork (for lumping algorithm, see Materials and methods) has only core metabolites, biomass building blocks and by-products on both reactants and products sides. This representation is similar to Neidhardt’s definition, since he also described the stoichiometric expenditure of core metabolites in his estimations. Similar to our analysis, the values that Neidhardt et al. reported for the synthesis of histidine are different than the linear route that is reported in databases. However, Neidhardt reported only 1 lumped reaction for each biomass building block, and they are not overall stoichiometrically balanced (S2 Table). lumpGEM allows us to build alternative subnetworks and corresponding lumped reactions for the same BBBj. In this specific case, with the minimum subnetwork size Sminj being 21, lumpGEM generated 12 alternative subnetworks, and 3 unique lumped reactions. This signifies that the overall lumped reactions of different subnetworks can be the same. This has been observed also in a previous study that focuses on pathway generations for an industrial chemical [41].
A comparison between lumpGEM results and Neidhardt tables for the common metabolites (ATP, NADH, NADPH, etc) indicates that they are close to each other, however the overall stoichiometry that lumpGEM reports for the biosynthesis of histidine is more detailed and includes more metabolites as co-substrates and co-products (Table 1). The main reason for this discrepancy is that Neidhardt did not report elementally balanced lumped reactions for any of the biomass building blocks. For instance, oxaloacetate appears as a precursor to balance the non-core metabolite L-aspartate in the subnetworks that participates in adenylosuccinate synthase reaction as a co-substrate, and is not reported in Neidhardt precursors (Table 1). Therefore, values reported by lumpGEM are more comprehensive, and account for the balancing of all metabolites in the synthesis pathways, along with the small molecules, such as inorganic metabolites and protons.
The differences between the alternative subnetworks may emerge from different reactions in the ‘linear’ pathway from main precursor to the biomass building blocks, or from the other non-core reactions, which are balancing the non-core metabolites in the linear route. These two sources of differences, and especially the latter, may result in an explosion in the number of subnetworks that can be generated for some of the biomass building block. For metabolites like amino acids, which are not so far from the core carbon metabolism, the number of alternative smallest subnetworks Sminj is relatively small (Table 1), whereas for big molecules, such as lipids, there exists hundreds of alternative routes (S1 Table). The small number of alternative subnetworks for amino acids also shows that the number of non-core metabolites that appeared along the linear synthesis route is small, since the main explosion in the number of subnetworks emerges from these reactions. As an example, all 12 alternative subnetworks for histidine include the linear 10 steps route from R5P to histidine and alternative subnetworks are generated by other non-core reactions. Moreover, there is a slight correlation between the number of alternative subnetworks and the size of the subnetworks. Most of the amino acids that have more than 2 alternative subnetworks require more than 10 steps for their biosynthesis. A big molecule, Phosphatidylglycerol (dihexadecanoyl, n-c16:0) is a lipid with a Sminj of 40 reactions, and has 127 alternative subnetworks with 16 unique lumped reactions. In the first subnetwork generated by lumpGEM, within the 40 reactions, 34 of them are part of linear synthesis route and 6 of them are balancing non-core metabolites. However, despite the increase in the number of subnetworks, the number of unique lumped reactions remains relatively small (S1 Table).
When we analyse the alternative lumped reactions, in some cases, we only observe a difference in the stoichiometry of the cofactors in the reactant and product side. For example, the small change between the two alternative L-arginine lumped reactions is caused from alternative reactions that are decomposing pyrophosphate. In one case, pyrophosphate is decomposed into phosphate and water by the inorganic diphosphatase enzyme and in the second subnetwork it is decomposed into ATP, phosphate and proton using ADP as co-substrate by polyphosphate kinase enzyme. Thus, the main carbon flows of the subnetworks are exactly the same, and the lumped reactions differ only in the stoichiometry of the cofactor metabolites.
Along with the biomass building blocks such as amino acids and lipids, the biomass formulation of iJO1366 includes growth associated maintenance (GAM) and the de novo synthesis of adenylate pool metabolites. The Varma core network (S2 File 2) is capable of hydrolyzing the amount of ATP for GAM in the biomass (53.95mmol/gDW). However, the stoichiometric coefficients of ATP and ADP are not the same in the biomass equation, which indicates that the biomass formulation of E. coli requires synthesis of adenylate pool metabolites. In order to identify the subnetwork(s) for the synthesis of these pool metabolites, we followed the same procedure that we did for biomass building blocks, and we built a GEM with an additional reaction with the coefficients of ATP, ADP, water, phosphate and proton from GEM biomass. Then, by forcing a flux through this reaction, and by minimizing the number of non-core reactions (See Material and methods), we generated minimal subnetworks Smin. The resulting networks are composed of 27 reactions with 24 alternatives, which synthesize adenylate pool metabolite ADP.
Along with the biomass building blocks, GAM and adenylate pool metabolites, the biomass equation of iJO1366 includes diphosphate as a by-product. By following the same procedure that we have followed for adenylate pool, we generated subnetworks that balance the diphosphate in the biomass formulation. There are 2 alternative subnetworks composed of only 1 reaction in RncGEM that can balance the diphosphate in biomass equation: Inorganic Diphosphatase (PPA) and Polyphosphate Kinase (PPKr). Both of these reactions are decomposing the diphosphate into phosphate and proton, the former with water, and the latter with ATP/ADP cofactor pair.
When we analysed the alternative lumped reactions for the same biomass building block, we observe different requirement of precursors, cofactors, nitrogen and sulphur sources. This behaviour is expected, and a detailed analysis could also suggest which lumped reaction is more suitable for specific studies. One of the main criteria to rank the lumped reactions is their capability to synthesize the BBB from the carbon source, specifically the yield of BBB on the carbon source. In order to calculate the yield per lumped reaction, we built a ‘mini’ core model for each of them, which is composed of Mcore—Rcore, VBBB and the lumped reactions under study. By optimizing the synthesis of the selected BBB and calculating the C-mole yield over the carbon source of interest, specifically glucose, we ranked the alternative lumped reactions for each BBB. Interestingly, different lumped reactions can produce different amounts of biomass building blocks over a wide range of yield amounts (Table 2).
dTTP is a deoxy nucleoside triphosphates and its main precursor for all the generated subnetworks is ribose-5-phosphate (R5P) from Pentose Phosphate Pathway. Consequently, the core network can supply the same amount of carbon to all the generated subnetworks and corresponding lumped reactions. One explanation for the differences in yields (Table 2) is the capability of the lumped reactions to direct all the carbon from R5P to dTTP. When we analyse the 4 lumped reactions with highest yield, we see that on the product side, the only compound other than inorganics and cofactor pairs like quinone/quinol, ATP/ADP-PI is dTTP. For the lumps with second highest yield, along with the pyruvate (C3) in the reactant side, we see AcCoA (C2) and CO2 (C1) on the product side. In the 3rd highest yield case, fumarate (C4) replaces pyruvate on the reactant side, and succinate (C4) appears on the product side as a by-product. The lowest yield producing lumped reaction has fumarate and pyruvate on the reactant side, and has AcCoA, CO2 and succinate on the product side of the equation. This signifies that the lumped reactions with lower yield are losing carbon through those core metabolites and the number of carbons in these metabolites defines the yield. However, it is not possible to generalize such a rule, since the metabolites appearing on the product side of the lumped reaction can be assimilated by the system and the capacity of the network for this re-assimilation will have a significant effect on the yield of the lumped reaction. Moreover, having a lower yield does not necessarily mean that these lumped reactions are not useful for metabolic modelling, since they can be used under sub-optimal growth conditions or under conditions when growth is not the main physiological optimality criterion or to provide flexibility under mutation. The use of these alternatives and their physiological interpretation deserves an in-depth study and lumpGEM can serve as a framework for systematic studies.
By generating subnetworks for GAM and diphosphate, lumpGEM took into account all the components of biomass formulation both on the product and reactant sides. By testing for yield, we have shown that all the generated lumped reactions are capable of producing their target BBBj. However, to produce biomass, these lumped reactions must be able carry flux under the same quasi steady state condition, in the same model with biomass as cellular objective. This requires generating a metabolic network composed of the defined core network, lumped reactions, the transport and sink reactions defined in GEM. We used the Varma model (S2 File 2) as core model and the GEM iJO1366 as defined in the previous sections and introduced a new mixed-integer linear programming (MILP) formulation to have a systematic way to choose the lumped reactions. We formulated the problem as the following: We generated a model with all the lumped reactions generated by lumpGEM and calculated the theoretical maximum yield for biomass. We then fixed this yield in the model, and minimized the number of active lumped reactions and include only these lumped reactions for further analysis. This network consists of 56 cytosolic enzymatic (Varma network), 144 transport reactions, 78 lumpGEM output reactions along with 64 sinks (343 in total with biomass formulation) and 153 unique metabolites. The metabolic network is capable of producing 0.951/hr specific growth rate with 10 mmol/gDWhr glucose uptake rate. This signifies that all the lumped reactions were capable of producing corresponding BBBj successfully under the given condition simultaneously. The specific growth rate of GEM for the same condition is 0.997/hr, which is close to the biomass accumulation of the model generated with the output of lumpGEM. This shows that lumpGEM can be used to generate networks, which are small, but comprehensive and able to mimic the GEM behaviour.
lumpGEM can be applied to any GEM with a proper biomass formulation, and in order to test its applicability, we used it on a eukaryotic, compartmentalized GEM, iMM904 [37] of S. cerevisiae. This yeast is one of the mostly studied unicellular organisms along with E. coli, and has many biotechnological applications [42]; thus making it a strong candidate for modelling approaches. Applying lumpGEM on this S. cerevisiae strain revealed the contribution of different possible precursors and cofactors for each of the biomass building blocks defined in GEM, moreover it revealed alternative subnetworks and lumped reactions for the same biomass building block. This can be interpreted as building ‘Neidhardt style’ tables for S. cerevisiae.
In order to define the core for S. cerevisiae, we used the Metabolic Flux analysis (MFA) model of Christen et al. [43] and mapped the reactions of this model with the iMM904 reactions. In addition, we mapped reactions of iMM904 that are in Varma network but not in MFA model and included them as core reactions for S. cerevisiae network. The generated core network is composed of (S2 File 2) 88 reactions and 67 unique metabolites along 2 compartments, cytoplasm and mitochondria. Following these steps, we have applied the lumpGEM algorithm to the GEM as described in Material Methods section. The main difference between E. coli and S. cerevisiae GEMs are that S. cerevisiae GEM is compartmentalized, however this does not bring any more complexity for lumpGEM since it treats the transport reactions between compartments as it treats single enzymatic reactions. These transport reactions are part of the generated subnetworks for biomass building blocks and can participate in lumped reactions. These lumped reactions can include metabolites from different compartments, if these compartments are included in the ad-hoc selected network. By the nature of the lumpGEM algorithm, the lumped reactions include only core metabolites, biomass building blocks and by-products. Since the core network has 2 compartments, mitochondria and cytosol, the resulting lumped reactions can have metabolites from these compartments. The synthesis pathways of common biomass building blocks of S. cerevisiae and E. coli such as amino acids are very similar. The sizes of the networks differ mainly from the transport reactions between the compartments. Another reason for the divergence is the non-core metabolites along the linear routes to biomass building blocks, because there are different enzymes in these two organisms that are balancing these non-core metabolites. Interestingly, different subnetworks between the two organisms do not necessarily produce different lumped reactions. When the synthesis pathway is in one compartment (mainly cytosol), the lumped reactions of E. coli and S. cerevisiae are very similar (Table 3). L-phenylalanine, L-methionine, L-serine, L-cysteine and L-tyrosine are some examples of these similarities. Small differences for these overall reactions emerge from different cofactor usage. A similar behaviour is also observed for subnetworks including more than 1 compartment. The main difference between E. coli and S. cerevisiae overall reactions emerges from metabolites in different compartments, which are also different due to the energetics cost of transport reactions. As an example, the L-arginine biosynthesis in E. coli and S. cerevisiae is very similar, however the main difference emerges from the transport reactions between the compartments in the yeast. A smallest subnetwork for E. coli is composed of 13 reactions, and for S. cerevisiae, this number rises to 17. Two additional reactions for S. cerevisiae are the transport reactions between cytosol and mitochondria for the metabolites L-aspartate and ornithine. Another difference is the bicarbonate equilibration reaction (HCO3E) converting carbon dioxide to bicarbonate, which is the co-substrate for the carbamoyl-phosphate synthase reaction converting glutamine to carbamoyl phosphate. The reaction that produces carbamoyl phosphate in E. coli model uses carbon dioxide instead of bicarbonate as co-substrate. Moreover, it does not use L-glutamine as co-substrate. Therefore, the last difference in the subnetworks emerges from the synthesis of glutamine by S. cerevisiae through the usage of glutamate synthase (NADH2) and glutamine synthetase enzymes.
As a general comparison, lumpGEM generated 2797 subnetworks and 615 unique lumped reactions for 102 biomass building blocks for E. coli, whereas it generated 155 subnetworks and 114 lumped reactions for S. cerevisiae for 44 biomass building blocks (S1 File). Even though the numbers seems very different, the main explosion of number of possible subnetworks for E. coli emerges from lipids, which are not common with S. cerevisiae. Moreover, E. coli GEM has a more detailed biomass building block definition compared to S. cerevisiae GEM. Thus, it is fairer to compare the common biomass building blocks between two organisms as we have shown for amino acids.
We have also performed an analysis to compare E. coli and S. cerevisae for the production of a valuable industrial chemical, succinate. In this case, we did not define any core, other than the media composition, and minimized the number of active reactions to produce succinate. The analysis indicated that S. cerevisiae requires at least 33 reactions to produce succinate with a 100% carbon-mole yield over glucose, whereas E. coli needs only 25. The overall lumped reaction for E. coli do not have any metabolite other than glucose, oxygen and cytosolic proton in the reactant side, and only succinate, water and periplasmic proton on the product side. The overall lumped reaction generated for the S. cerevisiae subnetwork also includes the same metabolites, with additional production of ATP along with succinate. When we apply the same formulation to the biosynthesis of L-malate, we observe that minimal subnetwork generated for S. cerevisiae (30 reactions) are still bigger compared to E. coli (24 reactions), however this time the lumped reaction does not include any metabolite other than L-malate, glucose, water, oxygen and proton. The E. coli lumped reactions for L-malate is very similar to the succinate case. This analysis is done only for 1 Smin subnetwork for both compounds, and can be extended by generating all the minimum subnetworks, and for every common metabolite between E. coli and S. cerevisiae.
The complexity of cellular metabolism necessitates the development of methods to better understand and investigate the metabolic capabilities of organisms. For this purpose, small sized metabolic models are developed and widely used to study cellular physiology, and are proven to be useful platforms for many applications. With the emergence of genome scale models (GEMs), studies on metabolism entered a new era, since GEMs encapsulate all biochemistry that occurs in an organism. However, it is still crucial to be able to focus on certain parts of the metabolism with a modular manner and to understand the capabilities of these modules. Within this scope, we developed lumpGEM, a systems biology tool that captures the minimal sized subnetworks that are capable of producing target compounds from a set of defined core metabolites. In this work, we applied lumpGEM on all biomass building blocks (BBB) of E. coli iJO1366 and S. cerevisiae iMM904 and generated different subnetworks that can participate in producing biomass. We also used lumpGEM to re-define the pathway and subsystem definitions for the biosynthesis of BBBs, and identified that many different subsystems participate for the production of a single BBB. Moreover, by lumping the generated subnetworks, lumpGEM allowed us to identify the individual contribution of core metabolites and cofactors for the synthesis of each BBB. This procedure is a very promising method for many applications, such as experimental studies like MFA [44] and in silico studies like FBA, TFA [45,46] and atom mapping analysis. The main advantage of lumpGEM is its capability of making the ad-hoc selected core models consistent with genome-scale model (GEMs) in terms of biomass requirements. Metabolic Flux Analysis (MFA) analysis can benefit from such approach, since lumpGEM identifies the expenditure of core metabolites for the biosynthesis of biomass building blocks, thus accounting correctly for the metabolic costs.
lumpGEM can also be used to build synthesis pathways for any metabolite defined in the metabolic network. This makes it a versatile tool to study the characteristics of industrial chemical production strains [47], since it identifies all the enzymes, either linearly connected or nested that participate in the biosynthesis of the target compound. This approach can be used to compare the similarities and differences between the host organisms to produce a target chemical, since lumpGEM can compare the metabolic costs and capabilities of different organisms for the biosynthesis of an industrially relevant molecule. Apart from pathway and subsystems/subnetworks analysis, lumping strategy reduces the complexity of the networks significantly. By exploiting this property of lumpGEM, we built a reduced model that has an ad hoc defined core with a biomass yield very close to its parent GEM model. With a systematic approach to define the core [48], we can generate representative reduced models that are consistent with their GEM for different studies, such as kinetic modelling [49–51], in where it is crucial to base the analysis on models that do not sacrifice stoichiometric, thermodynamic and physiological constraints.
The algorithm considers a core system of metabolites and reactions and identifies them in genome scale model (GEM) of the organism. Using this GEM, it decomposes the biomass composition of the organisms into individual biomass building blocks (BBB). lumpGEM source code is available in S3 File.
In lumpGEM, we introduce and use the following definitions:
Postulate 1: Binary control is unbiased to reaction directionality. This means that RiGEM that is controlled by zrxn, i can operate in both directions if the existing constraints (mass balance and thermodynamics) allow it.
Postulate 2: Any Mcore is a potential precursor for the biosynthesis of BBBj.
Postulate 3: Maximizing for the sum of zrxn, i results in the smallest subnetwork Sj to produce BBBj from Mcore. This subnetwork is not necessarily composed of only linear pathways as reported in databases such as KEGG [53], SEED [54] or EcoCyc [40] etc. and may include branches.
Postulate 4: The flux distribution for each generated subnetwork cannot guarantee an optimum flux distribution that will specify the individual stoichiometric contribution of each Mcore to synthesize BBBj due to the degrees of freedom (DOF) that the system has.
Moreover, only the defined core reactions and metabolites of the GEM built in Step f for generating subnetworks is constraint with thermodynamics, and the generated subnetworks are constrained by only mass-balance.
To test the thermodynamic feasibility of these subnetworks, we have built the following MILP formulation for each Sj:
Postulate 5: Minimizing the sum of net flux in Sj generates a stoichiometrically proportional flux distribution in the subnetwork Sj. This leads to the exact stoichiometric expenditure of each Mcore to synthesize BBBj.
To identify alternative subnetworks for BBBj, GEM is further constrained with the following integer cuts constraint after generating each Sj with an iterative manner [33].
Postulate 6: Since Rksub is active if only zRksub=0, the next solution will have at least 1 different reaction from the previous solution. Aftermath, the same procedure is applied for the newly generated Sj,2.
|
10.1371/journal.pntd.0000492 | Trachoma in Western Equatoria State, Southern Sudan: Implications for National Control | Trachoma is thought to be common over large parts of Southern Sudan. However, many areas of the country, particularly west of the Nile, have not yet been surveyed. The aim of this study was to confirm whether trachoma extends into Western Equatoria State from neighboring Central Equatoria, where trachoma is highly prevalent, and whether intervention with the SAFE strategy is required.
Population-based cross-sectional surveys were conducted using a two-stage cluster random sampling method to select the study population. Subjects were examined for trachoma by experienced graders using the World Health Organization (WHO) simplified grading scheme. Two counties thought to be most likely to have trachoma were surveyed, Maridi and Mundri. In Maridi, prevalence of one of the signs of active trachoma (trachomatous inflammation-follicular (TF)) in children aged 1–9 years was 0.4% (95% confidence interval (CI), 0.0%–0.8%), while no children showing the other possible sign, trachomatous inflammation-intense (TI), were identified. No trachomatous trichiasis (TT) was found in those aged under 15, and prevalence was 0.1% (95% CI, 0.0%–0.4%) in those aged 15 years and above. In Mundri, active trachoma was also limited to signs of TF, with a prevalence of 4.1% (95% CI, 1.4%–6.9%) in children aged 1–9 years. Again, no TT was found in those aged under 15, and prevalence in those aged 15 years and above was 0.3% (95% CI, 0.0%–0.8%).
Trachoma prevalence in the east of Western Equatoria State is below the WHO recommended intervention threshold for mass drug administration of antibiotic treatment in all villages. However, the prevalence of TF and TT in some villages, particularly in Mundri County, is sufficiently high to warrant targeted interventions at the community level. These results demonstrate that trachoma is not a major public health problem throughout Southern Sudan. Further studies will be required to determine trachoma prevalence in other areas, particularly west of the Nile, but there are presently no resources to survey each county. Studies should thus be targeted to areas where collection of new data would be most informative.
| Baseline data on trachoma prevalence is a prerequisite for intervention. Prior to the present study, all surveys in Southern Sudan reported trachoma prevalences that exceeded the threshold for large-scale intervention. This gave rise to the notion that the disease may be endemic throughout the country. The present study was conducted under the auspices of the National Program for Integrated Control of Neglected Tropical Diseases, to verify whether prevalences in two counties west of the Nile exceeded the WHO recommended intervention threshold for mass drug administration (MDA) of antibiotic treatment. The results show that trachoma prevalence at county level was below this threshold. However, prevalences in some communities within the county were above the threshold, meaning that they should be targeted with MDA of antibiotics, as well as with other interventions such as trichiasis surgery, health promotion and improved water and sanitation. This finding reminds us of the need for geographical targeting of resources, both for surveys and subsequent intervention. Current resources are insufficient to conduct population-based prevalence surveys for trachoma throughout Southern Sudan. Further surveys should thus be conducted in areas where collection of additional information will be most informative. We propose that a combination of risk-mapping and rapid assessments is used to identify such areas.
| Trachoma is caused by ocular infection with the obligate intracellular bacterium Chlamydia trachomatis. Ocular Chlamydia is spread through contact with eye discharge from the infected person and through transmission by eye-seeking flies [1]. The disease is associated with poor personal and environmental hygiene, in particular limited access to water and sanitation, overcrowding and poor socioeconomic conditions.
Trachoma is the leading infectious cause of blindness, estimated to be responsible for 3.6% of blindness worldwide [2]. It is endemic in 56 countries, mainly in poor rural areas, including parts of Central and South America, many African countries and some countries in the Eastern Mediterranean [3]. However, there is a lack of information from some major populations, including large parts of Southern Sudan, which remains an important obstacle to estimating the disease burden and to the implementation of control efforts [4]. The World Health Organization (WHO) has identified a need for more trachoma data, and it is recognized that such data are necessary for implementation of the SAFE strategy: Surgery for trichiasis, Antibiotics to treat infection, and Facial cleanliness and Environmental improvement to reduce transmission [5].
Trachoma has long been known to be prevalent in parts of Sudan [6],[7], but comprehensive data on distribution and burden particularly in Southern Sudan continue to be limited. To generate baseline data, the Carter Center and the Ministry of Health, Government of Southern Sudan (MoH-GoSS), have jointly conducted prevalence surveys in thirteen sites covering a large geographical area mostly to the east of the river Nile [8],[9],[10],[11]. In all of these locations the average prevalence of active trachoma (trachomatous inflammation-follicular (TF) and/or trachomatous inflammation-intense (TI)) in children aged 1–9 years was found to be well above the 10% threshold recommended by WHO for large-scale SAFE intervention [5],[11],[12]. Ngondi and colleagues have used prevalence data from Upper Nile and Jonglei to estimate that in these two States alone 3.9 million people need antibiotic treatment and 206,000 people are in need of immediate trichiasis surgery [9]. These estimates are based on the assumption that trachoma prevalence is homogenous over large areas, which remains to be confirmed for Southern Sudan.
In common with other neglected tropical diseases (NTDs) endemic to Southern Sudan, there is a need to conduct additional surveys to better understand the epidemiology of trachoma and identify areas requiring interventions [13]. In Southern Sudan, a National Program for Integrated Control of NTDs has recently been established, presenting a new opportunity to contribute towards trachoma control by integrating annual distribution of antibiotic treatment with mass drug administration (MDA) of preventive chemotherapy (PCT) for other common NTDs, namely onchocerciasis, lymphatic filariasis (LF), soil-transmitted helminth (STH) infection and schistosomiasis [14],[15]. The program has started to operate in geographic areas where community-directed treatment with ivermectin (CDTI) for onchocerciasis control has been conducted for a number of years, because it intends to use the CDTI approach for co-implementation of other interventions where feasible [16],[17].
Western Equatoria State is one of the two States that has been selected for initial integrated NTD intervention, because it has a large onchocerciasis focus and a well-established CDTI network, and anecdotal and past survey data indicate that LF, STH infection and schistosomiasis are also prevalent [12]. To date there are, however, no information or data available for this State on the distribution or prevalence of trachoma. The present study was therefore conducted to generate baseline data on the prevalence of trachoma in parts of Western Equatoria State to provide evidence as to whether annual MDA of PCT should include antibiotic treatment, and to contribute data to revise mapping of the burden of trachoma in the region [4].
During November 2008, two population-based prevalence surveys were conducted in Western Equatoria State, which lies in the South-West of Southern Sudan (Figure 1). The majority of people are agriculturalist, growing maize, cassava, groundnut, and fruit. The state capital is Yambio, and the other major towns are Tambura, Nzara, Maridi and Mundri. The total population of Western Equatoria State was estimated to be 845,989 in 2008, using data collected during National Immunization Day. This is approximately 8% of the total population of Southern Sudan, which is estimated to be around 10 million. One trachoma survey was conducted in Mundri county and the other in Maridi county (Figure 1). At the time of the surveys, 186,668 people were estimated to live in Mundri and 186,830 in Maridi; both counties consisted of six payams.
The study followed the standard MoH-GoSS protocol for trachoma prevalence surveys [18]. The protocol recommends that population-based prevalence surveys are conducted at county (rather than payam) level, which is the second (rather than the third) administrative level; the State being the first administrative level in Southern Sudan. Estimation at county rather than payam level was thought to be consistent with WHO guidelines, which recommend trachoma prevalence be estimated at district level or an administrative area corresponding to an average population size of 100,000 [19],[20].
It was estimated that in each county a total sample size of 2000 people (of all ages and sexes) was required. This allows for an estimated prevalence of 5% trachomatous trichiasis (TT) in adults aged 15 years and above (chosen because TT was likely to be the least prevalent indicator measured) within a precision of 2%, given a 95% confidence limit and a design effect of 2, and based on the assumption that adults aged 15 years and above comprise 50% of the population.
In each county 20 villages were sampled with probability proportional to the estimated population size of the payam, although not all of the payams in Maridi were accessible. Households were randomly selected using the sketch map and segmentation method [21]. All residents of the household were enumerated, and all those present who gave informed consent were examined.
Ophthalmic Clinical Officers and General Clinical Officers from the local payams were trained by an experienced ophthalmologist (K. Lewis) to use the WHO simplified grading system [22]. This scheme categorizes trachoma infection according to five grades: TF, TI, trachomatous scarring (TS), TT and corneal opacity (CO). Two stages of assessment were used to select the best trainees. In the first stage, trainee examiners identified trachoma grades using the WHO set of trachoma slides [22]. Those examiners who achieved at least 80% agreement then proceeded to the second stage of field evaluation. During field evaluation, a reliability study comprising 50 persons of varying age and sex were selected by the ophthalmologist to represent all trachoma grades. Each trainee examiner evaluated all 50 participants independently and recorded their findings on a pre-printed form. Inter-observer agreement was then calculated for each trainee using the ophthalmologists' observation as the “gold standard.” Only trainees achieving at least 80% inter-observer agreement after the field evaluation were included as graders.
All inhabitants of selected households who provided verbal consent were examined using a torch and a 2× magnifying binocular loupe. Each eye was first examined for in-turned lashes (TT), and the cornea was then inspected for CO. The upper conjunctiva was subsequently examined for inflammation (TF and TI) and scarring (TS). Both eyes were examined. Signs had to be clearly visible in accordance with the simplified grading system in order to be considered present. Trachoma signs only had to be present in one eye for the person to be categorized as suffering from a particular grade of trachoma. Alcohol-soaked cotton swabs were used to clean the examiner's fingers between examinations. Individuals with signs of active trachoma or bacterial conjunctivitis were treated with 1% tetracycline eye ointment and provided with information on face washing and good hygiene practices. Patients with TT or other significant eye conditions were referred to the nearest facility where free surgery is available (i.e. Juba Teaching Hospital).
The data was initially entered using Personal Digital Assistants (PDA, Palm Tungsten E2) in the field. A second data entry was conducted by the Trachoma Control Program, MoH-GoSS using Microsoft Office Excel. Consistency checks were performed in EpiInfo version 3.2.2 (Centers for Disease Control and Prevention [http://www.cdc.gov/EpiInfo]). Range and consistency checks were conducted for all variables. Data were analyzed in STATA 9.0 software (Stata Corporation, College Station, TX, U.S.A.). Individuals with missing data on sex and/or age were excluded from the analysis.
For each county, prevalences of trachoma signs were summarized by age, in relation to WHO recommendations for implementation of trachoma control activities [5]. Unadjusted exact binomial 95% confidence intervals (CIs) are presented, along with adjusted 95% CIs that account for potential clustering obtained using generalized estimating equation (GEE) modeling. Chi-squared tests, or Fishers test where appropriate, were used to examine evidence for differences in proportions.
The study protocol received ethical approval from the Directorate of Research, Planning and Health System Development, MoH-GoSS. Clearance to conduct the surveys was obtained from the State MoH, followed by County Health Departments and the local government. The study was explained to each member of the selected households. The household heads were asked to provide written consent for the entire household to participate in the study, and each inhabitant of the household who provided verbal consent was examined. Those individuals who did not provide verbal consent were not examined. Personal identifiers were removed from the dataset before analysis.
In the 20 study villages in each county, the number of households sampled and individuals examined were similar (Table 1). The average household size in Mundri was larger (p<0.001). There were three missing values for sex in Mundri. Age data were missing for approximately 10% of individuals in each county. There was no difference in the sex distribution by county (p = 0.672), whereas age distribution did differ by county (p<0.001), with more children aged 1–9 years and fewer people aged 15 or above in Maridi, compared to Mundri.
Consent refusal was very low; 48 individuals (2.8%) refused overall, of which 16 (2.5%) were in Maridi and 32 (2.9%) in Mundri, although some enumerated individuals were subsequently unavailable for examination. Approximately 90% and 80% of children aged 1–9 from study villages in Maridi and Mundri were examined for signs of active trachoma, respectively. Examination data for TS were available for 80% of individuals in Maridi and 70% in Mundri. Of those aged 15 and above, TT and CO examinations were conducted in 75% of individuals in Maridi and 65% in Mundri.
The overall prevalence of TF (adjusted 95% CI) in children aged 1–9 years in Maridi was 0.4% (0.0–0.8%) with cases occurring only in male children (Table 2). In Mundri, the prevalence of TF (adjusted 95% CI) in children aged 1–9 years was 4% (1.4–6.9%), with similar prevalence in male and female children. No cases of TI were found in either county, so prevalence estimates for active trachoma are the same as for TF.
For TT, the overall prevalence (adjusted 95% CI) in those aged 15 years and above in Maridi was estimated to be 0.1% (0.0–0.4%) based on just one female individual with TT (Table 2). In Mundri, three male individuals were found to have TT leading to a prevalence estimate of 0.3% (0.0–0.8%). The small number of individuals with TT also had some degree of corneal opacity. No trichiasis was found in individuals under the age of 15 in either county.
Little evidence of scarring was observed in either county, the highest prevalence (adjusted 95% CI) was in individuals aged 15 and over in Mundri; 0.9% (0.2–1.5%).
No villages in Maridi had a prevalence of TF above 5%, but one village did have a prevalence of TT of more than 1% (Table 3, Figure 1). In Mundri, two villages had a prevalence of TF between 5% and 9% and a further three villages had a prevalence of more than 10% (Table 3, Figure 1), suggesting a requirement for F and E components of the SAFE strategy in these five villages along with annual MDA of antibiotics in the three villages with higher prevalence. Two additional villages in Mundri had a prevalence of TT of more than 1% (Figure 1).
This study was conducted within the context of scaling up of the National Program for Integrated Control of NTDs, to decide whether antibiotic treatment for trachoma should be included in annual MDA of PCT for other NTDs in Western Equatoria State. Based on consultation with State and County Health Authorities, two counties were selected to represent those areas of the State thought to be worst affected by trachoma. Despite purposefully selecting these areas, we found that the overall burden of active trachoma in children aged 1–9 years was below the 10% threshold recommended by WHO for large-scale MDA of antibiotic treatment. Similarly, the overall prevalence of TT in people aged 15 years or over was below the recommended 1% threshold for large-scale SAFE intervention. However, some study villages in Mundri County did have TF prevalences that exceeded 5%, which warrants community-wide intervention by means of annual MDA of antibiotic treatment, health promotion and improvements of water and sanitation [5]. The three study communities with cases of trichiasis will need to be provided with access to surgery. It is likely that other communities, particularly in Mundri County, are also affected by active trachoma and/or TT and need to be provided with some or all components of the SAFE strategy. Community-by-community assessments will be needed to identify these [5].
The relatively low trachoma prevalences observed in Maridi and Mundri are in stark contrast to those reported by investigators who surveyed other parts of Southern Sudan, including two sites (Tali and Katigiri) adjacent to Mundri that were found to have 72.6% and 50.0% prevalence of active trachoma in children aged 1–9, respectively [9]. In fact, in all other sites previously surveyed throughout Southern Sudan, prevalence of active trachoma exceeded the 10% intervention threshold by up to eight times [8]–[11]. Present findings thus provide an important contribution to our understanding of trachoma epidemiology in Southern Sudan, showing that the disease is not highly endemic throughout the country and that prevalence can vary considerably between adjacent counties. Based on this observation it would be unwise to conclude from our data that the whole of Western Equatoria State has a low prevalence of trachoma, but given that we purposefully selected areas that were thought to be worst affected, we may assume this is the case. To verify this assumption and to decide whether further population-based prevalence surveys are required in this State and elsewhere, we propose that more evidence on the distribution and burden of trachoma is collected through a combination of risk-mapping, using a global information system (GIS), and rapid assessments [23]. For the purpose of risk-mapping, existing trachoma prevalence data for Southern Sudan, as well as environmental and other potential risk factors [23], should be incorporated into a model to predict areas at risk of trachoma transmission. Although GIS has so far not been used to predict the spatial distribution of trachoma, the approach has informed targeting of a number of other NTD interventions, such as schistosomiasis [24]–[26]. By taking account of the specific epidemiological characteristics of trachoma, it may be possible to develop a GIS-based model that could complement the existing methods for identifying trachoma endemic areas [27].
The challenging operating conditions in Southern Sudan meant that survey implementation deviated slightly from the approach recommended in the national trachoma survey protocol [18]. Due to a lack of security not all of the payams in Maridi were accessible. In both counties we had to exclude some of the randomly selected village, because they were inaccessible due to flooding or impassable roads, or we found that they were more than a 45-minute walk from the nearest vehicle access point. Furthermore, some villages that were recorded on the maps developed by the United Nations Joint Logistics Coordination and International Mapping Unit no longer existed. As we had used these maps to select villages, it meant that not all of the sites from our initial random selection could be surveyed. If a village was known to be inaccessible or non-existent in advance, an alternative site was randomly selected. If the village was found to be inaccessible or non-existent on arrival, an alternative site close to that geographic location was selected.
This slight deviation from a true random selection of villages may have introduced bias, thus affecting our estimation of overall trachoma prevalence in a county. For example, overall prevalence may have been underestimated if the villages inaccessible at the time of the survey had poorer access to healthcare, so that residents were less likely to have been treated for trachoma. However, our field observations indicated that healthcare provision was uniformly poor across the survey area, with no major variations in access or quality between villages. Healthcare delivery largely consisted of community health workers with little equipment or training. Because few of the randomly selected villages were excluded, we consider it unlikely that any significant bias was introduced to the overall estimate of trachoma prevalence for either county.
Of the individuals living in households recruited into the study, the overall percentage examined was 68.9% in Mundri and 76.2% in Maridi. This rate was somewhat less than that reported by other investigators, who examined 85%–90% of registered household occupants during surveys in other parts of Southern Sudan [10],[11],[28]. The main reason for non-response was that individuals were not present at the time of the survey; unfortunately it was not possible to follow these up due to the difficult operating conditions. The majority of those absent were men aged 15 years and above, meaning that significantly more females than males in older age groups were examined in both counties. Such absenteeism may have biased our overall findings, possibly resulting in an overestimate of trachoma prevalence if, for example, males with good eyesight had been more able to leave home. We were unable to investigate this further, because trachoma was not sufficiently endemic to make a valid comparison of prevalence between males and females. Lastly, we had to exclude some incomplete eye examinations from the analysis, as some individuals withdrew their consent after examination of the first eye. This was more common in children, and might have led to an underestimation of TI and TF infection, particularly if infected individuals experienced more discomfort during examination and were thus more likely to withdraw their consent.
Despite these operational limitations and their potential implications for the precision of our estimates, the low trachoma prevalence found clearly shows that annual MDA with antibiotics does not need to be targeted at all villages in the two counties, but only at a selection within these. This means that integration of trachoma treatment as part of annual MDA for the other diseases targeted by the National Program for Integrated Control of NTDs would not be practical.
There are a number of possible explanations for the difference in the results of the present study when compared to those of other investigators. Firstly, previous surveys on trachoma in Southern Sudan were conducted in areas where health workers had already reported trachoma to be a cause of blindness. In contrast, the selection of the present study sites was not based on such reports, but on the assessment of State and County health staff as to which areas are worst affected by trachoma, in a State that generally experiences little demand for trachoma treatment services. This difference in the selection of the survey area will have affected our results. Whereas investigators of previous studies had good evidence that trachoma was endemic before they undertook the surveys, there was no indication that it constituted a major public health problem in Mundri or Maridi.
The second difference between the present study and those previously conducted in Southern Sudan is that we used different survey staff. The graders used in the present study may have under-graded symptoms, when compared to grading done in previous studies, which in turn may have led to systematic bias in clinical grading. This could explain why there were seemingly no cases of active trachoma in villages with TT and vice versa (Figure 1). Though we can not exclude this possibility, it may also be possible that these findings are the result of extensive population movement in Southern Sudan, with the few diagnosed TT cases maybe having moved relatively recently from trachoma endemic areas to these three villages. It is unlikely that the absence of active cases in villages with TT is a result of recent socioeconomic development and associated improvements in access to safe water and proper sanitation, as such development is proceeding very slowly throughout the country.
To minimize the possibility of under- or over-grading of trachoma symptoms, we ensured that the clinical scoring conducted by our graders was consistent and agreed with a standardized set of photos. We also assessed the feasibility of verifying the graders findings by taking photographs and/or samples for analysis by nucleic acid amplification tests as suggested elsewhere [29], but established that both methods were beyond what was feasible with existing resources and in an environment with no infrastructure. For future trachoma studies, the possibility of implementing either or both of these two complementary approaches should be reassessed and budgeted for.
The third potential variation between the current and previous surveys is a difference in the livelihoods between the populations sampled, as well as their hygiene behavior. A detailed analysis of trachoma risk factors in the previously sampled populations [30] showed that cattle ownership was common (69.2% overall), and increased the risk of trachoma infection, along with other factors such as unclean face, face washing less than twice a day and high household fly density. The geographical area visited during the present surveys, however, is known to be largely inhabited by agriculturalists. This was confirmed in our household surveys, which showed that 86.7% and 75.3% of households in Maridi and Mundri, respectively, were mainly involved in agriculture (data not shown). The absence of large amounts of livestock and their dung might have resulted in lower fly densities and hence less transmission. In addition, Ngondi and colleagues [30] reported anecdotal evidence that people keeping cattle also dispose of human feces in the cattle pens, hence providing an even more favorable breeding environment for flies. As we did not investigate behavior in sufficient detail, this and other potential differences in daily hygiene and sanitation behavior between agriculturalist and pastoralist tribes in Southern Sudan remains hypothetical. Further sociological studies would help to better understand such differences and their potential implications on trachoma transmission [30].
Above explanations are likely to account for only part of the differences between study sites in Southern Sudan. The importance of geographical, climatic and socio-economic determinants, for example, warrants further investigation. Their possible association in Southern Sudan has not been sufficiently researched, although an earlier study in the Sudan provides evidence for an inverse correlation of TF/TI prevalence with humidity and rainfall [7].
The low prevalence of trachoma observed in the present study and the lack of reported cases from other counties in Western Equatoria State suggest that trachoma is also unlikely to be highly endemic in those counties not surveyed to date. Although this needs to be verified, we do not consider this a current priority. Instead, the limited resources available for surveys should be used to generate baseline data for counties where trachoma is known to be a problem, but where there is no data to monitor interventions. Among these, counties where intervention is feasible (in terms of access and funding) should be prioritized.
In addition, it will be important to complete the picture of trachoma epidemiology in Southern Sudan, to get a better understanding of the scale of the burden and the resources required to eliminating blinding trachoma from Southern Sudan by 2020. In the interest of conserving scarce resources, we propose that trachoma distribution is assessed in stages. In the first stage, the large amount of existing prevalence data should be used to develop a trachoma risk map for the whole of Southern Sudan, which will not only indicate where trachoma is certain to be highly endemic but also where prediction of endemicity is rather uncertain. Based on this information, rapid assessments should then be conducted in those areas with no data and high uncertainty of the level of endemicity. Based on the outcome of such assessments it can be decided whether one should move to the third stage - detailed prevalence surveys - to provide baseline data. In an iterative process, the data generated through rapid assessments and additional prevalence surveys should be used to verify the predictions of the risk map model, and fine tune and improve these, including estimates of the national trachoma burden.
|
10.1371/journal.pgen.1004155 | Genome-Wide Analysis of SREBP1 Activity around the Clock Reveals Its Combined Dependency on Nutrient and Circadian Signals | In mammals, the circadian clock allows them to anticipate and adapt physiology around the 24 hours. Conversely, metabolism and food consumption regulate the internal clock, pointing the existence of an intricate relationship between nutrient state and circadian homeostasis that is far from being understood. The Sterol Regulatory Element Binding Protein 1 (SREBP1) is a key regulator of lipid homeostasis. Hepatic SREBP1 function is influenced by the nutrient-response cycle, but also by the circadian machinery. To systematically understand how the interplay of circadian clock and nutrient-driven rhythm regulates SREBP1 activity, we evaluated the genome-wide binding of SREBP1 to its targets throughout the day in C57BL/6 mice. The recruitment of SREBP1 to the DNA showed a highly circadian behaviour, with a maximum during the fed status. However, the temporal expression of SREBP1 targets was not always synchronized with its binding pattern. In particular, different expression phases were observed for SREBP1 target genes depending on their function, suggesting the involvement of other transcription factors in their regulation. Binding sites for Hepatocyte Nuclear Factor 4 (HNF4) were specifically enriched in the close proximity of SREBP1 peaks of genes, whose expression was shifted by about 8 hours with respect to SREBP1 binding. Thus, the cross-talk between hepatic HNF4 and SREBP1 may underlie the expression timing of this subgroup of SREBP1 targets. Interestingly, the proper temporal expression profile of these genes was dramatically changed in Bmal1−/− mice upon time-restricted feeding, for which a rhythmic, but slightly delayed, binding of SREBP1 was maintained. Collectively, our results show that besides the nutrient-driven regulation of SREBP1 nuclear translocation, a second layer of modulation of SREBP1 transcriptional activity, strongly dependent from the circadian clock, exists. This system allows us to fine tune the expression timing of SREBP1 target genes, thus helping to temporally separate the different physiological processes in which these genes are involved.
| Circadian rhythmicity is part of our innate behavior and controls many physiological processes, such as sleeping and waking, activity, neurotransmitter production and a number of metabolic pathways. In mammals, the central circadian pacemaker in the hypothalamus is entrained on a daily basis by environmental cues (i.e. light), thus setting the period length and synchronizing the rhythms of all cells in the body. In the last decades, numerous investigations have highlighted the importance of the internal timekeeping mechanism for maintenance of organism health and longevity. Indeed, the reciprocal regulation of circadian clock and metabolism is now commonly accepted, although still poorly understood at the molecular level. Our global analysis of DNA binding along the day of Sterol Regulatory Element Binding Protein 1 (SREBP1), a key regulator of lipid biosynthesis, represents the first tool to comprehensively explore how its activity is connected to circadian-driven regulatory events. We show that the regulation of SREBP1 action by nutrients relies mainly on the control of its subcellular localization, while the circadian clock influences the promoter specific activity of SREBP1 within the nucleus. Furthermore, we identify the Hepatocyte Nuclear Factor 4 (HNF4) as a putative player in the cross-talk between molecular clock and metabolic regulation.
| Mammals possess an internal circadian clock which allows them to anticipate and adapt to daily environmental changes [1]. The molecular mechanism underlying the cell-autonomous circadian rhythms relies on a network of feedback loops in which BMAL1, CLOCK, Neuronal PAS domain (NPAS) protein 2 and Retinoic acid receptor-related Orphan Receptor (ROR) proteins act as transcriptional activators and period homolog proteins (PER1, 2 and 3), cryptochromes (CRY1 and 2) and REV-ERBs function as inhibitors producing the self-sustained oscillating production of their target genes, including themselves. At central level, the expression of clock genes is dictated by a pacemaker localized in the hypothalamic suprachiasmatic nucleus (SCN) which synchronizes the phase in nearly all body cells. However, in peripheral organs such as the liver, oscillations are also entrained by the feeding and fasting cycle [2], [3], [4]. This sophisticated regulatory system contributes to coordinate many physiological processes, such as sleep-wake cycles, locomotor activity, body temperature, hormone secretion and energy metabolism that all display circadian rhythms. In particular, the importance of the connection between circadian clock and metabolism regulation is emerging. Epidemiological studies have shown an increased incidence of obesity, diabetes, and cardiovascular disease, in addition to certain cancers and inflammatory disorders in night workers [5]–[7]. Accordingly, in genetic mouse models the disruption of the clock alters metabolic homeostasis at different levels (reviewed in [8]), suggesting a still unresolved relationship between nutrient state and circadian homeostasis.
The Sterol Regulatory Element Binding Protein 1 (SREBP1), a basic Helix-Loop-Helix-Leucine Zipper (bHLH-LZ) transcription factor, plays a key role in the regulation of lipid biosynthesis, which is one of the most feeding-related function in the liver [9]. SREBP1 is synthesized as an inactive precursor, anchored to the ER-membrane and its N-terminal fragment is released into the nucleus after proteolytic cleavage in response to cholesterol depletion [10] or to activation of the insulin signalling pathway [11], [12]. Two SREBP1 isoforms, 1a and 1c, are obtained through alternative splicing of the same gene [13], [14]. The liver expresses mostly the SREBP1c isoform that mediates the insulin-driven lipogenic activity [15]. Besides being under the control of the feeding-fasting cycle, SREBP1 translocation to the nucleus is also influenced by one of the master clock regulators, REV-ERBα [16]. Nevertheless, in absence of a functional clock, such as in cry1−/−;cry2−/− mice, a normal expression pattern of several SREBP1 target genes can be restored by an imposed rhythmic food intake [4], suggesting a dominant role of the feeding-fasting cycle in the regulation of SREBP1.
To systematically understand how the interplay of circadian clock and nutrient-driven rhythm regulate SREBP1 activity, we evaluated the genome-wide binding of SREBP1 to its targets along the day in wild-type mice. Our results define SREBP1 binding pattern in the physiological context of both rhythmic food absorption and circadian rhythm and they give the first tools to comprehensively explore how SREBP1 activity is connected to circadian-driven regulatory events.
To evaluate the genome-wide dynamics of SREBP1 binding to its target sites in a physiological context, we prepared liver chromatin from C57BL/6 mice, collecting samples each 4 hours during one day (see Material and Methods). ChIP-seq with an antibody that recognizes both SREBP1 isoforms was performed at each time point. The SREBP1 antibody was tested extensively (Figure S1) and has also been used in previous studies [17]. We obtained an average of 38 millions sequence reads by time point by ultra-high-throughput sequencing (Table S1). The mapping allowed the identification of 448 bona fide SREBP1 binding peaks, above the background. As shown in Figure 1A, the binding of SREBP1 is overall oscillatory, with the maximum for most of the sites at Zeitgeber Times (ZT) 14 or 18 (light is on at ZT0 and is off at ZT12). To systematically evaluate the rhythmicity of SREBP1 recruitment to its targets, a cosine function was fitted to the temporal profile of the binding (see Material and Methods). This allowed to calculate, for each peak, the binding phase and the amplitude of the oscillation together with its associated P-value, as exemplified for the two sites found on the Srebp1 gene itself (Figure 1B). 53% of SREBP1 binding sites were found to be rhythmic (P<0.1 for the amplitude). Four clusters of targets were clearly distinguishable based on binding kinetics (Figure 1A), the first with a phase distributed around ZT15–ZT17, whereas the other ones with the phase peaking around ZT11–ZT12, as determined by cosine function (Figure 1C). The observed kinetics of SREBP1 binding, especially for cluster A peaks, was consistent with its gene expression and nuclear localization, (Figure 1D, 1E and 1F). Collectively, these results show that the activity of SREBP1 oscillates with a pronounced circadian rhythm, in agreement with the previously reported daily variations of its RNA and protein levels [16], [18]–[20].
SREBP1 binding sites identified in this study are grouped in four clusters with a slightly shifted phase. To better investigate the features of these sites we calculated the number of nucleotides spanned by each peak and found that sites belonging to cluster A (236 out of 448) were narrow, with a typical length of about 200 nucleotides (Figure 2A). These peaks were also closer to the nearest annotated transcription start site (TSS) than peaks belonging to clusters B, C or D, whose distance to the nearest annotated TSS roughly matches randomly picked genomic locations (Figure 2B). Moreover, the amplitude of the binding oscillation along the 24 hours was greater for cluster A peaks (Figure 2C). These observations suggest that cluster A sites may be more relevant in the regulation of transcription mediated by SREBP1. In agreement with this hypothesis, a MEME [21] motif search analysis clearly identified the canonical SREBP1 consensus motif in more than 60% of the sites belonging to cluster A (Figure 2D), but only in 6% of the sites assigned to clusters B, C and D. Within cluster A, the MEME analysis also showed that motifs for SP1 and NFY, two transcription factors (TFs) known to cooperate with SREBP1 to regulate the transcription of its target genes, were overrepresented [17]. The consensus motif for the Hepatic Nuclear Factor 4 (HNF4) was also identified in 61 out of the 236 cluster A sites. The discovered motifs were enriched in cluster A peaks with an empirical p-value<0.001, as shown in Table S2. In addition, in the regions belonging to cluster B, C and D, we determined only several highly repetitive sequences as top-scoring motifs (for example ACACACACA in 73 sites out of 212) that could not be associated to any known consensus motif for TFs. This result suggests that SP1, NFY and HNF4 may participate to SREBP1-mediated transcriptional regulation and further supports the functional importance of SREBP1 binding sites assigned to cluster A. Thus, we opted to focus the following analyses on these regions, although we cannot exclude that the other sites might contribute to mediate SREBP1 activity in mouse liver potentially through genome loops.
To explore the cellular processes that are regulated by SREBP1 along the day, we annotated each site with the nearest Ensembl transcript. We used DAVID [22], [23] to identify clusters of genes enriched with functional annotations. As expected, we identified lipid biosynthetic processes and fatty acid metabolism as the most prominent pathways controlled by SREBP1 (Table S3). In addition, we found a significant enrichment of genes involved in carbohydrate metabolism, in the response to nutrient levels, in mitochondrial and endoplasmic reticulum functions and in coenzyme metabolism. In a previous genome-wide study performed in human HepG2 cells, it was shown that unique combinations of SREBP1, SP1 and NFY target distinct functional pathways [17]. Since we found a good enrichment within the SREBP1 binding sites of the consensus motifs for NFY and SP1, but also HNF4, we explored whether a network among these three transcription factors could be highlighted in mouse liver. As shown in Table 1, genes involved in lipid biosynthesis and in the regulation of fatty acid and steroid metabolism were highly represented in all categories. In some cases, however, one biological function was targeted by a unique combination of regulators. For example, the biosynthesis of coenzymes was selectively represented within the genes bearing only the site for SP1, whereas both SP1 and HNF4 motifs were present in genes involved in apoptosis. Likewise, a combination of HNF4 and NFY motifs marked most of the genes involved in immunological processes. Finally, pathways related to carbohydrate metabolism and mitochondria were particularly enriched in genes without NFY or SP1 motifs, suggesting that SREBP1 may cooperate with other regulators at the promoter of these genes (the complete functional annotation clustering is reported in Table S4). Our results suggest that in mouse liver, in physiological conditions, the network SREBP1-SP1-NFY-HNF4 may be important in order to determine the functional effect of SREBP1 binding (Figure 2E).
In Figure 1, we showed that SREBP1 binds to target sites belonging to cluster A with a sharp phase between ZT15 and ZT17. To investigate the functional effects of SREBP1 binding on gene transcription, we checked in our previously reported data set [24] the 24 hours profile of RNA polymerase II (Pol II) recruitment in the proximity of SREBP1 target genes. Importantly, most of the SREBP1 binding sites belonging to cluster A were co-occupied by Pol II (Figure 1A), further supporting the functional relevance of these regions. We next evaluated Pol II binding to the promoter and in the gene body of all putative SREBP1 target genes. In parallel, we measured mRNA levels of the same genes by microarray analysis (Table S5). More than 85% of SREBP1 target genes show an expression level above the median expression level of all the transcripts, suggesting that they are transcribed. Our analyses revealed three clusters of target genes, that we called A1, A2 and A3, with distinct temporal profile of transcription and expression (Figure 3A and Figure S2). In cluster A1, the peak of Pol II binding was concomitant, or even slightly earlier than SREBP1 binding. In contrast, for genes belonging to cluster A2, Pol II association to both promoter and gene body strictly followed SREBP1 binding. Lastly, Pol II recruitment to the genes of the A3 group was shifted by about +8 h with respect to SREBP1. For all clusters, the temporal profile of gene expression was consistent with the dynamics of Pol II association. The distribution of all the expression phases obtained for the genes belonging to the three groups confirmed that SREBP1 target genes are expressed in different moments of the day, in spite of the concomitant binding of the transcription factor (Figure 3B). This observation suggests that other factors participate in the regulation of the various SREBP1 target genes in order to assure their appropriate expression timing. Interestingly, genes that were mainly expressed during the fed state (clusters A1 and A2) were functionally enriched in the regulation of lipid and coenzyme biosynthetic processes, as well as in the response to hormones, such as insulin. In contrast, SREBP1 target genes involved in mitochondrial oxidation and apoptosis were enriched during the fasting period (Table 2). Thus, the promoter specific events that determine the different temporal expression profile of SREBP1 target genes contribute to define the set of cellular functions that are active at a given time.
To understand the molecular mechanism underlying the different temporal expression of SREBP1 target genes, we first explored the possible involvement of the network SREBP1-SP1-NFY-HNF4 in determining the functional effect of the binding of SREBP1 to its targets. To check for the presence of a pattern characterizing the three groups of SREBP1 target genes identified earlier (see Figure 3), we evaluated the presence of different combinations of SP1 and NFY motifs and their orientation with respect to the SREBP1 binding sites (data not shown). However, we could not establish any significant correlation. In contrast, we found that HNF4 motifs were significantly overrepresented (P-value<0.02) in the regions under SREBP1 peaks of the genes expressed during the fasting period (cluster A3), compared to the other clusters (Table S6). The actual recruitment of HNF4 to these putative binding sites was assessed by ChIP on randomly selected SREBP1 Responsive Elements (SREs) (Figure 3C). Besides HNF4, other transcription factors, such as the cAMP response element-binding protein (CREB) or Forkhead box proteins O (Foxo), are important players in the hepatic metabolic regulation upon fasting [25], [26]. However, their known consensus motifs were not found in the proximity of cluster A3 SREBP1 peaks. These observations strongly suggest a specific cross-talk between HNF4 and SREBP1 in the regulation of these genes. To further investigate which control processes dictate the distribution of SREBP1 target gene expression along the day, we then considered the possible role of the circadian rhythm in this regulation. To test this hypothesis, it was necessary to uncouple the circadian rhythm from the response to nutrients. Thus, we fed mice lacking BMAL1 (Bmal1−/−) only during the darkness period for one week before collecting liver samples every four hours. Upon this experimental conditions, the circadian clock was completely disrupted in Bmal1−/− mice, as demonstrated by the flattened expression of key core and output components of the clock, such as Clock1, Cry1, Cry2, D site albumin promoter binding protein (Dbp), Rev-Erbα and Kruppel-like factor 10 (Klf10) (Figure 4B). Body weight, daily food intake and glycemia were unchanged in Bmal1−/− mice (Figure S3). Importantly, the imposed rhythmic food intake restored an oscillatory nutrient response, as shown by the levels of circulating insulin that were comparable to the wild type (Figure 4A). Accordingly, in Bmal1−/− mice SREBP1 translocated to the nucleus from ZT18 onwards (Figure 4D and 4E) and its expression was still cycling, although the phase was delayed by about 6 h compared to the wild type (Figure 4C). We next checked the dynamics of SREBP1 binding to a panel of the targets previously identified in wild type mice and we found that SREBP1 was recruited to all the tested sites in an oscillatory way, but with an average phase shift of about 4 hours (figure 4F). Finally, to assess the impact of the circadian oscillator impairment on SREBP1-driven transcription, we globally evaluated the expression of SREBP1 target genes in Bmal1−/− mice (Table S5). Most SREBP1 target genes were still scored as oscillating in Bmal1−/− upon temporal restricted feeding (57% have a P<0.05). However, the heatmap rendering of their expression patterns (Figure 5A) revealed a temporal profile that was perturbed in Bmal1−/− compared to WT mice, as most of the genes now had a maximum expression at ZT18, coinciding with the binding of the transcription factor (Figure 5C). Accordingly, the expression phases of the genes belonging to the clusters A1 and A3 identified earlier (Figure 3) were now largely concomitant with those of genes belonging to cluster A2, therefore mostly grouped between ZT14 and ZT24 (Figure 5B and Figure S4). This phase shift was not due to a selective decrease of the number of cycling genes in clusters A1 and A3, as the percentage of the significantly oscillating genes was comparable in the three clusters in WT and Bmal1−/− mice (cluster A1: 84% in WT vs 94% in Bmal1−/−; cluster A2: 71% in WT vs 66% in Bmal1−/−; cluster A3: 62% in WT vs 64% in Bmal1−/− with P<0.05).
To explore how the core clock components participate to this regulation we checked in published data sets whether key transcription factors such as BMAL1, CLOCK1, CRY1, CRY2, PER1, PER2, NPAS2 and REV-ERBs can differentially bind to the promoters of the three clusters of SREBP1 target genes [27]–[29]. Interestingly, we found that most SREBP1 peaks (148 out of 236, ≈63%) have an overlapping REV-ERBα and/or REV-ERBβ peak. In addition, a REV-ERBα/β binding site was detected also in another 17% of the promoters of SREBP1 target genes, but in a non-overlapping position. This strong occupancy of SREBP1 targets by REV-ERBs is consistent with the previously reported involvement of REV-ERBα in the regulation of lipid metabolic genes [30], and suggests the existence, in mouse liver, of a SREBP1-REV-ERBs network in physiological conditions. The frequency of REV-ERBs recruitment was comparable in clusters A1, A2 and A3 (data not shown), thus arguing against the possible role of these nuclear receptors in determining the distinct phase of expression of these genes. However, due to the presence of REV-ERB binding sites in many SREBP1 target genes, the flattened REV-ERB expression observed in Bmal1−/− may perturb, at least in part, the phase of several SREBP1 target genes. Indeed, in WT mice, the temporal expression profile of SREBP1 and REV-ERBs is very different (the phases of expression are ZT15 and ZT8, respectively [16]), and these factors are not expected to compete for binding at the same time to the same genes. The other transcription factors tested were recruited to a lesser extent on SREBP1 target gene promoters and for none of them we observed a significant enrichment in clusters A1, A2 or A3 (data not shown).
Taken together, our results confirm that SREBP1 activity is strongly dictated by the rhythmicity of nutrient intake. In addition, our observations indicate that a functional circadian core clock is necessary to assure the correct temporal expression profile of SREBP1 target genes and suggest a role for HNF4 in dictating the phase of expression of genes whose mRNA levels peak when SREBP1 binding is low. Further studies will aim at understanding whether and how the lack of circadian rhythm perturbs HNF4 activity.
SREBP1 is a highly circadian transcription factor whose activity is strongly regulated by nutrient availability through the insulin signaling pathway. In mouse liver SREBP1 expression displays a daily rhythm with a peak in the nocturnal feeding period under standard housing condition of mice [16], [18]–[20]. In this study we evaluate the dynamics of SREBP1 recruitment to DNA by determining its genome wide cis-acting targets (cistrome) in the liver along an entire day. SREBP1 binds to 448 sites with an oscillatory profile that is temporally coherent with the phase of its maximal expression. Within SREBP1 binding sites, four distinct groups are clearly distinguishable. The first set (cluster A) contains peaks that are likely the more relevant in the transcriptional regulation mediated by SREBP1 as they are the closest to TSS and they are bound more rhythmically by SREBP1. Importantly, in more than 60% of these sites we identified the direct repeat 5′-ATCACCCCAC-3′ that was described as the Sterol regulatory proteins Responsive Elements (SRE) in several promoters, such as the human LDL receptor promoter [13], [31]–[33]. This direct repeat variant of the canonical E-box inverted repeats 5′-CAnnTG-3′ was shown to be specifically recognized by SREBP proteins due to the presence of a tyrosine residue in a position that corresponds to an arginine in all the other bHLH-LZ proteins and that is critical for high affinity contacts with the SRE [34], [35]. Furthermore, our ChIP-seq results highlighted the presence of predicted binding sites for SP1, NFY and HNF4 in 60%, 30% and 25% of cluster A sites, respectively. Since SREBP1c, the major SREBP1 isoform in the liver, is a weak transcriptional activator, this observation is consistent with earlier studies demonstrating that the transcription factors SP1, NFY and CREB cooperate to regulate different SREBP1-responsive promoters [36]–[39]. In 2009, Seo et al have published a list of liver SREBP1 target genes obtained from a genome-wide study of mice subjected to 24 hours fasting followed by 12 hours refeeding with high carbohydrate diet [40], thus creating a condition where a very high SREBP1c activity is expected. In this study, a functional variant of the direct repeat SRE (5′ACTACANNTCCC-3′) was identified as a preferred site for SREBP1 binding, and no enrichment of the predicted NFY binding site was identified. The difference between this data set and ours can most likely be attributed to the difference in the specific experimental conditions, acute challenge on the one hand and physiological condition on the other hand (present study). Consistent with our study, both SP1 and NFY proteins were recruited on more than 30% of SREBP1 target genes in a genome-wide analysis of SREBP1 binding in HepG2 cell line [17]. In addition to SP1 and NFY, here we identify HNF4 as an important player of the interconnected regulatory circuit that may assure the specific regulation of SREBP1 target genes with distinct functions. Interestingly, consensus motifs for SP1, NFY and HNF4 were found to be overrepresented in the promoters of cycling genes in the liver [41], [42]. Since ≈70% of SREBP1 targets show a circadian gene expression, our results are in line with these bioinformatics predictions, supporting the involvement of these transcription factors in the complex transcriptional regulation of circadian rhythm in liver. The second set of SREBP1 target sites falls in the three clusters B, C and D. These peaks are not enriched in regions proximal to TSSs for mapped genes, nor in predicted motifs for known transcription factors. Furthermore the temporal profile of SREBP1 binding to these sites is flattened compared to the first set of SREBP1 target sites (cluster A). Additional studies are required to understand whether these peaks have a functional role, or whether they are bound in a secondary manner by SREBP1 due to the formation of DNA loops.
In the liver the expression of several known SREBP1c target genes is decreased in fasted mice, when the levels of SREBP1 are very low and increased upon refeeding, when both SREBP1 expression and nuclear translocation are induced [40], [43]. Accordingly, our analysis of Pol II recruitment on SREBP1 putative target genes, coupled with the measurement of their mRNA levels, revealed a maximum of transcription and expression during the fed state, namely between ZT12 and ZT24, for the majority of these genes. This is consistent with the binding of SREBP1 to DNA, which is higher at this time of the day. Nevertheless, a large set of SREBP1 target genes (cluster A3) displays a temporal expression profile strongly shifted with respect to SREBP1 recruitment. However, the different phase of expression observed in these genes is coherent with the dynamics of Pol II association to their promoter and gene body, thus arguing against a major involvement of post-transcriptional mechanisms in this delay and suggesting the existence of promoter specific events that determine the different temporal expression profile of SREBP1 target genes. In particular, the low level of expression of genes belonging to cluster A3 when SREBP1 is bound raises the question whether SREBP1 itself, or through interaction with coregulatory proteins, can act as transcriptional repressor for these genes. Interestingly, it was proposed that SREBP1 may act as negative regulator of the cytosolic phosphoenolpyruvate carboxykinase (Pck-1) gene by impairing the recruitment of the transcriptional coactivator Peroxisome Proliferator-Activated Receptors γ Coactivator -1 (PGC-1) on HNF4α [44]. To explain the negative effect of SREBP1 on this gene, a second mechanism was put forward by Chakravarty and colleagues, suggesting an interference between the binding of SREBP1 and SP1, due to the orientation on the opposite DNA strands of the two binding sites [45]. Although we find an enrichment of motifs for NFY and SP1 in the close proximity of SREBP1 peaks from the first set of target sites (clusters A1, A2 and A3), we do not observe a different presence and/or orientation of these sites with respect to SREBP1 peaks, among these three groups of target sites. Conversely, the high frequency of the HNF4 motif in the cluster A3 suggests that the cross-talk between HNF4 and SREBP1 may be a general mechanism through which SREBP1 negatively affects the transcription of a sub-set of its target genes. In agreement with this hypothesis, among the targets of SREBP1 expressed upon fasting we detect the Peroxisome Proliferator-Activated Receptor α (Pparα) gene, that was shown to be crucially regulated by HNF4 [46].
In recent years, growing evidences have highlighted the impact of circadian gene networks on nutrient balance and, on the other hand, the regulation of the circadian clock by metabolism and food consumption [47], [48]. Thus, circadian clock and metabolism converge in numerous ways to control the activity of a number of transcription factors that are essential for maintaining metabolic homeostasis, although the exact contribution of each input remains to be deciphered. Several studies demonstrated that upon restricted feeding (RF), namely when time and duration of food availability is limited in time, mice adjust to the feeding period within a few days, they display food anticipatory behavior and consume their daily food intake during that limited time [49]–[52]. This feeding regimen drives rhythms in arrhythmic and clock mutant mice or in animals with SNC ablations, thus uncoupling the circadian clock, synchronized by SCN, from the periphery [3], [4]. Many physiological activities that are normally dictated by the SCN master clock, such as hepatic P450 activity, body temperature, locomotor activity, and heart rate, are restored by RF. In the liver of cry1−/−;cry2−/− mice, RF restores the oscillatory circadian expression profile of a number of “feeding driven” transcripts, although with a small delay in their phase of expression, showing that the circadian clock anticipates changes in the feeding state and accelerates the transcriptional response to an acute activation or repression by feeding [4]. This is consistent with our observation that in Bmal1−/− mice, rhythmic SREBP1 expression and activity, that are drastically flattened when mice are fed ad libitum (data not shown), are reinstated upon RF, although with a deferred phase. Notably, growth and metabolic defects that were reported in Bmal1−/− mice either at older age or under different feeding regimens [53]–[58], are negligible in the experimental conditions adopted in our study, suggesting that the role of the circadian clock in the regulation of SREBP1 could be evaluated in the absence of major confounding pathologies. As an example, Bmal1−/− mice at 8–10 month of age have an impaired insulin release due to the absence of a functional clock in pancreatic beta-cells [59]. However, in our case the delayed SREBP1 activation cannot be attributed to a reduced insulin release, since we detect normal glucose and insulin levels in Bmal1−/− mice at 3 month of age upon RF. Conversely, the expression of REV-ERBα, that is directly regulated by BMAL1, is constantly downregulated. This event leads to the derepression of the Insulin Induced 2 (Insig2) gene, encoding a trans-membrane protein that sequesters SREBP proteins to the endoplasmic reticulum membranes, thus interfering with the proteolytic activation of SREBPs, in agreement to what was shown in REV-ERBα−/− mice [16]. SREBP1 activity in the nucleus reflects also the rate of its proteosomal degradation after DNA binding [60], [61], a process that is strongly sensitive to the insulin-mediated inactivation of the glycogen synthase kinase 3 (Gsk3) [62]. In Bmal1−/− mice the phase of SREBP1 recruitment to DNA is shifted, but we do not observe a longer SREBP1 accumulation on its targets, suggesting that the absence of a functional clock is not significantly altering its degradation process. In conclusion, our results show that besides the nutrient-driven regulation of SREBP1 nuclear accumulation, a second layer of modulation of SREBP1 transcriptional activity exists and is strongly dependent from the circadian core clock. This system allows to fine tune the expression timing of SREBP1 target genes, thus helping to temporally separate the different physiological processes in which these genes are involved. Thus, SREBP1 is situated at the interface of the circadian and the metabolic regulation and its study promises to shed light on the emerging association between diabetes, obesity, sleep, and circadian timing.
All animal experiments and procedures were approved by the Swiss Veterinary Office (authorisation VD-1453.4). C57BL/6 male were purchased from Charles River. Bmal1−/− mice were a kind gift from Dr. Frédéric Gachon and were generated as previously described [63], [64]. 12–14 week old (at time of sacrifice), mice were housed in a 12 h light/12 h dark (LD) regimen for 2 weeks with food and water freely available during night and day. They were then phase-entrained to a 12 hr/12 hr LD regimen with food access between ZT12 and ZT24 for 7 days (ZT = Zeitgeber time; ZT0 is defined as the time when the lights are turned on and ZT12 as the time when lights are turned off). At each ZT2, ZT06, ZT10, ZT14, ZT18 and ZT22 three to five mice were anesthetized with isoflurane and decapitated. Mice were killed under dim red light at ZTs during the dark phase. The livers were perfused with 2 ml of PBS through the spleen and immediately collected. A small piece of liver tissue was snap-frozen in liquid nitrogen. The remaining liver tissue was immediately homogenized in PBS containing 1% formaldehyde for chromatin preparation.
Perfused livers were processed for chromatin preparation as previously described [65]. The chromatin samples from the mice of the same ZT were then pooled, frozen in liquid nitrogen and stored at −80°C. The following antibodies were used: anti-RPB2 (Santa Cruz Biotechnology, H-201), anti-SREBP1 (Santa Cruz Biotechnology, H-160), anti-HNF4 (Santa Cruz Biotechnology, C-19). Chromatin was subjected to immunoprecipitation of Pol II as described [24]. For SREBP1, the samples were diluted ten times in “sonication buffer” containing 50 mM HEPES (pH 7.9), 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Na-deoxycholate, 0.1% SDS and proteinase inhibitors (Roche). 1 ml of diluted chromatin was immunoprecipitated with 10 µg of antibody as described [66]. Briefly, the immune complexes were collected by adsorption to ten µl of protein-A-Sepharose (25% slurry in sonication buffer), pre-blocked with 10 µg/ml of salmon sperm DNA and BSA at 4°C overnight. The beads were washed twice with “sonication buffer”, twice with sonication buffer containing 500 mM NaCl, twice with 20 mM Tris, pH 8.0, 1 mM EDTA, 250 mM LiCl, 0.5% NP-40, 0.5% Na-deoxycholate and twice with TE buffer. The immunocomplexes were eluted with 50 mM Tris, pH 8.0, 1 mM EDTA and 1% SDS at 65°C for 10 min., adjusted to 200 mM NaCl and incubated at 65°C overnight to reverse the cross-links. After successive treatments with 10 µg/ml Rnase A and 20 µg/ml proteinase-K, the samples were extracted with NucleoSpin Kit (Macherey- Nagel). The DNA concentration was determined by fluorometry on the Qubit system (Invitrogen). 10–12 ng DNA were used for the preparation of the library. Libraries for ultra-high throughput sequencing were prepared with the ChIP-Seq DNA sample kit (Illumina) as recommended by the manufacturer.
About 100 mg of snap-frozen liver tissue were used for RNA preparation with the TRIzol reagent (Invitrogen) followed by purification with miRNeasy Mini Kit (Qiagen), according to manufacturer's instructions. For microarray analysis 500 ng of total RNA from each liver sample at the same time point were pooled and analyzed on Mouse Gene 1.0ST arrays according to the manufacturer's instructions (Affymetrix). All statistical analyses were performed with the statistical language R and various Bioconductor packages (http://www.Bioconductor.org). Normalized expression signals were calculated from Affymetrix CEL files using RMA normalization method. For quantitative RT-PCR analysis, the retrotranscription has been done using iScript cDNA synthesis kit (Bio-Rad, Laboratories, Hercules, CA) and following the manufacturer's instructions. The primers sequences are shown in Tables S7 and S8. Real-time monitoring of PCR amplification of cDNA was performed using the FastStart Universal SYBR Green Master (Roche Applied Science, Indianapolis, IN) in an ABI Prism 7900 Sequence Detection System (Life Technologies, Carlsbad, CA). The PCR arbitrary units of each gene were defined as the mRNA levels normalized to the 36b4 and the Rps9 expression level in each sample using the qBase Software.
Nuclear extracts were prepared by the NUN procedure as described previously [67], and Western blotting was performed according to standard protocols using the antibody for SREBP1 indicated above. U2AF and Lamin A were used as loading control (anti-U2AF and anti-Lamin A were from Sigma-Aldrich).
At sacrifice, blood was taken for determination of biochemical parameters and circulating hormones. Insulin levels were determined with ELISA kit from Mercodia, Uppsala, Sweden, following manufacturer's instructions.
At each time point, DNA sequenced reads were mapped to the mouse genome (Mus musculus NCBI m37 genome assembly (mm9; July 2007)) using Bowtie [68] with three mismatches and at most five hits allowed on the genome. When computing genomic read densities, each alignment contributed 1/(total number of hits) to the local density. If several reads in the same library mapped at the same genomic position and on the same strand (redundant tags), we kept only one read for the rest of the analysis. The total numbers of reads per time point are given in Table S1.
The mapped reads were shifted to account for the length of the inserts based on the average fragment size, namely 190 184 204 199 202 208 and 176, for each of the seven libraries. The fragment size was divided by two and half of the read length was subtracted resulting in a shift of 55, 52, 62, 60, 61, 64, and 48 nucleotides from ZT02 to ZT26.
We fragmented the genome into 500 nucleotide blocks and collected the counts within each block for SREBP1 as well as for input experiments. We kept all blocks for which we had a signal equivalent to 40 tags in at least one time point of the SREBP1 experiment. We log transformed the data after adding 1 pseudo count and quantile normalized both SREBP1 and input experiments. We selected blocks with a log2 signal SREBP1/input greater than 2, i.e. at least a four-fold enrichment of the SREBP1 signal in comparison to the input signal in at least one time point. We repeated this procedure by shifting the block definition of half the length of the blocks and merged the overlapping blocks that passed these criteria.
To define proper “peaks” in these wide regions, we look for shorter regions accounting for most of the counts. For this, we repeatedly consider the two borders along 50 nucleotides, and discard the one with less read counts if we keep 75% of the total reads in the remaining region. This operation is repeated with shorter borders until no further refinement is possible.
We used the MEME suite [21] to identify enriched motifs in the sequences corresponding to the refined peaks. We first performed several motif discovery, on all peaks, and on the subset associated or not with Pol II. We searched for 15 motifs between 6 and 10 nucleotides long. The discovered motifs have been associated to known transcription factors (from the TRANSFAC database) with STAMP. To retrieve these motifs in the subsets where they have not been discovered, we searched for them in all sequences (using FIMO). To assess the relevance of the number of observed motifs in our dataset, we counted the occurrence of the same motifs in random sequences. These sequences are selected in the proximity of the TSS of genes expressed in our samples (among the top 10% in the microarray data). Their size and distance to TSS are in the same range as that of our SREBP1 peaks.
Before fitting a cosine function to estimate the amplitude and the phase of the oscillation in a 24 hour period, counts in the refined peaks were quantified and normalized according to the total number of non redundant mapped reads for each given library.
We used the function x(t) = b0+b1*cos(b3+2π*t/24) to perform a least squared fitting of temporal profiles. The parameter b0 represents the mean signal, b1 the amplitude of the oscillation, and b3*24/2π the phase. These parameters were estimated by nonlinear least-squares using the Gauss-Newton algorithm.
For the microarray temporal profile analysis we used the function
while when we compared WT and Bmal1−/− mice we used the functionwhere b4 and b5 represent the different batch effects, GT is a dummy variable that indicates the mouse genotype (WT or KO) and b0gt, b1gt, b3gt the associated coefficients.
Illumina sequencing data for the ChIP-seq are available at GEO as the GSE48375. Additional processed data and visualization tools are provided at http://cyclix.vital-it.ch.
|
10.1371/journal.pgen.1006028 | Mitochondrial Polyadenylation Is a One-Step Process Required for mRNA Integrity and tRNA Maturation | Polyadenylation has well characterised roles in RNA turnover and translation in a variety of biological systems. While polyadenylation on mitochondrial transcripts has been suggested to be a two-step process required to complete translational stop codons, its involvement in mitochondrial RNA turnover is less well understood. We studied knockdown and knockout models of the mitochondrial poly(A) polymerase (MTPAP) in Drosophila melanogaster and demonstrate that polyadenylation of mitochondrial mRNAs is exclusively performed by MTPAP. Further, our results show that mitochondrial polyadenylation does not regulate mRNA stability but protects the 3' terminal integrity, and that despite a lack of functioning 3' ends, these trimmed transcripts are translated, suggesting that polyadenylation is not required for mitochondrial translation. Additionally, loss of MTPAP leads to reduced steady-state levels and disturbed maturation of tRNACys, indicating that polyadenylation in mitochondria might be important for the stability and maturation of specific tRNAs.
| The polyadenylation of cellular RNAs is a well-studied signal for gene expression, with a defined function in either RNA turnover or translation, in the majority of systems. In mammalian mitochondria the role of polyadenylation is less clear, and can to date only be attributed to completing the translational stop signal on several mitochondrial transcripts. Previous work though demonstrated that mitochondrial polyadenylation requires a certain length and shortening of the poly(A) tail signal has detrimental effects on mitochondrial function. In this study we deleted the mitochondrial polymerase responsible for polyadenylation in the fly, Drosophila melanogaster, and demonstrate that the mitochondrial poly(A) tail is essential for preserving the 3’ ends of mitochondrial transcripts, with no other enzyme capable of completing stop signals. Our study also shows that polyadenylation does not regulate transcript stability nor is it required for translation, but might be involved in the maturation of certain mitochondrial tRNAs. We therefore conclude that besides completing translational stop signals, mitochondrial polyadenylation protects the 3’ termini from degradation.
| The mitochondrial genome (mtDNA) is well conserved among metazoans, encoding a core set of 11 mRNAs, 22 tRNAs and 2 rRNAs in all species. In general, these genomes are very compact circular molecules, present in almost all metazoan cells as multiple copies within the mitochondrial matrix. Genes are intronless and a single major noncoding region contains regulatory elements, such as promoters and an origin of replication. Transcription from this region results in large polycistronic transcripts, often spanning the majority of the genome and requiring the recruitment of specific processing machineries to release the individual transcripts. Mt-tRNAs are disbursed throughout the genome and are proposed to form the structural basis for the successive cleavage of the 5' and 3' ends by RNaseP and RNaseZ complexes, respectively, at tRNA-mRNA junctions [1–4]. All mitochondrial RNAs require additional post-transcriptional modifications, catalysed by highly specialised enzymes [5,6]. The extent of this transcript maturation varies among species, but the majority of transcripts undergo polyadenylation, with several mRNAs requiring the addition of adenines to complete a translational stop codon. The mammalian ribosomal subunits seem to contain no to little poly(A) tail, while both subunits in flies are polyadenylated. Similarly, both human [7] and murine [8] MTND6 contain no poly(A) tail, while MTND6 of Drosophila melanogaster (Dm) is polyadenylated [9,10].
Mitochondrial polyadenylation is catalysed by a designated polyadenylic acid RNA polymerase (MTPAP) [11], transferring on average 40–50 adenines to almost all mitochondrial transcripts [2]. The crystal structure revealed that MTPAP functions as a homodimer without the need for additional protein co-factors [12,13]. However, unlike cytosolic or bacterial polyadenylation signals [14,15], the role of the mitochondrial poly(A) tails is less clear. Silencing of human MTPAP in cell culture failed to reveal a uniform function of polyadenylation for transcript stability, but rather led to the suggestion of a transcript-specific response. Loss of poyadenylation appears to stabilise the transcripts for the complex I subunits MTND1 and MTND2, while destabilising mitochondrially encoded complex IV transcripts. On the other hand, other transcripts do not seem to be affected at all by decreased polyadenylation [11,16]. Similar results were obtained in fibroblast cells from a patient with spastic ataxia and optic atrophy due to a pathogenic N478D substitution in MTPAP [17]. This mutation disrupts the fingers domain of MTPAP, severely affecting adenylase function and leading to increased and decreased mRNA steady state levels [18]. None of the above studies though completely abolished MTPAP function, resulting in the presence of short oligoadenylase signals. The failure to detect mitochondrial transcripts devoid of any adenine addition lead to the suggestion that mitochondrial polyadenylation is a two-step process, requiring an initial oligoadenylation by a yet unknown oligoadenylase [5–7,11,19,20]. On the other hand, low processivity, or a balanced activity of adenylation and deadenylation, as reported in Arabidopsis [21], could also explain these results. Nevertheless, in all the above models, mt-mRNAs retained their stop codons, but oligoadenylation was not sufficient for adequate translation in mitochondria [18], raising the question why oligoadenylation is not sufficient for mitochondrial translation. In contrast, a mild increase in poly(A) tail length was observed in cell lines upon silencing or overexpression of mutant components of the proposed mitochondrial degradosome, although no functional consequences were associated [16,20,22].
We previously reported that a number of different steps in the mitochondrial gene expression process can influence the polyadenylation of mt-RNA, with varying consequences on mitochondrial translation [8,10,23]. Silencing the mitochondrial helicase SUV3 in D. melanogaster led to the accumulation of processing intermediates of mitochondrial primary transcripts, increased mt-mRNA steady-state levels, disrupted translation, and was accompanied by reduced poly(A) tail length [23]. A similar effect on polyadenylation was also observed in knockout and knockdown models of the mammalian leucine-rich pentatricopeptide repeat containing (LRPPRC) protein [8] or its fly homolog bicoid stability factor (BSF) [10], but resulted in increased de novo translation. Mutations in LRPPRC have been associated with a severe neurological autosomal recessive neurodegenerative disorder, Leigh syndrome French-Canadian variant [24], and thus it is plausible that reduced polyadenylation plays a part in disease aetiology. LRPPRC forms a complex with the SRA stem loop interacting RNA-binding protein, SLIRP [8,18,25], but unlike LRPPRC/BSF, SLIRP is dispensable for polyadenylation and mt-RNA stability [26]. Thus, it remains unclear why mitochondria require a poly(A) signal of certain length and in what way polyadenylation is involved in mitochondrial translation.
To address these questions we deleted MTPAP in D. melanogaster. We demonstrate that mitochondrial poly(A) tail formation is exclusively performed by MTPAP and its disruption leads to trimming of the 3' termini and the addition of short heterooligomers to mt-mRNAs. Additionally, we show that, unlike its function in the cytosol, polyadenylation in mitochondria is not required for translation, but the trimmed transcripts fail to encode proteins that result in functional OXPHOS complexes. Finally, our data suggest that mitochondrial polyadenylation does not regulate transcript stability, but rather protects mRNA integrity and might be required for the maturation of mt-tRNACys.
We performed a standard BLAST search for the human MTPAP ortholog in Dm and identified a single candidate, encoded by CG11418, sharing a 30% identity on protein level (S1 Fig). Mitochondrial localisation was predicted in silico, using either Mitoprot II (0.87) or Target P (0.84) software and confirmed by transiently expressing a GFP-tagged CG11418 fusion protein in both Schneider and HeLa cells (Fig 1A). To assess whether CG11418 has an important role in mitochondrial polyadenylation, we used the UAS-GAL4 enhancer trap system in two independent RNAi lines to down regulate the expression of CG11418 in flies. Constitutive expression of GAL4 resulted in significant silencing of CG11418 (Fig 1B), causing reduced larval body weight (Fig 1C) and death at the ferrate stage or soon after eclosure (Fig 1D), suggesting that CG11418 is essential for fly survival. We determined the poly(A) tail length of circularised mitochondrial transcripts in control and knockdown larvae by cloning and sequencing. As previously observed [10,23], the majority of control transcripts showed polyadenylation with 35–50 adenines, including 16S rRNA, while 12S rRNA showed a much shorter poly(A) tail. In contrast, transcripts from CG11418 knockdown larvae revealed a severe reduction in polyadenylation (Fig 1E), prompting us to conclude that CG11418 is indeed the fly ortholog of human MTPAP and will therefore be called DmMTPAP henceforth.
As mentioned above circumstantial evidence suggested that mitochondrial polyadenylation might be a two-step process [5–7,11,19,20], and in agreement with this, cloning and sequencing of mitochondrial transcripts from silenced DmMTPAP larvae disclosed a polyadenylation signal compatible with oligoadenylation in MTND5, MTCOX1 (Fig 1E) and 16S rRNA transcripts (Fig 1F). In contrast, the majority of cloned transcripts of MTATP6/8, MTND1, MTND4/4L and MTND6 had no poly(A) tail signal, or transcripts with trimmed 3' termini (Fig 1E). Surprisingly, polyadenylation of 12S rRNA did not seem to be affected by MTPAP silencing and retained on average four adenines on full-length transcripts (Fig 1F).
Gene silencing thus left a certain amount of ambiguity for the polyadenylation process of mitochondrial transcripts, and we therefore generated functional DmMTPAP knockout (DmMTPAPKO) flies by homologous recombination, without disturbing the expression of flanking genes or a gene situated within the CG11418 locus (for cloning strategy and targeting see materials and methods and S2A–S2F Fig). CG11418 is X-linked and heterozygous DmMTPAPKO female flies were viable and fertile, while male DmMTPAPKO larva died at the 3rd instar larval stage (Fig 2A). DmMTPAP expression levels confirmed that heterozygous female flies retained 50% mtPAP transcript levels, while DmMTPAPKO larvae had negligible levels (Fig 2B). Cloning and sequencing of the 3' termini of circularised mitochondrial transcripts from DmMTPAPKO larvae failed to show any sign of oligoadenylation, with the exception of one MTATP6/8 clone and one MTCOX1 clone that contain 4 or 1 adenine extensions, respectively (Fig 2C). Further, the majority of clones were shortened at the 3' termini by 1–20 nt (Fig 2C). The only exceptions to this were a single MTND5 clone, three MTND4/4L clones, three MTCOX1 clones and one MTCYTB clone, which showed no adenine additions but retained their annotated full-length reading frame (Fig 2C). Some 16S ribosomal transcript subunits still retained full-length poly(A) tail signal, although the majority had reduced or no poly(A) tail signal but no increased amount of 3' trimming. In agreement with the results obtained by DmMTPAP silencing, 12S rRNA transcripts were adenylated comparable to controls (Fig 2D). Surprisingly, approximately 30% of the sequenced clones, including the full-length clones from MTCOX1 and MTCYTB, had short oligonucleotide additions, consisting predominantly of cytosine and adenosine nucleotides (S1 Table). To confirm that these additions were indeed 3' extensions, the 3' ends of MTCOX1 and MTCYTB were analysed through a rapid amplification of cDNA ends (RACE), confirming both 3' shortening and unconventional extension of mitochondrial mRNAs upon loss of DmMTPAP (Fig 2C and 2D, S1 Table). Thus, our results demonstrate that MTPAP is the sole adenylase in mitochondria, and loss of DmMTPAP leads to mitochondrial transcripts with compromised integrity.
Silencing of human MTPAP has previously been suggested to both increase and decrease mitochondrial transcript steady-state levels, while retaining a short oligoadenylation signal [11,16]. Concurrently, the knockdown (Fig 3A) and knockout (Fig 3B and 3C) of DmMTPAP in larvae led to an increase of mitochondrial mRNA steady-state levels, with the exception of MTCOX1 and MTCYTB, which showed either unchanged or decreased steady-state levels (Fig 3A–3C). Northern blot analysis also revealed additional shortened transcripts of MTCOX1, MTCYTB, MTCOX3 and MTND4/4L, suggesting that in the absence of a poly(A) tail degradation intermediates of these transcripts accumulate (S3A Fig). Interestingly, these mRNAs also had comparatively mild changes in steady-state levels in DmMTPAPKO larvae (Fig 3B and 3C). To confirm the shortened transcripts are indeed the result of mRNA degradation, rather than accumulation of antisense RNAs, we used single-stranded probes against MTCYTB confirming the accumulation of shortened transcripts (S3A Fig). The general increase in mRNA steady-state levels in DmMTPAPKO larvae can be a consequence of de novo transcription and indeed, we observed a mild increase in newly synthesized transcripts (Fig 3D) and mtDNA steady-state levels (Fig 3E), however not to such an extent that it can explain the increased steady-state levels of mitochondrial transcripts. Thus, despite loss of 3' integrity, mitochondrial mRNAs were not degraded by the mitochondrial degradosome, suggesting that polyadenylation is required for sufficient degradation of some mitochondrial transcripts.
Transfer RNA steady-state levels are sometimes seen as a proxy for de novo transcription, and in agreement with in organello experiments, mitochondrial tRNA levels were all increased in DmMTPAPKO larvae. The only exception was tRNACys, which showed a marked reduction and an accumulation of shortened RNAs (Fig 4A and 4B). A similar trend was observed in RNAi knockdown larvae (S3B Fig), although no increase in de novo transcription (S3C Fig) or mtDNA levels (S3D Fig) was observed. To address the difference in stability of tRNACys, we cloned and sequenced tRNACys transcripts, showing that the majority of transcripts lacked the usual CCA addition in DmMTPAPKO larvae (Fig 4C, S2 Table). This maturation defect was confirmed by a lack of aminoacylation of tRNACys, with tRNATyr being the only other tRNA affected (Fig 4D).
Several mitochondrial transcripts require polyadenylation to complete a functional stop codon, but even oligoadenylation was insufficient for normal translation in human mitochondria from a patient with mutant MTPAP, suggesting a requirement of polyadenylation for translation [18]. We were therefore interested to investigate the effects on translation in the absence of full-length mRNAs. To our surprise, we observed an aberrant pattern with increased levels of de novo translation for the majority of peptides, both in KD (Fig 5A and S4 Fig) and KO (Fig 5B) mitochondria. We also observed some truncated peptides, suggesting aberrant translation for some peptides (Fig 5A and 5B, asterisk). Interestingly, translation seemed to correlate with transcript steady-state levels, with decreased MTCOX1/MTND4/MTCYTB polypeptides. Despite the increased de novo translation, we observed decreased steady-state levels of the nuclear-encoded complex I subunit NDUFS3, both in RNAi (Fig 5C and 5D) and DmMTPAPKO samples (Fig 5E and 5F), as well as of the mtDNA-encoded complex IV subunit MTCOX3 (Fig 5E and 5F), suggesting impaired complex assembly.
To assess OXPHOS complex formation, we performed blue-native gel electrophoresis (BN-PAGE) experiments, revealing reduced complex assembly and decreased complex I and IV in-gel activities, both in RNAi (S5A and S5B Fig) and DmMTPAPKO (Fig 6A) mitochondria. Additionally, Western blot analysis also exposed a complex V assembly defect (Fig 6B and S5C Fig). Finally, we performed mitochondrial oxygen consumption measurements, demonstrating that the loss of polyadenylation leads to a combined complex I and IV defect (Fig 6C), confirmed by decreased isolated respiratory chain enzyme activities in both DmMTPAPKO and RNAi larvae (Fig 6D and 6E).
Despite extensive efforts, the role of mitochondrial polyadenylation remains elusive, with both stabilising and destabilising roles assigned to various mitochondrial transcripts [11,16–18]. How these differential roles are supposedly regulated or executed is unclear, and therefore the only known function of mitochondrial polyadenylation is that it serves to complete the stop codon of several mitochondrial-encoded transcripts. To gain insights into the role of mitochondrial polyadenylation we identified the Drosophila melanogaster homolog CG11418 as the fly ortholog to human MTPAP, and show that polyadenylation is essential for mitochondrial function and fly development.
Silencing of MTPAP leads to reduced poly(A) tail length in human cell lines [11,16], while cells from a patient with a mutation in MTPAP also severely affected polyadenylation [17,18]. In these experiments, all transcripts retained short adenine tails, leading to the proposal that mitochondrial polyadenylation is a two-step process, requiring an oligoadenylase to initiate polyadenylation. However, although silencing of DmMTPAP also retained very short polyadenylation signals for some transcripts in flies, deletion of DmMTPAP resulted in a complete loss of polyadenylation signal in all mitochondrial mRNAs investigated. The only transcripts retaining some degree of polyadenylation in KO flies were the two ribosomal transcripts. We interpret this that maternal ribosomal transcripts are protected in the monosome from degradation and therefore retained their tail. This is supported by the observation that 16S rRNA transcripts showed one population of transcripts with poly(A) tail—albeit shortened—and one population with no tail at all. On the other hand, the presence of poly(A) tails on ribosomal RNAs could also be explained by the presence of an rRNA-specific adenylase. This interpretation is supported by the fact that 12S rRNA showed no change in poly(A) tail length, with no transcripts completely lacking polyadenylation. Interestingly, 12S rRNA has naturally a short poly(A) tail favouring the idea of a 12S rRNA-specific oligoadenylase. Both of these interpretations are not exclusive, but our results reveal that no oligoadenylase is capable of compensating for the loss of MTPAP on mRNAs.
Previous reports from human cell lines and our observations made in flies demonstrate that silencing MTPAP can result in short poly(A) tails. In the absence of an oligoadenylase, it is intriguing that reduced MTPAP levels do not result in either non-polyadenylated transcripts, or full-length poly(A) tails, but rather in shortened tails. MTPAP therefore might have low processivity in vivo or requires a stabilising factor, similar to polyadenylation in the nucleus, where the polyadenylate-binding nuclear protein, PABPN1, binds to the newly synthesised poly(A) tail to increase PAP affinity to RNA [27]. However, no such poly(A) tail binding protein has been identified in mitochondria, and targeting the cytosolic PABPC1 to mitochondria not only inhibited mitochondrial translation and OXPHOS function but also reduced poly(A) tail length [28]. Polyadenylation is suggested to be regulated by the activities of MTPAP and the 2'-phosphodiesterase, PDE12, proposed to deadenylate mitochondrial mRNAs [29]. Disruption of MTPAP might disturb this balance, and the overexpression of PDE12 in human cell lines resulted in reduced poly(A) tail length of some transcripts, although the majority of transcripts were unaffected, suggesting that additional factors might be required. A similar mechanism has been seen in Arabidopsis, where a balance between the mitochondrial poly(A)-specific deadenylase, AHG2, and the polyadenylase, AGS1, seem to regulate RNA steady-state levels by modulating the poly(A) signal of mitochondrial mRNAs [21]. Interestingly, PDE12 overexpression also resulted in trimmed 3' termini, suggesting that PDE12 might be able to degrade into the 3' ends passed the poly(A) tail signal [29]. We also observed 3' trimming after deleting DmMTPAP in all mRNAs investigated, suggesting that polyadenylation protects the 3' end of mitochondrial RNAs from degradation.
Polyadenylation has previously been proposed to either stabilise or destabilise mitochondrial transcripts, with a stabilising effect on transcripts that require polyadenylation for stop codon formation. Deletion of DmMTPAP did not confirm this function, with the lowest steady-state levels observed in MTCOX1 and MTCYTB transcripts, which do encode a functional stop codon in mtDNA, while highest steady-state levels were observed in transcripts both requiring polyadenylation for stop codon formation (MTND2, MTCOX2, MTND5) or already encoding a stop codon (MTND3). Additionally, we observed no agreement with previous results of polyadenylation and its effects on transcript stability, suggesting that the poly(A) tail does not differentially regulate mitochondrial transcript stability.
Polyadenylation in bacteria promotes degradation via a short poly(A) tail of ~30nt that resolves the 3' stem loop structure of bacterial mRNAs and a similar function is observed in plant mitochondria [30]. Increased steady-state levels and trimmed 3’ ends of the majority of messenger transcripts in the absence of MTPAP favours the hypothesis that the metazoan poly(A) tail has a similar function. In fact, a previous report suggested that a combination of PNPase and/or MTPAP activity adds short homo- or heterooligomers to mRNAs during the degradation process of mitochondrial transcripts [7]. Failure to add such short tails would potentially allow for certain mitochondrial transcripts to adopt a structural confirmation at the 3’ termini that prevents their degradation. Our results are compatible with such a model, as failure to polyadenylate resulted in increased steady-state levels of the majority of mitochondrial transcripts and transcripts that were truncated by maximal 20nt. An involvement of a structural configuration in the degradation of mitochondrial transcripts is supported by our observation that in the absence of polyadenylation many transcripts did contain an oligomer extension, but unlike the extensions proposed to occur during degradation [7], the extensions observed here carried the signature of the tRNA nucleotidyl transferase (TRNT1), responsible for the CCA-addition during tRNA maturation [31]. Interestingly, the CCA additions by TRNT1 prevent re-processing by RNaseZ [32] and thus might be a protective mechanism. However, our results are also compatible with the presence of factor(s) regulating transcript stability upon polyadenylation. For instance, the Fas-activated serine/threonine kinase, FASTK, has recently been suggested to interact with the 3' termini of MTND6 in human cells, regulating transcript stability [33], while other members of the FASTK family have also been reported to affect transcript steady-state levels of some mt-RNAs [34]. Human MTND6 is not polyadenylated and whether polyadenylation was required for other factors is not known.
In human mtDNA, tRNATyr and tRNACys overlap by a single base pair and it has recently been proposed that MTPAP activity is required to correctly resolve this overlap at the 3' end of tRNATyr [35]. In the fly these two tRNAs do not overlap, but we still observed maturation defects for both tRNAs. tRNACys was also the only tRNA investigated with reduced steady-state levels and also presented with incorrect CCA modifications. The mechanism behind this is not clear, but suggests that MTPAP is required for maturation of this cluster in additional species. Despite low steady-state and aminoacylation levels of tRNACys mitochondrial translation was upregulated. This is particularly surprising, when considering that the majority of mRNAs encoded incomplete 3' termini, demonstrating that the poly(A) tail is not required for mitochondrial translation. However, these transcripts do not encode functional peptides, which fail to assemble into complexes. Thus, mitochondria might not possess a nonsense mediated decay pathway to remove incorrect mRNAs, and it will be interesting to identify ribosomal release mechanisms in these flies.
In summary, we demonstrate that polyadenylation of metazoan mitochondrial mRNAs is not dependent on a mitochondrial poly(A) oligoadenylase, and that polyadenylation is not required for the stability of transcripts. We rather suggest that the poly(A) tail protects mRNA integrity preserving the 3' termini from degradation.
DmMTPAP cDNA was cloned (for primers see S3 Table) into pEGFP-N3 plasmid (Clontech) to generate a dmmtpap-GFP fusion construct, which was subsequently subcloned into pAc5.1/V5-His A plasmid (Life Technologies). Constructs were used to transfect HeLa or Schneider 2R+ cells, respectively. HeLa cells were cultured in high-glucose DMEM (Life Technologies) supplemented with 10% foetal bovine serum at 37°C in a 5% CO2 atmosphere. Schneider 2R+ cells were cultured in Schneider’s Drosophila Medium (Life Technologies) supplemented with 10% foetal bovine serum at 25°C. For co-localization studies, HeLa cells or Schneider 2R+ cells were transfected using a calcium phosphate transfection kit (Sigma-Aldrich), following the manufacturer’s instructions. 48 hours after transfection HeLa cells were fixed with 4% PFA and decorated with anti-TOM20 antibody (Santa Cruz, sc-11415) Schneider 2R+ cells were stained with 50 nM Mitotracker Red (Life Technologies). Images were obtained in a Nikon Confocal Microscope at the Live Cell Imaging unit, Karolinska Institutet.
The two non-overlapping UAS-RNAi lines w;UAS-MTPAP RNAi#1; (#11418-R1) (NIG-Fly Stock Centre (Japan)) and w;UAS-MTPAP RNAi#2; (#31497) (Vienna Drosophila Resource Centre) were used for in vivo knockdown studies. In all studies, ubiquitous down-regulation of CG11418 expression was driven by driver daughterlessGAL4 (w;;daGAL4) analogous to previous reports [10,23,36]. Experimental samples are labeled as follows throughout the text: daGAL4 Control (w;;daGAL4/+), RNAi #1 Control (w;UAS-MTPAP RNAi#1/+;), DmMTPAP RNAi #1 (w;UAS-MTPAP RNAi#1/+;daGAL4/+), RNAi #2 Control (w;UAS-MTPAP RNAi#2/+;) and DmMTPAP RNAi #2 (w;UAS-MTPAP RNAi#1/+;daGAL4/+).
The DmMTPAP knockout line was generated by ends-out homologous recombination as described previously [36,37]. The CG11418 locus contains a second gene between exons 2 and 3 of CG11418, Tsp2A, and the knockout strategy therefore only targeted exon 1 and the 5' end of exon 2 of CG11418, removing an 817 bp DNA region that contains the initiation of transcription, the ATG start codon and up to the first in frame ATG codon in position 342 of CG11418 mRNA. Approximately 5 Kb upstream and 3 Kb downstream of the CG11418 locus (BAC clone from BACPAC Resource Centre, Oakland, California, USA) were cloned by ET recombination into the pBluescript II SK+ vector (Stratagene). Both homology arms were sequenced and subsequently subcloned into the pGX-attP vector [37], generating pGX-attP/DmMTPAPKO. Primer sequences and restriction sites are listed in S3 Table. pGX-attP/DmMTPAPKO was injected into D. melanogaster embryos via P-element-mediated germ line transformation using The BestGene Drosophila Embryo Injection Services (Chino Hills, California, USA). Crosses for ends-out homologous recombination were carried out as described [37]. Homologous recombination events were identified by PCR and Southern Blot analysis. Primers for PCR screening and cloning of the Southern Blot probe are detailed in S3 Table.
All fly stocks were backcrossed for at least 6 generations into the white Dahomey (without Wolbachia) background (w;;). All fly lines were maintained at 25°C and 60% humidity on a 12h:12h light:dark cycle on a standard yeast-sugar-agar medium.
To estimate the eclosure rates of adult flies, eggs were collected during a 3-hour time window and transferred to vials (100 eggs/vial) to ensure standard larval density. Eclosure of adult flies was monitored in regular time intervals.
Genomic DNA was isolated with the DNeasy Blood and Tissue Kit (Qiagen), following manufacturer’s instructions. For Southern Blot mapping of DmMTPAPKO larvae, 1 μg of each DNA sample was digested with SalI and precipitated, followed by separation on a 0.8% agarose gel and blotting to Hybond-N+ membranes (GE Healthcare). Membranes were hybridized with [32P]-labelled double stranded DNA probes and exposed to PhosphorImager screens. dsDNA probes were labelled with [32P]-dCTP (Perkin Elmer) following the PrimeIt II kit protocol (Stratagene). qPCR quantification of mtDNA levels was performed in triplicates on a QuantStudio 6 (Applied Biosystems), using 5 ng of DNA and Platinum SYBR Green qPCR supermix-UDG (Life Technologies). Primers for probes and qPCR are listed in S3 Table.
Total RNA was isolated using the ToTALLY RNA kit (Ambion) and quantified with a Qubit fluorometer (Life Technologies) unless otherwise stated. Reverse transcription for qRT-PCR analysis was performed using High Capacity cDNA Reverse Transcription Kit (Life Technologies). qRT-PCR was performed on a QuantStudio 6 (Life Technologies) and/or 7900HT (Applied) with Taqman probes (Life Technologies) or Platinum SYBR Green qPCR supermix-UDG (Life Technologies) and gene-specific primers. Primers and Taqman probes used for qRT-PCR are listed in S3 Table.
RNA circularisation and RT-PCR was performed as previously described [9,10,23]. In brief, mitochondrial RNA extracts were treated with TURBO DNase (Life Technologies) to remove contaminant DNA. 12 ng of mitochondrial RNA were circularised with T4 RNA ligase 1 (New England Biolabs), precipitated and cDNA synthesis was performed, using the GeneAmp RNA PCR kit (Life Technologies) and gene-specific primers. PCR products were cloned into pCRII-TOPO and transformed in One Shot TOP10 E. coli (Life Technologies) following manufacturer's instructions. The plasmid was subsequently purified and the insert was sequenced using M13 forward and M13 reverse primers. Primer sequences for RT-PCR to amplify the region containing the poly(A) tails have been previously described [10,23].
Linker ligation was essentially performed as previously described [38]. In brief, a phosphorylated oligonucleotide linker was ligated to 2.5 μg of a total RNA sample using T4 RNA ligase 1 (New England Biolabs). RNA was precipitated and cDNA synthesis was performed using a primer complementary to the linker sequence (anti-linker) and the GeneAmp RNA PCR kit (Life Technologies). The 3' end of the mitochondrial RNAs was PCR amplified using the anti-linker and gene-specific primers. The PCR products were cloned into pCRII-TOPO and transformed in One Shot TOP10 E. coli (Life Technologies) following manufacturer's instructions. The plasmids were then purified and the insert was sequenced using M13 forward and M13 reverse primers. Linker, anti-linker and gene-specific primers for 3' RACE experiments are listed in S3 Table.
Steady-state levels of mitochondrial transcripts were determined by Northern blot analysis, using 3 μg of total RNA. RNA samples were separated in 1% MOPS-formaldehyde agarose gels and transferred to Hybond-N+ membranes (GE Healthcare). To analyse mitochondrial tRNA steady-state levels, samples were separated in neutral 10% PAGE and transferred to Hybond-N+ membranes (GE Healthcare). Membranes were hybridised with either randomly [32P]-labelled dsDNA probes or in vitro transcribed single-stranded RNA probes to detect mRNAs and rRNAs or with strand-specific [32P]-end labelled oligonucleotide probes to detect tRNAs. Membranes were exposed to a PhosphorImager screen and the signal was quantified using a Typhoon 7000 FLA and the ImageQuant TL 8.1 software (GE Healthcare). Primers used to generate dsDNA probes and oligonucleotide probes have been previously described [10,23,36].
RNA was isolated from 4-day ael larvae, using TRIzol Reagent (Life Technologies) and resuspended in 0.3 M NaOAc (pH 5), 1 mM EDTA. 2 μg of RNA were loaded on acidic 6.5% polyacrylamyde, 8 M urea, 0.1 M NaOAc pH 5 gels and run for 48h at 4°C. Gels were transferred to Hybond-N+ membranes (GE Healthcare) and tRNAs were blotted as described above. To deaminoacylate tRNAs, samples were incubated in 0.5 M Tris pH 9 at 70°C for 10 minutes before loading on the gels.
Mitochondria were isolated from third instar larvae and in organello transcription assays were performed as previously described [10,23,36]. In brief, 200 μg of fresh mitochondria were incubated for 45 min in transcription buffer (30 μCi [32P]-UTP, 25 mM sucrose, 75 mM sorbitol, 100 mM KCl, 10 mM K2HPO4, 50 μM EDTA, 5 mM MgCl2, 1 mM ADP, 10 mM glutamate, 2.5 mM malate, 10 mM Tris-HCl pH 7.4 and 5% (w/v) BSA), followed by RNA extraction, separation on a 1% MOPS-formaldehyde agarose gel and transferring to Hybond-N+ membranes (GE Healthcare). Mitochondrial de novo translation in isolated mitochondria was assayed as previously described [10,23,36], using EXPRESS protein labeling mix easy-tag (Perkin Elmer). Equal amounts of total mitochondrial protein were separated on 17% SDS-PAGE gels, followed by staining with 1 g/L Coomassie Brilliant Blue in a 20% ethanol, 10% acetic acid solution. Gels were then destained, dried and exposed to a PhosphorImager screen to visualise the mitochondrial translation products.
Western blot analyses were performed using whole fly or mitochondrial protein extracts according to the Cell Signaling Technology protocol (CellSignaling). Protein extracts were separated on 4–12% or 12% NUPAGE acrylamide gels (Invitrogren) and after transfer to PVDF membranes (Millipore) decorated with the following antibodies: Complex I-subunit NDUFS3 (Mitoscience MS112, dilution 1:1000), complex IV-subunit COX3 (Mitosciences, MS406, 1:500), complex V (Mitoscience MS504, dilution 1:5000) and VDAC (Mitoscience MSAO3, dilution 1:1000–2000). Protein bands were visualized with ECL western blotting reagents (Bio-Rad).
BN-PAGE and in-gel staining for complex I and IV activities was performed as previously described [10,23,36]. In brief, mitochondria were pelleted and lysed in 1% digitonin, 0.1 mM EDTA, 50 mM NaCl, 10% glycerol, 1 mM PMSF and 20 mM Tris pH 7.4 for 15 minutes on ice. After removing undissolved material by centrifugation, samples were loaded on 4–10% polyacrylamide gradient gels. In-gel complex I activity was determined by incubating the BN-PAGE gels in 2 mM Tris-HCl pH 7.4, 0.1 mg/ml NADH and 2.5 mg/ml iodonitrotetrazolium chloride. In-gel complex IV activity was determined by incubating the BN-PAGE gels in 50 mM phosphate buffer pH 7.4, 0.5 mg/ml 3.3’-diamidobenzidine tetrahydrochloride (DAB), 1 mg/ml cytochrome c, 0.2 M sucrose and 20 μg/ml catalase. Staining was performed at room temperature.
Oxygen consumption from third-instar larvae was measured at 25°C, using an oxygraph chamber (OROBOROS), as previously described [10,36]. Respiratory chain enzyme activities were determined on isolated mitochondria, as previously described [39].
All data are presented as mean ± standard error of the mean (SEM) or standard deviation (SD) as indicated. A one-way ANOVA with Dunnett's multiple comparison test was used for statistical analysis, except for transcript tail length cloning and sequencing data, where a Mann-Whitney test was used.
|
10.1371/journal.pntd.0001361 | Rodent Abundance Dynamics and Leptospirosis Carriage in an Area of Hyper-Endemicity in New Caledonia | Widespread but particularly incident in the tropics, leptospirosis is transmitted to humans directly or indirectly by virtually any Mammal species. However, rodents are recognized as the most important reservoir. In endemic regions, seasonal outbreaks are observed during hot rainy periods. In such regions, hot spots can be evidenced, where leptospirosis is “hyper-endemic”, its incidence reaching 500 annual cases per 100,000. A better knowledge of how rodent populations and their Leptospira prevalence respond to seasonal and meteorological fluctuations might help implement relevant control measures.
In two tribes in New Caledonia with hyper-endemic leptospirosis, rodent abundance and Leptospira prevalence was studied twice a year, in hot and cool seasons for two consecutive years. Highly contrasted meteorological situations, particularly rainfall intensities, were noted between the two hot seasons studied. Our results show that during a hot and rainy period, both the rodent populations and their Leptospira carriage were higher. This pattern was more salient in commensal rodents than in the sylvatic rats.
The dynamics of rodents and their Leptospira carriage changed during the survey, probably under the influence of meteorology. Rodents were both more numerous and more frequently carrying (therefore disseminating) leptospires during a hot rainy period, also corresponding to a flooding period with higher risks of human exposure to waters and watered soils. The outbreaks of leptospirosis in hyper-endemic areas could arise from meteorological conditions leading to both an increased risk of exposure of humans and an increased volume of the rodent reservoir. Rodent control measures would therefore be most effective during cool and dry seasons, when rodent populations and leptospirosis incidence are low.
| In this study, we surveyed rodents and their Leptospira carriage in an area where human leptospirosis is hyper-endemic. We evidenced the well-known associations between specific rodent species and particular leptospires in both mice and rats. Overall, the observed Leptospira prevalence was in the range 18–47% depending on species, similar to other descriptions. However, significant variations were observed both in the abundance of rodents and their Leptospira carriage, one hot period with heavy rain being associated with both a highest abundance and an increased prevalence. Similar meteorological conditions could lead to increased leptospires dispersal by the rodent reservoir and increased exposure of humans to risk situations (e.g. flood, mud). Because rodent control measures were demonstrated elsewhere to be cost-effective if correctly planned and implemented, this contribution to a better knowledge of rodent and leptospires dynamics provides useful information and may in turn allow to develop relevant rodent control actions aimed at reducing the burden of human leptospirosis.
| Leptospirosis is an endemic bacterial disease in many tropical and sub-tropical areas. Various Leptospira strains, maintained in different animal species, are excreted in the urine of asymptomatic chronically infected individuals [1]–[3]. Humans get infected when abraded skin or mucous membranes come into contact with contaminated kidneys, urine or urine-contaminated environments [2]. The detailed epidemiology of leptospirosis, both a zoonosis and an environmental disease, both an occupational and a recreational disease, is highly complex. Though veterinary leptospirosis is often studied, little is usually known on how wild or feral Mammals contribute to leptospirosis dynamics [4]. Virtually any Mammal species can act as a reservoir of a co-adapted Leptospira strain [1], but among animal reservoirs, rodents are recognized as the most significant Mammal species maintaining and disseminating leptospires worldwide [2], [3], [5]. The Norway (or brown) rat Rattus norvegicus is notably known to be a reservoir of Leptospira interrogans of the serogroup Icterohaemorrhagiae, whereas the domestic mouse (Mus musculus) is a reservoir for Leptospira borgpetersenii of the serogroup Ballum [2], [3]. In New Caledonia, Mammal biodiversity is low: no native terrestrial Mammal is known, except 9 bat species (both micro- and megabats) [6]. However, four rodent species are known to be present, all resulting from importation by the early human settlements: three rat species (Rattus exulans, Rattus rattus, and R. norvegicus) and the domestic mouse (M. musculus) [7]. Leptospirosis in farm animals has been well studied (see [4] for review) but feral Mammals have not been investigated. Even though several strains and serovars are involved in human cases [8]–[10], Icterohaemorrhagiae is the most frequent serogroup, pointing to the importance of rodents as a major reservoir.
In most tropical regions where leptospirosis is known to be endemic, a seasonality is observed, with highest incidence during hot rainy periods, particularly after tropical storms and floods [11], [12] or during the monsoons [13]. During seasonal or post-cyclone outbreaks, there are particular areas of New Caledonia where leptospirosis incidence can reach 500 annual cases per 100,000 habitants [14]. Patchy distributions of leptospirosis have been described in New Caledonia [14], [15] but also in Brazil [16]–[18].
How human exposure to environmental contamination, reservoir abundance (especially rodents) and their Leptospira prevalence each contribute to such outbreaks and “leptospirosis hot spots” remains unknown. The aim of our study was to determine rodent abundance and the dynamics of Leptospira prevalence in their populations, in a site previously determined as a leptospirosis hot spot.
Based on previous studies [9], [19], the municipality of Bourail (see Figure 1) was known as a place of high incidence of leptospirosis. Using the diagnostic data of the early 2008 outbreak [14], we more precisely described the probable contamination area of these cases. Demographic data obtained from the Institut de la Statistique et des Études Économiques (http://www.isee.nc/index.html) were also used to evaluate the incidence of leptospirosis in each district. This incidence was then plotted on a map using PopGis version 1.0 (Secretariat of the Pacific Community). This allowed identifying three hot spots within the municipality of Bourail (Figure 1). The districts of Pötê and Buiru were chosen for our survey, based on their limited surface area and the good acceptance of our project by custom chiefdom and populations. These two study sites correspond to two Melanesian tribes where many outdoor activities are part of the everyday life, including fishing and bathing in freshwater streams, agriculture, maintenance of backyard pig pens, hunting (deers and wild hogs). Most of the households have one or more dogs that freely stray from the houses. Most of the inhabitants go bare foot and know the presence of rodents in and around their houses. After adequate contacts with the customs authorities, the study period was determined as a two-year period, one survey being conducted for each hot (January–April) and each cool (July–October) seasons.
Relative rodent abundance was evaluated both in the hot season (March) and during the cooler months (September-October). The sampling strategy was aimed at evaluating both the rodent abundance and the prevalence of Leptospira in the kidneys of the different rodent species. The method used was based on the standard index trapping technique developed in New Zealand for the study of rodents [20]: Hundred snap traps were placed by pairs (one rat- and one mouse-sized lethal snap trap) every 25 meters along a transect line close to the households within the study site. In Pötê, two transects were used with 50 traps each: one along the stream close to the households, the other one starting close to a farmer's house and extending in a cattle pasture along the same stream (Figure 2). In Buiru, the transect also extended along the streams close to the household, thus dividing at a stream confluence. Trap stations were set for 3 consecutive nights and baited with cheese and peanut butter. The processing of the trapped rats used the methods of Cunningham & Moors [20] and included the identification of species based on measurements for head-body length (HBL) and tail length (TL), assignment to a developmental stage (either adult or juvenile), and sex determination. Any animal that had been damaged by predators was identified to the species level, sex and age class was determined only if possible. Traps were checked and re-baited daily, captures and whether baits were taken or the trap sprung was also recorded. Trapping success was corrected for sprung and bait-taken traps by subtracting half a trap night for each such occurrence as described [21]. This allowed calculating an index of abundance, expressed as a number of captured animals per 100 corrected trap-nights.
Meteorological data were kindly provided by Météo-France (http://www.meteo.nc/) and corresponded to the closest automated meteorological station (Bourail). Rain (monthly accumulated rainfall) and temperature (monthly average of daily minima and monthly average of daily maxima) were plotted for the two years of the study (2009 and 2010).
Rodents were killed when caught by the snap traps. Each individual rodent was aseptically dissected and one kidney was immediately put in 95% alcohol for postponed DNA extraction. During the first survey (March 2008), one small piece of kidney was also aseptically transferred into a EMJH culture tube supplemented with 300 µg.ml−1 5-Fluoro-Uracile as an inhibitor of contaminant bacteria [22] for Leptospira isolation. A few rodents captured had been attacked by cats or birds at the time of collection; some could therefore not be completely identified or dissected. These were considered for abundance calculations but not studied for Leptospira prevalence.
Back at the lab (at the end of the 3-day sampling period), EMJH culture tubes were incubated aerobically for 14 weeks at 30°C with weekly dark field microscopic observation. Positive cultures were immediately subcultured in fresh EMJH and then frozen at −80°C with 10% glycerol.
A small piece of each individual kidney (ca. 10 mg) was aseptically dissected, rehydrated in 3 successive ultrapure sterile water baths for 6–12 hours each. It was then grinded in 50 µL sterile Phosphate Buffer Saline, pH 7.4 (Sigma) and 180 µL ATL Buffer (QIAamp DNA mini kit, QIAGEN) using a sterile single-use Piston Pellet (Kimble Chase). DNA was then extracted using QIAamp DNA mini kit (QIAGEN) following the manufacturer's instructions for tissue. The proteinase K digestion step was set for 4 hours. Additional proteinase K (20 µL) was added to samples incompletely digested at this time and incubation prolonged until complete tissue lysis (up to 8 hours). One millilitre of each Leptospira culture was centrifuged and extracted using the same QIAamp DNA mini kit (QIAGEN) using the manufacturer's instructions for cultured cells. The elution volume was 100 µL for either rodent kidney or Leptospira isolate.
Leptospira carriage in the rodent kidneys was assessed using two previously described diagnostic real time PCR assays, both using SYBR Green I technology, namely the detection and amplification of lfb1 [23] or lipL32 [24]. To check the absence of PCR inhibitors that would lead to false negative results, all kidney DNA extracts were also amplified with a “universal” 18S rDNA PCR, using primers previously described [25] and SYBR Green I technology. All oligonucleotide sequences are shown in Table S1.
The lfb1 PCR products amplified from positive kidneys were purified using the MinElute PCR Purification Kit (Qiagen) and directly sequenced as previously described [10]. For Leptospira isolates, a MLST scheme [26] was used as described before [10]. Alignments and phylogenies were then obtained using previously described techniques [10].
All sampling points were referenced using a handheld GPS device (Garmin). Data were transferred to MapInfo version 7.0 (Pitney Bowes Software Inc.) on a 1/10,000 map, kindly provided by the Direction des Infrastructures, de la Topographie et des Transports Terrestres – Gouvernement de la Nouvelle-Calédonie. All captured animals were similarly plotted on the map.
Qualitative variables were expressed as percentages or proportions. Groups were compared using Fisher's exact test. A test with a p value lower than 0.05 was considered statistically significant. Statistical analyses were performed by using Stata 11.0 (StataCorp LP, College Station, TX, USA).
Because rodents are introduced invasive Mammals in New Caledonia, they are legally classified as “dangerous detrimental species” and no particular authorization is required for their capture and study [27]. Protocols for animal experiments were prepared and conducted according to the guidelines of the Animal Care and Use Committees of the Institut Pasteur. The protocol was approved before the start of the experiments by a scientific committee and an animal care committee of the Institute Pasteur in New Caledonia.
To obtain the permission of conducting surveys in tribal lands, custom chiefs were met and the project of the research explained during a public information meeting on leptospirosis. A custom council (directed by the tribe and senior council chiefs) gave the necessary agreement for working on their land. All field studies were conducted with the help of a salaried tribe guide.
Fieldwork surveys were carried out in highly contrasted meteorological conditions with major differences in monthly accumulated rainfalls and temperatures (Figure 3): 2009 had a very rainy hot season, while 2010 had a relatively dry hot season. When compared to historical ten-year average rainfall and temperature data, the hot season in 2009 had a large rainfall excess, and stream overflows and floods actually occurred, whereas the 2010 hot season was both cooler and much dryer than a ten-year average. Although the cool season was warmer in 2010 than in 2009, both cool seasons had quite similar patterns, especially regarding (near normal) rainfall intensities.
All four rodent species known to be present in New Caledonia, namely the three rat species (R. rattus the black rat, R. norvegicus the brown or Norway rat and R. exulans the Polynesian rat) and the domestic mouse (M. musculus) were actually captured during our study. A total of 239 rodents were captured, out of which 231 could be identified (species) and 210 could be sampled for Leptospira carriage. Similarly due to predation of the captures, the age class could only be ascertained for 213 individuals. The black rat R. rattus was the most frequently captured species, accounting for 60.6% of the captures (140 rats), whereas mice M. musculus accounted for 25.5% (59 mice), Norway rats R. norvegicus for 9.1% (21 individuals), the rarest species being Polynesian rats R. exulans (4.8%, 11 rats). The greatest number of captures was achieved during the hot season 2009, with 113 rodents (47.3%). During the 2009 cool season, 2010 hot and cool seasons, the numbers of rodents captured were 28, 56 and 42 respectively. A greater number of rodents were systematically captured in Buiru (a total of 142 captures, 59.4%) than in Pötê (a total of 97 captures, 40.6%). The corresponding values in terms of abundance, an index used to compare rodent densities between different places and seasons and expressed as a number of captured rodents per 100 corrected trap × nights, are shown in Figure 4.
A significantly higher proportion of juveniles was found in all rat species in hot seasons (63.6%) when compared to cool seasons (18.9%) (p<0.001). Contrastingly, in mice, a lower proportion of juveniles was captured in hot seasons (14.3%) compared to cool seasons (46.7%) (p = 0.027).
From the universal 18S amplification, only one kidney DNA extract demonstrated PCR inhibition. This rodent was therefore considered for abundance calculations but excluded from prevalence studies. The detailed results are shown in Figure 5 and Table S2. In total, 56 rodents out of 210 (26.7%) were found as carrying Leptospira in their kidneys. This prevalence however considerably varied according to species, age class and seasons. The Leptospira prevalence was significantly higher in mice (24/51 = 47.1%) and Norway rats (7/19 = 36.8%) than in black rats (23/129 = 17.8%) or Polynesian rats (2/11 = 18.2%). Considering all species together, adults were more frequently carrying Leptospira (40/119 = 33.6%) than juveniles (15/88 = 17%) (p = 0.01), though this difference was not significant in every individual species. As an example, the prevalence was 21.2% in adult black rats, whereas it was 14.3% in juveniles (p = 0.36). Similarly, it was 75% (6 out of 8) in adult Norway rats and only 10% (1/10) in juveniles (p = 0.01).
The culture technique, used only during the first survey (March 2008) allowed the collection of 8 Leptospira isolates, from 6 mice (6 isolates), one black rat and one Norway rat. Because of its low yield and frequent failure in Leptospira positive kidneys, it was not used in further surveys.
The difference in Leptospira prevalence between the different surveys (see Figure 5) was not significant. However, trends could be evidenced in all rodent species, regardless of species or age class. The prevalence was higher in hot seasons (30.1%) than in cool seasons (19.4%) this difference however is not significant (p = 0.13). This difference was highest between the hot and rainy season in 2009 (33.7%) than all other seasons grouped together (21.2%) (p = 0.59) and was significant when considering only adult rodents (20/43 = 46.5% vs. 20/76 = 26.3% p = 0.0284). Similarly, the Leptospira prevalence was significantly higher during the hot and rainy season in 2009 (31/92 = 33.7%) than during the cool season in 2010 (5/40 = 12.5%) (p = 0.011).
The 8 isolates obtained from mice (6), one Norway rat (1) and one black rat (1) were typed using a MLST scheme described previously [10], [26]. The results pointed to L. borgpetersenii putatively belonging to the serogroup Ballum for mice isolates and L. interrogans putatively belonging to the serogroup Icterohaemorrhagiae for the isolates from both rat species.
In the 2 Polynesian and 7 Norway rats, the lfb1 PCR product sequence presumptively identified L. interrogans belonging to the Icterohaemorrhagiae serogroup as the infecting Leptospira [10]. Identical lfb1 sequences pointing to the serogroup Icterohaemorrhagiae were obtained from 20 out of 26 (76.9%) black rats, the remaining 6 (23.1%) were infected with a L. borgpetersenii with an lfb1 sequence presumptively pointing to the serogroup Ballum. From 24 Leptospira-infected mice, 22 were also infected with a L. borgpetersenii presumptively belonging to the serogroup Ballum, one with a L. interrogans presumptively identified as serogroup Icterohaemorrhagiae, whereas the last one was infected with an unidentified Leptospira sp. Actually, its kidney DNA extract gave a positive PCR amplification using the lipL32 PCR [24] but failed to be amplified using the other diagnostic PCR targets tested, namely lfb1 [23], secY [28], even if using degenerated primers (see Table S1) or different primers targeting a larger product [29]. Similarly, the TaqMan-based lipL32 assay [30] gave negative results using this DNA extract. The lipL32 PCR product was purified and sequenced, yielding a 352 bp sequence (Accession Number JN092329) that did not match any known Leptospira strain when compared with sequences available in GenBank using the Blast algorithm. Attempts to specifically amplify the 16S rRNA gene using Leptospira specific primers [31] for species identification were conducted on a gradient thermocycler but despite many attempts only allowed the sequencing of 207 bp of this gene. This 16S rDNA sequence (Accession Number JN092330) demonstrated a highest identity of 98% with L. kirschneri. The phylogenetic position of this uncultured Leptospira as deduced from this short 16S ribosomal sequence is shown in Figure S1.
We were able to capture all four rodent species (the black rat R. rattus, the Norway rat R. norvegicus, the Polynesian rat R. exulans and the mouse M. musculus) present in New Caledonia. The Norway rat (R. norvegicus) was very rarely captured in Pötê while this species was captured in all four seasons in Buiru. Whatever the site, the black rat (R. rattus) was the species most frequently captured and accounted for about 60% of captures. This coexistence of the black rat with Polynesian rats (R. exulans) and domestic mice (M. musculus) is consistent with previous studies realized in other locations in New Caledonia in uninhabited sclerophylls or rainforests. The sympatric behaviour of these rodent species in New Caledonia is regarded as a peculiarity. The four introduced rodents are usually not been found to coexist in the same habitat notably in New Zealand except possibly on the Chatham Islands [32]. In New Caledonia however, like in Hawaii [33], we found the four species to be sympatric. Our surveys allowed sampling rodents during highly contrasted seasons: the hot season 2009 was especially wet and warm, contrasting with the hot season 2010 which was much drier. The cool seasons 2009 and 2010 were quite similar, except that succeeding to either a wet (2009) or a dry (2010) hot season. Significant differences were noted in rodent abundance, highest abundances being observed during the hot and rainy period in 2009. A marked seasonality of rodent dynamics is well-known and has notably been considered as a major concern when considering rodents as reservoirs of infectious diseases [34], [35] and was modelled for an African rodent in the context of leptospirosis [36]. It is also recognized that seasonal factors must be considered when rodent control programs are to be implemented [37], [38].
The overall prevalence of Leptospira spp. in our rodent sample was 26.7%, a finding in accordance with former studies [39], [40]. No correlation was shown between sex and prevalence but age had a major impact on prevalence, adult animals being much more frequently infected (33.6%) than juveniles (17%) (p = 0.01), as already described in other locations [41]–[43]. Mice were more frequently infected than rats (p<0.001), no difference being evidenced between the three rat species (p = 0.16). Interestingly, when considering the ecological habits of the different rodent species, mice and Norway rats that are considered as commensal species (living closer to humans) have a higher prevalence (44.3%) than the more sylvatic black and Polynesian rats (17.9%) (p<0.001). Higher Leptospira prevalence in mice and Norway rats compared to black rats was frequently observed in some locations [44], [45], though contrasting results were reported in others [46].
As expected and already largely recognized, mice appeared to maintain L. borgpetersenii strains, the DNA sequences pointing to Ballum as the putative serogroup, Norway rats maintaining L. interrogans presumptively identified as belonging to the Icterohaemorrhagiae serogroup, both being evidenced in black rats in which Ballum appeared to be less frequent (23%). The simultaneous carriage of these two leptospires in a single (and probably panmictic) black rat population was also already described, e.g. in New Zealand [43] or Argentina [42]. Oppositely, no L. borgpetersenii carriage was detected in Norway rats, as was sometimes observed in Hawaii [41] or New Zealand [43]. Only two Polynesian rats were found as carrying leptospires, both from the species L. interrogans and presumptively identified as belonging to the serogroup Icterohaemorrhagiae, while leptospires from the serogroup Ballum found in other R. exulans populations [41] were not evidenced in our captures.
Unexpectedly, an unknown leptospire was also detected using various PCR techniques. This leptospire was found in the kidney of a domestic mouse. Its sequences clearly point to a species belonging to the pathogenic cluster of Leptospira spp. (see Figure S1) but its exact species identification cannot be ascertained. Interestingly, this strain could not be detected using the lfb1 PCR [23] routinely used for diagnosis in New Caledonia, nor using the TaqMan-based lipL32 technique [30] or the secY technique [28], all supposed to detect all pathogenic Leptospira spp. This surprising finding not only highlights the rich biodiversity of the Leptospira phylum but also questions about the existence of other pathogenic Leptospira species in New Caledonia that would be undetected using several of the PCR techniques described and currently used for diagnosis.
Water and rodents are known to play pivotal roles in the epidemiology of leptospirosis. Taken together, Icterohaemorrhagiae and Ballum serogroups have been responsible for more than 75% of human leptospirosis cases in New Caledonia [10], again highlighting the major contribution of rodents to human leptospirosis. The increased incidence of human leptospirosis in hot rainy seasons observed in New Caledonia [9], [14] and elsewhere [3] could result from the combined effects of an increased exposure of humans to mud and surface waters and of an increased Leptospira contamination of these environments. This latter would also result from both a higher survival probability of leptospires in wet environments during hot rainy periods and higher leptospire abundance due to increased seeding by reservoir populations. We actually evidence a higher rodent abundance and an increased Leptospira prevalence in rodent populations during one hot period with heavy rainfall. The results of our study are therefore in agreement with this global scheme, notably suggested as a factor contributing to a leptospirosis epidemics in Guadeloupe, West Indies [47] and with the rural model proposed by Holt and colleagues [36]. Additionally, our study in an area of leptospirosis hyper-endemicity highlights a higher Leptospira prevalence in mice and Norway rats, both rodent species which ecology and behavior bring in closer contact to humans compared to the more sylvatic black and Polynesian rats. Taken together, our data strongly suggest that all parameters studied might contribute to the occurrence of human leptospirosis epidemics during hot periods with heavy rainfalls: increased rodent populations with higher Leptospira carriage, leading to an increased contamination of an environment more favorable to leptospire survival.
Our data, though in complete agreement with prior knowledge on rodent dynamics elsewhere, only rely on two consecutive years and even more significantly only one season with heavy rain. Because interactions between climate variables, reservoir hosts and the pathogen are especially complex, additional surveys are needed to ascertain the influence of climate on rodents and their Leptospira carriage dynamics in the context of New Caledonia. With regard to rodent control measures, our results are also in agreement with previous knowledge and model predictions [36]–[38]. Should the impact of climate and meteorological variability be confirmed, the best rodent management strategy to minimize leptospirosis burden in New Caledonia would probably be the use of rodenticides before the start of a hot rainy period, a situation similar to rodent control for agricultural crops [38], therefore at times of low rodent density and low leptospirosis incidence, also corresponding to periods of low political and public awareness. Nevertheless, because economical modeling tends to demonstrate a similar cost-benefit effect of rodent control measures compared to post-exposure treatments [48], a better control of rodent populations should be increasingly considered as a possible approach for leptospirosis management in endemic areas.
Similarly to a study in Guadeloupe [47], the climatic conditions leading to leptospirosis epidemics in New Caledonia are under strong influence of the El Niño Southern Oscillation [9], [49]. The major advances in the modeling and prediction of this climatic phenomenon probably provides opportunities for predicting leptospirosis epidemics in some regions (e.g. [50]), in turn permitting to implement leptospirosis prevention measures (like river dredging, drainage or rodent control actions) in areas of high leptospirosis endemicity, in a timely manner.
|
10.1371/journal.pcbi.1003443 | Compact Modeling of Allosteric Multisite Proteins: Application to a Cell Size Checkpoint | We explore a framework to model the dose response of allosteric multisite phosphorylation proteins using a single auxiliary variable. This reduction can closely replicate the steady state behavior of detailed multisite systems such as the Monod-Wyman-Changeux allosteric model or rule-based models. Optimal ultrasensitivity is obtained when the activation of an allosteric protein by its individual sites is concerted and redundant. The reduction makes this framework useful for modeling and analyzing biochemical systems in practical applications, where several multisite proteins may interact simultaneously. As an application we analyze a newly discovered checkpoint signaling pathway in budding yeast, which has been proposed to measure cell growth by monitoring signals generated at sites of plasma membrane growth. We show that the known components of this pathway can form a robust hysteretic switch. In particular, this system incorporates a signal proportional to bud growth or size, a mechanism to read the signal, and an all-or-none response triggered only when the signal reaches a threshold indicating that sufficient growth has occurred.
| A large number of proteins in the cell are modified post-translationally by phosphorylation at multiple specific locations. This can help bring about interesting dynamical behaviors such as bistability or all-or-none responses to stimuli. Such behaviors are in turn important for cellular decision-making, differentiation, or the regulation of cellular processes. In this paper we propose the use of a specific technique for modeling allosteric multisite proteins, which can be thought of as the reduced version of a more detailed set of reactions. After validating this technique computationally by comparison to more detailed systems, we apply it to a new model of a signal transduction pathway with several multisite proteins. The model is concerned with a mechanism in budding yeast that is thought to measure the extent of daughter bud growth and send a signal to initiate mitosis when sufficient growth has occurred. Using the given framework we derive an analytically tractable model that creates the desired all-or-none signal. Overall, we give quantitative support to a newly discovered biochemical pathway, and we show the utility of the new modeling framework in the context of a realistic biological problem.
| Protein phosphorylation is a common form of post-translational modification frequently used in nature to alter protein activity, for instance by changing the electrostatic properties of the protein or its spatial structure. The phosphorylation of the same protein at multiple different aminoacid residues is also very common, and it is found in proteins such as p53 [1], Sic1 [2], EGFR [3], Wee1 [4], Ste5 [5] and many others [6].
The differences in the function of single-site vs. multisite phosphorylation are not completely understood. Many multisite proteins are involved in regulatory processes that can benefit from the presence of bistability, hysteresis, or limit cycles, which require sufficiently nonlinear interactions in addition to the right type of feedback [7], [8]. A reasonable hypothesis is that multisite phosphorylation can give rise to ultrasensitive dose responses, in a way that would not be possible in a comparable single-site system [9]–[13]. Many detailed mechanisms have also been proposed to explain the role of multisite systems in the emergence of bistability (for examples, see [14], [15]).
On the other hand, such detailed multisite mechanisms are normally not used as part of actual mathematical models of biochemical interactions. This is because explicitly modeling multiple sites usually involves the introduction of numerous variables, one or more for each phosphorylation state, and realistic models are often too complex already to justify this additional effort. Systems that attempt to model biochemical reactions explicitly often use the assumption that the protein has two states, one active and one inactive, with a simple reaction to transform one into the other, effectively assuming that the protein only has one site. Other models are more phenomenological in nature and include, for example, Hill function terms in the equation that are less clearly tied to the actual biochemical reactions [16], [17].
In this paper we describe a simple mechanistic approach for modeling multisite allosteric proteins. This approach, named modified fraction (MF) modeling, is capable of describing ultrasensitive dynamics without introducing a large number of additional variables. Under this framework one keeps track of the fraction of modified sites in the protein, and the concentration of active protein over time is estimated from this information. The protein is activated in a way that requires the phosphorylation of several but not all of the sites. The approximation becomes increasingly precise as the number of sites increases, with good estimates already for around four or more sites. In a sense, this mechanism can be considered a one-variable, quasi-steady state reduction of a model similar to the Monod-Wyman-Changeux allosteric system [18], although uses of MF modeling outside of MWC are also possible. The MF framework can also be extended to other types of multisite modification such as ligand binding, multisite transcription factor regulation, multisite methylation or acetylation, ubiquitination, etc [19], [20].
Perhaps the best way to test the versatility of a modeling tool such as the one proposed is to implement it in an actual biochemical system. In the current work we describe a detailed mechanistic model of a cell size checkpoint. Cell size checkpoints halt the cell cycle at specific points until sufficient cell growth has occurred [21], [22]. The mechanisms by which cell size checkpoints operate are poorly understood, and it is unclear whether they monitor actual cell size or parameters more closely related to the extent or rate of growth. In budding yeast, growth of a new cell is initiated when a daughter bud is formed on the surface of the cell [23]. The daughter bud initially grows in a polar manner, with all growth directed to the bud tip. Growth of the bud eventually switches to isotropic growth, in which the bud grows over its entire surface (see Figure 1A) [24]. The timing of the switch determines the duration of polar growth, which influences cell size and shape. It has been proposed in recent work by one of the authors that a cell size checkpoint controls the timing of this switch [25]. This checkpoint is the subject of our model. The variables are illustrated in Figure 1B and described in more detail below. See Tyson and Novak [26] for an accessible introduction to the systems-wide modeling of cell cycle checkpoints.
The MF approximation equation was first developed in [27], in the context of multisite systems with independent modification sites, with an emphasis on the estimation of the Hill exponents of sequential and nonsequential systems and on the comparison of their qualitative behavior. A major advance of the current paper, beyond the application to the cell cycle checkpoint, is to extend this work to cooperative and allosteric systems. Such systems are by definition non-independent, since the modification of one site accelerates the rate of modification of its neighbors. Cooperative systems are also more common and much better characterized than independent ones. The validation of the approximation in cooperative systems is ultimately based on a computational comparison of the MF reduction with detailed cooperative models having or variables.
In the first two Results sections we carry out a description of the modified fraction method to model multisite systems, and we compare simulations of the reduced model with those of a detailed mechanistic model. In the remaining two Results sections we carry out a mathematical analysis of the proposed checkpoint signaling pathway. We hypothesize that this interaction pathway has the capacity to produce a bistable signal responsible for a sudden switch from polar to isotropic growth, once the bud has undergone sufficient polar growth. The model presents several desirable qualities for a checkpoint, in particular a clear downstream signal when a sufficient polar bud growth has occurred.
We start by describing the assumptions on our model in the context of multisite phosphorylation (although it could also be applied to other irreversible covalent modifications as well as noncovalent ligand binding). Suppose that a protein substrate is phosphorylated by a kinase at possible sites and dephosphorylated by a phosphatase . The system is assumed to be nonsequential, so that there are different phosphoforms of , and the number of sites is thought to be relatively large e.g. . The system is cooperative in the sense that site phosphorylation accelerates the phosphorylation of neighboring sites. Since the number of sites is relatively large, the activation is thought to be cumulative and the effect of any individual site is assumed to be small. The sites are assumed to be equivalent to each other, in the sense that the rate of phosphorylation and dephosphorylation is similar across all sites and that no site has a stronger effect on substrate activation than other sites. The activation of the substrate may be due to binding to another molecule or body, such as the cell membrane. It could also be due to an internal structural change that allows the substrate to interact differently with other proteins. Thus the active protein concentration can be defined as the concentration of the protein bound to a particular molecule or in a particular molecular state.
Suppose that the phosphorylations lead to the concerted, redundant activation of the protein. That is, multiple phosphorylations are necessary for activation (concerted), and not all sites need to be phosphorylated for full activation (redundant). We define the activity function such that the fraction of active protein with phosphorylated sites is given by .
There are many systems that likely fall within this general framework. For instance, Ste5 is a scaffold protein in budding yeast with phosphorylation sites, which relays a pheromone response only when it is bound to the membrane [28], [29]. When phosphorylated by Cdk1, it tends to unbind from the membrane, shutting down its activity. The sites are predicted to lie on an unstructured region of the protein and appear to act by changing the protein's bulk electrostatic properties. In the paper [29], it was shown through site-directed mutagenesis that around five or more phosphorylations are necessary and sufficient for deactivation. Another recent example is the multisite phosphorylation of Cdc25 by Cdk1 in fission yeast, which was similarly studied in detail by mutating individual sites [30].
According to the modified fraction framework, we estimate the concentration of a particular protein state from the overall fraction of modified sites. For instance, if the protein has sites and the fraction of phosphorylated sites is , then the fraction of protein with only the first and last sites phosphorylated is roughlyHere is the total protein concentration. This is not an equality since cooperative effects introduce correlations among the sites, i.e. the sites are not independent of each other, but it is an approximation assuming cooperative effects are sufficiently weak. Multiplying on both sides by and adding over all possible phosphoforms with phosphorylations, the concentration of proteins with exactly phosphorylated sites out of a total of sites isOne can estimate the overall concentration of active protein asA key aspect of this formula is that if we denote the right hand side by , it can be shown that converges to as increases [31]. This gives the approximation(1)which becomes increasingly precise for large . Notice that the different quantities in this formula can potentially be measured in the lab - the active protein concentration via an activity assay, the total protein concentration via Western blot, and the activity function through site-directed mutagenesis.
A timescale decomposition argument can be made to use this approximation away from steady state. If is the fraction of active sites over time, and the timescale of protein activation is much faster than the rate at which changes, then one can approximate at any given time using the same formula. This produces a convenient method for modeling multisite systems under the given assumptions by keeping track of the variable , without creating , let alone variables. On the other hand, if , or any other process affecting protein activation, is at least as fast as protein activation itself, then nontrivial dynamics might take place such as limit cycle oscillations, and the approximation can introduce errors.
It is necessary to calculate the fraction of phosphorylation itself. Assuming linear rates of phosphorylation and dephosphorylation for , one obtains the system
That is, , where is the fraction of inactive sites. This is the default for the model, although any other rate equation for can be used, including Michaelis-Menten complex formation at the level of the individual sites.
As for the activity function , any sigmoidal function can be used, including functions measured directly by experiments. By default we assume the following form, which we will derive in the next section:
See also [5], [9], [27] for other uses and derivations of this formula in the literature. The function can actually be highly switch-like for large , which illustrates how small increases in the kinase can result in large activity changes in the protein , unlike linear rate models with only one site. Other forms for the activity function have effectively been considered by by Kapuy et al [32], and also by Wang et al [10].
In Figure 2A we show the relationship between the fraction of phosphorylated sites and the active protein concentration for a particular choice of the MF parameters using the approximation formula (and chosen to fit the detailed model described in the next section for ). Notice that the activation is concerted and redundant, in that a minimal threshold of phosphorylation is required for activation, and activation is reached for less than full phosphorylation.
A validation of the performance of this model for is now shown in Figure 2 in the context of a system similar to the classical and widely used Monod-Wyman-Changeux model of an allosteric multisite protein [18]. The original MWC model describes the binding of oxygen to the different sites of hemoglobin and the allosteric transitions of this protein between two different states. Rather than modeling oxygen binding, a protein with phosphorylations is assumed to change between an active conformation and an inactive conformation . Each of these forms can also be phosphorylated or dephosphorylated at the rates given in the diagram in Figure 2B. Although the model is interpreted in a different way from MWC, from a mathematical point of view it is almost identical. The coefficient accounts for an assumption that the protein is phosphorylated at a faster rate when it is active than when it is inactive. The coefficients etc represent the fact that this is still a nonsequential model: for instance, can be phosphorylated at different sites, so the phosphorylation rate is multiplied by . See the derivation of this model from more basic principles in Text S1, and a recent review on multisite systems by one of the authors [28].
Using this multisite model, we now derive parameters for a corresponding MF model. For instance, one can define the ‘average’ phosphorylation rate at a given phosphorylation site, regardless of whether the protein is active or inactive, , and the dephosphorylation rate . By way of derivation of the activity function , suppose that a protein with phosphorylations is switching between active and inactive form,At steady state, we assume that this exchange is balanced and calculate . Then the fraction of active sites with phosphorylations at steady state isIn other words , where and . See Text S1.3 for more details. In particular, the ultrasensitive behavior of the function generally increases with the number of sites.
At any given time, the active protein concentration of the full model is defined as . In Figure 2C we compare the full 12-variable MWC model for with the corresponding MF approximation. For every value of the input kinase concentration, the resulting concentration of active protein is plotted at steady state. Notice the close similarity between the two graphs, which is even more surprising since MF is essentially a one-variable model.
For comparison, we also plot the behavior of an overly simplified but all too commonly used model, in which the substrate is assumed to have a single phosphorylation site instead of sites, and it is modeled according to the reactionusing linear reaction rates. Notice that the behavior of this single-site model in Figure 2C is very different from that of the MWC model, and that any switch-like behavior in the response is lost. This can have important consequences regarding the existence of multiple steady states, hysteresis, oscillations etc in the context of larger systems, which will be illustrated below. It is easy to show that , i.e. it corresponds to when . It should be noted that if the single site system is modeled using Michaelis-Menten reactions rather than linear rates, it could have strongly ultrasensitive behavior in the saturation regime via zero-order ultrasensitivity [33]; see the Discussion section for more details.
We carried out a calculation of the distance between and for many different combinations of the number of sites and the allosteric parameter . For every such set of parameters, the two graphs were plotted at steady state as a function of , and the error was calculated in Figure 2D. Notice that the approximation is within 1% precision for arbitrary and . On the other hand, in order to obtain high ultrasensitivity it is required that be relatively large and/or be small (Figure 2E). See also Figure S2, where additional parameter variations are explored over four orders of magnitude using the same type of graphs, with similar results.
It is worth comparing this methodology with the approach known in the literature as rule-based modeling, where a series of chemical reactions is defined using a streamlined algorithm, and high-powered computing is used to handle the resulting large number of variables; see e.g. BioNetGen [34]. The advantage of this method is that a large number of reactions can be defined and handled this way, including complex parameter optimizations. One disadvantage is that the combinatorial explosion resulting from combining reactions can sometimes exceed the computational power. Another is that the large number of equations makes any mathematical analysis difficult, if at all possible.
It is interesting that the MWC model can actually be described in terms of rule-based modeling. In Figure S1A we describe a series of chemical reactions, over all possible phosphoform states, and we show in Text S1 that this system is in fact equivalent to the MWC model. Thus MF can also be seen as the 1-variable reduction of a system with variables and a much larger number of reactions.
In this section, we will embed the MF system within increasingly complex systems of equations. We consistently use upper case for proteins and lower case for modified fractions of sites. However, we will first provide some technical experimental background regarding this specific pathway.
In order to find the steady states of both subsystems together, recall that each one can be reduced to a single equation, so that the steady states correspond to the joint solution of the two equations. For fixed , the solutions of the full model form the intersection of the graphs for the equations (2), (3). This is illustrated in Figure 3F, where the graphs in Figure 3C and Figure 3E are superimposed on the same plane. From a control perspective, the upstream and downstream systems have each an input and an ouptut, and they feed back into each other (see the two dotted boxes in Figure 1B). The active Zds1/PP2A dimer also acts as the overall output of the system, since it triggers the downstream response to cell cycle regulatory proteins.
Although it is natural that an increase in the Rho1 flow can eventually trigger the activation of the pathway, the main focus here is not in the flow but in the overall Rho1 concentration at the bud tip. Given that Rho1 has a rate of growth proportional to and a linear rate of degradation, at steady state one can show that and are proportional, . This follows from adding the ODE rate equations at steady state, . In this way one can use the total Rho1 concentration as a bifurcation parameter at steady state even though it is simultaneously a variable in the system. Alternatively, since enzymatic reactions and dimer formation are fast processes compared with Rho1 flux and Rho1 degradation, one can let be the slow variable in the system and carry out a timescale decomposition analysis using as a constant [39].
Let's look at how the system has a hysteretic response for increasing values of the flow signal and the corresponding total Rho1 concentration at steady state. An increase in these values has the effect of raising the graph associated with the upstream system, as shown in Figure 4A. For smaller values of , the intersection of both graphs includes three positive steady states (notice the two graphs don't quite intersect at the origin). But when increases over a certain threshold, the intersection of the two graphs contains a single positive steady state, with a large value of . This can cause an abrupt change in the qualitative behavior of the system, triggering a sudden increase in the Zds1/PP2A output. Once this change has taken place, the concentration of the output stays high even if the input decreases.
Figure 4B shows a sample timecourse of the system for a time-variable vesicle flow (dotted line). The total Rho1 concentration increases over time with the inflow of vesicles. At a certain timepoint the active Rho1 concentration abruptly increases, due to the switch at the Pkc1/Rho1 upstream level. An increase in Rho1/Pkc1 concentration some time before this can be seen in Figure 4C. At a later time the switch between Zds1 and PP2A is also triggered, leading to a sudden increase in PP2A/Zds1 concentration. Even under variable flow, the total Rho1 concentration roughly corresponds to the membrane accumulated at the bud, except for a certain amount of variability due to Rho1 degradation. Lowering the Rho1 degradation rate can decrease this difference. Notice that the vesicle flow oscillations do not correspond to cell division, but to oscillations in the rate of growth, for instance due to varying food availability.
In Figure 4D we plot the output signal PP2A/Zds1 as a function of total Rho1 at steady state and overlay the solution of the timecourse simulation (red stars). This graph also illustrates the hysteretic behavior of the system, in that once a critical threshold of Rho1 concentration is reached, the output signal is dramatically increased. This change would constitute a clear signal that the bud has reached a large enough size for crossing the polar/isotropic growth checkpoint.
Since both the downstream (PP2A, Zds1) switch and the upstream (Rho1,Pkc1) switch are driven by positive feedback loops, it is valid to ask which of the two loops is more relevant for the overall system behavior. We argue that it is the downstream loop that is more essential, using the bifurcation analysis in Figure 3 and Figure 4A. If the upstream system is not bistable but has a single steady state for every input , then the graph in Figure 3E is replaced by a single-valued decreasing function. Nevertheless this (green) line can still have one or three intersections with the (blue) downstream dose response in Figure 4A, indicating hysteresis for the overall system. On the other hand if the downstream system is not bistable, then the blue curve in Figure 4A is replaced by a single-valued, increasing function, which would be unlikely to have three intersection points with the (green) upstream dose response. Thus the downstream switch is essential, while the upstream switch is not.
Notice that this system contains the standard elements of a signal transduction pathway, including an initiating signal (vesicles), a sensor (Rho1), a series of transducers (Pkc1, PP2A, etc), and an effector (active Zds1/PP2A). Total Rho1 is a proxy for the membrane concentration, even if bud growth slows for a period of time, and the cascade of reactions allows the signal to be transduced from the membrane to Zds1/PP2A and ultimately the cyclin dependent kinase. To ensure the high fidelity of the signal transmission [40], the downstream signal is sent abruptly after total Rho1 concentration reaches a particular size. Notice that longer periods of inactivity can potentially reduce the Rho1 concentration significantly – one possible prediction is that after such a period the bud grows longer than expected.
In this paper we have introduced a simple and compact framework to describe the dynamics of allosteric multisite phosphorylation systems, and we have applied this tool to a new molecular model of a size checkpoint in budding yeast. Multisite phosphorylation modeling can be problematic because ignoring the multiple sites can have significant effects in the dynamics, while introducing many auxiliary phosphoform variables can be cumbersome in more realistic models. The modified fraction approach is intuitive and flexible (model the sites instead of the protein), and it only introduces one additional variable per protein.
The components of the MF model, namely the function and the rates of phosphorylation and dephosphorylation of individual sites, can potentially be subject to direct experimental measurement, unlike the use of more abstract Hill function terms. This can allow to carry out ‘raw-data modeling’ e.g. to use an experimentally measured activity function directly in the model rather than using it to derive parameters. This methodology is also useful out of equilibrium when the timescale of phosphorylation is sufficiently fast compared with other timescales in the system.
There are several reasons why the MF method might be particularly suitable for modeling many multisite phosphorylation systems. Nonsequential phosphorylation is likely more common in nature than the more often modeled sequential systems, since enforcing sequential phosphorylations would require an additional mechanistic effort. Bioinformatic data suggests that most phosphorylation sites in multisite proteins are located in unstructured and unconserved protein regions [6], suggesting that often it is the collective effect that matters rather than the individual sites. There is also experimental evidence in yeast signal transduction that certain proteins, such as Ste5, are activated in a concerted and redundant manner, although this type of information is still unknown for most proteins. Notice that the approximation formula would still hold if the protein activation is not concerted or redundant. In that case the formula will just approximate a dose response that may not be ultrasensitive.
One of the best known mechanisms for ultrasensitive dose responses is zero-order ultrasensitivity, as suggested by Goldbeter and Koshland [33], [41]. Its main assumption is that substrate concentration needs to be in the saturation regime i.e. large compared to the value of the enzymes. The MF method does not pose any constraint on , in fact the linear regime we used can be found when substrate concentrations are small compared to values. Moreover, MF also applies when the enzymatic reactions involve complex formation, by writing a Michaelis-Menten equation for . Therefore zero-order ultrasensitivity can be used in synergy with MF in the saturation regime, and MF can be used regardless of value. A zero-order dose response could likely replicate the behavior of the MWC model as shown in Figure 2C, however it could not be considered a short hand notation for MWC since the two mechanisms are fundamentally different.
In the case of the checkpoint pathway, the active proteins Pkc1 and PP2A have been found to have an approximate of 0.5 [42] and 1.2 [43], respectively, for specific targets. The overall concentrations of their substrates in the cell are much lower - however these proteins tend to localize at the bud, so that the resulting local concentrations are unknown and it is unclear whether a zero-order approach would apply. A recent paper by Martins and Swain [44] points out that zero-order ultrasensitivity often results from low enzyme to substrate ratios, and localized proteins that act as enzymes and substrates for each other would likely not satisfy such ratios. That paper proposes instead a mechanism involving an allosteric model analogous to MWC, using enzyme sequestration to obtain ultrasensitivity. The paper by Kapuy et al [32] also proposes a mechanism for bistability through ultrasensitive effects, and this mechanism is applied to a detailed model of the budding yeast G1-S transition in Barik et al [45]. Other mechanisms for ultrasensitivity involve competition among substrates for the same enzyme [46] and protein localization [12], among others [28], [47].
More generally, in cell regulatory networks there is a need to implement nontrivial dynamics such as bistable switches and hysteresis, which requires some form of nonlinear response in addition to the right feedback interconnections. It has been observed that several regulatory proteins have multiple phosphorylation sites, and there are many open questions regarding their intended function. Together with the onerous nature of modeling several multisite proteins using sequential networks and multiple variables each, it can be seen why a one-variable reduction such as MF can allow for much-needed simplicity.
The actual mechanisms regulating the interactions between cell size and cell division remain largely unanswered in many cases. This has left few alternative options apart from somewhat heuristic approaches in otherwise very detailed models, see e.g. [48]. The present model is an attempt, based on recent experiments, to construct a detailed mechanistic model in the context of the polar to isotropic bud transition in yeast. Notice that if the proteins PP2A, Pkc1, Zds1 had only one site each, then according to the argument in the first Results section, and the downstream and upstream models could never be bistable (see equations (2) and (3)). The multiple sites are providing the underlying nonlinearity so that the models can have interesting dynamical behaviors. This is consistent with the work by Yang et al [49], which reached the same conclusion through randomized parameter searches in multisite cell cycle models. Also, the equations (2), (3), which represent the steady states of the downstream and upstream systems, have the same qualitative behavior for a wide range of parameters. In this sense one can say that the switch-like nature of the checkpoint is robust to many parameter changes, provided that a few key qualities are satisfied. The bistability in each subsystem is due in part to positive feedback loops in each subsystem, one between Pkc1 and the Rho1/Pkc1 dimer, and another between Zds1 and the PP2A/Zds1 dimer. Notice that while Rho1/Pkc1 activates PP2A, the downstream PP2A/Zds1 inactivates Pkc1, forming a negative feedback loop. This feedback could serve to reduce the activity of the pathway before a sufficient Rho1 signal has accumulated. Notice that the switch-like activation of Cdk1 is a complex process that may well be regulated by other mechanisms in conjunction with the switch discussed, and that this overall regulation also depends on the organism studied.
The MF framework eliminates several parameters such as the number of phosphorylation sites (as long as it is sufficiently large), the transition rates and the cooperativity coefficient . The remaining parameters, such as the shape of the activity function , can potentially be measured in the lab using site-directed mutagenesis and activity assays. However this is still a formidable task and one that is yet to be done for most proteins involved in cell cycle regulation.
Since it is assumed in the derivation of the formula (1) that the sites are roughly independent from each other, one might think that the MF framework doesn't work for allosteric or cooperative systems. However the detailed model in Figure 2 is allosteric, and yet the model closely describes its dynamics. In simulations we find that the accuracy of the representation is increased when is large (e.g. ) and/or the cooperativity is weak. The use of a MWC-type model for multisite phosphorylation has been pointed out in the past, see for instance [5] and the more recent [44].
Questions for future work include the following: if phosphorylation and dephosphorylation of the multisite protein is not faster than other processes in the system, can one still approximate away from equilibrium? This might be possible by defining a simple differential equation for instead of the algebraic equation (1). Also, the linear dynamics used to calculate the fraction can be replaced by more complex models such as a Michaelis-Menten reaction, which may be explored in detail, including the interaction with zero-order mechanisms. This might lead to bistable behavior in the full multisite model, which raises the question of how the corresponding model reduction might be, possibly involving multivalued functions .
For convenience we include in one location all chemical reactions of the model, mass conservation laws, the definition of auxiliary variables following the multisite modeling formalism, and a self contained set of differential equations after eliminating additional variables. Recall that for multisite proteins represents the fraction of active sites, the active monomer concentration, the total concentration including active and inactive forms, and the total amount of in the system including dimer and monomer forms. Also recall that Rho1 and Pkc1 are denoted by , , Zds1 and PP2A by , , and the Rho1/Pkc1 and Zds1/PP2A dimers by , , respectively.
Since most quantitative information about the pathway is unknown, we make educated estimates on the order of magnitude of the parameters. Since parameters are clustered in equations (2) and (3), dependence on the parameters is more limited. Protein concentrations usually range from 0.001 to 10 in the cell. The concentrations of total PP2A (), Zds1 () and Pkc1 () are set between 0.01 and 0.1 as indicated in Table 1. Define for all . The dissociation rate has been observed to be quite low in experiments since most Zds1 has been found bound to PP2A. We set it as 0.001 , which is in the range of drugs binding to their target proteins. is set higher at 0.1 . The unit-less parameters are set to 0.1 and 1 respectively, indicating the steady state ratio of inactive to active substrate when the two antagonistic enzymes are in similar concentration. The rates are set to 0.01 , indicating that when e.g. there are equal amounts of active and inactive substrate at steady state. is set at 0.0002 .
Very little is known about the values of the individual rates . Fortunately as it is shown in the analysis in Text S1, most of the dynamic rate constants appear only in the form , instead of individually. These steady state ratios are generally easier to estimate experimentally than the individual parameters [50]. However the actual rates determine the transient behavior of the system and to some extent determine also its steady state values. Since a majority of the reverse rates share the same units of , we set the values of these parameters and then find the corresponding to fit the given ratio . For simplicity we set [14], [51]. We set for maximal protein concentration , that is, . The Rho1 degradation rate is set to 0.0001 ; it can be further decreased in order to stabilize the Rho1 protein.
Regarding the activity functions , we assume that the ultrasensitive behavior of these graphs increases with the number of phosphorylation sites; see the derivation of in the Results and also [5]. Since PP2A, Zds1, and Pkc1 have been found to have around 3, 5, and 8 sites respectively, we implement this with parameters that produce the graph observed in Figure 3A. See Table 1 for a list of parameter values.
The initial conditions used in the model correspond to the system in the off state. They are equal to zero for all variables, except and .
|
10.1371/journal.pcbi.1000165 | A Genomewide Functional Network for the Laboratory Mouse | Establishing a functional network is invaluable to our understanding of gene function, pathways, and systems-level properties of an organism and can be a powerful resource in directing targeted experiments. In this study, we present a functional network for the laboratory mouse based on a Bayesian integration of diverse genetic and functional genomic data. The resulting network includes probabilistic functional linkages among 20,581 protein-coding genes. We show that this network can accurately predict novel functional assignments and network components and present experimental evidence for predictions related to Nanog homeobox (Nanog), a critical gene in mouse embryonic stem cell pluripotency. An analysis of the global topology of the mouse functional network reveals multiple biologically relevant systems-level features of the mouse proteome. Specifically, we identify the clustering coefficient as a critical characteristic of central modulators that affect diverse pathways as well as genes associated with different phenotype traits and diseases. In addition, a cross-species comparison of functional interactomes on a genomic scale revealed distinct functional characteristics of conserved neighborhoods as compared to subnetworks specific to higher organisms. Thus, our global functional network for the laboratory mouse provides the community with a key resource for discovering protein functions and novel pathway components as well as a tool for exploring systems-level topological and evolutionary features of cellular interactomes. To facilitate exploration of this network by the biomedical research community, we illustrate its application in function and disease gene discovery through an interactive, Web-based, publicly available interface at http://mouseNET.princeton.edu.
| Functionally related proteins interact in diverse ways to carry out biological processes, and each protein often participates in multiple pathways. Proteins are therefore organized into a complex network through which different functions of the cell are carried out. An accurate description of such a network is invaluable to our understanding of both the system-level features of a cell and those of an individual biological process. In this study, we used a probabilistic model to combine information from diverse genome-scale studies as well as individual investigations to generate a global functional network for mouse. Our analysis of the global topology of this network reveals biologically relevant systems-level characteristics of the mouse proteome, including conservation of functional neighborhoods and network features characteristic of known disease genes and key transcriptional regulators. We have made this network publicly available for search and dynamic exploration by researchers in the community. Our Web interface enables users to easily generate hypotheses regarding potential functional roles of uncharacterized proteins, investigate possible links between their proteins of interest and disease, and identify new players in specific biological processes.
| Establishing a functional network is invaluable to furthering our understanding of gene function, pathways, and systems-level properties of an organism and can be a powerful resource in directing targeted experiments. The availability of diverse genome-scale data enables the prediction of networks encompassing all or at least most of the proteins in an organism. In Saccharomyces cerevisiae, probabilistic models have been used to predict the genomewide protein–protein functional interactions by integrating diverse data types [1]–[6]. Such probabilistic approaches have also been used in mammals to predict physical interactions [7],[8] and to generate expression networks [9]–[13]. In human, functional relationship networks have also been generated by integrating diverse interaction data [14]. However, it is still challenging to predict functional relationships through integrating diverse genomic data in mammalian model systems, due to the intrinsic complexity of these genomes and functional biases in individual datasets. Yet recent accumulation of both traditional targeted experiments, including protein physical interactions [15]–[17], gene-disease/phenotypic associations [18] and genome-scale data including gene expression and tissue localization [19]–[21], phylogenetic and phenotypic profiles [22],[23], as well as data retrieved based on homology [2],[24] provides the basis for establishing a global functional relationship network in the laboratory mouse [25].
We describe here a functional network in mouse generated by integrating a wide range of data types. In contrast to interactomes that represent physical interactions, our functional network predicts the probability that two proteins are involved in the same biological process and thus represents a more comprehensive combination of physical, genetic and regulatory linkages (Figure 1A). We demonstrate the utility of our network to predict gene functions and pathway components by both computational and experimental approaches. Further, we demonstrate how it can be used to further our understanding of the systems-level features of the mouse functional network. Our global functional network for the laboratory mouse is a valuable resource for analysis and annotation of the mouse proteome and can be used as a means of generating biological hypotheses for subsequent experimental validation, especially through the interactive public web interface available at http://mouseNET.princeton.edu.
Bayesian networks have been used successfully for integrating diverse data sources in many biological settings, including protein function prediction [3],[6], prediction of genetic interactions [26], physical interactions [4],[7] and most relevant to this work, prediction of functional networks in S. cerevisiae [2],[5],[6] and human [14]. The Bayesian approach is especially well-suited to our problem, where many genome-scale data have missing values and collections of individual investigations may not be a complete representation of genome profiles. Based on a Bayesian framework, we designed a method that combines redundant datasets, processes continuous data, minimizes over-fitting and finally, integrates all experimental evidence (Table 1) in a confidence-based manner to estimate the genomewide pair-wise probabilities of functional linkage (Figure 1A). The resulting mouse interactome includes 20,581 genes, with edges representing the probability of functional relationship between each pair (Figure 1B). As demonstrated below, creation of this functional network through integrating diverse data sources can facilitate identification of novel pathway components and represents a powerful resource for understanding genetic diseases and network evolution.
A key application of a functional network prediction is to uncover novel pathway components. We first evaluated the accuracy of our predicted network through cross-validation analysis on known functional linkages (co-annotations of proteins to specific Gene Ontology [27] terms), which is the standard for unbiased computational evaluation. In short, cross-validation can be used to assess the accuracy of predictions by evaluating the system's accuracy in recovering subsets of known annotations withheld during the training process. Our integrated network is substantially more successful in predicting known functional linkages than any of the individual datasets and making more correct predictions (demonstrating higher precision) at every confidence cutoff (Figure 2A). This result is robust to using a different annotation standard, i.e., co-annotation to the same Kyoto Encyclopedia of Genes and Genomes [28] (KEGG) pathways (Figure 2B). Notably, although the relative performance of datasets varies with different standards, the consistently good performance of our results suggests that the integrated predictions are robust to variations in the annotation standard.
A common pitfall of many global integration schemes is the tendency to make precise predictions over only a limited set of biological processes [29]. Thus we evaluated the functional composition of our integrated results using KEGG, which is an accurate representation of our current knowledge of different pathways. The integrated network exhibits a balanced representation of a large group of pathways, even though many individual datasets have significant functional biases (Figure S1, the complete statistics of this functional composition analysis are included in the Dataset S1). For instance, the protein–protein interaction data obtained from the Biomolecular Interaction Network Database (BIND) [15] is significantly skewed towards the processes of focal adhesion. In contrast, given the broad functional coverage of the integrated network, we expect our approach will be useful in further characterization of a variety of pathways.
The high accuracy in predicting co-annotation to KEGG pathways (Figure 2B) by our network and its broad functional coverage (Figure S1) suggest that mouseNET can accurately capture pathway-based functional linkages for a variety of processes. We thus focused specifically on the predicted functional network for the major conserved signaling pathways related to development, including Hedgehog, Wnt, MAPK, TGF-β, Notch, and Toll-like receptor signaling pathways. We find that in addition to recovering known pathway components (Figure S2), these networks include a number of proteins not previously annotated to the pathway. Many of these novel predictions have reasonable experimental support in the literature. For example, in the 40 most tightly connected nodes surrounding known MAPK pathway proteins (Figure 3), 14 of them are annotated as the canonical pathway components in KEGG (p<10−10, hypergeometric distribution). Furthermore, two of the other nodes (Kit, MGI:96677 and Shh, MGI:98297) are not annotated to the MAPK pathway in KEGG but are annotated in the Gene Ontology [27] to be MAPK-related. Another nine unannotated predictions in the cluster of 40 have been suggested in literature to be involved in the MAPK pathway (Table S2 and Text S1). Thus, our system not only recovers well-established knowledge but also implicates novel pathway components, and therefore could be a powerful tool for generating hypotheses for experimental approaches.
Our genomewide prediction of protein function based on the integrated network produced 689 novel annotations with an estimated 80% precision. A subset of these new predictions was evaluated through examination of the literature by MGD curators and the precision estimate was confirmed (Dataset S2). Of these, 17 predictions were confirmed based on literature evidence at the level sufficient for annotation in MGI, and another six were found to have some support in the literature, but at a level not yet sufficient for GO annotation. For example, Retn (MGI:1888506), which does not have a GO biological process or KEGG pathway annotation, was predicted with high confidence (over 0.8) to be involved in glucose homeostasis (GO:0042593). The loss of Retn was indeed found to improve glucose homeostasis in leptin deficiency [30], confirming the prediction. This evaluation demonstrates that through integrating information from diverse sources, the system is capable of making accurate novel predictions on genes not previously annotated in GO or KEGG.
To further validate novel functional relationships predicted by our integrative network, we investigated proteins predicted to cluster around the homeobox transcription factor Nanog (MGI:1919200), which is an essential gene responsible for maintaining embryonic cell fate. Specifically, we experimentally down-regulated the expression of Nanog, and observed the nuclear protein expression changes of the top functional interactors in our predicted network by mass spectrometry. Five of the top 10 Nanog interactors predicted by mouseNET (Figure 4A) were detected in the nuclei and thus, we could evaluate their expression following Nanog down-regulation. We observed that after Nanog down-regulation, expression levels of four of them either significantly increased (DNA (cytosine-5-)-methyltransferase 3-like, Dnmt3l, MGI:1859287 and DNA methyltransferase 3B, Dnmt3b, MGI:1261819) or decreased (transformation related protein 53, Trp53, MGI:98834 and POU domain, class 5, transcription factor 1, Pou5f1, MGI:101893) (p<0.1 when compared to the overall distribution of the nucleus-detected proteins, Figure S8). Of those, Pou5f1 has also been previously shown to be involved in ES cell regulation [31],[32] and it has significant overlap in genomic binding targets with Nanog [33],[34]. Furthermore, the change in expression for these four proteins is consistent for different time points after Nanog knock-down, and increases consistently over the time course (Figure 4B). This experimental verification demonstrates that our system is a powerful tool which can aid researchers in generating accurate hypotheses for discovery of proteins involved in a specific cellular process.
Our functional network can also highlight information about physical interactions and transcriptional binding sites. For example, the 17 physical interactions with Nanog identified by Wang et al. were highly enriched in pairs of high functional relationship confidence (Mann-Whitney U test p = 0.00069). In addition, on the transcription level, the Nanog binding loci associated genes [34] were also highly enriched in high confidence functional interactors of Nanog predicted by our network (U test p = 3.98E-18). Therefore, by integrating a diverse collection of data, mouseNET enables users to explore variety types of functional associations, including physical interactions and transcriptional level regulation.
MouseNET provides a valuable resource to characterize the systems-level features of a model organism, which is a critical issue in understanding the organization and dynamics of the proteome. In the mouseNET network, the majority of proteins have only a small number of connections (Figure 5A), yet the presence of a few highly connected nodes (Figure 1B) implies central modifiers of the proteome. These ‘hub’ genes (at confidence cutoff 0.6) are enriched in regulation of response to stress, DNA metabolic process and cell cycle, (Bonferroni-corrected p<1.0E-9) (Table 2). Additionally, these hubs were significantly enriched (Bonferroni-corrected p = 8.3E-10) for ‘chromosome organization and biogenesis’, which is in agreement with a previous study in C. elegans that identified a class of genetic interaction hubs, all six of which were chromatin regulators [35].
We further analyzed the topology of the functional network surrounding these hubs and found distinct characteristics that correlate with their role in the cell. Proteins with high connectivity may appear in densely connected modules, or alternatively, they could be linkers of multiple functional modules and participate in several pathways [36]. To investigate these two classes, for each gene we computed the clustering coefficient, C, which gives the probability that its interactors are connected to each other. We found that low clustering coefficients, when controlled for node degree, are critical indicators of proteins participating in more biological pathways (Figure 5B). This trend is robust against different confidence cutoff levels for the interactions (Figure S3). For example, both nucleolar protein 1 (Nol1, MGI:107891) and paxillin (Pxn, MGI:108295) have 50 functional linkages with more than 0.6 confidence in interactions (Figure 5C and 5D). However, the former, which has a C of 0.44, is involved in only the rRNA processing pathway, while the latter, with a C of 0.06, is known to be involved in multiple biological processes, including activation of MAPK activity, branching morphogenesis of a tube, cell adhesion and protein folding. Furthermore, we found that the set of proteins with low clustering coefficients, but not the set of all proteins with only high node degree, is highly enriched for ‘signal transduction’ (Table 2), probably because proteins involved in signal transduction are central to cross-talk among multiple pathways and the cell's diverse response to various stimuli. Thus, the topology of the functional network contains important clues to the global organization of the proteome; and in addition to connectivity, we demonstrate that the clustering coefficient is a critical factor characterizing modifiers of multiple biological pathways.
Global modeling of functional linkages provides a general framework to analyze the relationship between local network properties and functional consequences of individual gene perturbations. For example, previous studies have predicted that the network connectivity is correlated with the propensity of a protein to be essential [37],[38]. Recently, however, there has been debate over whether this relationship is indeed true in yeast or human [39],[40], the main issue being whether high connectivity is truly a property of the underlying network or simply an effect of intense study of the essential gene set (i.e., annotation or investigational bias).
To address this question in the mouse functional network and control for investigation bias, we constructed two networks: one including all input data except knock-out phenotype information, and one including only whole-genome datasets. To avoid the caveat that not all gene knock-outs have been constructed, only genes that have been knocked out or targeted were included in all statistical analyses. For the first functional network, essential genes or disease-associated genes are significantly more connected than average (p<10−18 for perinatal lethality, p<10−9 for postnatal lethality, and p<10−6 for disease-associated genes, Mann-Whitney U test) (Figure S4A). However, in the functional network based on only whole-genome datasets, the difference between essential and non-essential sets was not significant, nor was that between disease-related set and the genome average (Figure 6A), suggesting the observed relationships between essentiality and network connectivity are likely to be explained by investigational biases in our case. This result is consistent with a previous study [41] which suggested that the vast majority of disease genes show no tendency to encode physical interaction hubs in human data. We further considered whether connectivity and local topology in our functional network relate to other perturbation phenotypes. Although most phenotype-responsible gene groups (Table S1) have a higher than average connectivity based on all available input data (Figure S4B), only proteins involved in tumorigenesis, embryogenesis still have significantly higher connectivity than average (p<0.05) on the whole-genome-data-only network (Figure 6B). This result highlights that the variation in intensity of study for genes can cause significant biases in the conclusions reached when comparing the connectivity of different groups of genes.
We observed that all groups of phenotype-associated genes have a lower clustering coefficient than average, and most participate in more biological pathways (Figure 6C). This conclusion holds true when controlling for investigational biases. For example, Trp53, with very high connectivity (Figure 1B) and particularly low clustering coefficient (0.02252), is essential during both embryonic perinatal and postnatal stages and plays a role in tumorigenesis, the reproductive system, and has ten other high level phenotypes (Table S1) according to the Mouse Genome Informatics (MGI) database [18]. This result implies that hubs with low clustering coefficient and participating in multiple pathways are important buffers of the genome, and that mutations or other disruptions of these genes are likely to be related to a detrimental phenotypes and, likely, disease.
Genome evolution on the sequence level has been studied intensively during the past decades. Studies of functional evolution on the genome-scale, on the other hand, require comprehensive profiling of proteins, which is difficult due to largely incomplete annotation of protein function in most organisms. Here, we demonstrate that mouseNET is a valuable resource for cross-species functional evolution studies by comparing it to the S. cerevisiae network [2]. To avoid circularity caused by integration of sequence similarity information, we generated a functional network that excludes all orthology-based input data. Given these mouse and yeast networks, we first investigated whether functional linkages are conserved between pairs of orthologs as identified through InParanoid [23]. Our results indicate that high-confidence functional linkages in S. cerevisiae are strongly predictive of functional linkages between orthologous gene pairs in mouse (Figure 7A for statistical analysis).
We also investigated the conservation of functional neighborhoods in the mouse and yeast networks. To make the datasets comparable, we included only orthologous pairs in the conservation statistical analysis. We found that the two networks vary from a high degree of conservation to almost no conservation (Figure 7B and 7C). Functional linkages between proteins involved in response to stress, response to endogenous stimulus, catabolic process, DNA metabolism, cell cycle, and other core biological processes and components were highly conserved between yeast and mouse (Table 3), e.g., the ribosomal protein L15 (Rpl15, MGI:1913730; Figure 7B and 7C). In contrast, functional relationships in processes specific to higher organisms, including, behavior, embryonic development, multicellular organismal development and anatomical structure morphogenesis were limited to the mouse network (Table 4). For example, the HtrA serine peptidase 1 (Htra1, MGI:1929076) plays a role in BMP signaling pathway [42], but its ortholog in yeast, YNL123W (Nma111, SGD: S000005067) is involved in apoptosis and lipid metabolic process [43],[44] (Figure 7B and 7C). The newly generated interactions for these mouse-specific functional networks originated through a combination of orthologous pairs in yeast and novel connections with existing genes or genes that have no ortholog in yeast (Figure 7B and 7C). Interestingly, ion transport was among the list of enriched processes for both conserved and unconserved subgraphs. We found that in conserved subgraphs, these genes were enriched in energy-coupled proton transport, which is conserved from yeast to mammals. In contrast, in the unconserved subgraphs, this enrichment of ion transport was due to genes involved in metal-ion or chloride transport, probably because of their involvement in the neural system. Details regarding the enrichment statistics are available in the Dataset S3.
Comparative analysis of interactomes between species, such as that presented above, is no doubt a promising approach for answering a number of fundamental biological questions [45]. Previous studies, e.g., [40], have demonstrated the sparsity of our current knowledge of physical interactions in many organisms, which has led to a very limited set of identified conserved interactions. As demonstrated here, the comparison of higher-coverage functional networks based on probabilistic models for integrating diverse genomic data provide an alternative solution for studying the evolution of functional linkages between proteins.
In this study, we combined diverse genetic and genomic data using a probabilistic framework to generate a functional network for the laboratory mouse. Our network accurately predicts functional linkages between mouse genes and covers a broad range of biological processes. We expect this view of the mouse proteome will be an invaluable resource in identifying novel pathway components and understanding system-level organization.
We have demonstrated several applications of our network in this study. First, we characterized the topology of the network and demonstrated that local network topology correlates with biological functions. Also, we used this genomewide view of functional linkages to investigate the relationship between diverse phenotypes and the local configuration of subnetworks. Finally, although network comparison across several species is limited by the sparsity of our current knowledge of physical interactions [40], generation of a functional network based on diverse data types also allowed us to examine the conservation of subnetworks on a global system level.
We provide a searchable interface for the exploration of the mouse functional network (http://mouseNET.princeton.edu). The interface also presents a full analysis of the functional enrichment of networks surrounding the genes(s) of interest and the disease genes in the local network. Through our interface, users could identify the original evidence supporting for specific functional linkages. The website includes integration results generated for the purpose of topological studies (controlled for investigational biases) and of cross-species network alignment studies (by excluding homology data) (http://mouseNET.princeton.edu/supplement/supplemental_data.htm). In the future, new publicly available genome-scale data will be added to our system, which will provide up-to-date support for hypothesis generation for questions ranging from individual protein function prediction to characterization of diverse system-level features.
In this study, we focused on the generation of a global functional network of mouse and demonstrated its wide applicability. Availability of tissue-specific datasets should allow us to generate tissue, cell, and developmental stage-specific network predictions using similar probabilistic frameworks. These tissue or developmental stage-specific networks will be more targeted and will be invaluable to the researchers of individual fields of study.
To build a functional network of proteins, we have collected a diverse set of evidence from several databases (Table 1). In order to predict pair-wise protein–protein relationships, all data were preprocessed, as described below, into pair-wise scores, reflecting the similarity between protein pairs. The databases included in our analysis are:
In the following section, we applied a naïve Bayes network to integrate all data sources and to predict pair-wise functional relationships. However, the application of a naïve Bayesian framework requires a non-trivial assumption of independence between individual evidence sources, which correspond to different evidence nodes in the naïve Bayes network. To address this issue, we evaluated the conditional independence between datasets and those with significant dependence were merged into a single evidence node. To determine whether two datasets should be merged, we calculated the likelihood ratio of each combination of datasets with and without the assumption of independence.(3)(4)where E is the score of the protein pair in dataset i or j, a FRY means a positive functional relationship (FR = 1) in gold standard, and FRN means a negative functional relationship (FR = 0).
Two conditionally independent datasets will have similar likelihood ratios calculated by the above two approaches (Figure S6A). In contrast, highly dependent datasets tend to have erroneously high likelihood ratios (Figure S6B) when they are treated as independent ones. After a complete analysis of the independence properties between every dataset pair, we found that phenotype data from MGI and disease data from OMIM are highly dependent on each other. As a result, we treated these phenotype and disease data as a single evidence node in the Bayesian network, and each of the remaining datasets as an individual evidence node.
As data sources are different in their accuracy of measurement as well as relevance for predicting protein functions, creating an accurate network for functional linkages requires a systematic approach that weights and integrates information from individual datasets. We applied a Bayesian network to integrate diverse data and make the final functional linkage predictions (Figure 1A). Specifically, we computed the posterior probability of a functional relationship given all available evidence as follows:(5)where FR represents functional relationship, Ei represents the score of the pair in each dataset i and Z is a normalization factor. Intuitively, this probability FRij for two proteins i and j represents how likely it is, given existing data and accuracy and coverage of each input dataset, that proteins i and j participate in the same biological process.
To learn the parameters in this Bayesian framework, we established a gold standard that approximates a true set of functionally related proteins. Mouse Genome Informatics (MGI) maintains curated annotations of Gene Ontology (GO) for mouse [53]. The sources of these annotations include (1) hand annotation from primary literature, (2) electronic annotation based on gene name and symbols, (3) annotation from SwissProt keywords, (4) Enzyme Commision (EC) numbers. These annotation sources are reasonably accurate for our analysis. We defined positive as pairs of proteins that are co-annotated to a specific Biological Process GO term (less than two hundred genes annotated to this GO term) and negatives as those in which both members of the pair have specific annotations but do not share any of them.
To model the posterior distribution given a set of data, we grouped the pair-wise values from each dataset into discrete groups. For binary datasets, for example, physical interactions, it is easy to separate the two categories where 0 means that there is no interaction between the pair, and 1 means that the interaction exists. Continuous pair-wise scores (e.g., expression profiles and phenotype/disease data) require a binning approach for discretization. We observed that for each dataset, the posteriors generally decreases with small fluctuation as the pair-wise score decreases (Figure S7). Thus, to avoid over-fitting to noise in the datasets, discretization was done so as to force the posteriors of the discretized bins to decrease as the average pair-wise score of those bins decreases.
An important application of such a functional network is to predict novel pathway components. We therefore applied our network to predict pathway components in KEGG [28]. For a specific pathway, during each iteration, 10 known genes were seeded into the weighted network and the rest of the genes were treated as unknowns. Thus for every other gene, we compute an adjacency to the 10 seeds. This process was repeated three hundred times with random samplings of the seed set. We then calculated the average adjacency for each gene:(6)where wi represents the weight of each gene and j represents the seed genes, and wijk represents the confidence, as estimated by our integration, of the functional relationship between protein i and j in iteration k. ni is the number of times gene i was not one of the seed genes. The top components and recovery curves were generated based on the ranking of wi.
To characterize the topology of the functional network, we calculated the connectivity and clustering coefficient C of all proteins. The clustering coefficient of a protein gives the probability that its neighbors are connected to each other. In a densely connected module or clique, C is close to one. C for each of the proteins was calculated as follows [54]:(7)where n denotes the number of links between k direct interactors.
We obtained GO annotations [27] from the Mouse Genome Informatics (MGI) [18] on Jan 18, 2007. The enrichment of each GO term was found using a hypergeometric distribution. The most enriched GO terms were represented by the lowest Bonferroni-corrected p value [55].
To facilitate wide access to the integrated functional network by the biology community, we implemented a web interface (http://mouseNET.princeton.edu) that allows the users to browse our predictions based on single or multiple protein queries. We have implemented a probabilistic algorithm that searches the direct or indirect neighbors with the largest adjacency to the query set [2]. GO term enrichment was calculated for the top neighbors, which facilitates fast discovery of unknown gene function.
We also provide the community with a list of gene function predictions based on our network for proteins with no currently known function. Specifically, we calculated the GO term enrichment of the top 40 nearest neighbors of each gene using the hypergeometric distribution. Then the per-function enrichment of each gene's top neighbors is reported as a Bonferroni-corrected p-value and thus their putative function is deduced.
The Nanog controllable embryonic stem cell lines were set up and tested by Natalia Ivanova, and were cultured as described [56]. The feeder cells, primary mouse embryonic fibroblasts, were removed before use. To down-regulate Nanog, we withdrew the doxycycline (1 g ml−1) from the media, but still supplied the cells with all the routine ES cell nutrients (DMEM with 15% FBS (Hyclone), 100 mM MEM non-essential amino acids, 0.1 mM 2-mercaptoethanol, 1 mM l-glutamine (Invitrogen), and 103 U ml-1 of LIF (Chemicon). For the nuclear protein measurement, nuclear protein samples were prepared with nuclear/cytosol fractionation kit (BioVision, catalog number: K266-100). The samples from four different time points were labeled by different isotope (iTRAQ) and then analyzed at a single run of mass spectrometry. We used ProQUANT (Applied Biosystems) and the ProGROUP (Applied Biosystems) software to identify proteins. The experiment was repeated three times. Proteins detected more than twice were included in the analysis and the average values were used.
|
10.1371/journal.pmed.1002865 | Progression to type 2 diabetes mellitus and associated risk factors after hyperglycemia first detected in pregnancy: A cross-sectional study in Cape Town, South Africa | Global data indicate that women with a history of hyperglycemia first detected in pregnancy (HFDP) are at up to 7 times risk of progressing to type 2 diabetes mellitus (T2DM) compared with their counterparts who have pregnancies that are not complicated by hyperglycemia. However, there are no data from the sub-Saharan African region, which has the highest projected rise in diabetes prevalence globally. The aim of this study was to determine the proportion of women who progress to T2DM and associated risk factors 5 to 6 years after HFDP in Cape Town, South Africa.
All women with HFDP, at a major referral hospital in Cape Town, were followed up 5 to 6 years later using a cross-sectional study. Each participant had a 75 g oral glucose tolerance test; anthropometric measurements and a survey were administered. A total of 220 participants were followed up. At this time, their mean age was 37.2 years (SD 6.0). Forty-eight percent (95% CI 41.2–54.4) progressed to T2DM, 5.5% (95% CI 3.1–9.4) had impaired fasting glucose, and 10.5% (95% CI 7.0–15.3) had impaired glucose tolerance. Of the participants who progressed to T2DM, 47% were unaware of their diabetes status. When HFDP was categorized post hoc according to WHO 2013 guidelines, progression in the diabetes in pregnancy (DIP) group was 81% (95% CI 70.2–89.0) and 31.3% (95% CI 24.4–39.3) in the gestational diabetes mellitus (GDM) category. Factors associated with risk of progression to T2DM were; at follow-up: waist circumference (odds ratios [OR] 1.1, 95% CI 1.0–1.1, p = 0.007), hip circumference (OR 0.9, 95% CI 0.8–1.0, p = 0.001), and BMI (OR 1.1, 95% CI 1.0–1.3, p = 0.001), and at baseline: insulin (OR 25.8, 95% CI 3.9–171.4, p = 0.001) and oral hypoglycaemic treatment during HFDP (OR 4.1, 95% CI 1.3–12.9, p = 0.018), fasting (OR 2.7, 95% CI 1.5–4.8, p = 0.001), and oral glucose tolerance test 2-hour glucose concentration at HFDP diagnosis (OR 4.3, 95% CI 2.4–7.7, p < 0.001). Our findings have limitations in that we did not include a control group of women without a history of HFDP.
The progression to T2DM in women with previous HFDP found in this study highlights the need for interventions to delay or prevent progression to T2DM after HFDP. In addition, interventions to prevent HFDP may also contribute to reducing the risk of T2DM.
| International research shows that when a woman has diabetes detected in pregnancy, which may resolve after the pregnancy, she remains at high risk of future diabetes.
However, we do not know what proportion of women progress to type 2 diabetes in Africa, because there is no research that has been done before, despite the rapid increase in the number of people with diabetes.
We recalled 220 women 5 to 6 years after they had diabetes first detected in pregnancy and tested them for type 2 diabetes in Cape Town, South Africa.
We found that almost half of the women (48%) at follow up had type 2 diabetes. Of the women with type 2 diabetes, 47% did not know that they had type 2 diabetes.
We also found that being obese at follow-up and having higher blood glucose concentrations at the time the women were tested for diabetes in pregnancy (DIP) increased the chances of progressing to type 2 diabetes 5 to 6 years after the pregnancy.
A large proportion of South African women who have diabetes first detected in pregnancy may develop type 2 diabetes at an early age and within 6 years after the pregnancy.
It may be necessary to change their lifestyle after the pregnancy so they can reduce the chance of progressing to type 2 diabetes.
Screening for type 2 diabetes after the pregnancy needs to be more often so women who develop diabetes are diagnosed and receive treatment earlier.
Further research is needed, because we did not include women with normal blood glucose during pregnancy in this study.
| Sub-Saharan Africa, compared with other regions, is expected to have the greatest increase in the number of people living with diabetes by the year 2040, with more than half the people affected unaware of their diabetes status [1]. Since 2015, diabetes has already risen to be the second leading cause of death, after tuberculosis, in South Africa [2]. The prevalence of obesity, the strongest known risk factor for type 2 diabetes, has increased across the world, and more so in African women, with a recent meta-analysis showing that in this group, mean body mass index (BMI) increased from 22 kg/m2 in 1980 to 25 kg/m2 in 2014 [3]. In South Africa, the combined obesity and overweight prevalence increased from 29% to 40% in men and 57% to 70% in women during the period of 2002 to 2016 [4,5]. Other drivers of the diabetes epidemic, such as poor nutrition and decreased physical activity, have also increased during the last 2 decades [6]. Further, HIV antiretroviral therapy–induced lipodystrophy may also increase risk of diabetes, especially in women of childbearing age, who are disproportionally affected by HIV, compared with their male counterparts [6]. In view of the current high burden of diabetes and the expected rise in diabetes prevalence, it is imperative to identify populations at elevated risk and introduce risk-lowering interventions.
Women with a history of hyperglycemia first detected in pregnancy (HFDP), including gestational diabetes mellitus (GDM), are at high risk of future development of T2DM [7]. Initially, GDM was defined based on the risk of developing T2DM, but this may have resulted in the inclusion of women with undiagnosed diabetes in the GDM subgroup. Following the recommendations of the International Association of Diabetes and Pregnancy Study Group (IADPSG) [8] and the publication of the findings from the Hyperglycaemia and Adverse Pregnancy Outcomes (HAPO) Study [9]—a multicenter study with participants from 10 countries—WHO [10], in 2013, defined HFDP as either diabetes in pregnancy (DIP) or GDM. According to WHO, GDM is now diagnosed as glucose intolerance in pregnancy with fasting glucose values between 5.1 and 6.9 mmol/L and/or oral glucose tolerance test (OGTT) 2-hour glucose concentrations between 8.5 and 11.0 mmol/L, whereas women with blood glucose values diagnostic of type 2 diabetes first discovered in pregnancy are classified as having DIP. The HAPO Study demonstrated associations between fasting glucose concentrations, as low as 5.1 mmol/L at HFDP diagnosis and adverse fetal outcomes at birth, whereas the HAPO follow-up study [11] and others [12] showed a high risk for T2DM in women in the postpartum period as well as long term and adiposity in their offspring. Notably, neither the HAPO Study nor the follow-up studies included data from an African cohort. Despite the absence of data from African countries, it is expected that lower fasting glucose concentration cut-offs for HFDP diagnosis, in addition to increased awareness and improved screening as well as increasing calls for universal screening for HFDP, may result in a higher prevalence of HFDP worldwide, especially in transitioning populations, such as South Africa. In China, for example, a 4-fold increase in GDM prevalence was noted when universal screening was introduced [13].
Prior to the introduction of the term HFDP, most studies used the term GDM to describe any hyperglycemia first detected during pregnancy. In this article, we use the term HFDP where the studies may have used the term GDM, using older criteria in which the DIP subgroup was possibly included. The prevalence of HFDP varies in different populations, although this is complicated by the use of different diagnostic criteria as well as different screening methods for hyperglycemia during pregnancy [14]. HFDP prevalence from a systematic review of the small number of available studies in Africa ranged from 0% to 14% [15]. Recent studies that used the IADPSG [8] criteria for GDM diagnosis reported prevalence of 8.9% in Nigeria [16], 2.9% in Kenya [17], and in South Africa, 9.1% in Soweto [18] and 25.8% [19] in Johannesburg. The Johannesburg estimate of 25.8% may have included women with DIP and therefore could be an estimate of HFDP. Using the conservative Soweto estimate, if 1 in every 11 pregnancies is complicated by GDM, then public health interventions are required to prevent or delay T2DM in these women post the index pregnancy in South Africa. However, the paucity of data on the prevalence of and associated risk factors for T2DM, in women after GDM, in sub-Saharan Africa, and in South Africa may hinder effective development and planning of interventions and policies.
Data from meta-analyses of studies, mostly from high-income countries, show that women with previous GDM have up to 7-fold risk of developing T2DM [7,20], increased risk of long term cardiovascular disease [7], and for the offspring, increased risk of immediate adverse perinatal as well as future cardiometabolic disease risk [7] compared with those with nondiabetic pregnancies. Further, the risk of progression to T2DM is highest during the period 3 to 6 years post GDM [20]. However, the estimated risks may be overestimates because most of the included studies used older GDM criteria that included women with DIP. In addition, there is a great degree of heterogeneity in the risk for T2DM, with relative risks ranging from 2.7 in Germany to 38.4 in Sweden [14]. The estimates of risk vary by country and within countries, by ethnicity and region, which may be due to differences in the distribution of risk factors of T2DM in different populations. Different follow-up times and different study designs may also contribute to the differences in the risk estimates.
Progression to T2DM post HFDP varies widely, from a low 6% in Australian nonindigenous women [21] to 42% in Indian women [22], using IADPSG criteria. Risk factors for progression also vary widely—ethnicity, increased BMI, family history of T2DM, increased waist circumference and severity of GDM at diagnosis being some of the most frequently identified [23]. In Africa, apart from a single study that followed up 77 women up to 12 weeks post HFDP in Cape Town [24], to our knowledge, there are no data on the progression to T2DM post HFDP or associated risk factors. This study in women 5 to 6 years post HFDP provides the only data to date on the proportion of women who progress to T2DM beyond the postpartum period, as well as factors associated with risk of progression, in Africa, specifically in Cape Town, South Africa. We also investigated the proportion of women who progressed to T2DM in the GDM and DIP groups using the modified WHO 2013 criteria, applied retrospectively.
The study was carried according to an ethics approved study protocol (S1 Doc). Data on all women managed for HFDP at Groote Schuur Hospital (GSH) during the period of 1 September 2010 to 31 August 2011 were routinely collected during the index pregnancy [25]. During that time, in the Western Cape province of South Africa, GDM screening and diagnosis was based on the provincial guidelines [26]. Screening was based on selective risk factors—maternal age ≥40 years, BMI ≥40 kg/m2, previous GDM, previous fetal birth weight ≥4.5kg, previous unexplained miscarriage, acanthosis nigricans and polycystic ovarian syndrome—whereas GDM was diagnosed using the United Kingdom National Institute for Health and Care Excellence (NICE) 2008 criteria (fasting glucose above 5.5 mmol/L and OGTT 2-hour glucose over 7.8 mmol/L) [27].
A cross-sectional study of the same participants (n = 498) was undertaken 5 to 6 years later during the period of 1 January 2016 to 31 Jan 2017. We contacted and invited participants through letters mailed to their last known address, calls to their telephone or cell phone numbers in the hospital record or next of kin and finally, and home visits when all other attempts failed. Women who were pregnant at follow-up were excluded from the study.
On the day of testing, participants underwent a standard 75-gram OGTT after fasting for 8 to10 hours. Blood was drawn for glycated haemoglobin A1c (HbA1C) as well as glucose and insulin at fasting and 120 minutes post OGTT glucose load. The blood samples were kept on ice, aliquoted within 4 hours of collection, and stored at −80° until analyzed. Participants on treatment for T2DM were not required to do either the OGTT or the HbA1C. A trained fieldworker administered a questionnaire (S2 Doc) to obtain sociodemographic information, reproductive history, self-reported personal and family medical history, and psychosocial health and lifestyle factors such as physical activity (modified WHO Global Physical Activity Questionnaire), smoking, and diet using a 2-week food frequency questionnaire.
Height, weight, waist, and hip circumference and blood pressure were measured using standardized procedures. Waist-hip ratio was calculated as the ratio of each participant’s waist circumference to their hip circumference. BMI was grouped according to WHO criteria for underweight (<18.5 kg/m2), normal weight (18.5–24.9 kg/m2), overweight (25–29.9 kg/m2), obese (30–39.9 kg/m2), and morbidly obese (>40 kg/m2) [28]. Outcomes for each participant were T2DM, impaired fasting glucose (IFG), and impaired glucose tolerance (IGT) using WHO 2006 criteria [29].
Plasma glucose was measured using the Randox RX Daytona Chemistry Analyzer. HbA1C was measured using turbidimetric inhibition immunoassay (D10TM Haemoglobin A1c Program; Bio Rad Laboratories, Hercules, CA). The precision and trueness of the Randox RX Daytona Chemistry Analyzer were verified using the Clinical and Laboratory Standards Institute document EP15. Coefficients of variation calculated from running 40 separate samples at 3 different times were 3.0% for glucose and 1.6% for HbA1C.
The sample size for this study was based on the main aim: to estimate the proportion of participants who progressed to T2DM by the time of follow-up. Most studies found a prevalence of T2DM during the first 5 years after GDM diagnosis between 20% and 50% [12,30]. Using Open Epi sample size calculator for a proportion (http://www.openepi.com/SampleSize/SS), assuming that 35% of our participants would have progressed to T2DM and using the range 20% to 50% from literature (i.e., 15% either side of our assumed proportion), the minimum sample size required was 154. We anticipated difficulties in following-up women in our setting and therefore decided to include all women who we could contact and who agreed to participate.
All statistical analysis was carried out using Stata 15 statistical software [31]. For all hypothesis testing and comparisons, significance was set at 0.05, whereas 95% CIs were reported for the prevalence of T2DM as well as all odds ratios (ORs). Means and SDs were presented for normally distributed measured variables, medians and interquartile ranges (IQRs) for variables that were not normally distributed, and for categorical variables, frequencies and proportions were reported.
To compare variables between participants who progressed to T2DM and those who did not, chi-squared test and Fischer’s Exact (small frequencies) were used for hypothesis testing for categorical data, whereas the t test for independent groups (or Wilcoxon rank sum test if data were not normally distributed) were used to compare measured data.
The analysis for factors associated with T2DM at follow-up was redone after input from journal reviewers, with the main change being using continuous variables (BMI, age, waist and hip circumference) in their raw, and not categorized forms. We carried out a multiple logistic regression model that included variables that have been shown to be associated with risk of T2DM. Variables included from data measured at follow-up were age, anthropometry (BMI, hip and waist circumference), socioeconomic variables (education and employment), comorbidities (self-reported dyslipidemia and high blood pressure), total physical activity from the Global Physical Activity Questionnaire (GPAQ) and family history of diabetes. Variables included from baseline measurements were OGTT glucose concentrations at diagnosis of HFDP and type of treatment for HFDP. Stopping alcohol because of health reasons (n = 58 with responses) was not included in the multivariate regression because there were too many missing or “not applicable” data. We also did not include waist-hip ratio in the model because of the very wide 95% CI. Further OGTT 1-hour glucose at HFDP diagnosis was also not included because of its limited clinical utility and because most health facilities in South Africa do not measure it. For logistic regression model diagnostics, we assessed the following: linearity assumption using the Lowes graph, multicollinearity using variance inflation factors, model specification using the C-statistic, and confirmed the fit of the model using the Hosmer-Lemeshow goodness of fit test. We also checked for outliers as well as influential observations.
This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 STROBE Checklist).
The study received ethics clearance from the Human Research Ethics Committees of the University of Cape Town (Reference 656/2015) as well as permission to conduct research at the GSH. Participants gave written informed consent. If found to have undiagnosed T2DM, participants were referred for treatment.
Of the 498 eligible women, 220 (44.2%) participated in the follow-up study, 234 (47.0%) could not be contacted, and 44 (8.8%) could not participate (Fig 1). There were no major differences between participants who were followed up and those lost to follow-up, except that participants who followed up had a higher mean BMI at booking and a lower mean OGTT 2-hour glucose concentration at the time of HFDP diagnosis (S1 Table).
As seen in Table 1, at baseline during the index pregnancy, the mean age was 30.8 years (SD 5.9), most participants (142 [65.4%]) were of mixed ancestry, followed by 68 (31.3%) who self-identified as Black. More than three-quarters had a first-degree family history of diabetes, whereas most participants were either obese (96 [44.9%]) or morbidly obese (49 [22.9%]). Just under a third of the participants (27.7%) received oral hypoglycemic treatment, and 23.6% had insulin therapy for HFDP. At HFDP diagnosis, 70 participants had FPG ≥7 mmol/L and/or 2-hour blood glucose concentrations ≥11.1 mmol/L and were retrospectively classified as DIP, and the remaining 150 (68.2%) had FPG between 5.6 and 6.9 mmol/L and/or 2-hour blood glucose concentrations 7.8 to 11.0 mmol/L, retrospectively, classified as GDM. Our classification of GDM differed from WHO 2013 guidelines in that our cohort did not include women with fasting glucose values lower than 5.5 mmol/L, whereas we included women with OGTT 2-hour blood glucose between 7.8 and 8.4 mmol/L.
Table 2 shows the characteristics of the participants at follow-up. At follow-up, the mean age of the participants was 37.2 years (SD 6.0). Most of the participants (167 [75.9%]) had secondary or matric level education. More than two-thirds of the participants were either obese (96 [44.9%]) or morbidly obese (49 [22.9%]).
At the time of follow-up, 47.7% (n = 105, 95% CI 41.2–54.4) progressed to T2DM, of these 47.1% were not previously diagnosed, 12 participants had IFG (5.5%, 95% CI 3.1–9.4), and 23 participants (10.5%, 95% CI 7.0–15.3) had IGT (Fig 2). Using an HbA1C ≥6.5 and an established T2DM diagnosis, the T2DM prevalence was 49.5% (95% CI 42.8–56.2). When HFDP was categorized post hoc, according to the modified WHO 2013 criteria, progression to T2DM in the DIP group was 82.9% (95% CI 72.0–90.1) and 31.3% (95% CI 24.4–39.3) in the GDM category.
Participants who progressed to T2DM compared with those who did not progress had significantly higher median glucose concentrations (in mmol/L) at all times of the OGTT at HFDP diagnosis—fasting: 6.4 (IQR 5.7–7.2 versus 5.6 (IQR 4.9–5.9), p < 0.001; 1-hour: 11.0 (IQR 10.0–12.2) versus 9.8 (IQR 8.5–10.6), p < 0.001, and 2-hour glucose: 10.1 (8.6–11.1) versus 8.6 (8.1–9.3), p < 0.001)—and were more likely to be on either oral hypoglycemic (36.2% versus 20.0%, p = 0.007) or insulin therapy (41.0% versus 7.8%, p < 0.001) during HFDP (Tables 1 and 2).
At follow-up, compared with participants without T2DM, participants who progressed to T2DM were significantly less likely to have a tertiary level education but more likely to have primary school level education (tertiary: 8.7% versus 18.3%, primary: 14.6% versus 5.2%, respectively, p = 0.017), more likely to report having dyslipidaemia (21.0% versus 6.1%, p = 0.001), more likely to have stopped drinking alcohol for health reasons (8.6% versus 4.4%, p = 0.038), and more likely to have gained less BMI (in kg/m2, [median 0.0 (IQR −3.0 to 2.8) versus 1.6 (−1.1 to 4.0), respectively, p = 0.019]). Box plots comparing fasting and OGTT 2-hour glucose levels at HFDP diagnosis, waist, and hip circumferences, and waist-hip ratio at follow-up, by T2DM status at follow-up, are shown in S1 Fig, S2 Fig, S3 Fig, S4 Fig, and S5 Fig, respectively.
Fig 3 shows the results of multiple logistic regression for variables independently associated with T2DM. Baseline variables significantly associated with risk of progression of T2DM were fasting glucose at HFDP diagnosis (OR 2.7, 95% CI 1.5–4.8, p = 0.001) and OGTT 2-hour glucose concentration at HFDP diagnosis (OR 4.3, 95% CI 2.4–7.7, p < 0.001), oral hypoglycaemic treatment for HFDP (OR 4.1, 95% CI 1.3–12.9, p = 0.018), and insulin treatment during HFDP (OR 25.8, 95% CI 3.9–171.4, p = 0.001). The following variables measured at the time of follow-up were significantly associated with progression to T2DM: having primary school education only, compared with tertiary education (OR 16.2, 95% CI 1.1–244.3, p = 0.044), self-reported dyslipidaemia diagnosis (OR 72.0, 95% CI 7.6–682.6, p < 0.001), self-reported hypertension diagnosis (OR 5.0, 95% CI 1.6–15.6, p = 0.006), BMI (OR 1.1, 95% CI 1.0–1.3, p = 0.001), waist circumference (OR 1.1, 95% CI 1.0–1.1, p = 0.007), and hip circumference (OR 0.9, 95% CI 0.8–1.0, p = 0.001).
The model consisted of 200 observations after the removal of outliers (n = 17) and the omission of participants with missing data (n = 3). In the final model, the p-values for the C-statistic (_hatsq) and the Hosmer-Lemeshow statistic were 0.123 and 0.809, respectively, confirming good fit for the model. There was no significant collinearity because pairwise correlations resulted in variance inflation factors between 1.04 and 1.82. The Lowes graph confirmed the linear model assumption.
Our major findings are that in women 5 to 6 years post HFDP, only 36.4% had normal glucose tolerance; 47.7% had progressed to T2DM, of whom 47% were previously undiagnosed, 5.5% had IFG, and 10.9% IGT. When we further categorized the HFDP post hoc using modified WHO 2013 GDM criteria, progression to T2DM was 83% and 31% in the DIP and GDM categories, respectively. Factors associated with risk of T2DM were fasting and OGTT 2-hour glucose concentration at HFDP diagnosis, oral hypoglycaemic and insulin treatment during HFDP, primary school education, BMI, and waist and hip circumferences at follow-up.
A key consideration for this study is the impact of GDM definition changes on the progression to type 2 diabetes. Recommendations of the IADPSG [8], based on findings from the HAPO Study were adopted by WHO in 2013 [10], and since then, most regional bodies have moved towards adopting WHO guidelines. Consequently, most studies published before 2013 used GDM definitions, such as the 1999 WHO guidelines on the diagnosis of GDM, which included both women with GDM and women with DIP and may therefore have overestimated the progression to T2DM proportion. In S2 Table, we have listed studies that investigated progression to T2DM in the medium- to long-term postpartum period, published during the period of 2000 to 2019, and the proportion of women who progressed to T2DM from each study. Most of these studies [32–46] used either WHO 1999 guidelines or other criteria, whereas only 4 studies used the IADPSG or equivalent criteria [11,21,22,47] for the diagnosis of GDM. Therefore, any comparisons of our findings with published data will need consideration of the heterogeneity of HFDP and GDM definitions.
In the South African context, the T2DM prevalence in our study population is 4 times higher than that of South African women overall (11%) [48] and higher than the T2DM prevalence in black women aged between 25 and 74 years (13.8%) [49] or women of mixed ancestry aged over 30 years (28.2%) [50] in Cape Town. Clearly, South African women with a history of HFDP are a vulnerable population and require intervention to delay or prevent progression to T2DM. The high proportion of women who progressed to T2DM (48%) could be explained partly by the possibility that, for some of the women, their glucose never returned to normality because they were not evaluated 6 weeks post the index pregnancy. This highlights the need for postpartum screening in these women. Global data indicates that postpartum screening is between 24% and 58% [51], whereas, in South Africa, less than 30% attend the recommended 6 weeks postpartum OGTT [24]. Several barriers from both the health system and patient perspectives hinder the 6-week postpartum screening. The South African health system is overburdened [52], and postpartum screening for diabetes at 6 weeks using the recommended 2-hour OGTT would add significantly to the burden. Although there are no South African data on barriers to postpartum screening, other studies have shown that the inconvenience of the OGTT and lack of time are the main reasons women do not attend the postpartum screening [51]. There is ongoing debate on the utility of either fasting glucose only or the HbA1C for the postpartum screening [51]. Research is required to establish the optimum method to replace the OGTT, and for the HbA1C, both optimum timing for screening as well as cut-offs for the diagnosis of type 2 diabetes in African women. In the Western Cape, after delivery, the women must attend diabetes screening at a separate clinic while taking their offspring to a well-baby clinic for vaccination and follow-up, which may result in most women prioritizing the baby’s care over theirs. Studies investigating the barriers to postpartum screening as well as optimum screening methods in South African women are needed.
We found very different proportions of women who progressed to T2DM between the GDM and DIP groups. The proportion of women who progressed to T2DM in the GDM group was 31% when we recategorized the women using modified WHO 2013 criteria. The high proportion of women who progressed to T2DM in the DIP group (83% on OGTT alone but 96% on both HBA1C and OGTT) suggests that they may possibly have had T2DM before the pregnancy. However, they clearly had more severe glucose intolerance during the pregnancy compared with the GDM group. When HbA1C assessment was added, only 3 (4%) women in the DIP group did not progress to T2DM. Further analysis of the DIP group showed discordance between the OGTT results and HbA1C; 2 of the 7 participants with either impaired glucose intolerance or IFG had HbA1C levels above 6.5% (7.2% and 8.6%), whereas 3 out of 6 participants with normal GTT had HbA1C levels of at least 6.5% (6.5%, 6.5%, and 6.6%). The remaining 3 participants with normal GTT had HbA1C levels below 6.5%. Our data, although in a small sample of women with DIP, suggests the need for more structured follow-up for assessment for T2DM after the pregnancy.
Comparisons of our findings with other studies that have investigated progression to T2DM is complicated by several issues. Firstly, there are no African studies that have investigated progression to T2DM post HFDP; the HAPO studies did not include an African cohort. Secondly, and more importantly, the heterogeneous definitions used for HFDP and GDM in the published studies (S2 Table) make it difficult to compare proportions of women who progressed to T2DM. Lastly, comparisons with published data are further complicated by the different study designs and different lengths of follow-up from the different studies. The proportion of women who progressed to T2DM of 31% in our GDM group, classified according to a modified WHO 2013 criteria, is somewhat high, compared with the 4 studies [11, 21, 22, 47] that used either the IADPSG or other criteria almost similar to it. This may be due to the cut-offs we used for GDM. Our study population is slightly different in that we did not include women with fasting glucose values between 5.1 and 5.5mmol/L, whereas we included women with OGTT 2-hour glucose values between 7.8 and 8.4mmol/L. The women in our study population, in terms of diagnostic glucose values, would have been almost similar to those included by Chamberlain and colleagues [21], in which widely different proportions of progression to T2DM for indigenous (25.5%) and nonindigenous women (5.7%) at 5 years post partum were reported. The proportion of women who progressed in the indigenous women was fairly similar to our study. A study from India by Gupta and colleagues [22] in women diagnosed using the IADPSG GDM criteria found that 25% and 42% of women aged 20 to 29 years and 30 to 39 years, respectively, progressed to T2DM in 5 years. The remaining 2 prospective cohort studies that used the IADPSG criteria for the diagnosis of GDM had follow-up periods that are very different from ours. In Japan, Inoue and colleagues [47] found that 22% progressed to T2DM 2 years post GDM, whereas 7.9% progression was observed in the HAPO Study [11] after a median follow-up of 11.4 years. It seems that progression to T2DM is heterogeneous, even when similar criteria for GDM diagnosis are used.
Identifying risk factors for the risk of progression to T2DM is a necessary step when designing interventions to delay or prevent T2DM. The risk factors for progression in our study are largely similar to findings from previous studies: fasting [53] and 2-hour OGTT glucose concentration [47,54] at HFDP diagnosis, and, at follow-up, BMI, waist and hip circumferences [39,54,55]. Insulin and oral hypoglycaemic treatment during HFDP are an indicator of the severity of HFDP and, in our study, 77% of women who had insulin treatment were classified as DIP. Of the women with dyslipidemia, 66% were already on treatment for diabetes at the time of follow-up, and it is well known that uncontrolled diabetes is associated with higher triglyceride and lower high density lipoprotein cholesterol (HDL-C) levels [56], and the participants with an established diabetes diagnosis were more likely to have been screened for dyslipidemia as part of standard care [57]. Although no other long term follow-up studies have shown a similar association between T2DM and education as ours, Gante [58] found an association between lower education and persistent postpartum glucose disorders in Portuguese women at 6 weeks follow-up. In our study and our setting, education is a good indicator of socioeconomic status, and therefore may be associated with an inability to access healthier lifestyle options such as better diets. Women with lower education may also not be able to access information on reducing T2DM risk after HFDP; therefore special interventions may be required for this group.
Preventing T2DM can be achieved through either population-wide approaches, such as the sugar tax, or interventions targeted at high-risk populations. The latter requires the screening and identification of high-risk individuals and offering interventions. Various diabetes prevention programs in both high-income and low-to-medium-income countries [46–49] have shown that lifestyle interventions can reduce the risk of T2DM in high-risk populations, such as people with IGT, although screening for IGT in a population can be expensive and difficult. Our study highlights the notion that women with a history of GDM are an obvious and easily accessible target for prevention because they are diagnosed as part of routine care in the health system. An added benefit of this approach is that by targeting these women, there is a real chance of decreasing the risk of intergenerational transmission of T2DM to the offspring. In South Africa, there are increasing calls for universal screening for GDM [18,19], which is costly and adds to the workload of health workers compared with the risk-factor–based screening, which leaves a substantial proportion of women with GDM unscreened. Regardless of the screening approach used, research on the efficacy or effectiveness of lifestyle interventions in preventing or delaying progression to T2DM in women post HFDP in South Africa would provide much-needed data.
Our study has several limitations. We were only able to follow-up with 44.2% of women after 5 to 6 years, comparable with other studies [46,47,53] and partly explained by a highly mobile population in the Western Cape, where in-and-out migration is common [59]. The women who participated were more likely to book 2 weeks early (15 versus 17 weeks), had a higher BMI at booking by 2 units (34.6 versus 32.7 kgm2), and had lower 2-hour OGTT blood glucose at HFDP diagnosis (9.0 versus 12.0 mmol/L) compared with the women who were lost to follow-up and therefore not completely representative of our study population. Due to the design, our study did not follow up women until diabetes developed, and therefore we do not have time to development of diabetes, as well as being unable to establish temporality for any of the risk factors we identified. The lack of a control group of women with normoglycemic pregnancies at the same time as our sample is a further limitation. However, when compared with recent T2DM prevalence in similar aged women in the Western Cape, our data indicate a high T2DM prevalence in women with a history of HFDP. More robust studies, with control groups, may be needed to further investigate our findings.
Almost half of the women with a history of HFDP progress to T2DM within 5 to 6 years, with almost half of them undiagnosed, in Cape Town, South Africa. There is a need for postpartum screening and interventions to reduce the risk of progression.
|
10.1371/journal.pntd.0007061 | Population genetic structure and geographical variation in Neotricula aperta (Gastropoda: Pomatiopsidae), the snail intermediate host of Schistosoma mekongi (Digenea: Schistosomatidae) | Neotricula aperta is the snail-intermediate host of the parasitic blood-fluke Schistosoma mekongi which causes Mekong schistosomiasis in Cambodia and the Lao PDR. Despite numerous phylogenetic studies only one DNA-sequence based population-genetic study of N. aperta had been published, and the origin, structure and persistence of N. aperta were poorly understood. Consequently, a phylogenetic and population genetic study was performed, with addition of new data to pre-existing DNA-sequences for N. aperta from remote and inaccessible habitats, including one new taxon from Laos and 505 bp of additional DNA-sequence for all sampled taxa,.
Spatial Principal Component Analysis revealed the presence of significant spatial-genetic clustering. Genetic-distance-based clustering indicated four populations with near perfect match to a priori defined ecogeographical regions. Spring-dwelling taxa were found to form an ecological isolate relative to other N. aperta. The poor dispersal capabilities suggested by spatial-genetic analyses were confirmed by Bayesian inference of migration rates. Population divergence time estimation implied a mid-Miocene colonisation of the present range, with immediate and rapid radiation in each ecogeographical region. Estimated effective population sizes were large (120–310 thousand).
The strong spatial-genetic structure confirmed the poor dispersal capabilities of N. aperta—suggesting human-mediated reintroduction of disease to controlled areas as the primary reason for control failure. The isolation of the spring-dwelling taxa and ecogeographical structure suggests adaptation of sub-populations to different habitats; the epidemiological significance of this needs investigation. The large effective population sizes indicate that the high population densities observed in surveyed habitats are also present in inaccessible areas; affording great potential for recrudescence driven by animal-reservoir transmission in remote streams. Mid-Miocene colonisation implies heterochronous evolution of these snails and associated schistosomes and suggests against coevolution of snail and parasite. Heterochronicity favours ecological factors as shapers of host-parasite specificity and greater potential for escape from schistosomiasis control through host-switching.
| The disease Mekong schistosomiasis poses a threat to the health of about 1.5 million people living near the Mekong river and its tributaries in Cambodia and Laos. It is a water-borne parasite transmitted by direct contact with water in which freshwater snails of the species Neotricula aperta live. Control of the snails is an effective approach to control of the parasite; however, because many suitable habitats for N. aperta occur in remote and inaccessible areas, knowledge of N. aperta population sizes and interconnectivity is insufficient for the design of effective snail control interventions. Although much of the region is difficult to survey by conventional means, population genetics can be used to estimate population structure and total size from small samples of accessible populations. The study added to existing data-sets, to give more population samples and longer DNA-sequences, together with improved analytical approaches to provide a better overview of N. aperta. The findings suggest that N. aperta in different kinds of habitats are also genetically different, with very low levels of migration between them; this genetic clustering is greater than expected from spatial distance alone. Further work is needed to determine if these different clusters vary in ability to transmit the parasite. The overall population size estimates were very large; thus suggesting that high snail population densities observed in accessible habitats are also characteristic of inaccessible populations—parasites are therefore more likely to return after disease control by immigration from remote areas. Finally, the timing of evolutionary events for snails and parasites was found to differ; this implies that the parasite may not be as strongly restricted to one species of snail as originally thought, which has implications for avoidance of parasite control by host-switching.
| Mekong schistosomiasis is a debilitating disease caused by infection with the parasitic blood-fluke Schistosoma mekongi Voge, Bruckner & Bruce 1978 [1]. An estimated 1.5 million people in Cambodia and Laos are at risk of infection by this schistosome [2]. The life-cycles of schistosomes require a snail-intermediate host, often species within a particular genus; however, transmission of S. mekongi is known to be restricted to only a single strain of the caenogastropod snail Neotricula aperta (Temcharoen 1971) (Pomatiopsidae: Triculinae) [3]. Transmission of Mekong schistosomiasis is highly focal and known only from seven foci involving the Mekong and three tributary river systems; thus the total range of the parasite is a mere 300 km section of the lower Mekong drainage between Khong Island (Lao PDR or Laos) and Kratié (Cambodia) [4]. By contrast the range of the snail-intermediate host is much greater, although still markedly discontinuous and geographically limited. Prior to 2014, N. aperta was known from 31 localities in Cambodia, Laos and Thailand, involving nine river systems of the lower Mekong basin [4]; however, the snail was not known from the Mekong river, or its tributaries, upstream of Khammouanne Province in Central Laos. In 2014 N. aperta was reported from the Mekong river at Ban Tha Kathin, Sri Chiang Mai District (Nong Khai Province, Thailand), which is over 200 km upstream of any previously recorded population. Interestingly, the snails in Nong Khai were found on rocks incorporated within concrete as part of anti-erosion defences along the river [5]. N. aperta had been previously found only in naturally-sited smooth stones in shallow areas of the rivers, or on submerged wood, and never on anthropogenic constructs [6].
The three strains of N. aperta, recognised on the basis of body size and mantle pigmentation [7], do not occur in sympatry and are remarkably limited in their distribution. The α-strain, the largest, being 3.5 mm in shell height on average [6], is found only in two ephemeral pools that form beside the main channel of the Mekong river during the dry-season at Khemmarat in Northeast Thailand. The β-strain is found only in the Mul river of Northeast Thailand, close to the Mul-Mekong confluence. The γ-strain is found at all other localities, and is the smallest (1.8 mm [6]) in shell height [8]. Genetic variation in N. aperta (expressed as DNA-sequence based phylogenies) does not track polytypy (see ‘Indications of earlier population phylogenetic studies’ below). N. γ-aperta is found only in shallow, well oxygenated, gently flowing waters, with silt-free smooth rock platforms. The snail exhibits poor dispersal capabilities and does not survive well outside its habitat (even if encased in damp mud) [9,10].
Surveillance to date indicates isolation of N. aperta to Cambodia and Laos (deployment in Thailand is limited to the border region with Laos), and it is difficult to explain the presence of only one other species of Schistosoma Weinland 1858 endemic to the lower-Mekong (namely Schistosoma malayensis Greer et al. 1988, which is transmitted by species of Robertsiella Davis & Greer 1980, also Triculinae, in peninsular Malaysia). Theories accounting for the distribution of S. mekongi have been based on an assumption of isochronicity and phylogenetic tracking by the parasite on the snails, with both taxa arriving in the region off the Indian craton, via Tibet, in the mid-Miocene (~18 Ma) [11]. The radiation of the snails and Schistosoma is described as mirroring the divergence of the main rivers of Tibet, as they cut their way through China and Southeast Asia to the sea [12]. Such a vicariance theory is at odds with more recent estimates of divergence times for these taxa. For example, a Bayesian estimate using a Yule tree model and an uncorrelated log normal relaxed clock suggested a divergence time of a mere 3.8 Ma (Pliocene) for S. mekongi and its sister taxon Schistosoma japonicum Katsurada 1904 which colonised China [2]. In contrast, the divergence of the snail-intermediate hosts of these parasites has been dated at 10 Ma [13]. Similarly, a divergence time of 5–9 Ma [14] estimated for Robertsiella and Neotricula Davis 1986, based on the general invertebrate clock for cox1, are at odds with the 2.5 Ma estimated for S. malayensis and S. mekongi using a Bayesian approach as described above [2]. Similarly difficult to reconcile with the vicariance model, is that despite a supposed 18 Ma of contact between the endemic Southeast Asian Schistosoma and the Triculinae, there are over 90 species of Triculinae in this region, but they transmit only four species of Schistosoma. The problem can be resolved by decoupling the snail and schistosome histories. It is important to note that the most recent common ancestor linking Neotricula and Robertsiella is in Hunan (China), and not in the area of Tibet [15]. The fact that Robertsiella shows derived character states, whereas Neotricula is morphologically a conserved member of the Triculinae, and S. mekongi appears to be derived from an S. malayensis-like lineage in DNA-sequence-based phylogeniesis, is also evidence for phylogenetic incongruence [2]. Further, at least five species of Neotricula are reported from Hunan, but only one from Cambodia and Laos, and none from Tibet, Yunnan (China) or Myanmar [16]. Consequently, an alternative phylogeography was proposed, with proto-S. malayensis and Neotricula entering Vietnam, from Hunan, via the Red river valley [16], which had Pliocene connections with the Yangtze [17]. Both Triculinae and members of the S. malayensis clade are proposed to have entered Southeast Asia using a Vietnam to Cambodia route, but the colonisations were independent and heterochronous. In support of this, molecular dating indicates major divergence events occurring across the known range of this snail between 4 and 6.5 Ma (in response to the final Indosinian orogeny) [16], and a radiation of S. mekongi into Cambodia and northwards to Khong Island (Lao PDR) around 1.3 Ma [2]. It is hypothesised that S. malayensis-mekongi diverged from the Schistosoma sinensium Bao 1958 clade (comprising mainly of species limited to rodents in China and the upper Mekong drainage) during their migration from Hunan on a part of the Sunda shelf now located under the sea off the Vietnam coast [16].
DNA-sequence based phylogenetics detected cryptic taxa within N. γ-aperta from Northeast Thailand; these were γ-I and γ-II, with the former clustering with Cambodian and Lao snails and the latter with snails from the Xe-Bang-Fai river in Khammouanne Province of central Laos. Consideration of the life-cycle of the snails, the larger shell height and greater within clade genetic variation of γ-II, led to the proposal that this clade was in fact comprised mainly of colonists arriving in the Thai Mekong river from tributaries in central Laos early in the dry-season, whereas γ-I comprised of a predominance of locally recruited snails [6]. Following the discovery in 2003 of 11 new populations, occurring in six river systems of Lao PDR and Cambodia [18], an effort was made to sequence DNA from individual snails in these, and previously known N. aperta populations and to thereby estimate population sizes, histories, and migration rates and routes [16]. The genetic clustering method used in this analysis was a Nested Clade Analysis on a network estimated by Statistical Parsimony [19]. The study generated partial sequences for two mitochondrial genes, cox1 and rrnL; however, the cox1 locus showed incompatibilities with the infinite sites model (then a strict requirement for the approach used to date divergence events) at many sites, and could not be united into an unambiguous network by Statistical Parsimony. Consequently, the study was based solely on the rrnL locus. This 2008 study found two monophyletic clades within N. aperta, a spring-dwelling form of northern Lao PDR and a more widespread larger-river dwelling form of southern Laos and Cambodia. Divergence of these clades was dated to 9.3 Ma, with further divergence into sub-clades around 5 Ma. The largest estimated population sizes were found in the Mekong river clades, and these were among the fastest growing populations (followed by those of eastern Cambodia). In keeping with the Red river hypothesis described above, gene-flow was in a predominantly South to North direction.
A more recent DNA-sequence based phylogeny was published (Limpanont et al., 2015) [5] based on one new population (previously unsampled), four new samples of previously studied populations, and data previously published (Attwood & Johnston, 2001) [20] for five Mekong river populations, one Xé Bang Fai river (Khammouanne Province, Laos) N. γ-aperta population, one population of N. β-aperta, and one of N. α-aperta. The newly sequenced population was from the Mekong river at Ban Tha Kathin and, as mentioned above, this represents the first record of N. aperta upstream of Khammouanne Province in Laos. The only clade to retain monophyly in the Limpanont et al. (2015) [5] phylogeny was N. β-aperta. In contrast, Attwood & Johnston (2001) [20] (based on a sub-set of the 2015 data-set) estimated a phylogeny in which N. α-aperta was basal, and the β-strain was also basal to a clade containing all N. γ-aperta sampled. In the Limpanont et al. (2015) [5] phylogeny, all of the newly sampled snails fell into one clade, which also included γ-II (Khemmarat) of the Attwood & Johnston (2001) [20] study, except for Thai γ-I taxa of the newly sampled snails which clustered with the Limpanont et al. (2015) [5] samples of N. γ-aperta from Khemmarat and Khong Jeum (probably Khong Chiam or Khong Jiam resort, a.k.a Ban Dan in earlier publications). The new population from Nong Kai clustered with a (Xé Bang Fai river,γ-II) clade also found in the Attwood & Johnston (2001) [20] study. The Limpanont et al. (2015) [5] study also recovered the ((Khong Island, γ-I), Kratié) clade reported in 2001. Nevertheless, the 2015 study found differently composed γ-I and γ-II clades, such as a break down of the distinction between Thai and Lao+Cambodian taxa in Limpanont et al. (2015) [5].
To address discrepancies between the phylogenetic studies of Attwood & Johnston (2001) [20] and Limpanont et al. (2015) [5], and to estimate better the population genetic parameters obtained by Attwood et al. (2008) [16], the present study used a larger number of characters (1050 bp), a modern Bayesian approach to phylogeny estimation, a more flexible combined partitioning scheme and model testing approach, with greater consideration to starting parameter values in modelling nucleotide substitution and cladogenesis, and vastly greater levels of resampling for error assessment. TMRCA estimation in Attwood et al. (2008) used approaches implemented in Genetree [21]; these assume a either simple constant size or simple exponential growth coalescent model [22] and determine likelihood of summaries of the data under these models by simulation. For example, Genetree assumes the same Ne and θ for both populations and requires multiple runs with one parameter fixed whilst searching for the Maximum Likelihood Estimate (MLE) of another; thus by alternatively fixing and searching for the ML within pairs of parameters joint maxima are found. Consequently, analyses may proceed with MLEs at false maxima. Simultaneous maximisation is possible but is locally inaccurate, and estimates for more than two populations were computationally impractical in 2008 and remain demanding even today. Although Genetree reports posterior distributions for TMRCAs, the approach is not conventionally Bayesian as these distributions are only computed at the MLEs of the population parameters (i.e. empirical-Bayesian [23]). The present study therefore used a true Bayesian approach, also with the capacity to incorporate a greater variety of clock models, to estimate TMRCAs, as implemented in BEAST 1.8.3 [24]. The 2008 study used MIGRATE v.0.97 [25], which assumes a stable sub-population size structure over time, to estimate gene-flow among populations. Although Bayesian, MIGRATE suffers several of the aforementioned problems cited for Genetree. The present study used a more realistic Isolation with Migration model in a Bayesian estimation of posterior probability distributions for migration parameters.
Wang et al. (2014) [26] also reported a phylogeny for N. aperta; however, they focused on sampling in the lower Mekong river around Kratié, used a short sequence from the cox1 gene (342 bp) and a phylogeny estimated by Neighbour-Joining (NJ), which is a distance based method and therefore decouples estimation of genetic distance from estimation of the tree. Consequently, such a phylogenetic method is greatly inferior to ML (or other tree+model based estimators), unlikely to work well with short sequences, and is now reserved mainly for very big data (as it is fast). Wang et al. (2014) [26] reported a phylogeny featuring a basal taxon, which the authors described as ‘another race of N. aperta or even an independent species’. The finding of such disjunction in the pattern of genetic variation in these snails further supports the need for judicious population genetic study of this snail. In addition, to the issues described above, none of the previous studies incorporated geospatial data into the analysis—the present study was the first to incorporate such data.
For reasons of clarity, from here on this account will, when referring to elements of the present study, reserve the term ‘(sub-)population’ for those clusterings found to be genetically distinguishable, validated, (sub-)populations, and use ‘taxon’ for all other samples/collection sites or suspected but unevaluated populations. The term clade will be used for populations or clusters thereof in a phylogeny. The findings of Attwood et al. (2008) [16] suggested that history, rather than ecology, might best explain the absence of S. mekongi from most of Laos; this implied that a spread of Mekong schistosomiasis into central and northern Laos (and even into Thailand) was possible. Unfortunately, the computing power and analytical approaches (e.g. a network based clustering approach) available to the 2008 study were far inferior to those available today. In addition, more realistic modelling of the cox1 data is now possible, enabling their use in such a study. Further, the dating techniques used in the 2008 study were rather simplistic. Consequently, the present study was performed to apply modern analytical approaches to the full data-set, i.e. including both cox1 and rrnL, plus one additional taxon, representing a previously unsampled river. Unlike Attwood et al. (2008) [16], the present study focused on N. γ-aperta, as low gene-flow between the strains may violate assumptions of the analyses. The samples of the Nong Kai population newly reported in 2015 [5], were not available to this study; however, the phylogeny published in 2015 indicates that this population is derived from a major river population from Khammouanne in Laos. As the latter are well sampled in this study, the addition of samples from Nong Kai was not considered critical to addressing project aims.
The aim was to evaluate the findings of Attwood et al. (2008) [16] and to consider the colonisation of Southeast Asia by S. mekongi. Of particular significance to public health is to confirm the South to North migration of the snails, the extent of gene-flow between snails in transmission foci on the Mekong river and tributaries draining into the Mekong from Laos and eastern Cambodia, and the extent of divergence between spring-dwelling and river-dwelling snail taxa. Such questions relate to the potential for expansion of the range of N. aperta (and thereby of S. mekongi) northwards into Laos unrestricted by lack of potential habitat, the role of source populations in tributaries as sources in restoration of Mekong river snail populations involved in transmission following the annual flood (and in reintroduction of S. mekongi from elsewhere), and the potential for spring-dwelling snails to act as intermediate hosts for S. mekongi and support transmission in habitats currently assumed to be free of schistosomiasis. The findings of the study are therefore of considerable value in the design and planning of future schistosomiasis control and for risk-mapping in the Mekong region.
The snails involved in the study were collected in Cambodia, Laos and Thailand, as reported earlier [16], with the addition of a new taxon in Laos sampled de novo. Table 1 gives details of taxa sampled, sampling dates, strain of N. aperta collected, and sample codes used. Sampling procedures were as previously reported [16]. Briefly, partial sequences of the cytochrome oxidase subunit I gene (cox1) and the large ribosomal-RNA gene (rrnL) were obtained, both loci being situated in the mitochondrial genome. In addition to those properties of mitochondrial loci (e.g. maternal pattern of inheritance, and smaller effective population size) that make them suited to population genetic studies, the use of these two loci allowed incorporation of a large pre-existing data-set covering almost all of the known range of N. γ-aperta. Full details of sequencing procedures, primers and justification for choice of loci have been published elsewhere [27]. Sites with indels (totalling 3 sites) were excluded from the analyses.
DNA-sequence data used in the analyses have been deposited in the GenBank as follows:
For those analyses requiring an outgroup, the snail Robertsiella silvicola was chosen, as phylogenies for the Pomatiopsidae indicate that it lies at the root of the clade containing all Mekong river Neotricula [27]. The corresponding sequences, taken from the GenBank, were AF531550 and AF531548 (cox1 and rrnL, respectively).
Using SeqTrace 0.9.0 [28], DNA Sequencing Chromatograms were converted into quality controlled sequences, and a consensus produced for each sample, from paired forward and reverse reads, so as to maximise final sequence quality. The resulting sequences were aligned using CLUSTAL 2.1 [29], alignments visualised using Aliview 1.17.1 [30] and trimmed using GNU Bash 4.3.48(1) (commands and extensions thereof) [31]. A concatenated cox1+rrnL ‘both loci’ alignment was created using pyfasta 0.5.2 [32] to select all cox1 sequences for which there was a corresponding rrnL sequence. The reading frame of the protein coding locus was determined using ExPASy Translate [33]. The apparently optimum partitioning strategy and corresponding evolutionary models were determined using PartitionFinder 1.0.1 [34], under a BIC criterion.
The generation of a resistance surface is a requisite for spatial analysis of these data. Consequently, the CRAN R 3.2.3 [35] OpenStreetMap 0.3.2 package [36] was used to obtain a high resolution map of the sampling area. The water courses (excluding those know to be unsuitable habitats for N. aperta) were then expanded (to allow for anthropochory and zoochory of snails between closely adjacent drainages). A break was introduced into the Nam Theun river to simulate the barrier now posed by the Nam Theun 2 dam, which was not then incorporated into the OpenStreetMap data-set. The image was next desaturated and brightness minimised. The negated image was then imported to R as a raster array, the grey-scale information extracted and rasterised, coordinates were then reassigned to the image, which was finally pre-projected (to the original projection matching the data) to create a cost-surface that was used to express resistance to dispersal in subsequent analyses. To account for the fact that downstream migration is more likely than upstream, a NW to SE (decreasing) cost gradient was applied to the array (mirroring the predominant flow direction and current of the rivers in the region). In effect, this resistance surface was telling the analyses that dispersal between rivers was highly unlikely unless the rivers were very close together (within 3 km), and dispersal upstream (especially in fast flowing highland streams) was less likely than downstream.
Before attempting a detailed analysis of migration patterns and phylogeography, the data were interrogated for population genetic structure. First, inter-taxon FST values were estimated using the R package hierfstat 0.04–22 [37]. The data were next tested for signs of Isolation By Distance (IBD), as the presence of significant IBD may confound interpretation of spatial-genetic variation observed, and violate the assumptions of common clustering algorithms. To achieve this, a matrix of inter-taxon distances was generated using the distcalc function [38], with slight modification (to allow data input as R variables), and the resistance surface; this provides more meaningful distances in a study of freshwater snails than simple euclidean distances. The DNA-sequence data were read into R and manipulated using R packages adegenet 2.0.2 [39] and ape 3.4 [40]. A Mantel test, implemented in R package ade4 1.7–3 [41], was then used to assess the correlation between Edwards’ distances [42] and ‘river’ distances for individual snails. The test applied Monte-Carlo sampling and 5000000 replicates. Distant and differentiated populations may show a similar pattern of genetic structure to that caused by the continuous clines of classical IBD. A simple approach to distinguish the two processes is to plot both distances [43]. To avoid granularity due to binning, 2-dimensional kernel density estimation (Parzen method) was used together with the MASS package [44] and adegenet (in R) to plot local densities, so as to distinguish true IBD from differentiation of distant populations [39].
A simple assessment of geographical distribution followed by multivariate methods was used to establish the presence of distinct populations in N. γ-aperta; the use of multivariate analysis provides a rapid and flexible (i.e. independent of population genetic models) approach to determining if significant deviations from panmixis occur across the range of this snail, before attempting a more involved phylogeographical analysis. In particular, the use of linked loci (mitochondrial) violates the assumptions of model based population genetic clustering approaches such as those used in STRUCTURE [45] and BAPs [46]. Similarly, such haploid loci preclude tests based on heterozygosity deviations.
A table of (binomially transformed) standardised allele frequencies was subjected to Principal Component Analysis (PCA), as implemented in adegenet. Here PCA eigenvalues represented the amount of genetic diversity accounted for by each Principal Component (PC), and a sudden decrease in eigenvalues may correspond to the point where true structure in the data gives way to noise [39]. In the present study, the first PC, and possibly also the second, were found most likely to exhibit a relevant biological signal (as judged by inspection of the corresponding scree-plot). Again kernel density estimation is used to best depict the distribution of the genotypes on the PCs. Following common practice [39], the informative PCs (in this case the first and second) were next plotted onto the geographical space, as a first step in uniting spatial and genetic variation. The plotting was implemented using ade4. Next, Moran’s I [47] test of spatial autocorrelation along the PCs was used to evaluate any spatial clustering (i.e. populations) suggested by the plots. To achieve this, spatial connectivity was defined as the resistance surface described in the previous sub-section. Moran’s test was implemented using the spdep 0.6–13 package [48] in R. As the test requires a spatial weights matrix, costs were obtained using geographical distances extracted from the river network in the resistance surface; this was achieved using R packages gdal 2.2.1 [49], maptools 0.9–2 [50], and rgeos 0.3–23 [51]. In order to assess the spatial association of each snail with its neighbours and the precision of the sampling regime, a Moran scatter plot [52] was performed; this involves plotting the standardised haplotypic data against their lagged values (i.e. the weighted average of those of their neighbouring samples) along each PC. The values in the plot correspond to standard deviations and therefore provide a useful insight into how individuals differ from one another at increasing scales. The plot was again implemented in spdep. In addition, a geographical (NW-SE) blue to red colour gradient was applied to the snail samples on the Moran plot using R’s grDevices.
The above PCA-based tests, restrict the analyses to few out of the many PCs, and therefore fail to utilise the data fully. Spatial structures in the whole data can be detected using a Mantel permutation test [53] for correlation between the genetic and geographic distance matrices. Here the test was implemented using the R package ade4. The test uses a matrix containing Euclidean distances among individual snails calculated from the scaled genetic data (as used in the PCA above), and a matrix of the corresponding spatial distances calculated from the resistance surface (i.e. effectively ‘by river’ distances). Following a suggestive result from the Mantel test, a spatial PCA (sPCA) [54] was performed to determine what proportion of the genetic variability may be spatially structured, given the data. In addition to optimising the variance of the PCs, sPCA also optimises their spatial autocorrelation through Moran’s I (i.e. the eigenvalues are composite) [55]. As with Moran’s I test, the input for the sPCA was the standardised haplotype data and the matrix of spatial proximities obtained from the resistance surface. Implementation was through adegenet.
Two randomisation tests are available in adegenet to determine which scale of structures are interpretable for the data [54]. In these tests the standardised haplotype frequency matrix used in sPCA, and list of weights derived from the resistance surface, is regressed onto Moran’s Eigenvector Maps (MEMs), and a mean R2 is obtained for each MEM (the highest R2 is taken as the test statistic and compared to a reference distribution obtained by Monte-Carlo resampling of randomly permuted frequency matrices). The global test involves decomposition of the matrix into global MEMs, and the local test into local MEMS [39]. Here 9999 permutations were used. Following these tests for interpretable structure, the spatial genetic patterns were visualised by plotting lagged global scores onto geographical space (the use of lagged scores reduces noise) [54]. The colorplot function in adegenet was used to obtain similar plots, but with a better contrasting colour range, such that similarity of colour indicates genetic affinity. Finally, Ward’s clustering method (implemented by hclust in base R) was used to assign taxa to populations inferred by minimising the energy distance between cluster groups on the first PC of the sPCA using the Ward1 algorithm [56]. The clustering was based on the differentiation of sharp changes along the PC, which define population boundaries.
Having obtained an estimate of population structure within N. γ-aperta by multivariate approaches, the amount and direction of gene-flow, and thereby inferred rates of migration, among these sub-populations can be estimated. To enable this, the R package phangorn 1.99.14 [57] was used to estimate a phylogeny (using NJ with a stochastic rearrangement algorithm) and then to fit HKY+G+I model parameters; these parameters were then used in the subsequent analyses where such starting values were required. As the sPCA did not suggest more than five populations, the Ward’s clustering was repeated with a permitted maximum of five sub-populations. N. aperta has been described as a metapopulation, and the analysis should therefore avoid methods such as those implemented by Migrate-N that assume stable sub-population structure over time. Nevertheless, Migrate-N 3.2.8 [58] was used in a preliminary assessment of support for combining the four populations found by Ward’s clustering into the two or three indicated by the sPCA. The Migrate-N analyses assumed a prior distribution of haplotype frequencies over populations (i.e. was Bayesian).
Several series of test runs were performed in order to select starting parameter values, settings (e.g. burn-in length, adaptive versus static chain heating schemes, random or UPGMA starting trees, etc) and prior distributions (e.g. on starting theta and migration rates (mis), being uniform or random uniform or exponential) for the Migrate-N analyses. Initially, starting values for thetas and mis were based on Fst, with later runs started using estimates of theta and mi from the output of the currently best performing test-run. The relative performance of each run was assessed through Bayes-Factor (BF) tests comparing MArginal Likelihoods (MALs) determined by thermodynamic integration coupled with the use of Bezier-curves to improve approximation where a low number of heated chains is used (i.e. four here) [58]. Natural log BFs, and the guidelines of Kass and Raftery [59], in the context of theoretical expectations, were used to compare alternative runs. In addition to BFs, Effective Sample Sizes (ESSs) for estimated parameters, consistency of estimates (between replicate runs), and shape of posterior distributions, were examined for signs of good mixing and chain convergence. Runs with increasing chain length were also compared in order to optimise sampling. Consequently, analyses were performed with 20 short-chains, two replicate long-chains of 50000000 generations (taking 5000000 samples from each), a burn-in of 1000000, FST starting values and uniform priors on theta and mi, UPGMA starting-tree, a static heating scheme (with default values), and posterior generation by SLICE sampling. Runs with the above settings, and BFs, were then used to test alternative hypotheses regarding the number of distinct sub-populations (panmictic units) present in N. aperta.
The Migrate-N based analyses were complemented by interrogation of the data using IMa2p [60]. IMa2p might be better able to model a metapopulation as it can accommodate changes in sub-population size over time; however, this involves estimation of additional parameters and greatly increases computation time. Consequently, the analyses began with a three-population model suggested by the Migrate-N tests as best explaining the data. A per locus mutation rate of 1.3566e-04 per year was obtained from the overall meanRate of the three main BEAST runs, with the mean and 95% Highest Posterior Probability Density (HPD) passed to IMa2p as a prior. The priors on spltting time and genealogy described an island model, that is an ancestral population splitting into two in the deep past, followed by fairly constant levels of gene-flow through to the present. Test runs were initiated with a prior on this divergence of 0.8 Ma (the last major interval of river flow reversals and tectonic upheaval, the cessation of which could have segregated N. aperta populations between northern and southern Laos [27]).
Test runs, with different starting parameter values, were used to determine appropriate values to initialise the main runs. At the same time trendline plots were check for persistence of obvious trends and multiple runs, that differed only in random-number seed, were performed with the expectation that parameter estimates (and posterior probabilities) should be similar; these measures provided an indication that the Markov chains had converged in distribution. ESS values among the parameters >>50 were considered a sign of adequate mixing among parallel chains. Plots of posterior density were used to assess influence of priors (e.g. distribution greatly truncated by prior maximum, suggests prior is too low), together with runs sampling from the prior distribution alone. The full range of IMa2p’s appropriate run options (see Table 2) were evaluated, where similar posteriors were found with different settings, the simpler combination was used (e.g. fewest parameters estimated).
BEAST was used to estimate a phylogeny for the individual snails sampled, and to estimate divergence times for major clades through a Bayesian approach. Bayesian phylogenetics does not assume approximate normality or large sample sizes as would general ML methods [62], and is therefore statistically superior to approaches based on unintegrated likelihoods [63]. In addition, Bayesian methods consider the posterior probability of the model (with parameters) and tree after observing the data; this is proportional to the product of the prior probability of an hypothesis and the probability of observing the dataset given the hypothesis (i.e., its likelihood), and, unlike direct ML, allows incorporation of prior information about the phylogenetic process and dates of divergence. Such incorporation of prior distributions more realistically accommodates the uncertainty associated with calibration points and estimated rates used in the analyses [64]. Although Bayesian methods have been known to erroneously converge on long-tree solutions, the present data are not partitioned and the number of parameters to be estimated is relatively low, so that the occurrence of such erroneous convergence is unlikely [65,66]. Finally, Bayesian methods are also preferred over direct ML because of their speed in terms of computing time (for analyses with an equivalent level of confidence). BEAST, which uses a Markov chain Monte Carlo (MCMC) approach to approximate the posterior probability distribution of parameters in a phylogenetic model, was chosen because of its incorporation of divergence dating and phylogenetic modelling. Molecular clock methods in BEAST generally outperform other dating approaches (e.g., non-parametric methods such as NPRS [67] or penalized likelihood methods [68]) particularly for divergences with a low time depth, as they not only allow for uncertainty in dates assigned to calibration points (through priors), but also avoid reliance on untested assumptions about the pattern of clock rate variation among lineages [69].
Initially a dated phylogeny was estimated with calibration priors on key cladogenic events. BEAST allows specification of a large number and variety of priors, starting-values and models. BFs enabled comparisons for a series of short (48 million generations) test runs used to determine optimum settings. For each test run a MAL was recorded; this MAL was the log marginal likelihood (using stepping stone sampling) obtained from pathLikelihood.delta. In brief, the tests examined the effect of changing or removing priors on substitution rates (ucld.stdev and meanRate), population size history (tree priors), offset on tree height, and date calibrations (TMRCAs and root). The purpose of such tests was to ensure that no prior was overly determinant (i.e. shaping) of the posterior distribution; this was further established by running the final analyses without the data (i.e. using the priors only), which can also reveal unpredicted problematic interactions between priors. The posterior distributions were examined using Tracer 1.5 [70]. and the run settings were only accepted if no distribution was seen to be markedly cut off by its prior or show signs of failure to converge (rising likelihood). In addition, all combinations of clock model and rate prior-distribution, implemented in BEAST, were compared (i.e. strict, fixed/random local, and Uncorrelated-Relaxed Models (URMs), with CMTMC (strict/local only), exponential*, gamma*, invgamma, Jeffry’s, lognormal, normal, and uniform prior distributions—*the only distributions tested with URM). Finally, the effect of doubling the chain length was determined for the most promising clock model and prior combinations. Indicators of likely Markov Chain convergence were trace values that reached, and then varied around, a constant log-likelihood from early in the run (just after the burnin) and thereafter, and ESSs greater than 200. Testing for the optimum length of long-chains was based around the experiences of earlier work [71] involving a similar analysis (e.g. same loci, number of sequences, etc), with adjustment for differing length of dimensions of the data. S1 Table lists the final priors, models and other run settings chosen on the basis of BFs; the MAL of this run was -8010.8614 (BF 30.43 cf. next best run). A plot of MAL against generation number (from 6 to 1450 million generations), for runs with all other settings as in S1 Table, was used to ensure that the likelihood had reached a stable plateau suggestive of stationarity. A chain length of 600 million was thereby found to be optimal (S1 Fig) (BF cf. worst and next best run respectively, 19.4583 and 9.6710). Subsequently, three such runs were performed with different random number seeds in order to confirm further convergence. No outgroup taxon was required, as the use of a relaxed clock model provided an estimate of the position of the root of the tree.
The BF based model testing suggested that substitution rates were best represented by an Uncorrelated relaxed Clock with branch-specific rates drawn from an Exponential Distribution (UCED). An exponential distribution implies that most of the branches have rates at the lower end of the range, with a few branches showing high rates. The choice of an uncorrelated model implies an episodic mode of evolution [72], which is not inconsistent with an historical biogeography described for N. aperta as dominated by cladogenesis following diastrophic events, changes in river courses and orogenies [73]. In addition, an uncorrelated clock is perhaps more realistic over the relatively short time-depth of this phylogeny and the environmental stochasticity just referenced, which are likely to overwhelm variance contributed by inherited factors. Modelling of the distribution of substitution rates across branches, for model averaging of the UCED, used an array of positive continuous parametric distributions (such as Gamma and Inverse Gaussian (IG)). Such a continuous parameterisation is considered to better accommodate rate heterogeneity; for example, the long upper tail of the IG distribution permits some taxa to have relatively high rates [74], a feature also consistent with the exponential distribution of branch-rates. It was envisaged that transition from small streams to major rivers would effect a jump in rates.
Priors on the times of divergence events were used to guide calibration of the molecular clock, as an alternative to specifying a mean clock rate. The divergence times were MLEs of Time to Most Recent Common Ancestor (TMRCAs) taken from an earlier population phylogenetic study of N. aperta [16]. The priors were applied as normally distributed calibration dates with Standard Deviation (SD) set to achieve twice the range of the Confidence Interval (CI) of the MLE. The lower tail of these prior distributions was truncated at 0.2 Ma (megaannum or million years). Three such date priors were applied, with reference to three clades previously recognised [16] as major in the evolutionary history of N. aperta. The correspondence of these clades to sub-populations used in the migration studies of the present investigation (see Results: ‘Migration rates among sub-populations’ below) are as follows: Cambodian Eastern Rivers (CER), BOL; Northern Spring Populations (NSP), CKS+CKR; and Cambodian Lower Mekong (CLM) i.e. the southernmost Mekong river populations of Cambodia (which is MEK in part). The geographical deployment of the clades referred to in the present study are shown in Fig 1. The priors are given in S1 Table. It must be noted that these priors resulted from analysis of part of the dataset of the present study. Consequently, they are not entirely independent, but their use is equivalent to the employment of a training data sub-set to obtain parameter estimates.
The phylogeny is presented as a 50% majority rule consensus tree representing the Maximum Clade Credibility (MCC) trees of three replicate estimates, and produced using sumtrees Version 4.1.0 [75]. The MCC tree represents the topology yielding the highest product of posterior probabilities sampled for its individual constituent clades. The first 1000 trees (6 million generations) of each run were discarded as a burnin, and node support values were averaged across replicates. Phylogenetic trees were visualised using figtree 1.4.2 [76]. An array of colour ramp values was generated for geographical coordinates, running from Northeast to Southwest, using R, and the terminal branches of nexus trees, plotted using ape, were then coloured, using the plotrix package [77] in R, according to geographical location of the corresponding tip. The marginal densities of estimated parameters (e.g. TMRCAs) were examined using Tracer, and means and 95% CIs (as 95% Highest Density Interval (HDI) for the posterior distribution) are reported for the combined samples of all three replicate runs.
The data-set, after editing for quality control and removal of taxa lacking quality sequence data for both loci, comprised 1050 characters (nucleotide sites or base-pairs) and 260 individuals (excluding the outgroup), representing 27 taxa of the Mekong river and nine of its tributaries in Cambodia, Laos and Thailand (Fig 2). Within the data-set, the first 545 bp corresponded to cox1 and sites 546–1050 to the rrnL locus. The first character represented a third codon position. A single partition, with a HKY+I+G, model was found, by PartitionFinder, to be the simplest scenario that appeared to be least inadequate in representing the evolution of the data (BIC 17053.344).
FST values ranged from 0.024 to 0.946 (mean 0.579). The extreme FST values corresponded to DIL versus OHG and KRK versus TOT; these corresponded to almost, but not exactly, the most extreme river distances. Similarly, the Mantel test suggested high levels of IBD (P < 2x10-7) S2 Fig. A plot of the distance measures, however, showed clumping, which suggests IBD caused by well differentiated pops geographically far apart, rather than classical IBD (S3 Fig).
Fig 2 indicates a loose geographical clustering of the snail taxa; however, such a simple plot is a poor expression of density as it cannot account for near overlapping samples. Consequently, 2-dimensional kernel density estimation was again used to visualise better the spatial clustering. The resulting plot (Fig 3) suggested that the snail taxa were grouped into one northern and one or two southern clusters; however, none of these clusters was entirely discrete (especially the two southern groupings) and an individual-based analysis appeared more appropriate for the spatial-genetic stage of the investigation, as individuals could not be unambiguously pooled into groups.
The PCA for genetic structure (genetic diversity among genotypes) indicated a number of discernible potential populations (S4A Fig). A corresponding loading plot revealed that this structure was the result of possession of many original alleles between these (S4B Fig). Moran’s I test was highly significant at both the first and second PC (S5 Fig); thus also supporting the presence of spatial-genetic clustering in the data. In both cases Moran’s I was extremely positive, suggesting that snails were likely to be surrounded by others with genotypes closely similar to their own. The Moran scatter plot for the first PC (Fig 4A) shows a concentration of Khammouane spring-dwelling snails throughout the upper-right quadrant, i.e. these snails tend to be surrounded by their close relatives and haplotype distribution is contagious (strictly, they show positive spatial autocorrelation greater than the sample mean). In contrast, snails from the center of the range show little spatial-genetic autocorrelation (i.e., they are rather panmictic), whereas those from the southeastern limit (red) are less likely to be surrounded by individuals bearing similar haplotypes, than the mean for the sample, (as they fall into the lower left quadrant). The reverse situation is seen in the Moran scatter plot for the second PC (Fig 4B); however, in this case points are clustered more around zero; this may reflect that the second PC is expected to be much less informative than the first (see scree-plot, inset to S4A Fig). Neither PC indicated negative autocorrelation (i.e. few points fell in the upper-left or lower-right quadrants), suggesting that across the range the snail haplotype distributions tend to be positively correlated, but to varying degrees. The points in both plots mostly fell close to the regression line, which suggested that there were no significant problems with the weights matrix, and that the observation (sampling) scale was sufficient for the scale of the spatial structure present.
The Mantel test for population structure was highly significant (S6 Fig). Consequently, spatial-genetic structure was investigated further by sPCA. Examination of the scree-plot for the sPCA indicated that the first 26 eigenvalues were positive, with only minor PCs negative (S7A Fig); this suggested overall positive spatial-genetic autocorrelation in these snails. Moran’s I ranged from -0.2 to 0.95 and the composite eigenvalue for the first PC clearly exceeded all others (S7B Fig, decomposed sPCA λ1). Consequently, the data contained interpretable variability and spatial structure, and only the first PC should be retained in subsequent analyses. Now that interpretable spatial-genetic structure had been detected, data regression onto MEMs was used to determine if global or local spatial structures should be interpreted. The results indicated that, whilst the test for interpretable local structure was not significant (observed value 0.009, P = 0.9994), that for global structure was highly significant (observed value 0.158, P = 0.0001).
Plotting lagged scores onto the geographical space (Fig 5) revealed a notable North-South divide among the snail taxa; this divide is located between the Khammouane (Lao) and Ubon (Northeast Thailand) taxa in the North, and all other taxa to the South. The second PC also suggested at least a third cluster of taxa may be present in the rivers of Cambodia (Fig 5B). Fig 6A depicts genetic affinity through taxon plot colour, and indicates a general North to South cline along the Mekong river, with a moderately distinct population in the Xe Kong river and the rivers of Northeast Cambodia and a highly distinct spring-dwelling population in the Northwest. Fig 6B shows the dendogram resulting from minimum variance clustering using the first PC of the sPCA; this shows a major division in N. γ-aperta, with a clade containing all spring-dwelling taxa plus those of the upper reaches of the rivers sampled in Khammouane, and a second, larger, clade comprising all other taxa. The larger clade is divided into a sub-clade including all taxa of the rivers of northeastern Cambodia and a sub-clade containing all Mekong river taxa (including taxa of the lower reaches of the rivers in Khammouane). Finally, an interpolation map, using lagged principal scores from the sPCA, was produced (using the R package akima 0.5–12 [78]) and is provided in Fig 7 for further visualisation of genetic clines present in N. γ-aperta—the North to South cline mentioned above is seen clearly therein (i.e. the blue to red transition).
Ward’s clustering with a maximum limit of five taxa indicated four ecogeographical populations (see Table 3 for the sub-population membership): TOT or Cammon karst small spring (CKS); Cammon karst river and large primary stream (CKR); and two flood-plain major river populations Mekong (MEK) and Bolaven (BOL). CKR comprises mainly riverine snails occurring to the South of the karst ridge in Khammouane Province (Laos)—the Cammon highland region. The Bolaven Plateau lies in southern Laos between the Mekong river on the West and the Annamite Mountain Range in the East. The upper Xe Kong river is found on this plateau. The MEK sub-population here refers to taxa of the Lao and Cambodian Mekong river itself. The four ecogeographical regions correlated well with the inferred sub-populations; however, exchange of 4 snails was required to fit the taxa precisely to the ecogeographical categories (one MOP and one KSC were moved from CKR to MEK; one DKX and one XKM were moved from MEK to BOL). These exchanges were made because the primary interest was to assess migration rates of snails among geographical regions. The relationship between the four sub-populations and those inferred by sPCA is depicted in Fig 5.
MigrateN starting parameters were as detailed in Methods: ‘Inference of migration rates’ above. Despite a Gelman’s R < 1.2, high acceptance ratios and ESS values ranging from around 70k to 1 million, replicate runs did not give consistent migration rate estimates. Consequently, the CKR and MEK populations were combined into a single population covering the karst rivers of Khammouanne immediately to the North (e.g. Xe Bang Fai drainage) and the Mekong river of Laos and Cambodia, into the northern part of which the rivers of CKR drain. The results of three replicate runs of 50 million generations each are given in Table 4 which shows considerable variance among replicates despite very similar likelihood values. The high variance of parameter estimates suggests that the data are insufficient to estimate parameters at this resolution (i.e. level of population structure); however, the particularly large variance in θBOL estimate, where sample size was not especially low (N = 100), could be partly due to further (cryptic) population structure and cross-population sampling. Greatly unequal sample sizes, e.g. TOT→WBR, would lead to unequal underestimation of θ, with θTOT being most underestimated in the example given. The observation of a very high migration rate TOT→WBR, with a relatively very low rate WBR→TOT might be due to the greater probability of dispersal, during floods, from TOT into the Nam Yom and onward downstream to MEK; however, the spring population is a phylogentically distinct ecotype, biogeographically isolated, and inflation of the θWBR/θTOT ratio is an alternative explanation. Unusually high migration is also seen TOT→BOL despite these populations being isolated by highland and considerable distance (indeed, replicate 3 may have stumbled upon the true value for TOT→BOL). Overall it is likely that migration levels among all populations are between zero and ten snails per generation.
Although perhaps unable to reliably estimate all parameters of these gene-flow models, Migrate-N was used in model testing with respect to number of distinct sub-populations present. Table 5 shows the results of the testing. The model with greatest posterior probability was that with a combined CKR+MEK population and, as the log Bayes Factor with the next ranked model was -11.0098 (which is < -2 [59]), the suggestion is that this simpler model is a better fit to the data than the original four population model.
As mentioned above, Migrate-N assumes constant population structure over time. Consequently, migration rates were also estimated using IMa2p and the metapopulation model chosen using Migrate-N. The CKS (TOT) taxon had to be excluded because runs with all three taxa (BOL, WBR and CKS) failed to converge within 48x(12x3.4GHz) hxCPUs during test runs or crashed after the burnin. Again this was possibly a consequence of the smaller sample and effective-population size of TOT relative to the lotic populations. TOT was not simply merged with WBR because TOT had appeared as a highly discrete population throughout the population structural analyses. Its merger with any other data partition was therefore unjustified and its exclusion saved computational time. No significant autocorrelation was found among any of the parameters estimated (suggesting that the data were sufficient to distinguish each of them), and ESS values for all exceeded 150. Table 6 shows that both population size parameters were consitently estimated (especially that for WBR) and imply effective population sizes of around 310000 (WBR) and 120000 (BOL). The greater size of WBR is expected as it includes 300 km of the lower Mekong river itself, whereas BOL is dominated by smaller highland streams. Although indicating the same relative proportions, the estimates of Neμ from Migrate-N (Table 4) are much lower; this may be explained by the use of a mutation rate prior (from BEAST) in the IMa2p analyses. As BEAST was focused on estimating mutation rate associated parameters, whereas Migrate-N estimated both mutation and migration rates, and population size, the mutation rate from BEAST was considered to be a better indicator of the true rate.
In contrast to population size, the migration rate estimates were less consistent (Table 6) and the HPDs appeared dependent on the prior (i.e. the HPD did not tail off toward zero at either the upper or lower limit of the prior). Nevertheless, the greater population migration rate is seen in the BOL→ WBR direction, which may be expected consdiring that rivers such as the Sre Pok and Xe Kong (of BOL) drain into the lower Mekong river (MEK of WBR). In contrast, much less variation was seen in estimates of splitting time. The analyses indicated that the BOL and WBR populations split 117793 (71797,183988) years prior to sampling. The mean TMRCA across all three main runs was quite consistent, being 43.22±1.16 (±SD); this approximates to 320000 generations ago.
Uninformative data can appear to describe an island model if a high upper bound is set as a prior on the splitting time of the ancestral population, as such analyses tend to infer a splitting time at that upper bound. To exclude this possibility, runs were performed with -t set to 25% of the value indicated by the test runs. Despite the reduction in the prior, the same high estimate of splitting time was returned; this suggested that the data were consistent with an island model.
Fig 8 shows the 50% majority rule consensus tree from phylogenetic estimation using BEAST. The tree depicts a phylogeny in which the four ecogeographical populations described in the previous sub-section, as a result of sPCA, did not appear monophyletic, except for CKS which formed an inclusive sub-clade within one of the two CKR clades; however, BOL and MEK were paraphyletic. BOL appears tripartite, divided into a large clade made up of snails from the Mekong-tributary rivers of northeastern Cambodia, a smaller clade of snails from the Attapeu region of southeastern Laos, and four monotypic clades (RAM1, RAM2, JND and DIL), all arising from a common ancestor with MEK-2 in an unresolved polytomy. MEK-2 comprised of the Mekong river taxa of the southerly limit of the range in Cambodia. MEK-2 showed three sub-clades whose relationships are unresolved (i.e. forming a trichotomy from a common ancestor with the six BOL clades). Interestingly, MEK-2 did not include any snails from the rivers of northeastern Cambodia that drain into the Mekong in this region (these clustered with some MEK-I (SDO) taxa and the larger BOL sub-clade).
The MEK-1, (CKR-1,(CKS,MXL)), CKR-2 clades form a basal polytomy. Within the CKR-1 clade, a trichotomy describes a monotypic clade (MOP), trifurcate clade (YOM), and a polytomic clade (six branches), within which CKS (TOT) forms a monophyletic clade and KLR and TKN appear respectively monophyletic on a bifurcate clade. The remaining four branches comprise of MXL and YOM. The two CKR clades differ in that CKR-1 includes CKS and TKN, and CKR-2 includes BKV taxa. The main clade of CKR-2 comprises a (BKV,(YOM)) clade, an (YOM,(MXL)) clade, three monotypic MXL clades, with two monotypic BKL clades basal to this main clade.
The standard deviation on the mean branch rate estimate for the uncorrelated relaxed exponential clock used was low (0.0019); this suggested that meaningful divergence times may be obtained by sampling the posterior. The marginal probability densities of the divergence time estimates showed HDIs (95% highest posterior probability densities) of around 4 Ma; their means are given in S2 Table. The mean divergence time priors (S1 Table) used fell below the lower limit of the HDIs for all estimates (suggesting that these distributions were not simply reflecting the priors). The HDIs of all three divergence time estimates showed partial overlap (~50%) and so could not be reliably ordered in time. Nevertheless, the CER divergence may have followed the CLM and NSP divergences; the latter being apparently isochronous at around 15 Ma. The divergence of the CER clade was estimated as having occurred around 13 Ma. The divergence time estimates are also given on Fig 8 at the root of the relevant clades. A per-lineage mean clock rate of 0.85% (substitutions/site/Ma) was estimated across the time depth of the phylogeny (see S2 Table).
The study found that population genetic variation corresponded with snail ecology or habitat. The fact that Ward’s clustering almost exactly recovered the eco-geographical structure of the samples, strongly suggests that N. aperta is not panmictic and that there is limited gene-flow among these regions. PCA indicated a marked clinal disjunction between the North and South of the range, with perhaps two, less well defined, sub-populations in the South (Fig 3). The disjunction was located between Khammouanne in Laos and sub-populations south of the northern boundary of the Bolavens plateau (including populations of Ubon, Northeast Thailand). Parzen plots of genetic distance versus ‘river distance’ suggested that the North-South divide was a consequence of well differentiated populations geographically far apart, rather than true IBD along extended clines. The contagious distribution indicated by the Moran scatter-plot for the Khammouanne spring-dwelling snails is consistent with their being endemic populations restricted to their host spring and possessing poor dispersal capabilities. The finding does not suggest that snails found in these springs were recently swept in from major-river populations during flood events. In contrast, snail populations occupying the central part of the range (e.g. the Mekong of southern Laos) appear far more panmictic. Such panmixis is most likely the result of mixing among sub-populations during the annual flood, and the ready interconnection of the rivers involved. Nevertheless, these populations still showed significantly positive spatial auto-correlation, which implies that each snail is more likely to be surrounded by individuals more closely related to it than would be expected if the distribution were random (i.e. spatial-genetic affinity in the snails showed granularity). Snails of the most southeasterly populations (i.e. the rivers of northeastern Cambodia) appeared to be surrounded by individuals less closely related to them than expected in the absence of population structure. This may be due to a lack of migration among the rivers of this region despite their geographical proximity, which leads to the existence of spatially close but genetically distant individuals.
sPCA and data regression onto MEMs indicated highly significant global population structure, but much less local structure. The North-South divide suggested by PCA was reinforced by plotting of lagged principal scores onto geographical space. The existence of a third sub-population comprising the sampling sites of Northeast Cambodia and the Xe Kong drainage of Laos (effectively the Bolovens sub-population, BOL) was supported by the second PC of the sPCA, and to a lesser extent by the minimum variance clustering dendogram for the first PC (of the sPCA), where it appears as a sub-clade of the major southern clade (Fig 6B).
Ward’s clustering identified four ecogeographical populations whose membership corresponded almost directly with the snails sampled within the geographical region of each of the four inferred clusters. Such a remarkable correspondence raises the possibility that ecogeographical differences sub-divide N. aperta and that some degree of ecological, as well as physical, isolation might be present. The sub-populations inferred by sPCA were also consistent with these ecogeographical clusters (Figs 1 and 6).
In contrast to Ward’s clustering, phylogenetic estimation using BEAST, found only CKS to be monophyletic (and only as a sub-clade of one of two CKR clades). It should be noted, however, that whilst PCA and related methods are based solely on genetic differences, BEAST attempts to recover history and could give a better estimate of the ancestral relationships of the sub-populations. The BEAST phylogeny (Fig 8) suggests two Mekong river clades, one of southern Laos (MEK-1) and the other of Cambodia (MEK-2). Although the Cambodian Mekong river clade did not contain any of the taxa from populations of the Northeast Cambodian rivers that drain into the Cambodian Mekong, it did appear to share a common ancestor with taxa from these rivers and from the Attapeu region of Laos in the North (i.e. with the two Bolovens clades). Nested clade analysis [79] previously published for an earlier (more limited) data-set [16] also suggested a disjunction between Lao and Cambodian Mekong river populations, with the latter having closer affinity with a Bolovens clade. Although the two stretches of the Mekong river are separated by the cataract at Khone falls, this is unlikely to represent a significant barrier to downstream dispersal. Consequently, either historical or ecological factors may explain the relatively deep evolutionary divergence between MEK-2 (and BOL) from all other taxa. Average levels of dissolved oxygen (2002–2014) were reportedly higher in the MEK-2 stretch of the Mekong river [80], and it could be that MEK-1 populations have acquired adaptations to lower oxygen levels (or associated habitat differences) such that they are disadvantaged when migrating southwards into MEK-2 habitats. In terms of phylogenetics, the Red river hypothesis [81] proposes a Pliocene colonisation of Laos, by proto-Neotricula from Hunan, via Cambodia, with snails radiating northwards into Laos from northeastern Cambodia. The two hypotheses are not mutually exclusive and are difficult to distinguish using the present data; however, the use of rapidly evolving markers such as microsatellites may shed light on the processes involved (by emphasising the effects of very recent or ongoing (i.e. ecological) events).
Confidence intervals for divergence time estimates (from BEAST) were wide, such that it is not possible to order the times of the MRCAs of all three major clades in Fig 8. Nevertheless, the common ancestor of the Bolovens clade may be more recent, at around 13 Ma, than those of MEK-2 (CLM) and CKS+CKR (NSP), which appear isochronous at just over 15 Ma. These values are much greater than the estimated TMRCA for all of the taxa (excluding TOT) from IMa2p, which was around 0.3 Ma. Similarly, the splitting time between BOL and WBR was estimated at just over 0.1 Ma. It should be noted that whilst IMa2p estimated splitting times between contemporary populations, BEAST estimated dates for divergence events among ancestral populations. Further, the TMRCA can be much older than the most basal node of the phylogeny, which in the IMa2p analysis is that representing the ancestor of BOL and WBR if the depth of the phylogeny is less than 4-7Ne [82] (here this is around 10x greater at 1.7–3 Ma).
As in the 2008 study (based on the rrnL data only and fewer sampling sites), the order of divergence events appears to be near simultaneous divergence of the spring-dwelling Cammon karst and Cammon karst stream clades (NSP), and of MEK-2, followed by that of BOL. In contrast, Attwood et al. (2008) [16] dated these events at 10, 10, and 6.5 Ma rather than the 15.2, 15 and 13.3 inferred here. Nevertheless, the relative intervals are similar as are the absolute values when the wide confidence intervals are considered. In contrast to the earlier study, good congruence was observed between ecogeographical populations and genetic variation, whereas the 2008 paper did not infer a monophyletic BOL clade, instead BOL taxa were divided between MEK-1 and MEK-2. The pectinate appearance of the major clades in Fig 8 suggests rapid radiation of the snails soon after arrival of the founding population in each region (MEK-1, MEK-2, BOL, etc). This is in keeping with the idea that the Southeast Asian triculinae entered the region via the exposed Sunda shelf (now off the Vietnam coast) and founded the Cammon karst spring populations; thus retaining the same habitat as ancestral Triculinae in Hunan [73]. The founding population may have diverged into riverine forms and colonised Khorat basin, Bolovens plateau and Cambodia just prior to the uplift of the Bolovens plateau, around 15 Ma [83], and major isolating orogenic events.
Values of the population migration rate (Table 6) were much less than the threshold value of 1, at which point gene-flow begins to create the appearance of a single population [84]; therefore it was extremely unlikely that N. aperta exists as a single panmictic population. Nevertheless, unexpectedly high levels of migration were observed between CKS (TOT) and BOL and WBR (i.e. MEK+CKR); these were attributed to systematic error caused by the relatively small sample size and underestimated θ of CKS. In support of this, Attwood et al. (2008) [16], who also used Migrate (but not also IMa2p), and did not cluster CKS apart from CKR, reported values for NeM ranging from 0.000 to 0.005 between CKS+CKR and BOL+MEK1 or BOL+MEK2. Exclusion of CKS, allowed estimation of NeM between BOL and WBR which was found to be close to zero (0.001) WBR→BOL and almost ten times higher BOL→WBR; this is expected as N. aperta is thought to show very poor survival out of water. Even dispersal downstream (BOL→WBR) is likely to occur at low rates because habitats suitable for this snail are highly discontinuous [9]. The higher rate BOL→WBR is consistent with some colonisation of the Cambodian Mekong by snails from the rivers of Northeast Cambodia.
Estimates of Ne between 81084 and 313281 are very high. Even for dense stands of gregarious molluscs such as Ostrea edulis (the European flat oyster) Ne is reportedly only around 23000 [85]. N. aperta is much smaller than the aforementioned oyster, and lithic substrata provide a vast surface area along their ridges and crevices; therefore N. aperta populations may be very large. Although there is often no linear relationship between snail population density (or even density of infective snails) and prevalence of schistosomiasis in humans, the existence of very large numbers of snails around human settlements, each snail shedding few cercariae, is likely to favour transmission of S. mekongi. It is worth noting in this context, that in comparison with Biomphalaria glabrata (Say 1818) transmitting Schistosoma mansoni Sambon 1907, where one snail may shed over 2000 cercariae per day [86] (and prevalence of infection in the populations can be over 75% [87]), the cercarial ouput of N. aperta is much lower (as few as 23 per day per snail, with prevalence of 0.22% [88]). Most of the areas predicted to harbour N. aperta are inaccessible. The habitats that have been surveyed display extremely high population densities; for example, densities greater than 5000 snails per m2 have been reported for N. aperta in the Xe Kong river [89] (note this was prior to completion of the Nam Theun 2 dam). The present observations suggest that such high densities may be a feature of N. aperta populations across most of its range, including unsurveyed reaches of the upper rivers.
The findings of the present study were not inconsistent with a colonisation of Laos and Cambodia via southern Vietnam and into the Bolovens region; however, an earlier date (Miocene) was estimated relative to that reported by Attwood et al. (2008) [16]. In this respect the present date estimates agree with those of earlier phylogenetic studies [13,14]. The observation that, what little gene-flow there is in N. aperta was westwards from Bolovens to the other regions offers some support to a Sunda shelf-Vietnam colonisation; however, this pattern of gene-flow is relatively recent, with phylogenetics suggesting rapid colonisation of the entire range from the Cammon region and into Bolovens, just prior to isolating events associated with uplift of the Bolovens plateau (c.a. 15 Ma). Consequently, the results imply that the present distribution of N. aperta is more a result of history than of ecology, and that this snail is not currently necessarily currently limited by the distribution of suitable habitats. Nevertheless, dispersal capabilities appear to be very low (for example, closely parallel river courses in Cambodia harbour divergent populations) and expansion of the range is likely to occur at very low rates, even with human activity. Hydropower expansion, with 133 dams completed or proposed for the lower Mekong basin [90], is likely to effect gradual changes the distribution of snails, as changes in water depth create new suitable habitats [9]. The results confirm that the radiation of these snails and their associated schistosomes was heterochronous; this has implications for understanding of the snail-parasite association. Heterochronous evolution gives less opportunity for co-evolutionary arms races as proposed for other human schistosomes [91]. In this sense the findings support the conclusions of studies on S. japonicum, that ecological factors (e.g. schistosomes evolve to avoid snails that would release their cercariae in habitats that do not favour transmission) are perhaps more important than phylogenetics, and that host switches are more likely than previously thought [92]. The likelihood of host-switching in S. mekongi and other Asian species, relates to the chance of the parasite escaping the effects of snail control efforts and requires further investigation.
The present study supports view that the spring-dwelling N. aperta of the Cammon plateau in Laos are distinct ecotype, and that there is a marked lack of gene-flow between the northern and southern halves of the snail’s range. Consequently, further investigation is required to assess the potential for spring-dwelling snails to act as intermediate hosts for S. mekongi and to support transmission in habitats currently assumed to be free of schistosomiasis. The remarkable correspondence between ecogeographical area and snail population genetic structure suggests that some degree of ecological adaptation, in addition to physical barriers, inhibits introgression between these regions. In turn this suggests that persistence of snail populations in the Mekong river through spate periods is not achieved primarily by colonisation from seeder populations in tributary highland streams that experience a less severe flood cycle, which has been proposed as an explanation for snail persistence [93]. The finding suggests that snail control in snail habitats in the upper reaches of tributaries will have little impact on persistence of schistosomiasis transmission in the Mekong river itself. The only exception might be the Bolovens populations, as there is some gene-flow apparent between those and Mekong river snail foci. The observation of fine-scale clustering (granular spatial-genetic associations) in the snail populations also has implications for disease control. The finding Implies that snail-mediated reintroduction of schistosomiasis, from outside of local snail or parasite control intervention areas, is unlikely because snail populations are made up of many small and relatively discrete micro-populations.
In summary, correspondence between ecogeographical sub-populations and clades identified phylogenetically, and by genetic distance based clustering, in the present study, illustrates the value of improved sampling (both geographical and genomic). The study has shown that N. aperta exists as a metapopulation at multiple scales, including down to a micro-population-level granularity; this has implications for the design of schistosomiasis control interventions. Similarly, the lack of gene-flow between tributary populations (except perhaps those of Bolovens) and those of the Mekong implies that the effects of Mekong river and highland stream interventions will be independent. Interpretation of phylogenetic reconstructions implies that history shapes the current distribution of N. aperta and some expansion of the range is possible, especially after hydropower development alters the regional hydrology. The study has confirmed limited gene-flow between Cammon plateau populations in the North of the range and other populations, and that spring-dwelling snails are probably a distinct ecotype. Further work is needed to assess the epidemiological significance of the spring-dwelling taxa. The extent of the analyses was limited by the number of characters sampled and future studies should use more extensive sampling of the genome, which is now becoming more practical as technology improves. The present work used pre-existing data that included some very remote and/or now extinct sub-populations, as well as newly published data. The work therefore provides a record of the population-genetics of N. aperta prior to the impact of extensive hydropower development in the region, as well as an indication of the potential for range expansion as well as predictions of responses to schistosomiasis control.
Field Research: Material from Laos was collected with the permission of the Ministry of Public Health Lao PDR. Material from Cambodia, with the permission of the Cambodia National Malaria Center. Material from Thailand, with the permission of Mahidol University (Faculty of Science).
|
10.1371/journal.ppat.1006854 | Ebselen alleviates testicular pathology in mice with Zika virus infection and prevents its sexual transmission | Despite the low case fatality, Zika virus (ZIKV) infection has been associated with microcephaly in infants and Guillain-Barré syndrome. Antiviral and vaccine developments against ZIKV are still ongoing; therefore, in the meantime, preventing the disease transmission is critical. Primarily transmitted by Aedes species mosquitoes, ZIKV also can be sexually transmitted. We used AG129 mice lacking interferon-α/β and -γ receptors to study the testicular pathogenesis and sexual transmission of ZIKV. Infection of ZIKV progressively damaged mouse testes, increased testicular oxidative stress as indicated by the levels of reactive oxygen species, nitric oxide, glutathione peroxidase 4, spermatogenesis-associated-18 homolog in sperm and pro-inflammatory cytokines including IL-1β, IL-6, and G-CSF. We then evaluated the potential role of the antioxidant ebselen (EBS) in alleviating the testicular pathology with ZIKV infection. EBS treatment significantly reduced ZIKV-induced testicular oxidative stress, leucocyte infiltration and production of pro-inflammatory response. Furthermore, it improved testicular pathology and prevented the sexual transmission of ZIKV in a male-to-female mouse sperm transfer model. EBS is currently in clinical trials for various diseases. ZIKV infection could be on the list for potential use of EBS, for alleviating the testicular pathogenesis with ZIKV infection and preventing its sexual transmission.
| Male-to-female sexual transmission of ZIKV has been reported more than the other sexual modes. Here we observe the disease progression of ZIKV in the testicular tissue of mouse models by gaining access to interstitial cells, the basement membrane of seminiferous tubule, spermatogenic cells, and sperm. In addition, we develop an animal model to study male-to-female sexual transmission with a high transmissibility through seminal transfer that allows us to identify a possible therapeutic intervention. Our data indicate that testicular oxidative stress and pro-inflammatory response may contribute to testicular pathology as well as successful sexual transmission of ZIKV; thus, this could serve as a potential target for therapeutic purpose. We evaluated several potential antioxidants and found that ebselen (EBS) treatments in male mice display a better property to alleviate testicular pathogenesis and prevent seminal transmission of ZIKV than vitamin C or quercetin.
| Zika virus (ZIKV) is a single-stranded RNA virus that belongs to the Flaviviridae family [1]. The recent outbreak of ZIKV infection has created a public health emergency of international concern [2]. ZIKV infection displays nonspecific clinical features and generally causes mild symptoms in humans [1]. Although the case fatality with ZIKV infection is low, its infection has been linked to congenital microcephaly and Guillain-Barré syndrome [2, 3]. Currently, there is no approved antiviral drug or vaccines for ZIKV infection. Preventing ZIKV transmission is a significant strategy for disease control and management.
ZIKV is transmitted to people primarily via the bite of an infected Aedes species mosquito [1]. However, recent studies have suggested the sexual transmission of ZIKV in humans [4–6]. Infectious ZIKV in semen has been reported [7, 8]. Moreover, unlike in serum or urine samples, ZIKV RNA can still be detected in semen up to 62 to 188 days after the onset of symptoms [9–11]. Studies have reported more male to female transmission than with other sexual modes [12]. Notably, sexual transmission of ZIKV may occur before, during, or after the onset of symptoms [2]. This may suggest the complexity of risk factors for sexually transmitted ZIKV. Other indications that ZIKV is sexually transmissible come from animal studies. ZIKV infects and damages the testes of infected mice, which results in the loss of germ cells, degeneration of seminiferous epithelium, and poor-quality sperm [13, 14].
Oxidative stress plays an important role in the pathogenesis of both RNA and DNA viruses [15, 16]. Notably, in many cases of testicular dysfunction and infertility, oxidative stress appears to be a common underlying factor [17, 18]. Ebselen (EBS), an antioxidant currently in clinical trials for preventing and treating various disorders, has been shown to reduce oxidative stress and improve histopathological features in a testicular injury study [19, 20]. EBS catalyzes the reduction of reactive oxygen species (ROS) in a mode similar to glutathione peroxidase 4 (GPx4) [21]. GPx4 is highly expressed in spermatogenic cells and plays a dual role as an antioxidant enzyme and a structural protein [22]. In addition, EBS plays a role in inhibiting the catalytic activity of nitric oxide synthase (NOS); thus, it may also reduce the level of nitric oxide (NO) and NO-associated inflammatory cytokines [23, 24].
Here, we demonstrate that ZIKV increased testicular oxidative stress and pro-inflammatory response. Moreover, we propose a possible therapeutic intervention by using the antioxidant EBS to alleviate the testicular pathogenesis and prevent the sexual transmission of ZIKV.
AG129 mice lacking interferon-α/β and -γ receptors were challenged with ZIKV (strain PRVABC59) by subcutaneous route in the footpad and the progression of ZIKV infection in the testes and sperm was prospectively evaluated. On 3 days post-infection (dpi), ZIKV envelope protein (ZIKV-E) was found predominantly in interstitial cells and less so in the basement membrane of the seminiferous tubule (SNT) (Fig 1A; 3-dpi). On 6 dpi, ZIKV-E was detected in interstitial cells, the basement membrane of the SNT, and in spermatogenic cells (Fig 1A; 6-dpi). Notably, ZIKV-E expression in spermatogenic cells was even more prominent on 9 dpi (Fig 1A; 9-dpi). Autofluorescence of red blood cells was noted in histological sections, including in mock-infected testes, as previously reported, although we used paraformaldehyde fixation to minimize the interference [25]. It is intriguing to speculate that testicular infection of ZIKV is initially presented in interstitial cells adjacent to the endothelium and then spread to the basement membrane of the SNT to further infect the spermatogenic cells. In addition, ZIKV-E and double-stranded RNA (dsRNA) were detected in the head, middle piece, and tail of sperm of infected mice as early as 3 dpi (Fig 1B and S1 Fig). Moreover, infectious ZIKV was detected in sperm using a plaque-forming assay. The level of infectious ZIKV in the sperm gradually increased from approximately 2 log10 plaque-forming units (PFU)/ml on 3 dpi to 5 log10 PFU/ml on 9 dpi (Fig 1C, gray bar). In agreement with a previous study [13], the viremia level was significantly decreased and undetectable on 6 and 9 dpi, respectively (Fig 1C, striped bar). These data indicate that ZIKV infects testes and sperm of mice.
We also evaluated the pathological features of ZIKV infection in testes. ZIKV infection on 3 dpi did not greatly affect the histological structure of testis as compared with age-matched mock controls. The architecture of the SNT remained intact, with normal germinal epithelium and accumulation of sperm in the lumen (S2A Fig). On 6 and 9 dpi, ZIKV infection caused pathological features in the testis, including involution of the SNT, degeneration of spermatocytes, and the appearance of multinucleated giant cells in the lumen (S2A Fig).
In addition, we checked the expression of TRA98 and ETV5. TRA98 is a germ cell marker, whereas ETV5 plays a role in mediating the blood—testis barrier (BTB) [26, 27]. In agreement with previous study [13], the expression of TRA98 and ETV5 was gradually impaired by ZIKV infection (S2B Fig). TRA98 and ETV5 expression was reduced on 6 and 9 dpi as compared with controls (S2B Fig). These testicular pathological features suggest that ZIKV may impair spermatogenesis.
To confirm this notion, we evaluated sperm quality in terms of sperm count and morphology [28]. ZIKV did not affect sperm count on 3 dpi, but on 6 and 9 dpi, sperm number was significantly lower in ZIKV-infected than control mice (Fig 1D). Also, as early as 3 dpi, the proportion of abnormal sperm was greater in ZIKV-infected than controls (Fig 1E) and was even greater on 6 and 9 dpi in ZIKV-infected mice (Fig 1E). Confocal images showed abnormal sperm with short tail, hairpin tail, and bent middle piece in ZIKV-infected mice (S2C Fig). Overall, these data suggest that ZIKV infection damages testes and reduces sperm quality.
Taking into account that reactive oxygen species (ROS) play a role in viral pathogenesis as well as testicular pathology [15–18], we measured whether ZIKV infection increases testicular oxidative stress. Semen parameters have been widely used to gauge testicular pathology associated with oxidative stress [18]. A high level of ROS in semen is linked to poor sperm quality [29]. Moreover, a low expression of antioxidant enzymes including GPx4 is associated with abnormal spermatogenesis [30, 31]. We prospectively evaluated testicular oxidative stress of ZIKV-infected mice with or without treatment with the antioxidant ebselen (EBS) as outlined in Fig 2A. EBS at 10 mg/kg body weight (bw) has been shown to reduce testicular oxidative stress and testicular damage in animal study [20]. ROS levels were higher in sperm from ZIKV-infected mice than controls as early as 3 dpi (Fig 2B; gray vs. white bar). Notably, therapeutic treatments with EBS (10 mg/kg bw/mouse/ip/day) after ZIKV challenge significantly decelerated the elevated ROS levels in sperm on 6 and 9 dpi (Fig 2B, striped vs. gray bar). We further evaluated the level of nitric oxide (NO) in sperm because EBS has been demonstrated to inhibit the catalytic activity of nitric oxide synthase [23]. Consistently, ZIKV infection significantly increased NO level in sperm and treatment with EBS could repress the level (S3A Fig).
Moreover, the expression of the scavenging enzyme GPx4, the most highly expressed GPx isoform that supports the middle piece structure of sperm [22], was also affected by ZIKV infection. Control sperm showed normal morphology with expression of GPx4 in the middle piece (S3B Fig). ZIKV infection increased GPx4 expression in sperm on 3 dpi, presumably due to the altered cellular response associated with viral replication (Fig 2C and S3B Fig; Mock vs. ZIKV). However, on 6 and 9 dpi, GPx4 expression was lower in infected than control sperm (Fig 2C and S3B Fig). Remarkably, sperm of ZIKV-infected mice receiving EBS treatment showed significantly improved GPx4 expression on 6 and 9 dpi (Fig 2C and S3B Fig; ZIKV vs. ZIKV+EBS). In addition, we evaluated the spermatogenesis-associated-18 homolog (SPATA18/MIEAP) that also mediates both oxidative stress and structural stability of sperm. SPATA18 is involved in eliminating oxidized mitochondrial proteins and reducing ROS generation [32, 33]. In addition, poor expression of SPATA18 adversely affects the spermatogenesis and morphology of sperm [34, 35]. SPATA18 was detected in the head, middle piece and upper tail of control sperm but its expression was lower in sperm of ZIKV-infected mice as early as 3 dpi (Fig 2D and S3C Fig; Mock vs. ZIKV). This adverse effect of ZIKV infection on SPATA18 expression could be significantly alleviated by EBS treatment (Fig 2D and S3C Fig; ZIKV vs. ZIKV+EBS).
We further assessed testicular inflammatory response because high level of ROS may act as signaling molecules to provoke up-regulation of pro-inflammatory cytokines [36]. On 9 dpi, ZIKV significantly increased the production of seminal inflammatory cytokines including IL-1α, IL-1β, IL-6, IFN-γ and G-CSF (Fig 3A; white vs. gray bar). EBS treatment significantly reduced the levels of ZIKV-induced pro-inflammatory cytokines particularly IL-1β, IL-6 and G-CSF (Fig 3A, gray vs. striped bar). In addition, a massive infiltration of CD45+ and IL-1β+ cells was observed in the SNT of ZIKV-infected testes (Fig 3B and 3C), while EBS treatment reduced the testicular infiltration of CD45+ and IL-1β+ cells (Fig 3B and 3C). Taken together, ZIKV infection increases testicular oxidative stress and inflammatory response, which can be alleviated by treatment with the antioxidant EBS.
The role of ROS and pro-inflammatory cytokines in tissue injury has been noted [37]. So, we evaluated whether the inhibitory effect of EBS on testicular oxidative stress and inflammatory response may affect testicular pathology of ZIKV-infected mice on 9 dpi. The testes of control mice showed a regular structure of SNTs, with normal germinal epithelium and accumulation of sperm in the lumen (Fig 4A). In contrast, ZIKV-infected mice showed involution of SNTs, with degeneration of spermatogenic cells in the lumen (Fig 4A). Notably, ZIKV-infected mice receiving EBS treatment during 1–6 dpi displayed a minor pathological feature of SNTs, with few degenerated spermatogenic cells in the lumen (Fig 4A).
In addition, the expression of TRA98 and ETV5 was lower in ZIKV-infected than control testes (Fig 4B). The expression of TRA98 and ETV5 was improved in ZIKV-infected mice receiving EBS versus infected mice receiving solvent (Fig 4B). Sperm parameters were also significantly improved with EBS treatment. Total sperm count was more than three-fold higher in ZIKV-infected mice receiving EBS than those receiving solvent (Fig 4C). This improvement was accompanied by a better sperm morphology profile. EBS- and solvent-treated mice showed about 30% and 70% sperm with abnormal morphology, respectively (Fig 4D and 4E). Hence, EBS treatment attenuated the ZIKV-induced testicular pathology.
To address the impact of lacking interferon signaling on our findings with AG129 mice, we further performed ZIKV infection study by use of wild-type C57BL/6 mice with anti-IFNAR1 antibody pretreatment as outlined in Fig 5A. ZIKV infection significantly increased NO and ROS levels in sperm and both of them can be reduced by treatment with EBS after ZIKV challenge (Fig 5B and 5C). In addition, ZIKV infection significantly induced seminal inflammatory cytokines including IL-6, IL-10, G-CSF and GM-CSF (Fig 5D; ZIKV vs. Mock). Treatment with EBS repressed the levels of inflammatory cytokines particularly IL-6, G-CSF and GM-CSF (Fig 5D; ZIKV vs. ZIKV+EBS). Slight variation of seminal cytokine profile was noted between C57BL/6 and AG129 mouse models, probably due to the difference on genetic background and/or susceptibility to ZIKV. Viremia could be detected on 2 dpi and EBS treatment slightly reduced the viremia level (Fig 5E). Moreover, ZIKV-E was detected in interstitial cells and primary layer of SNT where germ and sertoli cells are resided (Fig 5F; ZIKV), whereas EBS treatment greatly limited the expression of ZIKV-E in the interstitial cells (Fig 5F; ZIKV+EBS). The expression of ZIKV-E in SNT was accompanied with the downregulation of TRA98 (Fig 5F; ZIKV), while TRA98 remained largely unaffected in EBS-treated mice (Fig 5F; ZIKV+EBS). ZIKV infection also caused degeneration of spermatogenic cells in the lumen of SNT (Fig 5G; ZIKV). In contrast, testis of EBS-treated mice displayed a relatively normal histological structure (Fig 5G; ZIKV+EBS). Taken together, ZIKV-induced testicular oxidative stress, inflammatory response, and pathology noted in the wild-type C57BL/6 mice can also be alleviated by EBS treatment.
To address whether the protective effect of EBS on testicular pathology is associated with its antiviral potential against ZIKV, we evaluated the anti-ZIKV activity of EBS. First, we determined the non-cytotoxic doses of EBS with lactate dehydrogenase (LDH) release assay on human microglial CHME3 cells. Treatment with EBS up to 25 μM had no cytotoxicity (Fig 6A), but at a higher dose (50 μM) EBS displayed significant cytotoxicity on CHME3 as reported earlier on human hepatoma HepG2 cells [38]. Minor reduction of ZIKV progeny production and viral E protein expression was noted in cells treated with non-cytotoxic 25 μM EBS (Fig 6B and 6C). Furthermore, in AG129 mice challenged with ZIKV, treatment with EBS on 1–6 dpi displayed a minor effect on overall animal survival and viremia level (Fig 6D and 6E). However, EBS treatment greatly limited the expression of ZIKV-E in testes as compared with solvent control, which showed a disperse expression of ZIKV-E in interstitial and spermatogenic cells (Fig 6F). Moreover, as analyzed by western blotting, EBS treatment reduced the expression of ZIKV-E in testis (2.2 fold reduction) and brain (1.4 fold reduction), but not in the spleen (S4 Fig). Thus, EBS exhibited a weak anti-ZIKV activity in culture cells and a tissue-specific antiviral activity in challenged mice. Therefore, the protective effect of EBS on testicular pathology is predominantly associated with its property to reduce testicular oxidative stress and inflammatory response.
We then examined whether EBS treatment may also prevent sexual transmission of ZIKV. Animal models for studying sexual infection of ZIKV have been reported [39–41]. Here, we performed sperm—vaginal transfer from one male to one female mouse as outlined in Fig 7A. Briefly, 50 μl of sperm sample was collected from EBS-treated or solvent-treated ZIKV-infected mice at the indicated time and used for vaginal inoculation into female mice by use of a bent, blunt-end 22-gauge needle to avoid uterine injury. Before this experiment, we evaluated the effect of EBS treatment on ZIKV level in sperm by plaque-forming assay. Treatment with EBS significantly reduced the ZIKV level in sperm as compared with solvent treatment on 6 dpi (Fig 7B; 6-dpi). Although EBS treatment was stopped on 6 dpi, the level of ZIKV in sperm remained significantly lower than in controls on 9 dpi (Fig 7B; 9-dpi).
Next, we used this infectious sperm to challenge female mice by vaginal inoculation. The survival of female mice receiving sperm from solvent-treated mice collected on 6 dpi was 50%, whereas no female mice receiving sperm from EBS-treated mice died (Fig 7C; black vs. empty circle). Sperm from solvent-treated mice collected on 9 dpi led to 0% survival of recipient female mice, whereas 83.33% of female mice receiving sperm from EBS-treated mice survived (Fig 7C; black vs. empty triangle). High viremia was observed in all of the female mice receiving sperm from solvent-treated mice. In contrast, viremia could only be detected in one of the female mice receiving sperm from EBS-treated mice (S5 Fig), well correlating with the survival data.
We also tested the efficacy of other antioxidants, vitamin C and quercetin, to improve testicular pathology and prevent sexual transmission of ZIKV (S6 Fig). Vitamin C and quercetin are the most abundant antioxidants found in dietary sources [42]. Treatment with vitamin C and quercetin did not affect the overall survival of recipient female mice. However, the median survival time (T50) of female mice receiving sperm from vitamin C-treated mice was significantly improved than those receiving sperm from solvent-treated mice (S6D Fig; black vs. empty circle; T50 = 13.5 vs. 17.5 days; P = 0.011). The T50 did not differ between female mice receiving sperm from quercetin-treated mice and controls (S6D Fig; black circle vs. triangle; T50 = 13.5 vs. 12 days; P = 0.677). These data suggest that different antioxidants may have different efficacy in preventing the sexual transmission of ZIKV. Finally, we examined tissue from female mice receiving sperm transfer to confirm the systemic infection of ZIKV. RNA of ZIKV was detected in brain, ovary-fallopian tubes, and spleen (Fig 7D). Collectively, these data suggest that EBS treatment may prevent seminal transmission of ZIKV.
Here we show the disease progression of ZIKV in the testicular tissue of AG129 mice lacking interferon receptors and also in wild-type C57BL/6 mice pretreated with anti-IFNAR-1 antibody; both animal models are useful for studying ZIKV replication, tropism, immunity, and transmission [41, 43, 44]. In addition, we developed an animal model to study male-to-female sexual transmission of ZIKV and identified a possible therapeutic intervention to minimize the transmission. ZIKV progressively infects testis by gaining access to interstitial cells, the basement membrane of the SNT, spermatogenic cells, and sperm. ZIKV infection leads to several testicular pathologies, including involution of the SNT, degeneration of spermatogenic cells, oligospermia, a high proportion of sperm with abnormal morphology, poor expression of TRA98 and ETV5. In agreement with a previous study [14], our data suggest that ZIKV infection may cause male infertility. Moreover, ZIKV infection increased ROS levels and impaired cellular antioxidant of sperm. ROS is considered a risk factor for some diseases including cancer and infection [45, 46]. Our data indicate that high ROS levels may be associated with high level of seminal pro-inflammatory cytokines, poor testicular outcome and high level of infectious ZIKV in sperm. ROS derived as a byproduct of viral replication can accelerate pro-inflammatory response to cause endothelial barrier disruption and tissue injury [15, 37]. Thus, oxidative system may be one of the cellular factors attributed to testicular pathogenesis and sexual transmission of ZIKV. In testicular dysfunction, reducing ROS levels is a common therapeutic strategy to improve the disease prognosis [17, 18]. We used this strategy to alleviate testicular pathogenesis and prevent seminal transmission of ZIKV.
Sexual transmission of ZIKV could be as critical as mosquito-borne transmission in the context of maternal infection. By August 2016, the World Health Organization had documented the sexual transmission of ZIKV in 11 countries worldwide [12]. The presence of infectious ZIKV in the semen poses a potential risk of sexual transmission. However, one may argue a sustainable transmission of ZIKV via sexual intercourse [47]. Our data indicate that seminal ROS and inflammatory cytokines may contribute to successful sexual transmission of ZIKV. Despite the substantial level of ZIKV detected in the sperm of EBS-treated mice on 6 dpi (Fig 7B; ZIKV+EBS, 6-dpi), the sperm failed to establish ZIKV transmission in female mice (Fig 7C; empty circle). Furthermore, the sperm of EBS-treated mice on 9 dpi caused only 16.67% mortality in female mice, although the ZIKV level in sperm did not greatly differ from that of solvent-treated mice on 6 dpi, which killed 50% of the transmitted female mice (Fig 7B; ZIKV+EBS [9-dpi] vs. ZIKV [6-dpi], P = 0.37). This finding can be explained in part by a relatively lower level of seminal ROS in EBS- than solvent-treated mice (Fig 2B; ZIKV+EBS [9-dpi] vs. ZIKV [6-dpi], P = 0.03). Correlation study and regression modeling are warranted to determine whether the level of ROS and inflammatory cytokines can be a predictor of successful sexually transmitted ZIKV. It has recently become apparent that semen acts directly on tissues in the female reproductive tract. Indeed, a substantial level of seminal pro-inflammatory cytokines and ROS may provoke inflammatory responses in the female reproductive tract and cause tissue injury [37, 48]. Consequently, these processes may create access to and favorable conditions for seminal infectious ZIKV to establish an initial infection in the female reproductive tract before its systemic infection.
EBS displays a weak effect on ZIKV replication in culture cells and ZIKV-induced lethality in challenged animals. However, a tissue-specific restriction on ZIKV was noted in the infected mice. We speculate that the protective effect of EBS on testicular pathology and sexual transmission of ZIKV is associated with its efficacy to reduce the level of oxidative stress and pro-inflammatory cytokine in testes. EBS, a synthetic antioxidant, has a unique characteristic as compared with the other antioxidants. EBS may reduce the production of NO by inhibiting the catalytic activity of nitric oxide synthases (NOS) [23]. Consequently, EBS may also suppress the level of NO-associated inflammatory cytokines [49]. Recently, EBS has been demonstrated to reduce the level of multidrug-resistant staphylococcal-associated inflammatory cytokines including IL-1β and IL-6 [50]. A substantial level of NO activates IL-6 production through MAPK signaling pathway and increased IL-6 mRNA stability [24]. In addition, the protective effects of EBS against oxidative stress rely on its GPx-like activity [19]. It is an effective scavenger of lipid hydroperoxides mimicking the action of GPx4 [21]. GPx4 is one of the most important GPx isoforms in a testicular context and is abundantly expressed in spermatogenic cells, particularly sperm [22, 51]. Therefore, EBS may have beneficial properties for the treatment of testicular pathology while more effective testis-specific drugs or antioxidants remain to be discovered. Our data indicate that EBS also displays efficacy to reduce ZIKV load in brain. Interestingly, GPx4 plays a role in brain development and neuronal apoptosis [52]. GPx4 deficient mice die in utero at midgestation with a major defect in embryonic brain development [53]. The implication of GPx4 in female reproduction remains unknown and further studies to evaluate whether EBS may alleviate ZIKV pathogenesis in female reproductive system are warranted.
Antioxidants have been used to improve the prognosis of viral diseases. In a case report, high doses of intravenous vitamin C given over 2 days may have resolved symptoms of a patient with Chikungunya [54]. In addition, vitamin C given over 3 days in increment doses up to 75 g substantially improved symptoms of a patient with ZIKV infection [55]. The mechanism of action of antioxidant vitamin C remains controversial. However, studies suggest that vitamin C reduces inflammatory responses and ROS levels in a glutathione-independent mechanism [56, 57]. Similarly, a flavonolic antioxidant, quercetin, displays its antioxidant properties in part by suppressing the inflammatory response and scavenging peroxyl radicals [58, 59]. However, we did not see a potential use of quercetin (10 mg/kg bw) to prevent sexual transmission of ZIKV. We speculate that the efficacy of antioxidants to prevent sexual transmission of ZIKV may vary.
The dosage and treatment duration of antioxidant as a therapeutic agent must be carefully interpreted. Antioxidants have beneficial effects at a physiologic dose. However, at a relatively high dose, they may have unfavorable effects [42]. The dosage of 10 mg EBS/kg bw/day for mice is equivalent to 0.811 mg/kg bw/day for humans, as calculated by a published method [60]; thus, for a 60-kg person, the dosage would be 48.6 mg/day. We used the treatment duration of EBS of 6 days because viremia can still be detected on 6 dpi. EBS is being used in clinical trials for various diseases [19]. A study to evaluate the safety and pharmacokinetics of 200 mg EBS oral capsule (SPI-1005) in healthy adults has been completed (NCT01452607). Therefore, the present study strongly argues for a potential clinical use of EBS to alleviate the testicular pathogenesis with ZIKV infection and its sexual transmission.
The mouse experiments were conducted according to the guideline outlined by Council of Agriculture Executive Yuan, Republic of China. This animal protocol was approved by the Academia Sinica Institutional Animal Care and Use Committee (Protocol no. 16-06-966) and were performed in accordance with the guidelines. Infection was performed in mice under isoflurane anesthesia and all efforts were made to minimize animal suffering.
Five-week-old male AG129 mice with interferon-α/β and -γ receptor knockout were subcutaneously infected in the footpad with 5×104 plaque-forming units (PFU) of ZIKV per mouse. Testes, sperm, and sera were collected on 3, 6, and 9 days post-infection (dpi). Five-week-old male C57BL/6 mice were pre-treated with purified anti-mouse IFNAR-1 antibody (1 mg/kg body weight/mouse/intraperitoneal), one day prior to subcutaneous inoculation of ZIKV (1×105 PFU/mouse) in the footpad. Testes, sperm, and sera were collected on 9 dpi. To study the effect of the antioxidant ebselen (EBS, Cayman Chemical, CAS 60940-34-3) on testicular pathology, sexual transmission, and ZIKV-induced lethality, mice were treated with EBS (10 mg/kg body weight/mouse/intraperitoneal/day) or solvent control (phosphate buffered saline [PBS]+10% polyethylene glycol as a co-solvent) on days 1–2 (Group: 3-dpi), 1–5 (Group: 6-dpi), or 1–6 (Group: 9-dpi) after ZIKV infection. To study the effect of the antioxidants vitamin C (L-Ascorbic acid, Sigma, A4403) and quercetin (Alfa Aesar, A15807), mice were treated with vitamin C (Vit C, 10 mg/kg body weight/mouse/intraperitoneal/day), quercetin (Quer, 10 mg/kg body weight/mouse/intraperitoneal/day), or solvent control (PBS) on day 1–6 after infection. To study the sexual transmission of ZIKV, we used a modified protocol of artificial insemination in mice [61]. Estrous cycle may interfere the susceptibility of mice to transgenital transmission of ZIKV [44]. Therefore, four-week-old female AG129 mice were used to minimize the effect of hormonal staging in evaluating the efficacy of EBS. Briefly, 50 μl sperm of one male mouse was transferred into the vagina of one female mouse. A 120°-bent, blunt-end 22-gauge needle (Taiwan Vacuum Technology, D6622-45) was used for vaginal inoculation to avoid uterine injury. The mice were checked daily for severe symptoms including limb paralysis. Euthanasia of mice with severe limb paralysis was performed as the endpoint of animal survival. The brain, spleen, ovaries and fallopian tubes were immediately isolated after the animal died for estimating ZIKV RNA level.
ZIKV PRVABC59 strain (2015 Puerto Rico strain, Genbank accession: KU501215) was kindly provided by the Centers for Disease Control, Taiwan. The virus was propagated in C6/36 mosquito cells (ATCC: CRL-1660) and the level of infectious virus particles was measured by plaque-forming assay (PFU/ml) in Vero cells (ATCC: CRL-1587) as described [62]. Total RNA was extracted from homogenized animal tissues by using the RNeasy kit (QIAGEN) and level of ZIKV RNA was analyzed by real-time RT-PCR (Roche LightCycler 2.0). Briefly, cDNA was reverse-transcribed from 1 μg of RNA with random hexamers by using the ThermoScript RT kit (Invitrogen). PCR involved use of the LightCycler FastStart DNA Master PLUS SYBR Green I kit (Roche) with the primers for ZIKV (5′-CCGCTGCCCAACACAAG-3′ and 5′-CCACTAACGTTCTTTTGCAGACAT-3′) and β-actin (5′-TCCTGTGGCATCCACGAAACT-3′ and 5′-GAAGCATTTGCGGTGGACGAT-3′). ZIKV RNA level was normalized to that of actin for relative quantification.
Sperm was isolated as described [61] with modification. To isolate mature sperm, cauda epididymides were placed in 100 μl Vitro Fert medium (Cook Medical, K-RVFE 50) and gently sliced by using an 18-gauge needle. After a few minutes, epididymides were removed and suspension (sperm) was used for evaluating sperm quality, reactive oxygen species (ROS) and nitric oxide (NO) level, cytokine profile, plaque-forming assay, and sexual transmission study.
Sperm quality (sperm count and morphology) was evaluated according to the guidelines of the World Health Organization [28]. Trypan blue staining was performed and sperm number was determined by using a hemocytometer. Air-dried sperm smears were used for morphological and immunostaining assays. To evaluate sperm morphology by microscopy, sperm smears were fixed with 4% paraformaldehyde (PFA) and stained with Hoechst (Invitrogen, H21492) for nuclei and with CellTracker Orange (Invitrogen, C2927) for cytoplasm without permeabilization. For immunofluorescence assay, sperm smears were permeabilized with 0.5% TritonX-100 for 10 min, blocked with 5% skim milk in PBS for 1 hour and immunostained overnight with anti-flavivirus E [63], anti-dsRNA (Scicons, J2), anti-GPx4 (Abcam, Ab125066), or anti-SPATA18 (Abgent, AP13960a) antibodies. Confocal analysis was performed by using confocal laser scanning microscope, ZEISS LSM 700. Fluorescence intensity of confocal images was determined by using ImageJ (US National Institutes of Health). To determine the level of infectious ZIKV, sperm was subjected to five rapid freeze-thaw cycles to release virus particles from sperm before plaque-forming assay.
The level of ROS and NO in sperm was determined by using the OxiSelect Intracellular ROS (Cell Biolabs, STA-342) and NO (Cell Biolabs, STA-800) Assay kit, respectively. The fluorescence probe DCFH-DA was used to measure hydroxyl, peroxyl, or other ROS level within a cell, while NO fluorometric probe was used to measure the level of intracellular NO. Briefly, sperm was plated on 96-well black microplates at 5x104 sperm/well, then incubated with 1X DCFH-DA or 1X NO fluorometric probe (diluted in Hank’s Balanced Salt Solution without phenol red) for 45 min at 37°C. Sperm treated with 500 μM H2O2 was a positive control. Fluorescence intensity was determined by using a fluorescence reader at 480/530 nm (excitation/emission) [64]. For cytokine assay, sperm samples were centrifuged and supernatants (10 μl) were subjected to mouse inflammatory cytokines multi-analyte ELISArray (Qiagen, MEM-004A).
Testes were collected immediately after death and fixed overnight in Bouin’s solution (Sigma, HT10132), washed with 50% alcohol and then underwent tissue processing and embedding. Haematoxylin-eosin staining was performed on 3-μm-thick testes sections for histology. For immunohistochemistry, testes sections were immunostained with anti-flavivirus E [63], anti-TRA98 (Abcam, ab82527), anti-ETV5 (Abcam, ab102010), anti-CD45 (Novusbio, NB100-77417), and anti-IL1β (Abcam, ab2105) antibodies. Testis, brain, and spleen were digested in cold RIPA lysis buffer, sonicated at 21% amplitude for 2 min, and then incubated on ice for 30 min. Homogenate was centrifuge at 10,000 rpm for 10 min at 4°C and supernatants were collected for western blot analysis to evaluate the expression of ZIKV-E protein.
Human microglial CHME3 cells [65] were grown in DMEM containing 10% fetal bovine serum. Cytotoxicity of EBS was determined by using lactate dehydrogenase (LDH) assay (Roche). Briefly, cells treated with EBS at the indicated doses for overnight underwent LDH assay. Sample absorbance was determined by using an ELISA reader (molecular device) at 490 nm. For antiviral study, cells in 12-well plates were adsorbed with ZIKV at 0.1 of multiplicity of infection for 2 hours with the indicated doses of EBS, washed thoroughly to remove unbound viruses, then incubated for another 24 hours in the presence or absence of EBS. The antiviral effect of EBS was evaluated by immunofluorescence and plaque-forming assays.
Data were compared by Mann-Whitney or Kruskal-Wallis Bonferroni post-hoc test. Statistical significance was set at P < 0.05. Survival curves were descriptively analyzed by using SigmaPlot v10.0 (Systat Software) and compared by Log-rank test with use of Prism v5.0 (GraphPad Software).
|
10.1371/journal.pgen.1002200 | The Caenorhabditis elegans GATA Factor ELT-1 Works through the Cell Proliferation Regulator BRO-1 and the Fusogen EFF-1 to Maintain the Seam Stem-Like Fate | Seam cells in Caenorhabditis elegans provide a paradigm for the stem cell mode of division, with the ability to both self-renew and produce daughters that differentiate. The transcription factor RNT-1 and its DNA binding partner BRO-1 (homologues of the mammalian cancer-associated stem cell regulators RUNX and CBFβ, respectively) are known rate-limiting regulators of seam cell proliferation. Here, we show, using a combination of comparative genomics and DNA binding assays, that bro-1 expression is directly regulated by the GATA factor ELT-1. elt-1(RNAi) animals display similar seam cell lineage defects to bro-1 mutants, but have an additional phenotype in which seam cells lose their stem cell-like properties and differentiate inappropriately by fusing with the hyp7 epidermal syncytium. This phenotype is dependent on the fusogen EFF-1, which we show is repressed by ELT-1 in seam cells. Overall, our data suggest that ELT-1 has dual roles in the stem-like seam cells, acting both to promote proliferation and prevent differentiation.
| Stem cells can both produce differentiated cells and self-renew, producing more stem cells. Choosing between these opposing options is critical for development. Here, we have investigated the molecular genetics underlying this choice in the nematode worm, C. elegans, using the seam cells as a model of stem cell divisions. The transcription factor RNT-1 works together with BRO-1 (homologues of mammalian RUNX and CBFβ genes, respectively) to regulate proliferation of the seam cells, reflecting the roles of RUNX/CBFβ in mammalian stem cells. To better understand how bro-1 is regulated, we looked for conserved regions of non-coding DNA, likely to be of functional importance. We identified a 122 bp conserved non-coding element that is necessary and sufficient for bro-1 expression. Subsequent analysis suggested that the GATA transcription factor ELT-1 directly regulates bro-1. We have found that ELT-1 actually performs two distinct roles, promoting proliferation of seam cells while also preventing them from inappropriately fusing with surrounding tissue and losing their stem-like properties. Furthermore, we propose a link between the retention of stem cell properties and the maintenance of seam cells in a distinct compartment, in which they are protected from differentiation.
| The regulation of the decision between cell proliferation and differentiation is a key aspect of metazoan development, ensuring that the correct number and type of cells are present. The RUNX family of transcription factors (RUNX1, 2 and 3), together with their binding partner CBFβ, are key players in the control of stem cell proliferation in haematopoiesis [1]–[3], osteogenesis [4]–[6] and neurogenesis [7], [8]. Moreover, mutations in these genes are known to cause a variety of diseases, with both CBFβ and RUNX genes having the potential to act as either oncogenes or tumour suppressors, depending on the nature of the mutations and the context in which they act [9].
In the nematode worm, Caenorhabditis elegans, RUNX and CBFβ are each represented by a single gene, rnt-1 and bro-1 respectively. Reflecting their role in mammalian tissues, RNT-1 and BRO-1 are involved in the regulation of the division patterns of the stem cell-like seam cells [10]–[12]. The seam comprises a specialized epithelial tissue of lateral hypodermal cells located along each side of the worm and provides a tractable model system for studying the balance between proliferative and differentiative developmental decisions. The seam cells are considered stem cell-like because of their ability to both self-renew and to produce a variety of differentiated cell types; in addition to increasing the number of seam cells during some divisions, the lineage contributes cells to the hypodermis (an epithelial tissue), as well as giving rise to neurons and glial cells [13]. Seam cells do not self-renew throughout adult life, and have not yet been shown to reside in a classic “niche”, hence the qualification stem-like, but they do provide a useful simplified model for the stem cell mode of division throughout larval development that is well established [14]–[21].
The choice between the two very different developmental alternatives of proliferation and differentiation is intrinsically linked to the way in which the seam cells divide. Symmetrical divisions allow the expansion of the stem cell pool, as both daughters of the division retain the seam (stem cell-like) fate. Conversely, asymmetrical divisions initiate developmental pathways that result in the production of differentiated cell types. Whilst one daughter of the division (the posterior daughter, in the case of cells V1-6 in the seam lineage) retains the seam fate, the other loses its ability to proliferate further and instead proceeds to differentiate into one of a number of different cell types, most commonly assuming the hypodermal fate and fusing with the hyp7 syncytium. In this case, the differentiation event involves fusion with a different cell type, although seam cells can differentiate independently of fusion, for example taking on neuronal fates [13]. At the end of larval development, seam cells undergo homotypic fusion, producing the fully differentiated adult seam cell syncytium that secretes the alae [13]. The seam cells divide repeatedly throughout larval development, and the form that the division takes (symmetrical or asymmetrical) is critically dependent on the developmental stage of the worm. The nature of these stem cell-like divisions, and thus the fate of the daughter cells arising from them, therefore requires precise regulation both spatially and temporally.
In C. elegans, rnt-1 and bro-1 are rate-limiting regulators of seam cell divisions, with both genes being required for proliferation in this tissue. In rnt-1 and bro-1 mutants, the number of seam cells present in individuals is reduced due to division failures within the seam lineage [12]. Furthermore, over-expression of rnt-1 and bro-1 results in hyperplasia of the seam [10]. In this study, we have examined the mechanisms underlying the regulation of bro-1 expression. We identified a small (122 bp) conserved non-coding element (CNE) within the first intron of bro-1, which we show to be both necessary and sufficient for bro-1 expression in the seam. In order to identify direct upstream regulators of bro-1, a yeast one-hybrid screen was performed using this conserved element as bait. This, coupled with an in vitro binding assay, demonstrates that the GATA transcription factor ELT-1 directly regulates bro-1 through this CNE. The involvement of both GATA and RUNX/CBFβ factors in regulating a stem cell-like lineage in C. elegans thus mirrors the situation in mammalian systems, where the activities of these proteins are intricately linked with the specification of stem cell populations (for example haematopoietic stem cells [22]).
We then sought to further investigate the role of ELT-1 in seam cells. Analysis of the elt-1 RNAi phenotype suggests that ELT-1 has both bro-1-dependent and independent functions in the seam. ELT-1 functions through bro-1 to promote proliferation of seam cells and through eff-1 to repress inappropriate differentiation and fusion with the hypodermal syncytium. We also show that apical junction components themselves are important for the maintenance of seam cell fate, providing the correct contacts and therefore creating an environment in which cells are protected from differentiation signals emanating from surrounding tissues. Any disruption to the continuity of this environment contributes to the initiation of differentiation and loss of the seam fate.
Comparative sequence analysis revealed high levels of conservation between exonic regions of bro-1 in C. elegans and C. briggsae (73, 85 and 75% identity for exons 1, 2 and 3 respectively, compared to an average of 53% identity for introns 1 and 2). In contrast, we found no significant blocks of conservation in the 0.8 kb between bro-1 and the next upstream gene. However, a 122 bp region within the first intron was found to be highly conserved (3 species alignment shown in Figure 1B) (69% identity compared to 48% for the rest of the introns). We termed this the “bro-1 Conserved Non-coding Element” (bro-1 CNE).
Next, we tested the ability of the bro-1 CNE to act as an enhancer for cell-specific expression. When used in conjunction with the pes-10 minimal promoter (bro-1 CNE::gfp), the CNE was able to drive seam cell-specific GFP expression in worms (Figure 2A–2C). A vector containing a similar-sized fragment of a different, non-conserved region of intron 2 failed to drive any discernable GFP expression (data not shown). Furthermore, when the intergenic region between bro-1 and the nearest upstream gene was used, no GFP expression was evident (data not shown). To further characterise the importance of the CNE in regulating bro-1 expression, we first deleted the 122 bp region from the bro-1::dsRED2 construct [10]. The wild type construct (containing the full bro-1 genomic region) drives expression in seam cells and rescues the bro-1 mutant male tail phenotype caused by division failures in V and T lineage seam cells (Figure 2J). In contrast, the mutagenised bro-1::dsRED2 construct did not express in seam cells, and was unable to rescue the bro-1 male tail phenotype (Figure 2K). Secondly, we used the CNE (together with the pes-10 minimal promoter) to drive expression of bro-1 cDNA::gfp (see Text S1 and [23]). This construct drove seam cell expression, and rescued the bro-1 mutant phenotype (Figure 2L). Thus, the bro-1 CNE is both necessary and sufficient for bro-1 expression in seam cells.
Next, we used a yeast one-hybrid system to determine which transcription factors are able to bind to the bro-1 CNE. Three tandem copies of the CNE were used as bait and screened with a mixed stage C. elegans transcription factor cDNA library. Of 120 positive colonies, 110 contained the clone encoding the GATA factor ELT-1, demonstrating that ELT-1 protein binds to the bro-1 CNE in this assay.
The bioinformatics software Patch [24] revealed the presence of possible binding sites for GATA transcription factors in the CNE (Figure 1B), two of which are relatively well conserved across Caenorhabditis species [25], [26]. Site A matches the GATA consensus sequence WGATAR, while site B is slightly different (AGATTA), but matches the target sequence for the human GATA family member GATA6 [26]. To investigate the significance of these sites, both were deleted separately from the bro-1::dsRED2 rescuing construct. While deletion of site A had no effect on tail rescuing ability, deletion of site B abolished the ability of the construct to rescue bro-1 mutant tails (Figure 3A).
To confirm this interaction, an Electrophoretic Mobility Shift Assay (EMSA) was performed using in vitro translated ELT-1 protein and a portion of the bro-1 CNE containing GATA site B as a probe. ELT-1 protein was able to bind the CNE, resulting in a shift in the position of the labelled probe (Figure 3B). Cold competitor probe was able to compete with the identical labelled probe, resulting in a diminution in intensity of the shifted band. However, cold probe with a mutation in GATA site B (AGATTA to ATAGTA) was unable to compete with the wild type labelled probe, suggesting that ELT-1 protein interacts directly with GATA site B.
Other C. elegans members of the GATA family of transcription factors were tested for their ability to bind the bro-1 CNE using the same band shift assay; all failed to shift the labelled probe (Figure 3C). Taken together, these data suggest that amongst the members of the GATA family in C. elegans, ELT-1 alone mediates bro-1 transcriptional activity in the seam by direct binding to GATA site B within the CNE.
The fact that bro-1 seam expression disappears when ELT-1 binding site B within the bro-1 regulatory region is deleted suggests that ELT-1 is an activator of bro-1 expression. To confirm this, we monitored bro-1::gfp expression in animals subjected to elt-1 RNAi and observed a decrease in signal intensity (data not shown). It was necessary to reduce the level of elt-1 expression by RNAi as elt-1 alleles are embryonic lethal; therefore only the effects of a small reduction in elt-1 expression could be analysed. To quantify this, we measured endogenous elt-1 and bro-1 transcript levels by qRT-PCR in animals surviving the elt-1 RNAi treatment. These animals (most of which exhibited seam defects) were found to have a 1.5 fold reduction in elt-1 transcript levels and a 1.7 fold decrease in bro-1 transcript levels, demonstrating that ELT-1 plays a major role in regulating bro-1 expression in vivo.
ELT-1 has previously been reported to be essential for seam differentiation and maintenance, with elt-1(RNAi) animals having fewer seam cells as assayed by the seam cell marker scm::gfp [27]. Significantly, the average number of seam cells in elt-1(RNAi) animals (13.6 seam cells per side, ±0.3, n = 83) is very similar to that of bro-1 mutants (14.0 seam cells per side, ±0.3, n = 91) and significantly lower than animals fed control HT115 bacteria containing the empty feeding vector L4440 (15.8 seam cells per side, ±0.1, n = 30). The cellular basis of this phenotype has not been previously described, therefore lineage analysis was performed to elucidate the cellular mechanism of seam cell loss.
elt-1(RNAi) worms have variable seam division failures (Figure 4A). These defects were observed during both the symmetrical and asymmetrical divisions of L2, suggesting that the role of ELT-1 is not limited to one type of division. However, given the difficulties in pursuing the lineage analysis past L3 (these worms are very sick), we cannot exclude the possibility that the role of ELT-1 in regulating seam cell division is limited to the L2 stage. Division failures were not observed during the L1 asymmetric division, although embryonic developmental abnormalities often meant that the number of seam cells present at the start of L1 was lower than that of wild type worms. Nevertheless, the ability of the remaining seam cells to undergo the asymmetric L1 division parallels the situation in rnt-1 and bro-1 mutants, where division failures are restricted to the subsequent divisions. These division defects are distinct from the classic retarded heterochronic phenotypes, in which stage-specific stereotypical division “cassettes” occur at the wrong stage. In the case of lin-4, the L1 pattern of division is re-iterated at later stages, resulting in fewer seam cells overall [28]. In the case of elt-1(RNAi) and rnt-1/bro-1 mutants, the defects seen were not representative of stereotypical division patterns occurring at the wrong stage, but rather involved outright division failures, or occasionally symmetrisation towards the hypodermal fate.
To assess whether the division failures observed in elt-1(RNAi) worms were correlated with loss of seam fate, the integrated scm::gfp reporter wIs51 [29] was used as a seam cell marker. We observed that even worms which do not show gross morphological abnormalities, and which therefore can be followed by lineage analysis throughout development, frequently exhibit loss of scm::gfp from usually one or two seam cells from the late L1 stage onwards. While the seam cell remains clearly visible under DIC (retaining its distinctive eye-shaped seam morphology), GFP expression fades over 30–60 minutes until the cell shows no fluorescence at all (Figure 4B). To confirm the identity of these ‘seam’ cells that fail to express scm::gfp, cell boundaries were visualised using the ajm-1::gfp reporter (which marks the apical junctions between seam cells and the surrounding hyp7 syncytium, in which it is not expressed). These cells were always bounded by AJM-1::GFP, confirming their seam identity (as suggested by their morphology). Thus, when seam cells are counted in elt-1(RNAi) animals using scm::gfp, not all of the seam cells present will be observed. In addition to division failures, therefore, the loss of expression of the ‘seam’ marker scm::gfp also likely accounts for the apparent progressive loss of seam cells previously reported in elt-1(RNAi) worms [27].
Interestingly, a similar phenotype is observed in bro-1 and rnt-1 mutants (data not shown). However, while loss of this marker has been attributed to degeneration of seam cells [27], we have not observed this phenomenon in either rnt-1/bro-1 mutants or elt-1(RNAi) worms; seam cells retain their characteristic “eye” shape and position within the seam line, but may lose scm::gfp in these backgrounds. Given the prevalence of using scm::gfp as a marker of seam fate, we were surprised to observe that scm::gfp expression does not in fact seem to be tightly correlated with proliferative potential; lineage analysis revealed that cells that stop expressing scm::gfp often undergo the correct number of divisions at the appropriate times. Conversely, we observed division failures in seam cells strongly expressing scm::gfp. Taken together, this suggests that while scm::gfp may mark seam cells during normal development, it is not an infallible marker for seam cell fate, perhaps because particular elements of seam ‘fate’ are uncoupled in elt-1(RNAi) and rnt-1/bro-1 mutant animals.
Cells that remained in the seam line but failed to divide, always retained expression of ajm-1::gfp, regardless of whether or not they expressed scm::gfp (Figure 4B and 4C). This suggests that their failure to divide results not from a change of fate per se (from seam to hypodermis), but rather from a loss of proliferative ability. This is supported by dpy-7p::yfp expression, which marks cells that have adopted a hypodermal fate but is not expressed in seam cells; cells bounded by AJM-1::GFP never expressed dpy-7p::yfp, even when they fail to undergo scheduled divisions (data not shown).
Unlike the situation in bro-1 mutants, however, observation of elt-1(RNAi) animals carrying the integrated ajm-1::gfp construct revealed that some cells, at the same time as losing scm::gfp expression, lose AJM-1::GFP, used here as a marker of seam cell boundaries (data not shown and Figure 5A–5D). This may be predictive of inappropriate fusion with the hyp7 syncytium. Cells were never observed that lost ajm-1::gfp but retained scm::gfp. The ‘disintegration’ of the AJM-1::GFP continues until only vestigial traces are evident between the two seam cells on either side of the affected cell, as shown in Figure 5D.
In the hours subsequent to losing ajm-1::gfp expression, we observed that the seam cell tends to become rounder and move out of the line of the seam, with both the morphological and positional changes being indicative of the acquisition of hypodermal fate. To test this, a strain carrying the integrated transgenes ajm-1::gfp and dpy-7::yfp was used. In several cases, cells were observed during lineage analysis losing ajm-1::gfp and then acquiring dpy-7::yfp expression. Gaps in ajm-1::gfp expression in the seam were correlated with the presence of dpy-7::yfp-expressing nuclei which had not yet, or only partially, moved out of the line of the seam (Figure 5E). Thus, elt-1 RNAi causes an additional phenotype not observed in rnt-1 or bro-1 mutants, whereby some seam cells differentiate inappropriately by fusing with the hypodermal syncytium.
Overall, our data suggests three distinct phenotypes observed in elt-1(RNAi) animals. The first, in common with rnt-1 and bro-1 mutants, involves division failure in the absence of a permanent change in cell fate (i.e. not involving fusion with the hypodermis). Secondly, loss of SCM::GFP is also observed in elt-1(RNAi) animals (as well as in rnt-1 and bro-1 mutants), but we found this to be independent of division failure. Thirdly, in elt-1(RNAi) animals but not in rnt-1/bro-1 mutants, we observed inappropriate adoption of the hypodermal fate as shown by the acquisition of DPY-7::YFP, preceded by loss of scm::gfp expression and of AJM-1::GFP from the apical junctions. While these cells always lose SCM::GFP, fusion with the hypodermis is not always the consequence of SCM::GFP loss. However, AJM-1::GFP is very tightly linked to the seam fate; loss of AJM-1::GFP is always coupled with fusion to the hyp7 syncytium and acquisition of the hypodermal fate.
If cells are lost from the seam in elt-1(RNAi) animals because of inappropriate fusion with the hyp7 syncytium, we would anticipate that ectopic eff-1 expression would be evident, as the fusogen EFF-1 is known to be required for this fusion event [30]. To test this, a transgenic strain carrying both an eff-1 transcriptional reporter, which normally expresses in dorsal and ventral hypodermis but not in seam cells, and ajm-1::mCherry was used. In elt-1(RNAi) animals, we observed ectopic eff-1p::gfp expression in seam cells that had lost their AJM-1 boundary (Figure 6A). Furthermore, we found that elt-1 RNAi induced fusion of seam cells with the hyp7 syncytium was suppressed in eff-1 mutants (Figure 5A, 5C, 5D and Figure 6C). Taken together, these data suggest that ELT-1 represses eff-1 expression in seam cells in order to prevent fusion of these cells with the surrounding hypodermal syncytium, thus maintaining their distinct stem cell-like fate.
In order to address more directly the role of seam cell boundaries in determining the stem cell-like identity of seam cells, we knocked down the expression of components of apical junctions by RNAi. Given the essential functions of these proteins in a variety of cell types throughout development, it was necessary to use a seam-specific RNAi approach [31]. Knockdown of either ajm-1, let-413 or dlg-1 gave a similar phenotype, involving breaks in the AJM-1::GFP boundary, associated loss of scm::gfp expression and withdrawal from further division (Figure 6D; let-413 and dlg-1 data not shown). When ajm-1 was knocked down by RNAi, 66% of animals displayed breaks in the seam boundary as marked by ajm-1::mCherry (n = 50), whereas in controls not subjected to RNAi 5% of animals displayed breaks (n = 36) (the odd break in the seam would be expected due to mosaicism of the ajm-1::mCherry scm::gfp array). Therefore, seam cell membrane integrity, as marked by apical junctions, is essential for the maintenance of seam stem cell-like fate.
Next, we wanted to examine whether the presence of an intact boundary around the seam cells is sufficient to specify seam fate. This could happen in two ways: the boundary could either promote seam fate, or block differentiation signals from the surrounding environment. To test this, we used an eff-1 mutant in which anterior daughters retain their AJM-1::GFP marked boundary and remain in contact with the seam line instead of moving into the hypodermis. In other words, ectopic AJM-1::GFP boundaries are present in this strain. To assess whether these cells inappropriately retain the seam fate we monitored scm::gfp expression, as well as dpy-7::yfp expression. Firstly, we found that ectopic AJM-1::GFP bordered cells frequently do not express scm::gfp (Figure 6B), suggesting that these cells do not retain all aspects of seam fate. Perhaps this is not surprising, given that these cells are the anterior daughters of asymmetric seam divisions and would have already received the instruction to withdraw from further proliferation. However, we never observed dpy-7::yfp expression in these cells (Figure 6E), indicating that differentiation is repressed. This suggests that fusion of seam cells with hyp7 (mediated by EFF-1) is essential as a trigger for differentiation in this context. These cells are therefore in developmental ‘limbo’, being neither completely seam nor hypodermis. Thus, although the anterior daughters of asymmetric seam divisions are destined to differentiate at division, the nature of the differentiation event is not specified at this stage and requires further inputs; it is not until cell fusion and membrane breakdown occurs, as marked by the loss of apical junction boundaries, that the hypodermal fate is adopted.
Overall, therefore, seam cell boundaries may not be sufficient to specify all aspects of seam fate, but they are sufficient to prevent differentiation. Our data therefore suggest that the seam stem cell-like fate is retained by the blocking of “hypodermalizing” signals in cells that are bounded by apical junctions. ELT-1, therefore, plays dual roles in specifying the stem cell-like properties of the seam cells; on the one hand activating proliferative potential via bro-1, and on the other preventing inappropriate differentiation through the repression of eff-1 and consequent maintenance of seam boundaries (Figure 7).
C. elegans seam cells are an excellent model for stem cell-like modes of division, sitting as they do at the crossroads between the developmental decision to proliferate (thus renewing the pool of pluripotent precursors), or to pursue a differentiation pathway towards one or more specialised cell types. We used comparative genomics coupled with a yeast one-hybrid screen and promoter deletion analysis to define upstream regulators of bro-1 expression; a gene known to be essential for seam cell proliferation. Surprisingly, we found that sequences necessary and sufficient for the expression of bro-1 lie exclusively within an intron. Our evidence adds strength to the concept that highly conserved non-coding sequences correlate with biologically important enhancer elements [32], [33]. Multiple lines of evidence suggest that ELT-1 binds to this intronic sequence; in a yeast one-hybrid screen using the CNE as bait, ELT-1 was identified in 92% of positive clones, whilst an EMSA band shift experiment not only confirmed this finding, but demonstrated that the putative GATA site towards the 3′ end of the CNE (GATA site B) is essential for ELT-1 binding, an observation confirmed by rescue experiments.
Taken together, evidence from our yeast one-hybrid screen, in vitro binding assay, rescue experiments and quantitative RT-PCR experiments suggest that ELT-1 is a direct upstream transcriptional regulator of bro-1. This model is supported by the similarities between elt-1 (RNAi) and bro-1 mutant phenotypes in terms of failures of seam cell divisions, and loss of scm::gfp expression. It therefore seems highly likely that elt-1 and bro-1 (as well as rnt-1) function within the same developmental pathway to promote seam cell proliferation. It is interesting to note that the L1 (asymmetric) division is robust and appears unaffected in both elt-1 RNAi and bro-1/rnt-1 mutant worms. This division is most likely regulated by a separate pathway. The involvement of both the CBFβ/RUNX complex and a GATA transcription factor in the regulation of the proliferation of the stem cell-like seam cells is reminiscent of the situation in Drosophila and in mammals, where GATA and RUNX factors work together to regulate blood cell formation [34]. Indeed, mammalian Runx1 is known to be transcriptionally activated by Gata2 [22]. Thus, the direct regulatory relationship between ELT-1 and bro-1 suggests yet another mode of interaction between members of these two gene families.
elt-1(RNAi) worms display a striking phenotype which is not observed in bro-1 mutant animals, however. This involves loss of the apical cell junction marker ajm-1::gfp from some seam cells, followed by movement of the cell out of the seam and subsequent differentiation into the hypodermal fate. It has previously been argued that seam cell phenotypes (loss of scm::gfp positive cells and gaps in ajm-1::gfp expression) in elt-1(RNAi) worms do not result from inappropriate fusion of the seam with the hypodermis, on the basis that scm::gfp-expressing cells are never observed in the hypodermal syncytium [27]. However, here we argue that cells do indeed fuse with hyp7 in elt-1(RNAi) animals, but first lose their scm::gfp and AJM-1 boundary, as well as changing their morphology as they undergo the transition from seam to hypodermal fate and begin to express hypodermal markers. Thus, we suggest that the breaks in AJM-1::GFP result not from degeneration of the seam cells but are the result of a transition between two cell fates; the proliferative seam fate, which is associated with the AJM-1::GFP boundary, and the differentiated syncytial hyp7 fate, which is not.
In terms of its bro-1-independent role in the seam, elt-1 appears to function upstream of EFF-1. The role of the fusogen eff-1 is critical in the seam [30] and is responsible for promoting the fusion of hypodermal seam daughters with the hyp7 syncytium by causing the formation of pores in the membrane [35]. Indeed, eff-1 over-expression has been shown to result in inappropriate fusion of seam cells with hyp7 [36]. This phenotype is strikingly similar to that seen in elt-1(RNAi) animals, and suggests that ELT-1 acts to repress eff-1 in the seam, thereby preventing seam cells from fusing with hyp7. Our finding that ectopic eff-1 expression is observed in elt-1(RNAi) animals confirms this.
Inappropriate fusion of the seam cells with the hypodermis has been reported previously; this phenomenon has been observed in embryos and newly hatched animals in which elt-5 and elt-6 (which act redundantly) had been knocked down by RNAi [29]. Moreover, similar to the observations described here for elt-1(RNAi) animals, fusion in the case of elt-5/elt-6 RNAi was accompanied by gradual dissolution of the AJM-1::GFP boundary around the seam cells [29]. We therefore suggest that there is a network of GATA factors acting to prevent inappropriate differentiation of seam cells throughout development. In addition, other transcription factors have been shown to regulate seam cell development, for example NHR-25, BAF-1 and CEH-16 as well as heterochronic regulators like LIN-14 and LIN-29 [14], [15], [28], [37]–[39]. The interactions between all these genes will be an interesting area for future study.
The apparent close relationship between the presence of intact apical junctions and the stem cell-like properties of seam cells is suggestive of similarities with stem cells in Drosophila. In the testes and ovaries of Drosophila, germline stem cells (GSCs) are retained in what has been termed a niche; the niche concept, introduced over 30 years ago [40], describes how stem cells can be maintained in a proliferative state by signals from a microenvironment, consisting of cells and the extracellular components they produce. The niche is essential for the stem properties of such cells and, as these cells move out of the niche, so they lose these properties in favour of differentiated cell fates. In this way, far from merely creating an inert environment in which the Drosophila GSCs reside, the cells around these stem cells provide cues which regulate the maintenance of the stem cell pool, physically anchor the GSCs to the niche, and even control the polarity of the stem cells, determining the positions of the daughters of GSC divisions relative to one another and to the niche. The mechanisms underlying the complex interactions between GSCs and their niche microenvironment involves extracellular signalling [41]–[45] as well as physical adhesion of stem cells to the niche [46], with the DE-cadherin and Armadillo/β-catenin apical junction complex being both important in recruiting GSCs to the niche and required for the maintenance of the stem cell pool. Loss of either of these proteins results in dramatic depletion of GSCs from the niche. Here, we show that C. elegans apical junction proteins are required to maintain the undifferentiated stem cell-like fate of the seam cells.
Perhaps there is something analogous to a seam stem cell “niche” in C. elegans, in the sense that cell contacts (marked by, and dependent on, apical junction proteins) are required to maintain a microenvironment in which the seam cells are prevented from differentiation. The importance of cellular contacts for seam development has previously been recognized. For example, the developmental fate of the V5 seam cell has been shown to be dependent on correct seam cell contacts either side [47], and proliferation of the seam cells has been shown to be perturbed when contacts between seam cells are not properly re-established following division [15]. Thus, both the proliferation and differentiation of the seam cells has been shown to be dependent on signals from the surrounding microenvironment, raising the question of whether they do in fact reside in a niche. In support of the niche concept, we find that cells that have withdrawn from the proliferation programme as a result of asymmetric division, but which have failed to fuse with the hyp7 syncytium (as a result of eff-1 mutation), do not express markers of differentiation like dpy-7. In other words, the boundary provides protection from differentiation signals. However, this notion of a niche has to remain speculative in the absence of defining the nature of these signals.
Intriguingly, the regulation we have discussed in seam cells, in which ELT-1 represses eff-1 in order to prevent fusion and differentiation of cells that have the proliferative fate, mirrors the situation in vulval precursor cells (VPCs). In the developing vulva, the 6 VPCs P3.p–P8.p are prevented from fusing with the hyp7 syncytium in early larval stages (in L3 this exclusion from hyp7 is limited to P5.p–P7.p) [48]. Fusion with hyp7 acts to limit the developmental potential of Pn.p cells (as it does with anterior seam daughters) and is restricted to those that flank the developing vulva [48]. LIN-39 acts to prevent this fusion by repressing eff-1 in P3.p–P8.p during L1 and in P5.p–P7.p during L3 [49]–[51]. In eff-1 mutants, unfused VPCs fail to differentiate into the hypodermal fate, retaining their AJM-1 boundary and at least some aspects of the vulval fate. These cells fail to proliferate, however, suggesting that other signals are required for the induction of normal vulval development. This is analogous to the situation with unfused seam cells in eff-1 mutants, which also fail to divide once they leave the seam line, remaining in developmental “limbo”. In both cases, however, fusion with hyp7 and associated breakdown of cell boundaries is required for cells to take on the differentiated hypodermal fate. In both the seam and the developing vulva, therefore, only those cells that are prevented from fusing with hyp7 (thereby retaining their boundaries) are protected from differentiation and retain further developmental and proliferative potential.
Overall, we present a model (Figure 7) in which the GATA factor ELT-1 plays important dual roles in maintaining the ‘stemness’ of the seam cells, by both promoting the proliferative fate and preventing differentiation. Firstly, ELT-1 acts directly through bro-1 to promote proliferation and self-renewal of the seam. Secondly, ELT-1 is essential for maintaining the integrity of the seam cell compartment, as marked by apical junctions. In fulfilling this latter role, ELT-1 works through EFF-1. When eff-1 is repressed, the boundaries around the seam cells are maintained, and thus differentiation is prevented. We also find that apical junction components themselves are important for maintaining seam cell fate, but are not arguing that ELT-1 acts directly on apical junction components. Indeed, as has been previously suggested, apical junction breakdown could be a relatively late event in the cell fusion process [35], [52]. Taken together, these data suggest that the seam cells reside in a microenvironment in which they are protected from differentiation by the boundary that separates them from the hyp7 syncytium. Thus, the seam microenvironment may satisfy the criteria of a niche in certain respects, protecting seam cells from influences that would otherwise trigger differentiation. The GATA factor ELT-1 works through bro-1 to promote seam cell proliferation and through eff-1 to maintain seam cells in the undifferentiated state.
All strains used were derived from the wild type N2 Bristol strain. Manipulations and maintenance of strains were performed as previously described [53]. Strains used are described in Table S1.
Lineage analysis was performed as previously described [12]. For lineage analysis of elt-1(RNAi) animals, 3 µl of a freshly prepared solution of M9 and HT115 E. coli cells expressing the elt-1 dsRNA (scraped from a 2 mM IPTG NGM plate, on which they had been growing at 20°C for several days) was placed on the pad. For each slide, a single worm was transferred into this drop with an eyelash pick. A coverslip was then slowly lowered on top of the worm, and microscopy performed with Nomarski (DIC) optics and a 100× oil immersion objective (Zeiss). Photomicrographs were taken using a 63× Zeiss oil immersion objective and Axiovision software (Release 4.5).
elt-1 knockdown by RNAi was performed as described previously [27], using an identical feeding construct to pPM88, named pAW565. The seam specific RNAi strategy is described in Text S1 and all other RNAi was performed by feeding as previously described [54].
Plasmids used in transgenic strains are described in Table S1 and detailed cloning strategies described in Text S1.
Injections were performed as described previously [55] using the unc-119+ (pDP#MM016β) transformation marker [56]. Constructs were injected at 10–20 ng/µl.
cDNAs of elt-1, elt-2, elt-3, elt-5 and elt-6 were amplified from a mixed stage cDNA preparation using Phusion polymerase (Finnzymes). The PCR products were cloned into the pCR®-XL-TOPO vector (Invitrogen) and the TNT® Quick Coupled Transcription/Translation kit (Promega) was then used for in vitro transcription and translation. To make the labelled probes, oligonucleotides covering ‘GATA site B’ were synthesised (WT probes: CB204 gatccgacaagattacaatccacat, CB206 atgtggattgtaatcttgtcggatc; mutant probes: CB205 gatccgacaatagtacaatccacat; CB207 atgtggattgtactattgtcggatc), annealed by heating and gradual cooling, and labelled with [γ-32P] dATP (for hot WT probes) or without [γ -32P] dATP (for cold, competitor probes). The DNA binding reaction was carried out on ice for 30 minutes before the reaction mixture was loaded onto a 7% non-denaturing polyacrylamide gel and run at 4°C in 0.5×TBE.
The entire 122 bp bro-1 CNE was used as bait in the yeast one-hybrid screen; three copies were inserted in the forward direction into the Clontech Matchmaker vectors pHisi-1 and pLacZi and integrated into yeast strain YM4271 [57] as described in Text S1. YM4271 [pbro-1HIS; pbro-1LAC] was then transformed with a mixed stage transcription factor cDNA library and plated onto –his, -leu, -ura 15 mM 3-AT SD plates. All colonies that grew within 3 days were assayed for lacZ expression [58], and after re-isolating plasmids and re-checking, positive clones were sequenced.
qRT-PCR was performed on synchronised worms obtained by bleaching gravid animals and seeding the eggs onto elt-1 RNAi or L4440 control plates. Larvae were harvested after 1–2 days by washing with M9. RNA was extracted by the hot phenol method [59] and mRNA levels of elt-1/bro-1 and a normaliser (nuo-2, expressed in all seam cells) were assessed using SYBR-Green and a Qiagen Rotor-Gene Q machine. Expression levels were assayed by the 2−ΔΔCT method [60].
|
10.1371/journal.pntd.0002997 | Identification of Giardia lamblia DHHC Proteins and the Role of Protein S-palmitoylation in the Encystation Process | Protein S-palmitoylation, a hydrophobic post-translational modification, is performed by protein acyltransferases that have a common DHHC Cys-rich domain (DHHC proteins), and provides a regulatory switch for protein membrane association. In this work, we analyzed the presence of DHHC proteins in the protozoa parasite Giardia lamblia and the function of the reversible S-palmitoylation of proteins during parasite differentiation into cyst. Two specific events were observed: encysting cells displayed a larger amount of palmitoylated proteins, and parasites treated with palmitoylation inhibitors produced a reduced number of mature cysts. With bioinformatics tools, we found nine DHHC proteins, potential protein acyltransferases, in the Giardia proteome. These proteins displayed a conserved structure when compared to different organisms and are distributed in different monophyletic clades. Although all Giardia DHHC proteins were found to be present in trophozoites and encysting cells, these proteins showed a different intracellular localization in trophozoites and seemed to be differently involved in the encystation process when they were overexpressed. dhhc transgenic parasites showed a different pattern of cyst wall protein expression and yielded different amounts of mature cysts when they were induced to encyst. Our findings disclosed some important issues regarding the role of DHHC proteins and palmitoylation during Giardia encystation.
| Giardiasis is a major cause of non-viral/non-bacterial diarrheal disease worldwide and has been included within the WHO Neglected Disease Initiative since 2004. Infection begins with the ingestion of Giardia lamblia in cyst form, which, after exposure to gastric acid in the host stomach and proteases in the duodenum, gives rise to trophozoites. The inverse process is called encystation and begins when the trophozoites migrate to the lower part of the small intestine where they receive signals that trigger synthesis of the components of the cyst wall. The cyst form enables the parasite to survive in the environment, infect a new host and evade the immune response. In this work, we explored the role of protein S-palmitoylation, a unique reversible post-translational modification, during Giardia encystation, because de novo generation of endomembrane compartments, protein sorting and vesicle fusion occur in this process. Our findings may contribute to the design of therapeutic agents against this important human pathogen.
| The flagellated protozoan parasite Giardia lamblia is a major cause of non-viral/non-bacterial diarrheal disease worldwide. This parasite can cause asymptomatic colonization or acute or chronic diarrheal illness and malabsorption [1]. Infection begins with the ingestion of Giardia in its cyst form which, after exposure to gastric acid in the host stomach and proteases in the duodenum, gives rise to trophozoites. The inverse process is called encystation and begins when the trophozoites migrate to the lower part of the small intestine where they receive signals that trigger synthesis of the components of the cyst wall. The encystation process is tightly regulated but the exact mechanism that controls this process is still obscure. Expression of the three Cyst Wall Proteins (CWP) and the glycopolymer biosynthetic enzymes, is largely upregulated. In addition, several other proteins, whose roles in encystation are yet to be discovered, are upregulated at the transcriptional level [2], [3]. Various protein posttranslational modifications (PTM) have been implicated in the development of encystation, such as phosphorylation [4] and deacetylation [5], among others [6], [7], [8]. There is also some evidence of the role of PTM in gene regulation for the control of this process [9].
Protein S-palmitoylation (hereafter referred to as palmitoylation), the post-translational addition of palmitic acid (16∶0) to cysteine residues of proteins, is a PTM essential for proper membrane trafficking to defined intracellular membranes or membrane sub-domains, protein stability, protein turnover, and vesicle fusion [10], [11], [12]. Unlike the other lipid modifications, palmitoylation is potentially reversible, providing a regulatory switch for membrane association [13], [14]. Palmitoylation is catalyzed by a family of protein acyltransferases (PATs), which transfer a palmitoyl moiety derived from palmitoyl-CoA to a free thiol of a substrate protein to create a labile thioester linkage [15], [16]. The discovery of these enzymes came through studies in yeast that identified the PATs Erf2 and Akr1, which are active against Ras and casein kinase, respectively [17], [16]. These enzymes are polytopic integral membrane proteins which share the conserved Asp-His-His-Cys (DHHC) - cysteine-rich domain (CRD). The general membrane topology predictions indicate that the core structure of a PAT is four transmembrane domains (TMDs), with the N- and C- terminus in the cytoplasm [18]. The signature feature DHHC-CRD, which is indispensable for palmitoylating activity, is located in the cytoplasmic loop between the second and third TMDs [19]. There is a small group of PATs that display six TMDs with an extended N-terminal region encoding ankyrin repeats. The yeast PAT called Akr1 is a member of this group [16], [20]. All these findings were crucial in defining palmitoylation as an enzymatic process and led to subsequent identification of protein acyltransferases in many other organisms, such as mammals [21], [22], plants [23], and protozoan parasites like Toxoplasma gondii [24], [25], Plasmodium [26], [25], and Trypanosoma brucei [27].
There is scarce knowledge about palmitoylation in Giardia, but some findings indicate that this PTM may play an important role in pathogenesis. It was shown that α19-giardin, one of the major protein components of the Giardia cytoskeleton, can be both myristoylated and palmitoylated [28] and that the variant-specific surface proteins (VSPs) may be palmitoylated within their C-terminal domains [29], [30]. Later, Touz et al. determined the exact site of palmitoylation of the VSPs, characterized the enzyme responsible for this modification, and determined the participation of palmitoylation during antigenic variation [31], a process in which the trophozoite continuously changes its surface antigen coat [32]. Antigenic variation and encystation are two distinctive mechanisms of defense that the parasite has developed to survive in hostile environmental conditions during its life cycle, and it has been suggested that both are mechanistically related processes [33].
Accumulation of material in membrane vesicles followed by transport and vesicle fusion and secretion are some of the main events involved in Giardia encystation. Because palmitoylation has been reported to play a key role in these events in other cell types [12], [10], [34], [35], [36], it is likely that this PTM may also play a role in Giardia encystation. In this work, we address the question of whether PATs and palmitoylation itself are involved in Giardia encystation. We provide evidence about the role of palmitoylation in Giardia encystation biology by inhibiting this PTM with 2-bromopalmitate (2-BP) or 2-fluoropalmitate (2-FP). Using bioinformatics, we identified the potential PATs (hereafter called DHHC proteins) in the Giardia lamblia proteome and performed a phylogenetic analysis of these proteins. We evaluated the expression of the total collection of DHHC proteins in trophozoites and encysting parasites. Using dhhc transgenic Giardia parasites, we revealed the intracellular localization of DHHC proteins and their influence in CWP expression and cyst yield when parasites were induced to encyst. Our data suggest a role of palmitoylation and DHHC proteins in encystation, providing an insight into the impact of this PTM in Giardia survival.
Trophozoites of the isolate WB, clone 1267 [37], were cultured in TYI-S-33 medium supplemented with 10% adult bovine serum and 0.5 mg ml−1 bovine bile (Sigma, St. Louis, MO) as described [38]. GL50806_40376 (High Cysteine Non-variant Cyst protein; HCNCp), GL50803_1908, GL50803_2116, GL50803_16928, and GL50803_8711 open reading frames (ORF) were amplified from genomic DNA. GL50806_40376 was cloned into the vector pTubV5-pac [39] to generate pHCNCp-V5 plasmid. GL50803_1908, GL50803_2116, GL50803_16928, and GL50803_8711 were each one cloned into the vector pTubHA-pac [39] to generate the pDHHC-HA plasmids. Trophozoites were transfected with the constructs by electroporation and selected by puromycin (Invivogen, San Diego, CA) as previously described [40], [41], [42]. Trophozoites transfected with empty pTubHA-pac or pTubV5-pac plasmids were used as control. Primer sequences used for DHHC proteins cloning are depicted in table S1. Encystation was induced by growing trophozoites for one culture cycle in TYI-S-33 medium without bile (pre-encystation). Bile-deficient medium was poured off along with unattached trophozoites and replaced with warmed encysting medium containing 0.45 mg ml−1 porcine bile (Sigma, St. Louis, MO) and 0.25 mg ml−1 lactic acid (Sigma, St. Louis, MO), pH 7.8, and incubated at 37°C for 48 h [43]. Total encysting cultures were harvested at 48 h by chilling and centrifugation, and subsequently used for palmitoylation assay, RNA extraction, western blot, immunofluorescence, or flow cytometry.
The assay followed the procedure described by Papanastasiou et al. and Corvi et al. [29], [44]. Briefly, 8×106 growing and encysting wild-type or dhhc transgenic parasites were washed, suspended in 1 ml of RPMI (Gibco, Invitrogen, Carlsbad, CA) containing 200 µCi of [9,10-3H(N)]-palmitic acid (Perkin-Elmer, MA), previously conjugated to BSA fatty acid free (1∶1, mol∶mol ratio), and incubated for 4 h at 37°C. The samples were then suspended on SDS–PAGE loading buffer without any reducing agent and loaded onto SDS-PAGE gel. The gel was then incubated for 30 min in ddH2O and for 30 min more in 1M sodium salicylate pH 6.5. The gel was then incubated with 3% glycerol, 10% acetic acid, and 40% methanol for 30 min, dried for 2 h at 80°C using a gel dryer machine, and exposed to autoradiographic film for a month. For hydroxylamine treatment, the gel was soaked in either 1 M NH2OH- NaOH pH 7.0 or 1 M Tris-HCl pH 7.0 (Control) for 48 h. Finally, the gel was incubated for 30 min in ddH2O and for 30 min more in 1M sodium salicylate pH 6.5, dried as described above, and exposed to autoradiographic film for a month.
Total cellular palmitoylated proteins from growing and encysting wild-type or transgenic (overexpressing HCNCp) parasites, were purified following the procedure described by Wan et al. [45]. Briefly, 5×107 trophozoites or 48 h encysting parasites were harvested and lysed with Lysis buffer (LB; 50 mM Tris-HCl pH 7.4, 5 mM EDTA, 150 mM NaCl) with 10 mM N-Ethylmaleimide (NEM; Thermo Scientific Pierce Rockford, IL) plus protease inhibitors. After sonication, 1.7% of Triton X-100 was added to each sample and incubated for 1 h at 4°C under shacking. The samples were then centrifuged at 500× g for 5 min at 4°C. The supernatant was collected in a new tube and solubilized proteins were precipitated with chloroform-methanol. Proteins were resolubilized in 4% SDS buffer (SB; 4% SDS, 50 mM Tris-HCl pH 7.4, 5 mM EDTA) with 10 mM NEM by incubating at 37°C under shacking. Each sample was then diluted with 3 vol of LB with 1 mM NEM, protease inhibitors, and 0.2% Triton X-100 and incubated overnight at 4°C under shacking. Proteins were then precipitated by three sequential chloroform-methanol extractions after which each sample was dissolved in SB and split into two equal fractions: one for neutral pH hydroxylamine treatment (hyd+) and the other for neutral pH Tris buffer treatment (hyd−). The hyd+ portion was diluted with 4 vol of hyd+ buffer (1M hydroxylamine pH 7.4, 150 mM NaCl, 1 mM HPDP-Biotin, 0.2% Triton X-100, protease inhibitors), and the hyd- portion with 4 vol of the hyd- buffer (50 mM Tris-HCl pH 7.4, 5 mM EDTA, 150 mM NaCl, 1 mM HPDP-Biotin (Thermo Scientific Pierce, Rockford, IL), 0.2% Triton-X-100, protease inhibitors) and incubated for 1 h at room temperature under shacking, followed by chloroform-methanol precipitation. The samples were then resuspended in SB at 37°C under shacking. Protein pellets were solubilized in LB containing 0.2% Triton X-100. Streptavidin-agarose (Thermo Scientific Pierce, Rockford, IL) was added at concentration of 25 µl beads ml−1 and the lysate and samples were incubated for 1 h at room temperature. Unbound proteins were removed by four sequential washes with LB containing 0.2% Triton X-100. Samples were finally eluted with 100 mM DTT containing 0.2% Triton X-100. Each eluate was then analyzed by Western blotting.
Giardia trophozoites were cultured as described above. 2-bromopalmitate (2-BP) (Sigma-Aldrich, St. Louis, MO) or 2-fluoropalmitate (2-FP) (Cayman Chemical, Ann Arbor, MI) were added to the media for 48 h to reach a final concentration of 10, 20, 40, 50, 75 or 100 µM for 2-BP, and 100, 150 or 200 µM for 2-FP. The inhibitors were diluted in DMSO (Sigma-Aldrich, St. Louis, MO) following manufacturer indications. The parasites were then analyzed by staining them with Trypan blue to distinguish live from dead cells and by counting them in a Neubauer chamber. To perform a growth curve, parasites from three independent experiments were counted. Parasites were induced to encyst as described above. 2-BP or 2-FP were added with encysting media for 48 h to reach a final concentration of 10, 20 or 40 µM for 2-BP, and 100 µM for 2-FP. The inhibitors were diluted in DMSO as mentioned above. For immunofluorescence the parasites were subcultured onto 12 mm round glass coverslips (Glaswarenfabrik Karl Hecht, Sondhein, Germany) in 24-well culture plates for 1 h, fixed with 4% paraformaldehyde in PBS for 20 min at 4°C, washed twice in PBS and blocked with 10% normal goat serum (Invitrogen, Carlsbad, CA) in 0.1% Triton X-100 in PBS for 30 min at 37°C. The samples were then incubated with FITC labeled anti-CWP1 mAb (Waterborne Inc., New Orleans, LA) diluted 1∶250 in PBS containing 3% normal goat serum and 0.1% Triton X-100 for 1 h at 37°C or anti-CWP1 mAb and DAPI diluted in PBS (dilution 1∶500) (Sigma, St. Louis, MO). The coverslips were then mounted onto glass slides using FluorSave reagent (Calbiochem, La Jolla, CA). Fluorescence was visualized in a Zeiss Axiovert 200 microscope (Carl Zeiss, Jena, Germany). To quantify the percentage of encysting parasites, 55 cells from three separate experiments were counted and classified as encysting I, encysting II, or cyst according to the cell shape, membrane staining, and number and size of the encystation-specific vesicles. The average was taken in each of the three groups.
A proteome database was constructed gathering complete proteomes for 25 Metazoa (Amphimedon queenslandica (aqu), Anolis carolinensis (aca), Apis mellifera (apm), Bombyx mori (bmo), Caenorhabditis elegans (cae), Canis familiaris (cfa), Ciona intestinalis (cin), Danio rerio (dre), Daphnia pulex (dpu), Drosophila melanogaster (dme), Equus caballus (eqc), Felis catus (fca), Gallus gallus (gga), Gorilla gorilla (ggo), Homo sapiens (hsa), Ixodes scapularis (ixs), Mus musculus (mmu), Nematostella vectensis (nve), Ornithorhynchus anatinus (oan), Petromyzon marinus (pma), Pteropus vampyrus (pva), Rattus norvegicus (rno), Schistosoma mansoni (sma), Sus scrofa (ssc) and Xenopus tropicalis (xtr)), 18 Fungi (Aspergillus nidulans (and), Batrachochytrium dendrobatidis (bde), Botryotinia fuckeliana (bfu), Candida albicans (clb), Encephalitozoon cuniculi (ecu), Gibberella zeae (gze), Leptosphaeria maculans (lem), Nematocida sp (nsp), Neurospora crassa (ncr), Pichia pastoris (ppa), Puccinia graminis (pug), Saccharomyces cerevisiae (sce), Schizosaccharomyces pombe (szp), Sclerotinia sclerotiorum (scl), Tuber melanosporum (tme), Ustilago maydis (uma), Vittaforma corneae (vco) and Yarrowia lipolytica (yli)), 12 Plants (Arabidopsis thaliana (ath), Brachypodium distachyon (bdi), Glycine max (gmx), Medicago truncatula (met), Oryza sativa (osa), Physcomitrella patens (php), Populus trichocarpa (pot), Selaginella moellendorffii (smo), Solanum lycopersicum (sly), Solanum tuberosum (stu), Sorghum bicolor (sbi) and Vitis vinifera (vvi)), 1 Brown alga (Aureococcus anophagefferens (aan)), 1 Red alga (Cyanidioschyzon merolae (cym)), 3 Green algae (Ostreococcus taurii (ota), Chlamydomonas reinhardtii (chr) and Chlorella variabilis (chv)), and 24 Protists (Babesia bovis (bbo), Bigelowiella natans (bna), Chlamydomonas reinhardtii (chr), Chlorella sp (chl), Cryptosporidium parvum (cpv), Dictyostelium discoideum (ddi), Entamoeba histolytica (ehi), Giardia lamblia (gla), Guillardia theta (gth), Leishmania major (lma), Paramecium tetraurelia (pat), Perkinsus marinus (pem), Phaeodactylum tricornutum (pht), Phytophthora capsici (pcs), Phytophthora ramorum (pra), Plasmodium falciparum (pfa), Polysphondylium pallidum (pop), Tetrahymena thermophila (tet), Thalassiosira pseudonana (thp), Theileria parva (thp), Toxoplasma gondii (tgo), Trichomonas vaginalis (tva), Trypanosoma brucei (trb) and Trypanosoma cruzi (tcz)) from Ensembl, the Joint Genome Institute (JGI) and the NCBI databanks. zf-DHHC HMMer profile was obtained from Pfam [46], and used to search the proteomes database [47]. Incomplete sequences or those that did not start with the M residue were deleted from the dataset. Also, 90% similar amino acid sequences were clustered using CD-HIT web server with default settings, to reduce the redundancy of the set [48]. The final dataset contained 1034 amino acid sequences. Multiple sequence alignment of DHHC-CRD amino acid sequences was carried out using PROMALS3D online server with default settings [49]. Following manual curation using GeneDoc software [50], sequences lacking conservation in the regions of interest (i.e., DPG, DHHC-CRD and TTxE) were removed. Block Mapping and Gathering with Entropy (BMGE) [51] was used to select columns suitable for phylogenetic inference with the following settings: m = BLOSUM30, g = 0.2, b = 4.
Phylogenetic analysis was performed by Maximum Likelihood (ML) using PhyML [52] with approximate likelihood-ratio test (aLRT), in combination with the LG+G amino acid replacement matrix, which was determined by ProtTest to be the model of protein evolution which best fit the data [53]. Phylogenetic trees were generated and edited with Itol [54].
RNA from WB1267 trophozoites or 48 h encysting WB1267 was extracted and purified using TRIzol reagent (Invitrogen, Carlsbad, CA) and SV total RNA Isolation System (Promega, Madison, WI). Total RNA were reverse transcribed using Revertaid reverse transcriptase according to the manufacturer's specifications (Fermentas, Thermo Scientific, PA). DNA contamination was tested by performing PCR in a “-RT” control (a mock reverse transcription containing all the RT-PCR reagents, except the reverse transcriptase. For PCR, 30 cycles (30 s at 94°C, 30 s at 55°C and 1 min at 72°C) were used ending with a final extension of 10 min at 72°C. The expression of the Giardia glutamate dehydrogenase (gdh) gene was assayed for positive control. Aliquots (50 µl) of the RT-PCR reaction were size-separated on 1% agarose gel prestained with SYBR Safe (Invitrogen, Carlsbad, CA). Primers sequences used in RT-PCR are displayed in table S2. These assays were performed four times in duplicates.
RNA from WB1267 trophozoites, 48 h encysting WB1267 or dhhc transgenic 48 h encysting cells (GL50803_1908, GL50803_2116, GL50803_16928, GL50803_8711) was extracted and purified as described above. 2 µg of total RNA were reverse transcribed using Revertaid reverse transcriptase according to the manufacturer's specifications (Fermentas, Thermo Scientific, PA). DNA contamination was tested as described above. cDNA samples were stored at −80°C until use. Control samples were prepared as above using nuclease-free ddH2O in place of RNA. Primers for PCR were designed using Primer express 3.0 software (Applied Biosystems, Forster City, CA) and were synthesized by Invitrogen, Inc. (Carlsbad, CA). Amplification was performed in a final volume of 20 µl, containing 2 µl of each cDNA sample which were previously diluted 1∶1000 (for dhhc genes) or 1∶10000 (for cwp genes), and 10 µl of SYBR Green Master Mix (Applied Biosystems, Foster City, CA). qRT-PCR was performed in a StepOne thermal cycler (Applied Biosystems, Foster City, CA). The mRNA levels of the genes studied were normalized to the expression of the Giardia glutamate dehydrogenase (gdh) gene. The relative-quantitative RT-PCR conditions were: holding stage: 95°C for 10 min, cycling stage: 40 cycles at 95°C for 15 s, 60°C for 1 min and melt curve stage: 95°C for 15 s, 60°C for 1 min, and 95°C for 15 s. Expression data were determined by using the comparative ΔΔCt method [55]. Primer sequences used in qRT-PCR are displayed in table S3.
For Western Blot assays, parasite lysates or purified palmitoylated proteins were incubated with 2× Laemmli buffer, boiled for 10 min, and separated in 10% Bis-Tris gels using a Mini Protean II electrophoresis unit (Bio-Rad). Samples were transferred to nitrocellulose membranes (GE Healthcare Biosciences, Pittsburgh, PA), blocked with 5% skimmed milk and 0.1% Tween 20 in PBS, and later incubated with anti-HA mAb or anti-V5 mAb (Sigma, St. Louis, MO; dilution 1∶1000 or 1∶50 respectively) diluted in the same buffer for 1 h. The membrane was then washed, incubated with IDRye 800CW conjugated goat anti-mouse Ab (LI-COR, Lincoln, NE; dilution 1∶10000) for 1 h, and analyzed on the Odyssey scanner (LI-COR, Lincoln, NE). For the analysis of VSPs expression, blockage was performed with 5% skimmed milk and 0.1% Tween 20 in TBS, and then incubated with 5C1 anti-VSP1267 mAb diluted in the same buffer for 1 h. After washing and incubation with an enzyme-conjugated secondary antibody, proteins were visualized with the SuperSignal West Pico Chemiluminescent Substrate (Pierce, Thermo Fisher Scientific Inc., Rockford, IL, USA) and autoradiography. Controls included the omission of the primary antibody, the use of an unrelated antibody, or assays using non-transfected cells.
For immunofluorescence assays (IFA), trophozoites or encysting cells cultured in growth medium or encysting medium, respectively, were harvested and washed two times with PBSm (1% growth medium in PBS, pH 7.4) and allowed to attach to multi-well slides in a humidified chamber at 37°C for 30 min. After fixation with 4% formaldehyde (Sigma, St. Louis, MO) in PBS for 40 min at room temperature, the cells were washed with PBS and blocked with 10% normal goat serum (Invitrogen, Carlsbad, CA) in 0.1% Triton X-100 in PBS for 30 min at 37°C. Cells were then incubated with the anti-HA mAb (Sigma, St. Louis, MO; dilution 1∶500) in PBS containing 3% normal goat serum and 0.1% Triton X-100 for 1 h at 37°C, followed by incubation with Alexa 546-conjugated goat anti-mouse (dilution 1∶500) secondary antibody at 37°C for 1 h. Encysting cells were also incubated with FITC-conjugated anti-CWP1 mAb (Waterborne Inc., New Orleans, LA; dilution 1∶250). Alternatively, cells were incubated with 9C3 anti-BiP mAb (marker for ER) [56] or 5D2 anti-AP2 mAb (marker for peripheral vacuoles) [57] in PBS containing 3% normal goat serum and 0.1% Triton X-100 for 1 h at 37°C, followed by incubation with Alexa 546-conjugated goat anti-mouse (dilution 1∶500) secondary antibody at 37°C for 1 h. Samples were then incubated with FITC-conjugated anti-HA mAb (Sigma, St. Louis, MO; dilution 1∶100). Preparations were stained with DAPI diluted in PBS (dilution 1∶500) (Sigma, St. Louis, MO). Finally, preparations were washed with PBS and mounted in Vectashield mounting medium (Vector Laboratories, Burlingame, CA). Fluorescence staining was visualized with a motorized FV1000 Olympus confocal microscope (Olympus UK Ltd, UK), using 63× or 100× oil immersion objectives (NA 1.32). The fluorochromes were excited using an argon laser at 488 nm and a helio-neon laser at 543 nm. Detector slits were configured to minimize any cross-talk between the channels. Differential interference contrast images were collected simultaneously with the fluorescence images, by the use of a transmitted light detector. Images were processed using Fiji software [58] and Adobe Photoshop 8.0 (Adobe Systems) software. The colocalization and deconvolution were also performed using Fiji.
For the analysis of the amount of cyst yield in dhhc transgenic trophozoites by flow cytometry, the parasites were induced to encyst for 48 h. Trophozoites, encysting cells, and cysts were collected from confluent cultures. Parasites were pelleted by centrifugation at 1455 g for 15 min at 4°C, resuspended in cool sterile ddH2O and placed at 4°C overnight. Mature water-resistant cysts were then processed following the protocol for immunofluorescence (see above) without permeabilization. Briefly, parasites were washed two times with PBSm (1% growth medium in PBS, pH 7.4). After blockade with 10% normal goat serum, the parasites were labeled with anti-CWP1 mAb (Waterborne Inc, New Orleans, LA; dilution 1∶250) diluted in PBSm for 1 hour at 4°C. Cells were then washed twice in PBS and fixed with 4% formaldehyde (Sigma, St. Louis, MO) in PBS for 40 min at room temperature. Unlabeled samples were used to determine background fluorescence, and subsequently, fluorescently labeled cysts were analyzed in triplicate on a FACSCanto II flow cytometer (Becton & Dickinson, New Jersey, NY). All samples were analyzed in parallel by IFA to assess encystation efficiency.
Results were analyzed for statistical significance (defined as p<0.05 and indicated by asterisks in figures) by performing unpaired, two-sided Student's t-test with GraphPad Prism 5 Data Analysis Software (GraphPad Software, Inc., La Jolla, CA). Mean and standard error of mean (SEM) values were calculated from at least three biologically and technically independent experiments.
It has been shown that protein palmitoylation actively participates in cell differentiation in a variety of cells [59], [60], [61]. The analysis of the expression of palmitoylated proteins, using metabolic labeling with [3H] palmitic acid, showed that encysting Giardia parasites displayed a different pattern of total protein palmitoylation than growing parasites (Figure 1A, T-ET/hyd−). The results showed a band of ∼60 kDa in trophozoites that may correspond to the expressed VSPs [31] (Figure 1A, T/hyd−). However, when Giardia encysting cells were analyzed, the assay displayed a larger amount of palmitoylated proteins, as can be judged by the larger number of bands displayed compared to trophozoites (Figure 1A, ET/hyd−). When we performed neutral treatment with hydroxylamine, almost complete removal of the attached palmitates was observed in both growing and encysting parasites (Figure 1A, T-ET/hyd+). This confirms that palmitate is attached through a labile thioester linkage (S-palmitoylation) in Giardia, as has been observed in other cell types including parasites [62], being most common among palmitoylated proteins [63]. Protein S-palmitoylation reversibility makes it a flexible, rapid and precise way of protein activity regulation [64] which may be crucial in the encystation process. The fact that the amount of total S-palmitoylated proteins was higher in encysting cells compared to trophozoites suggested that this PTM may play an important role during Giardia differentiation. This observation is in accordance with previous reports showing an important role of protein S-palmitoylation in controlling several crucial processes in parasites such as invasion or motility [44]. During Giardia encystation, the cyst wall proteins (CWPs) are sorted, concentrated within encystation-specific vesicles (ESVs), and exported to the nascent cyst wall [65], [66], [67]. Thus, the larger amount of palmitoylated proteins observed in encysting parasites (Figure 1A, ET/hyd−) may be explained by this additional requirement of protein sorting and export during this stage. In addition to the CWP1, 2 and 3, another type of cyst wall protein has been identified, a High Cysteine Non-variant Cyst protein (HCNCp) [68]. HCNCp belongs to a large group of cysteine-rich, non-VSPs, Type I integral membrane proteins (HCMp) [68]. The palmitoylation prediction algorithm CSS-Palm 3.0 [69] strongly predicts that HCNCp is palmitoylated at cysteines 1602 (CSS-Palm score 6.57, high stringency cut-off 0.31) and 1603 (CSS-Palm score 4.99, high stringency cut-off 0.31), which are located in the transmembrane region and in the cytosolic tail respectively (HMMTOP, (http://www.enzim.hu/hmmtop/) [70], [71]). In order to find out whether HCNCp is palmitoylated or not, we performed the following approach: first, we expressed full length HCNCp as a fusion protein containing a C-terminal V5-tag and a tubulin promoter [39]. The expression of the ∼169 kDa HCNCp protein was equally observed in hcncp-V5 transgenic growing and encysting parasites, together with fragments of 21, 42 and 66 kDa already reported by Davids et al. [68] (Figure S1). Second, hcncp-V5 transgenic trophozoites (HCNCp T) and encysting (HCNCp ET) parasites were subjected to acyl biotin exchange (ABE) as described in Methods. Parallel plus- and minus-hydroxilamine (hyd) samples were analyzed by Western blotting using an anti-V5 mAb (Figure 1B). Only the samples that were treated with hydroxylamine had free cysteine residues able to be detected by biotin/streptavidin (see Methods). When we assayed HCNCp T purified samples, we observed three bands (169, 66 and 21 kDa) and a weak band of 42 KDa (Figure 1B, HCNCp T/hyd+). Also, the four bands (169, 66, 42, and 21 kDa) were observed for HCNCp ET purified sample compared to the control (hyd−), showing that not only the full length but also the smaller epitope-tagged fragments of the HCNCp protein were palmitoylated in encysting parasites (Figure 1B, HCNCp ET/hyd+). The presence of these four bands may account, at least in part, for the bands shown in figure 1A (Figure 1A, ET/hyd−). Although we showed that the constitutively expressed HCNCp can be palmitoylated during growth and encystation, it was clearly reported that HCNCp is almost exclusively expressed during encystation when its expression was analyzed at the mRNA and protein (expression under its own promoter) levels [68]. Altogether, these results suggest that HCNCp is likely important during encystation, while the machinery necessary for its palmitoylation remains unaltered during growth and differentiation. Despite the need of additional assays to accurately identify additional palmitoylation substrates, it seems that this PTM is more frequently founded in encysting cells compared to trophozoites. In parallel to HCNCp T and HCNCp ET samples, we also performed ABE in wild-type trophozoites and encysting parasites and analyzed the purified samples by Western blotting using anti-VSP1267 mAb (Figure 1C). The results showed the specific protein band of VSP1267 (MW ∼60 KDa), in both growing and encysting parasites, suggesting that this PTM may be important for VSP function during the entire Giardia life cycle.
Further analysis using ABE or click chemistry [72] assays, together with different methods for Mass spectrometry-based proteomics, including Multidimensional protein identification technology [45], will expand our knowledge about other palmitoylated proteins in Giardia, defining the palmitoyl proteome of this parasite and shedding light on the role of this PTM in its life cycle.
The fact that Giardia encysting cells displayed a large amount of palmitoylated proteins prompted us to find out whether inhibition of protein palmitoylation would influence Giardia encystation. Several compounds have been reported to block protein palmitoylation [73]. The 2-bromopalmitate (2-BP) [74] and the 2-fluoropalmitate (2-FP) [73] inhibitors are non-metabolizable palmitate analogs that block palmitate incorporation into proteins using a still unclear mechanism. These two compounds have been widely used, act as broad inhibitors of palmitate incorporation and do not appear to selectively inhibit the palmitoylation of specific protein substrates. To test the effect of these inhibitors during encystation, Giardia wild-type trophozoites were induced to encyst together with the addition of either 2-BP or 2-FP. It has been reported that 2-BP is not well tolerated by in vitro cultured cells and causes cell death even after a brief exposure to 100 µM of 2-BP [75]. Thus, a growth curve was performed to determine the optimal concentrations that do not affect Giardia growth (10, 20 or 40 µM for 2-BP and 100 µM for 2-FP), observing that trophozoites died under concentrations higher than 50 µM of 2-BP or 150 µM of 2-FP (Figure 2A). After 48 h of encystation, treated or control parasites were harvested, permeabilized, stained with anti-CWP1 mAb and analyzed by fluorescence microscopy (Figure 2B). Wild-type encysting trophozoites were classified as encysting I (EI) (corresponding to 6 h of encystation [76]), encysting II (EII) (corresponding to 12 h of encystation [76]), and cysts (corresponding to 24–48 h of encystation [76]) (Figure 2B, upper panel), based on the following features: cell shape, membrane staining, and number and size of the ESVs. As shown in figure 2B (lower panel), there was a significant reduction in the amount of cysts when parasites were treated with 2-BP (20 µM or 40 µM) or 2-FP (100 µM).
The effect of 2-BP as a generic palmitoylation inhibitor has been reported in a wide variety of cells [77], [74], [78] including parasites like Toxoplasma gondii [62], although the concentrations used were much higher than the ones we used in this work. Interestingly, with 20 and 40 µM of 2-BP, there was an increase of the encysting II parasites compared to the control, reaching its highest levels when the concentration of 2-BP was 40 µM and resulting also in a diminution of encysting I cells (Figure 2B, lower panel). Thus, the decrease in the amount of cysts may be at the expense of the arrest of the cells at the encysting II stage of differentiation. In order to find out whether the treatment with palmitoylation inhibitors affect DNA replication, we analyzed the number of nuclei in the population of EII cells that were increased, observing no differences compared to the control (Figure 2C). Although a pleiotropic effect of 2-BP cannot be excluded, it is very likely that the observed decrease in cyst formation is associated with the inhibition of palmitoylation and the subsequent defect in ESVs docking and fusion, as was shown to be the case for other cells [79], [80].
Some results have suggested that palmitoylation in cells may occur nonenzymatically, i.e. spontaneous formation of thioester linkage in the presence of palmitoyl-CoA [81]. However, studies in yeast showed that DHHC protein family-mediated palmitoylation accounted for most of the palmitoylated proteins found in this organism [79]. Therefore, we decided to explore the Giardia proteome to study the presence of DHHC proteins in this parasite.
PATs, the discovery of which has been crucial for the enzymology of palmitoylation, are a widespread evolutionary family of proteins [16], [82] ranging from eight in Saccharomyces cerevisiae [82], twelve in Trypanosoma brucei [27], eighteen in Toxoplasma gondii [25], twelve in Plasmodium [26], [25] to twenty-three members in humans [82]. To identify the complete set of Giardia putative PATs, we performed a HMMER search against the Giardia complete proteome using a DHHC PAT HMMer profile from Pfam (zf-DHHC). As shown in figure 3A, we found nine DHHC proteins in the Giardia proteome that displayed conserved sequences when compared to other organisms: i) the DHHC-CRD domain, ii) the two short motifs DPG (aspartate-proline-glycine) and iii) TTxE (threonine-threonine-any-glutamate) motif [20], [82]. One protein (gla_8711) contained a DHYC amino acid motif, instead of the canonical DHHC motif. However, this DHYC motif has been reported to be functional in the yeast PAT Akr1 [16].
We next analyzed the molecular identity of Giardia DHHC proteins with bioinformatics tools. In agreement with previous reports for other PATs [20], [18], [25], Giardia DHHC proteins were predicted to be polytopic membrane proteins, mainly harboring between three and six TMDs with the DHHC domain facing the cytosol (Figure 3B). There is a small group of DHHC proteins, including yeast DHHC protein Akr1, displaying the conserved 33 amino acid ankyrin repeats, which are frequently involved in protein-protein interactions [83]. By contrast, none of the Giardia DHHC proteins showed ankyrin repeats in their structure. Moreover, gla_8619 displayed a coiled coil structure and gla_96562 a signal peptide. As already described for other organisms [18], [25], Giardia DHHC proteins displayed a conserved structure, sharing domains and motifs that are present across all members of this enzyme family.
The names used in this paper, GiardiaDB, NCBI, and UniProt accession numbers for Giardia DHHC proteins are indicated in table 1.
In order to elucidate the phylogenetic relationship among the PATs and to infer the evolutionary history of Giardia DHHC proteins, we retrieved 1034 DHHC-CRD protein sequences from 84 completely sequenced eukaryotic genomes, including the Giardia lamblia genome (Assemblage A, isolate WB), by means of the DHHC PAT HMMer profile from Pfam (zf-DHHC). A Multiple Sequence Alignment was constructed with PROMALS3D [49], and Block Mapping and Gathering with Entropy (BMGE) [51] was used to select columns suitable for Maximum Likelihood (ML) phylogenetic inference. Maximum likelihood phylogenetic trees were calculated using PhyML [52], and Branch support was evaluated by approximate likelihood-ratio test (aLRT) [84]. The resultant phylogenetic tree can be divided in six monophyletic clades (MC), three of which together contain almost 90% of all sequences (MC D, E and F). Four MC have Giardia DHHC proteins: MC A and D contain one DHHC sequence each, while MC E and F contain five and two Giardia sequences respectively (Figure 4A and figures S2, S3, S4, S5). Without any further consideration than the topology of the tree and the early divergent phylogenetic status of Giardia, it can be argued that the Most Recent Common Ancestor of Giardia and the rest of the eukaryotic lineage (MRCA) had a minimum of four and a maximum of six groups of PATs. However, of the two Giardia-lacking MC one is almost entirely composed of Plant paralogues (MC C). Moreover, many MC contain subclades composed mostly or even only by Plant paralogues, suggesting that gene duplication have largely taken place in this group. All these can be seen as an indication of functional diversification among Plants, which also constitutes a plausible evolutionary mechanism for the origin of the MC C.
If we hypothesize that all DHHC sequences evolve from 4 PATs groups in the MRCA, we should be able to explain, in a parsimonious way, the MC lacking Giardia sequences as examples of evolutionary innovation. As we mentioned before, this is suitable in the case of the MC C, but not for the MC B (the other Giardia sequences-lacking MC). This is because MC B is composed of sequences from a greater variety of organisms compared to MC C, making the possibility of a common functional diversification very unlikely. Nevertheless, it is possible for the MC B to be the result of reductive evolution, meaning that Giardia lost sequences during its adaptation to a parasitic lifestyle, since the more stable environment provided by the host can cause relaxation or loss of selective constraints.
We tested gene loss across DHHC-CRD protein family by examining the heavily duplicated genomes of Trichomonas vaginalis, given that duplicated genes are most likely to be released from functional constraints (Figure 4B). For this, we retrieved all DHHC sequences from Trichomonas (http://trichdb.org/trichdb/) using the same pipeline described above, except that this time no sequences were excluded from the posterior analysis. Variations in the HC, C and DHHC portions of the DHHC-CRD domain were extracted from the MSA, and mapped onto a phylogenetic tree. Contrary to what is found in Plants, there is a substantial presence of poorly conserved sequences among Trichomonas genome that cluster together in the tree. Moreover, we found a strong correlation between the degree of conservation in the HC, C and DHHC portions of the DHHC-CRD domain within each sequence.
Altogether, our findings suggest that the MRCA had five groups of DHHC sequences from which the other sequences eventually evolved by functional diversification, and that Giardia lost at least one representative sequence presumably during its adaptation to a parasitic lifestyle.
We also determined the orthology relationships between sequences from different assemblages. For this, we retrieved DHHC sequences from Giardia isolates WB, GS and P15 (Assemblages A, B and E, respectively; http://giardiadb.org/giardiadb/), following the pipeline described above. As expected, every DHHC sequence in the isolate WB has a highly similar ortholog in the other isolates, which cluster together in the tree (Figure 5). Only one WB sequence, EAA36893, escapes this pattern, but this probably constitutes a case of defective annotation in isolates GS and P15.
Semi-quantitative RT-PCR indicated that all the dhhc genes were expressed in trophozoites and in encysting parasites (Figure S6). This prompted us to explore further the expression levels of these genes in growing and encysting parasites by performing qRT-PCR analysis of mRNA expression from these cells. As shown in figure 6, many of the dhhc transcripts were present at relatively constant levels, but gla_8619, gla_1908, and EAA36893 were downregulated in encysting parasites while gla_2116 was upregulated in 48 h encysting cells. Considering that Giardia contains minimal systems, either as a result of reductive processes associated with a parasitic lifestyle, as a reflection of basic evolutionary characteristics, or both [85], [86], the fact that the nine dhhc genes found by bioinformatics were expressed in vegetative and encysting parasites suggests that protein palmitoylation and the PATs themselves may be playing a key role during the entire life cycle of this parasite.
We next sought to characterize four of the nine DHHC proteins that are expressed in Giardia based on their expression profile. We chose two that are expressed at similar levels in growing and encysting parasites (gla_8711 and gla_16928), one that is downregulated during encystation (gla_1908), and one that is upregulated in encysting parasites (gla_2116).
To further analyze these DHHC proteins, we expressed full-length gla_1908, gla_2116, gla_16928 and gla_8711 as fusion DHHC proteins containing C-terminal HA-tag [39] and evaluated their protein expression profiles by Western blotting using an anti-HA mAb (Figure 7). Analysis by semi-quantitative RT-PCR indicated that the overexpression of these fusion proteins was 2 to 3-times higher in transgenic cells, as reported for protein expression using a similar vector [9]. Immunofluorescence assays showed that HA-tagged gla_1908, gla_2116, and gla_16928 partially co-localized with BiP in the endoplasmic reticulum (ER) or around the nuclei of transgenic trophozoites (Figure 8, trophozoite). Our results confirmed the localization of gla_16928 already shown by Touz et al. [31]. Analysis of intracellular localization of yeast and mammalian DHHC proteins revealed that the majority of these localize to the ER and Golgi [20], [87]. However, there are a few exceptions, including human DHHC5 protein [87] and Giardia DHHC protein (EAA36893) [31], which localize to the plasma membrane. Also, we found that gla_8711 partially co-localized with the adaptor protein AP-2 [57] at the lysosomal-like peripheral vacuoles (PVs) as well as in plasma membrane and flagella (Figure 8, trophozoite). Ongoing experiments intended to knock-down this protein may reveal its importance during the Giardia life cycle.
The hallmark of encystation in Giardia is the synthesis of CWP1, CWP2, and CWP3 [88]. These proteins are expressed and concentrated within the ESVs before they are targeted to the cyst wall [89], [6], [90]. To address the influence of the overexpression of these HA-tagged DHHC proteins during encystation, dhhc-ha transgenic trophozoites were induced to encyst in vitro. The localization of DHHC-HA proteins as well as CWP1 expression, intracellular localization, and vesicle formation were addressed by IFA. To examine in detail the results obtained, we decided to analyze each dhhc-ha transgenic cell following the protocol described above, in which the cells were classified as encysting I, encysting II, and early cyst. We observed that gla_1908 (Figure 8A), gla_2116 (Figure 8B), and gla_8711 (Figure 8D) transgenic parasites displayed normal encystation. It was noteworthy that gla_16928 (Figure 8C) had enlarged ESVs, with co-localization between gla_16928-HA and CWP1 observed in those vesicles (Figure 8C, inset). Additionally, it was noted that gla_16928 early cysts had a larger size and an abnormal shape compared with wild-type cells (not shown) and other transgenic early cysts.
When CWP expression was analyzed in dhhc transgenic parasites by qRT-PCR, we observed that, except for gla_2116 transgenic cells, which displayed similar levels or even moderate decrease in the mRNA expression of CWPs compared to the control, the other dhhc-ha transgenic parasites showed increased expression of CWP1, CWP2, and CWP3 (Figure 9A). Several transcription factors have been described as involved in the regulation of cwp gene transcription [91], [92], [93], [94], [95], [96], [97]. However, the mechanisms underlying transcription control in this parasite have not been completely elucidated. It has always been assumed that the mobilization mechanism for transcription factors in many organisms is based on proteolytic processing [98], [99], [100], [101]. Nevertheless, there is a group of lipid-modified transcription factors whose mobilization mechanism to the nucleus is not based on proteolytic processing but on reversible palmitoylation [102]. If that were the case for the transcription factors involved in Giardia encystation, DHHC proteins would be palmitoylating different transcriptions factors that, in turn, may regulate CWP expression. It would be interesting to explore the molecular architecture of Giardia transcription factors to find out whether palmitoylation is involved in regulating their shuttling between the cytoplasm and the nuclei.
Analyzing the amount of water-resistant cysts, we observed that gla_1908 and gla_8711 transgenic cells yielded a significantly higher amount of cysts than the control (Figure 9B). In contrast, gla_2116 transgenic cells, while displaying an apparently normal encystation process (Figure 8B) and CWP expression (Figure 9A), produced a reduced number of mature cysts (Figure 9B). A likely explanation is that gla_2116 may be involved in the palmitoylation of a protein in charge of turning encystation-specific genes off and ending the encystation process. In the case of gla_16928 transgenic parasites, these cells produced a low percentage of cysts (Figure 9B) although the CWP expression was increased (Figure 9A). These findings, in addition to the large ESVs seen in figure 8C (encysting II) and the large size of early cysts (Figure 8C, early cyst), may be explained by a high rate of synthesis of CWPs in gla_16928 transgenic parasites, which may exceed the mechanisms of vesicle discharge regulation, leading to the formation of immature non-water-resistant cysts. Further experiments using knock-down strategies are needed to completely address the role of each DHHC protein in the encystation process. Table 2 summarizes the main features of the Giardia DHHC proteins analyzed in this work.
The different localization of DHHC-HA proteins in trophozoites and the differential effect of DHHC overexpression in encystation prompted us to evaluate the palmitoylation pattern in the dhhc transgenic parasites (Figure 10). gla_1908, gla_2116, gla_16928, and gla_8711 transgenic trophozoites or encysting parasites displayed a similar global protein palmitoylation pattern compared to wild type (Figure 1A). Mass spectrometry-based proteomics analyses will be necessary to accurately identify any differences in the palmitoylation substrates among the dhhc transgenic parasites.
This work presents a detailed analysis of Giardia lamblia DHHC protein structure and phylogeny and reveals a possible role of palmitoylation in Giardia encystation. Our data, suggesting the presence of DHHC proteins in growing and encysting parasites, reinforced the idea that this PTM has conserved and important functions in cell-signaling, protein-sorting and protein-export throughout evolution. Without being able to assign a specific substrate candidate to each Giardia DHHC proteins, we showed that overexpression of these enzymes had consequences on CWP expression and on the amount of cysts produced. Proteomic analysis of Giardia palmitoyl proteome would be a great contribution to elucidating the mechanisms by which palmitoylation participates in encystation biology. Finally, the suggested role of palmitoylation in Giardia encystation, a key event that enables the parasite to survive in the environment, infect a new host and evade the immune response [1], [103], could open new ways to intervene in the process of Giardia infection.
|
10.1371/journal.pcbi.1000428 | Analysis of the Free-Energy Surface of Proteins from Reversible Folding Simulations | Computer generated trajectories can, in principle, reveal the folding pathways of a protein at atomic resolution and possibly suggest general and simple rules for predicting the folded structure of a given sequence. While such reversible folding trajectories can only be determined ab initio using all-atom transferable force-fields for a few small proteins, they can be determined for a large number of proteins using coarse-grained and structure-based force-fields, in which a known folded structure is by construction the absolute energy and free-energy minimum. Here we use a model of the fast folding helical λ-repressor protein to generate trajectories in which native and non-native states are in equilibrium and transitions are accurately sampled. Yet, representation of the free-energy surface, which underlies the thermodynamic and dynamic properties of the protein model, from such a trajectory remains a challenge. Projections over one or a small number of arbitrarily chosen progress variables often hide the most important features of such surfaces. The results unequivocally show that an unprojected representation of the free-energy surface provides important and unbiased information and allows a simple and meaningful description of many-dimensional, heterogeneous trajectories, providing new insight into the possible mechanisms of fast-folding proteins.
| The process of protein folding is a complex transition from a disordered to an ordered state. Here, we simulate a specific fast-folding protein at the point at which the native and denatured states are at equilibrium and show that obtaining an accurate description of the mechanisms of folding and unfolding is far from trivial. Using simple quantities which quantify the degree of native order is, in the case of this protein, clearly misleading. We show that an unbiased representation of the free-energy surface can be obtained; using such a representation we are able to redesign the landscape and thus modify, upon site-specific “mutations”, the folding and unfolding rates. This leads us to formulate a hypothesis to explain the very fast folding of many proteins.
| It is commonly believed that, with sufficient computer time and accurate models, the energy landscape of any protein could be mapped out from its sequence by running and analysing folding simulations, thus making possible prediction of both folding mechanism and native structure. This is not yet possible: folding events have only been observed in simulations of very small, fast (sub µs) folders [1],[2]. The main reason for this limitation is the computational expense of accurate protein models, which typically allow only a few ns of dynamics to be generated within a reasonable timescale of weeks or months. Another obstacle may be the models themselves, whose accuracy is difficult to assess for the very same reason. Nevertheless, with the development of faster processors, new sampling techniques and improved force-fields, equilibrium simulations of accurate protein models are likely to become achievable in a not-too-distant future. The analysis of such equilibrium simulations, however, poses another problem. Determining and representing the free-energy surface, which underlies the thermodynamic and dynamic properties of the model, from an equilibrium simulation in a meaningful way is a complicated task, and numerous studies have been devoted to this task [3]–[10]. Most commonly, the free energy surface has been projected on a small number (usually one or two) progress variables, such as the root mean square distance (RMSD) from the native structure, the radius of gyration Rg or the number of native contacts. Integrating over all other degrees of freedom induces a free energy landscape as a function of these coordinates, which typically exhibits a maximum (the transition state) at some point between the minima representing the ensembles of denatured states and the native state. This enormous projection is highly problematic, as features inherent to the multi-dimensional nature of the true folding space, such as the presence of local minima, can be lost. Most importantly, the existence and height of free-energy barriers in these projections are often inaccurate. One solution to this problem is provided by a recently proposed method to determine and represent unprojected free-energy surfaces [11],[12]. Based on disconnectivity graphs [13], the method aims to group conformations into free-energy minima not using geometrical criteria but equilibrium dynamics. More recently this method has been extended to determine a one-dimensional projected free-energy surface in terms of a reaction coordinate that preserves the free energy barrier, and the coordinate dependent diffusion coefficient [14]. This method has previously been applied to model systems such as a 20-residue designed peptide that folds to a double hairpin [10] and a coarse-grained model of a protein under mechanical force [15].
The problem of how best to analyse an equilibrium folding trajectory cannot be addressed with detailed models for the reasons mentioned above. Reversible folding trajectories can, however, be obtained with structure based models, hence their broad popularity in computational folding studies [16]–[25]. Using these models a sequence can fold from a random extended conformation to the native structure, reach equilibrium and unfold and refold a large number of times in a typical trajectory. Depending on the target structure, the free-energy barrier for unfolding may still be exceedingly large and folding too slow to be observed. Such models disfavour non-native interactions, and are therefore strongly biased towards native interactions. Consequently their accuracy in describing the folding behaviour of real proteins has been debated [26],[27]. Nevertheless they predict features which are believed to be characteristic of the folding landscapes of real proteins, such as the presence of intermediates [28]–[30] and downhill folding [31]–[36], and are undoubtedly useful for understanding the general features of landscapes. Structure based models are also easily malleable and sensitive to individual interactions [37]–[39], allowing the effects of perturbations of the free energy landscape to be investigated.
In this paper we use both geometric projections and the unprojected representation described above to extract free energy surfaces from reversible folding simulations. The specific landscape which is probed is that of a model of the N-terminal domain of phage λ-repressor protein [40] at its melting temperature. We chose this five-helix bundle protein (Figure 1) because it has been extensively studied experimentally [41]–[49], and has been shown to be a very fast (∼3600 s−1 at 37°C and 0 M urea), two-state folder [40]. The two analyses are compared, and states which are hidden by the geometric projection are discovered. In particular, hidden parallel pathways and intermediates are found to play an important role in the fast folding of the model. Removing these features by perturbing the model results in a more than two-fold reduction in the folding rate. The aim of this work is not to discuss the merits of structure-based models for reproducing known experimental properties of proteins, but rather to demonstrate the importance of a thorough analysis of equilibrium kinetics which is not biased by the choice of arbitrary projection variables.
Simulations of λ-repressor and two variants have been performed using the force-field of Karanicolas and Brooks [50],[51] implemented in the program CHARMM [52]. In this structure-based Cα model, interactions are attractive if they are present in the experimental native state and repulsive otherwise. The magnitude and range of the interactions depend on the chemical properties of the residues and their separation in the experimental structure. The dihedral part of the potential is sequence-specific.
The force-field was modified to generate two variants, A and B. In variant A, the magnitudes of the non-bonded interactions between residue 73 and residues 80, 81 and 84 were increased by factors 1.75, 2.5 and 1.75, respectively. In variant B, attractive non-bonded interactions were introduced between residues 43 and 48, and residues 44 and 47.
To maintain a constant temperature, Langevin dynamic simulations were performed with a timestep of 15 fs and a uniform friction coefficient of 1 ps−1 acting on all particles. We verified that the friction coefficient corresponds to the regime in which rates are proportional to the friction coefficient, i.e., we use a friction low enough to guarantee the generation of a large sample of folding/unfolding events, but which is not in a ballistic, low friction regime [53].
Simulations of each protein were performed over a broad range of temperatures, and the Weighted Histogram Analysis Method (WHAM) [54] used to calculate specific heat capacity curves. The temperature at which the specific heat reached a maximum was identified as the melting temperature . Longer (30 µs) simulations were run at this temperature, with coordinates being saved every 7.5 ps. More than 600 folding events were observed for the wild-type protein.
The equilibrium trajectories are first analysed by projection onto the geometric coordinates RMSD from the native structure and fraction of native contacts formed (). Contacts are considered to be present if two Cα atoms are separated in sequence by more than 4 residues and are less than 12 Å apart, and the native contact map is constructed from the experimentally determined native structure.
The further analysis consists of three stages. First, the trajectory is used to build a network, the equilibrium kinetic network (EKN), which describes the system kinetics at equilibrium. This is obtained by clustering the trajectory in the principal component space defined by the distance between selected atom pairs, and counting the number of transitions between clusters (see Text S1 for details).
Once such network has been determined, its free energy profile (FEP) is built using a procedure which is described in detail elsewhere [14],[55] and in Text S1. The FEP is plotted as a function of a “natural coordinate” which is constructed so that the diffusion coefficient is constant along the profile, and the mean first passage times (MFPTs) between any two points can be calculated using Kramer's equation [10]. For sequential folding pathways, the heights of the barriers on the FEP of the system are exact. If parallel pathways are present, however, usually only the highest barrier is exact. To overcome this problem, any two states can be chosen and the FEP between only these two states built, giving an exact barrier height. The third stage of the process is to use the FEP to iteratively partition the network into basins to generate a simplified EKN (SEKN) which describes the system kinetics. The procedure by which the SEKN is generated is described below.
The simplified equilibrium kinetic network (SEKN), which describes the inter-basin kinetics, is constructed by iteratively partitioning the EKN into basins. To do this, notable barriers are first identified in the FEP. Two representative nodes on either side of the barrier are selected in the EKN, and the network divided by computing the “minimum cut” [11],[12] between these two nodes. This procedure is applied iteratively until there are no notable internal barriers in any of the basins. The number of effective transitions between each pair of directly connected basins is then computed by assuming diffusive dynamics and using Kramers' equation to estimate the mean first passage time from one basin to the other [55].
For all the analyses shown below, we assessed the convergence by repeating the analysis for the first and second half of the trajectories. The networks are in all cases identical and the populations of basins differ at most by 10% (see Text S1 for details).
At first glance, the folding behaviour of the structure-based model of λ-repressor appears to be two state. The specific heat profile shows a sharp peak at the melting temperature (), indicating highly cooperative folding behaviour. Timeseries' of geometric coordinates such as the number of native contacts QN and RMSD (shown in Figure 2A) switch rapidly between two states: one is characterised by high QN and low RMSD (i.e. a native-like state) and the other by low QN and high RMSD (a denatured-like state). According to these coordinates, therefore, folding of the model is a two-state process. More than 600 folding events occur within the simulation time of 30 µs. Figure 2B shows free energy profiles built from projections of the trajectories onto the two coordinates. Clearly two stable states are present, separated by a small barrier. The relative stabilities of the two states, however, differ according to the coordinate used: while on the RMSD projection the native state is marginally more stable than the denatured state, the opposite is true when QN is used as the reaction coordinate. The size of the barrier for the folding transition also differs from in the RMSD projection to in the QN projection. These differences highlight the difficulties involved in analysing trajectories by projecting them onto single geometric reaction coordinates. A better solution may be to project onto a plane defined by several reaction coordinates: the top left panel of Figure 3 shows a projection of the trajectory at onto both RMSD and QN. This projection appears to be more reliable, with the two states being clearly separated, and an energy barrier of around . However, as we will show in the next section, even this projection hides detail which is important in understanding the folding process.
Figure 4 shows the results of the more detailed analysis of the trajectory at . Panel B shows the free energy profile (FEP) as a function of the “natural coordinate” described previously: five stable states are identifiable. These five free energy basins are plotted as a function of RMSD in panel C. At low values of RMSD (∼2 Å), two native basins are present, labelled n1 and n2. Two intermediate states i1 and i2 lie at slightly higher RMSD (∼4 Å). The denatured state d is a broad basin with a minimum at RMSD ∼15 Å. Figure 3 shows the positions of the five states on a projection onto the two-dimensional reaction coordinate (RMSD, QN): the two native and two intermediate states overlap considerably, making them indistinguishable in the overall projection. The SEKN, which provides information about the populations and kinetics of the network, is shown in of Figure 4A. Two parallel pathways can be identified as the main folding routes: d→i1→n1 and d→i2→n2. Folding also occurs through i1→n2 and i2→n1, but at a much slower rate. Interchange between the two native states (n1→n2) and between the two intermediate states (i1→i2) is rapid, suggesting that they are separated only by small free energy barriers. From FEPs plotted between the states the size of these barriers can be estimated as 3 and 2.5 kBT for the native and intermediate states respectively. Exchange between the native and intermediate states (i.e. n1→i1 and n2→i2), is also fast, and these states are separated by energy barriers of only ∼2 kBT. The rate limiting step in folding is the transition between the denatured and intermediate states, for which the energy barrier is ∼5 kBT.
The distribution of folding times from d to n1/n2 is shown in Figure 5. The curve fits a single exponential distribution: the equilibration of the native and intermediate states is sufficiently fast compared to the d to i1/i2 step that a single time constant can be used to describe the folding with reasonable accuracy. This has the consequence that, should the folding pathways described above be representative of the real protein, a kinetic experiment would not reveal the presence of the intermediate state, or indeed the parallel pathways.
Panels A and B of Figure 6 show matrices of average inter-residue distances for the n1 and i1 states. The two are similar, with local contacts being present in the helical regions (residues 9–23, 33–39, 44–51, 61–69 and 79–85), as well as several regions of non-local contacts. The differences between the two states lie in the helix 5 region, in which the non-local contacts are significantly reduced. This can be more clearly seen in the matrix of differences between the pairwise distances (Figure 7C): helix 5 moves away from the rest of the protein during the transition from state n1 to state i1. The distance matrices for the n2 and i2 states, which are not shown, reveal an analogous change. The secondary structure propensities for the native and intermediate states are shown in Figure 8. Whilst all five helices are always present in the two native states, the helicity, and particularly the helicity of helix 5, is slightly diminished in the intermediate states: in both i1 and i2 helix 5 is only present in around 75% of structures. The positional root mean fluctuations (RMSF) of each residue (Figure 9) for the intermediate and native states also indicate that the largest differences are in the helix 5 region, in which the flexibility is significantly larger in the intermediate states than in the native states. Analysis of contact probabilities reveals that 12 attractive native contacts are lost (or present in at least 50% fewer structures) in the transitions from n1 to i1 (or n2 to i2), and these are all made by residues in helix 5 and the loop between helices 4 and 5. Together these analyses give a clear picture of the two intermediate substates. In state i2 helices 1–4 are native-like, and helix 5 is generally formed but detached from the rest of the structure. State i1 is similar, but with a slightly frayed helices 1–4. Figure 10B shows representative structures of the i2 state.
As non-native interactions are not included in the model, entropy must play an important part in stabilising the intermediate states. In this case the loss of enthalpy that results from breaking the long-range native contacts made by helix 5 is balanced by the increased entropy associated with the freedom of the helix.
The differences between the n1 and n2 states and between the i1 and i2 states are more subtle. The left-hand panel of Figure 7 shows the changes in average pairwise distances between the two native states; the differences are very clearly localised in the region of residues 42–47 (part of the loop region between helices 2 and 3, and the N terminal of helix 3). This difference can also been seen in the secondary structure propensities of the two states (Figure 8): helix 3 is slightly shorter in state n1, commencing at residue 47 rather than residue 44. The region between the two helices, which has no secondary structure elements in state n2, is classified as a bulge or a turn in state n1.
Figure 9A shows the RMSF for each residue in the two native states (n1 and n2). Again, the differences are localised in the same area, with state n1 being more flexible in this region than n2. The increased entropy associated with the increased flexibility in state n1 is compensated for by a loss of attractive contacts: Table 1 shows that several attractive contact probabilities, all in the residue 42–47 region, are significantly reduced in n1 compared to n2. Figure 10A shows representative structures of states n1 and n2.
These analyses show that the two native states arise from a careful balance of enthalpy and entropy: whilst n1 loses out in enthalpic terms by having fewer attractive contacts than n2, it gains entropy from increased flexibility of the loop. This is also the case for the two intermediate states: again the changes are localised to the same loop region (Figure 7B), and the increased entropy associated by the flexibility of the loop in i1 (Figure 9B) is balanced by a loss of contacts in this region (Table 2).
Figure 3 shows that the denatured state identified by the unprojected analysis is very similar in terms of RSMD and QN to the denatured state identified by projection onto these coordinates. The enthalpic destabilization and high heterogeneity of the denatured state make it intrinsically difficult to study, both in experiment and simulation, and it is therefore interesting to characterize it to some extent here. As stated previously, the aim of this paper is not to reproduce the experimental properties of λ-repressor, or to debate the accuracy of coarse grained models. Nevertheless, it is a valuable exercise to make some comparison with experiment, as such a comparison could point in directions in which the model could be improved. The average radius of gyration of the denatured state in the simulation is 20.5 Å; this compares well with the value determined experimentally for a mutant of the same protein of 23±2 Å [44]. Both the experimental and simulation values of Rg are smaller than the value (26 Å) expected for a random coil [56], indicating that there are residual interactions in the denatured state. Certainly this is the case in the simulation: the average pairwise distance matrix for the denatured state (Figure 6) shows that although no long range interactions are present, a number of local contacts are formed, indicating the presence of some secondary structure. This can also be observed in the secondary structure propensity of this state (Figure 8): whilst the helices are diminished in this state, all five are present to some extent. Evidence of secondary structure in the denatured state has been found for a number of proteins [57],[58]. In fact, a recent NMR study of a mutant of λ-repressor in which the denatured state is populated under non-denaturing conditions showed that significant helical structure was present [48],[49]. In contrast to the simulation results presented here, however, the helicity was limited to the N-terminal region of the protein. This disagreement indicates that the high helicity observed in the simulation may well be an artifact of the model.
The malleability of the Go-like model, together with the above information about the folding mechanism, allow modifications of the model which alter the folding pathway. Such modifications are useful as, by comparing the folding rates of the wild-type and modified proteins, it may be possible to identify those features in the folding landscape of the wild-type which make it a fast folder. Here, two modifications have been made: one which removes the intermediate states from the pathway, and another which removes the parallel pathways.
The first modification (A) was designed destabilize the intermediate state: the interactions of residue 73 with residues 80, 81 and 84 are strengthened. This should clamp helix 5 into its native position, and thus destabilise the intermediate state, in which helix 5 is not docked. The melting temperature of the modified model is slightly higher than the wild-type (327 K compared to 323 K) i.e., the modification marginally stabilises the native state. The FEP (Figure 11B) calculated from simulations at shows only three stable states; from the RMSD plot (Figure 11C) they can be identified as two native substates (n1 and n2), and the denatured state. The intermediate states have been destabilized sufficiently that they are no longer significantly populated. Interchange between the native substates is rapid (see SEKN, Figure 11A SEKN), but the barrier between n1/n2 and d is rarely crossed.
The second modification (B) was designed to force the model to fold via a single, rather than parallel, pathway. The above analysis shows that the native and intermediate substates differ mainly in the region of residues 42–47. Introducing attractive interactions between those pairs of residues which form contacts in state n2 but not in n1 should stabilise n2 relative to n1 and thus channel the flux into a single pathway. Two interactions were introduced in the design of model B: between residues 43 and 48, and 44 and 47. The SEKN for this model (Figure 12A) shows that the design was successful: the protein now folds via the pathway .
Folding rates for the wild-type and two modified proteins, taken from the SEKN, are shown in Table 3. Folding rates are for the d to i1/i2 transition for the wild-type and model B, as this is the rate limiting step, and for the d to n1/n2 transition for model A. Both models fold significantly more slowly than the wild-type. This result is important as it shows that both the intermediates and parallel pathways are at least partially responsible for the observed fast folding of the wild-type model.
In this paper we have investigated several ways of analysing equilibrium simulations: traditional geometric analysis, in which the trajectory is projected onto one or several reaction-coordinates, and a recently proposed method which uses an unprojected representation of the free energy landscape. In particular we have focused on the folding of a structure-based model of a small, fast-folding five-helix bundle, λ-repressor, which has been widely studied experimentally. Fluorescence and NMR measurements indicate that λ-repressor is a two state folder which can be transformed into a barrierless folder via specific mutations. The simulations agree with experiment when analysed using RMSD and QN as reaction coordinates: the model appears to fold quickly via a two state transition. The unprojected analysis, however, reveals more complexity: an obligatory intermediate state is present in the pathway, and the native and intermediate states are split into two “sub-states”. The intermediate states, which cannot be distinguished from the native states in projections over conventional geometrical coordinates, are stabilised by a balance of enthalpy and entropy: helices 1–4 are natively docked and helix 5 is generally formed but detached.
The characterisation of the different states on the folding pathway revealed by the detailed analysis allowed the design of “mutants” of the model which fold via different mechanisms. In one mutant, the intermediate states were destabilised so that they were no longer populated i.e., folding occurred directly from the denatured state to the two native substates. The role of intermediates in folding has been widely debated: it appears that, depending on their stability [59] they may act as kinetic traps and thus slow folding [60], or as an important stepping stone, channeling flux to the native state and thus accelerating folding [61],[62]. The analysis of the folding of both the “wild-type” model and the “mutant” showed that the rate of folding was significantly smaller for the mutant. This indicates that, for our model, the intermediate state guides the protein towards the native state, thus accelerating folding. Another mutant was designed to fold via a single pathway i.e., the native and intermediate substates of one pathway were stabilized so that the other pathway was no longer significantly populated. The resulting folding rates were smaller than the wild-type, and approximately equal to the rate that could be predicted from considering only one path of the wild-type. This result demonstrates that, at least for this model of λ-repressor, the fast observed folding rates are at least partially due to the presence of parallel pathways.
It is well known that experimental probes of protein folding are often localised and therefore may not be sensitive to structural changes in distant parts of the protein. In this paper we have shown that an analogous problem exists in simulation: the projection of reversible trajectories onto geometric reaction coordinates can hide important features of the folding pathway. Such features can, however, be uncovered by a more detailed analysis such as the unprojected representation used here. This detailed analysis reveals important characteristics of the folding landscape of a structure-based model of a fast-folding protein which help to explain how it folds so quickly.
|
10.1371/journal.pntd.0006859 | Estimating the current burden of Chagas disease in Mexico: A systematic review and meta-analysis of epidemiological surveys from 2006 to 2017 | In Mexico, estimates of Chagas disease prevalence and burden vary widely. Updating surveillance data is therefore an important priority to ensure that Chagas disease does not remain a barrier to the development of Mexico's most vulnerable populations.
The aim of this systematic review and meta-analysis was to analyze the literature on epidemiological surveys to estimate Chagas disease prevalence and burden in Mexico, during the period 2006 to 2017. A total of 2,764 articles were screened and 36 were retained for the final analysis. Epidemiological surveys have been performed in most of Mexico, but with variable study scale and geographic coverage. Based on studies reporting confirmed cases (i.e. using at least 2 serological tests), and taking into account the differences in sample sizes, the national estimated seroprevalence of Trypanosoma cruzi infection was 3.38% [95%CI 2.59–4.16], suggesting that there are 4.06 million cases in Mexico. Studies focused on pregnant women, which may transmit the parasite to their newborn during pregnancy, reported an estimated seroprevalence of 2.21% [95%CI 1.46–2.96], suggesting that there are 50,675 births from T. cruzi infected pregnant women per year, and 3,193 cases of congenitally infected newborns per year. Children under 18 years had an estimated seropositivity rate of 1.51% [95%CI 0.77–2.25], which indicate ongoing transmission. Cases of T. cruzi infection in blood donors have also been reported in most states, with a national estimated seroprevalence of 0.55% [95%CI 0.43–0.66].
Our analysis suggests a disease burden for T. cruzi infection higher than previously recognized, highlighting the urgency of establishing Chagas disease surveillance and control as a key national public health priority in Mexico, to ensure that it does not remain a major barrier to the economic and social development of the country's most vulnerable populations.
| In Mexico, estimates of Chagas disease prevalence and burden vary widely due to the ecology and epidemiology of this disease resulting of many geographical, ecological, biological, and social interactions. Better data are thus urgently needed to help develop appropriate public health programs for disease control and patient care. In this study we performed a meta-analysis from published data on T. cruzi infection seroprevalence in Mexico between 2006 and 2017. This systematic review shows a national estimated seroprevalence of T. cruzi infection of 3.38% [95%CI 2.59–4.16], with over 4.06 million cases in Mexico, which is higher than previously recognized. The presence of T. cruzi infection in specific subpopulations such as pregnant women, children and blood donors also informs on specific risks of infection and calls for the implementation of well-established control interventions. This work confirms the place of Mexico as the country with the largest number of cases, highlighting the urgency of establishing Chagas disease control as a key national public health priority.
| Chagas disease or American trypanosomiasis is an infection caused by the protozoan parasite Trypanosoma cruzi, which is mainly transmitted to humans and other mammals through the contaminated feces of hematophagous bugs called triatomines (family Reduviidae). However, it can also be spread via non-vectorial routes, such as blood transfusion, congenital transmission, organ transplantation, ingestion of food and beverages contaminated with T. cruzi, or laboratory accidents [1]. Over the years, infection with T. cruzi can cause heart failure or sudden death associated with progressive heart damage [2]. Some patients may also suffer from digestive, neurological or multiple alterations. This disease, classified by the World Health Organization (WHO) within the group of Neglected Tropical Diseases, is a major public health problem in Latin America where it is estimated that 6 to 7 million people are currently infected [1]. Due to human migrations, Chagas disease is emerging in other regions (Europe and United States principally) [3]. Estimates suggest that 80,000 to 120,000 T. cruzi-infected immigrants live in Europe, and 300,000 live in the United States [4], and the disease is a growing concern in these regions [5]. The global economic burden of Chagas disease is more than US$7.2 billion per year, exceeding the costs of other diseases of health impact such as certain cancer (US$6.7 billion for uterine cancer, US$4.7 billion for cervical cancer, and US$5.3 billion for oral cancer) or rotavirus infections (US$2 billion) [6,7].
In Mexico, estimates of Chagas disease prevalence and burden vary widely, which has complicated the establishment of a strong National Chagas Disease Program for vector control as well as for patient detection and care in the country. For the past several years, the Ministry of Health only reports a few hundred cases per year [8], suggesting that the disease has an anecdotal burden in terms of public health. On the other hand, other estimates suggest that there are about 1.1 million individuals infected with T. cruzi in Mexico, and 29.5 million at risk of infection [9,10]. Higher estimates of up to 6 million cases have also been proposed [11]. The annual cost for medical care for patients in the outpatient setting in this country is estimated between US$4,463 and US$9,601, and annual costs for patients admitted via an emergency care unit is between US$6,700 and US$11,838 [12].
There are also important regional differences in prevalence levels or number of cases reported in Mexico. For example, between 1928 and 2004 the states with the highest number of human cases reported were Chiapas, Guerrero, Jalisco, Morelos, Nayarit, Oaxaca and Queretaro. Conversely, few cases were reported in the states of Chihuahua, Coahuila, Guanajuato and Estado de Mexico [11]. It is not clear if such differences in prevalence are reflecting true differences in eco-epidemiological conditions, as Mexico is home to an extensive diversity of triatomine species, habitats, and socioeconomic conditions, or if there are bias in disease surveillance among regions [11].
Such wide discrepancies are important to reconcile to ensure that Chagas disease does not remain a major barrier to the development of Mexico's most vulnerable populations. Updating and improving surveillance data for Chagas disease in Mexico is therefore an important public health priority. In this context, the aim of this systematic review and meta-analysis was to estimate Chagas disease prevalence and burden in Mexico. We focused our study on the period from 2006 to 2017, to define current disease status rather than historical/cumulative burden, but our results are nonetheless compared with past reviews [11,13,14] to shed light on possible temporal trends on the status of Chagas disease in the country.
The current study was conducted in accordance with the PRISMA statement [15] (Supporting information). Potential data sources were identified and selected in different bibliographic databases. The ISI Web of Science (v5.13.1) was chosen because it incorporates many relevant databases including the SciELO Citation Index from 1997 onwards (provides access to leading journals from Latin America, Portugal, Spain and South Africa) and the Web of Science’s Core Collection from 1980 onwards (https://webofknowledge.com/). A part of the literature was selected from the LILACS database (lilacs.bvsalud.org/en/), which is the most important index of scientific and technical literature of Latin America and the Caribbean. Finally, the BibTri database (https://bibtri.cepave.edu.ar/) was also used because it integrates scientific literature specifically related to Chagas disease.
We restricted our search to the period from January 2006 to December 2017, to obtain information on the current status of Chagas disease in Mexico rather than on its historical/cumulative status, which has been summarized in previous reviews [11,13,14]. Selection was made using the search terms ‘Chagas disease in Mexico/Enfermedad de Chagas en México’ and with the equivalent keywords obtained via Medical Subject Headings (MeSH) website (https://meshb.nlm.nih.gov/search), i. e. American Trypanosomiasis, Chagas' Disease, Trypanosoma cruzi Infection, Trypanosomiasis in South American.
For all these articles, titles and abstracts were screened for any indication that the study contained data related to the seroprevalence of T. cruzi infection in human populations from Mexico. Typically, this excluded studies of, for example, therapeutic options for patients with chronic Chagas disease, molecular studies of lab strains of the parasite, or experimental model developments (Fig 1). In the second step of the process, full text copies were obtained and articles containing quantitative data on T. cruzi infection seroprevalence were retained. Extreme care was taken in cross-validating whether the information contained in each study was unique and not duplicated elsewhere.
The ultimate step was to extract the relevant information contained in the selected articles which included 1) publication data (bibliographic information), 2) sampling dates, 3) sampling strategy (archive, random, volunteers, etc…), 4) geographic area covered by the study, 5) studied population (blood donors, patients, pregnant women, newborns, children, random populations), 6) laboratory techniques used (ELISA tests; IHA, PCR…) and the number of laboratory techniques used to validate the cases detected, 7) total sample size, number of human cases. Studied populations were then divided into subgroups to allow for analysis of the seroprevalence of T. cruzi infection at different levels, including the general population (population sampled in different geographic locations), pregnant women, children (under 18 years old), and blood donors (number of patients).
We calculated two general estimated prevalences 1) the first considering all studies, irrespective of the inclusion of confirmatory diagnostic, and 2) the second including only the studies in which at least 2 serological tests were used (as recommended by the WHO for an accurate identification of cases, see results part). We further calculated 95% confidence intervals (95%CI) based on the reported data and sample sizes [16,17]. We assessed the extent of publication bias in the selected studies through a funnel plot. Next, we performed a meta-analysis to calculate the effect size estimate, and the weighted effect size (% weight), based on a random-effects model [18]. For each estimated prevalence, a test for heterogeneity among studies (Q) and the variability of the effect size due to variation between observations (I2) was calculated, and forest plots were elaborated. The estimated prevalence obtained in each population were then compared to the data reported by the Ministry of Health and with other reviews [8].
A total of 2,764 articles were screened and 36 were retained for the final analysis (see Fig 1). All the articles included in this study corresponded to serological surveys in different populations and settings, including general or specific populations such as pregnant women or blood donors, published between 2006 and 2017 (Table 1). Research on Chagas disease seroprevalence has been performed in most of the Mexican Republic (Fig 2). The states with more studies were Veracruz, Yucatan, and Queretaro. The extent of publication bias in the selected studies (with a total of 79 observations, each of the 36 studies may have several observations of different states, populations…) was assessed through a funnel plot (Fig 3), and the symmetric distribution of data points indicated a lack of publication bias or systematic heterogeneity of the dataset.
We first considered the studies in the general population (population sampled in different geographic locations) and in pregnant women (which can be considered as highly representative of the general population as well [55]) to obtain a national estimate. When considering all studies, irrespective of the inclusion of confirmatory diagnostic (i.e. based on a single serological test, 28 studies), the total number of human cases reported in the literature during the period 2006–2017 was 884 with a national estimated prevalence (calculated according to the sample size between studies) of 3.28% [95%CI 2.52–4.03] (Fig 4). The seroprevalence of infection varied between 0.21% and 9.13% depending on the state. Only two studies (in Jalisco and Colima) were based on a single test and when considering only the studies in which at least 2 serological tests had been performed (26 studies), hence cases had been confirmed as currently recommended by the WHO for an accurate identification of cases, the national estimated seroprevalence was 3.38% [95%CI 2.59–4.16], with seroprevalences varying between 0.21% and 12.01% depending on the state (Fig 5). The highest seroprevalence levels were reported in the states of Jalisco, San Luis Potosi, Chiapas, Estado de Mexico, Queretaro, and Oaxaca. Based on a national population of nearly 120 million (National census of 2015), this seroprevalence level would correspond to 4.06 million cases in the country [95%CI 2.45–4.50 million]. On the other hand, the number of cases of T. cruzi infection reported by the national program of epidemiologic surveillance of the Ministry of Health during 2006–2017 period reached 8,687 ([8] and Table 2), with a regular increase in the number of cases detected with time.
A few of these studies (7 studies) focused on pregnant women, which may transmit the parasite to their newborn during pregnancy [56]. While these studies only covered 7 states (Fig 6), a total of 212 T. cruzi-infected pregnant women were detected, for a global estimated seroprevalence of T. cruzi infection of 2.21% [95%CI 1.46–2.96] in this specific population. The highest seroprevalence levels in pregnant women were reported in the states of Jalisco, Oaxaca, and Estado de Mexico. Based on current birth rate in Mexico (2,293,000 births in 2016), this would correspond to 50,675 births from T. cruzi infected pregnant women per year. With a congenital transmission rate of 6.3% [22], there may be 3,193 cases of congenitally infected newborns per year in the country.
Some studies also focused on or included data on children under 18 years (6 studies), which may indicate more recent transmission. These covered only 4 states (Fig 7), with a global estimated seroprevalence of T. cruzi infection of 1.51% [95%CI 0.77–2.25].
Several additional studies also evaluated T. cruzi infection in blood donors (9 studies), and seropositive human cases were detected in every state of the Mexican Republic, except for Baja California Sur, Sinaloa, and Zacatecas (Fig 8). The total number of blood donor cases reported was 2,300 corresponding to a national estimated seroprevalence of 0.55% [95% CI 0.43–0.66]. The highest seroprevalence was observed in the states of Quintana Roo, Tabasco, Puebla, Campeche and Nayarit.
Finally, we also examined the type of serological test performed in these studies. The most widely used tests were indirect hemagglutination, followed by Chagatest ELISA from Wiener lab, and immunofluorescence assays (Table 3), which represented 67.8% of all tests used. Several other commercial ELISA tests were also used (24.7% of tests), and in-house tests including ELISA, and western blot represented 3.7% of the tests used.
Chagas disease remains one of the most relevant parasitic disease in the Americas, but its epidemiology in Mexico is still poorly understood. Better data are thus urgently needed to help develop appropriate public health programs for disease control and patient care. In this study we analyzed published data on T. cruzi seroprevalence of infection in Mexico between 2006 and 2017. A total of 36 studies were identified, covering most of the country with the notable exception of the state of Michoacán. To take into account the sample size heterogeneity among studies, we analyzed the data based on meta-analysis techniques using a random-effects model. Due to discrepancies in previous studies often attributed to diagnostic methods and uncertainties about the confirmation of cases [57], current recommendations of health agencies request a minimum of 2 serological techniques for accurate diagnostic [10]. Based on this criterion, we found a national estimated seroprevalence of T. cruzi infection of 3.38%, corresponding to 4.06 million cases in the country. Only a few studies were discarded for lack of confirmatory testing, indicating that most of recent studies followed current guidelines for the accurate diagnostic of cases. This seroprevalence level can thus be considered rather conservative, but it is much higher than previous estimates. For example, the Pan American Health Organization (PAHO) estimated that 1,100,000 individuals were infected with T. cruzi in Mexico in 2006, and 29,500,000 were at risk of infection [58], and a national prevalence of 0.65% (with 733,333 cases) was established in 2010 by the Mexican Ministry of Health [59]. The most recent estimates from the WHO based on 2010 data reports 876,458 cases [10], corresponding to a national prevalence of 0.78%. Our analysis of data from the last decade thus suggests that the magnitude of T. cruzi infection in Mexico may have been underestimated in these previous reports. Based on ours data, annual cost for medical care in the outpatient setting was estimated between US$18 and US$39 billion, and annual costs in emergency care unit is between US$27 and US$48 billion [12]. In addition, recent studies pointing out a low sensitivity of commercial serological tests for T. cruzi diagnostic [22,32,60], some of which are widely used in Mexico (Table 3) also raise concerns that the seroprevalence of T. cruzi infection may even be higher than currently detected. Improvements in serological tests are thus urgently needed for a more reliable disease surveillance [61]. Our analysis nonetheless places Mexico as the country with the largest number of cases of T. cruzi infection as previously estimated [10], and highlights the urgency of establishing national priorities for the control of parasite transmission and patient care as well as improved epidemiologic surveillance.
Our results also point out to some regional differences in T. cruzi infection seroprevalence among states. The ecology and epidemiology of Chagas disease are the result of many geographical, ecological, biological, and social interactions [62], which may explain some of these differences. High seroprevalence levels have been previously reported for several states including Jalisco, Chiapas, Queretaro, Oaxaca, Veracruz, and Morelos [11], suggesting a well-established endemicity in these states. States with seroprevalence levels higher than previously reported also emerged through our study, in spite of limited sample sizes. These include San Luis Potosi, Estado de Mexico, Hidalgo and Guanajuato.
T. cruzi infection is also present at a significant estimated seroprevalence in pregnant women in Mexico. Despite the limited information available for this specific population, we could estimate that there are 50,675 births from T. cruzi infected pregnant women per year, corresponding to 3,193 cases of congenitally infected newborns per year in the country. This prevalence is again higher than previous estimates [63], strengthening the urgency of addressing congenital Chagas disease in the country. Because infected newborns can be effectively treated, the lack of specific screening programs to identify them is a missed opportunity for the control of the disease. Indeed, a recent health economic study in the US evidenced the large benefits of maternal screening for T. cruzi infection, as lifetime societal savings due to screening and treatment was estimated at $634 million saved for every birth year cohort [64].
Very limited information is available on T. cruzi infection in children. Nonetheless, we were able to identify a few studies in children up to 18 years, who presented an average prevalence of 1.51%. This may indicate more recent and active transmission compared to data on adult populations, and suggests that the incidence of T. cruzi infection has been fairly stable over time. Therefore, effective vector control programs tailored to the extensive diversity of triatomine species present in Mexico [65] are urgently needed to reduce vectorial T. cruzi transmission to human populations [66,67].
Blood transfusion has been considered the second most important mode of transmission of Chagas disease in Mexico [68]. In 1998, the screening of almost 65,000 blood donors from 18 government-run transfusion centers showed a 1.5% prevalence of anti-T. cruzi antibodies in blood donors [69]. The highest prevalence was detected in the states of Hidalgo, Tlaxcala, Puebla, Chiapas y Yucatan, as expected from previous reports, whereas the northern states of Nuevo Leon and Chihuahua, had the lowest seroprevalence in blood donors. For the period between 1978–2004, Cruz-Reyes et al. defined a national prevalence of positive serology in blood banks of 2.03% [11]. In our study, the national estimated prevalence detected in blood donors was lower with 0.55%. These differences can be explained by the increased reliability of serologic screening of blood donors with the passing of legislation making screening mandatory in the year 2000 [13,70]. The addition of a pre-screening questionnaire to exclude high-risk individuals may also have led to a lower prevalence in screened donors. The highest prevalence of 1.99% is detected in the state of Quintana Roo and the lowest in Baja California Sur, Sinaloa and Zacatecas (with a prevalence of 0%).
A major strength of our analysis was to consider the sample size heterogeneity among studies and the reliability of serological testing performed, and to ensure it followed WHO recommendations for confirmation of cases using at least a second test. Hence, our estimates of seroprevalence are robust and conservative. On the other hand, there are some limitations. First, some heterogeneity among study designs and particularly sampling strategies and recruitment of subjects may have generated some bias. For example, the difference in the number of studies performed per state can lead to over- or under-estimation of the prevalences. Also, while we did not detect major publication bias, there was an uneven coverage of the different states by research studies, which may be a confounding factor affecting differences in T. cruzi infection prevalence among states. This highlights the need for much improved nationwide disease surveillance to clearly identify geographic heterogeneities in T. cruzi transmission and Chagas disease epidemiology. Finally, the small number of studies/sample sizes for some of the subgroup analysis also add uncertainties to our estimates of T. cruzi estimated seroprevalence in specific subpopulations.
In conclusion, our systematic review and meta-analysis estimates a national seroprevalence of T. cruzi infection of 3.38%, with 4.06 million cases in Mexico, which is higher than previously recognized. It places Mexico as the country with the largest number of cases, highlighting the urgency of establishing Chagas disease control as a key national public health priority, to ensure that it does not remain a major barrier to the economic and social development of Mexico's most vulnerable populations. It remains essential to strengthen effective surveillance for Chagas disease in all the country to obtain more precise data. The presence of T. cruzi infection in specific subpopulations such as pregnant women, children and blood donors also informs on specific risks of infection, and calls for the implementation of well-established control interventions [56,67,71]. Finally, while our estimates are conservative and based on confirmed cases, the lack of sensitivity of current serological tests observed in Mexico suggest that the true magnitude of Chagas disease in the country may still be underestimated, and the development of more reliable diagnostic tests will be key for an effective identification of cases as well as improved patient care [61].
|
10.1371/journal.pcbi.1005605 | A new index for characterizing micro-bead motion in a flow induced by ciliary beating: Part I, experimental analysis | Mucociliary clearance is one of the major lines of defense of the respiratory system. The mucus layer coating the pulmonary airways is moved along and out of the lung by the activity of motile cilia, thus expelling the particles trapped in it. Here we compare ex vivo measurements of a Newtonian flow induced by cilia beating (using micro-beads as tracers) and a mathematical model of this fluid flow, presented in greater detail in a second companion article. Samples of nasal epithelial cells placed in water are recorded by high-speed video-microscopy and ciliary beat pattern is inferred. Automatic tracking of micro-beads, used as markers of the flow generated by cilia motion, enables us also to assess the velocity profile as a function of the distance above the cilia. This profile is shown to be essentially parabolic. The obtained experimental data are used to feed a 2D mathematical and numerical model of the coupling between cilia, fluid, and micro-bead motion. From the model and the experimental measurements, the shear stress exerted by the cilia is deduced. Finally, this shear stress, which can easily be measured in the clinical setting, is proposed as a new index for characterizing the efficiency of ciliary beating.
| Mucociliary clearance is the first line of defense of the human pulmonary airways. Mucus transporting debris, particles, microorganisms and pollutants is carried away by the coordinated motion of cilia beating at the surface of the airway epithelium. We present here an experimental, mathematical and numerical study aiming at defining a global index for assessing the efficiency of this beating. We measure experimentally the ciliary beat frequency, ciliary beat amplitude, and metachronal wavelength on ciliated edges obtained from nasal brushing. Properties of fluid motion are simultaneously extracted from micro-bead tracking next to the ciliated edge. A mathematical and numerical model is developed to describe the fluid motion induced by the cilia tips considered as a moving wall. Experimental and numerical results show that the bead velocity is a parabolic function of the distance to the wall. It allows us to infer the shear stress exerted by the cilia on fluid from micro-bead tracking. This quantity is proposed as a universal index characterizing the beating efficiency, which can be extracted in the current clinical setting.
| Mucociliary clearance is one of the major defense mechanisms of the respiratory airway system. The mucus layer coating the epithelial surface of the airways filters the inhaled air by trapping potentially harmful material (fungi, bacteria and other particles) [1–4]. This mucus layer is continuously carried away and out of the airways by the activity of motile cilia. Neighboring cilia beat in an organized manner with a small phase lag, their tips creating an undulating surface on top of the cilia layer which deforms in a wave-like fashion called the metachronal wave [5–7].
The beat pattern of an individual cilium displays a two-stroke effective-recovery motion [8]. During the effective stroke, cilia beat forwards and engage with the mucous layer, propelling it forward. In contrast, during the recovery stroke, they return to their initial position in the underlying periciliary fluid, minimizing thereby the drag on the mucus in the opposite direction. This asymmetry in the beat pattern is responsible for a net fluid flow in the direction of the effective stroke. In the airways, each mature ciliated cell may be covered with up to 200 cilia, with a surface density around 5–8 cilia/μm2 [6, 9]. Each cilium, approximately 6 μm long and of diameter around 0.2 μm, beats 12 to 15 times per second, resulting in a velocity of the mucus layer of approximately 1 mm per minute [10].
Defects in mucociliary clearance may result in chronic airway inflammation and infections causing injury and structural changes to the airway epithelium, leading to a variety of diseases like bronchiectasis and chronic sinusitis. Two main reasons may lead to impaired mucociliary clearance. The first one is related to alterations of the mucus properties, as in cystic fibrosis where the mucus become too thick and sticky to be moved properly by the cilia. The second one is linked to a dysfunction in ciliary motion or ciliary coordination. Such dysfunctions may be either inherited as in Primary Ciliary Dyskinesia (PCD) or acquired. Diagnosis of dyskinesia is difficult. For instance, in the diagnosis of PCD, a combination of transmission electronic microscopy, nasal nitric NO, genetic testing and use of high-speed video-microscopy is recommended [11].
In situ observation of ciliary beating and mucociliary clearance is almost impossible in patients at present stage, and one currently lacks a reliable and general method for evaluating mucociliary clearance in the clinical field. In the past, integrated assessment of mucociliary clearance was achieved through techniques using saccharine [12], a drop of blue marker [13], or clearance of radioactive tracers [14–17]. More recently, micro-optical coherence tomography was also proposed [18]. However, due to their various requirements (patient cooperation for the saccharine test, endoscopic examination, inhalation of radiopharmaceutical, …) these different methods are rarely used in the current clinical practice. Light microscopy observation of ciliated edge obtained by nasal or bronchial brushing is the most common method used to evaluate ciliary beating. This observation is often associated with High-Speed Video-Microscopy (HSVM) analysis [19–23] that quantifies only cilia motion and ciliary beat pattern. However, none of these tests provides information about the global efficiency of ciliary beating regarding mucociliary clearance.
This present study aims at a deeper understanding of the relationship between ciliary beat pattern, measured by ciliary beating analysis, and the induced motion of the surrounding fluid in order to assess the global efficiency of ciliary beating. An original method of micro-bead tracking (MBT) is presented where micro-beads act as markers of the flow generated by the cilia [24]. A numerical model is then developed (presented in detail in a companion paper [25]) based on envelope modeling approach, allowing us to simulate the flow generated by the cilia as in the MBT experiment. We finally propose a new and global index for characterizing the efficiency of the ciliary beating. One very appealing aspect of this index is that it does not require any modification of the present clinical practice of data collection (nasal or bronchial brushing).
We present here a brief overview of the two-dimensional model describing the system composed of the cilia, the surrounding fluid, and the micro-beads. The entire model is extensively detailed in [25]. Inspired by [27], we propose an approach in which the momentum transfer from discrete cilia to the fluid (induced by the ciliary beating) is modeled through an appropriate continuous boundary condition. The ciliated edge is chosen parallel to the x axis, the ciliary tip being located around y = 0 on average. Each cilium is assumed to undergo a periodic elliptic motion (see Fig 5, left). Taking the limit of a continuous cilia distribution, the cilia array is simplified as an undulating surface that covers the cilia layer, ignoring the details of the sub-layer dynamics, see Fig 5, right (inspired by Velez-Cordero et al. [28]).
The tip of a cilium located at the horizontal coordinate ξ* is assumed to follow a periodic elliptic trajectory centered in (ξ*, 0) during each elementary beat (Fig 6). The ‘*’ notation is used here to represent dimensional quantities, for consistency reason with [25]. At time t*, the tip coordinates ( X w * , Y w * ) satisfy
{ Xw*=ξ*−a cos(ωt*)Yw*=βa sin(ωt*) (2)
where β is a function of the ellipse eccentricity, 2βa is its minor axis in the y* direction, and 2a its major axis in the x* direction. For β > 0, the tip orbits clockwise, while for β < 0, the tip orbits counterclockwise.
In this picture, the metachronal wave is materialized by prescribing the motion of the envelope. This envelope is then used as a boundary condition to compute the fluid velocity field dominated by viscous forces, inside a channel of height h (see Fig 6). The fluid is assumed to be stagnant above this height. Indeed, the parameter h essentially summarizes in our 2D model (see companion paper) the effect of the external environment of the clump, this environment imposing a limit to the spatial extension of the cilia-induced flow field. This fact explains why this parameter h cannot be estimated a priori and has to be measured as an external free parameter.
The trajectories of micro-beads in this velocity field are calculated by solving the equation of motion with Stokes drag. This theoretical model provides us with a prediction of the effective speed of micro-beads as a function of their average altitude above the ciliated edge. This velocity profile is found to follow essentially a parabolic profile as a function of the height above the cilia wall. Finally, the velocity of the micro-beads extrapolated at the cilia wall gives a direct estimate of the shear stress exerted on the fluid, this index being proposed as an index of the ciliary beating efficiency [25].
78 movies were recorded, corresponding to a total of 24 ciliated edges. From these movies, the trajectories of 195 micro-beads were retained. Three examples of MBT movie are displayed online (see supporting information).
Micro-beads velocities are essentially oriented along the x* direction (parallel to the ciliated edge), as can be seen in Fig 7. Only 4% of the micro-beads exhibit a vertical velocity component larger than 25% of the horizontal component. Moreover, this vertical component is about equally distributed among positive and negative values (98 vs 97). This result suggests that the bead velocities can be modeled in a good approximation as parallel to the ciliated edge.
Micro-bead velocities range from 0.0 to 253.8 μm.s-1 (mean = 42.2 μm.s-1, std = 35.0 μm.s-1). This wide variation is explained by the spread of the distances of the micro-beads to the ciliated edge which range from 0.3 to 70.9 μm (mean = 12.7 μm; std = 11.1 μm). Indeed, as observed in Fig 8, left which presents micro-bead velocity measurements on three different ciliated edges, velocities appear to be strongly correlated to these distances, the fastest micro-beads being the ones closest to the ciliated edge. To confirm this observation, velocities of all micro-beads were separated into two groups, respectively above and below to the median value of the distance (Fig 8, right). The comparison between the two groups was performed with a statistical software package using non-parametric tests (Mann-Whitney U-test). A p value <0.05 was considered significant. Here again, one can distinctly observe a clear link between the value of the micro-bead velocity and its distance to the ciliated edge.
In order to assess the parameters that influence significantly the micro-bead velocity, a multiple linear regression analysis was performed between the values of the micro-bead velocities (the dependent variable in the regression) and five exploratory variables, 4 of which are determined from measurements of ciliary beating (CBF, CBA, λ, ρc), and the fifth being the distance y0 to the cilia. The resulting multiple regression equation of micro-bead velocities is given by:
V eff ( y 0 ) = ( 6 . 96 × CBF ) + ( 230 . 6 × ρ c ) + ( 11 . 95 × CBA ) - ( 1 . 22 × y 0 ) - ( 0 . 83 × λ ) - 243 . 35 , (6)
with a coefficient of determination equal to 0.68 and a probability smaller than 10−6 (F-test). Moreover, all probabilities associated to regression coefficients (t-test) are smaller than 10−6, except for the wavelength (p = 0.048). One can remark that this regression implicitly assumes a linear dependency on the exploratory variables, while a further and finer analysis using our mathematical model reveals a more complex behavior, as for instance the micro-bead velocity profile above the edge which is shown in fact to be essentially parabolic.
The shear stress τw exerted by the cilia wall on the fluid is given by (see [25]):
τ w = 2 μ h U w * (9)
where U w * is the extrapolated velocity at the ciliated wall, h is the horizontal intercept (vanishing velocity) of the fitted parabolic profile, and μ is the dynamic viscosity of the fluid. We propose to use this force τw as an index for assessing the efficiency of the ciliary beating. MBT measurements enable us to retrieve U w * and h, and therefore to compute the shear stress.
In Table 1 we report the results of the analysis performed on 11 patients. For each patient, one or several ciliated edges have been analyzed. For each edge, the number of tracked micro-beads is reported, followed by measured values of the input parameters of the model. The last 3 columns show the values of h and U w *, retrieved from the parabolic profiles, and the value of the deduced shear stress.
The experimental MBT results (see Fig 10, left and right), allowed us to validate the numerical model. The model predicts that the micro-bead velocity is an increasing function of CBA and CBF while it decreases with λ and ϕ (the last dependency means that micro-bead velocity increases with ρc). These behaviors are consistent with the various signs of the regression coefficients of Eq 6 deduced from MBT experiments.
The good agreement observed between the numerical model and experimental MBT results led us to propose the steady component of the shear stress exerted by the cilia wall on the fluid as a new index of the ciliary beating efficiency. To our knowledge, there exist very few evaluations of the shear stress exerted by the respiratory cilia in the literature. If we assume a cilia density per unit area of 5 cilia/μm2 [9], we find a force per cilium equal to 0.013±0.009 pN. Such a value is much lower (at least 3 orders of magnitude) than the ones reported in a study on human bronchial epithelial cell culture [9], or in a study on culture grown from frog esophagus [33]. However, in these two last cases, the measured quantity was an oscillating force and not the steady component resulting from the entire beating as in our study. This suggests that these two types of measurement should not be directly compared. Moreover, the shear stress we find is of the same order of magnitude than the shear stress induced by the so-called Couette flow between two parallel flat plates:
τ Couette = μ e V , (10)
where e and V are the distance and the relative velocity between the two plates, respectively. If we put on a par e with h, and V with U w *, respectively, one finds the formula for the two stresses to be very similar. The factor 2 observed in our mode originates from the difference of velocity profiles (linear for a Couette flow, parabolic in our case).
The parameter β, i.e., the ratio between minor and major axis is very difficult to measure on a large part of our movies. We decided to set β at 0.14 for all simulations as it was the mean value observed on our movies. The measured values of all other parameters obtained by ciliary analysis in this study were consistent with the literature. As an example, our values of metachronal wavelength fall in the same range than the few numbers that can be found in the literature for other cellular models (paramecium [34, 35], frog oesophagius [36] or rabbit trachea [5]).
In contrast with other analyses of the ciliary beating [20], or evaluations of the beat pattern [19], τw provides a direct estimation of the force that is potentially applied by the ciliated epithelium on the surrounding fluid, hence on the mucus. As such, it appears as a very good candidate for a global index of the potential ciliary beat efficiency. It encompasses the beating of the ciliated edge as a whole rather than focusing on an individual cilium. In comparable experimental conditions (here the cell survival medium at room temperature), τw allows us to compare the ciliary beat efficiency of different patients with each other. Moreover, this index that can be easily obtained from a MBT experiment via a simple parabolic fitting requires neither a subjective interpretation by an operator, nor a specific procedure in terms of human data sampling, but only a classical nasal (or bronchial) brushing.
Low values of τw, corresponding to a low ciliary beat efficiency can be explained by several distinct causes: ciliary beating parameters alterations (CBF, CBA, metachronal wavelength), reduced cilia density, or a loss of coordination between cilia. Indeed our model assumes, via the metachronal wave, a perfect cilia coordination as well as a pure rectilinear geometry of the edge. As a consequence, high values of τw are expected to correspond to well coordinated cilia without degradation of ciliary beating parameters (CBF, CBA, metachronal wavelength). However, for low values of τw, a ciliary beat pattern analysis may be required to find the reason of a degraded efficiency (ciliary beating parameters, loss of coordination, …).
Table 1 displays values of τw measured in several patients. These values exhibit a relatively high intra-patient heterogeneity which is not surprising as it was already observed in other characteristics of the ciliary beating, such as the “the distance traveled by the cilium tip in one second weighted by the percentage of beating ciliated edges” [20]. Despite its intra-patient variability, this distance is able to discriminate non-PCD from PCD patients with a specificity and sensitivity above 0.95 [20]. Similarly, using τw as a screening index in clinical studies would probably require analyzing several edges per patient.
Interestingly, the patient exhibiting the highest values of τw (patient №1, see Table 1) seems to be also the patient with the most normal clinical state. To this date, it is difficult to pinpoint a precise threshold of τw that would allow one to discriminate between clinically healthy and pathological edges. Determining this threshold would require a clinical study including a larger cohort of control patients compared to patients with well defined pathological ciliary beating phenotype, a study far beyond the scope of the present study.
Ciliary beating parameters (CBF, CBA, pattern, …) depend on the characteristics of the fluid at least in part. In our MBT experiments, cilia are beating in a fluid whose physical properties are very similar to water. Such conditions are far from the real conditions of airways coated with mucus. One may wonder how the modification of the surrounding fluid influences cilia beating. Several numerical studies have investigated the way viscosity influences ciliary beating and the difference between a Newtonian vs. non Newtonian surrounding fluid. Jayathilake et al. have explored numerically the effect of an increased PCL viscosity on the motion of cilia embedded, and have found that, for a given beating frequency (set around 10 Hz, consistent with our measurements), the velocities of the induced PCL flow were almost unaffected by a 5-fold increase of the viscosity [37]. Other models taking into account the thermodynamic characteristic of the mucus (generalized Newtonian fluid, …) can be found in the literature [28]. More recently, Sedaghat et al. have studied in detail the two layer structure (PCL and mucus), PCL being modeled as a Newtonian fluid and mucus as a viscoelastic fluid [38]. Their study showed firstly that mucus viscosity has a very limited effect on the mucus flow, due to the dominating part of the elastic part in the viscosity, and secondly that ciliary beating frequency plays a major role in the mucus velocity. These results were confirmed by very recent 3D numerical simulations [39]. In these works, the beating frequency was identified as a major determinant of the mucus flow, before the mucus viscosity itself. However, in all simulations, this frequency is always a prescribed parameter whereas in reality it results from a delicate force and momentum balance in the fluid-structure interaction. As a consequence, it is very complicated to really assess the net influence of mucus rheological properties on cilia beating, considering the extreme difficulty to reproduce in vitro the in vivo situation. The essential point here lies in the value of the beating frequency, which appears to be very similar between in vivo data from the literature and our ex vivo experiments.
Replacing the surrounding fluid with water may also induce difference in biochemical interactions. The mucociliary clearance process is generally described with three main components: mucins secreted by goblet that will give mucus, cilia that move the mucus and an ion transport process allowing maintaining an adequate aqueous environment on the airway epithelium [40]. These components can be, at least partially, controlled by local agonist, extracellular nucleotides, and nucleosides released from the epithelium. For example, it is known that ATP, UTP, and adenosine increase the ciliary beat frequency while ATP and UTP stimulate the secretion of mucins.
However, to our knowledge, the relation between these local agonists and the cilia beat pattern remains to be established. In 2000 Chilvers et al. proposed to use digital high speed video microscopy to visualize the beat pattern of human nasal cilia in the absence of mucus [19]. Since 2009, the European Respiratory Society strongly recommended to combine this technique (i.e., in the absence of mucus) with classical tests in order to ensure primary ciliary dyskinesia diagnosis [23, 41]. At this time, studying cilia beat pattern in the presence of mucus for patient specific evaluation would be very difficult to carry out because it would require tracheal or bronchial explants which are ethically problematic in clinical practice.
Clearly, the experimental procedure described in this study is not intended at evaluating the effect of possible alterations of the cilia environment (change in rheological properties of the mucus, modification of the nucleotides in the periciliary layer…). Our MBT experiments, the developed numerical model and the proposed index aim essentially at evaluating pathologies impairing mucociliary clearance resulting from a defect in ciliary motion (ciliopathies). In future studies, one could even envision adding exogenous drug to the survival medium to test the potential effect of a drug treatment. For now, the goal is to simulate the conditions of our experiment corresponding to a measurement that can be realized in a clinical setting.
We have presented here a first experimental validation of the numerical model introduced in [25], describing the Newtonian fluid flow induced by cilia motion above a ciliated edge. This study shows that the computational model satisfactorily predicts the profile of velocities of micro-beads in the Newtonian flow, this profile being essentially parabolic. Recovering this profile allows us to assess the shear stress locally applied by the ciliated edge onto the fluid. Our model suggests that this shear stress at the cilia wall characterizes the momentum transfer between the cilia and the fluid, and thus the efficiency of the ciliary beating. Interestingly, the estimation of this index does not require any modification of the present clinical practice of data collection (nasal or bronchial brushing). This study opens the broad perspective of using this index in the future to characterize ciliary function of transport during normal and pathological conditions of either congenital origin such as PCD or acquired dyskinesia secondary to pathogenic invasion and/or occupational exposures.
|
10.1371/journal.ppat.1005934 | The Replicative Consequences of Papillomavirus E2 Protein Binding to the Origin Replication Factor ORC2 | The origin recognition complex (ORC) coordinates a series of events that lead to initiation of DNA strand duplication. As a nuclear double stranded DNA plasmid, the papillomavirus (PV) genome resembles a mini-chromosome in infected cells. To initiate its replication, the viral E2 protein binds to and recruits the E1 DNA helicase at the viral origin. PV genome replication program exhibits three stages: initial amplification from a single genome upon infection to a few copies per cell, a cell cycle linked maintenance phase, and a differentiation dependent late stage where the genome is amplified to thousands of copies. Involvement of ORC or other pre-replication complex (pre-RC) factors has not been described. We report that human PV (HPV) and bovine PV (BPV-1) E2 proteins bind to ORC2, however, ORC2 was not detected at the viral origin. Depletion of ORC2 enhanced PV replication in a transient replication model and in keratinocytes stably maintaining viral episomes, while there was no effect on copy number in a cell line with integrated HPV genomes. Consistent with this, occupancy of E1 and E2 at the viral origin increased following ORC2 silencing. These data imply that ORC2 is not necessary for activation of the PV origin by E1 and E2 but instead suppresses E2 replicative function. Furthermore, we observed that over-expression of HPV E2 decreased ORC2 occupation at two known mammalian origins of replication, suggesting that E2 restricts pre-ORC assembly that could otherwise compete for host replication complexes necessary for viral genome amplification. We infer that the ORC2 complex with E2 restricts viral replication in the maintenance phase of the viral replication program and that elevated levels of E2 that occur during the differentiation dependent amplification stage subvert ORC loading and hence DNA synthesis at cellular origins.
| Papillomavirus genome replication occurs during three distinct stages that are linked to the differentiation state of the infected epithelium. The viral proteins E1 and E2 recognize the viral origin and initiate a process that attracts host DNA replication factors. The origin recognition complex (ORC) coordinates initiation of chromosome duplication. While ORC2 binds to the E2 protein, its depletion does not impair PV genome replication. Instead, depletion of ORC2 stimulates viral replication, while over-expression of E2 protein decreases ORC2 occupancy at mammalian origins. We propose that the relative abundance of E2 and ORC2 in complex regulates viral and cellular origin licensing.
| Papillomaviruses (PV) are medically important pathogens especially as specific genotypes carry a high risk of progression to cancer, most commonly of the uterine cervix and oropharynx. Because PVs have limited protein coding capacity in their typically 8 kilobases (kb) genome, these viruses do not encode a DNA polymerase and must rely on host DNA replication factors. The viral genome replicates and is maintained as circular covalently closed double stranded, histone coated DNA plasmids in infected cells, thus resembling multi-copy mini-chromosomes. The viral genome replicative program consists of three stages [1, 2]. Upon virus infection, its genome enters the nucleus of basal level epithelial cells and establishes a low copy number (1 to perhaps 50). In the second ‘maintenance’ stage, these episomes duplicate as host epithelial cells replicate and depart the basal cell and suprabasal compartments [3, 4]. Monolayer keratinocyte cultures that harbor viral episomes reflect this stage of virus replication. During this stage, the autonomous viral genomes segregate in mitosis as a kinetochore independent mini-chromosome. E2 protein binding to ChlR1 and Brd4 was shown to mediate attachment of the viral DNA to host chromosomes that is necessary for mitotic partitioning and nuclear retention of viral episomes [5, 6]. The third ‘amplification’ stage occurs in upper epithelial strata where non-dividing epithelial cells persist in a prolonged S/G2 phase [7]. In these cells, the viral episomes replicate to hundreds of episomes that are packaged into nascent virion particles.
Many of our insights into PV replication proteins emerged from studies of bovine papillomavirus type-1 (BPV), which is maintained as a stable replicating episome in murine NIH3T3 and C127 cell lines. Its E2 protein is composed of an N-terminal 220 amino acid transactivation domain (TAD), a non-conserved hinge region, and a C-terminal dimerization and DNA binding domain [8]. The TAD mediates interactions with several cellular proteins necessary for transcriptional activation and replication such as Brd4, TaxBP1, and GPS2/AMF-1 [6, 9–11]. The E2 protein binds with high affinity to an inverted palindromic sequences present in all PVs, which serves to regulate viral transcription and replication [12]. E2 binds to and recruits E1, an ATP dependent replicative helicase, to these E2-binding motifs [13]. Together with an adjacent E1 binding site and short polyA tract, these composite sequences define and function as the origin of replication (ori) [13, 14]. Antibodies to BPV E1 and E2 have been used in ChIP experiments to document localization of E2 protein to this region in G1/S phase [15], however, analogous reagents for the HPV protein counterparts have not been available. Assembly of E1 into double hexamers requires release from E2, which has been reported to be a consequence of cyclin dependent kinase mediated phosphorylation of E1 [16, 17]. Mutations in the viral E1 or E2 genes result in integration or loss of the viral genome [18].
Eukaryotic origin licensing encompasses a highly orchestrated series of precise steps that limits firing to once per cell cycle [19]. DNA synthesis begins when the seven-subunit origin recognition complex (ORC) assembles on DNA. In contrast to the ~30 base pair (bp) PV origin, the sequence requirements for metazoan replication origins definition are poorly defined, and these may span chromosomal regions of perhaps 100,000 kb [20–23]. Mammalian replication initiates at the predicted 104–105 origins in a processive cascade that is coordinated by loading of the pre-replication complex (pre-RC) onto chromatin [24–26]. The experiments performed herein include chromosomal regions that are recognized to contain potential origins for comparisons to their PV counterparts operative in the same cell. The ORC1 protein is the only ORC member to contain a bromo-adjacent homology domain (BAH) necessary for interaction with DNA replication origins [27]. ORC1 is bound to DNA during G1 and released during S-phase and DNA synthesis, while ORC proteins 2–6 may be attached to DNA throughout the majority of cell cycle. ORC2 association with chromatin is regulated by cell cycle dependent phosphorylations at threonines 116 and 226, which lead to its dissociation from chromatin during the transition from S to M while still in complex with the other ORC proteins [28]. After ORC binding to chromatin in late M through G1, the multiple mini-chromosome maintenance complex (Mcm2-7) DNA helicase is loaded onto chromatin following binding of Cdc6 to ORC.
Very little is known of which host replication factors are required during each stage of the viral replicative program. We describe here that BPV and HPV E2 proteins bind to ORC2, however, this pre-RC factor is not present at the viral replicon and its depletion did not suppress viral replication but instead increased the presence of E1 and E2 at the viral origin. This led to the hypothesis that E2 association with ORC2 regulates viral and cellular origin loading such that at low levels of E2, its complex with ORC2 suppresses viral replication. When high levels of E2 are achieved in the differentiation dependent viral genome amplification stage, E2 restricts cellular origin recognition by ORC, which would otherwise complete for factors necessary for viral DNA synthesis. These results illustrate a novel mechanism by which PV E2 orchestrates the viral genome replication program in a stratifying epithelial environment.
The current study began as we questioned whether ORC proteins assemble at the PV origin. While E1 and E2 proteins are sufficient to activate the PV ori, it is possible that a pre-ORC factor also participates in viral ori licensing. For example, the Epstein-Barr Virus (EBV) EBNA-1 protein binds to and recruits ORC2 to oriP [29]. This leads to entry of ORC2-4 along with the helicase component MCM2 [30]. E2 has structural and functional similarities to EBNA-1 [31, 32], however EBV does not utilize a viral DNA helicase that resembles E1 at oriP. We initiated this survey using immunoprecipitation of the BPV-1 E2 protein followed by immunoblotting for the pre-RC mammalian factors ORC1 and ORC2. C33A cells were transfected with plasmids expressing BPV-1 E2 or the truncated E2R that lacks the 160 initial amino acids of the TAD [33] along with mouse FLAG-ORC1 or FLAG-ORC2. Full length BPV-1 E2 but not E2R co-immunoprecipitated ORC2 protein but not ORC1 (Fig 1A). Next we determined whether the TAD was sufficient for association with ORC2. BPV-1 E2 full length, TAD, and E2R along with mouse ORC2 were expressed in C33A cells. Since full length BPV-1 E2 increases transcription of most co-transfected genes [34], cells were treated with 10 μM MG132 for 6 hr except for those expressing full length E2 to achieve comparable ORC2 protein levels among all the experimental groups. Both E2 and its TAD individually co-immunoprecipitated ORC2, whereas the E2R form did not (Fig 1B). In analogous experiments, HA-tagged human ORC2 (HA-hORC2) co-immunoprecipitated with FLAG-HPV-31 E2 (Fig 1C) but not with the HPV-31 E1 protein (Fig 1D). The presence of HPV-31 E1 did not interfere with ORC2 pull down with HPV-31 E2 (Fig 1D). To diminish the possibility that other mammalian factors bridge E2 to ORC2, baculovirus lysates containing human ORC2 protein were combined with bacterially expressed 6-histidine tagged BPV-1 E2 TAD or control HPV-16 E6 proteins. Nickel bead collection of these his-tagged proteins showed that ORC2 specifically interacted with the E2 TAD (Fig 1E), implying that in the absence of other mammalian replication factors, ORC2 may bind directly to E2. Subsequently we confirmed the protein-protein interaction between E2 and ORC2 in vivo using in-situ proximity ligation assays (PLA). As a positive control, we observed fluorescent puncta in cells transfected with FLAG-HPV-31 E2 and HA-HPV-31 E1 (Fig 2B). C33A cells transfected with FLAG-HPV-31 E2 and HA-hORC2 showed fluorescent spots compared to untransfected cells (Fig 2A), indicating association between two proteins in vivo. We also detected PLA interaction between FLAG-HPV-31 E2 and endogenous levels of ORC2 (Fig 2C).
Because E2 engages the viral ori, the expectation was that ORC2 would be present near the PV ori or the flanking sequences of the long control region (LCR). After confirming that commercial ORC2 antibodies immunoprecipitate the native ORC2 protein using baculovirus expressed human ORC2 lysate (Fig 3A), we performed a series of chromatin immunoprecipitation (ChIP) experiments using HPV-BP and CIN612-9E cells, which maintain low copy HPV-16 and HPV-31 episomes respectively. These cell lines, expressing E1 and E2 from native viral promoters, were synchronized to G1/S with double thymidine to capture putative ORC complexes at or near the PV ori. Repeated ChIP assays failed to detect endogenous ORC2 near the HPV-16 or HPV-31 LCR, while ORC2 was detectable near the mammalian origin of replication present within the GM-CSF gene (Fig 3B and 3C). A 6 kb region upstream of the MCM4 upper regulatory region (Exon 9) and the associated primers, previously reported to be free of ORC2 [35], served as a negative control. Commercially available ORC2 antibodies do not have high affinity for mouse ORC2, prohibiting analogous ChIP experiments in C127 cell lines that stably replicate BPV-1 episomes. Because ORC2 was not present near the viral ori, other regions within and flanking the LCR were probed for the presence of ORC2 and compared to antibodies directed against the viral E1 or E2 proteins as positive controls. ChIP was performed for HPV-31 E2 and ORC2 using lysates from CIN612-9E cells enriched in G1/S with double thymidine. In these experiments, we used an anti-peptide antibody to HPV-31 E2 (epitope aa 97–109; Fig 4A) and HPV-16 E1 antibodies confirmed to cross-react with the highly conserved HPV-31 E1 protein (Fig 4C). Target DNA immunoprecipitation with HPV-31 E2 was at least 15 fold greater than with negative control EE antibodies, while ORC2 could not be detected at these regions (Fig 4B). HPV-31 E2 enrichment was greatest near the region of the E2 binding sites between the LCR2 and LCR4 primer sets. To perform analogous experiments with a second HPV genotype, we used W12 cells that contain HPV-16 episomes. Consistent with our previous results, these ChIP reactions demonstrated the presence of HPV-16 E1 at the LCR while ORC2 was not detected at this location in the HPV genome (Fig 4D).
Previous studies observed that E1 and E2 co-expression leads to formation of nuclear foci in CV-1, C33A, and primary human foreskin keratinocyte (HFK) cells in the presence and/or absence of the HPV origin [8, 36–38]. These are considered to be ‘replication centers’ where E2 and E1 assemble on viral genomes. Several DNA repair proteins have been shown to accumulate at these foci, including ɣH2AX, pATM, pChk2, NBS1, and Rad51 [39, 40]. While the ChIP data implied ORC2 does not directly participate in viral ori licensing, we questioned whether it was present at these nuclear foci. CV-1 cells were transfected with expression vectors for HPV-31 E1, HPV-31 E2, and YFP-ORC2 with or without an HPV-31 ori containing plasmid. HPV-31 E2 and YFP-ORC2 expression alone are both diffusely nuclear. YFP-ORC2 did not co-localize with E1 and E2 nuclear foci in the presence (S1A–S1C Fig) or absence of the HPV-31 origin (S1D–S1F Fig). We also confirmed these observations using BPV E1 and E2 proteins, which were detected in discrete nuclear spots, while ORC2 was dispersed throughout the nucleus (S2 Fig). This observation is consistent with our ORC2 co-IP data in which ORC2 interacted with E2 but E1 was not in this complex (Fig 1D).
Because ORC2 co-immunoprecipitated with E2 yet was not detectable at the viral ori or at viral replication foci, this inferred that ORC2 is not necessary for viral replicon licensing. To provide experimental evidence for this, we performed transient replication assays in which firefly luciferase activity is proportional to episomal copy number as corroborated by PCR for viral DNA content [41–45]. Further confirmation of this assay is sensitivity to DNA replication inhibitors [46]. For the analyses, we normalized firefly luciferase levels to co-transfected renilla levels to control for non-specific transcriptional effects. Due to relatively high levels of E1 and E2 protein expression, this experimental transient transfection model likely reflects the amplification stage of viral genome replication [47]. ORC2 depletion resulted in increased replication of the luciferase-linked HPV-31 ori plasmid driven by co-transfected E1 and E2 (Fig 5A). ORC2 knockdown did not change luciferase levels when the ori was transfected with E1 or E2 alone (S3A Fig). Using BPV-1 E1 and E2 with a luciferase-reporter ori plasmid, ORC2 knockdown increased BPV-1 replication though not to same extent as with HPV-31 (S3B Fig). The non-specific transcriptional activity of BPV-1 E2 may contribute to its higher basal luciferase levels [48]. The ORC2 shRNA hairpin had no effect on luciferase activity with co-expression of replication incompetent E2R with BPV-1 E1.
We also investigated the consequences of ORC2 depletion on the HPV replication at endogenous E1 and E2 protein levels maintained in monolayer cultures by co-transfecting the ORC2 shRNA together with the HPV-31 ori reporter plasmid into CIN612 cell lines. These conditions represent the maintenance phase of the viral replication program. The reporter plasmid system was necessitated because direct measurement of viral copy number by quantitative PCR would include the large number of CIN612-9E cells that were not transfected with the shRNA. Consistent with our previous studies, ORC2 shRNA induced increased luciferase production and displayed a corresponding reduction of ORC2 protein (S3C Fig). for comparison, we investigated the consequences of ORC2 depletion on the pREP4-EBNA1, an EBV based episomal oriP plasmid containing the EF1α promoter and luciferase reporter gene. Transfection of the ORC2 shRNA suppressed EBNA1 induced replication (Fig 5B), consistent with the observation that EBV replication is ORC2 dependent [29]. As ORC2 shRNA knockdown had the opposite effect on luciferase levels with the EBV reporter compared to the HPV reporter, we concluded that the increased luciferase levels in the HPV system was not due to a direct effect of ORC2 shRNA on luciferase protein expression but reflective of PV copy number. In more efficiently transfected HPV-BP cells, knockdown of ORC2 showed a ten fold increase in HPV-16 DNA episome number measured by quantitative PCR with primers specific to the HPV-16 LCR (Fig 5C and 5D). In contrast, ORC2 knockdown did not produce a significant change in HPV-16 DNA content in SiHa cells, which contain a single integrated copy of the HPV-16 genome. While it might be argued that ORC2 knockdown might change luciferase mRNA expression or stability, these experiments demonstrated the same effects on cells with endogenous HPV genomes.
The decrement in ORC2 protein levels following ORC2 shRNA was relatively small, as we assumed that a dramatic reduction in ORC2 levels would be detrimental to cell proliferation. Nonetheless, an ORC2 siRNA duplex was obtained to attempt more effective reduce the ORC2 expression in CIN612-9E cells. The ORC2 siRNA caused a more dramatic reduction of ORC2 levels compared to the previously used ORC2 shRNA at 48 h (Fig 6A). ORC2 silencing had no effect on the cell cycle profile at 48 h (S4A Fig). We then tested the consequences of ORC2 depletion on HPV-31 replication in a model of replication maintenance. First, CIN612 cells were transfected with the HPV-31 ori reporter plasmid in the absence or presence of ORC2 siRNA. As seen in Fig 6B, ORC2 depletion enhanced luciferase levels at 48h. In the following series of experiments, we evaluated the effects of ORC2 depletion on viral replication during keratinocyte differentiation [49, 50]. CIN612 cells expressing endogenous levels of HPV-31 E1 and E2 were transfected with the HPV-31 ori reporter plasmid. To induce differentiation, cells were placed in 10% FBS DMEM + 2 mM CaCl2 for 48 h. Levels of the differentiation marker involucrin protein were elevated after calcium treatment (Fig 6C). As shown in Fig 6D, luciferase expression increased approximately 1.5 fold during differentiation, which is consistent with previous studies using Southern blotting to measure copy number at 48 h in CIN612-9E cells [40, 51]. Subsequently, control and ORC2 siRNAs were transfected into CIN612 cells along with the HPV-31 ori reporter plasmid in the presence or absence of 48 h calcium and normalized to renilla levels. Consistent with the shRNA experiments, ORC2 silencing enhanced HPV-31 ori luciferase expression (Fig 6E). Finally, we measured HPV-31 DNA genome levels in CIN612 cells after ORC2 depletion and calcium induced differentiation. HPV-31 DNA content was significantly higher with ORC2 knockdown compared to the control at 72 h (Fig 6F). Consistently, the ORC2 depleted cells showed increased HPV-31 E2 and E1 occupancy near two of the three viral origin regions tested by ChIP assay (Fig 7).
Taken together, these results implied that ORC2 is not necessary for HPV ori licensing, so the logical question became what is the biological significance of E2 binding to ORC2. A clue is that high level expression of the HPV E2 protein has been reported to inhibit cellular proliferation [52, 53]. This led us to hypothesize that E2 association might interfere with ORC2 function. To test this, we queried the effects of E2 expression on ORC2 occupancy at a frequently utilized mammalian origin. These ChIP experiments are challenging since it is well established that origin utilization is variable in both site usage and timing [54, 55] and because of the magnitude and repetitive composition of the human genome [56]. These experiments included U2OS T-REx cells with a doxycycline-inducible HPV-31 E2 expression cassette (i31E2) and U2OS cells that stably express the HPV-16 E2 protein (16E2). ORC2 protein levels were unaltered in the presence of E2 (Fig 8A). It was reported that HPV-16 E2 expressing U2OS cells are capable of progression into and through S-phase although this occurs at a slower rate [57]. We analyzed the cell cycle between the E2 expressing cell lines and their controls and found no difference between their cell cycle profiles (S4B Fig). U2OS, 16E2, and i31E2 cells were enriched in G1/S phase content and ChIP assays were performed with ORC2 and control antibodies. Occupation of ORC2 at the GM-CSF origin was reduced in cells expressing HPV-31 E2 and HPV-16 E2 (Fig 8B). The presence of ORC2 on the lamin B2 origin [35] was also decreased in i31E2 cells following dox-induced HPV-31 E2 expression (Fig 8C).
For the timely and efficient duplication of an entire metazoan genome, DNA synthesis must begin at an estimated 10–100,000 replication origins that must be activated once and only once per cell cycle [20]. Ori firing is coordinated by ORC binding followed by a cascade in which the MCM2-7 hexamer, Cdc6, and Cdt1 assemble during G1 and become activated by protein kinases during S-phase [58]. While metazoan origins span tens of thousands of kilobases, small viral genomes cannot accommodate such vast sequences. Furthermore, papillomaviruses do not encode any DNA polymerases and yet persistently infect and autonomously replicate their genomes in host cells. Functionally, the PV origin consists of recognition sites for the high affinity and high specificity E2 protein, a contiguous binding site for E1, and a flanking A-rich tract. Upon entry of a single papillomavirus genome into a basal epithelial cell nucleus, it undergoes replication to a few copies. The viral E1 and E2 proteins are not present in the virion, so viral transcription with accompanying synthesis of these proteins is assumed to occur very early after infection. A recent study found that replication begins at the viral ori and proceeds through a theta structure model, although subsequent replication appeared to be unidirectional and did not initiate at a specific site [59]. Viral copy number is maintained in proliferating basal keratinocytes and in vitro cultured cells that carry PV episomes, which presumably requires viral origin licensing during G1/S after which the replicated genomes partition without triggering the mitotic spindle dependent checkpoint [60]. Viral genome segregation involves E2 interaction with Brd4 and ChlR1 [6, 61, 62]. The genome amplification stage occurs in response to an unknown differentiation associated cue. Activation of the ATM/ATR DNA damage pathway has been shown to stimulate viral genome amplification [2, 36] but is not necessary for episome maintenance [63]. TopBP1, a protein involved in the DNA damage response pathway, is essential for initiation of viral replication and may play a crucial role for genome amplification [64, 65]. Our goal is to characterize the cellular replication factors that are necessary for each stage of the PV tri-phasic replicative program.
In experiments to investigate involvement of ORC factors in PV replication, we discovered that ORC2 co-immunoprecipitates with BPV-1 and HPV-31 E2. The E2 protein regulates viral transcription by interacting with cellular factors and targeting these to the viral genome via its C-terminal DNA binding domain [8]. Its N-terminal trans-activation domain (TAD) was sufficient for ORC2 association. We were able to detect by ChIP endogenous levels of HPV-31 E2 and HPV-16 and -31 E1 proteins at the viral ori using a cell line that stabily maintains HPV episomes. Unexpectedly, ORC2 was not present at the HPV origin region in these cells under conditions where ORC2 was detectable at a commonly utilized cellular origin. It seems unlikely these antibodies might not access the ORC2 complex at the viral DNA. These data inferred that ORC2 is not necessary for viral origin licensing. Consistent with this observation, YFP-tagged ORC2 protein was not visualized at E1 and E2 nuclear replication foci but instead appeared diffusely throughout the nucleus. However, we cannot exclude the possibility that the ORC2 epitope is not available or that a few molecules of ORC2 protein are present in these foci.
An RNAi ORC2 knockdown strategy was pursued with the goal of reducing but not eliminating this key pre-RC factor. ORC2 depletion increased viral copy number in transient replication assays and in keratinocytes that maintain episomal HPV-16 genomes relative to genomic DNA content. ORC2 depletion had no effect on HPV content in SiHa cells with an integrated HPV-16 genome. Consistently, ORC2 knockdown increased HPV-31 DNA content in HPV-31 episomal cell lines. In experiments in which ORC2 expression vectors were co-transfected along with a PV-replicon reporter or E2 vector, the cells were unable to be established, indicating that over-expression of ORC2 is deleterious for cell proliferation, and hence effects on viral replication could not be studied. We found that short term ORC2 knockdown did not alter the cell cycle prolife, implying sufficient levels of ORC2 are present to sustain growth. Previous reports have found depletion of ORC2 has different outcomes on the cell cycle dependent on cell type. For example, ORC2 siRNA decreased S-phase DNA content in MCF10A breast cancer cells [66]. In HeLa cells, ORC2 depletion showed a slower S-phase and an increased M-phase [67]. However in HCT116 with a hypomorphic ORC2 mutation that expressed 10% of ORC2 compared to wild-type replicated normally once is S-phase [68]. There is evidence for post-translation modifications of ORC2, which may regulate ori licensing [28].
Utilization of re-replication (pre-RC) proteins has been discovered with several large DNA viruses including two gamma herpesviruses. EBV genomes utilize the mammalian replication licensing mechanism for conservation of viral copy number in latently infected cells. EBNA1 binds to and recruits ORC2 to the EBV oriP [29, 30]. Consistent with its role in EBNA1 regulated EBV replication, our ORC2 knockdown experiments disclosed a comparative decrease in EBNA-induced transient replication. ORC2 was reported to bind to Kaposi’s sarcoma-associated herpesvirus (KSHV, HHV-8) terminal repeats where LANA dependent replication initiates [69]. Depletion of ORC2 was reported to decrease KSHV latent replication [69]. In contrast and similar to our observations, reduced levels of ORC2 increased lytic replication of beta herpesvirus human cytomegalovirus (HCMV, HHV-5) [70], which conceptually resembles PV amplification that occurs in differentiated epithelial cells.
We conclude that PV ori licensing does not require ORC2. The data shown here indicate that ORC2 suppresses PV genome replication during its maintenance phase. Levels of the E2 protein are very low in the maintenance phase of the viral life cycle such that the ORC2 protein may interfere with E2 loading at the viral origin. Increased viral replication following ORC2 depletion is likely due to the increased occupancy of E1 and E2 detected at the viral origin.
The ratio of ORC2 to E2 changes during the viral amplification stage, which resemble a G2/M arrest state imposed by high levels of the HPV E4 protein in the upper epithelium [71, 72]. While this has been proposed to allow for continued viral DNA synthesis in the absence of cell cycle progression, tens of thousands of human origins persist and might be licensed and activated. Our hypothesis is that the high levels of E2 that occur in differentiated epithelial strata [49, 50] inhibit ORC2 function and thereby restrict firing of host replicative origins that would otherwise effectively compete for replication enzymes necessary for viral genome amplification. Several studies have shown that over-expression of E2 induces growth arrest in both HPV positive and negative cell lines [47, 53, 73–75]. The E2-mediated growth arrest in HPV positive cells is mediated through expression of E6 and E7 [76], but the mechanism in HPV negative cells is unclear. In HaCat cells, over-expression of HPV-16 E2 induced apoptotic cell death, however, surviving cells maintained long-term low expression of E2 and appeared to be terminally differentiated [77]. Primary human foreskin keratinocytes infected with HPV-31 E2 adenovirus arrested in S-phase [52]. Inducible expression of HPV-31 E2 protein in U2OS cells resulted in reduced ORC2 bound at the GM-CSF ori. The cell cycle profile in the cell lines with E2 over-expression did not change suggesting that reduced ORC2 bound at the origin was not due to cell death.
These experiments illustrate the dual nature of the E2-ORC2 interaction and how the virus senses and manipulates its cellular environment. At low levels of E2 present in basal cells, ORC2 inhibits viral replication and thereby prevents a lytic infection. During epithelial differentiation when elevated levels of E2 are achieved, its complex with ORC2 restricts host cell origin licensing thus promoting replication of the HPV amplicon.
Codon optimized FLAG HPV-31 E2 [36] was cloned between the BamHI and HindIII sites of pcDNA3. Codon optimized triple FLAG HPV-31 E1 [41], pCG-BPV-1 E1 Eag123 [78], the ori-luciferase plasmids for the PV transient replication assay [41], pLKO.1 ORC2 shRNA [64], YFP-ORC2 [40] used were previously reported. For shRNA experiments, pCMV-GIN-Zeo (Open Biosystems) was used as the control plasmid. Human ORC2 cDNA was cloned between the NheI and ApaI sites in pcDNA3 containing a HA tag. Mouse ORC1 and ORC2 cDNA were amplified by PCR from mouse cDNA with FLAG tags and cloned into the MluI and SalI sites of the pCI backbone. The BPV-1 expression vectors and GST vectors have been described elsewhere [10, 79]. pREP4 which contains the EBV oriP and EBNA1 was modified to contain a E1F-α minimal promoter and used to measure EBV replication. To generate inducible HPV-31 expression, codon optimized FLAG HPV-31 E2 was cloned into the HindIII and ApaI sites of pcDNA4/TO-luc creating pcDNA4/TO-FLAG 31 E2.
The following antibodies were used: Mouse anti-FLAG (M2; Sigma), mouse anti-HA (HA-7; Sigma 3F10; Roche, 12CA5), mouse anti-human ORC2 (MBL), rat anti-human ORC2 (Cell Signaling), mouse anti-HPV-16 E6 (Arbor Vita), mouse anti-EE and anti-β-actin (Sigma). BPV-1 E2 was identified with B201 (a mouse monoclonal antibody with an epitope between amino acids 160–220) and II-1 rabbit antisera [12]. The sheep anti-HPV-16 E2 antibody was described elsewhere [62]. With the prior approval of the Johns Hopkins University Animal Care and Use committee (protocol RA14M47), two Lewis rats were immunized three times with 50 μg of GST-tagged E1 N terminal fusion protein (GST-E1N); (1–209 amino acids of HPV16 E1 protein) formulated with TiterMax (TiterMax USA, Inc.) adjuvant and the fourth immunization only GST-E1N was used. Three days following the final immunization, splenocytes were harvested, pooled and fused with SP2/0 myeloma cells using polyethylene glycol (PEG). Supernatants from wells containing viable hybridomas were screened at 10 to 14 days after fusion by ELISA using bacterially expressed GST-E1N and GST-tagged E2 (GST-E2) derived from HPV16 as a coating antigen. Selected positive clones that recognized GST-E1N but not GST-E2 were re-cloned three times by limiting dilution (1H4-4E-32.7.3.1 and 11B63G.6.1), and the antibodies secreted were purified using protein G Sepharose (GE Healthcare). An 18mer peptide array of E1 peptides within the 1–206 regions of HPV16, each overlapping by 15 residues was utilized to map the MAb epitopes by direct ELISA. Diluted hybridoma supernatants or purified antibodies were added to the blocked peptide-coated plates and incubated for 1 h at 37°C. Plates were washed three times with PBS containing Tween-20 again and then incubated with HRP conjugated goat anti-rat IgG for 60 min at 37°C. Unbound MAb was then washed away with 0.01% Tween 20 in PBS and signal detected with ABTS (Roche, Basel Switzerland). Anti-E1 MAb, 1H4-4E-32.7.3.1 isotype IgG2a, recognized only two overlapping E1 peptides comprising residues 4–21 and 7–24, its epitope likely within in 7–21 (TNGEEGTGCNGWFYV). Another anti-E1 Mab of isotype IgG2a, 11B63G.6.1, recognized 5 overlapping E1 peptides (4–21, 7–24, 10–27, 13–30 and 16–33) and, its epitope likely resides within residues 16–21 (NGWFYV). In addition, a rabbit was immunized 5 times with GST-E1N antigen (amino acid 1–209 of HPV16 E1) generated rabbit-anti-E1 serum (Generated by Proteintech Group, Inc.). Rabbit anti-HPV-31 E2 (10008) was generated to amino acids 97–109 (TSLELYLTAPTGC; manuscript in preparation).
HEK293TT (from John Schiller and Chris Buck), CV-1 (from Douglas Lowy), SiHa (ATCC), U2OS T-REx (from Feng Yao), U2OS–HPV-16 E2 (from Iain Morgan), and C33A (from Douglas Lowy) were cultured at 37°C in Dulbecco’s Modified Eagle Medium (Life Technologies) with 10% fetal bovine serum (Atlas Biologicals) and penicillin/streptomycin (100 U/ml; Life Technologies) at 5% CO2. CIN612-9E (from Laimonis Laimins) a clonal cell line which maintains HPV-31 genomes were grown in E-medium with J23T3 fibroblast feeders (from Howard Green). W12 (from Margaret Stanley and Paul Lambert) a cervical carcinoma cell line with maintains HPV-16 genomes were grown in F-medium with J23T3 fibroblast feeders. HPV-BP cells (from Aloysius Klingelhutz) a clonal cervical keratinocyte cell line containing HPV-16 episomes were maintained in Keratinocyte SFM containing human recombinant Epidermal Growth Factor 1–53 (EGF 1–53), Bovine Pituitary Extract (BPE) and penicillin/streptomycin (100 U/ml; Life Technologies) at 5% CO2 [80]. High Five cells were grown in SF9 media (Life Technologies) at 30°C and 5% CO2. U2OS T-REx cells were transfected with 1 μg pcDNA4/TO-FLAG HPV-31 E2 using Lipofectamine 2000. 24 hrs later, 400 μg/ml of Zeocin was added for about two weeks until cell colonies were established. 800 ng/ml of doxycycline was added to the media to induce expression of FLAG-HPV-31 E2.
In-situ PLA was performed using the PLA Red kit (Olink Biosciences). C33A cells were transfected with FLAG-HPV-31 E2 and HA-hORC2 or FLAG-HPV-31 E2 and HA-HPV-31E2 constructs. 24 h later, cells were fixed in 4% paraformaldehyde for 10 min, permeabilized for 15 min in 0.5% Triton-X 100/PBS, washed in PBS, blocked with 5% goat serum in 0.2% Triton-X 100/PBS, then incubated overnight with primary antibody combinations (E2–rabbit FLAG or ORC2–Mouse anti-human ORC2; or E2-M2 FLAG or E1-Rabbit 16E1) at 4°C. The PLA then followed the manufacturer׳s protocol. Briefly, cover slips were washed in buffer A, incubated with PLA probe PLUS and MINUS for one hour at 37°C, washed twice and the probes were ligated for 30 min. Amplification was performed for 100 min at 37°C, washed twice in buffer A and B. Cover slips were mounted in Duolink in-situ mounting media. Cells were analyzed with a Nikon microscope.
Cells were synchronized (G1/S) using 2.5 mM thymidine for 16 hr, released for 6 hr, and retreated with 2.5 mM thymidine for 16 hr before harvest. For the ORC2 siRNA experiments, CIN612-9E were transfected with 30 nM of the siRNAs as described below. ChIP used the ChIP-IT express chromatin immunoprecipitation kit (Active Motif). Real time PCR was performed using Sso Fast Evagreen Mastermix (BioRad). Primers for the HPV-16 LCR in Fig 2 were 5’ GGGTGTGTGCAAACCGTTTTGGGTTA 3’ and 5’ CCGATTTCGGTTACGCCCTTAGTTT 3’ and primers for the HPV-31 LCR were 5’ CCTGCTCCTCCCAATAGTCAT 3’ and 5’ AAACGGACCGGGTGTACAA 3’. We used primers for human GM-CSF ori #2 (#23) [81]. The ‘exon 9’ primers amplify a region six kb upstream of the upper regulatory region of the MCM4 gene at which pre-ORC factors are not enriched [35]. Primer sets for the HPV-31 and HPV-16 LCR region in Fig 3 are listed in Table 1 below.
Results were analyzed using BioRad CFX manager software. Each experiment was performed at least three independent times.
Cells were synchronized (G1/S) using 2.5 mM thymidine for 16 hr, released for 6 hr, and retreated with 2.5 mM thymidine for 16 hr before harvest. At harvest, cells placed in 90% ethanol. For processing, cells were washed in PBS, treated with 50 μg/ml RNase A, 0.05 mg/ml propidium iodide for 20 min, run on FACS Caliber and analyzed with FlowJo software.
Cells were transfected using Lipofectamine 2000 or PEI (2 mg/ml) according to manufacturers’ instructions. For experiments using MG-132, cells were treated 24 hr post transfection at 10 μM for 6 hr. After 24–48 hr, cells were lysed in 1% NP-40 containing 50 mM NaCl, 10 mM HEPES, pH 7.4, 10% glycerol, 2 μl benzonase (Millipore) and protease inhibitor cocktail (Sigma). To each reaction, 40 μl of 50% protein A/G slurry (Invitrogen) and 1 μg of antibody was added and rotated for 2 hr or overnight at 4°C. Beads were washed 2 times in PBS and 2 times in 500 mM NaCl, 0.1% SDS, 1% NP-40, 2 mM EDTA, 20 mM Tris-HCl, pH 8.0. Proteins were separated by SDS-polyacrylamide gel electrophoresis (SDS-PAGE). Gels were transferred onto PVDF membranes (Millipore), blocked in 5% PBS-Tween (0.1%) milk, and probed with antibodies. Chemiluminescence substrates (Thermo Scientific) were used to detect antibody signal.
Purification of 6X His E2 1–216 [82] and MBP-16E6 [83] proteins was performed as described [82, 84]. ORC2 baculovirus and ORC2 protein was produced as described elsewhere [85]. High Five cells infected with ORC2 baculovirus were lysed in 1% NP-40 buffer. Lysates were added to 6X His E2 1–216 and MBP-16E6 proteins with nickel Dynabeads (Invitrogen) beads in 1% NP-40 containing 50 mM NaCl, 10 mM HEPES, pH 7.4, 10% glycerol, 2 μl benzonase and protease inhibitor cocktail.
C33A cells were grown to 50% confluence in 6-well dishes. 3 μg shRNA plasmid (Control or ORC2 shRNA), 15 ng pRL (Rluc), 75 ng pFLORI31 or pFLORIBPV-1, 0.3 μg pE1 and 0.3 μg pE2 or 0.1 μg pREP4 (Life Technologies) using Lipofectamine 2000. For the BPV replication assays, pCG-BPV-1 E1 Eag1235 was used for E1 expression. 48–72 hr later cells were lysed and luciferase activity was measured using Steady-Glo or Dual Glo. Firefly luciferase levels were normalized to renilla luciferase levels. CIN612-9E cells were cultured to 50% confluence in 6-well dishes. Feeder J2 3T3 cells were removed before transfection. 0.5 μg shRNA plasmid (Control shRNA or ORC2 shRNA), 15 ng pRL (Rluc), 75 ng pFLORI31 were transfected into CIN612-9E cells using FugeneHD (Invitrogen). J2 3T3 cells were added the next day. 48 hr later, cells were lysed and luciferase activity measured using Dual Glo luciferase assay reagent (Promega). Firefly luciferase activity was normalized to renilla luciferase levels.
CIN612-9E cells, without feeders, were plated in 6 cm plates with the addition of the transfection reaction. The transfection reaction contained Lipofectamine 2000 (Invitrogen) and either control siRNA (Santa Cruz; sc-37007) or ORC2 siRNA duplexes (UUGAAGAAGGAGCGAGCGCAGCUUU [66], IDT) at a final concentration of 5 or 15 nM with 250 ng pFLORI31 and 50 ng pRL (Rluc) per well. Five hours post transfection, media containing the transfection was removed and replaced with E-medium or 10% FBS DMEM + 2 mM CaCl2 to induce differentiation. 48 hr later, cells were lysed and luciferase activity measured using Dual Glo luciferase assay reagent (Promega). Firefly luciferase activity was normalized to renilla luciferase levels.
HPV-BP and SiHa cells were plated in 10 cm plates. ShRNA knockdown hairpins were transfected using Fugene HD. 24 hrs post transfection cells were maintained in 1 μg/ml puromycin (Life Technologies) for 3–5 d. At about day 10, cells were lysed into Steady Glo (Promega) lysis buffer and western blot was performed as described above or cells were lysed in phenol:chloroform:isoamyl alcohol (25:24:1, Fisher Scientific). DNA was quantified using the Nanodrop and 30 ng was used for Real Time PCR as described above. HPV-16 DNA content was measured using primers to HPV-16 LCR (listed above-Fig 3). HPV-16 DNA content was normalized to a region of genomic DNA (GM-CSF ori # 2, listed above). This experiment was performed three independent times.
CIN612-9E cells, without feeders, were plated in 6 cm plates with the addition of the transfection reaction. The transfection reaction contained Lipofectamine 2000 (Invitrogen) and either control siRNA (Santa Cruz; sc-37007) or ORC2 siRNA duplexes (UUGAAGAAGGAGCGAGCGCAGCUUU [66], IDT) at a final concentration 15 nM. 4 hr post transfection, the media was replaced with 10% FBS DMEM + 2 mM CaCl2 to induce differentiation. 72 hr later, cells were lysed in TE with 0.1% SDS with 10 ng/μl RNase. DNA was isolated with phenol:chloroform:isoamyl alcohol (25:24:1, Fisher Scientific) and Real Time PCR was performed as described above. HPV-31 DNA content was measured using primers to HPV-31LCR3 (listed above-Fig 3). HPV-31 DNA content was normalized to a β-actin DNA primer set—5’ GAGGCACTCTTCCAGCCTTC 3’ and 5’ CGGATGTCCACGTCACACTT 3’.
CV-1 cells were transfected with 1 μg FLAG-tag HPV-31 E2, 1 μg FLAG-tag HPV-31 E1, and 8 μg YFP-ORC2 mammalian expression vectors with and without 50 ng HPV-31 origin or transfected with 2ug pCG-BPV-1 E2, 2ug pCG-BPV-1 E1, and FLAG mouse ORC2 mammalian expression vectors using Lipofectamine 2000. 24–48 hr later, coverslips were washed twice in PBS and fixed in 4% paraformaldehyde for 15 min followed by 3X PBS washes. The cells were permeabilized in PBS/0.2% Triton-X (PBS-T) for 10 min followed by 2X PBS washes. Coverslips were placed in 10% goat serum in PBS-T for 1 h, followed by primary antibody in PBS-T for 2 hr, washed 3X with PBS, incubated in 1:5000 dilution of goat anti-mouse 594 or goat-anti-rabbit 488 (Invitrogen) for 1 hr, washed 3X PBS, and mounted onto slides with Vectashield mounting media with DAPI.
All animal studies were carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and with the prior approval of the Animal Care and Use Committee of Johns Hopkins University under protocol RA14M47.
Two-way or one-way t-test was used for analysis. Means are expressed +/- SEM. * indicates p-values ≤ 0.05.
|
10.1371/journal.ppat.1001096 | Cyclin-Dependent Kinase Activity Controls the Onset of the HCMV Lytic Cycle | The onset of human cytomegalovirus (HCMV) lytic infection is strictly synchronized with the host cell cycle. Infected G0/G1 cells support viral immediate early (IE) gene expression and proceed to the G1/S boundary where they finally arrest. In contrast, S/G2 cells can be infected but effectively block IE gene expression and this inhibition is not relieved until host cells have divided and reentered G1. During latent infection IE gene expression is also inhibited, and for reactivation to occur this block to IE gene expression must be overcome. It is only poorly understood which viral and/or cellular activities maintain the block to cell cycle or latency-associated viral IE gene repression and whether the two mechanisms may be linked. Here, we show that the block to IE gene expression during S and G2 phase can be overcome by both genotoxic stress and chemical inhibitors of cellular DNA replication, pointing to the involvement of checkpoint-dependent signaling pathways in controlling IE gene repression. Checkpoint-dependent rescue of IE expression strictly requires p53 and in the absence of checkpoint activation is mimicked by proteasomal inhibition in a p53 dependent manner. Requirement for the cyclin dependent kinase (CDK) inhibitor p21 downstream of p53 suggests a pivotal role for CDKs in controlling IE gene repression in S/G2 and treatment of S/G2 cells with the CDK inhibitor roscovitine alleviates IE repression independently of p53. Importantly, CDK inhibiton also overcomes the block to IE expression during quiescent infection of NTera2 (NT2) cells. Thus, a timely block to CDK activity not only secures phase specificity of the cell cycle dependent HCMV IE gene expression program, but in addition plays a hitherto unrecognized role in preventing the establishment of a latent-like state.
| Cyclin-dependent kinases (CDKs) control the cell division cycle. Many viruses employ CDK activity to control critical steps of their own replication cycle and to synchronise their replication with the cell cycle dependent availability of vital cellular enzymes and molecular building blocks. Here we show an unexpected antiviral function of CDK activity at a very early stage of human cytomegalovirus (HCMV) infection, the onset of immediate early (IE) gene expression. HCMV is unique amongst herpesviruses in being unable to initiate IE gene expression during the S/G2 phase of the cell cycle. CDK inhibition by either DNA damage-dependent induction of the cellular CDK inhibitor p21 or by the pharmacological CDK inhibitor roscovitine overcomes this limitation and makes S/G2 cells fully permissive for HCMV. Importantly, in undifferentiated NTera2 (NT2) cells, which normally establish a quiescent, latent-like HCMV infection, CDK inhibition also relieves the block of IE gene expression, suggesting a more general role for CDK activity in the control of this important human pathogen.
| Human cytomegalovirus (HCMV) is a wide-spread human pathogen causing serious disease in immunocompromised patients and neonates [1]. As with all herpesviruses, HCMV exists either in a latent, asymptomatic state or undergoes poductive replication leading to lysis of the host cell. Lytic replication starts with the onset of viral immediate early (IE) gene expression. IE gene products, especially the major IE (MIE) proteins IE1 and IE2, have essential functions in host cell regulation and in activating the subsequent cascade of viral early and late gene expression [2]. In latently infected cells, MIE gene transcription is silenced and consequently viral gene expression is restricted to only very few genomic loci [3], [4], [5], [6]. Reactivation from latency is achieved by mechanisms that trigger desilencing of the MIE promoter/enhancer [7], [8], [9]. Thus, control of MIE gene expression is pivotal to the outcome of infection and, therefore, represents a main focus of HCMV research. In addition, MIE gene expression as the initial step in HCMV replication is considered a prime target for antivirals and an IE2-specific antisense-RNA (fomivirsen) has already proven to be effective in the local treatment of HCMV retinitis [10].
Interestingly, latent infection is not the only situation where HCMV replication is blocked at the level of MIE gene expression. For primary fibroblasts it has been shown that the cell cycle state at the onset of infection determines whether viral gene expression is initiated or not. In G0/G1, IE gene expression starts immediately while in S/G2 phase, transcription of IE1 and IE2 is efficiently suppressed [11], [12]. However, infection of S/G2 fibroblasts does not fully prevent but rather delays the onset of the lytic cycle until cells have completed cell division and reentered the next G1 phase. The physiological relevance of the cell cycle dependent regulation of HCMV is not understood. Furthermore, it is unclear what makes S/G2 cells non-permissive for MIE gene expression and whether the underlying mechanism also plays a role in the establishment of HCMV latency.
Here we analyzed the molecular determinants of cell cycle dependent repression of HCMV major IE genes. We found that inhibition of cyclin dependent kinase activity either by checkpoint activation or the chemical inhibitor roscovitine was sufficient to fully restore virus permissiveness in S/G2. Moreover, CDK inhbition was also successful in antagonizing the silencing of lytic gene expression during quiescent, latent-like infection of undifferentiated NTera2 (NT2) cells, suggesting a mechanistical link between cell cycle and latency-associated repression of IE gene transcription.
The block to MIE gene expression in S and G2 phase has been described using the highly attenuated HCMV laboratory strains Towne [11], [12] and AD169 [13]. We first decided to test whether cell cycle dependent IE expression also applies to low-passage strains and clinical isolates, as this would also suggest an in vivo relevance for HCMV infection. To this end, we infected human fibroblasts with either high-passage strains AD169, Towne and Davis, low-passage strain Toledo or clinical isolates Merlin, TB40/e and VHL/e and analyzed the cells for MIE gene expression and cell cycle distribution (figure 1A). All strains clearly failed to initiate IE1 and IE2 expression in cells with an S/G2 DNA content. While typically more than 60% of infected G1 cells were IE1/IE2-positive at 3 hpi, only a maximum of 13% S/G2 cells supported MIE expression. Thus, the cell cycle-dependent block to MIE gene expression is a genuine feature of HCMV and not a mere consequence of a reduced virulence attributed to fibroblast adapted laboratory strains.
In contrast to AD169 and Towne, the TB40/e isolate has retained the broad spectrum of cell tropism characteristic for HCMV infections in vivo. This enabled us to test whether the S/G2-specific block of MIE expression also accounts to cell types other than fibroblasts. We analyzed primary endothelial cells (HUVEC), two transformed cell lines of epithelial origin (HeLa, HEK293) and the glioblastoma cell line U373-MG (figure 1B and C). All cell types tested were permissive for MIE gene expression in G1 but not S or G2 and hence behaved like primary fibroblasts in this respect. Together these results underscore the general relevance of the cell cycle state for virus-host interaction and in addition, allow more flexibility in choosing informative cellular systems for further specific experimental approaches.
To identify molecular determinants of the cell cycle-dependent block to HCMV gene expression, we first turned to examine a possible role of cellular DNA synthesis. The rational being that DNA synthesis is the hallmark of S phase in its own right and the observation that the block to IE gene expression was often found more pronounced in S rather than G2 (see figure 1A). Inhibition of DNA replication was achieved by aphidicolin or hydroxyurea treatment. When applying these inibitors together with or a few hours prior to infection we only saw a slight increase of MIE-positive cells in S/G2 compared to DMSO-treated control cells (figure 2) suggesting that ongoing DNA replication is not a cause per se for the inhibition of viral gene expression. However, preincubating cells with the inhibitor for 18 h or longer prior to the 5 h infection period, raised the percentage of MIE-positive cells in the S/G2 compartment to at least the same high level as in G1 cells. This indicates that prolonged inhibition of cellular DNA replication induces secondary events that are able to effectively overcome the S/G2-specific block to HCMV gene expression.
Chemicals that disrupt replication, such as aphidicolin and hydroxyurea, are known to elicit an ATR-dependent DNA damage response [14], [15]. Thus, the positive influence of replication blockage on MIE gene expression in S/G2 could be a consequence of checkpoint activation rather than the absence of DNA synthesis per se. To address this point, we directly induced DNA damage and chose two independent ways to do so, namely doxorubicin treatment and ultraviolet (UV) irradiation. To cover both early and delayed effects of checkpoint activation on viral gene expression, cells were infected at different time points after induction of DNA damage (figure 3A). Cells that were infected immediately after DNA damage showed the same pattern of cell cycle-dependent MIE gene expression as untreated cells (figure 3B). In contrast, at 24 h and especially at 48 h after DNA damage many cells with an S/G2 DNA content had become permissive for MIE gene expression. This finding demonstrates two things. First, induction of the DNA damage checkpoint is indeed effective in creating a supportive environment for viral MIE gene expression in S/G2. Second, the events leading to increased permissiveness are likely to be part of the long-term and not the fast but transient response to cellular DNA damage.
The non-permissive nature of S/G2 cells causes a long delay in the HCMV replicative cycle, as viral gene expression cannot be initiated before cells have divided and reentered G1 [11], [12]. To investigate whether the DNA damage-induced rescue of MIE gene expression is also sufficient to trigger the consecutive gene expression program of the normal replicative cycle in S/G2 cells, we compared the kinetics of MIE, early and late gene expression in untreated and doxorubicin treated cells (figure 4). Consistent with a previous report [11], after infection of untreated S phase cells the program of lytic gene expression lagged at least 24 h behind that of infected G1 cells. In contrast, doxorubicin treated S phase cells were able to support not only MIE but also the subsequent steps of the viral gene expression program without delay. Of note, this also includes the pp28 gene. The fact that expression of true late genes like pp28 depends on the onset of viral DNA synthesis indicates that also viral DNA replication was initiated in S/G2 cells with normal kinetics. Taken together, these data clearly suggest that with the DNA damage-dependent release of MIE gene expression the major obstacle to HCMV replication is removed from S phase cells and infection proceeds with the normal kinetics known from infected G0/G1 cells.
The p53 tumor suppressor protein governs the long-term cellular response to genotoxic stress by inducing a p21-dependent cell cycle arrest or apoptosis [16]. Since the observed rescue of viral gene expression in S/G2 cells after DNA damage follows long-term kinetics, we next asked whether this rescue depends on p53-dependent signalling. To this end we made use of HEL fibroblasts with a stable p53 knockdown [17]. Using the same experimental setup as before, we treated p53 knockdown (p53-KD) and appropriate control (mock-transduced and GFP-KD) fibroblasts with aphidicolin or doxorubicin and infected them immediately (0 h) or 24 h after treatment (figure 5A). Immunoblot analysis confirmed that p53-KD but not control cells resist the aphidicolin and doxorubicin-induced upregulation of p53 and p21 (figure 5B). The increase of p53 expression in control cells between 0 and 24 h was accompanied by the expected increase in permissiveness for MIE gene expression in S/G2 (figure 5C). In contrast, p53-KD cells failed to upregulate MIE genes in S/G2 cells rather showing the typical G1-restricted pattern of MIE gene expression and this was regardless of whether infection was initiated at early or late times after DNA damage or inhibition of DNA replication. This clearly demonstrates that the rescue of HCMV gene expression in S/G2 relies on p53-dependent checkpoint signalling both, after induction by genotoxic (doxorubicin) or replicative (aphidicolin) stress.
Different to our approach, a study by Fortunato et al. employed 24 h aphidicolin treatment to synchronize cells at the G1/S-transition before HCMV infection [12]. At the time of infection they released the cells from the aphidicolin block, enabling cells to recover and cycle through S/G2/M/G1. As the continuous presence of the drug is needed to keep the replication checkpoint and p53-dependent signalling active, it is not contradictory to our findings that their synchronization protocol had no major effect on the percentage of MIE-positive cells in S/G2.
P53 is a short-lived protein whose abundance is controlled by Mdm2-mediated ubiquitination and subsequent degradation by the 26S proteasome [18]. It has been known for some time that the block to MIE gene expression in S phase can be overcome by proteasomal inhibitors [12]. However, the reason for that remained elusive. Given the above results we rationalized that with respect to MIE expression in S phase p53 might be the limiting target of proteasomal inhibition. Like DNA damage proteasomal inhibition leads to p53 stabilisation and consequently to checkpoint activation. To directly address this hypothesis, we infected p53-KD and control fibroblasts in the presence of the proteasome inhibitor MG132 (figure 6A). As expected, after 8 h of MG132 treatment both p53 and its target p21 had strongly accumulated in non-infected and HCMV-infected control cells and this effect was largely suppressed in p53-KD cells (figure 6B). The following analysis of cell cycle dependent viral MIE gene expression revealed that the p53 status was indeed crucial for the increased permissivenes of S phase cells after proteasome inhibition. In the presence of p53, MG132 treatment resulted in a marked increase of MIE-positive S phase cells from 20% to 50% (figure 6C). In contrast, in p53-negative cells proteasomal inhibition had no influence on MIE gene expression that even dropped from 15% to 13% in the S phase compartment. Thus, the finding that unrelated agents such as aphidicolin, doxorubicin and MG132 rescue MIE gene expression in S phase fibroblasts is a consequence of their shared ability to induce the p53 tumor suppressor protein.
One of the main effectors of p53-dependent signalling is the CDK inhibitor p21 [19]. To directly test if p21 is required for the p53-dependent release of MIE gene expression in S/G2 we focussed on the HCT116 cell system where somatic knockouts of both p53 and p21 are available [20]. Because HCT116 cells are of epithelial origin and therefore non-permissive for fibroblast adapted HCMV laboratory strains [21] we used the endotheliotropic TB40/e isolate for infection. As before, we applied doxorubicin to induce via DNA damage the p53-p21 checkpoint axis and analyzed MIE gene expression after infection at early (0 h) and late times (24 h) post doxorubicin treatment (figure 7A). It appeared that untreated HCT116 cells in G1 as well as in S/G2 only weakly support MIE gene expression (see left part of figure 7B). For wild type (wt) and p53−/− cells, the proportion of MIE-positive cells was below 2% in all cell cycle phases, only p21−/− cells appeared to be a slightly more supportive of viral gene expression with up to 5.4% MIE-positive cells in G1. However, after exposure to doxorubicin the permissiveness of HCT116 cells increased remarkably (see right hand panels of figure 7B). A first very minor increase was already visible at 0 h, yet proved to be independent of p53 and p21. After 24 h, the number of permissive wt cells was increased to 35% in G1 and 20% in S/G2. This late effect of doxorubicin treatment was almost completely prevented in p53 and p21 knockout cells.There, compared to wt cells, the percentage of MIE-positive cells was 3-fold (p53−/−) and 3.5-fold (p21−/−) lower in G1 and even 13-fold (p53−/−) and 6-fold (p21−/−) lower in S/G2. First, these results show that in the case of HCT116 cells, viral gene expression in G1 seems to be subject to the same repressive mechanism as in S/G2. Second, in addition to p53 the rescue of MIE gene expression also and equally depends on p21, suggesting that p21 represents the critical effector of p53 in this context.
The best understood checkpoint effector functions of p21 are the inhibition of DNA replication by PCNA binding and the inhibition of CDK2 and CDK1 activity [19]. To test whether the p21-dependent effect on MIE gene expression relies on CDK inhibition, we infected fibroblasts in the presence of the pharmacological CDK inhibitor roscovitine (also known as CYC202 or seliciclib). In a first set of experiments, roscovitine was applied in parallel with viral infection, i. e. left on the cells from the beginning of infection until cell harvest at 4 hpi (figure 8A). The effects we observed were clearly dose-dependent. Low doses (5 µM) of roscovitine resulted in a modest increase from 8% to 18% MIE-positive cells in S/G2 (figure 8B) and at medium concentrations (15 µM) the number further increased to 34%. However, at the same time and independent of the cell cycle position the average MIE expression level per cell dropped significantly with increasing concentrations of roscovitine such that high doses (50 µM) finally led to an almost complete loss of MIE gene expression in all cell cycle phases. Thus, roscovitine exerts two opposing effects during the first hours of HCMV infection. The cell cycle independent, negative effect on MIE gene expression has been described [22] and attributed to inefficient viral transcription caused by the roscovitine-mediated inhibition of CDK7 and CDK9 [23], [24]. However, more pertinent to the question examined here and in good agreement with the above data there is a clear positive effect of roscovitine on MIE gene expression. In a previous study roscovitine (15 µM) was left on HCMV-infected cells until harvest at 12 hpi and did not increase the percentage of IE-positive cells in S/G2 [12]. In this setting, the negative influence of CDK inhibition on ongoing viral gene expression probably outweighed the positive effect on its initation.
In order to better discriminate between the two opposing activities of CDK inhibition, we next modified our experimental setup. To this end, we only allowed roscovitine on infected cells for two hours before harvesting cells at 5 hpi (figure 8C). Thus, roscovitine should have exerted its positive effect on the initiation of MIE without negatively affecting further viral MIE protein accumulation. Indeed, this setting avoided the negative impact of CDK inhibition resulting in a further doubling of MIE-positive cells in S/G2 (from 34% to 68%) when increasing the roscovitine concentration from 15 to 25 µM (figure 8D). This further supports the view that CDK activity acts as a strong inhibitor of the initiation of viral MIE gene expression during the onset of lytic infection. Roscovitine treatment also rescued S/G2 phase-specific MIE gene expression in U373 and HUVEC cells (figure 8E). Even at high concentrations of roscovitine (50 µM in the case of U373) MIE gene expression was not impaired to any extent in these cells reaching high and comparable levels in the S/G2 and G1 cell cycle phases.
It has been shown that, in addition to CDK inhibition, roscovitine treatment can also trigger p53-dependent signalling [25], [26]. Although the observed kinetics made it unlikely that the effect of roscovitine was mediated via p53 signalling, we aimed at formally excluding this possibility to further unravel the importance of a direct inhibitory CDK function on initiation of HCMV MIE gene expression. To this end we analyzed whether roscovitine – unlike doxorubicin, aphidicolin or MG132 (see above) – is still effective in the absence of p53. Using p53 positive control fibroblasts, the direct comparison to doxorubicin treatment revealed that short-term treatment with roscovitine was as efficient in creating a permissive state in S/G2 cells as the long-term response to DNA damage. However, while the effect of doxorubicin was completely abolished in p53-KD cells, the strong effect of roscovitine proved to be p53 independent (figure 9).
In summary, these data demonstrate that short-term treatment with roscovitine can overcome the block to HCMV MIE gene expression in S/G2. This roscovitine-mediated rescue resembles the p53 and p21 dependent MIE expression after long-term checkpoint activation but acts downstream from and independent of p53, directly targeting CDK activity. Thus, CDK activity exerts a strong negative effect at the initiation state of HCMV lytic infection.
Latent infection arguably represents the most relevant in vivo situation where MIE gene expression of HCMV is dominantly repressed. To address the important question as to whether CDK activity could also contribute to maintain the block to IE expression in latently infected cells, we made use of the embryonic carcinoma cell line NT2, a well-accepted in vitro model for HCMV latency [27]. In undifferentiated, pluripotent NT2 cells MIE gene expression is blocked and HCMV establishes a quiescent infection. The block can be relieved by chemical induction of neuronal differentiation [28], [29], [30], [31]. This resembles the differentiation state-dependent permissiveness of cells of the myeloid lineage, a major site of HCMV latency in vivo [32].
To analyze a possible function of CDKs for the permissiveness of HCMV in NT2 cells, we adjusted the experimental set up to the slower kinetics of viral gene expression in NT2 cells (figure 10A). Two positive controls were included in the experiment. Retinoic acid (RA)-induced differentiation (as indicated by the downregulation of the pluripotency marker Oct4) enabled MIE gene expression in 70% of all cells (figure 10B). Trichostatin A (TSA)-mediated inhibition of histone deacetylases led to 21% IE1/2-positive cells. However, in this case the rescue of MIE gene expression was differentiation-independent, which is consistent with previous reports [33], [34]. Intriguingly, transient CDK inhibition also increased the permissiveness of NT2 cells for MIE gene expression. This was shown for three different CDK inhibitors. Besides roscovitine we used another 2,6,9-substituted purine analogue (CVT313) and a structurally unrelated compound (SU9516). SU9516 proved to be most effective leading to up to 44% IE1/2-positive cells after 24 h of infection compared to only 1.4% in the DMSO-treated control. Of note, this induction was not a consequence of CDK inhibitor-induced differentiation since IE-expressing NT2 cells remained undifferentiated by means of undisturbed Oct3/4 expression. This result suggests that the inhibitory function of CDKs on HCMV lytic gene expression is not restricted to the cell cycle but is likely to have broader relevance for states of inhibited IE expression. In line with this reasoning, we were able to show that the induction of MIE gene expression by transient CDK inhibiton can occur in all cell cycle phases of NT2 cells even though S/G2 cells reacted most sensitive at low, suboptimal concentrations of inhibitors (figure 10C).
Given the cell cycle-independent nature of MIE induction by CDK inhibitors, we asked whether the permissiveness of NT2 cells might correlate with the expression of a constitutively active 38kD-form of Cyclin A2 that was recently described in mouse bone marrow and human myeloid precursor cells [35], [36]. Indeed this form was detectable in undifferentiated but not differentiated NT2 cells (figure 11A). In addition, the expression of full-length Cyclin A2 dropped significantly during retinoic acid-induced differentiation. This is consistent with published data [37] and might at least partially explain the observation that differentiated NT2 cells express MIE genes in a completely cell cycle-independent manner (figure 10C).
An already described determinant of MIE expression in NT2 cells is the differentiation-dependent nuclear localization of the viral tegument protein pp71 [38]. Nuclear localization of pp71 facilitates MIE gene expression by neutralizing the Daxx-mediated cellular intrinsic immune defense [39]. To address the question as to whether the rescue of MIE expression by CDK inhibition works via nuclear translocation of pp71 we compared the subcellular localization of pp71 under conditions of CDK inhibition and retinoic acid-induced differentiation (figure 11B). We found that the cytoplasmic localization of pp71 in untreated cells was not affected upon CDK inhibition. In contrast, retinoic acid-treated cells contained a large fraction of nuclear pp71 as expected. This suggests that CDK inhibition and pp71 nuclear translocation trigger independent mechanisms to induce MIE gene expression. Considering that after retinoic acid treatment MIE expression occurs in a cell cycle independent fashion (figure 10C) and pp71 is able to enter the nucleus (figure 11B and [38]) it appears that both mechanisms can respond to differentiation and therefore might act synergistically.
Here we show for the first time that CDK activity negatively controls the onset of HCMV gene expression. This finding was unexpected because HCMV, like many other viruses, was previously shown to be subject of positive regulation by CDKs. Apparently, the effect of CDK activity on HCMV varies depending on the phase of infection and the type of CDK (see figure 12). At a pre-immediate-early stage of infection, an S/G2-specific, probably Cyclin A2-dependent (see below) CDK prevents the initiation of IE gene expression at the level of transcription (this study and our own unpublished data). Once IE expression has been initiated, CDK activity is required for accurate processing and accumulation of viral transcripts [22]. This correlates with recruitment of CDK7 and CDK9 to the sites of viral transcription where they catalyze hyper-phosphorylation of the RNA polymerase II C-terminal domain [23], [24]. At later times CDK activity is needed for proper expression, modification and localization of pUL69, pUL83 and other HCMV proteins and for efficient production of viral progeny [40], [41], [42]. Both CDK2 [43] and CDK9 [40] but not CDK1 [44] are likely to contribute to these late effects. When CDKs are inhibited througout infection, the net effect is an almost complete suppression of HCMV replication [43], [44]. Accordingly, CDKs have been suggested as targets for anti-HCMV therapy [45], [46]. However, our data raise a possible caveat about the use of CDK inhibitors as antiviral drugs because they might favour HCMV induction in non-permissive cell types and during latent or latent-like infections.
Several lines of evidence argue that the CDK activity leading to inhibition of MIE gene expression in S/G2 is provided by Cyclin A2-CDK1/2. First, CDK1 and CDK2 are the only CDKs inhibited by both p21 (binds CDK1, 2, 4 and 6 [47]) and roscovitine (targets CDK1, 2, 5, 7 and 9 [48]). Second, Cyclin A2-dependent kinase activity is induced at the G1/S transition and remains high until early prometaphase [49], thus constituting the only Cyclin-CDK activity profile that matches the non-permissive cell cycle window for HCMV. Third, HCMV induces high levels of Cyclin E and B-dependent kinase activity but represses Cyclin A2 in an IE2-dependent manner [11], [50], [51], [52]. This points towards a model whereby HCMV – as soon as MIE gene expression has started – evades the inhibitory influence of Cyclin A2-CDK activity. Also in the case of latent-like infection the Cyclin A2-CDK status appears to perfectly correlate with the ability of cells to support MIE gene expression. Retinoic acid-induced differentiation which prevents the establishment of MIE quiescence in NT2 cells is also known to cause downregulation of Cyclin A2 (see also figure 11A), induction of the CDK inhibitor p27 and inhibition of CDK2 activity [37]. Work is ongoing in our laboratory to determine the role of Cyclin A2-dependent kinase activity during lytic infection as well in experimental models of HCMV latency.
To understand the mechanism of CDK dependent inhibition of HCMV MIE gene expression it will be essential to know the CDK substrate mediating this control. One possibility is that a viral regulator of MIE gene expression changes its activity, expression or localization depending on the CDK status of the host cell. Examples for such viral sensors of cellular CDK activity are the bovine papillomavirus (BPV) protein E1 and the apoptin protein of chicken anemia virus. BPV-E1, an essential factor for the initiation of viral replication, is inactivated and marked for nuclear export in S/G2 by Cyclin A2-CDK2-dependent phosphorylation [53]. In contrast, apoptin, a tumor selective inducer of apoptosis, requires phosphorylation by Cyclin A2-CDK2 for nuclear localization and cell death induction [54]. In the case of HCMV, the CDK substrate needs to be present at pre-IE times of infection, so the best viral candidate factors would be tegument proteins which are delivered as part of the HCMV virion to the host cell. Importantly, a number of tegument proteins have already been decribed to exert functions prior to MIE gene expression [55] and therefore represent possible targets for such a claimed CDK-dependent mechanism. Candidates include pUS24 [56], pUL26 [57], [58], pUL28/29 [59], pUL35 [60], [61], pUL47 [62], pUL76 [63]. The differentiation-dependent subcellular localization of pp71 that regulates its Daxx-neutralizing function in NT2 and THP-1 cells [38], [39] already provides a proof of principle for the control of HCMV tegument proteins by cellular factors immediately after infection. Moreover, the example of the herpes simplex virus protein VP16 demonstrates that the availability of a tegument protein can be decisive for the cell cycle sensitivity of herpesviral IE gene expression [64].
How might HCMV benefit from a CDK-sensitive mechanism controlling MIE gene expression? In a simple view, it enables the virus to synchronize the onset of its lytic cycle with G0/G1 - the cell cycle phase that is considered to be most supportive for virus replication. An alternative scenario is that CDK activity is a downstream constituent of a p53 sensitive switch that can operate between latent and lytic infection. In this case cell cycle dependency of HCMV may be an additional by-product. P53 is a coordinator of cellular responses to different kinds of stress including inflammatory [65] and oncogenic stress [66]. Given that HCMV is frequently found reactivated in inflammatory diseases and cancer [67], the possibility that activated p53 triggers lytic gene expression via CDK inhibition appears an attractive option.
Human embryonic lung (HEL) fibroblasts (Fi301, obtained from the Institute of Virology, Charité, Berlin, Germany) were maintained in Eagle's minimum essential medium (EMEM) supplemented with Earle's balanced salt solution, 25 mM HEPES, 1 mM sodium pyruvate, 2 mM L-alanyl-L-glutamine, nonessential amino acids, 0.75 ‰ (w/v) sodium bicarbonate, 50 µg/ml gentamicin and 10% fetal bovine serum (FBS). Human umbilical vein endothelial cells (HUVEC) were obtained from Lonza (Walkersville, MD, USA) and maintained in EGM medium (Lonza) following the manufacturer's instructions. The generation of stable p53 and GFP knockdown derivatives was described elsewhere [17]. U373-MG, HEK293 and HeLa cells (all from ATCC, Manassas, VA, USA) were maintained in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% FBS, 2 mM L-alanyl-L-glutamine, 100 U/ml penicillin and 100 µg/ml streptomycin. The human colon carcinoma cell line HCT116 and its p53−/− and p21−/− derivatives were obtained from Bert Vogelstein (Baltimore, MD, USA) and maintained in McCoy's 5a medium supplemented with 10% FBS, 5 mM glutamine, 100 U/ml penicillin, and 100 µg/ml streptomycin. The human teratocarcinoma cell line NTERA-2 (NT2) was obtained from DSMZ (Braunschweig, Germany) and cultivated on gelatin-coated dishes in DMEM supplemented with 10% FBS, 5% horse serum, 100 units/ml penicillin, and 100 µg/ml streptomycin. Differentiation of NT2 cells was induced with 10 µM retinoic acid (Sigma-Aldrich, St. Louis, MO, USA) as described elsewhere (Plotkin, 1984). Where indicated, the following reagents were added to the cell culture medium: aphidicolin (Sigma-Aldrich, final concentration: 5 µg/ml), hydroxyurea (Sigma-Aldrich, 1 mM), roscovitine (Millipore-Calbiochem, 5–50 µM), CVT313 (purchased as CDK2 inhibitor III from Millipore-Calbiochem, 10–50 µM), SU9516 (Santa Cruz Biotechnology, Santa Cruz, CA, USA, 2–10 µM), trichostatin A (Sigma-Aldrich, 100 ng/ml), MG132 (Sigma-Aldrich, 2.5 µM). To stop treatment, cells were washed several times with normal growth medium.
The HCMV strains AD169, Davis and Towne were purchased from ATCC (Manassas, VA, USA). Merlin and Toledo strains were a gift from Gavin Wilkinson (Cardiff, UK). The endotheliotropic isolates TB40/e and VHL/e were a gift from Christian Sinzger (Tübingen, Germany). All strains were grown on HEL fibroblasts. Virus titers were determined by IE1/IE2-fluorescence, essentially as described [68]. Briefly, quiescent HEL fibroblasts were infected with various dilutions of virus stocks. After 24 h of incubation, cells were fixed and stained with IE1/IE2-specific antibody. Subsequently, the number of positive cells was determined by flow cytometry and used to calculate viral titers. Unless otherwise stated, a multiplicity of infection (MOI) of 5 IE protein forming units (IEU) per cell was used for infection experiments. Where indicated, cells were pulse-labelled (60 min) with 10 µM 5-ethynyl-2'-desoxyuridine (EdU, Invitrogen, Carlsbad, CA, USA) before infection to label S phase cells. Non-incorporated EdU was removed by several washes with normal growth medium.
Doxorubicin (Sigma-Aldrich) was added to the cell culture medium to a final concentration of 1 µM. After two hours, cells were washed once with phosphate buffered saline (PBS) and fed with fresh culture medium. UV-irradiation was carried out using an UV-Stratalinker 2400 (Stratagene) equipped with 254 nm UV-light bulbs. For the duration of UV-exposure the culture medium was replaced with PBS (0.04 ml/cm2) to avoid the generation of toxic medium-derived photoproducts. A dosage of 10 J/m2 UV-C light was applied which leads in the HEL fibroblasts we used to high levels of p21 induction (data not shown).
Cells were harvested by trypsinization, fixed and permeabilized by incubation in 75% ethanol for at least 12 h at 4°C and stained with specific antibodies and propidium iodide as described previously [13]. The following primary antibodies were used: anti-IE1/IE2 (clone E13, Argene, Verniolle, France), anti-gB (1-M-12, Santa Cruz), anti-pp28 (CH19, Santa Cruz), Oct3/4 (C-10, Santa Cruz). An Alexa Fluor 488-conjugated goat anti-mouse IgG antibody (Invitrogen) was used as secondary reagent. Isotype-specific antibodies were used for co-staining of IE1/2 (Alexa Fluor 546-conjugated goat anti-mouse IgG11, Invitrogen) and Oct3/4 (Alexa Fluor 488-conjugated goat anti-mouse IgG12b, Invitrogen). EdU-positive cells were detected using the Click-iT EdU Alexa Fluor 647 imaging kit (Invitrogen) according to the manufacturer's instructions. Cells were analyzed on FACScan or FACSCanto2 flow cytometers (BD Biosciences, San Jose, CA, USA) using CellQuest and FACSDiva software packages respectively. Cell doublets and aggregates were gated out of analysis. All experiments were performed at least thrice and only representative results were shown.
Cells were lysed by sonication in 50 mM Tris-Cl (pH 6.8)–2% sodium dodecyl sulfate–10% glycerol–1 mM dithiothreitol–2 µg/ml aprotinin–10 µg/ml leupeptin–1 µM pepstatin–0.1 mM Pefabloc. The lysates were clarified by centrifugation at 17,500×g and protein concentration was determined using the Bio-Rad DC protein assay (Bio-Rad Laboratories, Hercules, CA, USA). Lysates were then adjusted to equal protein concentration, supplemented with 100 mM dithiothreitol and bromophenol blue, and heated to 95°C for 3–5 min. Sodium dodecyl sulfate-polyacrylamide gel electrophoresis and immunoblotting were performed accorded to standard procedures. The following primary antibodies were used: anti-p53 (clone DO-1, Santa Cruz), anti-p21 (C-19, Santa Cruz), anti-Cyclin A2 (C-19, Santa Cruz), anti-GAPDH (mAbcam 9484, Abcam, Cambridge, UK). HRP conjugated goat anti-mouse-IgG and goat anti-rabbit IgG (both Santa Cruz) served as secondary antibodies.
NT2 cells were grown on glass coverslips and treated as described above. After harvest of the coverslips, cells were washed, fixed, permeabilized and immunostained exactly as described [38]. A mouse monoclonal antibody against pp71 (clone 2H10, kindly provided by Tom Shenk, Princeton, NJ, USA) and an Alexa Fluor-488- coupled goat anti-mouse IgG (Invitrogen) were used as primary and secondary reagents. Nuclei were counterstained by the use of 4′,6-diamidin-2-phenylindol (DAPI). Images were acquired by an Eclipse A1 laser scanning microscope using NIS-Elements software (Nikon Instruments, Tokyo, Japan). Equal microscope settings and exposure times were used to allow direct comparison between samples.
|
10.1371/journal.pcbi.1002889 | Temporal Adaptation Enhances Efficient Contrast Gain Control on Natural Images | Divisive normalization in primary visual cortex has been linked to adaptation to natural image statistics in accordance to Barlow's redundancy reduction hypothesis. Using recent advances in natural image modeling, we show that the previously studied static model of divisive normalization is rather inefficient in reducing local contrast correlations, but that a simple temporal contrast adaptation mechanism of the half-saturation constant can substantially increase its efficiency. Our findings reveal the experimentally observed temporal dynamics of divisive normalization to be critical for redundancy reduction.
| The redundancy reduction hypothesis postulates that neural representations adapt to sensory input statistics such that their responses become as statistically independent as possible. Based on this hypothesis, many properties of early visual neurons—like orientation selectivity or divisive normalization—have been linked to natural image statistics. Divisive normalization, in particular, models a widely observed neural response property: The divisive inhibition of a single neuron by a pool of others. This mechanism has been shown to reduce the redundancy among neural responses to typical contrast dependencies in natural images. Here, we show that the standard model of divisive normalization achieves substantially less redundancy reduction than a theoretically optimal mechanism called radial factorization. On the other hand, we find that radial factorization is inconsistent with existing neurophysiological observations. As a solution we suggest a new physiologically plausible modification of the standard model which accounts for the dynamics of the visual input by adapting to local contrasts during fixations. In this way the dynamic version of the standard model achieves almost optimal redundancy reduction performance. Our results imply that the dynamics of natural viewing conditions are critical for testing the role of divisive normalization for redundancy reduction.
| It is a long-standing hypothesis that the computational goal of the early visual processing stages is to reduce redundancies which are abundantly present in natural sensory signals [1], [2]. Redundancy reduction is a general information theoretic principle that plays an important role for many possible goals of sensory systems like maximizing the amount of information between stimulus and neural response [3], obtaining a probabilistic model of sensory signals [4], or learning a representation of hidden causes [3], [5]. For a population of neurons, redundancy reduction predicts that neuronal responses should be made as statistically independent from each other as possible [2].
Many prominent neural response properties such as receptive field structure or contrast gain control have been linked to redundancy reduction on natural images [2]. While an appropriate structure of linear receptive fields can always remove all redundancies caused by second order correlations, they have only little effect on the reduction of higher order statistical dependencies [6], [7]. However, one of the most prominent contrast gain control mechanisms—divisive normalization—has been demonstrated to reduce higher order correlations on natural images and sound [8]–[10]. Its central mechanism is a divisive rescaling of a single neuron's activity by that of a pool of other neurons [8, see also Figure 1a].
Recently, radial factorization and radial Gaussianization have been derived independently by [11] and [12], respectively, based on Barlow's redundancy reduction principle [1]. Both mechanisms share with divisive normalization the two main functional components, linear filtering and rescaling and have been shown to be the unique and optimal redundancy reduction mechanism for this class of transformations under certain symmetry assumptions for the data. Radial factorization is optimal for a more general symmetry class than radial Gaussianization [11], [13] and contains radial Gaussianization as a special case. As a consequence, radial factorization can achieve slightly better redundancy reduction for natural images than radial Gaussianization but the advantage is very small.
Here, we compare the redundancy reduction performance of divisive normalization to that of radial factorization in order to see to what extent divisive normalization can serve the goal of redundancy reduction. Our comparison shows that a non-adapting static divisive normalization is not powerful enough to capture the contrast dependencies of natural images. Furthermore, we show that (i) the shape of contrast response curves predicted by radial factorization is not consistent with that found in physiological recordings, and (ii) that for a static divisive normalization mechanism this inconsistency is a necessary consequence of strong redundancy reduction. Finally, we demonstrate that a dynamic adaptation of the half-saturation constant in divisive normalization may provide a physiologically plausible mechanism that can achieve close to optimal performance. Our proposed adaptation mechanism works via horizontal shifts of the contrast response curve along the log-contrast axis. Such shifts have been observed in experiments in response to a change of the ambient contrast level [14].
We now briefly introduce divisive normalization, radial factorization, and the information theoretic measure of redundancy used in this study.
In this study we have demonstrated that a static divisive normalization mechanism is not powerful enough to capture the contrast dependencies of natural images leading to a suboptimal redundancy reduction performance. Static divisive normalization could only exhibit close to optimal performance if the contrast distribution of the input data would be similar to a Naka-Rushton distribution that we derived in this paper. For the best fitting Naka-Rushton distribution, however, the interval containing most of the probability mass is too narrow and too close to zero compared to the contrast distribution empirically found for natural image patches. A divisive normalization mechanism that uses the -norm as in equation (3) instead of the Euclidean norm would suffer from the same problem because the Naka-Rushton distribution for -norms other than would have similar properties. However, the good performance of extended divisive normalization demonstrates that it is not necessary to model the contrast distribution perfectly everywhere but that it would be sufficient to match the range where most natural contrasts appear (Figure 1C).
Not every mapping on natural contrasts that achieves strong redundancy reduction is also physiologically plausible: We showed that the extended static mechanism yields physiologically implausible contrast response curves. Extending the static mechanism of divisive normalization for better redundancy reduction simply makes it more similar to the optimal mechanism and, therefore, yields implausible tuning curves as well. We thus suggested to consider temporal properties of divisive normalization and devised a model that can resolve this conflict by temporally adapting the half-saturation constant using temporal correlations between consecutive data points caused by fixations.
Another point concerning physiological plausibility is the relationship between divisive normalization models used to explain neurophysiological observations, and those used in redundancy reduction studies like ours. One very common neurophysiological model was introduced by Heeger [8] which uses half-squared instead of linear single responses:(6)In order to represent each possible image patch this model would need two neurons per filter: one for the positive part and one for the negative part . Of course, these two units would be strongly anti-correlated since only one can be nonzero at a given point in time. Therefore, taking a redundancy reduction view requires considering the positive and the negative part. For this reason it is reasonable to use as the most basic unit and define the normalization as in equation (2). Since and are just two different representations of the same information, the multi-information between is the same as the multi-information between different tuples . Apart from this change of viewpoint, the two models are equivalent, because the normalized half-squared response of equation (6) can be obtained by half-squaring the normalized response of equation (2). Therefore, a model equivalent to the one in equation (6) can be obtained by using the model of equation (2) and representing its responses by twice as many half-squared coefficients afterwards.
Previous work on the role of contrast gain control for efficient coding has either focused on the temporal domain [26], [27], or on its role in the spatial domain as a redundancy reduction mechanism for contrast correlations in natural images [9], [11], [12]. Our results emphasize the importance of combining both approaches by showing that the temporal properties of the contrast gain control mechanism can have a critical effect on the redundancies that originate from the spatial contrast correlations in natural images. Our analysis does not commit to a certain physiological implementation or biophysical constraints, but it demonstrates that the statistics of natural images require more degrees of freedom for redundancy reduction in a population response than a classical static divisive normalization model can offer. Our heuristic mechanism demonstrates that strong redundancy reduction is possible with an adaptation mechanism that faces realistic conditions, i.e. has only access to stimuli encountered in the past.
As we showed above, biologically plausible shapes of the contrast response curve and strong redundancy reduction cannot be easily brought together in a single model. Our dynamical model offers a possible solution to this problem. To what extent this model reflects the physiological reality, however, still needs to be tested experimentally.
The first aspect to test is whether the adaptation of the half-saturation constant reflects the temporal structure imprinted by saccades and fixations as predicted by our study. Previous work has measured adaptation timescales for [14], [28]. However, these measurements are carried out in anesthetized animals and cannot account for eye movements. Since our adaptation mechanism mainly uses the fact that contrasts at a particular fixation location are very similar it predicts that that adaptive changes of should be seen from one fixation location to another when measured under natural viewing conditions.
The mechanism we proposed is only one possible candidate for a dynamic contrast gain control mechanism that can achieve strong redundancy reduction. We conclude the paper with defining a measure that can be used to distinguish contrast gain control mechanisms that are likely to achieve strong redundancy reduction from those that do not. As discussed above, a necessary condition for strong redundancy reduction is that the the location and the width of the distribution of implied by a model must match the distribution of unnormalized responses determined by the statistics of natural images. In order to measure the location and the width of the distributions in a way that does not depend on a particular scaling of the data, we plotted the median against the width of the ––percentile interval (Figure 5). For the empirical distributions generated by the statistics of the image data we always found a ratio greater than . We also included a dataset from real human eye movements by Kienzle et al. to ensure the generality of this finding [29] as real fixations could introduce a change in the statistics due to the fact that real observers tend to look at image regions with higher contrasts [30]. All models that yield strong redundancy reduction also exhibit a ratio greater than . Thus, the ratio of the median to the width of the contrast distribution is a simple signature that can be used to check whether an adaptation mechanism is potentially powerful enough for near-optimal redundancy reduction.
The code and the data are available online under http://www.bethgelab.org/code/sinz2012.
Both the divisive normalization model and the optimal radial factorization consist of two steps: a linear filtering step and a radial rescaling step (Table 1). In the following, we describe the different steps in more detail.
We use the multi-information to quantify the statistical dependencies between the filter responses [38]. The multi-information is the -dimensional generalization of the mutual-information. It is defined as the Kullback-Leibler divergence between the joint distribution and the product of its marginals or, equivalently, the difference between the sum of the marginal entropies and the joint entropy(9)The multi-information is zero if and only if the different dimensions of the random vector are independent. Since the joint entropy is hard to estimate we employ a semi-parametric estimate of the multi-information that is conservative in the sense that it is downward biased.
For the marginal entropies , we use a jackknifed estimator for the discrete entropy on the binned values [39]. We chose the bin size with the heuristic proposed by Scott [40]. We obtain an estimate for the differential entropy by correcting with the logarithm of the bin width (see e.g. [7]).
In order to estimate the joint entropy, we use the average log-loss to get an upper boundSince the average log-loss overestimates the true entropy, replacing the joint entropy by in equation (1) underestimates the multi-information. Therefore, we sometimes get estimates smaller than zero. Since the multi-information is always positive, we set the value to zero in that case. For computing errorbars on the multi-information estimations, we use the negative values but a mean zero in such cases, which effectively increases the standard deviation of the error.
Since we want commit ourselves as little as possible to a particular model, we estimate by making the assumption that is -spherically symmetric distributed but estimating everything else with non-parametric estimators. If is -spherically symmetric distributed, the radial component is independent from the directional component [32] and we can write(10)The entropy of the radial component is again estimated via a histogram estimator. The term is approximated by the empirical mean.
Putting all the equations together yields our estimator for the multi-information under the assumption of -spherically symmetric distributed where are the univariate entropies estimated via binning.
Since the optimal value of for filter responses to natural image patches is approximately we use that value to estimate the multi-information of .
When estimating the multi-information of the responses of either divisive normalization or radial factorization, we use the fact thatwhere is the Jacobian of the normalization transformation. The mean is estimated by averaging over data points. The determinants of radial factorization, divisive normalization, and extended divisive normalization are given byAll multi-information values were computed on test data.
For the dynamically adapting model, the for each data point is sampled from a -distribution whose parameters are determined from the previous value and the posterior over obtained from the mixture of Naka-Rushton distributions. Since changes from step to step it becomes part of the representation and should be included when computing the multi-information (i.e. the redundancy) between the outputs . Therefore, the redundancy for the dynamically adapting model is measured by . For its computation, we use that , where is the mutual information between and . In the following, we write if . Under the assumption that both and are spherically symmetric distributed, we can decompose respective random variables into the uniform (on the sphere) and the radial part: and . This yieldswhich means that we can restrict ourselves to the mutual information between the two univariate signals and , which we estimate from a two-dimensional histogram with bins.
|
10.1371/journal.pntd.0000637 | Influence of Exposure History on the Immunology and Development of Resistance to Human Schistosomiasis Mansoni | Previous studies suggest that humans can acquire immunity to reinfection with schistosomes, most probably due to immunologic mechanisms acquired after exposure to dying schistosome worms.
We followed longitudinally two cohorts of adult males occupationally exposed to Schistosoma mansoni by washing cars (120 men) or harvesting sand (53 men) in Lake Victoria. Men were treated with praziquantel each time S. mansoni infection was detected. In car washers, a significant increase in resistance to reinfection, as measured by the number of cars washed between cure and reinfection, was observed after the car washers had experienced, on average, seven cures. In the car washers who developed resistance, the level of schistosome-specific IgE increased between baseline and the time at which development of resistance was first evidenced. In the sand harvesters, a significant increase in resistance, as measured by the number of days worked in the lake between cure and reinfection, was observed after only two cures. History of exposure to S. mansoni differed between the two cohorts, with the majority of sand harvesters being lifelong residents of a village endemic for S. mansoni and the majority of car washers having little exposure to the lake before they began washing cars. Immune responses at study entry were indicative of more recent infections in car washers and more chronic infections in sand harvesters.
Resistance to reinfection with S. mansoni can be acquired or augmented by adults after multiple rounds of reinfection and cure, but the rate at which resistance is acquired by this means depends on immunologic status and history of exposure to S. mansoni infection.
| Schistosomiasis is a parasitic blood fluke infection of 200 million people worldwide. We have shown that humans can acquire immunity to reinfection after repeated exposures and cures with the drug praziquantel. The increase in resistance to reinfection was associated with an increase in schistosome-specific IgE. The ability to develop resistance and the rate at which resistance was acquired varied greatly in two cohorts of men within close geographic proximity and with similar occupational exposures to schistosomes. These differences are likely attributable to differences in history of exposure to Schistosoma mansoni infection and immunologic status at baseline, with those acquiring immunity faster having lifelong S. mansoni exposure and immunologic evidence of chronic S. mansoni infection. As many conflicting results have been reported in the literature regarding immunologic parameters associated with the development of resistance to schistosome infection, exposure history and prior immune status should be considered in the design of future immuno-epidemiologic studies.
| Schistosoma mansoni age-infection curves in endemic human populations characteristically show a peak prevalence in children and early adolescence and then a decline beginning in the late teenage years to lower levels of prevalence among adults [1]. This has led many researchers to hypothesize that humans can acquire immunity to S. mansoni, leading to partial resistance against reinfection [2]. Since the natural lifespan of S. mansoni worms is approximately 5–10 years [3],[4], the decline in prevalence coincides with the time at which worms acquired in early childhood would naturally begin to die in persons living in endemic areas. One theory holds that upon worm death, either naturally or as a result of treatment, critical schistosome antigens not normally or appropriately encountered by the host during chronic infection are released. The release of these antigens alters the immune response patterns that result from exposure to intact worms [5],[6], and it is hypothesized that these changes in immune responses are responsible for the increased resistance to reinfection [2].
We previously reported the age-independent development of immunological resistance to reinfection with S. mansoni in a cohort of adult males occupationally exposed, by washing cars in Lake Victoria, undergoing repeated cycles of reinfection and praziquantel-induced cure [7]. Resistance to reinfection by all three of the schistosome spcies that cause most human disease has been associated with both cellular [8],[9],[10] and humoral immune responses, most notably IgE in response to parasite-specific antigens [11]–[16]. In turn, variations in these immune responses have been related to factors such as age, stage of disease, and duration of infection [17]–[24].
More recently, we have expanded our studies to include a second cohort of men who are also exposed to infectious water through their occupation of harvesting sand in Lake Victoria. Upon discovering differences in the two cohorts in the number of treatments and cures needed before increased resistance to reinfection was demonstrated, we explored demographic and immunologic factors that may explain the discrepancies.
All participants in this study were adult males occupationally exposed to S. mansoni by washing cars or harvesting sand on the shores of Lake Victoria near Kisumu, Kenya. The car washers stand ankle- to knee-deep in the lake to wash cars that have been driven into the shallow water at the edge of the lake. Enrollment of car washers began in June 1995, and follow-up continued until January 2009. With the exception of the period between January 2000 and September 2003, enrollment of new car washers was continuous throughout the duration of the study, so follow-up time varies for each individual.
The sand harvesters stand waist- to chest-deep in the water to shovel sand off the bottom of the lake. After filling their boats with sand, they then transport the sand to shore and stand in the water at the edge of the lake while they unload the sand onto the shore. Recruitment of sand harvesters began in March 2005, and follow-up continued until January 2009. Both groups of men are ethnically homogeneous, with 90% of the car washers and 98% of the sand harvesters belonging to the Luo tribe.
The car washing and sand harvesting sites are shown in Figure 1. The carwash is adjacent to the city of Kisumu, and the site is a busy area also populated with fishermen, fish merchants, and various other vendors. Although located only 5.2 km around the lakeshore and 3 km across the lake, the sand harvesting site differs considerably as it is located off the shores of the small fishing village of Usoma, a rural community separated and distinct from the city of Kisumu. The presence of S. mansoni-infected Biomphalaria sudanica snails has been confirmed at both exposure sites [25],[26].
All study participants gave written informed consent prior to enrollment. Study procedures were approved by the institutional review boards of the University of Georgia and the Centers for Disease Control and Prevention, the Scientific Steering Committee of the Kenya Medical Research Institute (KEMRI), and the KEMRI/National Ethics Review Committee of Kenya.
Upon enrollment, men were tested for S. mansoni eggs by the modified Kato-Katz method using two slides from each of three consecutive stool samples. Individuals positive for S. mansoni were treated with 40 mg/kg praziquantel (PZQ), and follow-up stool samples were taken 4–6 weeks later to assess for cure. If necessary, men were re-treated with PZQ until cure was demonstrated by three consecutive stool samples that were negative for schistosome eggs. Upon becoming stool negative, men were continually followed and retested for S. mansoni eggs at 4-week intervals. Each time a new infection was found, the study participant was treated with PZQ until he demonstrated cure.
Blood samples were taken every six months for subjects enrolled in 2003 or later and approximately yearly for car washers enrolled prior to 2003. Blood was tested for schistosome-specific antibodies, HIV-1 specific antibodies, and the ability of their peripheral blood mononuclear cells (PBMCs) to produce cytokines [7],[27]. The prevalence of malaria and soil-transmitted helminths in these populations was low. In the rare event malaria or soil-transmitted helminths were found the subjects were offered appropriate treatment.
Water exposure was measured by the number of cars washed or the number of days worked in the lake harvesting sand. Daily records of the number of cars washed by each car washer or the number of hours worked each day harvesting sand for each sand harvester were kept by on-site members of the carwash and sand harvester consortia who were employed as field workers for the present study. Since the number of hours spent in the water each day for sand harvesters was highly consistent (mean 5.3±0.9 hours), and sand harvesters likely receive most of their exposure to schistosomes as they are standing near the edge of the lake unloading the sand from their boats rather than when they are harvesting sand in waist- to chest-deep water away from the shore, we have chosen to use days worked rather than hours worked in water exposure calculations for the sand harvesters. Sand harvesters were given credit for one day of work for each day that they worked for at least one hour. It is important to note that one car washed is not equivalent to one day of work harvesting sand, thus direct comparisons between the two groups of men are not appropriate,
Isolation of PBMCs and cell cultures were performed as previously described [28]. Briefly, PBMCs were separated from venous blood using the ficoll-hypaque technique. PBMCs were washed and resuspended in RPMI containing 5% AB+ normal human sera, antibiotics and L-glutamine. The cells were incubated with 10 µg/ml soluble worm antigen preparation (SWAP) or 5 µg/ml soluble egg antigens (SEA) for five days at 37C in 5% CO2 and the supernatant fluids collected. PBMC production of the cytokines interleukin (IL)-5, IL-10, IL-13, and IFN-γ in response to SWAP and SEA was measured by capture ELISA using commercially-available kits (R&D Systems, Minneapolis, MN) according to manufacturer's instructions. Cytokine production was only performed on blood samples obtained after October 2003.
Anti-SWAP IgE isotype ELISAs were performed on plasma from the venous blood samples as previously described [29],[30]. External positive and negative controls (EC) comprised of pooled samples of high responders and normal human serum (NHS) from non-endemic volunteers, respectively, were run on each plate. Anti-SWAP IgE values for each sample were standardized according to the following formula:
IgE-specific ELISAs against the recombinant antigens ‘tegument allergy like’ (TAL)-1 (formerly Sm22.6) and TAL-2 (formerly Sm21.7) [31] were performed on a subset of baseline samples from 23 car washers and 20 sand harvesters. TAL-1 and TAL-2 were cloned and purified as previously described [32],[33]. ELISA plates were coated with recombinant antigen at 2 µg/ml. Following incubation with plasma samples (1∶20 dilution), antigen-specific IgE binding was measured using directly conjugated mouse anti-human IgE (Southern Biotech, Birmingham, AL).
Since almost all measurements were non-normally distributed, the Wilcoxon rank sum test was used for group comparisons, and the Wilcoxon sign rank test was used for paired comparisons of the same subjects at different time points. An alpha level of 0.05 was considered statistically significant for all comparisons. All analyses were performed with GraphPad Prism 5 or SAS version 9.1.
The number of cars washed or days worked harvesting sand between each cure and reinfection was estimated in an accelerated failure time model with the LIFEREG procedure in SAS [34]. Each infection interval was defined as the time between the documentation of cure and subsequent reinfection. Thus, “interval 1” is the interval between the first cure after study entry and the first reinfection following the first cure, “interval 2” is the interval between the time of the second cure and second reinfection, and so forth. Interval number was entered into the model as a categorical variable with interval 1 as the reference category. Thus the length of each cure-to-reinfection interval was statistically compared to that of the first interval.
The LIFEREG procedure can accommodate failure time data that is right- or left-censored. The first interval was considered left-censored for subjects negative at study entry. Intervals during which the subject left the study or follow-up ended before reinfection occurred were considered right-censored. Censored observations accounted for 59 of 570 total intervals (10.4%) among the car washers and 30 of 144 total intervals (20.8%) among the sand harvesters. Intervals during which more than three months elapsed between the last negative stool and a subsequent positive stool were excluded from the analyses, though other intervals from that same subject could be included. Entire subjects were excluded from the analysis if they did not have at least one complete infection interval—i.e. left the study without ever becoming egg-negative or after the initial cure but before the first reinfection.
Because daily records of car washing activities are incomplete prior to 1999, subjects whose entire follow-up occurred before February 1999 are not included in this analysis. For subjects enrolled before February 1999 and followed further, the cure-to-reinfection intervals occurring after February 1999 are included, beginning with the numbered interval that the subject had reached at that point.
The final study population consisted of 120 car washers with a mean follow-up time of 74.4 months (range: 9.1–165.5) and 53 sand harvesters with a mean follow-up time of 37.9 months (range: 12.6–61.1). The mean number of cure-to-reinfection intervals was 6.5 (range: 1–18) and 3.0 (range: 1–8) for the car washers and sand harvesters, respectively.
For each car washer, the number of reinfections per 100 cars washed (RCW) during the at-risk time over the course of follow-up was calculated as an indication of relative resistance to S. mansoni reinfection. For the sand harvesters, this measure was calculated as the number of reinfections per 100 days worked harvesting sand (RDW) during the at-risk time over the course of follow-up. At-risk time is the time between cure and reinfection. Cars washed (or days worked) in the time between infection and cure are not included in the RCW or RDW calculations. As the RCW or RDW is averaged over the entire duration of follow-up, in theory those men who enter the study with a higher level of resistance or develop resistance over the course of the study will have a lower RCW or RDW than men who retain a high degree of susceptibility over the course of follow-up. For some analyses, car washers and sand harvesters are dichotomized based on the mean RCW or RDW of each respective group. For ease of discussion, men with a below-mean number of reinfections are referred to as “more resistant phenotype,” and men with an above-mean number of reinfections are referred to as “more susceptible phenotype.” Factors associated with having a more resistant phenotype were evaluated in a logistic regression model.
The RCW or RDW for each car washer and sand harvester is plotted in Figure 6. The mean RCW for the car washers was 0.29 infections per 100 cars washed, with the individual RCWs uniformly distributed around the mean. Conversely, the sand harvesters exhibited a skewed pattern of resistance indexes, with the majority of the individual RDWs concentrated below the mean of 0.79 infections per 100 days worked, and only a few men with higher outlying RDWs.
Figure 7 shows the median number of cars washed in the intervals between each successive cure and reinfection. The figure depicts all car washers together (Figure 7a), and also stratified into more resistant (Figure 7b) and more susceptible (Figure 7c) phenotypes based on being below or above the mean RCW, respectively. For the entire cohort of car washers, the number of cars washed before reinfection was relatively stable until the seventh cure, at which point the number of cars washed between cure and reinfection begins to progressively increase with each successive cure. In the seventh cure-to-reinfection interval, and each interval thereafter, the number of cars is significantly greater than the number of cars washed in the interval between the initial cure and the first reinfection. When the car washers were stratified based on the RCW, those with the more resistant phenotype (Figure 7b) showed a pattern of increasing cure-to-reinfection intervals similar to that seen in the overall cohort. With the exception of interval ten (p = 0.0922), the median number of cars washed in each cure-to-reinfection interval after the eighth cure in the more resistant phenotype group was significantly greater than the initial interval (p-value range: 0.0005–0.0195). However, a pattern of increasing number of cars per cure-to-reinfection interval was not seen in the group of men with the more susceptible phenotype (Figure 7c). While some later intervals were significantly greater than the initial interval, overall these men did not, by the end of the study, exhibit a consistent pattern of increased resistance to reinfection upon repeated cures.
The median number of days worked in Lake Victoria between each cure and reinfection for all sand harvesters are shown in Figure 8a. The number of days in the interval between the second cure and second reinfection was significantly increased relative to the initial interval (p = 0.0118). Thus, as opposed to the car washers, the increase in resistance occurred in the sand harvesters after having experienced only two previous cures. This pattern was true for men with both more resistant and more susceptible phenotypes (Figures 8b–8c), though the more susceptible men started with a lower initial days worked to reinfection, and days to reinfection remained lower throughout follow-up. Although the graph appears to show a trend towards increased susceptibility after three previous cures, intervals 3–5 do not have significantly fewer days worked than interval two, and the apparent decrease is likely due to low numbers of subjects and high numbers of censored observations in intervals three and above.
Previous research by our group has shown that among men similarly exposed to S. mansoni by virtue of their occupation as car washers in infectious waters of Lake Victoria, a portion of the men developed resistance to reinfection after multiple rounds of cures and reinfections, while others remained susceptible despite equal or greater numbers of cures and reinfections [7]. We now show that the same observation holds true with a modified definition of resistance, based on exposure rather than time-to-reinfection, and after the addition of another cohort of men at the same carwash and additional follow-up of the original cohort. If car washers were maximally immune or non-immune at study entry, we would not have observed a progressively increasing number of cars before each reinfection as we did in many of the cohort, suggesting that these men are actively developing resistance. Those car washers who developed resistance began to do so after experiencing an average of seven previous cures.
However, a different pattern of the development of resistance emerged in a different cohort of men who receive daily exposure to schistosomes by harvesting sand in the lake just three km across the lake from the car washing site. In these men, the interval between cure and reinfection significantly increased after only two previous cures, after which point there were no further increases in the number of days worked between cure and reinfection, suggesting that no further increases in resistance occurred.
The numerical values of reinfections per 100 cars washed and reinfections per 100 days worked harvesting sand are not directly comparable as the S. mansoni transmission situation is different for each cohort. The sand harvesters spend on average 5.3 hours in the lake each work day, while the car washers wash an average of 3.2 cars per work day. However, the water at the car washer site is probably more heavily contaminated with S. mansoni cercariae, as the prevalence of infection in B. sudanica snails collected at the car wash site is higher than prevalence in snails collected at the sand harvesting site in Usoma [26]. Also, much of the sand harvesters' time is spent in deeper water, away from the shore where snails are not as likely to be present, so the majority of their cercarial exposure likely occurs during the time they are unloading sand onto the shore. Despite the differences in exposure between cohorts, the definition of resistance for each cohort is valid for comparisons within that cohort, and overall patterns of resistance should be comparable between the two cohorts.
The car washers exhibited a wide range of overall resistance levels according to the distribution of number of reinfections per 100 cars washed, which were fairly symmetrically distributed from the low to the high end of the spectrum. Although some car washers in the more susceptible group appeared to develop resistance after multiple cures and some might have eventually become resistant with longer follow-up, the general pattern observed was a gradual increase in resistance among those men who experienced a below-mean number of reinfections and no apparent consistent increase in resistance among men who experienced an above-mean number of reinfections. In contrast, the distribution of number of reinfections per 100 days worked for the sand harvesters was much less uniform, with the majority of the sand harvesters clustering towards fewer reinfections, indicating that a majority of the cohort entered the study with similar relatively high levels of resistance. Albeit to a much less degree than those in the more resistant group, even those relatively more susceptible sand harvesters exhibited an increase in time-to-reinfection after two previous cures.
The different histories of S. mansoni exposure in these two groups of men prior to study enrollment likely explain the differences in development of resistance upon multiple rounds of treatment and reinfection. The car washers reported working in the lake a mean of 5.7 years, while the sand harvesters had worked in the lake for a mean of 11.1 years. Moreover, while the majority of car washers were lifelong residents of the city of Kisumu or immigrants from other areas of Kenya, almost all of the sand harvesters were born in Usoma, the lakeside village where they now harvest sand. S. mansoni infection has been seen in children in Usoma as early as one year of age, with >90% becoming positive for antibodies to schistosomes by age 10 (J. Verani, unpublished data). A similar situation has been reported among children in fishing villages along the Ugandan shoreline of Lake Victoria, where Odogwu and colleagues found S. mansoni infection in 25% and 86% of children aged <3 years in two endemic villages [35]. Thus, men from Usoma likely had exposure to the lake as children long before they began working as sand harvesters, were probably initially infected with S. mansoni at an early age, and had likely experienced the natural death of worms multiple times prior to being treated as part of this study. In contrast, S. mansoni infections present at study entry in car washers likely represent more recent infections, and they had likely experienced the death of no or few worms prior to treatment with praziquantel.
These two groups of occupationally exposed adult males also differed considerably in their immune response patterns to schistosome antigens, and these differences are also likely explained by their different histories of exposure to S. mansoni. The baseline immune responses are suggestive of more recent infections in car washers. PBMC cytokine production in response to SEA at the time of enrollment was higher in car washers than amongst sand harvesters by all four measured cytokines. High responses to SEA have been associated with early S. mansoni infection, and these responses then decrease as infection becomes more chronic and exposure to constantly released egg antigens leads to development of immunoregulatory mechanisms [19],[20],[22],[23].
Similar to other researchers who have shown no difference in humoral responses to crude worm antigens in patients with early and chronic schistosomiasis [19], baseline anti-SWAP IgE responses did not differ between our cohorts. However, in both car washers and sand harvesters, older men had significantly higher levels of anti-SWAP IgE than did younger men, independent of exposure history. While increases in parasite-specific IgE with increased age are usually attributed to longer exposure to infection, Naus et al also reported increased IgE responses against schistosome worm antigens in older age groups in an immunologically naïve immigrant population recently arrived to an S. mansoni-endemic area of Kenya, suggesting that the increase may be innately age-related and not dependent on duration of schistosome infection [18].
Although baseline differences between car washers and sand harvesters were not seen in IgE responses to the heterogeneous worm antigens present in SWAP, differential IgE responses to the recombinant S. mansoni antigens TAL-1 and TAL-2 were observed between the two groups. Fitzsimmons et al have shown that TAL-1 expression is concentrated primarily in the adult worm, while TAL-2 is expressed on all life cycle stages, including miracidia, cercariae, and eggs [33]. Levels of anti-TAL-1 IgE antibodies were increased after treatment of S. mansoni infected individuals in the Fitzsimmons et al study, while anti-TAL-2 IgE antibodies were unchanged by treatment. The authors hypothesized that TAL-1 worm antigens are sequestered during active infection and are only released upon worm death. Conversely, the immune system is continuously exposed to TAL-2 due to the constant release of eggs during S. mansoni infection [32], thus leading to down regulation of responses to TAL-2. The current finding of higher pretreatment levels of anti-TAL-1 IgE in sand harvesters than in car washers and similarly low anti-TAL-2 responses in both groups fits this hypothesis. As the natural lifespan of an adult S. mansoni worm is approximately 5–10 years [3],[4], the car washers had likely not been exposed to any or many dying worms before receiving PZQ treatment as part of the current study, while the sand harvesters had likely already experienced multiple episodes of naturally dying worms, based on exposure since early childhood.
Car washers who were HIV positive at study entry were less likely to develop resistance over the course of follow-up than were men who were HIV negative. We previously reported that patients with schistosomiasis and HIV coinfection had significantly lower production of the cytokines IL-4 and IL-10 than schistosome-infected persons who were HIV negative [28]. Other researchers have reported an association between IL-4 production in response to schistosome antigens and increased resistance to reinfection with S. mansoni [8], Schistosoma haematobium [9], and Schistosoma japonicum [10]. HIV infection was not related to the ability to develop resistance in the sand harvesters, most probably because they had already been infected with and developed protective immune mechanisms against schistosomes prior to becoming infected with HIV as adults. While neither age nor number of years worked in Lake Victoria prior to study entry were associated with resistance among the car washers, only age was independently predictive of a resistant phenotype among the sand harvesters. As most sand harvesters likely had lake exposure since childhood before they began working harvesting sand, length of time worked in the lake became insignificant in the analysis after adjustment for age, as age is a better predictor for duration of water exposure in this group.
Many previous studies have shown various immune responses to be correlated with resistance to reinfection with all three species of schistosomes, most commonly the production of parasite-specific IgE [11], [14]–[16]. While we did not find any baseline antibody or cytokine responses to be predictive of the ability to develop resistance among the car washers, this was not unexpected given that a change in resistance did not become apparent until the men had experienced on average seven previous cures. However, among those car washers that did eventually demonstrate an increase in resistance against reinfection, we have documented increases in anti-SWAP IgE production that parallel the development of resistance. Increases in anti-SWAP IgE production did not occur in those who remained susceptible. We did not see a similar increase in anti-SWAP IgE as the interval between cure and reinfection increased in sand harvesters.
In conclusion, we have again demonstrated that resistance to reinfection with S. mansoni can be acquired or augmented by adults after multiple rounds of reinfection and PZQ-induced cure. However, we now also show that the ability to acquire this resistance and the rate at which resistance is acquired is markedly different in two populations within close geographic proximity to one another that share high levels of occupational exposure to S. mansoni infested water. These differences are likely attributable to differences in history of exposure to S. mansoni infection and their resulting immunologic status at baseline. As many conflicting results have been reported in the literature regarding immunologic parameters associated with the development of resistance to schistosome infection, these factors should be considered in the design of future immuno-epidemiologic studies and eventual vaccine study design.
|
10.1371/journal.pntd.0000934 | DC-SIGN (CD209) Promoter −336 A/G Polymorphism Is Associated with Dengue Hemorrhagic Fever and Correlated to DC-SIGN Expression and Immune Augmentation | The C-type lectin DC-SIGN (CD209) is known to be the major dengue receptor on human dendritic cells, and a single nucleotide polymorphism (SNP) in the promoter region of CD209 (−336 A/G; rs4804803) is susceptible to many infectious diseases. We reason that variations in the DC-SIGN gene might have a broad influence on viral replication and host immune responses.
We studied whether the rs4804803 SNP was associated with a susceptibility to dengue fever (DF) and/or dengue hemorrhagic fever (DHF) through genotyping analysis in a Taiwanese cohort. We generated monocyte-derived dendritic cells (MDDCs) from individuals with AA or AG genotype of rs4804803 to study the viral replication and immune responses for functional validation. A total of 574 DNA samples were genotyped, including 176 DF, 135 DHF, 143 other non-dengue febrile illnesses (OFI) and 120 population controls. A strong association between GG/AG genotypes of rs4804803 and risk of DHF was found when compared among DF, OFI and controls (p = 0.004, 3×10−5 and 0.001, respectively). The AA genotype was associated with protection against dengue infection compared with OFI and controls (p = 0.002 and 0.020, respectively). Moreover, MDDCs from individuals with AG genotype with a higher cell surface DC-SIGN expression had a significantly higher TNFα, IL-12p40, and IP-10 production than those with AA genotype in response to dengue infection. However, the viral replication in MDDCs with AG genotype was significantly lower than those with AA genotype. With both genotypes, MDDCs revealed an increase in viral replication following the addition of anti-IP-10 neutralizing antibody.
The rs4804803 SNP in the CD209 promoter contributed to susceptibility to dengue infection and complication of DHF. This SNP with AG genotype affects the cell surface DC-SIGN expression related to immune augmentation and less viral replication.
| Dengue fever (DF) is an arthropod-borne disease that is prevalent in tropical and subtropical regions of the world. DC-SIGN [dendritic cell-specific intercellular adhesion molecule 3 (ICAM-3)-grabbing non-integrin] is a major receptor for dengue infection. DC-SIGN, also called CD209, expresses on dendritic cells (DCs) that bind to ICAM-3, which is expressed on T cells to facilitate the initial interaction between DCs and T cells. Variations in the CD209 promoter (−336 A/G; rs4804803) genotype are involved in the pathogenesis of human infectious diseases. Here we found that patients with dengue hemorrhagic fever (DHF) had a higher frequency of the AG or GG genotype of rs4804803 than DF or controls. Functional studies determined that monocyte-derived DCs (MDDCs) from individuals with AG genotype had significantly higher cell surface DC-SIGN expression, associated with higher TNFα, IL-12p40, and IP-10 production, but lower viral replication than those with AA genotype. An increase in DEN-2 replication in MDDCs was observed following the addition of anti-IP-10 neutralizing antibody. These findings highlight the fact that the rs4804803 SNP in the CD209 promoter is associated with DHF and correlated to DC-SIGN expression and immune augmentation.
| Dengue viruses (DEN) are arthropod-borne flaviviruses that cause dengue fever (DF) with significant morbidity and mortality in tropical and subtropical regions of the world. There are four serotypes of dengue viruses (DEN types 1–4). Classic DF is a self-limited illness characterized by fever, headache, myalgia, arthralgia, and abdominal pain. Since the 1950s, a more severe form of the disease, dengue hemorrhagic fever (DHF), has been recognized worldwide [1]. Patients who develop DHF typically have initial symptoms similar to those in DF patients, but develop cytokine-storm-like plasma leakage manifested by hemoconcentration, thrombocytopenia, ascites, and pleural effusion near the defeverence stage [2]. DHF pathogenesis has been attributed to viral virulence versus immune enhancement; however, that has been the subject of debate for many years [2], [3].
The innate immune system is the first line of host defense against pathogens and is involved in early recognition and uptake of microbes by the host's professional phagocytes such as dendritic cells (DCs) and macrophages, through germline-encoded receptors, known as pattern recognition receptors (PRRs) [4]. These PRRs recognize microbial antigens and initiate innate immune responses followed by adaptive immunity [5]. PRRs are involved in phagocytosis, antigen presentation, and they trigger intracellular signaling and cytokine secretion [5]. PRR polymorphisms may therefore affect virus entry, replication, and immunity.
Among the PRRs, the CD209 molecule, also known as DC-SIGN (dendritic cell-specific intercellular adhesion molecule-3 grabbing non-integrin), plays an important role in the early interaction of a pathogen with a dendritic cell [6]–[8] and has a key role in DC-T cell interaction [9], DC migration [10], [11], and pathogen uptake [12]. DC-SIGN is organized into three domains, the N-terminal domain is located in the cytoplasm, the transmembrane domain anchors to the cytoplasmic membrane, and the extracellular domain consists of a neck region formed by seven highly conserved 23 amino acid repeats and a carbohydrate domain for pathogen binding [13].
The CD209 gene is located on chromosome 19p13.2–3 and is highly polymorphic. Numerous single nucleotide polymorphisms (SNPs) have been reported [14]–[18]. One of these SNPs represents a guanine (G) to adenine (A) transition at position −336 within the CD209 gene promoter (rs4804803). This variant has been associated with an increased risk for parenteral acquisition of human immunodeficiency virus (HIV) infection [15], severity of dengue infection [16], and confered high susceptibility to tuberculosis in a South African cohort [17]. Nevertheless, Vannberg et al. found that G variant allele of rs4804803 was associated with protection against tuberculosis in individuals from sub-Saharan Africa [18]. This variant affects CD209 promoter activity with multiple transcription factor binding sites for the Sp1/GATA1/CACCC- and CAC-binding transcription factors in a transfection study [16]. As an in vitro study of promoter activity might not reflect an actual functional association, we aimed to test whether the rs4804803 SNP in the promoter region of CD209 was associated with the susceptibility to DF and/or DHF in Taiwanese, and whether monocyte-derived DCs from humans with various genotypes of rs4804803 would reveal differences in DC-SIGN membrane expression and implicate the viral replication and immune reactions after DEN infection.
This study was approved by the Institution Review Board (IRB) of Chang Gung Memorial Hospital-Kaohsiung Medical Center, Taiwan. The dengue patients were recruited as described previously in the 2002–2003 DEN-2 outbreak in Taiwan [19]–[23]. A larger retrospective cohort was designed and re-approved by an additional IRB review (Document No.: 97-2111B). To validate cell surface expression and immune functions of rs4804803 SNP, we obtained informed consent to collect blood leukocytes from normal volunteers with AA or AG genotypes of rs4804803.
DEN infection was confirmed by clinical dengue symptoms and signs along with detection of DEN-2 RNA by quantitative RT-PCR in blood, detection of IgM to DEN or at least a 4-fold increase in dengue-specific hemagglutination inhibition titers in convalescent serum compared with that in acute-phase serum [20], [21]. In those with DEN-2 infection, blood was drawn at least once a day subsequent to admission into the hospital to measure the platelet counts and hematocrit levels. A Chest X-ray and abdominal ultrasonography were performed routinely in individuals without evidence of hemoconcentration or hypoalbuminemia to refine the differential diagnosis of DHF vs. DF based on pleural effusion or ascites. A clinical diagnosis of DHF was assigned according to the DHF criteria of the World Health Organization (WHO); including a reduced platelet count (<100,000/mm3), petechiae, hemorrhagic manifestation, and plasma leakage showing hemoconcentration (peak hematocrit ≥20% above the mean for the population, or an increase in hematocrit of 20% or more), pleural effusion, ascites, or hypoalbuminemia [24]. Patients with DF were defined by detectable DEN-2 RNA by RT-PCR or DEN-specific IgM, but without evidence of DHF. Primary or secondary DEN infections were identified using previously established serologic criteria for IgM/IgG ELISAs [19].
Patients with other non-dengue febrile illnesses (OFI) were defined by febrile illness with no detectable DEN-specific IgM, no detectable DEN RNA, and no obvious or reported bacterial etiology for their illness during the same study period. Population controls were healthy, unrelated volunteers from the same community, with neither signs nor previous history of dengue infection, with a DEN IgG sero-positive rate of 1.37% (1/73).
Genomic DNA was isolated from heparin-anticoagulated blood samples using a standard phenol-chloroform extraction followed by 70% alcohol precipitation. Genotyping for the CD209 variant (−336 A/G; rs4804803) was carried out using Custom TaqMan SNP Genotyping Assays (Applied Biosystems, Foster City, CA, USA). The primer sequences were 5′-GGACAGTGCTTCCAGGAACT-3′ (forward) and 5′-TGTGTTACACCCCCTCCACTAG-3′ (reverse). The TaqMan minor groove binder probe sequences were 5′-TACCTGCCTACCCTT G-3′, and 5′-CTGCCCACC CTTG-3′. The probes were labeled with the TaqMan fluorescent dyes VIC and FAM, respectively. The PCR was conducted in total volume of 15 µL using the following amplification protocol: denaturation at 95°C for 10 min, followed by 40 cycles of denaturation at 94°C for 20 s, followed by annealing and extension at 60°C for one minute. After the PCR, the genotype of each sample was determined by measuring the allele-specific fluorescence in the ABI Prism 7500 Sequence Detection System, using SDS 1.1 software for allele discrimination (both Applied Biosystems). To validate the genotyping by real-time PCR analysis, 100 PCR products were subject to restriction fragment length polymorphism (RFLP) analysis with MscI restriction enzyme (New England Biolabs, Beverly, MA, USA) and showed 100% identical result between these two genotyping systems.
Peripheral blood mononuclear cells were collected from peripheral blood of 20 healthy, DEN-specific IgM or IgG seronegative volunteers with AA or AG genotype. CD14+ monocytes were isolated by positive selection according to the manufacturer's specifications using CD14 microbeads and a magnetic cell separator (MACS) (Miltenyi Biotec, Bergisch Gladbach, Germany). Enriched CD14+ cells (purity>95%) were cultured for 6 days in six-well plates in complete RPMI medium in the presence of 10 ng/mL rhGM-CSF and 5 ng/mL rhIL-4 at 37°C, and 5% CO2. On day 3, half of the medium was replaced with fresh medium supplemented with rhGM-CSF and rhIL-4. Expression of markers was measured by flow cytometer using specific antibodies and their corresponding isotype controls.
Unless otherwise stated, monocyte-derived dendritic cells (MDDCs) were infected with DEN-2 at a multiplicity of infection (MOI) of 5 for 2 h at 37°C and 5% CO2. Cells were washed twice to remove cell-free virus, and cultured in complete RPMI medium (without cytokines) at a density of 2×105 cells/ml in 48-well plates. Cells and supernatants were removed and analyzed at 24, 48, and 72 h post-infection. For the neutralization experiments, cells were incubated in the medium alone or in the medium with the addition of anti-human CXCL10/IP-10 antibody (R&D Systems, Minneapolis, MN, USA) at 10 µg/mL for 30 min. Cells and supernatants were harvested and analyzed 24 h post-infection.
Total RNA extracted from MDDCs was subjected to quantitative RT-PCR to assess levels of mRNA corresponding to CD209 and ß2–microglobulin (B2MG) using the ABI PRISM 7500 instrument (Applied Biosystems). The forward primer, reverse primer sequence for detecting CD209 and B2MG were 5′-AACAGCTGAGAGGCCTTGGA-3′, 5′-GGGACCATGGCCAAGACA-3′, and 5′-AATTGAAAAAGTGGAGCATTCAGA-3′, 5′-GGCTGTGACAAAGTCACATGGTT-3′, respectively. The PCR cycling parameters were 40 cycles of PCR reactions at 94°C for 20 s, and 60°C for one minute. The results were detected in real-time and recorded on a plot showing fluorescence vs. time. RT-PCR products were also visualized on ethidium bromide-stained 1.5% agarose (Pierce Co., Rockford, IL, USA) gel with a 100- bp ladder (Pharmacia Biotech, Piscataway, NJ, USA) as a reference.
To measure the CD209 cell surface expression, MDDCs were stained with FITC-conjugated mAbs specific for DC-SIGN (R&D Systems, Minneapolis, MN, USA). An isotype-matched FITC-labeled control, mouse IgG2b (clone MOPC195, Immunotech, Beckman Coulter, Fullerton, CA, USA) was included in each experiment.
Total RNA extracted from MDDCs was subjected to assess DEN-2 RNA viral copies. Fluorescent RT-PCR was performed in an ABI 7500 quantitative PCR machine (Applied Biosystems) for 40 cycles using TaqMan technology as previously described [21].
Cytokine/chemokine production and viral replication were determined at 24, 48, and 72 h post-infection. Cell-free culture supernatants TNFα and macrophage chemoattractant protein 1 (MCP-1) concentrations were measured using ELISA kits from eBioscience Inc. (San Diego, CA, USA); IL-12p40 and IFN-inducible protein 10 (IP-10) concentrations were measured using ELISA kits from R&D Systems as per manufacturer's instructions.
Data are presented as mean ± SEM values. Alleles and genotypes distribution of rs4804803 are presented as numbers (percentages). Conformance of the allele frequencies to Hardy-Weinberg equilibrium proportions was tested to compare the observed and expected frequencies of heterozygotes and homozygotes. Differences among patients with DEN, DF, DHF, OFI, and population controls were determined using two-sided Chi-Square test or Fisher exact test. Odds ratio (OR) values were calculated with 95% confidence intervals (CI). The Student's t-test or Mann-Whitney U test was used for statistical comparisons between continuous variables. The Wilcoxon signed-rank test was used for statistical comparison of the neutralization experiments. All analyses were performed using SPSS 13.0 (SPSS Inc. Chicago, IL, USA).
During a large DEN-2 outbreak in southern Taiwan between June 2002 and January 2003, a hospital-based case-control study was used to identify the risk immune parameters [19]–[23]. Employing the decoded DNA samples from that same cohort of the population that study has been extended to investigate the association of rs4804803 SNP with DF, DHF, viral replication, and immune response. Based on the previous case-control study design, we have included DNA samples from 135 DHF, 176 DF, and 143 OFI patients in this expanded study. The main characteristics of the study population are summarized in Table 1. There were no significant differences in sex or total leukocyte counts between patients with DF and those with DHF. However, age (41.7±1.6 years vs. 45.7±1.3 years, p<0.001), serum GOT levels (70.1±8.1 U/mL vs. 313.8±74.6 U/mL, p = 0.002) and GPT levels (67.1±11.3 U/mL vs. 142.7±21.9 U/mL, p = 0.003) were significantly higher in the DHF group (Table 1). A patient manifested with abdominal pain had ascites as evidenced by abdominal ultrasonography was classified as DF because the patient revealed no thrombocytopenia (<100,000/mm3), petechia or hemorrhagic manifestation during the admission period.
We investigated the association of rs4804803 SNP in the promoter region of CD209 with protection from dengue infection and the susceptibility of DHF. Genomic DNA obtained from DEN patients (n = 311), OFI patients (n = 143), and population controls (n = 120) was genotyped for rs4804803 SNP. We found that GG/AG genotypes in 16.0% of the DEN patients were significantly higher than OFI patients (5.6%, OR = 3.23, p = 0.002) and population controls (7.5%, OR = 2.36, p = 0.020; Table 2). Moreover, the GG/AG genotypes were significantly higher in DHF patients (23.0%) than OFI patients and population controls (OR = 5.03 and 3.68, p = 3×10−5 and 0.001), and also significantly higher than DF patients (10.8%, OR = 2.46; p = 0.004; Table 2).
Analysis of the allele distribution between DEN and OFI patients or population controls showed that the G allele frequency was higher in DEN patients (8.4%), compared with OFI patients (2.8%, OR = 3.17, p = 0.002) or population controls (3.8%, OR = 2.34, p = 0.018; Table 3). Moreover, the frequency of G allele of rs4804803 was significantly higher in DHF patients (12.2%) than OFI patients or population controls (OR = 4.84 and 3.57, p = 2×10−5 and 0.001), and higher than DF patients (5.4%, OR = 2.44, p = 0.002; Table 3). Few DHF patients (n = 6) had dengue shock syndrome in this cohort; one of them carrying AG genotype.
To investigate whether the rs4804803 SNP in the promoter region of CD209 associated with the primary and secondary DEN infection, we used serological methods to detect DEN antibodies for differentiation into primary and secondary dengue infection. Of the 293 DEN patients, 141 (48%) had secondary DEN infections, based on detectable DEN-2 virus RNA and DEN IgG. As shown in Table 1, secondary DEN-2 infection was more frequently found in patients with DHF than in those with DF (65% vs. 36%, p<0.001). We found the rs4804803 GG/AG genotypes were found in 12.5% of patients with primary DEN infection and 16.3% of patients with secondary DEN infection, which did not reach significantly different (OR = 1.36; p = 0.352). As shown in Table 4, there was no association between rs4804803 SNP and primary or secondary dengue infection in DF patients (OR = 0.64; p = 0.409), or in DHF patients (OR = 1.80; p = 0.251). In addition, there was no association between allele distribution and primary or secondary dengue infection in DF and DHF patients (data not shown).
Due to the low frequency of GG genotype in our population (2 cases, 0.6%), we could not recall the patients because the data file had been decoded for identification. We examined DC-SIGN (CD209) expression in both mRNA level of MDDCs and protein level on their cell surface from healthy subjects with AA or AG genotype by quantitative RT-PCR and flow cytometry, respectively. A significant increase in CD209 mRNA expression was detected in the MDDCs from individuals with AG genotype than those from individuals with AA genotype (p = 0.032, Fig 1B). Similarly, individuals with AG genotype had a significantly higher cell surface DC-SIGN expression (p = 0.029; Fig 1D). However, the surface DC-SIGN expression declined rapidly along with DEN-2 infection on MDDCs from both genotypes' subjects, which showed no difference at 24, 48, and 72 h post-infection (Fig. 2A).
To investigate whether the rs4804803 SNP was correlated to viral replication, we measured DEN-2 RNA copies in MDDCs with AA or AG genotype of rs4804803 at 24, 48, and 72 h post-infection. DEN-2 replication was significantly higher in MDDCs from individuals with AA genotype than those with AG genotype at 48 h post-infection (1.07±0.45×106 copies/105 cells vs. 3.90±0.67×106 copies/105 cells, p = 0.006) and 72 h post-infection (4.83±0.70×105 copies/105 cells vs. 2.32±0.68×106 copies/105 cells, p = 0.003; Fig. 2B). Viral replication, as measured at 72 h post-infection, increased more remarkably in MDDCs at MOI of 5 and 10 (p<0.001 and 0.002, respectively; Fig. 2C).
To investigate whether higher cell surface DC-SIGN expression was correlated with immune response, we investigated kinetic cytokine/chemokine production by MDDCs from individuals with AA or AG genotype of rs4804803. Results showed that MDDCs with AG genotype had significantly higher TNFα production than those with AA genotype at 24 and 48 hr post-infection (303.51±66.75 pg/mL vs. 143.97±68.80 pg/mL and 202.35±19.35 pg/mL vs. 73.00±9.55 pg/mL; p = 0.021 and 0.002, respectively; Fig 3A). IL-12p40 production significantly increased by MDDCs with AG genotype than those with AA genotype at 24, 48, and 72 h post-infection (p<0.001, 0.007 and 0.001, respectively; Fig 3B). We also measured the concentration of two chemokines, MCP-1 and IP-10, which had been implicated in the recruitment and stimulation of monocytes, macrophages, dendritic cells, NK cells, and T lymphocytes [25]. It was found that IP-10, but not MCP-1, production was significantly higher by MDDCs with AG genotype than those with AA genotype at 24, 48, and 72 h post-infection (620.60±175.56 pg/mL vs. 243.02±41.64 pg/mL, 889.92±91.46 pg/ml vs. 168.02±24.02 pg/mL, and 614.44±49.16 pg/mL vs. 322.32±69.62 pg/mL; p = 0.034, 0.009 and 0.010, respectively; Fig 3C). The MCP-1 levels peaked at 48 hr in subjects with both genotypes', but there was no significant difference between AA genotype and AG genotype (550.72±60.73 pg/mL vs. 463.92±66.80 pg/mL, p = 0.157; Fig. 3D).
IP-10, produced by non-infected bystander DCs in response to DEN infection, is a potent chemoattractant for activated T and NK cells [26], and the modulation of adaptive immune response [27]. IP-10 has also been known to inhibit the binding ability of DEN in immortalized cells [28]. In our MDDC model, cells from individuals with AG genotype exhibited an augmented innate immune reaction, showing higher IP-10 production, post-infection (Figure 3C). Based on these results, we hypothesized that DEN-infected MDDCs with AG genotype produced higher levels of IP-10, which might block viral entry or viral replication in MDDCs. We used an anti-IP-10 neutralizing mAb to block endogenous IP-10 production by MDDCs. With both genotypes, the viral replication 24 h post DEN infection increased significantly more in the presence of neutralizing antibody than in the absence of neutralizing antibody (p = 0.034 and 0.040, respectively; Fig. 4A). IP-10 production by MDDCs from individuals with AG genotype significantly decreased (795.3±368.1 pg/mL vs. 273.8±87.8 pg/mL; p = 0.037), but it did not decrease in MDDCs from individuals with AA genotype (329.8±114.2 pg/mL vs. 201.8±87.0 pg/mL; p = 0.091; Fig. 4B). These results suggest that IP-10 produced by MDDCs is involved in the viral replication of DEN infection.
DC-SIGN has been shown to be an important receptor for DEN and a number of viruses, including HIV, Helicobacter pylori, and Mycobacterium tuberculosis and hepatitis C virus (HCV) [29]. Some studies have demonstrated that genetic variations of CD209 (rs4804803) were associated with the susceptibility to HIV [14], Mycobacterium tuberculosis [17], HCV [30], and dengue [16]. Few studies have demonstrated how the rs4804803 SNP is involved in viral replication or immune response. We are the first in the field to demonstrate the relationship among functional cell surface expression, viral replication, and immune responses in DEN-infected MDDCs from subjects with rs4804803 SNP. Here we found that rs4804803 SNP was strongly associated with the risk of DHF vs. DF and controls. Functional studies have determined that MDDCs from individuals with AG genotype have a significantly higher cell surface DC-SIGN expression than from those with AA genotype. MDDCs with AG genotype produced higher TNFα, IL-12p40, and IP-10 levels but lower viral replication in response to dengue infection.
Because the physiopathology of various manifestations of DHF is not fully understood, several studies have supported the supposition that secondary dengue infection [31], age [32], a number of preexisting chronic diseases such as diabetes and bronchial asthma [33], and host genetic factor [16], [22] increase the risk of progression to DHF. This indicates that multiple factors are involved in the development of DHF. Our findings regarding rs4804806 SNP associated with DHF vs. OFI control (p = 3×10−5; Table 2) in a case-control association study suggests that rs4804806 SNP contributes in part to the development of DHF.
Our study shows that the GG/AG genotypes of rs4804803 were associated with susceptibility to DHF, compared with DF, which is consistent with the observation of Sakuntabhai et al. [16]. In our study, the AA genotype was associated with protection against DHF, compared with OFI and population controls, while G allele was associated with protection against DF in Sakuntabhai's observation. The inconsistency between these studies regarding the protection for DHF or DF may result from two possibilities. First, the frequency of G allele in Chinese population is 3.8%; while in Thailand, it is 9.5–10.4% [16], [34]. Second, definition of DF and DHF might be also different. We defined DF and DHF according to WHO criteria, while in the study by Sakuntabhai et al., DF was defined by criteria of severe dengue fever syndrome with hemorrhage but no plasma leakage, excluding patients with flu-like symptoms or those having only fever. Moreover, the rs4804803 SNP was demonstrated to be in linkage disequilibrium with three other intronic polymorphisms in a Thai population, and these might also have contributed to the susceptibility of DHF [16].
Our results suggest that humans carrying the rs4804803 AG genotype have a higher DC-SIGN expression and lower DEN-2 replication in MDDCs. These results differ from a previous study by Loach et al. who demonstrated that the DC-SIGN expression levels on Raji cells after transfection of various DC-SIGN cDNA constructs were significantly correlated to the infection rate of DEN-1 [35]. DC-SIGN is an endocytic receptor shown to induce endocytosis of several pathogens, including dengue [36]–[38]. The difference between these two studies might be due to different cell types and ex vivo culture systems. In our study, it was found that MDDCs from subjects with rs4804803 AG genotype had higher surface DC-SIGN expression with higher production of chemokines such as IP-10, which could limit DEN-2 replication (Fig. 4A); however, the higher surface DC-SIGN expression in subjects with AG genotype decreased remarkably 24 h post-infection (Fig. 2A). In the study by Loach et al., ectopic expression levels of DC-SIGN on Raji cells enhanced DEN-1 replication, which might be related to a higher quantity of receptors or lower production of IP-10 favoring DEN replication. Results from these studies suggest that the correlation of viral replication to higher or lower DC-SIGN expression depends on genetic factors in the host, cell type, and dynamic changes in the receptor following DEN infection.
In functional studies of rs4804803 SNP, we determined that MDDCs with AG genotype had a higher DC-SIGN expression correlated to augmented immune responses with higher TNFα, IL-12p40, and IP-10, than those with AA genotype, but not MCP-1 production. DEN replication was significantly lower in individuals with AG genotype. The addition of anti-IP-10 neutralizing antibody blocked the production of endogenous IP-10 and significantly enhanced the replication of DEN-2 (Fig. 4A). This suggests that rs4804803 SNP was involved in the DC-SIGN expression associated with augmented immune response, such as the increase in the production of IP-10 that repressed the replication of DEN. This is supported by the fact that altered immune response, but not viral load, was observed in DHF patients [21], [39]. CLEC5A-mediated DEN infection in animals that was susceptible to DEN hemorrhagic infection also revealed augmented immune response [40].
In contrast, it is interesting to note that the viral replication in MDDCs from individuals with rs4804803 AA genotype was significantly higher than in individuals with AG genotype following DEN-2 infection. The mechanism by which rs4804803 SNP influences DEN replication in MDDCs is currently unknown. Chan et al. showed that certain polymorphisms of L-SIGN, a DC-SIGN homologue, mediated more efficient viral degradation of SARS-CoV [41]. The clinical implications of screening genotypes to prevent DEN infection might be supported if different viral loads could be demonstrated among humans with various genotypes of rs4804803 in future outbreaks of DEN.
The outcome of DEN infection is determined by a myriad of interactions among viral, immunological, and human genetic factors, as well as kinetic interactions between innate and adaptive immunity. This study provides new evidence that CD209 rs4804803 SNP, correlated to cell surface expression on dendritic cells, mediates augmented immune responses against DEN-2 infection and is implicated in the susceptibility of DHF. Further studies are warranted, particularly with regard to the genetic variants of CD209 on the DC polarization of adaptive immunity, and how they may promote or protect the development of DHF.
|
10.1371/journal.pgen.1006695 | Transaldolase inhibition impairs mitochondrial respiration and induces a starvation-like longevity response in Caenorhabditis elegans | Mitochondrial dysfunction can increase oxidative stress and extend lifespan in Caenorhabditis elegans. Homeostatic mechanisms exist to cope with disruptions to mitochondrial function that promote cellular health and organismal longevity. Previously, we determined that decreased expression of the cytosolic pentose phosphate pathway (PPP) enzyme transaldolase activates the mitochondrial unfolded protein response (UPRmt) and extends lifespan. Here we report that transaldolase (tald-1) deficiency impairs mitochondrial function in vivo, as evidenced by altered mitochondrial morphology, decreased respiration, and increased cellular H2O2 levels. Lifespan extension from knockdown of tald-1 is associated with an oxidative stress response involving p38 and c-Jun N-terminal kinase (JNK) MAPKs and a starvation-like response regulated by the transcription factor EB (TFEB) homolog HLH-30. The latter response promotes autophagy and increases expression of the flavin-containing monooxygenase 2 (fmo-2). We conclude that cytosolic redox established through the PPP is a key regulator of mitochondrial function and defines a new mechanism for mitochondrial regulation of longevity.
| There are a growing number of studies linking mitochondrial dysfunction to enhanced longevity, especially in the nematode C. elegans. The reasons for these pro-longevity effects have been elusive, but one current model is that adaptive responses to mitochondrial inhibition promote organismal health and stress resistance. Here, we report an intriguing example of mitochondrial stress induced by inhibition of a cytosolic metabolic pathway that extends lifespan in worms. We find that inhibition of the pentose phosphate pathway, which is essential for cytosolic redox homeostasis, affects multiple parameters of mitochondrial function and activates a starvation-like response that promotes longevity through recycling of damaged cellular components and induction of the enzyme flavin-containing monooxygenase 2. These results establish novel links between the pentose phosphate pathway, mitochondrial function, redox homeostasis, and organismal aging.
| Mitochondria are the primary sites of aerobic metabolism and energy production in the cell. The mitochondrial free radical theory of aging posits that reactive oxygen species (ROS) produced by mitochondria during oxidative metabolism cause damage to macromolecules which, over time, leads to the accumulation of cellular, tissue, and organismal declines, and ultimately death [1, 2]. In general, mitochondrial dysfunction is detrimental, and has been causally implicated in several age-related diseases, as well as severe, early-onset mitochondrial disorders. Paradoxically, however, inhibition of mitochondrial function has, in some cases, been associated with increased longevity in laboratory organisms from yeast to mammals [3]. This is particularly evident in C. elegans, where inhibition of mitochondrial respiration by mutation or knockdown of numerous electron transport chain (ETC) components usually increases lifespan [4, 5]. Mild oxidative stress can also increase lifespan [3], perhaps by inducing adaptive responses that compensate for these insults and provide cytoprotective effects to improve cellular stress resistance [6].
The mechanistic basis for lifespan extension in response to mitochondrial inhibition and mild oxidative stress in C. elegans is an active area of investigation. One mitochondrial stress pathway that has been associated with worm longevity in this context is the mitochondrial unfolded protein response (UPRmt) [7–9]. The UPRmt is a coordinated response to mitochondrial stress resulting in upregulation of mitochondrial chaperones, import machinery, and proteases, while negatively regulating expression of nuclear- and mitochondrial-encoded ETC components [10–12]. Activation of the UPRmt is regulated by the ATFS-1 transcription factor, which translocates to the nucleus in response to mitochondrial stress and directly activates transcription of several UPRmt target genes [11, 13]. Whether the UPRmt plays a direct role in determining longevity remains unclear. Lifespan extension by ETC inhibition or treatment with the ROS-generating compound paraquat is correlated with induction of the UPRmt [7, 10, 14]; however, deletion or RNAi knockdown of atfs-1 blocks induction of several UPRmt target genes but does not prevent or attenuate lifespan extension following inhibition of the ETC [15, 16]. Similarly, constitutive active alleles of atfs-1 cause activation of the UPRmt but do not extend lifespan [15–17].
There is experimental evidence supporting a role for several factors other than the UPRmt in lifespan extension downstream of mitochondrial inhibition in C. elegans, including the hypoxic response transcription factor HIF-1, CEP-1/p53, the CEH-23 transcription factor, components of the intrinsic apoptotic pathway, and the p38 MAPK PMK-3 [18–22]. A majority of these studies have been performed using mutants with defective ETC function, such as the Rieske iron-sulfur protein gene isp-1(qm150) allele and the ubiquinone biosynthetic gene clk-1(qm30) allele. With the possible exception of pmk-3, none of these factors is able to account for the full lifespan extension following RNAi knockdown of ETC genes such as the cytochrome c oxidase gene cco-1. This is consistent with a model proposed by the Hekimi lab that RNAi inhibition of ETC function promotes worm longevity by a mechanism distinct from mutations that impair ETC function [23].
Uncovering the genetic pathways and molecular mechanisms by which mitochondria influence aging and disease is critical both for developing better models of biological aging, as well as for identifying interventions to promote health and longevity. As mentioned above, low levels of oxidative stress can be beneficial to cellular health, but high levels can cause irreparable damage. This biphasic or non-linear relationship between mitochondrial ROS and survival is commonly referred to as mitohormesis, and posits that ROS act as signaling molecules to induce adaptive mechanisms [6]. This has been observed in C. elegans, where different levels of RNAi knockdown of a single mitochondrial gene can cause differential effects on lifespan and other physiological markers [24, 25]. The beneficial hormetic effects associated with elevated ROS are due to the contribution of multiple protective responses that are still being discovered. Therefore, we sought to identify and determine the interconnectivity of novel longevity pathways distinct from the UPRmt that are engaged by oxidative and mitochondrial stress.
Although the UPRmt does not appear to directly mediate lifespan extension, we reasoned that the partial correlation between activation of the UPRmt and longevity could be used to identify novel factors and mechanisms of action within the mitochondrial longevity network. To identify such factors, we performed a genome-wide RNAi screen for C. elegans genes that negatively regulate the UPRmt by looking for RNAi clones that activated the UPRmt reporter hsp-6p::gfp [15]. Some, but not all, of these genes also negatively affected lifespan such that RNAi knockdown increased longevity. One such gene is tald-1, which encodes the pentose phosphate pathway (PPP) enzyme transaldolase. The PPP pathway is a cytosolic metabolic pathway that functions to produce NADPH, ribose-5-phosphate, and interconvert 3–7 carbon sugars. The observation that tald-1(RNAi) induced the UPRmt reporter and increased lifespan intrigued us, as transaldolase is not a mitochondrial protein and has not been previously implicated in longevity control in any organism. Here we report that transaldolase deficiency indeed alters mitochondrial function, as evidenced by changes in mitochondrial morphology and direct measurement of mitochondrial respiration. The lifespan extension from tald-1(RNAi) is independent of the UPRmt, and instead involves activation of an oxidative stress response mediated by the p38 MAPK PMK-1 and JNK MAPKs JNK-1 and KGB-1, and a concomitant starvation-like response that signals through the transcription factor EB (TFEB) homolog HLH-30. Furthermore, we find that activation of the starvation-like response transcriptionally activates HLH-30-dependent autophagy markers, increases autophagic flux, and increases expression of the longevity-promoting flavin-containing monooxygenase 2 (fmo-2).
From an unbiased genome-wide RNAi screen for negative regulators of the mitochondrial unfolded protein response (UPRmt), we found that knockdown of either of the pentose phosphate pathway (PPP) enzymes transaldolase (tald-1) or transketolase (tkt-1) activates the UPRmt reporter hsp-6p::gfp in C. elegans [15]. These enzymes function in the non-oxidative branch of the PPP, generating ribose-5-P for nucleotide synthesis and interconverting three, four, five, six, and seven carbon sugars (Fig 1A). To determine if tald-1 and tkt-1 deficiencies specifically cause mitochondrial stress independent of the PPP, we tested if knockdown of other PPP enzymes not detected in the initial RNAi screen could also induce the hsp-6p::gfp reporter. RNAi knockdown of T25B9.9, which encodes the oxidative PPP enzyme 6-phosphogluconate dehydrogenase (6PGD), caused a significant increase in hsp-6p::gfp expression (+89%), albeit less robustly than tald-1(RNAi) (+187%) (Fig 1B and 1C). In addition, RNAi knockdown of Y57G11C.3 (gluconolactone hydrolase/GLH) or rpia-1 (ribose-5-phosphate isomerase/RPIA) slightly increased hsp-6p::gfp expression (+34%, +19%), while gspd-1 (glucose-6-phosphate dehydrogenase/G6PD) RNAi did not (S1A Fig). Inhibition of the PPP at multiple enzymatic steps, both oxidative and non-oxidative, is therefore sufficient to increase expression of a mitochondrial stress reporter. Next, we asked if inhibition of the enzymatic steps that robustly activate hsp-6p::gfp increase lifespan similar to other RNAi clones that induce this reporter. We found that knockdown of tald-1, tkt-1, and T25B9.9/6PGD all increased lifespan (Fig 1D). Since tald-1(RNAi) resulted in the strongest phenotypes among PPP enzymes tested, we chose to focus our studies on understanding the mechanisms by which tald-1 knockdown induces mitochondrial stress and enhances longevity.
The ATFS-1 transcription factor and the GCN-2 kinase, respectively, mediate the transcriptional and translational changes in response to mitochondrial stress that comprise the UPRmt [11, 26]. Loss of either ATFS-1 or GCN-2 does not prevent the lifespan extension from mitochondrial inhibition [15, 16]. These factors act in a compensatory fashion, however, and GCN-2 may be able to establish mitochondrial protein homeostasis in the absence of ATFS-1 or vice versa. Therefore, to convincingly assess whether the UPRmt regulates longevity from ETC or PPP inhibition, we examined if simultaneous loss of both atfs-1 and gcn-2 could prevent lifespan extension from RNAi knockdown of either tald-1 or the complex IV subunit cytochrome c oxidase 1 gene, cco-1. Both RNAi clones significantly increased the lifespan of atfs-1(tm4525); gcn-2(ok871) animals comparable to their effects in wild-type nematodes (Fig 1E and 1F). Similar results were observed in atfs-1(gk3094) mutant animals (S1B and S1C Fig). Thus, we conclude that neither ATFS-1 nor GCN-2 are required for lifespan extension, further supporting the model that mitochondrial stress or ETC inhibition affect lifespan independently of the UPRmt.
We next examined the temporal and genetic requirements for tald-1(RNAi) lifespan extension in the context of previously described C. elegans longevity pathways. Like RNAi knockdown of ETC genes [4, 24], tald-1(RNAi) only extended lifespan when knockdown occurred during development (feeding beginning at L1), and adult-specific knockdown (feeding beginning at ~L4/young adult) had no effect on longevity (Fig 1G). Knockdown of tald-1 also extended lifespan in animals carrying mutations of the FOXO-transcription factor daf-16, the AMP-activated protein kinase aak-2, and the germline-signaling factor glp-1 (S2A–S2C Fig), consistent with the reported effects of mitochondrial RNAi treatments [4, 5, 8, 27, 28]. Interestingly, tald-1(RNAi) resulted in a larger lifespan extension in animals lacking the hypoxic response transcription factor HIF-1 (S2D Fig), while loss of hif-1 attenuated the lifespan extension from cco-1(RNAi) (S2E Fig), as has been previously reported [20]. These data are consistent with a model that inhibition of the PPP extends lifespan by a mechanism that is overlapping but partially distinct from ETC inhibition.
Based on our findings that developmental knockdown of tald-1 induced the UPRmt, we asked whether other parameters of mitochondrial function are affected by tald-1(RNAi). First, we decided to use confocal microscopy to characterize any changes to intestinal mitochondrial morphology and content, since this tissue is particularly responsive to mitochondrial stress, as measured by the hsp-6p::gfp reporter. Using a mitochondrial-targeted GFP reporter whose expression is restricted to the intestine via the ges-1 promoter, we observed that tald-1(RNAi) caused a disruption to normal mitochondrial morphology in intestinal cells (Fig 2A and 2B). Mitochondria in these animals became thin and smaller in size, reflecting a potential change in mitochondrial dynamics. A similar change in morphology occurred following cco-1(RNAi) (Fig 2B). Interestingly, despite the smaller size of mitochondria following tald-1(RNAi) and cco-1(RNAi), there was increased GFP area per cell compared to controls (Fig 2C). This could indicate increased mitochondrial content; however, we did not observe any change in whole worm mitochondrial DNA abundance in these animals (S3A and S3B Fig), agreeing with previous studies reporting no change in mtDNA copy number from mitochondrial RNAi treatments [12, 29].
To better understand the effect of tald-1(RNAi) on mitochondrial morphology, we examined its interaction with factors known to regulate mitochondrial fusion and fission. As expected, knockdown of the fission factor dynamin-related GTPase drp-1 (DRP1/DNM1 homolog) caused intestinal mitochondria to swell and aggregate, while knockdown of the inner membrane fusion GTPase eat-3 (OPA1/MGM1 homolog) caused mitochondria to fragment and lack normal tubular structure (S4A Fig). Outer membrane fusion GTPase fzo-1 (MFN1/FZO1 homolog) knockdown also caused mitochondria to fragment, but morphology was remarkably similar to tald-1(RNAi) mitochondria, suggesting a mild pro-fission phenotype (S4A Fig). Accordingly, drp-1(RNAi) prevented the shift in mitochondrial morphology following tald-1 knockdown (Fig 2D), indicating that the core fission machinery is required for this response. In contrast, fzo-1 and the mitophagy components pdr-1 (PARK2 homolog) and pink-1 (PINK1 homolog) were not required for this phenotype (S4B and S4C Fig).
Since mitochondrial stress and mitochondrial fragmentation are associated with decreased mitochondrial function, we sought to directly measure metabolic activity in whole animals. The Seahorse XF24 Analyzer allows measurements of basal and real time changes in O2 consumption in C. elegans [30]. We found that knockdown of tald-1 caused an approximately 41% reduction in oxygen consumption, while, as expected [8, 31], knockdown of cco-1 caused a 67% reduction (Fig 2E). The reduction in oxygen consumption could not be fully explained by changes to worm length or density (S5A and S5B Fig), arguing that tald-1(RNAi) decreases basal mitochondrial respiration in whole animals.
To determine whether tald-1(RNAi) causes decreased mitochondrial respiration by altering ETC function or stability, mitochondria were isolated from animals and oxygen consumption of intact mitochondria was measured using malate, succinate, and TMPD/ascorbate as electron donors to drive complex I-, complex II-, and complex IV-dependent respirations, respectively. The mitochondria isolated in all trials retained normal coupling (P/O ratios) and a normal respiratory control index (State 3:State 4), indicating purification of healthy mitochondria (Fig 2F and 2G). As expected with Complex IV RNAi [32], cco-1(RNAi) decreased Complex I- and Complex IV-dependent respiration (Fig 2H). In contrast to cco-1(RNAi), tald-1(RNAi) did not cause a change in any rates measured (Fig 2H). Therefore, tald-1(RNAi) decreases whole animal respiration without altering maximal ETC capacity, potentially by reducing equivalents to the ETC in vivo.
Mitochondrial dysfunction has been proposed to extend lifespan in C. elegans through increased production of ROS and altered redox signaling [20, 33–35]. To specifically observe in vivo changes in redox environment, we utilized a transgenic strain expressing the ratiometric H2O2-specific biosensor HyPer, which is comprised of the regulatory domain of the bacterial transcription factor OxyR (OxyR-RD) fused to circularly permuted yellow fluorescent protein [36]. The OxyR-RD of HyPer is selectively oxidized by H2O2, generating a disulphide bridge that consequently alters the fluorescent properties of cpYFP. RNAi knockdown of either tald-1 or cco-1 significantly increased the oxidation of HyPer as measured via a plate reader assay, indicating elevated cytoplasmic H2O2 levels in these animals (Fig 3A). By confocal microscopy we observed similar results for tald-1(RNAi), but for cco-1(RNAi), oxidation of the reporter did not reach statistical significance (p>0.1) (S6A and S6B Fig). Since the PPP generates cytosolic NADPH, we hypothesize that oxidative stress in tald-1(RNAi) animals results from NADPH depletion and reduced ROS buffering capacity. Using LC-MS to measure NAD metabolites (Table 1), we found that tald-1(RNAi) decreased cellular NADPH levels, whereas cco-1(RNAi) did not (Fig 3B). In accordance with higher endogenous levels of oxidative stress, tald-1(RNAi) animals were sensitive to 10mM paraquat treatment (a high dose on the hormetic curve of paraquat treatment that decreases wild-type lifespan [37, 38]), which leads to the production of mitochondrial superoxide (Fig 3C). Thus, the presence of a functional PPP is required for normal resistance to exogenous oxidative stress.
Transaldolase deficiency in mammals causes a shift towards a more oxidative cellular redox status and compensatory activation of JNK MAPK signaling [39–41], prompting us to explore whether a similar response occurs in nematodes to regulate stress resistance and longevity. Remarkably, deletion of either jnk-1 or kgb-1, which encode C. elegans JNK MAPKs, fully prevented the lifespan extension from tald-1(RNAi) and significantly attenuated the lifespan extension from cco-1(RNAi) (Fig 4A–4D). This effect was specific to mitochondrial longevity, since daf-2(RNAi) robustly extended the lifespan of jnk-1(gk7) and kgb-1(um3) animals (S7A and S7B Fig). Although deletion of either jnk-1 or kgb-1 prevented lifespan extension in response to either tald-1(RNAi) or cco-1(RNAi), these mutations did not prevent the effects on mitochondrial respiration or UPRmt induction (S7C–S7E Fig).
The p38 MAPK PMK-1 has been implicated in mitohormesis-induced lifespan extension in response to reduced insulin/IGF-1-like signaling, metformin treatment, or glycolysis inhibition [33, 35, 42]. Interestingly, PMK-1 was also required for lifespan extension from tald-1(RNAi), but not from cco-1(RNAi) (Fig 4E and 4F). Therefore, despite some similar mitochondrial phenotypes and interactions with MAPK signaling, PPP inhibition and mitochondrial ETC RNAi longevity require both overlapping and distinct pathways. In addition, PMK-1 does not prevent UPRmt induction from tald-1(RNAi) or cco-1(RNAi) (S8A and S8B Fig), suggesting it is not upstream of mitochondrial stress. As previously reported [42, 43], we also found that PMK-1 regulates daf-2(RNAi) lifespan extension (S8C Fig) and is not specific to PPP inhibition.
The MAP3K ASK1 is a well-established factor upstream of p38 and JNK MAPKs that responds to oxidative stress via interactions with redox proteins [44–46]. In C. elegans, the ASK1 homolog NSY-1 was found to act upstream of PMK-1, JNK-1, and KGB-1 in various contexts [47–52]. Accordingly, we found that loss of NSY-1 attenuated the lifespan extension from tald-1(RNAi) or cco-1(RNAi), suggesting this factor responds to oxidative stress in both of these instances to promote longevity (Fig 4G and 4H). In agreement with NSY-1 regulating PMK-1 activity, we found that NSY-1 attenuated the lifespan extension from daf-2(RNAi) (S8D Fig). Therefore, NSY-1 is a MAP3K necessary for the activation of multiple longevity mechanisms, highlighting the importance of redox sensing in C. elegans longevity.
Since oxidative stress induces both p38 and JNK MAPK activity in mammalian cell lines, we predicted a similar response may occur in C. elegans. In order to test this, we treated animals with H2O2 and measured phosphorylation of JNK-1, KGB-1, and PMK-1 MAPKs. We found that as little as 5–15 minutes of H2O2 treatment is sufficient to activate these MAPKs (S8E Fig), demonstrating their high sensitivity to redox stress and further supporting for their role in longevity interventions associated with oxidative stress.
In addition to reducing in vivo respiration rates, we noted that tald-1(RNAi) and cco-1(RNAi) also caused dramatic reductions in intestinal fat levels, as assessed by Oil Red O (ORO) staining (Fig 5A and 5B). Such a response could reflect decreased lipid synthesis, increased fatty acid oxidation (associated with starvation), or decreased fatty acid absorption. Because C. elegans acquire the majority of lipid species from their bacterial diet and not from de novo fatty acid synthesis, with the exception of monomethyl branched-chain fatty acids [53], we focused on determining whether there were changes in expression of metabolic genes regulated by starvation including lipases, β-oxidation, monounsaturated fatty acid synthesis, and glyoxylate pathway genes [54].
First, we examined if decreased ORO staining might reflect degradation of cytoplasmic lipid droplets. The adipose triglyceride lipase ATGL-1 is an important lipase that is stabilized and localized to lipid droplets during fasting to mediate lipolysis [55]. Using the atgl-1p::atgl-1::gfp translational reporter, we found that tald-1(RNAi) dramatically increased ATGL-1::GFP levels, suggesting enhanced breakdown of lipid droplets in these animals (Fig 5C and 5D). The stearoyl-CoA desaturase fat-7 controls the relative abundance of saturated and mono-unsaturated fatty acids by converting stearic acid (18:0) to oleic acid (18:1). Expression of fat-7 is positively regulated by NHR-49 in fed conditions but is repressed during starvation, independent of NHR-49, to preserve saturated fatty acid levels [54, 56]. Using the fat-7p::gfp reporter we found that fat-7 expression was dramatically repressed in tald-1(RNAi) or cco-1(RNAi) animals (Fig 5E and 5F). This observation was also confirmed by qRT-PCR (Fig 5G). In a similar fashion, other metabolic genes known to be regulated by starvation [47, 54], such as genes involved in β-oxidation and the glyoxylate pathway, also change in tald-1(RNAi) animals and cco-1(RNAi) animals (Fig 5G). For example, we observed increased expression of carnitine palmitoyltransferase 4 (cpt-4) following tald-1(RNAi) or cco-1(RNAi), suggesting increased import of long-chain fatty acids into the mitochondria (Fig 5G). In addition, we observed increased expression of the bifunctional glyoxylate gene icl-1 with tald-1(RNAi) or cco-1(RNAi), indicating increased metabolism of fatty acids to promote gluconeogenesis and generation of succinate without concomitant NAD+ consumption and carbon loss (Fig 5G). In some cases, directionality or robustness of gene expression differed between tald-1(RNAi) and cco-1(RNAi) animals. For example, acs-2 expression is decreased by tald-1(RNAi) and increased by cco-1(RNAi), while acdh-1, acdh-2, and hacd-1 are downregulated by tald-1(RNAi), but not cco-1(RNAi) (Fig 5G). Multiple genes exist for each enzyme involved in β-oxidation in C. elegans and depending on the type of starvation response differential regulation of isoforms and even downregulation of certain β-oxidation genes occurs possibly owing to tissue-specific alterations or isoform preference for certain fatty acid chain lengths [47, 54, 57–59]. Thus, it is not surprising that metabolic gene expression profiles in tald-1(RNAi) and cco-1(RNAi) animals differ in some regards.
To explore if the starvation-like metabolic response underlies the pro-longevity effects of tald-1(RNAi) or cco-1(RNAi), we performed epistasis analyses with dietary restricted animals and starvation responsive transcription factors NHR-49 and HLH-30. We found that tald-1(RNAi) lifespan extension is slightly additive (mean +6% extension, median +0%) to complete removal of the bacterial food source in adulthood (bacterial deprivation, BD), suggesting that tald-1(RNAi) functions through a starvation response (Fig 5H). Supporting the notion that mitochondrial RNAi functions independently of dietary restriction to extend lifespan (mitochondrial RNAi acts during development [4, 24], whereas BD acts during adulthood [60, 61]), we found that cco-1(RNAi) was fully additive to BD lifespan extension (Fig 5I). This is intriguing since we observed that tald-1(RNAi) only extends lifespan when RNAi is initiated from development similar to mitochondrial RNAi. One possibility is that TALD-1 protein levels must reach a lower threshold to ensure hormetic benefits and a starvation response during adulthood, which is more likely if RNAi treatment begins from hatching. NHR-49 is a master regulator of gene expression changes that enable the mobilization of fat for energy metabolism, and HLH-30 regulates autophagy, fat storage, and has been previously implicated in lifespan extension downstream of dietary restriction and insulin/IGF-1-like signaling [62–64]. Interestingly, NHR-49 is not required for the lifespan effect of either tald-1(RNAi) or cco-1(RNAi) (Fig 5J and 5K). This agrees with a previous study that found reduced complex I, III, and IV activity caused NHR-49 dependent gene expression changes and increased lifespan independent of NHR-49 [65]. In contrast, tald-1(RNAi) caused nuclear localization of HLH-30 similar to starvation (Fig 6A and 6B), and also required HLH-30 for lifespan extension (Fig 6C). Importantly, tald-1(RNAi) did not affect food consumption, as measured by pharyngeal pumping rate (S9A Fig). In contrast to tald-1(RNAi), cco-1(RNAi) did not induce HLH-30 nuclear localization and the lifespan extension in this case was independent of HLH-30 (Fig 6A, 6B and 6D). Thus, transaldolase deficiency induces a starvation-like response and requires the autophagy regulator TFEB/HLH-30 for lifespan extension.
One major function of TFEB/HLH-30 is to promote autophagy [62, 63, 66], and this activity of HLH-30 is necessary for lifespan extension in response to dietary restriction and reduced insulin/IGF-1-like signaling [62]. Consistent with our observation that tald-1(RNAi) induces nuclear localization of HLH-30, we found that components of the autophagy pathway [62] were upregulated in a HLH-30-dependent fashion, including lgg-1 (LC3 homolog), sqst-1 (p62/SQSTM1 homolog), lmp-1 (LAMP1 homolog), and lysosomal subunit vha-17 (Fig 6E and 6F). In addition, autophagic flux is increased by tald-1(RNAi) (Fig 6G and 6H), as measured by a recently described LGG-1 reporter of lysosomal protease activity [67]. The reporter consists of LGG-1 tagged with two fluorescent proteins containing a flexible protease-sensitive linker. When the lysosome fuses with the autophagosome, lysosomal proteases cleave dFP::LGG-1 and release protease-resistant monomeric FP (mFP). Increases in autophagic flux are thereby reflected as an increase in the [mFP]/[dFP::LGG-1] ratio.
Another important target of HLH-30 recently implicated in longevity control is the flavin-containing monooxygenase FMO-2. FMO-2 is induced by both hypoxic signaling and starvation, and its induction by starvation is dependent on HLH-30 [68]. Utilizing an fmo-2p::mCherry transcriptional reporter, we found that tald-1(RNAi) also robustly induced fmo-2 expression, although not to the same extent as complete removal of the bacterial food source (Fig 7A–7C). Unexpectedly, whereas, tald-1(RNAi) or starvation causes intestinal fmo-2 expression, cco-1(RNAi) causes fmo-2 expression in the pharynx and cells proximal to the anterior bulb (S10A Fig). The increased expression of fmo-2 indicated by the reporter was confirmed by qRT-PCR (Fig 7D).
Since the regulation of fmo-2 is not well understood, we decided to test if HLH-30 is an essential regulatory factor of fmo-2 in multiple contexts. In support of this, we found that HLH-30 mediates fmo-2 expression from both BD and tald-1(RNAi) (Fig 7A and 7B). This observation was supported by qRT-PCR (Fig 7E). Thus, we decided to use the fmo-2 transcriptional reporter as a proxy for HLH-30 activity to determine genetic relationships between HLH-30 and the MAPKs that mediate tald-1(RNAi) lifespan extension. JNK-1 and KGB-1 were not required for fmo-2p::mCherry induction from tald-1(RNAi) or BD (S10B Fig), arguing that these MAPKs are not upstream of HLH-30. However, the p38 MAPK PMK-1 was required for induction of fmo-2p::mCherry from BD and there was a similar trend for tald-1(RNAi) (Fig 7A and 7C). Supporting this, fmo-2 induction by tald-1(RNAi) was attenuated in pmk-1(km25) animals by qRT-PCR (Fig 7E). Since loss of function in hlh-30 and pmk-1 cause similar effects with respect to fmo-2 expression and lifespan epistasis with tald-1(RNAi) and cco-1(RNAi), we tested if PMK-1 is upstream of HLH-30. Surprisingly, we found that pmk-1(km25) mutation did not alter HLH-30 nuclear localization from tald-1(RNAi) or BD (Fig 7F). Therefore, the simplest model is that PMK-1 functions in parallel with HLH-30 to activate fmo-2 expression.
To determine if FMO-2 activation contributes to the lifespan extension from transaldolase deficiency, we treated fmo-2(ok2147) mutants with tald-1(RNAi). We found that tald-1(RNAi) did not extend the lifespan of fmo-2(ok2147) animals (Fig 7G). In addition, cco-1(RNAi) longevity partially required fmo-2 (Fig 7H). In neither case did deletion of fmo-2 affect induction of the UPRmt reporter (S10C Fig). Consistent with the model that FMO-2 acts downstream of tald-1(RNAi) to promote longevity, tald-1(RNAi) did not further extend the lifespan of long-lived eft-3p::fmo-2 animals ubiquitously overexpressing fmo-2 (Fig 7I).
In this study, we found that the inhibition of the PPP enzyme transaldolase impairs mitochondrial respiration, induces a starvation-like metabolic response, and activates MAPK signaling pathways that together promote longevity in C. elegans. These observations define unexpected new connections between the cytosolic PPP, mitochondrial metabolism, and aging. Although our interest in transaldolase stemmed from the observation that tald-1(RNAi) induces the UPRmt, activation of this mitochondrial stress response does not appear to be involved in mediating the longevity phenotype. Instead, lifespan extension from tald-1(RNAi) likely involves at least two outputs previously associated with longevity: induction of autophagy and activation of the flavin-containing monooxygenase 2 (Fig 8).
The relationship between PPP activity and mitochondrial function is particularly intriguing. Our studies indicate that inhibition of the PPP is sufficient to reduce respiration rates in vivo and remodel the mitochondrial network by activating mitochondrial fission, but importantly, this is accomplished without apparent functional changes to the ETC itself, as evidenced by the normal in vitro activity of purified mitochondria. This mechanistically differentiates tald-1(RNAi) from the well-characterized long-lived ETC-deficient animals such as cco-1(RNAi) and isp-1(qm150), which directly impair ETC structure and function [32, 69]. Our findings also support mammalian literature where mitochondrial function is altered by transaldolase deficiency. For example, lymphoblasts isolated from transaldolase deficient patients exhibit decreased mitochondrial membrane potential, increased mitochondrial mass, and increased H2O2 levels, while transaldolase deficient mice are infertile due to mitochondrial defects in spermatozoa [39, 40]. Although the UPRmt is apparently not involved in mediating the lifespan effects, its activation clearly indicates mitochondrial stress in vivo in the tald-1(RNAi) animals. One potential source of this mitochondrial stress could be increased levels of ROS, as indicated by the HyPer reporter and the enhanced sensitivity of tald-1(RNAi) animals to paraquat.
These findings highlight the importance of the PPP not only as a key pathway involved in central carbon metabolism, but also as a signaling hub. This close monitoring of PPP activity is logical, as it lies at the intersection of nucleotide metabolism, fatty acid/sterol synthesis, redox regulation, and glycolysis. In this light, the starvation like-response to tald-1(RNAi) is of particular interest, since it suggests that decreased PPP flux is monitored by the cell and results in diminished growth signaling. We speculate this occurs at least partially through decreased mTORC1 signaling, as we observed increased autophagic flux and activation of HLH-30, which is negatively regulated by mTORC1 [63, 70–73]. Furthermore, this starvation-like response caused a metabolic shift that depleted intestinal fat stores and rewired lipid metabolism to downregulate the stearoyl-CoA desaturase (Δ-9-desaturase, SCD) fat-7, upregulate mitochondrial fatty acid import genes, and the glyoxylate gene icl-1, among others. A reduction in fat-7 expression limits monounsaturated fatty acid synthesis, which maintains saturated fatty acid levels, but could also alter cellular and membrane lipid composition, including that of the mitochondria [74]. Alternatively, decreased fat-7 levels may indicate one arm of a concerted effort to breakdown fats through gene expression changes, as fat-7 negatively regulates β-oxidation [56, 58]. We suspect this gene expression program promotes the mobilization and breakdown of fatty acids for both energy metabolism and gluconeogenesis through the mitochondrial glyoxylate pathway [75].
In this study, we implicated stress-activated MAPKs as one class of sensors that respond to reduced PPP activity and appear to be independent of HLH-30 activity. It is unclear whether direct interactions between enzymes or products of the PPP regulate MAPKs or if multiple indirect steps connect their activities. NADPH produced by the PPP not only maintains a reduced cytosolic redox environment, but also affects antioxidant systems such as thioredoxin, glutaredoxin, and peroxiredoxin that respond to oxidative stress via thiol-based chemistry to initiate downstream signaling events. For example, activity of the MAP3K ASK1/NSY-1 is fine-tuned via thiol-disulphide exchange reactions mediated by these redox proteins [76–80]. Thus, we speculate that a shift to a more oxidative cytosolic redox from PPP inhibition is coupled to activation of ASK1/NSY-1 and downstream p38 and JNK MAPK signaling. Accordingly, in a context dependent fashion, C. elegans p38 and JNK MAPKs regulate stress resistance from various oxidative insults and longevity from dietary restriction interventions such as intermittent fasting and metformin treatment [35, 47]. Our data further confirms that elevated cytosolic H2O2 correlates with MAPK mediated lifespan extension in novel and distinct contexts: RNAi knockdown of a PPP enzyme and an ETC Complex IV subunit. Interestingly, the MAP3K NSY-1 was required for the full lifespan extension from both interventions, but differences existed for downstream MAPK requirements. For example, tald-1(RNAi) required both the p38 MAPK PMK-1 and the JNK MAPKs JNK-1 and KGB-1 for lifespan extension, while cco-1(RNAi) only required the JNK MAPK branch. Furthermore, our discovery of an unreported role for the JNK MAPK pathway in mediating ETC RNAi longevity is intriguing, as no other genes outside hif-1 and the p38 MAPK pmk-3 have been reported to mediate these effects in C. elegans [20].
The simplest model for enhanced longevity downstream of tald-1(RNAi) is through activation of HLH-30, which has been previously shown to promote longevity downstream of dietary restriction, mTOR signaling, and insulin/IGF-1-like signaling [62]. Prior studies have focused primarily on activation of autophagy and lipophagy by HLH-30 [62, 63], but we recently reported that FMO-2 is another important pro-longevity HLH-30 target that is activated by both dietary restriction and the hypoxic response [68]. The exact role of FMO enzymes outside xenobiotic metabolism is not well understood, but they are induced by various redox stressors and are important for resistance to reductive stress, which affects endoplasmic reticulum protein homeostasis [68, 81, 82]. One proposed function of FMOs may be to counterbalance GSH-mediated redox buffering to promote an oxidative redox environment through O2- and NADPH-dependent oxidation of biological thiols [81–83]. Adding to the complexity of FMOs, we observed that both deletion and overexpression of fmo-2 extend lifespan at 25°C. Interestingly, HIF-1 shows a similar effect on longevity at 25°C, where both deletion and hyperactivation of HIF-1 extend lifespan [84]; these observations could be linked since fmo-2 is a target of HIF-1 [68]. In the case of hif-1 deletion at 25°C, lifespan extension requires daf-16 [84], demonstrating that longevity pathways compensate for each other to regulate organismal stress resistance and aging. Our data are consistent with the model that tald-1(RNAi) lifespan extension requires fmo-2, but we acknowledge that other factors downstream of either fmo-2 deletion (i.e. longevity factors induced by reduced fmo-2 expression) or HLH-30 could also be responsible.
One intriguing twist to this model is that, unlike either dietary restriction [60, 85] or activation of the hypoxic response [86], tald-1(RNAi) must occur during development in order to promote longevity. This is similar to the mitochondrial longevity mutants, which have previously been thought to be largely mechanistically distinct from these other longevity pathways. HIF-1 is known to be activated in some long-lived mitochondrial mutants in response to ROS and to mediate part of their lifespan extension [20]; however, HIF-1 is not required for lifespan extension from tald-1(RNAi). Thus, our data suggest that the PPP mediates a complex interaction between several portions of the overall longevity network in worms that have previously been studied as genetically distinct “pathways”. These interactions will be of interest for future studies of longevity and aging in C. elegans.
Given the highly conserved nature of the PPP and its interactions with cellular metabolism, redox balance, and stress resistance, it is interesting to consider the extent to which the observations reported here will translate to mammals. As previously mentioned, there is good reason to believe that transaldolase deficiency can similarly impact mitochondrial function, metabolism, and oxidative stress resistance in mammals. To the best of our knowledge, there are no reports of PPP or transaldolase inhibition extending lifespan in a mammal; however, the downstream effectors of tald-1(RNAi) in worms are likely to play a conserved role in aging, as numerous studies have implicated autophagy in mammalian aging [87] and FMO-2 orthologs are among the most consistently induced enzymes in numerous long-lived mouse models [88, 89].
In summary, we uncovered a novel role of the PPP not only as a central metabolic pathway, but also as a signaling hub that connects the UPRmt, p38 and JNK MAPK signaling, and a starvation response mediated by HLH-30 and FMO-2 to promote cellular homeostasis and organismal longevity.
RB967 (gcn-2(ok871)), ZG31 (hif-1(ia4)), CF1038 (daf-16(mu86)), CB4037 (glp-1(e2141)), VC8 (jnk-1(gk7)), KB3 (kgb-1(um3)), KU25 (pmk-1(km25)), VC1668 (fmo-2(ok2147)), STE68 (nhr-49(nr2041)), VC1024 (pdr-1(gk448)), SJ4100 (zcIs13[hsp-6p::gfp]), SJ4143 (zcIs17[ges-1p::gfpmt]), BX113 (waEx15 [fat-7p::gfp + lin-15(+)]), MAH235 (sqIs19 [hlh-30p::hlh-30::gfp + rol-6(su1006)]), KAE9 (eft-3p::fmo-2 + h2b::gfp + Cbr-unc-119(+)), and VS20 (hjIs67 [atgl-1p::atgl-1::gfp + mec-7::rfp]) were obtained from the Caenorhabditis Genetics Center (Minneapolis, MN). The atfs-1(tm4525) and hlh-30(tm1978) strains were obtained from the National BioResource Project (Tokyo, Japan).
The fmo-2p::mCherry reporter strain, a transcriptional reporter, was created by microinjecting RBW6699 worms with a solution of 50ng/μL of the BSP190 construct containing 2076 bp of genomic sequence preceding the ATG of the fmo-2 coding sequence followed by the mCherry coding sequence and the unc-54 3’ UTR. A single copy insertion was generated at the chromosome II ttTi5605 locus using the Mos1 mediated Single Copy transgene Insertion (MosSCI) protocol [90].
Fluorescence microscopy was performed using Zeiss SteREO Lumar.V12 and Nikon Eclipse E600 microscopes. Worms were immobilized using sodium azide, mounted onto 3% agarose pads, and imaged within a few minutes for reporter experiments. Levamisole was avoided for imaging hlh-30p::hlh-30::gfp animals, since it caused rapid HLH-30 nuclear localization. For reporter assays worms were developed on RNAi bacteria at 20°C and imaged on day 1 of adulthood, except for fmo-2 reporter experiments, where day 2 adults were imaged. At least three independent experiments with approximately 10 animals per condition per experiment were performed for each reporter with similar results. Statistical analysis for quantification of reporters was performed using student’s t-test with Bonferroni’s correction or ANOVA with Bonferroni’s post-hoc, * p<0.05, ** p<0.01, *** p<0.001.
Confocal microscopy was performed using the Zeiss 510 META Confocal (for imaging mitochondrial morphology) or Leica SP8X (for imaging the HyPer reporter). For imaging intestinal mitochondrial morphology, animals were immobilized using levamisole and mounted on 10% agarose pads to prevent movement during image acquisition. Mitochondria were imaged with a 100X oil objective and Z-stacks of the posterior intestinal cells were taken at 0.34 μm increments. Gain settings for each image were maximized without over-saturation to emphasize mitochondrial content regardless of GFP expression, import, and folding levels. For imaging processing, Z-stacks were deconvoluted using the Iterative Deconvolve 3D plugin in Fiji and 5 image slices were projected using max intensity projection. Mitochondrial content was analyzed by thresholding the 5 image slice projections of different animals for each condition and quantifying the % area of signal within cell boundaries. For quantification, multiple animals for each condition in at least 2 independent experiments were analyzed.
For imaging the HyPer reporter, we removed the animal autofluorescence by performing linear unmixing similar to previous work [36]. Rather than using mean autofluorescence values for age matched animals, however, we took advantage of fluorescence lifetime gating using HyD detectors. By combining the extremely short fluorescence lifetime of the HyPer reporter, and the long autofluorescence lifetime, we were able to use a simple channel subtraction approach to signal unmixing. HyPer transgenic worms were immobilized in 2 mM levamisole on a 3% agarose pad and the anterior of the worm was line-scanned with a 20x objective at 405ex500-550em [1.5–5.5 ns], 488ex500-550em [1.5–5.5 ns], and 405ex450-470em [7–11.5 ns] ("autofluorescence" channel). The fluorescence lifetime of HyPer is extremely short (<4ns), so by using an emission wavelength far from that of HyPer and a much longer lifetime gating [7–11.5ns], the autofluorescence channel contains no signal from HyPer. This simplifies the unmixing problem allowing us to determine R=FAFinHyPerFAFinAF where R is ratio of the autofluorescence as measured in each separate HyPer channel to the autofluorescence as measured in the autofluorescence channel. This ratio is determined using age matched wild-type animals fed EV(RNAi) and imaging in all three channels. The ratio is then determined by fitting a linear model to values for each pixel in all channels for multiple animals. To ensure that the any variation in imaging parameters or laser function is corrected for, this process is repeated for each of the three experimental groups. The HyPer fluorescence was then determined by removing the autofluorescence contribution to each HyPer channel by the following: FHyPer = FHyPer Raw − (R * FAF in AF). This allowed the autofluorescence to be removed pixel-by-pixel for each animal. The 488ex500-550em channel displayed negligible autofluoresence after fluorescence lifetime gating, so this was only done for the 405ex500-550em channel. The final HyPer redox values were determined as FHyPer488ExFHyPer405Ex following autofluorescence removal. For each worm the head was manually outlined, and the focal plane (z-slice) with the greatest combined fluorescence within the head was used for quantification. Intensity normalized ratiometric (INR) images were generated as previously described [36].
Synchronized eggs or L1 larvae were grown on NGM plates containing 4 mM IPTG, 25 μg/ml carbenicillin and seeded with RNAi bacteria. At the L4/young adult stage, worms were transferred to plates with 50 μM FUDR to prevent hatching of progeny. When necessary, worms were transferred to new plates with fresh bacteria. Lifespans were performed at 25°C for the majority of experiments, unless otherwise noted in the text and figures. Cohorts were examined every 1–3 days using tactile stimulation to verify viability of animals. Animals that displayed vulval rupture were included in analysis, since it is an age-related phenotype [84, 91]. Animals lost due to foraging or bagging were not included in the analysis. All lifespan analyses were replicated using independent cohorts on different dates with replicate statistics provided in S1 Table. p-values were calculated using the Wilcoxon rank-sum test.
To measure in vivo oxygen consumption in C. elegans, we utilized the Seahorse X24 Bioanalyzer (Seahorse Biosciences) as previously described [30, 92]. Worms were grown on concentrated RNAi bacteria (0.15 g/ml) for 3 days at 20°C starting from the L1 stage, washed from plates, and rinsed from bacteria with M9 buffer 4+ times, before being placed in Seahorse XF24 Cell Culture Microplates for analysis. Basal respiration for each condition was analyzed using the average respiration of 5 well replicates over the course of one hour. Respiration for each genotype was measured in at least 4 independent experiments.
To measure activity of the ETC, mitochondria were isolated from C. elegans treated with RNAi bacteria as previously described [93, 94]. To ensure sufficient material for mitochondrial isolations, worm populations were grown for three generations at 20°C. Initially, animals were grown on concentrated RNAi bacteria for two generations and then transferred into 4–6 250 ml liquid cultures. Liquid cultures were propagated for 4–5 days depending on condition and monitored for developmental progression of animals and bacterial density (maintained at ~2 x 1010 cells/ml) to avoid starvation. Animals treated with cco-1(RNAi) were grown for two generations on EV(RNAi) bacteria and then transferred to liquid cultures containing cco-1(RNAi), due to developmental and fecundity issues associated with multiple generations of cco-1(RNAi). Respiration of isolated mitochondria was measured in 4 independent experiments for each condition.
Measurement of in vivo H2O2 levels was performed using the transgenic HyPer reporter as previously described [36]. Worms were grown on concentrated RNAi bacteria for 4 days starting from L1 (due to growth delay of this strain), washed from plates, and rinsed from bacteria with M9 buffer. At least 3 replicates of 1,000 worms for each condition were pipetted into a black flat bottom 96-well plate. N2 animals grown on EV(RNAi) were used as a background control. Fluorescence measurements were made using a BioTek Synergy H1M plate reader.
Oil Red O staining and analysis was performed as previously described [95]. To quantify fat staining for each condition photos were converted to RGB color, a pseudo flat field correction was applied, images were separated into their respective RGB channels, and fat staining was thresholded in the green channel consistently across all images for a particular experiment. Fat content for each worm was quantified using the integrated density (limited to thresholded signal) of a 40 pixel diameter circle placed below the pharynx (i.e. over the anterior intestinal cells). Two independent experiments were obtained for quantification.
Levels of NAD+, NADH, NADP, and NADPH were determined via Ultra Performance Liquid Chromatography coupled with Mass Spectrometry as previously described [96] with some modifications. Briefly, L4 worms were homogenized in 20% HEPES-buffered methanol (pH 7.5) on dry ice. 5 μL of the extract was separated on a BEHAmide column (Waters, Milford MA) using a Acquity UPLC (Waters) and analyzed with a Xevo TQ (Waters, Milford MA) in multiple reaction monitoring mode (MRM). LC solvents were A: H2O with 10 mM Ammonium Acetate and 0.1% NH4OH, and B: 95:5 Acetonitrile H2O with 10 mM Ammonium Acetate and 0.1% NH4OH (Alkaline Gradient) for all metabolites. Unique transitions for each metabolite were employed as described previously [96]. The gradient was as in Table 1.
RNA was isolated from young adult worms using a TRIzol (Life Technologies) chloroform extraction and cDNA was prepared using iScript Reverse Transcription Supermix for qRT-PCR (Bio-Rad). qRT-PCR was used to measure the expression levels of target genes (iTaq Universal SYBR Green Supermix, Bio-Rad) and normalization controls pmp-3 and cdc-42 (TaqMan Gene Expression Assays, Life Technologies). The relative standard curve method was used to calculate gene expression. Primers of target genes are listed in S2 Table.
Protein was isolated from young adult/adult day 1 worms by flash freezing worm pellets in liquid nitrogen followed by extraction in lysis buffer [20 mM HEPES, pH 7.4, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1% (v/v) Triton X-100, and 1x Pierce Protease Inhibitor Mini Tablets, EDTA Free (88666, ThermoFisher Scientific)]. Proteins of interest were detected by immunoblot using anti-GFP (sc-9996; Santa Cruz Biotechnology), anti-p-JNK (Cell Signaling Technology), anti-p-p38 (Cell Signaling Technology), anti-p-KGB-1 (a gift from Drs. Naoki Hisamoto and Kunihiro Matsumoto), and anti-alpha-tubulin (Clone: DM1A, MS-581-P0, Neomarkers) antibodies at a 1:1000 dilution in 5% BSA TBS-T.
|
10.1371/journal.pcbi.1004475 | ClassTR: Classifying Within-Host Heterogeneity Based on Tandem Repeats with Application to Mycobacterium tuberculosis Infections | Genomic tools have revealed genetically diverse pathogens within some hosts. Within-host pathogen diversity, which we refer to as “complex infection”, is increasingly recognized as a determinant of treatment outcome for infections like tuberculosis. Complex infection arises through two mechanisms: within-host mutation (which results in clonal heterogeneity) and reinfection (which results in mixed infections). Estimates of the frequency of within-host mutation and reinfection in populations are critical for understanding the natural history of disease. These estimates influence projections of disease trends and effects of interventions. The genotyping technique MLVA (multiple loci variable-number tandem repeats analysis) can identify complex infections, but the current method to distinguish clonal heterogeneity from mixed infections is based on a rather simple rule. Here we describe ClassTR, a method which leverages MLVA information from isolates collected in a population to distinguish mixed infections from clonal heterogeneity. We formulate the resolution of complex infections into their constituent strains as an optimization problem, and show its NP-completeness. We solve it efficiently by using mixed integer linear programming and graph decomposition. Once the complex infections are resolved into their constituent strains, ClassTR probabilistically classifies isolates as clonally heterogeneous or mixed by using a model of tandem repeat evolution. We first compare ClassTR with the standard rule-based classification on 100 simulated datasets. ClassTR outperforms the standard method, improving classification accuracy from 48% to 80%. We then apply ClassTR to a sample of 436 strains collected from tuberculosis patients in a South African community, of which 92 had complex infections. We find that ClassTR assigns an alternate classification to 18 of the 92 complex infections, suggesting important differences in practice. By explicitly modeling tandem repeat evolution, ClassTR helps to improve our understanding of the mechanisms driving within-host diversity of pathogens like Mycobacterium tuberculosis.
| Within-host heterogeneity of an infection can arise through two distinct mechanisms: within-host mutation and reinfection. While current genotyping techniques based on MLVA (multiple loci variable-number tandem repeat analysis) can identify within-host diversity, standard methods for classifying the mechanism driving this diversity have limitations. We present ClassTR, a novel approach for classifying these types of complex infections. ClassTR uses optimization to resolve complex strains into simple strains and explicit models of tandem repeat evolution to classify the infections as clonal (due to within-host diversification) or mixed (due to reinfection). We illustrate ClassTR and validate its findings in the context of Mycobacterium tuberculosis infections. We construct simulated datasets to identify the best-performing variant of our method and find that it is significantly more accurate than the standard method of classification. We apply ClassTR to data from a study in South Africa and find substantial differences in the classifications produced by ClassTR and the standard method, demonstrating the real-world relevance of this approach. Our work suggests that an analysis of complex infections based on an evolutionary model improves our understanding of the drivers of within-host diversity.
| The genotyping technique known as MLVA (multiple loci variable-number tandem repeats analysis), which identifies the number of copies of tandem repeat regions at specific pre-selected loci, has benefited the study of many bacteria. Data produced by MLVA can be used to glean information about bacterial lineage, pathogenicity and relation to other bacteria of the same species [1]. Our study focuses on a specific bacterium, Mycobacterium tuberculosis, but our methods are generally applicable to a variety of bacteria.
Genetic and genomic approaches for interrogating the composition of Mycobacterium tuberculosis infections occurring within individuals has in some settings revealed an impressive degree of complexity, reflecting both within-host mutation and reinfection as distinct routes to complexity [2]. These complex infections, especially those comprising both drug-susceptible and drug-resistant isolates (i.e. heteroresistance), can undermine the effective treatment of individual patients [3–5], complicate laboratory testing and evaluation of treatment programs [2], and affect the transmission dynamics of disease in communities [6, 7]. While an individual’s clinical response to treatment may not depend on whether heteroresistance has arisen through within-host mutation or by reinfection, our ability to distinguish these mechanisms has profound implications for our understanding of the natural history of disease and for projections of disease trajectories. For example, high contributions of reinfection indicate limited immune protection associated with previous infection, and have implications for the impact of new and existing vaccines [8] and for the effectiveness of preventive therapy [9]. High contributions of within-host mutation would affect expected rates of acquired resistance and would have implications for optimal antibiotic dosing strategies [10, 11]. Accordingly, accurate estimates of the prevalence of complex infections among tuberculosis patients and new methods for distinguishing the relative contributions of within-host mutation and reinfection to within-host diversity would be valuable.
Mycobacterial interspersed repetitive unit-variable number tandem repeat (MIRU-VNTR), the specific name of the MLVA technique for Mycobacterium tuberculosis, is a currently favored approach for genotyping strains and offers advantages for detecting within-host heterogeneity over other methods such as spacer oligonucleotide sequencing (spoligotyping) and restriction fragment length polymorphism analysis (RFLP). These molecular genetic approaches for TB genotyping are reviewed in Mathema et al. [12] and an evaluation of their utility for detecting complex infections is described by Cohen et al [2]. MIRU-VNTR is a microsatellite typing system which produces a readout containing the number of copies of a repeat region at several pre-selected loci [13, 14]. These copy number variants (CNVs) can then be used to compare the Mycobacterium tuberculosis strain to other similarly typed strains. If a patient harbors a complex infection, there will frequently be 2 (and sometimes 3) different CNVs at a single locus, and these can result from a clonally heterogeneous or a mixed infection. This situation is illustrated in Fig 1.
Classifying complex MIRU-VNTR patterns as being due to either within-host mutation or reinfection is challenging. The current accepted approach for distinguishing clonal heterogeneity from mixed infection is a simple rule-based method: if two or more loci have multiple CNVs the infection is classified as “mixed”, whereas if only one locus has multiple CNVs the infection is classified as “clonally heterogeneous” [13–15]. This approach is sensible given that the more complexity observed within a particular genotype, the more likely it is to be due to reinfection with a distinct second strain. In addition, there are several sources of evidence which suggest that clonal evolution occurring over a relatively short period is unlikely to result in multiple complex loci [16–18]. Nonetheless, the rule this approach is based on suffers from several limitations. First, it does not take the context of the infection into account (namely, whether the constituent strains are present in other members of the population). Second, it does not distinguish between copy numbers that are a small genetic distance apart (such as 3 and 4) from ones that are far apart (such as 3 and 15), even though clonal heterogeneity is less plausible in the latter case. Third, this approach does not facilitate the resolution of mixed infections into their constituent strains.
We propose a new method, which we call ClassTR, to classify complex infections using MLVA data. Our method is based on an established model of tandem repeat evolution that accounts for the stepwise character of mutations, which we extend by using differential rates of evolution for different loci. ClassTR leverages the entire set of isolate genotypes collected in a population in order to resolve complex strains into simple strains (i.e. strains with only one CNV at each locus). Then, using a model of tandem repeat evolution it identifies the most likely sources of each simple strain to establish the probability of each patient having a mixed infection. We show that ClassTR outperforms the standard rule-based method, reducing its error rate by 61% on simulated data, and produces significantly different classifications than the standard method on a dataset collected from a community in KwaZulu-Natal, South Africa. ClassTR is implemented in the R Statistical Computing Language [19] in Supplementary Materials (S1 Code).
We say that a patient harbors a complex infection if at least one of the MIRU-VNTR loci contains 2 different CNVs. We assume that there are always 1 or 2 CNV per locus, and this is indeed what we usually observe in practice. The ClassTR algorithm classifies these complex infections as resulting from either clonal heterogeneity or mixed infection (when both clonal heterogeneity and mixed infection are present, ClassTR classifies the infection as mixed). It includes three steps, each of which is briefly discussed below and described in more detail in the Methods section. Briefly, ClassTR starts by creating an optimization problem to resolve the complex strains into their constituent simple strains. After solving this optimization problem, ClassTR uses the resulting simple strain representation of complex strains to infer the possible provenance of each of these complex strains. Finally, it computes the probability of clonal heterogeneity for each patient with a complex infection. We illustrate this process on a small example with 3 simple and 3 complex strains in Figs 2 and 3.
Following Aandahl et al [20] we define distances between strains based on explicit models of tandem repeat evolution: a constant model and a linear model. Both models assume that copy numbers evolve in a stepwise fashion, consistent with the process of slipped-strand mispairing [21]. The constant model assumes a Poisson process at each locus by which the copy number increases or decreases by 1 at a constant rate, while the linear model assumes a Poisson process at each copy, so that the rate of mutation is proportional to the current copy number. In both cases the distance between two strains represents the total number of mutation events required to go from one to the other. In addition to these basic models where different loci undergo mutations at the same rate, we also consider weighted models in which different loci mutate at different rates. In order to estimate these locus-specific mutation rates we use measures of locus diversity.
We say that a set of simple strains exactly covers a complex strain if the set of CNVs at each locus of the simple strains is precisely the CNVs at the corresponding locus of the complex strain. In general, a strain with one or more complex loci having 2 CNVs each can be exactly covered by 2 simple strains. However, in the absence of additional constraints there can be as many as 2q−1 possible such covers of a strain with q complex loci, which is 2048 possibilities for q = 12, the largest we observe in our data. We make a parsimony assumption and search for the covers of the complex strains that introduce the smallest possible number of additional simple strains (i.e. ones not observed in the original dataset). This defines an optimization problem which may have multiple solutions, especially when the cases are not densely sampled from the population. We narrow down alternative possibilities by a system of rewards for using a strain frequently observed in the dataset and penalties for strains that are far removed from any other simple strains in the dataset.
Once the optimization problem is solved, every complex strain is represented as a superposition of simple strains. These simple strains can be present in the original dataset or newly added. For each of the newly added simple strains we compute one or more predecessors among the original simple strains, defined as the closest among these strains according to the distance we chose. The final probabilities are then obtained by comparing the sets of predecessors of the two strains constituting a given complex strain; the more similar they are, the more likely the strain is to be the result of clonal heterogeneity.
The South African dataset we work with consists of data collected during a prospective study of within-host diversity of M. tuberculosis. Briefly, 500 adult, sputum smear-positive TB patients in a geographic cluster of participating clinics in KwaZulu Natal were sequentially recruited for participation at the time of diagnosis and before treatment was initiated. Additional pre-treatment sputum was collected from each participant and cultured in solid and liquid media. Bacterial DNA was isolated from both media and genotyping was done by 24 loci MIRU-VNTR according to standardized protocols [14]. Out of the 500 study participants, the isolates of 436 (87%) were successfully typed and included in this study. Of the 436 patients included in the study, 92 (21%) had complex MIRU-VNTR patterns. We note that, like many other South African communities, the one in this study has a high HIV prevalence.
The standard rule-based classification method designated 44 of 92 of the complex strains as clonal and the remaining 48 as mixed, whereas our method classified 50 of them as clonal and 42 as mixed. Only one patient got assigned a probability of 1/4 for clonality and 3/4 for mixed, and we ended up using the majority rule and classifying their infection as mixed. There were a total of 18 discrepant calls, 6 in which a strain was called clonal by the standard method but mixed by ours, and 12 in which the reverse occurred. Our results are summarized in Table 1.
In order to evaluate the performance of the 8 different variants of our method (defined by the constant or the linear metric, as well as the commonly used Hamming metric and Goldstein metric, and weighted or unweighted loci), as well as the standard rule-based method based on the count of complex loci, we produced 100 simulated datasets with characteristics similar to our South African dataset. The details of our simulations are described in the Supplementary Materials (S1 Text). We attempted to match the original dataset in terms of its strain clustering characteristics, distribution of the number of complex loci in strains, and distribution of the differences between the CNVs in a complex locus. To this end, we simulated the evolution of an initial population of strains with random but constrained mutation and reinfection events a large number of times, selected a number of subpopulations of appropriate size, and selected the final datasets according to their distance to the two target distributions.
We selected 100 datasets of N = 415 strains each, n = 83(20%) of which were complex; 42 were the result of clonal heterogeneity and 41 were mixed infections. We applied the standard rule-based method and our method on each dataset. We evaluated the accuracy of each method as the average probability they assigned to the correct classification for the 83 complex strains; namely, if a method returned a probability p of clonality, we scored p if the complex strain was actually clonal and 1 − p if it was actually mixed. The accuracy of the standard rule-based method averaged 48%, not significantly different from the 50% that would be expected from a random classification. On the other hand, the accuracy of ClassTR using our metric of choice, the linear weighted metric, was 80%, for a 61% reduction in error. We also evaluated the correctness of the resolution of complex strains into their constituent simple ones, and found that ClassTR produced the correct resolution in 88% on clonal infections and 95% on mixed infections. The results of running different variants of our method on classification accuracy are shown in Table 2.
In addition, we created 9 groups of 100 datasets each, with similar characteristics but not constrained to resemble the original dataset as closely. Each group corresponded to a combination of one of three mutation rates (low, medium and high) and one of three reinfection rates (low, medium and high). Our method outperformed the standard rule-based method on all of them except for the classification at the low mutation rate. The results of running ClassTR on those datasets are shown in Table 3 (for the resolution, using randomly picked resolutions as the baseline) and Table 4 (for the classification, using the standard method as the baseline). They suggest that although the resolution performance of ClassTR deteriorates quite substantially on mixed infections at high mutation rates, its classification performance remains consistently good, both in absolute terms and in comparison with the standard method. This finding is further substantiated by Table 5, which shows that ClassTR finds the correct classification when the correct resolution is given most of the time, with a slight deterioration at high mutation rates. However, it is also able to find the correct classification from an incorrect resolution quite frequently, as evidenced by a comparison of all three tables.
ClassTR is the first alternative method for classifying complex bacterial strains as either clonally heterogeneous or mixed infections. In contrast with the existing rule-based classification method, it includes an explicit model of tandem repeat evolution and utilizes information from other strains collected locally, provides probabilistic rather than deterministic classifications, and allows for the identification of individual strains within complex infections.
There are two computational problems that ClassTR solves. The first one, which we called the Parsimonious Resolution Problem and showed to be NP-complete, is reminiscent of haplotyping problems in eukaryotic genomes [22]. The key difference is the haploid nature of the bacterial genomes we analyze; the observed complexity in the strains is the result of the genotyping technique we use rather than an actual allelic variation due to recombination. The second one, which is the classification problem for complex bacterial strains, is reminiscent of the type of problems that arise in cancer genomics [23] in deciding whether particular tumor genotypes are related to one another (similar to clonal heterogeneity) or have arisen independently (similar to mixed infection). The key difference is the evolutionary model we adopt for tandem repeats, which may not apply to cancer.
At the current time we lack a “gold standard” approach for determining which infections are actually complex, and whether these complex infections are due to within-host mutation or reinfection. This makes it challenging to evaluate the relative performance of the standard rule-based approach and ClassTR. While future advances in genome sequencing are likely to provide additional data to test ClassTR against the standard rule-based approach, our results on simulated data suggest that ClassTR provides more accurate classifications than the standard approach, and the additional computational resources are justified by the improvement in classification accuracy. While the performance of ClassTR and the standard method for classification of complex infections is similar if the true mutation rate of MIRU-VNTR loci is at the lowest end of the plausible range, ClassTR outperforms the standard method under scenarios with higher mutation rates within this range. Furthermore, applying ClassTR to our data from KwaZulu Natal generates results that are substantially different from the standard rule-based method, which demonstrates that the difference between these approaches is not just theoretical. In addition, ClassTR provides the only available approach to extract the constituent strains involved in a mixed infection from MIRU-VNTR data alone.
There are several opportunities to modify the models we have used to even better reflect the evolutionary process driving copy number variation. First, our evolutionary model does not account for the fact that some evolutionary events may duplicate multiple segments in a single timestep. A more sophisticated model might allow for such duplication events to happen, albeit at a small rate, and this rate could be estimated from available data. Second, we model copy number increases and decreases symmetrically, whereas a more flexible model could allow these events to occur at different rates. Finally, an alternate model might be needed to account for the possibility that two strains may be simultaneously transmitted from one person to another or for the possibility of having more than two strains within a host, which may be relevant for certain types of infectious pathogens [24]. In addition, an intriguing opportunity for future work would be to investigate how accurate the classifications of complex infections as clonal or mixed could be at the time that each patient is admitted, rather than at the end of the study as we have done here, as well as to take into account the information about the strains found in a patient’s contacts, such as household contacts, perhaps by using these to constrain resolutions.
In conclusion, ClassTR is a tool which we believe will advance our capacity to identify the mechanisms underlying within-host heterogeneity in TB and other bacteria. By distinguishing within-host mutation from reinfection, we anticipate that this method will improve our understanding of the natural history of pathogenic infection at the individual patient level, and will improve our ability to project transmission dynamics and the effects of interventions in communities.
This part of the paper is organized as follows. First, we define simple and complex strains, and explain how simple strains can cover complex strains. These definitions allow us to formulate our first problem, that of resolving the complex strains into simple strains by introducing as few simple strains unobserved in the data as possible. In the Supplementary Materials (S1 Text) we show that this problem is NP-complete. We continue by showing how this problem can be efficiently solved using a mixed integer linear programming formulation. In the Supplementary Materials (S1 Text) we also explain how this problem can be simplified using graph decomposition, and how the number of solutions can be further reduced by using an idea from information theory known as the minimum description length. We conclude by describing the model of tandem repeat evolution and the method we use to classify complex strains as arising from clonal heterogeneity or mixed infection. In the Supplementary Materials (S1 Text) we further elaborate on some alternative models we have considered and how they influence our results.
In this section we define simple and complex strains and describe the principled way in which ClassTR separates complex strains into simple strains.
We formally define a simple strain as a string of length L (for MIRU-VNTR, L = 12 or L = 24) over the alphabet A consisting of all integers from 0 to some upper bound tmax. If s is a simple strain, we denote by sj its j-th symbol. We define a complex strain as a string of length L over the alphabet P ( A ), the power set of A, so that each of its symbols is a subset of A. If s is a complex strain, we call sj the content of s at position j.
A collection C of simple strains will be called a cover for a complex strain s if at each position 1 ≤ j ≤ L, we have sj ⊂ ∪c∈C cj. In other words, the content of the strain s at each locus is included in C. A collection C will be called an exact cover for a complex strain s if equality holds, i.e. sj = ∪c∈C cj ∀1 ≤ j ≤ L; in this case, C includes the content of s at each locus, and nothing else. A collection C will be called a minimal (exact) cover of s if C is an (exact) cover of s and no proper subset of C is. We always look for minimal exact covers for reasons of parsimony.
When a complex strain s has all contents of size 1 or 2, there exist minimal exact covers of size 2, and the number of such covers is 2q−1, where q is the number of positions with content of size 2. The value of q attains a maximum of 12 in our dataset, meaning that a single complex strain can have up to 2048 different covers.
Given the multiplicity of possible minimal covers for each complex strain, we use a global parsimony assumption to identify the ones that are actually present. Namely, we assume that, all other things being equal, the fewer simple strains we add to the ones in the dataset to cover all the complex strains, the better. Intuitively, this means that we attempt to explain complex infections in terms of strains we have observed as simple infections in the population. Thus we seek to cover all the complex strains by adding the smallest possible number of strains. Fig 2 presents a toy example of a dataset with its solution.
In this section we formalize the problem of resolving the complex strains by introducing as few new simple strains as possible, which we call the parsimonious resolution problem. In the Supplementary Materials (S1 Text) we show that the decision version of this problem is NP-complete, even in the case of all copy number variants being 0 or 1. As a corollary, our proof establishes that the parsimonious resolution problem for spoligotype data (where a 0 indicates the absence and a 1 the presence of a particular region), a version of which was studied by Lazzarini et al [25], is also NP-complete.
The decision version of the parsimonious resolution problem (PRP) can be stated as follows.
Given: an integer L; a finite alphabet A; a set of strings S = {s1, s2, …, sk} of length L over A ∪ A2, but not entirely over A (i.e. each position contains 1 or 2 elements of A, with at least one position containing 2 elements of A); a set of “free” strings F = {f1, …, fm} of length L over A; an integer K.
Decide: whether there exists a collection C = {c1, c2, …, cK} of K strings of length L over A, such that, for each string s ∈ S there exist 2 strings c and c′ in C ∪ F, such that c ∪ c′ = s (where the union is taken component-wise).
The correspondence between the PRP and the problem we are actually solving is as follows: L is the number of loci, A is the set of possible CNVs, S are the complex infections present in the data and F are the simple infections present in the data. Finally, K is the number of additional (new) simple strains we are seeking to add to F in order to resolve all the complex infections.
In this section we formulate the 0–1 integer linear program [26] for the parsimonious resolution problem. This integer linear program finds a set of simple strains that cover the complex strains in the dataset, paying for each simple strain that is not present in the dataset. It minimizes the total cost of these newly added simple strains. Its inputs are a set of simple strains that can be used “for free” and the set of complex strains to be covered. Its outputs are the variables corresponding to the new simple strains used in covering the complex strains.
Let N be the number of simple strains, n be the number of complex strains, and qi be the number of complex loci in the ith complex strain. For simplicity we assume that there are exactly 2 copy number variants at each complex locus, which is the case for our dataset. Let Si be the set of all simple strains that may be used to cover the ith complex strain, so that |Si| = Qi = 2qi. Let us also define Q ≔ ∑ i = 1 n Q i. Let S ≔ ∪ i = 1 n S i and q := |S|. Note that q ≤ Q.
We define two categories of variables, one to indicate usage, and the other to indicate coverage. The usage variables are denoted uj and are defined for every strain j in S. The value of uj is 1 if the simple strain j is used in the cover of at least one complex strain, and 0 otherwise. The coverage variables are denoted cij and are defined for every complex strain i and every simple strain j in Si. The value of cij is 1 if the simple strain j is used to cover the complex strain i, and 0 otherwise. For a complex strain i and a simple strain j in Si, we denote by i∖j the complement of j in i, namely, the simple strain that, together with j, covers i (here we use the assumption that every complex locus has exactly 2 CNVs). The complement i∖j can be obtained by taking the CNV in each complex locus of i that was not used in j.
For each simple strain j in S we also define the cost wj of adding it to the cover. The objective function is simply a linear combination of the usage variables uj with the costs wj as coefficients. We always take wj = 0 if the simple strain j is present in the dataset, because it is already available to be used in a cover. We also take wj = 1 for any simple strain j not present in the dataset, so the total cost ends up being the number of new strains used. The optimal solution is the one minimizing this total cost.
This leads us to the following integer linear program formulation:
Minimize ∑ j w j u j subject to (1) u j ∈ { 0 , 1 } ∀ j (2) c i j ∈ { 0 , 1 } ∀ i , j (3) c i j = c i ( i \ j ) ∀ i , j (4) c i j ≤ u j ∀ i , j (5) u j ≤ ∑ i c i j ∀ j (6) 1 ≤ ∑ j c i j ∀ i (7)
The first two sets of constraints, Eqs (2) and (3), ensure that all the variables take values 0 or 1. The next set of constraints, Eq (4), ensures that the simple strain j is used to cover the complex strain i if and only if its complement simple strain, i∖j, is also used to cover the complex strain i. The next two sets of constraints ensure that the simple strain j is marked as used if (Eq (5)) and only if (Eq (6)) it is used to cover at least one complex strain i. Finally, the last set of constraints, Eq (7), ensure that the complex strain i is covered in at least one way by simple strains.
The number of uj variables and constraints in Eq (2) is q ≤ Q; the number of cij variables and constraints in Eq (3) is Q; the number of constraints in Eq (4) is Q/2; the number of constraints in Eq (5) is Q; the number of constraints in Eq (6) is q ≤ Q; and the number of constraints in Eq (7) is n, for a total of Q + q ≤ 2Q variables and (5/2)Q + 2q + n ≤ (9/2)Q + n constraints. In particular, for our South African dataset, n = 92 and Q is roughly 8,000, while for our simulated datasets, n = 83 and Q varies from 5,000 to 25,000, so the total number of variables is always under 50,000 and the number of constraints under 100,000. Integer linear programs of this size can typically be solved to optimality in seconds by CPLEX (available from http://www-01.ibm.com/software/integration/optimization/cplex-optimizer). The total time required by ClassTR is about 5 minutes for the South African dataset with N = 436 strains. We additionally tested our method on a much larger dataset containing N = 4075 strains with n = 364 of them complex. Its processing took less than an hour on a single CPU, suggesting that our algorithm scales well with input size in practice.
In this section we define the four distances we use in ClassTR. These distances can be used to construct the predecessor sets which then allow us to calculate the probability of each complex strain being clonally heterogeneous or mixed.
We define the constant metric dC between two simple strains as
d C ( s , s ′ ) = ∑ j = 1 L | s j - s j ′ | .
This corresponds to the minimum number of mutation events needed to get from one strain to the other in the constant model of tandem repeat evolution defined by Aandahl et al [20]. Indeed, since the constant model assumes a Poisson process at each locus, |i − j| is precisely the number of mutations required to get from i to j.
We also define the linear metric dL between two simple strains as
d L ( s , s ′ ) = ∑ j = 1 L ∑ k = min ( s j , s j ′ ) + 1 max ( s j , s j ′ ) 1 k .
This corresponds to the expected number of timesteps needed to get from one strain to the other in the linear model of tandem repeat evolution defined by Aandahl et al [20]. Indeed, since the linear model assumes that a Poisson process takes place at each copy, it takes an expected 1/m timesteps to go from m to m − 1 copies.
Two other standard metrics we use are the Goldstein metric dG and the Hamming (categorical) metric dH, respectively defined as
d G ( s , s ′ ) = 1 L ∑ j = 1 L ( s j - s j ′ ) 2 and d H ( s , s ′ ) = ∑ j = 1 L [ s j ≠ s j ′ ] ,
where [I] is the Iverson bracket whose value is 1 if expression I is true and 0 otherwise. Note that the Goldstein metric is not a metric in the traditional sense because it does not respect the triangle inequality.
In addition we define weighted analogs of all these metrics, which are obtained by multiplying the contribution of each locus by its weight. To estimate the weight of each locus ClassTR uses the Simpson index [27], also known as the Hunter-Gaston index [28], reasoning that the more diverse a locus is, the faster it evolves and the less weight it should carry. These weights then allow us to compute the corresponding weighted distances in the constant or linear models defined below, which we denote d c w and d L w, respectively. In our datasets these weights ranged from 0.16 to 1.
Given the set of simple strains generated by the optimization, we describe how to produce the final soft classification of complex strains along the clonally heterogeneous to mixed spectrum in this section.
We start by choosing a distance function d on simple strains. Given a strain j, we define the predecessor set P(j) as the subset of the simple strains S present in the original dataset that are closest to s according to d. Formally,
P ( j ) ≔ arg min s ∈ S d ( j , s ) .
Of course, for any strain s ∈ S, the predecessor set only contains s itself (we do not take the presence of duplicate strains into account). We also note that the more highly resolving the distance, the smaller the predecessor sets are going to be. Thus, the unweighted constant distance could give rise to ties for the closest strain, but the weighted constant distance or the linear distance is less likely to yield a tie.
Intuitively, the more similar the predecessor sets of the constituent strains are to each other, the more likely the complex strain is to be clonally heterogeneous. For example, if two different covering strains are both very close to the same simple strain in the dataset, the complex strain composed of both of them is more likely to be clonal than if the two strains’ nearest matches in the data are two very different strains. We formalize this by using the Jaccard index [29] to evaluate the similarity of two sets A and B, defined as the size of their intersection divided by the size of their union:
J ( A , B ) ≔ | A ∩ B | | A ∪ B | .
Suppose that A and B are the predecessor sets of the constituent strains of a complex strain. Then we take the Jaccard index of A and B as the probability of the complex strains being clonally heterogeneous. Thus, a complex strain covered by two strains with identical predecessor sets will be classified as clonally heterogeneous, while one with two strains with non-overlapping predecessor sets (for example one covered by two distinct simple strains present in the original dataset) will be classified as mixed, with intermediate variants also possible as shown in Fig 3. This probability is the value we report as our final classification.
|
10.1371/journal.pgen.1004419 | The Transcription Factor TFII-I Promotes DNA Translesion Synthesis and Genomic Stability | Translesion synthesis (TLS) enables DNA replication through damaged bases, increases cellular DNA damage tolerance, and maintains genomic stability. The sliding clamp PCNA and the adaptor polymerase Rev1 coordinate polymerase switching during TLS. The polymerases Pol η, ι, and κ insert nucleotides opposite damaged bases. Pol ζ, consisting of the catalytic subunit Rev3 and the regulatory subunit Rev7, then extends DNA synthesis past the lesion. Here, we show that Rev7 binds to the transcription factor TFII-I in human cells. TFII-I is required for TLS and DNA damage tolerance. The TLS function of TFII-I appears to be independent of its role in transcription, but requires homodimerization and binding to PCNA. We propose that TFII-I bridges PCNA and Pol ζ to promote TLS. Our findings extend the general principle of component sharing among divergent nuclear processes and implicate TLS deficiency as a possible contributing factor in Williams-Beuren syndrome.
| DNA translesion synthesis (TLS) allows the DNA replication machinery to replicate past damaged bases, thus increasing cellular tolerance for DNA damage and maintaining genomic stability. Suppression of TLS is expected to enhance the efficacy of the anti-cancer drug, cisplatin. TLS employs a special set of DNA polymerases, including Pol ζ. The TLS polymerases are also involved in somatic hypermutation and proper immune response in mammals. Thus, it is critical to understand the underlying mechanisms of TLS. In this study, we have discovered the transcription factor TFII-I as a new Pol ζ-binding protein in human cells. We show that TFII-I is indeed required for TLS and DNA damage tolerance. We further delineate the mechanism by which TFII-I contributes to TLS. Our study significantly advances the molecular understanding of TLS, and provides a fascinating example of component sharing among disparate nuclear processes. Finally, because one copy of the TFII-I gene is deleted in Williams-Beuren syndrome (WBS), our work implicates TLS deficiency as a potential causal factor of this human genetic disorder.
| DNA bases experience many types of damage caused by both endogenous and exogenous factors. DNA repair pathways, such as the global genomic nucleotide excision repair (GG-NER) pathway, actively remove damaged bases [1]. In addition, when damaged bases are not completely removed, DNA translesion synthesis (TLS) allows replication past these lesions, thus increasing DNA damage tolerance and maintaining genomic integrity [2]. TLS requires a set of specialized DNA polymerases, including the Y family polymerases, Rev1, Pol η, ι, and κ, and the B family polymerase Pol ζ containing the Rev3 catalytic subunit and the Rev7 regulatory subunit [2]. Certain TLS polymerases, including Pol ζ, are involved in somatic hypermutation of immunoglobulin genes [3]–[5].
Recent advances have established that multiple polymerase-switching events occur during TLS, and have begun to elucidate the elaborate molecular mechanisms that regulate these steps. When replicative polymerases encounter damaged DNA bases, such as those crosslinked by UV, the sliding clamp PCNA is ubiquitinated [6]. Ubiquitinated PCNA recruits Pol η, ι, or κ through the adaptor polymerase Rev1 [7]–[13]. Pol η, ι, or κ inserts nucleotides directly opposite to the DNA lesion [14]–[16]. In a poorly understood second switch, Pol ζ is employed to extend DNA replication past the lesion. Rev1 can simultaneously bind to Pol ζ and one of the Y family polymerases, Pol η, ι, or κ, suggesting that this polymerase switching step might occur through simple repositioning of a large, multi-polymerase assembly on the DNA template [17]. After TLS is completed, replicative DNA polymerases re-engage with PCNA and resume high-fidelity replication.
TFII-I was first identified as a general transcription factor that bound to a pyrimidine-rich Initiator (Inr) sequence at the transcription start site and supported transcription in an in vitro reconstituted system [18]. TFII-I contains an N-terminal dimerization domain, six repeated domains (called R1-R6), and four AlkB homologue 2 PCNA-interacting motifs (APIM) motifs, among other features [19], [20]. Recent studies have suggested that TFII-I is not a general transcription factor required for all Inr-dependent transcription [21]. Instead, it has signal- and context-dependent regulatory roles in the transcription of specific genes. Interestingly, TFII-I is one of 26–28 genes affected by a hemizygous deletion of the chromosome 7q11.23 region in the rare human disorder, Williams-Beuren syndrome (WBS) [22]. WBS patients exhibit a wide spectrum of phenotypes, including distinctive craniofacial features, cardiovascular abnormalities, and mental retardation. Heterozygous mutant mice with the N-terminal 140 residues of TFII-I deleted show WBS-like craniofacial and neurobehavioral alterations, linking this region of TFII-I to a subset of WBS phenotypes [23].
In this study, we show that TFII-I physically interacts with the Pol ζ subunit Rev7 (also known as Mad2B). Functional studies reveal that TFII-I is indeed required for TLS and DNA damage tolerance in human cells. Depletion of TFII-I affects the transcription of a small number of genes, none of which are known to be involved in TLS, suggesting that the TLS function of TFII-I is independent of its role in transcription. Instead, both PCNA binding and homodimerization of TFII-I are required for TLS. We propose that TFII-I connects Pol ζ to PCNA and facilities TLS. Because a TFII-I mutant lacking its N-terminal dimerization domain is defective in TLS, our findings also implicate TLS deficiency as a potential contributing factor of WBS.
Rev7 shares sequence similarity with the spindle checkpoint protein Mad2 [24]. Both contain a HORMA (Hop1-Rev7-Mad2) domain that mediates protein–protein interactions [25]. Because of our long-standing interest in the structure and function of Mad2 [26]–[28], we examined the function and regulation of Rev7. We created 293T cell lines stably expressing human Rev7 fused at its N-terminus with a tandem affinity purification (TAP) tag and purified TAP-Rev7 complexes from these cells with or without UV irradiation (10 J/m2). TAP-Rev7 preparations from both samples contained prominent doublet bands at 140 kDa (Figure 1A). Mass spectrometry analysis of the unirradiated sample revealed that these bands belonged to the transcription factor, TFII-I, which was known to have multiple alternative splicing variants (Table S1). The sequence coverage of TFII-I was 52.4%. In addition to TFII-I, we also identified the known Rev7-binding protein, ZNF828/CAMP [29], with a sequence coverage of 13.6%. Another potential Rev7-binding protein was CAD (Carbamoyl-phosphate synthetase 2, Aspartate transcarbamylase, and Dihydroorotase), a key multifunctional enzyme in the pyrimidine biosynthetic pathway, suggesting a possible link between TLS and pyrimidine biosynthesis. Rev3L was not detected in Rev7 preparations, presumably due to its low abundance in cells.
Because TFII-I was the most abundant Rev7-binding protein in the TAP-Rev7 samples, we focused on the Rev7–TFII-I interaction in this study. Endogenous Rev7 and TFII-I proteins interacted with each other in human cells (Figure 1B). TFII-I did not interact with Pol ι or Pol η (Figure S1A). Recombinant GST-Rev7 bound to in vitro translated TFII-I (Figure 1C). A minimal Rev7-binding domain of TFII-I was mapped to its middle region containing R2-R4 repeats (Figure 1D and Figure S1). This minimal TFII-I domain, however, bound more weakly to Rev7 than the full-length TFII-I did, suggesting that additional regions of TFII-I might contribute to Rev7 binding. These results suggested that Rev7 physically interacted with TFII-I.
As a regulatory subunit of Pol ζ, Rev7 simultaneously binds to a small Rev7-binding motif (RBM) in Rev3L and the C-terminal domain (CTD) of Rev1, thus bridging an interaction between Rev1 and Rev3L (Figure S2A) [17]. When bound to Rev3L, Rev7 adopts the closed conformation and traps Rev3L RBM with a topological embrace through its “seat belt” [30]. The Rev1 CTD binds Rev7 at a site opposite of the Rev3L-binding site [17], [31]. We next tested whether TFII-I binding to Rev7 was compatible with Rev3L–Rev7 or Rev1–Rev7 interactions. A recombinant purified TFII-I fragment (residues 350–667) containing R2-R4 co-fractionated with Rev7 bound to Rev3L RBM (residues 1847–1898) with Rev1 CTD (residues 1140–1251) (Figure 1E and 1F). Based on gel filtration, the native molecular mass of this miniature TFII-I–Rev7–Rev3L–Rev1 complex was 78 kDa, which was consistent with the formation of a 1∶1∶1∶1 heterotetramer with an expected molecular mass of 81 kDa. With the small amount of each protein loaded, the Rev3L fragment was not visible by Coomassie blue staining. This fragment could only be visualized with large amounts of proteins loaded (Figure S2B). Thus, our results suggest that TFII-I can bind to the Rev3L–Rev7–Rev1 complex in vitro.
Inactivation of either subunit of Pol ζ, Rev3L or Rev7, reduces colony formation of mammalian cells treated with UV and cisplatin, presumably because they are required to bypass DNA damage induced by these agents [32]–[34]. We first confirmed that human 293T and U2OS cells depleted of Rev3L or Rev7 with small interfering RNA (siRNA) were indeed sensitive to UV or cisplatin using colony formation assays (Figure 2). Because antibodies that could detect endogenous Rev3L were unavailable, the efficiency of Rev3L depletion was indirectly inferred from the reduction of Rev3L mRNA as measured by quantitative PCR (Figure 2B). Depletion of TFII-I similarly resulted in UV and cisplatin sensitivity (Figure 2). Importantly, depletion of both TFII-I and Rev7 did not cause more severe phenotypes than depletion of either one alone did, suggesting that TFII-I might be required for DNA damage tolerance. Efficient depletion of Rev7 and TFII-I was confirmed by Western blots. There were no discernable cell cycle defects in cells depleted of TFII-I, Rev3L, or Rev7 in the absence of UV (Figure S3 and data not shown).
To obtain additional evidence for a role of TFII-I in DNA damage tolerance, we stained control and TFII-I RNAi cells for γ-H2AX, a DNA double-strand break (DSB) marker, at different times following the treatment of low-dose UV, and performed flow cytometry analysis. UV irradiation induced DNA damage in all samples. UV-induced DNA damage is expected to stall replication forks in S phase and indirectly produce DSBs. About 40% of all groups of cells were in S phase and positive for γ-H2AX staining at 2 hrs following UV treatment (Figure 3A and S3). At 12 hrs, the majority of these cells were γ-H2AX-positive and blocked in S phase (Figure S3). At 24 hrs after UV irradiation, few siControl cells were γ-H2AX-positive, indicating that they had progressed through S phase and effectively repaired their damaged DNA (Figure 3A and S3). In contrast, the majority of cells depleted of Rev3L or Rev7 remained blocked in S phase, and were γ-H2AX-positive (Figure S3), consistent with a known role of Pol ζ in DNA damage tolerance. Cells depleted of TFII-I were also less efficient in passing through S phase and in repairing DNA damage, as about 40% of TFII-I RNAi cells had positive γ-H2AX staining at 24 hrs after UV irradiation (Figure 3A, B). Most cells positive for γ-H2AX had DNA contents between 2C and 4C, indicating that they were blocked in S phase. Importantly, ectopic expression of siRNA-resistant Myc-TFII-I transgene at levels comparable to that of the endogenous TFII-I largely rescued the S phase block and DNA damage of TFII-I RNAi cells (Figure 3A-D), based on both flow cytometry and γ-H2AX immunostaining. These results indicate that, like Pol ζ, TFII-I is required for DNA damage tolerance, S phase progression, and genomic stability. Pol ζ has recently been suggested to play a direct role in DSB repair through homologous recombination [35]. Our results cannot distinguish between a direct role for TFII-I and Pol ζ in DSB repair and an indirect role for them in DNA repair through supporting TLS and DNA damage tolerance. In the future, it will be interesting to test whether TFII-I is directly involved in DSB repair.
We directly tested whether TFII-I was required for translesion synthesis. To do so, we performed a mutation frequency assay on the UV-irradiated SupF shuttle vector plasmid pSP189 [36]. As expected, depletion of Rev3L or Rev7 greatly reduced the mutation frequency of the SupF region in the UV-damaged shuttle vector (Figure 4A), consistent with their known roles in TLS. Depletion of TFII-I with multiple siRNAs also reduced the mutation frequency of SupF. Importantly, depletion of both TFII-I and Rev7 did not produce a stronger phenotype as did the depletion of either protein alone (Figure 4B), suggesting that TFII-I might work in the same pathway as Pol ζ.
DNA sequencing of the mutated SupF clones revealed that inactivation of TFII-I or Pol ζ did not alter the mutation spectrum (Figure 4C and Figure S4). In all samples, the majority of mutations were C∶G to T∶A transitions. Thus, TFII-I depletion reduces TLS efficiency in human cells. We note that there might be subtle differences in the mutation hotspots among different samples (Figure S4). In addition, mutations involving large deletions appeared to be absent in the siTFII-I cells. The significance and the underlying reasons for these apparent differences are unclear at present.
Because TFII-I has known functions in transcription, we tested whether the TLS defects of TFII-I RNAi cells were indirectly caused by a defect in the transcription of TLS genes. Using quantitative PCR, we first showed that TFII-I depletion did not substantially alter the mRNA levels of Rev1, Rev7, Rev3L, and Rad18, genes known to be involved in TLS (Figure 5A). Next, we performed gene expression profiling of HeLa Tet-On cells transfected with siControl, siTFII-I, or siRev7. Depletion of TFII-I reduced by two-fold the mRNA levels of only 48 genes (Figure 5B). None of these genes are known to be involved in TLS. Therefore, the TLS deficiency caused by TFII-I RNAi is not an indirect consequence of gross transcriptional defects, although we cannot rule out the possibility that subtle transcriptional defects of multiple genes cumulatively impact TLS. The fact that TFII-I depletion only affects the transcription of so few genes in HeLa cells is not surprising, as many TFII-I target genes are involved in neuronal functions or immune response [37]. Furthermore, two other TFII-I related genes, GTF2IRD1 and GTF2IRD2, might have compensated for the partial loss of TFII-I.
Likewise, depletion of Rev7 only decreased the mRNA levels of about 50 genes (Figure 5B). Moreover, only 12 genes were commonly suppressed in both siTFII-I and siRev7 cells. Therefore, Rev7 does not appear to have a major role in transcription. The primary function of the Rev7–TFII-I interaction is unlikely to be transcriptional regulation in HeLa cells.
We next explored the mechanism by which TFII-I contributed to TLS. TFII-I contains four APIM motifs [20], [38], which mediates its binding to PCNA (Figure 6A). Because PCNA has critical roles in mediating polymerase switching during TLS, we tested whether PCNA binding by TFII-I was required for TLS. We created a TFII-I mutant with all four APIM motifs mutated to alanine (TFII-I mAPIM). The endogenous TFII-I interacted with PCNA (Figure 6B). Myc-TFII-I wild type (WT), but not the Myc-TFII-I mAPIM mutant protein, interacted with PCNA in cells depleted of endogenous TFII-I (Figure 6B). Moreover, consistent with an earlier report [20], GFP-TFII-I WT, but not GFP-TFII-I mAPIM, was recruited to laser-induced DNA damage sites in U2OS cells, along with DsRed-PCNA (Figure 6C). This result confirmed that TFII-I mAPIM lost its functional interaction with PCNA. Importantly, Myc-TFII-I mAPIM still interacted with Rev7 (Figure 6B). Compared to Myc-TFII-I WT, the Myc-TFII-I mAPIM mutant protein was significantly less efficient in rescuing TLS defects caused by TFII-I RNAi (Figure 6D). Thus, the PCNA-binding activity of TFII-I is required for its function in TLS.
The simplest model to explain the involvement of TFII-I in TLS is that TFII-I binds simultaneously to both PCNA and Rev7, bridging an interaction between the two proteins and contributing to the recruitment of Pol ζ to DNA lesions. Unfortunately, we could not detect the recruitment of GFP-Rev7 to laser-induced DNA damage sites, barring us from testing this notion using cytological methods. We, therefore, tested this hypothesis using IP-Western methods. Interactions among PCNA, TFII-I, and Rev7 were detectable in unirradiated cell lysates (Figure S5A). These interactions were enhanced following UV irradiation. More importantly, depletion of TFII-I abolished the interaction between PCNA and Rev7 in both cases (Figure S5A). Expression of Myc-TFII-I WT, but not mAPIM, restored the interaction between PCNA and Rev7 (Figure 6B). These results suggest that TFII-I bridges an interaction between PCNA and Rev7 in human cells, and that this function of TFII-I requires its APIM motifs.
We next checked whether the recombinant purified TFII-I330–667 fragment could form a ternary complex with Rev7 and PCNA using gel filtration. To our surprise, we found that TFII-I330–667, Rev7, and PCNA did not form a ternary complex (Figure 7A). Addition of PCNA to the pre-formed TFII-I330–667–Rev7 complex produced a TFII-I330–667–PCNA binary complex and free Rev7. Thus, binding of PCNA and binding of Rev7 to a monomeric fragment of TFII-I are mutually exclusive.
On the other hand, TFII-I is known to homodimerize, and contains an N-terminal dimerization domain [19]. Indeed, a TFII-I fragment containing residues 1–667 fractionated with an apparent molecular mass of about 160 kD on gel filtration columns, which was consistent with it forming a homodimer (Figure 7B). By contrast, the TFII-I270–667 fragment that lacked the N-terminal dimerization domain fractionated as a monomer by gel filtration. Moreover, differentially tagged TFII-I, but not a TFII-I mutant protein lacking the first 90 residues (TFII-I Δ90), interacted with each other in human cells (Figure 7C), confirming that TFII-I could oligomerize in vivo. Finally, the monomeric TFII-I Δ90 mutant protein was defective in supporting TLS (see Figure 6D above). Consistently, this mutant could not restore the PCNA-Rev7 interaction in TFII-I-depleted cells (Figure 6B). Therefore, homodimerization of TFII-I is required for its function in TLS and for bridging the PCNA–Rev7 interaction. We propose that one monomer of a TFII-I dimer can bind to PCNA while the other can bind to Rev7 (Figure 7D). In this way, the TFII-I dimer bridges the interaction between PCNA and Rev7, and contributes to the recruitment of Pol ζ to DNA lesions during TLS.
Because Rev7 also interacts with the C-terminal domain (CTD) of Rev1, we tested whether recruitment of Rev1 to DNA damage sites was dependent on TFII-I or Rev7. GFP-Rev1 was recruited to laser-induced DNA damage sites in human cells (Figure S5B). Depletion of TFII-I or Rev7 did not alter this recruitment. Thus, Rev1 is recruited to DNA damage sites independently of TFII-I and Rev7. Taken together, our results suggest that TFII-I and Rev1 collaborate to recruit Pol ζ to DNA damage sites through TFII-I–Rev7 and Rev1-CTD–Rev7 interactions (Figure 7D).
Pol ζ plays critical roles in DNA translesion synthesis (TLS), cellular DNA damage tolerance, and the maintenance of genomic stability. In this study, we have discovered the transcription factor TFII-I as a new, functionally important interactor of Pol ζ in human cells. We found that PCNA binding and dimerization of TFII-I are required for efficient TLS. Our study thus provides key insights into the mechanism and regulation of Pol ζ in human cells.
We propose the following model to explain the involvement of TFII-I in TLS (Figure 7D). In this model, Rev1 and the TFII-I homodimer are independently recruited to ubiquitinated PCNA at DNA damage sites. This complex then simultaneously engages Rev7 and recruits Pol ζ to these lesions. Rev1 also anchors Pol η, ι, or κ to PCNA. After these Y-family polymerases insert nucleotides directly opposite to the DNA lesion, Pol ζ extends DNA synthesis past the lesion. Because TFII-I specifically interacts with the Pol ζ subunit Rev7, but not with Pol η or ι, we speculate that TFII-I might also mediate polymerase switching from Pol η/ι/κ to Pol ζ.
In support of a role of TFII-I in recruiting Pol ζ to DNA lesions, we showed that TFII-I bridges an interaction between PCNA and Rev7 in UV-irradiated human cells, using IP-Western experiments. We could not reconstitute a complex containing TFII-I, PCNA, Rev7, Rev3, and Rev1 in vitro using purified recombinant proteins, due to the difficulty of expressing full-length TFII-I and larger fragments of Rev3 and Rev1. Complex formation might also require DNA or additional accessory subunits of Pol ζ [39]-[41].
Rev3L has been reported to contain a putative APIM motif [20]. In addition, PolD3 (p66), an accessory subunits of Pol ζ, contains a functional PCNA-binding PIP motif [41], [42]. Furthermore, in addition to its ability to bind ubiquitin on ubiquitinated PCNA, Rev1 has been implicated in direct binding to unmodified PCNA [8], [43], [44]. Therefore, along with our finding that TFII-I binds to PCNA and Rev7, it is clear that the TLS machinery makes multiple contacts with PCNA. A cell-free system that can support PCNA- and Pol ζ-dependent TLS is needed to definitively establish the role of TFII-I in this process and dissect the relative contributions of the multiple PCNA-binding mechanisms. Finally, there are no known TFII-I orthologs in the budding yeast. It is possible that yeast Pol ζ uses distinct mechanisms to interact with PCNA.
We were unable to directly test whether TFII-I is required for Pol ζ recruitment to DNA damage sites, as we could not detect the enrichment of either endogenous Rev7 at UV-induced nuclear foci using immunofluorescence or the recruitment of GFP-Rev7 to laser-induced DNA damage sites. The underlying reason for the lack of Rev7 enrichment at DNA damage sites is unclear, but could be due to the transient nature of the TFII-I/Rev1-bridged interactions between PCNA and Pol ζ. Alternatively, the Rev1–Rev7 and TFII-I–Rev7 interactions are required, but are not sufficient, to recruit Rev7 to the site of DNA damage. Only the intact, functional Pol ζ (i.e. the Rev3L–Rev7 complex) can be efficiently recruited. Because Rev3L is a low-abundance protein in human cells, recruitment of Pol ζ to DNA damage sites might be below the detection limits of our cytological assays.
Two lines of evidence suggest that the TLS function of TFII-I is independent of its roles in transcription. First, depletion of TFII-I causes only a mild transcription defect in human cells. Of the few genes whose expression was down-regulated by TFII-I depletion, none had known roles in TLS. Second, the PCNA-binding APIM motifs of TFII-I are critical for TLS. These motifs do not have expected roles in transcription.
Williams-Beuren syndrome (WBS) is a rare genetic disorder caused by deletion of one copy of the chromosome 7q11.23 region, which contains TFII-I and about 25 other genes [22]. WBS patients have multiple symptoms, including distinctive craniofacial features, mild mental retardation, and cardiovascular defects. Different phenotypes have been linked to different genes in the 7q11.23 region. Mice with a heterozygous deletion of N-terminal 140 residues of TFII-I exhibit craniofacial and neurobehavioral alterations [23], implicating this region of TFII-I in WBS pathophysiology. In this study, we showed that the N-terminal region of TFII-I is critical for TLS, raising the intriguing possibility that defective TLS might underlie a subset of symptoms in WBS. It will be interesting to test whether cells derived from WBS patients exhibit sensitivity to UV irradiation and are defective in TLS, and may have defects in components of this TLS complex (Figure 7D).
In addition to TLS, Pol ζ is involved in somatic hypermutation, DNA interstrand crosslink repair, and DSB repair through homologous recombination [35], [45]. Future experiments are needed to test whether TFII-I also contributes to the functions of Pol ζ in these processes. Furthermore, inactivation of Pol ζ sensitizes human cancer cells to killing by the chemotherapeutic drug, cisplatin [32]. Chemical compounds targeting Pol ζ may enhance the efficacy of cisplatin. Our discovery of TFII-I as a novel Pol ζ regulatory factor presents new opportunities for the chemical inhibition of this important polymerase complex. Finally, the general transcription factor TFIIH has a well-established role in nucleotide excision repair [1]. Our findings linking TFII-I to TLS further strengthen the general principle of component sharing in diverse nuclear processes.
HeLa Tet-On, 293T, and U2OS cells were grown in Dulbecco's modified Eagle's medium (DMEM; Invitrogen) supplemented with 10% fetal bovine serum (FBS). Plasmid and siRNA transfections were performed with the Effectene reagent (Qiagen), Lipofectamine 2000, and Lipofectamine RNAiMAX (Invitrogen) for 48 hrs in the indicated cell lines before the desired analysis unless otherwise noted. To establish the TAP-Rev7 cell line, 293T cells were transfected with the pIRES-Puro-TAP (Clontech) or pIRES-Puro-TAP-Rev7 vectors, and selected with 2 µg/ml puromycin. Individual clones were isolated for further analysis. The following siRNAs were chemically synthesized at or purchased from Dharmacon: siControl (5′-GACCGUUAGGUACAGAAGAUU-3′), siLuc, (5′-UCAUUCCGGAUACUGCGAU-3′), siRev7-1 (5′-CGGACAUUUUAAAGAUGCA-3), siRev7-2 (5′UGCAUCUUUAAAAUGUCCG-3′), and siGENOME Smartpools against human Rev3L and TFII-I. For UV-C (254 nm) treatment, the growth medium was removed from the cells and reserved. Cells were washed twice with PBS. The plates (without PBS) were transferred to a UV cross-linker (Stratagene) and irradiated with the indicated UV doses. The UV-C dose delivered to the cells was confirmed with a UV radiometer (UVP, Inc.). The reserved medium was added back to cells. The cells were returned to the incubator.
For tandem affinity purification of TAP-Rev7, ten 150-mm dishes of 293T cells stably expressing TAP-Rev7 were harvested in the TAP lysis buffer (50 mM HEPES pH 7.5, 100 mM KCl, 2 mM EDTA, 10% glycerol, 0.1% NP-40, 10 mM NaF, 0.25 mM Na3VO4, 50 mM β-glycerolphosphate, 2 mM DTT, and 1X protease inhibitor cocktail). Cleared lysates were bound to IgG-Sepharose beads (GE Amersham) for 4 hrs at 4°C. Beads were subsequently washed three times with the lysis buffer and once with the TEV buffer (10 mM HEPES pH 8.0, 150 mM NaCl, 0.1% NP-40, 0.5 mM EDTA, 1 mM DTT, and 1X protease inhibitor cocktail). Protein complexes were cleaved off the beads by 70 µg TEV protease in TEV buffer overnight at 4°C. Supernatants were diluted in calmodulin-binding buffer (10 mM HEPES pH 8.0, 150 mM NaCl, 1 mM magnesium acetate, 1 mM imidazole, 0.1% NP-40, 6 mM CaCl2, 10 mM 2- mercaptoethanol) and incubated with calmodulin-sepharose beads (GE Amersham) for 90 minutes at 4°C. Captured protein complexes were washed three times with the calmodulin-binding buffer and the calmodulin rinse buffer (50 mM NH4HCO3 pH 8.0, 75 mM NaCl, 1 mM magnesium acetate, 1 mM imidazole, 2 mM CaCl2). Proteins were eluted in SDS sample buffer, boiled for 10 min, concentrated in microcon concentrators (Millipore), and subjected to SDS-PAGE. Gels were stained with colloidal Coomassie blue stain (Pierce) according to manufacturer's protocols. Unique bands were excised and in-gel proteolysis was performed using modified porcine trypsin digestion overnight. The resulting peptide mixture was dissolved and subjected to nano-LC/MS/MS analysis on a ThermoFinnigan LTQ instrument, coupled with an Agilent 1100 Series HPLC system. Peptide sequences were identified using the Mascot search engine (Matrix science). Those proteins identified in the TAP-REV7 purification with multiple peptides and not identified in the TAP-vector control pull-downs were considered hits.
The antibodies used in this study are: α-Myc (Roche), α-Rev7 (BD Transduction), α-TFII-I (Bethyl, A301-330A), α-Pol ι (Bethyl, A301-303A), α-Pol η (Abcam, ab17725), α-tubulin (Sigma), α-γH2AX (Millipore, 05-636), and α-PCNA (Santa Cruz, PC10). For immunoblotting and immunofluorescence, the antibodies were used at a final concentration of 1 µg/ml.
For immunoblotting, cells were lysed in SDS sample buffer, sonicated, boiled, separated by SDS–PAGE, and blotted with the indicated antibodies. Horseradish peroxidase-conjugated goat anti-rabbit or anti-mouse IgG (Amersham Biosciences) was used as the secondary antibodies. Immunoblots were developed using the ECL reagent (Amersham Biosciences) according to the manufacturer's protocols and exposed to film.
For immunoprecipitation, cells were lysed with the lysis buffer (50 mM Tris-HCl, pH 8.0, 250 mM NaCl, 5 mM MgCl2, 5 mM EDTA, 0.5% Triton X-100, 10 mM NaF, 80 mM β-glycerophosphate, 10% glycerol, 1 mM DTT, and protease inhibitor cocktail). The lysates were cleared by centrifugation for 30 min at 4°C at top speed in a microcentrifuge. Control IgG (Sigma) or α-TFII-I antibodies were covalently coupled to Affi-Prep protein A beads (Bio-Rad). The supernatants were incubated the antibody-coupled beads. The beads were washed with the lysis buffer. Proteins bound to the beads were dissolved in SDS sample buffer, boiled, separated by SDS-PAGE, and blotted with α-Rev7 and α-TFII-I antibodies.
For the immunoprecipitation of the PCNA complex, U2OS cells were fixed in PBS containing 0.25% formaldehyde for 10 min at room temperature, and the reaction was stopped by the addition of glycine to a final concentration of 0.125 M. After being washed twice with PBS, cells were resuspended in Lysis Buffer 1 (10 mM HEPES pH 6.5, 10 mM EDTA, 0.5 mM EGTA, 0.25% Triton X-100, 1X protease inhibitor) and kept on ice for 10 min. Following centrifugation at 1700 g for 10 min at 4°C, pellets were washed with Lysis Buffer 2 (10 mM HEPES pH 6.5, 200 mM NaCl, 10 mM EDTA, 0.5 mM EGTA, 1X protease inhibitor) and again pelleted at 1700 g for 5 min at 4°C. Pellets were then resuspended in Lysis Buffer 3 (25 mM Tris-HCl pH 6.5, 100 mM NaCl, 2 mM MgCl2, 1X protease inhibitor, 10% glycerol, 1 mM DTT, 10 mM BGP, 5 mM NaF, 3 mM NaVO4, Turbo nuclease), incubated on ice for 10 min, and sonicated. Lysates were then centrifuged at 14,000 rpm for 15 min at 4°C. The supernatant was incubated with Affi-Prep Protein A beads coupled to α-PCNA for 3 h at 4°C. Beads were washed five times with Lysis Buffer 3. Protein crosslinks were reversed by incubating the beads in SDS buffer at 95°C for 30 min. Proteins bound to beads were analyzed by SDS-PAGE and immunoblotting.
For immunofluorescence, HeLa Tet-On cells transfected with the indicated siRNAs were plated in four-well chamber slides (LabTek), treated with 10 J/m2 UV or left untreated, and fixed with 4% paraformaldehyde in 250 mM HEPES pH 7.4, 0.1% Triton X-100 at 4°C for 20 min. After 3–5 washes over 20 min in PBS, cells were permeabilized in PBS containing 0.5% Triton X-100 for 20 min, and then washed with PBS. The cells were blocked in PBS containing 5% BSA followed by a 2-h incubation with the primary antibodies. After 3–5 washes over 20 min with PBS, cells were incubated with fluorescent secondary antibodies (Alexa Fluor 488 or 647, Molecular Probes) for 30 min at room temperature. After incubation, cells were washed with PBS, and their nuclei were stained with DAPI (1 µg/ml). Slides were mounted and viewed with a 100X objective on a DeltaVision microscope. All images were taken at 0.2 µm intervals, deconvolved, and stacked. The images were further processed in ImageJ.
For GST pulldown assays, Myc-TFII-I or its fragments were in vitro translated in rabbit reticulocyte lysate in the presence of 35S-methionine and incubated with bacterially expressed GST or GST-Rev7 in the binding buffer (25 mM Tris-HCl pH 8.0, 2.7 mM KCl, 137 mM NaCl, 0.05% Tween-20) for 1 h at room temperature. Protein complexes were then bound to Glutathione-Sepharose beads for 30 min at room temperature. After 5 washes with the binding buffer, the proteins were eluted with SDS sample buffer, boiled, and subjected to SDS-PAGE. The bound proteins were analyzed with a phosphor imager (Fujifilm) to visualize 35S-labeled TFII-I and Coomassie staining to visualize GST and GST-Rev7.
Human His6-Rev7 R124A mutant bound to the Rev7-binding region of human Rev3L (residues 1847–1898) and untagged human PCNA were prepared as previously described [46], [47]. (Rev7 forms dimers in vitro, but in vivo function of this dimerization event is unclear. The R124A mutation disrupts Rev7 dimerization.) Human TFII-I fragments and the C-terminal domain (CTD) of Rev1 (residues 1140–1251) were expressed as GST-fusion proteins in bacteria and purified with the glutathione-Sepharose 4B resin (GE Healthcare). The eluted proteins were digested with the PreScission protease (GE Healthcare), and further purified with anion exchange and size exclusion chromatography. To assay for complex formation and to determine the apparent molecular weight of the complexes, the gel filtration standard (Bio-Rad, 151-1901), Rev7–Rev3L–TFII-I, PCNA–TFII-I, Rev7–Rev3L–Rev1–TFII-I, and Rev7–Rev3L–TFII-I in the presence of PCNA were fractionated on a Superdex 200 10/300 GL column (GE Healthcare).
293T cells were transfected twice with the indicated siRNAs in a 24 h period, and replated into six-well plates at 60 h after the first siRNA treatment, with 500, 2000, 10,000, and 40,000 cells per well. After another 24 h, cells were exposed to varying doses of UV (0, 4, 8, 12, and 16 J/m2). Twelve days later, colonies were fixed and stained in a solution containing 3∶1 methanol and glacial acetic acid plus 1% trypan blue (Sigma). Colonies containing 50 or more cells were counted. The surviving fractions for each group represent the plating efficiency for each treatment divided by the plating efficiency of the corresponding untreated control samples.
293T cells were transfected with the appropriate siRNAs. At 24 h after siRNA transfection, pSP189 plasmids were irradiated with UV (1000 J/m2) and transfected into the cells using Lipofectamine 2000 (Invitrogen). Cells were harvested 48 h later for plasmid purification using the DNA miniprep kit (QIAGEN). The purified plasmids were digested with DpnI and transformed into the bacterial strain MBM7070 by electroporation. Bacterial cells with wild-type SupF tRNA expressed functional β-galactosidase and formed blue colonies on X-gal plates, whereas bacteria with mutated SupF formed white colonies. The mutation frequency in the SupF gene was analyzed by counting the ratio between blue (wild-type) and white (mutant) colonies. Mutations in the SupF gene were confirmed by DNA sequencing.
HeLa Tet-On cells transfected with the appropriate siRNAs were irradiated with 10 J/m2 UV. Samples were taken at the indicated timepoints, fixed with 70% ice-cold ethanol, blocked with PBS containing 5% BSA and 0.25% Triton-X100, and stained with the anti-γ-H2AX monoclonal antibody. Cells were washed, incubated with the Alexa Fluor 488 donkey anti-mouse secondary antibody (Invitrogen), and counterstained with propidium iodide in PBS containing RNase A. Cells were analyzed with a BD FACSCalibur flow cytometer by using the CellQuest software. Data were processed with FlowJo (FloJo, Ashland, OR).
Total RNA was harvested from untreated and siRNA-treated HeLa Tet-On cells at 48 h after siRNA transfection using the RNeasy Kit (Qiagen). cDNA was synthesized from the total RNA, purified, and hybridized to a HumanHT-12 v4 BeadChip array at the UTSW Microarray Core facility. The arrays were then washed, stained, and scanned according to the manufacturer's protocol. For quantitative PCR (qPCR), cells were lysed in Trizol (Invitrogen). Total RNA was extracted by chloroform extraction and isopropanol precipitation. About 1–2 µg of total RNA was reverse transcribed with the high-capacity cDNA reverse transcription kit (Applied Biosystems) according to the manufacturer's instructions. Taqman probes for human Rev3L (Hs01076848_m1), Rev1 (Hs01019771_m1), Rev7 (Hs01057448_m1), and Rad18 (Hs00892551_m1), and GAPDH (Applied Biosystems) were used for qPCR in TaqMan master mix (Applied Biosystems) according to the manufacturer's protocol. Samples were run in triplicates with the appropriate negative controls.
U2OS cells were transfected with DsRed-PCNA and GFP-TFII-I or GFP-Rev1 along with the appropriate siRNAs. DSBs were introduced in the nuclei of cultured cells by microirradiation with a pulsed nitrogen laser (Spectra-Physics; 365 nm, 10 Hz pulse) [48]. The laser system was directly coupled (Micropoint Ablation Laser System; Photonic Instruments, Inc.) to the epifluorescence path of an Axiovert 200 M microscope (Carl Zeiss MicroImaging, Inc.) for immunostaining imaging or time-lapse imaging and focused through a Plan-Apochromat 63×/NA 1.40 oil immersion objective (Carl Zeiss MicroImaging, Inc.). The output of the laser power was set at 60% of the maximum. Time-lapse images were taken with an AxioCam HRm (Carl Zeiss MicroImaging, Inc.). During microirradiation, imaging, or analysis, the cells were maintained at 37°C in 35-mm glass-bottom culture dishes (MatTek Cultureware). The growth medium was replaced by CO2-independent medium (Invitrogen) before analysis. The images were further processed by ImageJ and Photoshop.
|
10.1371/journal.pgen.1004982 | Genomic Selection and Association Mapping in Rice (Oryza sativa): Effect of Trait Genetic Architecture, Training Population Composition, Marker Number and Statistical Model on Accuracy of Rice Genomic Selection in Elite, Tropical Rice Breeding Lines | Genomic Selection (GS) is a new breeding method in which genome-wide markers are used to predict the breeding value of individuals in a breeding population. GS has been shown to improve breeding efficiency in dairy cattle and several crop plant species, and here we evaluate for the first time its efficacy for breeding inbred lines of rice. We performed a genome-wide association study (GWAS) in conjunction with five-fold GS cross-validation on a population of 363 elite breeding lines from the International Rice Research Institute's (IRRI) irrigated rice breeding program and herein report the GS results. The population was genotyped with 73,147 markers using genotyping-by-sequencing. The training population, statistical method used to build the GS model, number of markers, and trait were varied to determine their effect on prediction accuracy. For all three traits, genomic prediction models outperformed prediction based on pedigree records alone. Prediction accuracies ranged from 0.31 and 0.34 for grain yield and plant height to 0.63 for flowering time. Analyses using subsets of the full marker set suggest that using one marker every 0.2 cM is sufficient for genomic selection in this collection of rice breeding materials. RR-BLUP was the best performing statistical method for grain yield where no large effect QTL were detected by GWAS, while for flowering time, where a single very large effect QTL was detected, the non-GS multiple linear regression method outperformed GS models. For plant height, in which four mid-sized QTL were identified by GWAS, random forest produced the most consistently accurate GS models. Our results suggest that GS, informed by GWAS interpretations of genetic architecture and population structure, could become an effective tool for increasing the efficiency of rice breeding as the costs of genotyping continue to decline.
| Genomic selection is a promising breeding technique that aims to improve the efficiency and speed of the breeding process. While it has been shown to be effective in crops such as wheat and corn, it has not yet been applied to rice breeding. Genome-wide association studies (GWAS), by contrast, are used to identify genes or QTLs that underlie traits of importance to breeding such as yield, flowering time, or plant height, and has been performed successfully in rice. Here, we experiment with applying genomic selection in conjunction with GWAS to a rice breeding program at the International Rice Research Institute in the Philippines and show that genomic selection can result in more accurate predictions of breeding line performance than pedigree data alone and that GWAS results can inform the results of GS. Our results suggest that GS could be an effective tool for increasing the efficiency of rice breeding.
| Over the next 30 years, the production of staple cereal grains including wheat, maize, and rice must to be doubled to keep pace with global population and income growth. At the same time, agriculture, in general, is imperiled by human-induced climate change, and plant breeders and farmers together must contend with increased biotic and abiotic stresses that are the direct result of climate unpredictability. Breeding rice varieties adapted to the Asian tropics is already a challenging and resource-intesive endeavor. The number of bacterial, fungal, viral and insect pests for tropical irrigated rice outnumber those for other major cereals. For non-irrigated rice, abiotic stresses such as flooding and drought also negatively affect production [1,2,3]. Rice breeders must therefore consider a large number of simple and quantitative traits in combination when developing new lines while, at the same time, maintaining and improving quality and ensuring yield improvements over existing varieties. Using coventional breeding methods, this process is extremely time consuming—on average, it takes up to ten years for elite varieties to be developed and identified [4].
The majority of public sector rice breeding programs in Asia still use conventional breeding schemes. By far, the most common way of breeding is the pedigree method, which involves visual selection and trait screening over several successive generations [2]. With advances in rice molecular genetics and genomics, however, other potentially faster breeding methods are being developed. Marker assisted selection (MAS), in which a small number of molelcular markers are used to tag genes-of-interest, has been implemented for rice improvement, but its overall impact on enhancing the efficiency of breeding has been limited [5]. MAS has been successfully used in rice to incorporate major genes and/or large-effect quantitative trait loci (QTLs) controlling abiotic stresses such as submergence, salinity and drought into new varieties [6]. However, most traits of interest to rice breeders are not controlled by just a few large-effect genes, but by many genes of small effect and/or by a combination of major and minor genes. MAS is far less suitable for these types of trait genetic architectures, so its utility to rice breeders is limited. Epistatic interactions and the effects of genetic background in rice furthermore make molecular breeding even more complicated.
Genomic selection (GS), introduced in 2001 by Meuwissen and colleagues, presents a new alternative to traditional MAS that has enormous potential to actually improve gain per selection in a breeding program per unit time, and thus breeding efficiency. In a GS breeding schema, genome-wide DNA markers are used to predict which individuals in a breeding population are most valuable as parents of the next generation of offspring [7]. These estimated values, termed the genome estimated breeding values (GEBVs), are output from a model of the relationship between the genome-wide markers and phenotypes of the individuals undergoing selection. The GEBVs are then used to select the best parents for making new crosses. The GS model itself is developed from a training population that resembles the population under selection (also referred to as the testing population); it is both genotyped and phenotyped, while the testing population is genotyped only. The testing population genotypes are then entered into the model to calculate the GEBVs of all the individuals in the population, even those that have not been phenotyped. Thus, the key difference between GS and traditional MAS is that genotyping is not limited to a selected set of markers that tag putative genes, but rather breeding value is predicted based on all available marker data to avoid ascertainment bias and information loss. Including all markers in the model regardless of effect size also makes it possible for the first time to track and select for small effect genes/QTL in addition to large effect genes/QTL. Statistical shrinkage, Bayesian, and/or machine learning methods are used to fit the many thousands of effects [7,8].
The advantage of GS over the widely-used traditional pedigree breeding method is thus one of breeding efficiency. Gain from selection during GS is proportional to GEBV accuracy. As a result, when GEBV accuracy is high enough, GS can reduce breeding time by increasing the proportion of high-performing offspring in a breeding population, thus accelerating gain from selection [9,10]. GS has been most successfully implemented in dairy cattle breeding, where its efficiency is proven: the replacement of progeny testing with the genotyping of young bulls has cut generation interval time in half [11]. Genetic heterogeneity is also low for Holstein-Friesian cattle, and GxG and GxE effects are limited, which makes prediction of breeding value simpler [5]. In plant breeding, these interactions present a challenge, as does the presence of structure within and between breeding populations, but GS still holds the potential to improve breeding efficiency. In temperate crops GS can accelerate gain from selection per unit time beyond that gained by the overall population improvement described above through the use of off-season nurseries [12,13], while in tropical crops like rice, GS can be used with one or more cycles of rapid generation advance [14] for a similar gain.
In plants most applied GS experiments to date have been in maize and small grains, and it is quickly generating interest as a breeding tool for those crops. High GEBV accuracies for grain yield and a variety of other quantitative traits have been obtained for both maize and wheat bi-parental and double haploid populations using experimental cross-validation [15,16], and GS has been demonstrated to outperform marker assisted recurrent selection (MARS) in at least one maize breeding program [17]. Moderate cross-validation prediction accuracies have also been obtained for yield and a variety of other traits in diverse germplasm collections and breeding populations of maize, wheat, oat, and barley [18,19,20,21,22,23]. Preliminary genomic selection research has also been published on several other crop plants including cassava, sugarcane, and sugar beet [24,25,26,27].
Several recent studies in maize, however, advise caution regarding the presence of hidden or known structure or family relatedness within a breeding population or germplasm collection when estimating GS accuracy. Windhausen et al (2012) found that within a diversity panel of 255 maize lines from eight distinct breeding populations any predictive ability in the dataset was a byproduct of the population structure, while Riedelsheimer et al (2013) found mean accuracies of 0% when trying to predict individuals in biparental families using data trained on the progeny of an unrelated cross [28,29]. To accurately predict the phenotypes of individuals in their biparental crosses, Riedelsheimer et al found it necessary to train the model using full sibs of the validation individuals, or half sibs representing both parents of the cross [28].
These limitations will likely also apply to rice, which is subject to deep population structure and is often bred in large, inter-related pedigree schemas. Rice is also frequently admixed, and many varieties contain introgressions from different subpopulations [30,31,32]. The work of Guo et al., (2014) evidences this need to control for subpopulation structure when performing GS in rice. In a rice diversity panel, Guo et al. (2014) found that ∼33% of the genomic heritability was explained by subpopulation structure, and that controlling for subpopulation structure when performing cross-validation significantly decreased prediction accuracy. When prediction was performed within a specific subpopulation, however, structure was found to have little effect on prediction accuracy [33]. This is fortunate for breeding programs, which generally work within a particular subpopulation, although introgressions are frequent.
Genetic architecture must also be taken into account when considering the implementation of genomic selection. GWAS results in maize have consistently found most agronomic traits to be controlled by many genes of small effect [34,35,36,37]. In rice, by contrast, many GWAS and QTL mapping studies have found large effect QTLs for agronomic traits, including grain yield, flowering time, plant height, aluminum tolerance, grain yield under drought stress, and submergence tolerance [38,39,40,41,42,43,44]. The difference in the genetic architecture between maize and rice, as well as the difference in the genetic architecture of different rice traits, could be expected to affect the relative efficacy of different genomic selection statistical methods.
To the best of our knowledge, no research on performing GS in a rice breeding population has yet been published. Here we report the results of performing GWAS and GS cross-validation using data on a collection of 363 elite breeding lines from the International Rice Research Institute's (IRRI's) irrigated rice breeding program. To assess GS accuracy, we performed five-fold cross validation to predict grain yield, flowering time, and plant height in the 2012 wet and dry season in Los Baños, Philippines and compared our prediction results using GS to those using only pedigree information as well as a traditional MAS model. For the GS models, the training population composition, marker number, and the statistical method for the calculation of GEBVs were varied to determine their effect on rice GS accuracy. Finally the GWAS results published in a companion paper allowed us to analyze the effect of genetic architecture on GS prediction accuracy [45].
384-plex Genotyping-by-sequencing (GBS) was used to discover and call SNPs on 369 advanced inbred breeding lines from IRRI's irrigated rice breeding program (methods). SNP calling was performed using the TASSEL3.0 GBS pipeline with physical alignment to the MSU v6.0 Nipponbare rice reference genome using Bowtie2 [46,47]. The resulting SNP data were imputed using the TASSEL 3.0 FastImputationBitFixedWindow plugin [48]. After imputation, SNPs with call rates < 90% were removed along with monomorphic markers to obtain a filtered SNP dataset containing 73,147 SNPs. Individuals with missing data > = 60%, a total of six individuals, were dropped for a total of 363 genotyped lines (materials and methods).
The majority of the 363 lines were known a priori from breeding records to belong to the indica or indica-admixed subpopulation groups. In order to identify outlier individuals belonging to the japonica or japonica-admixed groups, principle components analysis (PCA) was performed using the 73,147 SNPs. Thirty one outliers were identified and excluded based on this analysis (S1 Fig.). After removing these 31 outliers, the resulting PCA suggested no remaining subpopulation stratification within the dataset. Family structure, however, was presumed to still exist. As the presence of close relatives (e.g. full sibs) across training and testing folds in a cross-validation experiment can artificially inflate prediction accuracies, it was necessary to also control for this family structure. To do so, the remaining 332 lines were clustered using a partitioning around k-medoids algorithm (PAMK) based on the genotype matrix. k = 87 was found to be the most statistically favorable number of clusters in the dataset based on average silhouette width (S2 Fig.). Individuals in the same cluster (of 87) were then assigned to the same fold of 5 to form the five folds used for cross-validation. The most closely related individuals were thus placed within the same fold, making it impossible for them to be spread across training and testing groups [26] (materials and methods).
Four years of grain yield (kg/ha), flowering time (days to 50% flowering), and plant height (cm) data and related phenotypic covariates were curated from historical breeding trial records taken at a single location in Los Baños, Philippines for years 2009–2012, two seasons per year, dry and wet, with the exception of plant height in the 2009 wet season, which was not available (materials and methods). As the genotyped lines represent a subset of a working breeding program, substantial missing data are present in years 2009–2010 for all traits (S1 Table). Such an unbalanced design is typical of breeding trial data and to be expected in the practical implementation of GS. Correlations among years/seasons were calculated for all three traits using the trait least squares means. For grain yield, the 2011 and 2012 data were more tightly correlated than the earlier year data. Flowering time and plant height data was well correlated for all four years and seasons (S3 Fig.) (see methods).
Narrow-sense heritabilities were calculated on a per line basis for each trait for both validation seasons—the 2012 dry season (2012 DS) and the 2012 wet season (2012 WS) and ranged from 0.31–0.32 for grain yield, 0.30–0.35 for plant height, and 0.32–0.44 for flowering time (Table 1) (materials and methods). Heritabilities for all three traits were slightly higher in the dry season than the wet season.
Five-fold cross validation was performed using the full set of 73,147 markers to predict grain yield, flowering time, and plant height in the 2012 dry and wet seasons. The year and season data included in the training population were varied to determine which combinations of years/seasons were the most predictive of the 2012 dry and wet season (total of twelve different combinations). The GEBV accuracies were calculated as the correlation of predicted GEBV and observed phenotypes in the validation population.
Six statistical methods widely demonstrated to produce accurate genomics-assisted breeding models in a variety of crops were selected from the literature to test using our rice data. The selected methodologies were chosen to represent the variety of available approaches, and included one linear, parametric, and frequentist method: rrBLUP, one linear, parametric, and Bayesian method: Bayesian LASSO (BL), one non-linear semi-parametric method: Reproducing Kernel Hilbert Spaces (RKHS), and one non-linear machine learning method: Random Forest (RF) [19,23,49,50]. Multiple Linear Regression (MLR), in which a subset of significant markers are chosen to fit a linear model, has been shown to be effective for traits with a very simple genetic architecture, and served as our non-GS method control [51]. Finally, kinship BLUP was used to predict GEBV based on the pedigree A-matrix alone (ped) (methods) [52].
We estimated accuracies using three experiment types (CV1, CV2, and CV3). CV1 accuracies were calculated using training populations that included data from the validation year/season, i.e, if the validation population consisted of the 2012 dry season, then data on individuals from the 2012 dry season were included in the training population, excluding data on any individuals in the validation fold. However, this is likely to upwardly bias accuracy estimates by confounding GxE and line effects [21], so we worked to obtain an estimate of this bias by performing two other types of experiments. For CV2 accuracies we excluded the validation year/season from the training population. By removing these data from the training population, however, we introduce a different confounding factor to our accuracy estimate—a smaller training population size. We therefore performed cross validation experiment 3 (CV3) in which the data from the validation year/season were retained in the training population, but the equivalent data from the respective 2011 season were not included in the training population. The overall estimate of bias for a given permutation was subsequently estimated as accuracy of CV3—accuracy of CV2 [26] (materials and methods). The bias estimates were found to be very small and consistent for all tested traits and permutations (Table 2, S2–S4 Tables). It can thus be concluded that for the population and statistical methods tested here bias as a result of including data from the validation year/season in the training population is not a significant concern.
Grain yield. The highest prediction accuracies for grain yield in both the 2012 dry and wet seasons were 0.31, when the training populations consisted of data from all four years (2009–2012), both seasons per year. The peak dry season accuracy was obtained when rrBLUP was used to build the model, and the peak wet season accuracy was obtained when RF was used (Table 2, S2 Table). In general, however, prediction accuracies did not significantly vary depending on the combination of years or seasons in the training population (α = 0.05). These results indicate that the most recent and complete years (2011, 2012) are also the most predictive, but that adding data from earlier years to the training population and utilizing both seasons of data (as opposed to using only the dry season to predict the dry season or only the wet seasons to predict the wet season) can marginally increase accuracy (Table 2, S2 Table). These results make sense given the strong correlations between the wet and dry seasons within the same year and the weak correlations between the earlier and later years for grain yield (S3 Fig.). The lower relative importance of the earlier year data could also be due to the large proportion of missing data in the earlier years.
The statistical method used to build the prediction model had a significant effect on accuracy. RR-BLUP, Random Forest, and RKHS all performed significantly better than pedigree alone. RR-BLUP and RF, specifically, outperformed pedigree prediction by an average of ∼8%. Similar results have been documented in CIMMYT wheat populations where genetic markers have been found to add 7.7%-35.7% to the accuracy of grain yield predictions over a pedigree-only model depending on the population and environment [52]. The modest gains in accuracy of using markers to predict breeding value in our rice population suggest that larger training populations may be necessary to better model the effects of Mendelian segregation on yield, in addition to effects due to family relationships [53].
Some of the marker models performed worse than pedigree prediction. Bayesian LASSO performed significantly worse than prediction based on pedigree alone, while MLR performed worst of all. It is worth noting that the GWAS for grain yield in this population (unlike the GWAS for flowering time or plant height) did not identify any large effect QTL [45], which could explain why choosing a subset of markers to predict GEBV performed so poorly relative to the genomic selection methods (Table 2, S2 Table).
Flowering time. The prediction accuracies for flowering time were higher than those for grain yield at 0.63 and 0.54 for the best performing experiments in the 2012 dry and wet seasons, respectively. For the dry season, the most predictive training population was composed of the 2009–2011 data, dry seasons only, while for the wet season, the best training population included all seasons from 2010–2011. The prediction accuracies for flowering time in the 2012 dry season were significantly higher than those for the 2012 wet season across statistical methods and experiments (p < 0.0001), but the differences in the performance of different training populations were not significant within a given validation population (Table 2, S3 Table).
Unlike for grain yield, the best accuracies for predicting flowering time for both seasons were obtained using MLR. In fact, MLR significantly outperformed all other statistical methods and was more accurate than pedigree alone by 22% and 33% for the dry and wet seasons, respectively (Table 2, S3 Table). The higher accuracies for prediction of flowering time relative to predictions for yield, and also of the dry season predictions over the wet season predictions, can be explained by the higher trait heritabilities for flowering time of the 2012 dry season relative to the 2012 wet season (Table 1), and by the strong correlation in the phenotype data for all years and seasons (S3 Fig.). The outstanding performance of MLR, on the other hand, is best explained by the genetic architecture of flowering time. Multiple large effect QTL have been cloned for flowering time [43,44], and the GWAS performed on this population identified a single very large effect QTL on chromosome 3 that explained more than 40% of the variation in flowering time [45]. These results are also consistent with results for prediction of heading date using MLR versus GS in wheat [51]. Of the genomic selection methods tested (MLR is a non-GS method), random forest performed the best by a significant margin, and was the next best method of predicting flowering time after MLR. This is worth noting as the random forest algorithm is also effective at capturing large-effect QTL [54].
Overall, these results suggest that the presence of large effect QTL for specific traits in rice could improve the prediction accuracy of those traits, although it remains to be seen whether genomic selection models will be the most practical means of obtaining those predictions. One promising avenue of research would be to model the large effect QTL as fixed effects using a genomic selection method such as rrBLUP [55].
Plant height. The accuracies for plant height were similar to those for grain yield, 0.34 for the dry season when the 2009–2011 dry seasons served as the training population, and 0.32 for the wet season when all seasons and years served as the training population (Table 2, S4 Table). These results further suggest that heritability has an important effect on accuracy. Both grain yield and plant height had similar heritabilities, and similar prediction accuracies (Tables 1, 2, S4 Table.)
For predicting plant height, however, MLR was sometimes the best-performing statistical method, as was the case for the most accurate wet season experiment, described above, but for other experiments, MLR was the worst-performing method, as for the best performing dry season experiment, described above. Due to the inconsistent performance of MLR, the prediction method with the best performance over all experiments was random forest (Table 2, S4 Table). Across all experiments, random forest outperformed pedigree prediction by an average of 13.3%, an improvement in the performance of marker based prediction relative to pedigree prediction that is squarely in between the improvements seen for grain yield and plant height (Table 2, S4 Table).
These results suggest that large marker effects help to make up the genetic architecture for plant height, but that plant height genetic architecture is more complicated than the genetic architecture of flowering time. This inference is borne out by the GWAS results for plant height—four large effect QTL were identified, explaining ∼74% of the total variation [45]. While these effects are large, they are not as dramatic as the one super-QTL found for flowering time on chromosome three, which may explain the difference in the performance of MLR for the two traits. As for flowering time, future research in predicting plant height could explore fitting these QTL in linear models as fixed effects.
In order to determine the necessary number of markers for performing GS in a rice population of this type, we selected differently sized SNP subsets from the 73,147 SNP set. The subsets were selected in two ways: 1. to be evenly distributed across the genome (see materials and methods for details) or 2. at random. Ten selections were made for each subset size and type (i.e. random versus distributed), and five-fold CV was performed using each selection in combination with all five marker based models (materials and methods). For each trait, cross validation was run for both validation populations, with years 2009–2011, both seasons per year, serving as the training population. (Fig. 1, S4 Fig., S5 Table, S6 Table).
For all three traits and both validation seasons, it is clear from the marker subset results that 73,147 markers is more than is necessary to capture the QTL segregating in this population. For almost all traits, there was no significant difference in the best-performing GS method for a given trait or validation season when 7,142 SNPs (approximately 1 SNP for every 0.2 cM) were used versus when 13,101 SNPs (1 SNP for every 0.1 cM) or the full 73,147 SNPs were used. This was true for the randomly chosen SNPs as well as for the evenly distributed SNPs, however the accuracy variances were higher for the randomly chosen SNPs, so it is our recommendation that SNPs be evenly distributed across the genome when possible (Fig. 1, S4 Fig., S5 Table, S6 Table). Although it is possible that the variation in the call rates and minor allele frequencies of the randomly selected SNPs also contributed to the larger variations in accuracy in the random SNP subsets, it is still thought that the position of the SNPs was the most important contributor to prediction accuracy.
For all three traits and both validation seasons, prediction accuracies dropped significantly faster with decreasing numbers of markers when the markers were chosen at random versus when they were evenly distributed throughout the genome. The drop-off in prediction accuracy when random selections of SNPs were used was particularly acute for flowering time and plant height and is attributable to the presence of large-effect QTL for these traits; As the number of randomly chosen SNPs decreases, the odds of capturing the effect of any one QTL also decreases. The prediction results for grain yield, by contrast, did not differ as dramatically between the randomly and evenly distributed subsets as did those for flowering time and plant height. These results suggest that the genetic architecture for grain yield is more in line with an infinitesimal model, i.e., that there are many small effect QTL throughout the genome, and are in agreement with the grain yield GWAS results [45]. It thus follows that the effect of choosing SNPs at random would not be as detrimental for grain yield as it is for flowering time or plant height when accuracy crucially depends on capturing specific regions that explain a high proportion of the phenotypic variance.
At fewer than 7,142 SNPs, accuracies began to decrease for most traits and statistical methods, although the extent to which accuracies decayed depended on the prediction method used, the trait, and the validation season. For grain yield in the 2012 dry season, for example, there was no significant difference in the performance of rrBLUP at any marker set > = 3076 markers. For random forest, however, there was no significant difference in prediction accuracy all the way down to sets of markers > = 316 (random or distributed). While it would not be advisable to use such a small number of markers, as the smaller the number of markers, the larger the variation in prediction accuracy, these results do suggest that for grain yield, at least, random forest works better with smaller numbers of markers than does rrBLUP. The results for plant height were very similar to those for grain yield. For flowering time, when SNPs were evenly distributed, variances in accuracy were very small, again, most likely as a result of the super-QTL on chromosome three. These very small variances meant that for both MLR and random forest, accuracies were significantly lower for fewer than 7142 SNPs (distributed) or 1553 SNPs (random).
Taken collectively, these results suggest that using ∼1 SNP every 0.2 cM (∼6–7K SNPs), could be ideal for performing genomic selection in inbred rice breeding populations like the one at IRRI. Opportunely, two Infinium 6K SNP fixed arrays have recently been developed for use within specific rice breeding/research programs [56]. Fixed arrays have established advantages in rice, including robust allele calling, cost-effectiveness per data point, and speed of genotyping turn-around [56]. 6–12K fixed arrays could thus prove to be the most affordable and efficient means of genotyping for GS in rice, especially for smaller breeding programs with less genotyping informatics expertise. The best strategy, however, will likely be to have multiple genotyping platforms available and the flexibility to switch between them as needed. Genotyping turn-around time is ultimately key for GS because genotypes must be available in time for selections and the next generation of crossing. It should be noted that depending on the platform, genotyping individuals with more markers than is necessary could be detrimental to breeding progress if it overloads the bioinformatics and computational capacities of a breeding program.
The matrix of genotypes and phenotypes on a breeding population provides the opportunity to perform GWAS in addition to testing any GS models that are available. This paper describes the GS-side of a joint GS-GWAS project on a single rice breeding population, and is the first study to suggest that GWAS on a set of breeding lines might provide information about both the genetic architecture of the traits-of-interest and the population structure of the breeding materials. Specifically, our results on performing GS for grain yield, plant height, and flowering time demonstrate that performing GWAS using the inputs to GS can reveal the presence of large-effect QTL segregating in a breeding population, which can then be modeled accurately using GS.
Our results are promising for the implementation of GS in rice improvement. For all traits tested, GS outperformed prediction based on pedigree alone with the use of a reasonable number of markers (∼7000) suggesting that genomic selection is accessible for moderately-resourced public programs with minimal bioinformatics capacities. For yield, which appears to be controlled by many genes of small effect [45], RR-BLUP was the most computationally efficient of the best performing statistical methods. For plant height and flowering time, however, the highest accuracies were obtained using random forest and/or MLR, which suggests the presence of both large and small effect QTL for these traits, a hypothesis that is also supported by the GWAS results [45].
Currently, the most commonly used methods of rice improvement are pedigree breeding and traditional marker assisted selection, which mainly track large effect QTL. Our results suggest that genomic selection will make it possible for the first time to track, accumulate, and select for small effect QTL using genetic markers in addition to large effect QTL. One promising strategy is to build GS models in which large effect QTL are fit as fixed effects to capture the variance of large-effect QTL along with small effect QTLs located throughout the genome [55]. Future experiments in rice genomic selection should focus on building these models.
While genomic selection has yet to be integrated into applied breeding programs in rice as it has in maize and wheat, it would be feasible to undertake small pilot programs within specific rice breeding programs, especially for irrigated rice where growing environments are generally more uniform. Such pilot programs are needed, in particular, to determine when and how to incorporate genomic selection into existing breeding programs. An example of an irrigated rice breeding pipeline that incorporates genomic selection is presented in Fig. 2. Parents are selected and crossed and the resulting F1 progeny fixed over seven generations with selection of families for heritable traits. Traditionally, selection during pedigree line fixation would be based only on phenotype. Here, we propose incorporating selection based on GEBV at least once during fixation, as resources allow. Early generation GEBV-based selection would help to avoid eliminating families that carry beneficial rare or recessive alleles and would increase the proportion of top performers that are advanced to the observational yield trials (OYT). Late-generation selection based on GEBVs could be used to select fixed lines to advance to the OYT. The top lines advanced to the OYT based on GEBV could be used simultaneously as parents of the next generation of breeding (Fig. 2).
From the OYT, the best performing lines could be identified and advanced to the replicated yield trials (RYT) by a combination of phenotypic and genomic selection. Phenotypic selection by the breeder has the potential to compensate for the fact that the GS model is always a generation or more behind the current breeding population. This means that any favorable new GxG interactions will not be captured by the model and cannot be selected by GEBV alone. In species where the majority of the genetic variation under selection is controlled by many additive, small effect loci, this should not be a problem. However, in rice and other inbreeding crops, the genetic architecture of many important agronomic traits contains important non-additive features and transgressive variation is common [41,43,57,58,59,60]. The selected lines from the RYT are subsequently advanced to the multi-environment trials (MET) where the GEBVs can be used to select parents for the next generation of hybridization. In order to build or update the genomic selection model at any stage of selection, a training set consisting of a fraction of the breeding population (∼300 individuals) representing different families would need to be both phenotyped and genotyped. The rest of the lines would be genotyped only to calculate the GEBVs (Fig. 2).
The above genomic selection models would ideally account for multiple environments and GxE interactions, however current programs such as the one at IRRI and many other national breeding institutes do not make use of multi-environment data until very late stages of the breeding process, after the population has already been reduced to a manageable number of lines. Thus, even GS models that do not account for multiple environments, like those presented here, are of use to plant breeders and have the potential to improve breeding outcomes. The data from the Multi-environment trials on the IRRI breeding lines used in this experiment is currently being accumulated and vetted and will be a subject for future GS research.
In order to fully exploit the benefits of GS, however, new rice breeding schemes will need to be implemented to further reduce the breeding cycle and increase genetic gain. Heffner et al. (2010) proposed a GS scheme for winter wheat using rapid generation advance (RGA) to generate F5 lines, and multi-location field trials to test F5-derived material, which was further used to train the GS model [13]. A similar scheme should also be effective for rice and a modified scheme is currently being implemented at IRRI within the irrigated breeding program.
By using the genotype and phenotype inputs from pilot programs for both GS and GWAS, the accuracy of GS models could be improved while, at the same time, helping to answer basic biological questions about the genes underlying agronomic traits of interest. Ultimately, in order for genomic selection to be of practical use, it must be possible to select lines with combinations of phenotypes that are routinely measured by breeders, such as disease and insect resistance and grain quality. GEBVs could be used to select for traits that are either difficult or expensive to phenotype or are late in development (e.g. panicle or post-harvest traits), while phenotypic selection, such as in the OYT and RYT in Fig. 2, could be used for other important variety parameters. The use of multi-variate GS models or selection indices as GS phenotypes are other potential solutions to this problem, but both require additional research and computational/statistical inputs to implement.
In practice, determination of whether GS can cost-effectively increase genetic gains relative to utilizing pedigree data alone or simply phenotyping more lines requires a careful consideration of the relative cost of phenotyping compared to genotyping plus line development [61]. For GS to provide increased genetic gain in a pedigree breeding program, the prediction approach must either increase accuracy relative to phenotyping or permit a substantial increase in selection intensity. It is possible to increase selection intensity through the use of rapid generation advance, as mentioned above, but the selection intensity increase must be very large because the response of genetic gains to increasing selection is logarithmic rather than linear.
In this study, GEBV accuracy for yield averaged about 0.3 for the most effective prediction methods (Table 2). The corresponding accuracy for phenotypic selection is the square root of heritability, or about 0.55 for evaluation in a single three-replicate trial (Table 1). An accuracy of 0.3 corresponds to a heritability for yield evaluation of 0.1, which is roughly the accuracy achievable by screening for yield in a single unreplicated irrigated rice trial at IRRI (e.g. Bernier et al., 2007). Currently, the cost of phenotyping a single rice plot for yield and genotyping via GBS is roughly equivalent ($20-$30), so there is no clear advantage for GS over simply phenotyping more materials in unreplicated trials. However, genotyping costs are likely to continue to drop, whereas phenotyping costs are generally steady or rising. Furthermore, continued refinement of GS models by incorporating fixed effects and accumulation of high quality data over years and environments is expected to increase GEBV accuracy. As a result, we predict that in the near future, GS will become a cost-effective means of performing line selection in rice.
369 elite breeding lines were selected for genotyping from the International Rice Research Institute (IRRI) irrigated rice breeding program based on the planned inclusion of the lines in the 2011 Multi-Environment Testing Program and presence in the 2011 and 2012 Replicated Yield Trials (RYT) at IRRI (Los Baños). Approximately half of the lines were also included in the 2009–2010 RYTs at IRRI (S1 Table). The other lines were promoted from the observational yield trial (OYT) to the RYT in 2011.
Phenotypes for the replicated yield trials (RYT) were used for all the experiments and curated from the IRRI database for years 2009–2012, including wet and dry seasons each year. All of the RYT breeding lines, of which our selected 369 lines are a subset, were grown in a randomized complete block design with three replicates in the same field location at IRRI every season and year. The following data were curated for each year, with the exception that plant height data was not available for the 2009 wet season:
plant height: the actual measurement in cm from soil surface to tip of tallest panicle (awns excluded)
flowering time: days to when 50% of flowers were visible in whole plot
maturity date: days to when 85% of grains on panicle were mature
number of effective tiller or panicle per plant: count of the number of panicles on each plant
lodging score: percent of plants that lodged
grain yield (kg/ha): grain yield from a representative plot was harvested and weighed, from this sample the grain yield per hectare was calculated from an inner harvested area of the plot excluding border rows
rep: replication number of observation
The plant height, flowering time, and grain yield phenotypes were selected for prediction using the genomic selection models.
DNA extraction. Young leaf tissue was collected from each of the 369 breeding lines from plants grown in Gutterman Greenhouse in Ithaca, NY. DNA was extracted using the Qiagen 96-plex DNeasy kit as per the Qiagen fresh leaf tissue 96-plex protocol (www.qiagen.com/HB/DNeasy96Plant).
Library preparation. 384-plex genotyping-by-sequencing (GBS) libraries were prepared using the protocol by Elshire et al. 2011 [62], as described previously in Spindel and Wright et al 2013 [63].
GBS data analysis. SNPs were discovered and called from the raw 384-plex GBS data using the TASSEL3.0 GBS pipeline with physical alignment to the MSU version 6.0 Nipponbare rice reference genome using Bowtie2, as described in Spindel and Wright et al 2013 [47,63,64] (S5 Fig.). The IRRI breeding materials genotyped here are a collection of multi-parent related and unrelated inbred lines, so the GBS-PLAID algorithm for imputation, which was developed specifically for imputation of biparental rice mapping populations, was not useful [63]. Imputation of missing data was instead performed using the TASSEL3.0 FastImputationBitFixedWindow plugin with default settings [48]. The algorithm works by dividing the entire SNP dataset into small SNP windows, then identifying the most similar inbred line within each window to fill the missing data. The algorithm takes advantage of small IBD regions shared between pairs of inbred lines in the collection; if the window from the closest neighbor has more than 5% difference from the line being imputed, the data point is left as missing [48]. The imputation error rate using this algorithm was estimated for each chromosome in our dataset by masking a fraction of the un-imputed allele calls and comparing the imputed and actual calls. The average imputation error rate across the twelve rice chromosomes was estimated in this way to be less than 1%.
SNPs that still had 10% or more data missing after imputation (or call rates of < 90%) were removed from the dataset along with all monomorphic SNPs, for a total SNP set of 73,147 SNPs. After the SNP filtering described above, individuals with more than 60% missing data were dropped from the dataset, which resulted in the removal of six individuals that failed sequencing for the total of 363 genotyped lines used throughout the study (S5 Fig.).
The final dataset was then transformed from nucleotide genotype coding (i.e., 'A', 'C', 'T', 'G') to numeric coding (1, 0, -1 for class I homozygotes, heterozygotes, and class II homozygotes, respectively) to facilitate statistical analysis. The minimal remaining missing data were filled using the numeric genotype means of each line in order to perform PCA and genomic selection modeling (S5 Fig.).
The majority of the 363 lines were characterized a priori from pedigree records to belong to the indica or indica-admixed subpopulation groups. In order to identify outlier individuals belonging to the japonica or japonica-admixed groups, principle components analysis (PCA) was performed in R (version 3.0.1) using the imputed 73,147 SNPs, with remaining missing data filled using the line means. The first principal component of high density SNP data in rice can separate the indica and japonica subgroups [30], so by plotting the first four principal components using JMP Pro 10, 13 japonica outliers were identified as a tight cluster that was pulled apart from the rest of the 350 lines (S1A Fig.). These 13 lines were removed from the dataset, and a second PCA was performed using the same methodology as the first to identify any admixed outliers, i.e, outlier lines containing greater percentages of japonica derived SNPs. By plotting the first four principal components of the second PCA, another 18 lines were judged on a visual basis to be outliers and removed from the dataset, leaving a total of 332 lines to be used for the cross-validation experiments (S1B Fig.). A third PCA was performed using the remaining 332 to confirm that there were no additional subpopulation outliers.
It was also known from studying the breeding program pedigrees that differing degrees of family relatedness existed within the remaining 332 lines, including half sibs, full sibs, parents and offspring, and unrelated lines. The presence of highly related individuals in the dataset could have the effect of artificially inflating prediction accuracy if the most closely related individuals are randomly assigned to different folds, and one of those folds is then used as training, while the other is used as testing. Or, in other words, the training fold could end up as unusually predictive of the testing fold if, for example, a pair of full sibs is split across training and testing folds. To control for this possibility when designing our folds, we performed a partitioning around k-medoids analysis (pamk) using the R fpc package (function pamk) with the 73,147 SNPs. k values from 2 to 332 were tested to determine the most statistically probable k-value by average silhouette width (S2 Fig.). The largest average silhouette width was found to occur at k = 87 (S2A Fig.). Individuals found within same cluster of 87 were then assigned to the same fold, making it impossible for the most closely related individuals to be split across training and testing folds. Full clusters were assigned to one of five folds randomly, controlling only for cluster size in order to produce three folds of 66 individuals and two folds of 67 individuals. A similar procedure was used by Ly et al., 2013 [26].
For each cross validation experiment, one of the five folds served as the validation fold, and the other four folds served as the training folds. The process was repeated five times so that each fold served once as the validation fold, resulting in predicted GEBV values for all individuals. Accuracy was assessed as the mean Pearson Correlation of the predicted GEBV and observed phenotype in the validation population.
The cross validation experiments shown in Table 3 were performed in order to test all logical combinations of years and seasons in the training and validation populations. Note that a year's wet season was never used to predict the same year's dry season because in Southeast Asia, the dry season arrives first chronologically. We did, however, predict the 2012 wet season both with and without the 2012 dry season present in the training population. We tested scenarios in which both seasons per year were included in the training population as well as scenarios where only the data from the seasons matching the validation population were included in the training data (e.g., using only the wet season data to predict the wet season). We also sought to test scenarios using only more recent year data in the training population (e.g. only 2011, or 2010–2011) and scenarios using more historical year data in the training population (e.g. 2009–2011) (Table 3).
Cross validation experiment 1 (CV1) accuracies were calculated for all experiments with the validation year/season included in the training population, excluding individuals in the validation fold. Including the validation year/season in the training population can bias accuracies upwards by confounding GxE and line effects, however, so in order to obtain an estimate of this bias, we also performed cross validation experiments 2 and 3 (CV2, CV3) for CV permutations 1–5, see above table. For CV2, we excluded the validation year/season from the training population. These results are not directly comparable to those in which the training population contained the validation year/season (CV1), however, because the training population for CV2 is smaller than was used for CV1 and training population size can have an important effect on prediction accuracy. For this reason, we performed CV3, in which we included the validation year/season in the training population, but removed the equivalent seasons from 2011, e.g., for the first cross-validation permutation in the above table, CV2 would not include the 2012 dry season in the training population, and CV3 would include the 2012 dry season but would not include the 2011 dry season. Thus, the estimate of bias can be calculated for a given CV permutation experiment as CV3 accuracy minus the CV2 accuracy [26]. The bias was only estimated for the first five CV permutations because the bias estimates turned out to be small and similar to each other for all five CV permutations.
For all three traits, multiple years, seasons, and replicate yield entries existed along with the previously described covariates for all 332 individuals. In order to build genomic selection models, it was necessary to convert these raw yields into a single, adjusted yield for each individual. Adjusted yields, plant heights, or days to flowering were calculated for each year/season combination by fitting an initial linear model of the observations y, by line ID (GHID) x1, and phenotype covariates described above (e.g. lodging) x2…n for the given Year x Season in JMP. Non-significant covariates as determined by an F-test (α > = 0.05) or covariates that resulted in singularities were removed, and the model re-fit. When all covariates included in the model were statistically significant, the least squares mean yield for each line ID was exported as the adjusted yield. Missing phenotype data were coded as null data for the above analysis, or, in other words, no imputation or numeric filling of phenotypic values was performed.
The least square means for each year and season were also used to calculate a correlation matrix for each trait (S3 Fig.).
For each experiment, adjusted yields were calculated for each of the five training folds separately by fitting a linear model for each training fold as described above with the difference that data from all years and seasons for a particular CV experiment was including in the x matrices for all lines not in the validation fold. Year, season, and a year x season interaction were also included as covariates in the model, and subject to the same significance requirements as the other model covariates.
Six statistical methods were used for each experiment, including four genomic selection methods: RR-BLUP, Bayesian LASSO (BL), Reproducing Kernel Hilbert Spaces (RKHS), and Random Forest (RF), and two non-genomic selection methods: Multiple Linear Regression (MLR) and Pedigree-BLUP (PED). The four genomic selection methods were chosen based on their demonstrated success in accurately predicting GEBV in variety of crops and because they represent the different types of statistical methodologies used to build GS models, i.e., Linear parametric methods (RR-BLUP, BL), non-linear semi-parametric methods (RKHS), non-linear, non-parametric methods (RF), as well as Frequentist methods (RR-BLUP, RKHS), Bayesian methods (BL), and machine learning methods (RF) [19,23,49,50,65,66,67]. For an overview of the methods, see Lorenz et al., 2011[8].
Multiple linear regression using a subset of markers derived from single marker regressions (MLR), another linear, parametric statistical method was the fifth statistical method tested to predict breeding value, and served as our used as a non-GS marker-based prediction control. For each fold, single marker regression was run for all markers and p-values determined for each marker by f-test. Note that this is the statistical equivalent of a crude GWAS. Linear models were then tested using 1 through the first 100 most significant markers, and the model with the best fit was returned. The returned model was then used to calculate the accuracy for the given fold. For the marker subset experiments where the number of markers (p) was less than 100, models were tested using 1 through p markers. MLR has been shown to be effective for agronomic traits with very simple genetic architectures, but is otherwise not expected to perform well [51].
Prediction based on pedigree alone was the sixth statistical method and was performed in order to determine if a.) the fold design method properly controlled for family structure within the dataset, and b.) if GS could outperform prediction based on pedigree alone [52].
All statistical modeling was done in R. For the pedigree models an A-matrix was calculated using a three-generation pedigree file for all individuals in the training and validation populations using a custom R function. The models themselves were calculated using package rrBLUP (function kin.BLUP). RR-BLUP models were also calculated using package rrBLUP (function kinship.BLUP). RKHS models were calculated using kinship.BLUP, K.method = "GAUSS", modified so that parameter theta was always equal to 2.5, as per guidelines in the BGLR package documentation [68]. Random Forest was performed using package randomForest (function randomForest). Bayesian LASSO was performed using package BLR (function BLR).
Narrow sense heritabilities were calculated for each trait on a per line basis using the rrBLUP package, function mixed.solve, with the least square means for the complete validation populations used as input. The narrow sense heritabilities were calculated as the additive genetic variance divided by the total phenotypic variance. The set of 73,147 SNPs was used for all experiments with the exception of the marker subset experiments described below.
The cross-validation results were analyzed using ANOVA and pairwise student's t to determine:
a significant difference in the accuracy of prediction of the two validation populations across statistical methods, i.e., where yi (accuracy) = μ + xijβj + εij, and i is one RYT experiment and stat method for validation population j (e.g. xi = CV experiment 1 for method RR-BLUP and j = validation population 2012 DS).
b significant difference in the performance of the six statistical methods across the different experiments, i.e., where yi (accuracy) = μ + xijβj + εij, and i is one RYT experiment for stat method j (e.g. xi = CV experiment 1 and j = RR-BLUP).
c significant difference in the performance of each experiment across statistical methods, after excluding the three worst-performing statistical methods (Bayesian LASSO, MLR, and pedigree only), i.e., where yi (accuracy) = μ + xijβj + εij, and i is one statistical method for RYT experiment j (e.g. xi = RR-BLUP and j = CV experiment 1) (Table 1, S2–S4 Tables).
Distributed. To select subsets of SNPs that were evenly distributed across the genome, 11 bin parameters were selected: 25Kb (0.1 cM), 50 Kb (0.2 cM), 120 Kb (.5 cM), 240Kb (1 cM), 480 Kb (2 cM), 840 Kb (3.5 cM), 1200 Kb (5 cM), 1800 Kb (7.5 cM), 2400 Kb (10 cM), 3600 Kb (15 cM), 4800 Kb (20 cM). For each bin parameter, all SNPs in the 73,147 SNP set were placed into bins according to the bin parameter. To select subsets of SNPs for a given bin size, the SNPs in each bin were sorted first by minor allele frequency, largest to smallest, and then by call rate, largest to smallest. Ten selections of SNPs were made for each bin size—the first subset consisted of the top ranked SNP in each bin, i.e., the SNP with the highest MAF and call rate, the second subset consisted of the second ranked SNP in each bin, and so on for the top ten SNPs in each bin. If a bin had fewer than ten SNPs, then the top SNP in each bin was chosen for all ten selections.
Each subset was then used as the genotype matrix to perform five-fold cross-validation using the same folds as for the original RYT cross validation experiments. The RYT 2012 wet season and the RYT 2012 dry season served as the validation populations and RYT years 2009–2011, all seasons, served as the training population. The five marker-dependent statistical methods tested previously were used once more: RR-BLUP, RKHS, Random Forest, Bayesian LASSO, and MLR. Accuracy was calculated for each of the ten selections (for each bin parameter) as previously. A mean accuracy, standard deviation, and standard error for each bin parameter were also calculated by averaging the cross-validation results of the 10 selections for each bin parameter (S5 Table).
The average accuracies with standard error as the error bars were plotted versus the number of SNPs in each subset (as determined by the bin size parameter) using JMP (Figs. 1, S4). The results for full 73,147 SNP set were included on these plots as a reference, although these accuracies are not averages. ANOVA and pairwise students were used to test for significant difference in the performance of the five statistical methods across the different bin parameter sizes, and for significant differences in the performance of the various bin parameter sizes (and thus total SNP number) across the five statistical methods (S5–S6 Tables).
Random. Ten random selections of SNPs were chosen from the 73,147 SNP set for 15 subset sizes: 24, 48, 65, 83, 96, 109, 161, 212, 316, 448, 781, 1553, 3076, 7142, 13101 using a pseudo-random numbers generator. Subset sizes 83, 109, 161, 212, 316, 448, 781, 1553, 3076, 7142, and 13101 were chosen to match the number of SNPs in the distributed SNP subsets described above. The additional SNP subset sizes were included to improve resolution.
Cross validation experiments and analysis were performed for the random subsets as described above for the distributed subsets (Fig. 1, S4 Fig., S6 Table).
|
10.1371/journal.pbio.2003703 | Population genomics of Mesolithic Scandinavia: Investigating early postglacial migration routes and high-latitude adaptation | Scandinavia was one of the last geographic areas in Europe to become habitable for humans after the Last Glacial Maximum (LGM). However, the routes and genetic composition of these postglacial migrants remain unclear. We sequenced the genomes, up to 57× coverage, of seven hunter-gatherers excavated across Scandinavia and dated from 9,500–6,000 years before present (BP). Surprisingly, among the Scandinavian Mesolithic individuals, the genetic data display an east–west genetic gradient that opposes the pattern seen in other parts of Mesolithic Europe. Our results suggest two different early postglacial migrations into Scandinavia: initially from the south, and later, from the northeast. The latter followed the ice-free Norwegian north Atlantic coast, along which novel and advanced pressure-blade stone-tool techniques may have spread. These two groups met and mixed in Scandinavia, creating a genetically diverse population, which shows patterns of genetic adaptation to high latitude environments. These potential adaptations include high frequencies of low pigmentation variants and a gene region associated with physical performance, which shows strong continuity into modern-day northern Europeans.
| The Scandinavian peninsula was the last part of Europe to be colonized after the Last Glacial Maximum. The migration routes, cultural networks, and the genetic makeup of the first Scandinavians remain elusive and several hypotheses exist based on archaeology, climate modeling, and genetics. By analyzing the genomes of early Scandinavian hunter-gatherers, we show that their migrations followed two routes: one from the south and another from the northeast along the ice-free Norwegian Atlantic coast. These groups met and mixed in Scandinavia, creating a population more diverse than contemporaneous central and western European hunter-gatherers. As northern Europe is associated with cold and low light conditions, we investigated genomic patterns of adaptation to these conditions and genes known to be involved in skin pigmentation. We demonstrate that Mesolithic Scandinavians had higher levels of light pigmentation variants compared to the respective source populations of the migrations, suggesting adaptation to low light levels and a surprising signal of genetic continuity in TMEM131, a gene that may be involved in long-term adaptation to the cold.
| As the ice sheet retracted from northern Europe after the Last Glacial Maximum (LGM), around 23,000 years ago, new habitable areas emerged [1], allowing plants [2,3] and animals [4,5] to recolonize the Scandinavian peninsula (hereafter referred to as Scandinavia). There is consistent evidence of human presence in the archaeological record from approximately 11,700 years before present (BP) both in southern and northern Scandinavia [6–9]. At this time, the ice sheet was still dominating the interior of Scandinavia [9,10] (Fig 1A, S1 Text), but recent climate modeling shows that the Arctic coast of (modern-day) northern Norway was ice free [10]. Similarities in late-glacial lithic technology (direct blade percussion technique) of Western Europe and the oldest counterparts of Scandinavia appearing around 11,000 calibrated (cal) BP [11] (S1 Text) have been used to argue for an early postglacial migration from southwestern Europe into Scandinavia, including areas of northern Norway. However, studies of another lithic technology, the “pressure blade” technique, which first occurred in the northern parts of Scandinavia around 10,200 cal BP, indicates contact with groups in the east and possibly an eastern origin of the early settlers [7,12–15] (S1 Text). The first genetic studies of Mesolithic human remains from central and eastern Scandinavian hunter-gatherers (SHGs) revealed similarities to two different Mesolithic European populations, the “western hunter-gatherers” (WHGs) from western, central, and southern Europe and the “eastern hunter-gatherers” (EHGs) from northeastern and eastern Europe [16–24]. Archaeology, climate modeling, and genetics suggest several possibilities for the early postglacial migrations into Scandinavia, including migrations from the south, southeast, northeast, and combinations of these; however, the early postglacial peopling of Scandinavia remains elusive [1,4,6–19,25,26]. In this study, we contrast genome sequence data and stable isotopes from Mesolithic human remains from western, northern, and eastern Scandinavia to infer the early postglacial migration routes into Scandinavia—from where people came, what routes they followed, how they were related to other Mesolithic Europeans [17–21,27]—and to investigate human adaptation to high-latitude environments.
We sequenced the genomes of seven hunter-gatherers from Scandinavia (Table 1; S1, S2 and S3 Text) ranging from 57.8× to 0.1× genome coverage, of which four individuals had a genome coverage above 1×. The remains were directly dated to between 9,500 cal BP and 6,000 cal BP, and were excavated in southwestern Norway (Hum1, Hum2), northern Norway (Steigen), and the Baltic islands of Stora Karlsö and Gotland (SF9, SF11, SF12, and SBj), and represent 18% (6 of 33) of all known human remains in Scandinavia older than 8,000 years [30]. All samples displayed fragmentation and cytosine deamination at fragment termini characteristic for ancient DNA (aDNA) (S3 Text). Mitochondrial (mt) DNA-based contamination estimates were <6% for all individuals (confidence intervals ranging from 0% to 9.5%) and autosomal contamination was <1% for all individuals except for SF11, which showed approximately 10% contamination (Table 1, S4 Text). Four of the seven individuals were inferred to be males and three were females. All the western and northern Scandinavian individuals and one eastern Scandinavian carried U5a1 mt haplotypes, whereas the remaining eastern Scandinavians carried U4a haplotypes (Table 1, S5 Text). These individuals represent the oldest U5a1 and U4 lineages detected so far. The Y chromosomal haplotype was determined for three of the four males, all carried I2 haplotypes, which were common in pre-Neolithic Europe (Table 1, S5 Text).
The high coverage and Uracil-DNA-glycosylase (UDG)-treated genome (used in order to reduce the effects of postmortem DNA damage) [31] of SF12 allowed us to confidently discover new and hitherto unknown variants at sites with 55× or higher sequencing depth (S3 Text). Based on SF12’s high-coverage and high-quality genome, we estimate the number of SNPs hitherto unknown (not recorded in dbSNP [v142]) to be approximately 10,600. This number is close to the median per European individual in the 1000 Genomes Project [32] (approximately 11,400, S3 Text), although a direct comparison is difficult due to the lower sequencing depth, different data processing, and larger sample sizes in the 1000 Genomes Project. At least 17% of these SNPs that are not found in modern-day individuals were in fact common among the Mesolithic Scandinavians (seen in the low coverage data conditional on the observation in SF12), and in total 24.2% were found in other prehistoric individuals (S3 Text), suggesting a substantial amount of hitherto unknown variation 9,000 years ago (S3 Text). Thus, many genetic variants found in Mesolithic individuals have not been carried over to modern-day groups. Among the novel variants in SF12, four (all heterozygous) are predicted to affect the function of protein coding genes [33] (S3 Text). The “heat shock protein” HSPA2 in SF12 carries an unknown mutation that changes the amino acid histidine to tyrosine at a protein–protein interaction site, which likely disrupts the function of the protein (S3 Text). Defects in HSPA2 are known to drastically reduce fertility in males [34]. It will be interesting to see how common such variants were among Mesolithic groups as more genome sequence data become available. The genomic data further allowed us to study the physical appearance of SHGs (S8 Text); for instance, they show a combination of eye color varying from blue to light brown and light skin pigmentation. This is strikingly different from the WHGs—who have been suggested to have the specific combination of blue eyes and dark skin [18,20,21,23] and EHGs—who have been suggested to be brown-eyed and light-skinned [19,20].
In order to compare the genomic sequence data of the seven SHGs to genetic information from other ancient individuals and modern-day groups, data were merged with shotgun sequence data and SNP capture data from six published Mesolithic individuals from Motala in central Scandinavia, and 47 published Stone Age (Upper Paleolithic, Mesolithic, and Early Neolithic) individuals from other parts of Eurasia (S6 Text) [17–22,26,27,29,35–38], as well as with a world-wide set of 203 modern-day populations [18,32,39]. All 13 SHGs—regardless of geographic sampling location and age—display genetic affinities to both WHGs and EHGs (Fig 1A and 1B, S6 Text). One individual, SF11, seems to be a slight genetic outlier in the principal component analysis (PCA), which could be due to the lower coverage or driven by nuclear contamination (Table 1, S6 Text). Generally, the pattern of dual ancestry is consistent with a scenario in which SHGs represent a mixed group tracing parts of their ancestry to both the WHGs and the EHGs [17–19,22,24,40].
The SHGs from northern and western Scandinavia show a distinct and significantly stronger affinity to the EHGs compared to the central and eastern SHGs (Fig 1). Conversely, the SHGs from eastern and central Scandinavia were genetically more similar to WHGs compared to the northern and western SHGs (Fig 1). Using qpAdm [19], the EHG genetic component of northern and western SHGs was estimated to 48.9% (± 5%) and differs from the 37.8% (± 3.2%) observed in eastern and south-central SHGs. The latter estimate is similar to ancestry estimates obtained for eastern Baltic hunter-gatherers from Latvia [29] (33.7% ± 4.7%, Fig 1A). Although the difference in ancestry estimates between northern and western SHG, and eastern and south-central SHG is only marginally significant (Z = 1.87, p = 0.062), this pattern is in agreement with other analyses such as ADMIXTURE and TreeMix (S6 Text). Furthermore, the direct comparison using D statistics with Chimpanzee (Chimp) as an outgroup (D(Chimp, WHG; eastern or south-central SHG, northern or western SHG) < 0, Z = −5.14 and D(Chimp, EHG; eastern or south-central SHG, northern or western SHG) > 0, Z = 1.72) show that WHG are genetically closer to eastern and south-central SHG, whereas EHG tend to share more alleles with northern and western SHGs (S2 Fig). These patterns of genetic affinity within SHGs are in direct contrast to the expectation based on geographic proximity with EHGs and WHGs.
From about 11,700 cal BP, consistent archaeological evidence of human presence exists in southern Scandinavia following the retreat of the ice sheet [6,41,42] (S1 Text). Artifacts and tools found at these sites show similarities with the Ahrensburgian tradition of northern central Europe [15,43], suggesting that these hunter-gatherers likely had a southern origin from a WHG-like gene pool as no EHG ancestry has been found in central and western Europe [18,21,24,27]. Although this genetic component would have entered from today’s northern Germany and Denmark (Fig 2, Scenario a), it remains unclear how and where the EHG component entered Scandinavia (Fig 2, Scenarios b, c and/or d). The EHG-related migration likely took place after the migration of WHGs from the south as the earliest eastern-associated pressure blade finds postdate the southwestern-associated direct blade finds in Scandinavia (S1 Text). Two migrations with admixture at different time-periods would generate a genetic gradient with the highest contribution of a source close to its geographic region of entry. The observed genetic pattern is consistent with a migration of the EHGs from the northeast moving southwards along the ice-free Norwegian Atlantic coast where the two groups started mixing (Fig 2, Scenarios a and b), which would cause more EHG ancestry in western SHGs. If the EHG migration had crossed the Baltic Sea into Scandinavia, where it would meet and mix with a WHG-like population (Fig 2, combination of Scenarios a and c), a gradient with most EHG ancestry in eastern SHGs would have been created—exactly opposite to the observed pattern. A similar pattern would be expected if the EHG migration went around the Baltic Sea along current day’s Finnish west coast and down via today’s Swedish east coast (scenario not depicted in Fig 2). An EHG migration along the southern Baltic coast (Fig 2, Scenarios a and d) should cause a related pattern to a crossing of the Baltic Sea with more EHG ancestry in central and eastern SHGs. Furthermore, such a scenario would likely also make the Latvian Mesolithic hunter-gatherers the group with most EHG ancestry, which is in stark contrast to the empirical data in which the Latvian group shows the lowest proportion of EHG ancestry (33.7% ± 4.7%), and also not consistent with chronology, as the dated settlements east of the Baltic Sea are younger than the early settlements in Scandinavia (S1 Text). Thus, the only scenario consistent with both genetic and archaeological data is a migration of a WHG-related group migrating into Scandinavia from the south, followed by an EHG-related group migrating to Scandinavia from the northeast along the Norwegian Atlantic coast. Notably, such a migration along the Norwegian coast could have been facilitated by the use of the more specialized pressure blade technique (S1 Text) [12,14]. The individuals sequenced here postdate these migrations, but a genetic east-west gradient would be maintained over time in Scandinavia and only additional large-scale migrations from different sources would alter this pattern. This observation is important as the geographic pattern still holds without the chronologically much younger Steigen individual, which might represent local continuity or later migrations into north-western Scandinavia from the east.
Interestingly, stable nitrogen and carbon isotope analysis of northern and western SHGs revealed an extreme marine diet, suggesting a pronounced maritime subsistence, in contrast to the more mixed terrestrial and aquatic diet of eastern and central SHGs (S1 Text). Mobility is difficult to trace based solely on carbon and nitrogen isotope data; however, the patterns are consistent with a migration along the Norwegian Atlantic coast relying on local resources.
By sequencing complete ancient genomes, we can compute unbiased estimates of genetic diversity, which are informative of past population sizes and population history. Here, we restrict the analysis to WHGs and SHGs because only SNP capture data are available for EHGs (S7 Text). In modern-day Europe, there is greater genetic diversity in the south compared to the north. During the Mesolithic period, by contrast, we find lower levels of runs of homozygosity (RoH) (Fig 3A) and linkage disequilibrium (LD) (Fig 3B) in SHGs compared to WHGs (represented by Loschbour and Bichon [18,35]). By using a multiple sequentially Markovian coalescent (MSMC) approach [44] for the high-coverage, high-quality genome of SF12, we find that right before the SF12 individual lived, the effective population size of SHGs was similar to that of WHGs (Fig 3C). At the time of the LGM and back to approximately 50,000 years ago, both the WHGs and SHGs go through a bottleneck, but the ancestors of SHGs retained a greater effective population size in contrast to the ancestors of WHGs who went through a more severe bottleneck (Fig 2C), which is consistent across 100 bootstrap replicates (S2 Fig). These differences in effective population size estimates may be attributed to the admixture in SHGs as migration events can have delayed effects on estimates of effective population size over time [45]. Around 50,000–70,000 years ago, the effective population sizes of the ancestors of SHGs, WHGs, Neolithic groups (represented by Stuttgart [18]), and Paleolithic Eurasians (represented by Ust-Ishim [38]) align, suggesting that these diverse groups all trace their ancestry back to a common ancestral group, which likely represents the early migrants out of Africa.
With the aim of detecting signs of adaptation to high-latitude environments and selection during and after the Mesolithic period, we employed two different approaches that utilize the Mesolithic genomic data. In the first approach, we assumed that SHGs adapted to high-latitude environments of low temperatures and seasonally low levels of light, and searched for gene variants that carried over to modern-day people in northern Europe. Modern-day northern Europeans trace limited amounts of genetic material back to the SHGs (due to the many additional migrations during later periods), and any genomic region that displays extraordinary genetic continuity would be a strong candidate for adaptation in people living in northern Europe across time. We designed a statistic, Dsel (S9 Text), that captures this specific signal and scanned the whole genome for gene variants that show strong continuity (little differentiation) between SHGs and modern-day northern Europeans while exhibiting large differentiation to modern-day southern European populations [46] (Fig 4A; S9 Text). Six of the top 10 SNPs with greatest Dsel values were located in the TMEM131 gene that has been found to be associated with physical performance [47], which could make it part of the physiological adaptation to cold [48]. This genomic region was more than 200 kbp (kilo base pairs) long and showed the strongest haplotypic differentiation between modern-day Tuscan individuals (TSIs) and modern-day Finnish individuals (FINs) across the genome (S9 Text). The particular haplotype was relatively common in SHGs, it is even more common among today’s Finnish population (S9 Text) and showed a strong signal of local adaptation (S9 Text). Other top hits included genes associated with a wide range of metabolic, cardiovascular, and developmental and psychological traits (S9 Text) potentially linked to physiological adaptation to cold environments [48].
In addition to performing this genome-wide scan, we studied the allele frequencies in three pigmentation genes (SLC24A5, SLC45A2, which have a strong effect on skin pigmentation, and OCA2/HERC2, which has a strong effect on eye pigmentation) in which the derived alleles are virtually fixed in northern Europeans today. The differences in allele frequencies of those three loci are among the highest between human populations, suggesting that selection was driving the differences in eye color, skin, and hair pigmentation as part of the adaptation to different environments [50–53]. All of the depigmentation variants at these three genes are in high frequency in SHGs in contrast to both WHGs and EHGs (Fig 4B). We conduct neutral simulations of the allele frequencies in an admixed SHG population to estimate p-values for observing these allele frequencies without selection (S9 Text). The p-values for all three SNPs are lower than 0.2; the combined p-value [54] for all three pigmentation SNPs is 0.028. Therefore, the unique configuration of the SHGs is not fully explained by the fact that SHGs are a mixture of EHGs and WHGs, but could rather be explained by a continued increase of the allele frequencies after the admixture event, likely caused by adaptation to high-latitude environments [50,52].
By combining information from climate modeling, archaeology, and Mesolithic human genomes, we were able to reveal the complexity of the early migration patterns into Scandinavia and human adaptation to high-latitude environments. We disentangled two migration routes and linked them to particular archaeological patterns. We also demonstrated greater genetic diversity in Mesolithic northern Europe compared to southern and central Europe—in contrast to modern-day patterns—and showed that many genetic variants that were common during the Mesolithic period have been lost today. These findings reiterate the importance of human migration for dispersal of novel technology in human prehistory [14–20,27,40,55–58].
Genomic sequence data were generated from teeth and bone samples belonging to seven (eight, including SF13) Mesolithic SHGs (S1 Text). A detailed description on the archaeological background of the samples as well as post-LGM Scandinavia can be found in S1 Text. Additional libraries were sequenced for two previously published Neolithic hunter-gatherers, Ajvide58 and Ajvide70 [17] (S2 Text). All samples were prepared in the dedicated aDNA facilities at Uppsala University (SF9, SF11, SF12, SF13, SBj, Hum1, Hum2, Ajvide58, Ajvide70) and at Stockholm University (Steigen).
Bones and teeth were decontaminated prior to analysis by wiping them with a 1% Sodiumhypoclorite solution and DNA-free water. Furthermore, all surfaces were UV irradiated (6 J/cm2 at 254 nm). After removing 1 millimeter of the surface, approximately 30–300 mg of bone was powderized and DNA was extracted following silica-based methods as in [59] with modifications as in [57,60] or as in [61] and eluted in 25–110 μl of EB buffer. Between one and 16 extractions were made from each sample and one extraction blank with water instead of bone powder was included per six to 10 extracts. Blanks were carried along the whole process until quantitative PCR (qPCR) and/or PCR and subsequent quantification.
DNA libraries were prepared using 20 μl of extract, with blunt-end ligation coupled with P5 and P7 adapters and indexes as described in [57,62]. From each extract one to five double stranded libraries were built. Because aDNA is already fragmented, the shearing step was omitted from the protocol. Library blank controls, including water as well as extraction blanks, were carried along during every step of library preparation. In order to determine the optimal number of PCR cycles for library amplification, qPCR was performed. Each reaction was prepared in a total volume of 25 μl, containing 1 μl of DNA library, 1X MaximaSYBRGreen mastermix, and 200 nM each of IS7 and IS8 [62] reactions were set up in duplicates. Each blunt-end library was amplified in four to 12 replicates with one negative PCR control per index-PCR. The amplification reactions had a total volume of 25 μl, with 3 μl DNA library, and the following in final concentrations: 1 X AmpliTaq Gold Buffer, 2.5 mM MgCl2, 250 μM of each dNTP, 2.5 U AmpliTaq Gold (Thermo Fisher Scientific, Waltham, MA), and 200 nM each of the IS4 primer and index primer [62]. PCR was done with the following conditions: an activation step at 94°C for 10 min followed by 10–16 cycles of 94°C for 30 s, 60°C for 30 s, and 72°C for 30 s, and a final elongation step of 72°C for 10 min. For each library, four amplifications with the same indexing primer were pooled and purified with AMPure XP beads (Agencourt; Beckman Coulter, Brea, CA). The quality and quantity of libraries was checked using Tapestation or BioAnalyzer using the High Sensitivity Kit (Agilent Technologies, Cary, NC). None of the blanks showed any presence of DNA comparable to that of a sample and were therefore not further analyzed. For initial screening, 10–20 libraries were pooled at equimolar concentrations for sequencing on an Illumina HiSeq 2500 using v.4 chemistry, and 125 bp paired-end reads or HiSeqX, 150 bp paired-end length using v2.5 chemistry at the SNP & SEQ Technology Platforms at Uppsala University and Stockholm University. After evaluation of factors such as clonality, proportion of human DNA, and genomic coverage samples were selected for resequencing, aiming to yield as high coverage as possible for each library.
Based on the results of the non-damage-repair sequencing, the SF12 individual was selected for large-scale sequencing in order to generate a high-coverage genome of high quality where damages had been repaired using UDG. In addition to the 15 extracts previously prepared and used for non-damage repair libraries, another 111 extracts were made based on a variety of silica-based methods [27,57,59,60]. From these 126 extracts, a total of 258 damage-repaired double-stranded libraries were built for Illumina sequencing platforms. Libraries were built as above, except a DNA repair step in which UDG and endonuclease VIII or USER enzyme (NEB) treatment was included in order to remove deaminated cytosines [63]. qPCR was performed in order to quantify the number of molecules and the optimal number of PCR cycles prior to amplification for each DNA library. Furthermore, this step included extraction blanks, library blanks, and amplification blanks to monitor potential contamination. All of these negative controls showed an optimal cycle of amplification significantly higher to those of our aDNA libraries (>10 cycles) and they were thus deemed as negative. Our experimental results show minimal levels of contamination, which is in concordance with mt DNA and X chromosome estimates of contamination (see S4 Text and Table 1). Each reaction was done in a total volume of 25 μl, containing 1 μl of DNA library, 1 X MaximaSYBRGreen mastermix (Thermo Fisher Scientific), and 200 nM each of IS7 and IS8 [62], reactions were set up in duplicate. The PCRs were set up using a similar system as for the nondamage repair samples (in quadruplicates that were pooled prior to cleanup of the PCR products), except for using AccuPrime DNA polymerase (Thermo Fisher Scientific) instead of AmpliTaqGold (Thermo Fisher Scientific) and the following PCR conditions: an activation step at 95°C for 2 min followed by 10–16 cycles of 95°C for 15 s, 60°C for 30 s, and 68°C for 1 min, and a final elongation step of 68°C for 5 min. Blank controls, including water as well as extraction blanks, were carried out during every step of library preparation. Amplified libraries were pooled, cleaned, quantified, and sequenced in the same manner as non-damage repaired libraries. A small proportion of the libraries (n = 14) were also subjected to whole genome capture (WGC) using European MYbaits from MYcroarray, and following the manufacturers protocol as done in [64].
In order to sequence libraries to depletion, two to eight libraries were pooled together and sequenced until reaching a clonality of >50%; if sequencing was halted before reaching that clonality level, it was either because the library was classified as unproductive based on the genome coverage generated, or that the sequencing goal (>55 × coverage) was already reached and further sequencing was deemed unnecessary. Sequencing was performed as above.
Paired-end reads were merged using MergeReadsFastQ_cc.py [65]; if an overlap of at least 11 base pairs was found, the base qualities were added together and any remaining adapters were trimmed. Merged reads were then mapped single-ended with bwa aln 0.7.13 [66] to the human reference genome (build 36 and 37) using the following nondefault parameters: seeds disabled -l 16500 -n 0.01 -o 2 [17,18]. To remove PCR duplicates, reads with identical start and end positions were collapsed using a modified version, to ensure random choice of bases, of FilterUniqSAMCons_cc.py [65]. Reads with less than 10% mismatches to the human reference genome, reads longer than 35 base pairs, and reads with mapping quality higher than 30 were used to estimate contamination.
The genetic data obtained from the two bone elements SF9 and SF13 showed extremely high similarities, which suggested that the two individuals were related. Using READ [67], a tool to estimate kin-relationship from aDNA, SF9 and SF13 were classified as either identical twins or the same individual. Therefore, we merged the genetic data for both individuals and refer to the merged individual as SF9 throughout the genetic analysis.
All data show damage patterns indicative of authentic aDNA (S3 Text). Contamination was estimated using three different sources of data: (1) the mt genome [68], (2) the X chromosome if the individual was male [69,70], and (3) the autosomes [71]. The data mapping to the human genome can be considered largely endogenous, as the contamination estimates were low across all three methods (S4 Text).
Most population genomic analyses require a set of reference data for comparison. We compiled three different data sets from the literature and merged them with the data from ancient individuals (S6 Text). The three reference SNP panels were as follows:
These data sets were merged with ancient individuals of less than 15× genome coverage using the following approach: for each SNP site, a random read covering that site with minimum mapping quality 30 was drawn (using samtools 0.1.19 mpileup [73]) and its allele was assumed to be homozygous in the ancient individual. Transition sites were coded as missing data for individuals that were not UDG-treated, and SNPs showing additional alleles or indels in the ancient individuals were excluded from the data.
The six high-coverage ancient individuals (SF12, NE1 [27], Kotias [35], Loschbour [18], Stuttgart [18], and Ust-Ishim [38]) used in this study were treated differently, as we generated diploid genotype calls for them. First, the base qualities of all Ts in the first five base pairs of each read as well as all As in the last five base pairs were set to 2. We then used Picard [74] to add read groups to the files. Indel realignment was conducted with GATK 3.5.0 [66] using indels identified in phase 1 of the 1000 Genomes Project as reference [32]. Finally, GATK’s UnifiedGenotyper was used to call diploid genotypes with the parameters -stand_call_conf 50.0, -stand_emit_conf 50.0, -mbq 30, -contamination 0.02, and—output_mode EMIT_ALL_SITES using dbSNP version 142 as known SNPs. SNP sites from the reference data sets were extracted from the VCF files using vcftools [72] if they were not marked as low-quality calls. Plink 1.9 [75,76] was used to merge the different data sets.
We performed PCA to characterize the genetic affinities of the ancient Scandinavian genomes to previously published ancient and modern genetic data. PCA was conducted on 42 present-day west Eurasian populations from the Human Origins data set [18,39] using smartpca [77] with numoutlieriter: 0 and lsqproject: YES options. A total of 59 ancient genomes (52 previously published and 7 reported here) (S6 Text) were projected into the reference PCA space and computed from the genotypes of modern individuals. For all individuals, a single allele was selected randomly—making the data set fully homozygous. The result was plotted using the ploteig program of EIGENSOFT [77] with the–x and–k options.
popstats [78] was used to calculate D statistics to test deviations from a tree-like population topology of the shape ((A,B);(X,Y)) [39]. Standard errors were calculated using a weighted block jackknife of 0.5 Mbp. The tree topologies are balanced at zero, indicating no recent interactions between the test populations. Significant deviations from zero indicate a deviation from the proposed tree topology depending on the value. Positive values indicate an excess of shared alleles between A and X or B and Y, whereas negative values indicate more shared alleles between B and X or A and Y. Using an outgroup as population A limits the test results to depend on the recent relationships between B and Y (if positive) or B and X (if negative). Here, we used high-coverage Mota [37], Yoruba [32], and Chimp genome as (A) outgroups. popstats [78] was used to calculate f4 statistics in order to estimate shared drift between groups. Standard errors and Z scores for f4 statistics were estimated using a weighted block jackknife (Fig 1C).
A model-based clustering algorithm, implemented in the ADMIXTURE software [79], was used to estimate ancestry components and to cluster individuals. ADMIXTURE was conducted on the Human Origins data set [18,39], which was merged with the ancient individuals as described above. Data was pseudo-haploidized by randomly selecting one allele at each heterozygous site of present-day individuals. Finally, the data set was filtered for LD using PLINK [75,76] with parameters (--indep-pairwise 200 25 0.4), this retained 289,504 SNPs. ADMIXTURE was run in 50 replicates with different random seeds for ancestral clusters from K = 2 to K = 20. Common signals between independent runs for each K were identified using the LargeKGreedy algorithm of CLUMPP [80]. Clustering was visualized using rworldmap, ggplot2, SDMTools, and RColorBrewer packages of GNU R version 3.3.0. Starting from K = 3, when the modern samples split up into an African and eastern and western Eurasian clusters, the Mesolithic Scandinavians from Norway show slightly higher proportions of the Eastern cluster than Swedish Mesolithic individuals. This pattern continues to develop across higher values of K and it is consistent with the higher Eastern affinities of the Norwegian samples seen in the PCA and D- and f4 statistics. The results for all Ks are shown in S1 Fig.
In addition to ADMIXTURE, we assessed the admixture patterns in Mesolithic Scandinavians using a set of methods implemented in ADMIXTOOLS [39], qpWave [81], and qpAdm [19]. Both methods are based on f4 statistics, which relate a set of test populations to a set of outgroups in different distances from the potential source populations. We used the following set of outgroup populations from the Human Origins data set: Ami_Coriell, Biaka, Bougainville, Chukchi, Eskimo_Naukan, Han, Karitiana, Kharia, and Onge. We first used qpWave to test the number of source populations for Mesolithic west Eurasians (WHG). qpWave calculates a set of statistics X(u,v) = f4(u0, u; v0, v) where u0 and v0 are populations from the sets of test populations L and outgroups R, respectively. To avoid having more test populations than outgroups, we built four groups consisting of (1) genetically western and central hunter-gatherers (Bichon, Loschbour, KO1, LaBrana), (2) EHGs (UzOO74/I0061, SVP44/I0124, UzOO40/I0211), (3) Norwegian hunter-gathers (Hum1, Hum2, Steigen), and (4) Swedish hunter-gatherers (individuals from Motala and Mesolithic Gotland). qpWave tests the rank of the matrix of all X(u,v) statistics. If the matrix has rank m, the test populations can be assumed to be related to at least m + 1 “waves” of ancestry, which are differently related to the outgroups. A rank of 0 is rejected in our case (p = 3.13e-81), whereas a rank of 1 is consistent with the data (p = 0.699). Haak et al. [19] already showed, using the same approach, that WHG and EHG descend from at least two sources (confirmed with our data as rank 0 is rejected with p = 1.66e-86, whereas rank 1 is consistent with the data), and adding individuals from Motala does not change these observations. Therefore, we conclude that European Mesolithic populations, including Swedish and Norwegian Mesolithic individuals, have at least two source populations.
We then used qpAdm to model Mesolithic Scandinavian individuals as a 2-way admixture of WHG and EHG. qpAdm was run separately for each Scandinavian individual x, setting T = x as target and S = (EHG, WHG) as sources. The general approach of qpAdm is related to qpWave: target and source are used as L (with T being the base population), and f4 statistics with outgroups from R (same as above) are calculated. The rank of the resulting matrix is then set to the number of sources minus one, which allows to estimate the admixture contributions from each population in S to T. The results are shown in Fig 1. We calculate a Z score for the difference between Norwegian and Swedish SHG as Z=anor-asweSEnor2+SEswe2 where a are the ancestry estimates and SE are the respective block-jackknife estimates of the standard errors.
Heterozygosity is a measurement for general population diversity and its effective population size. Analyzing the extent of homozygous segments across the genome can also give us a temporal perspective on the effective population sizes. Many short segments of homozygous SNPs can be connected to historically small population sizes, whereas an excess of long RoH suggests recent inbreeding. We restricted this analysis to the six high-coverage individuals (SF12, NE1, Kotias, Loschbour, Stuttgart, Ust-Ishim) for which we obtained diploid genotype calls and we compared them to modern individuals from the 1000 Genomes Project. The length and number of RoH were estimated using Plink 1.9 [75,76] and the parameters--homozyg-density 50,--homozyg-gap 100,--homozyg-kb 500,--homozyg-snp 100,--homozyg-window-het 1,--homozyg-window-snp 100,--homozyg-window-threshold 0.05, and--homozyg-window-missing 20. The results are shown in Fig 3A.
Similar to RoH, the decay of LD harbors information on the demographic history of a population. Long-distance LD can be caused by a low effective population size and past bottlenecks. Calculating LD for aDNA data is challenging, as the low amounts of authentic DNA usually just yields haploid allele calls with unknown phase. In order to estimate LD decay for ancient populations, we first combine two haploid ancient individuals to a pseudo-diploid individual (similar to the approach chosen for conditional nucleotide diversity, S7 Text). Next, we bin SNP pairs by distance (bin size 5 kb) and then calculated the covariance of derived allele frequencies (0, 0.5, or 1.0) for each bin. This way, we do not need phase information to calculate LD decay because we do not consider multilocus haplotypes, which is similar to the approach taken by ROLLOFF [39,82] and ALDER [83] to date admixture events based on admixture LD decay. For Fig 3B, we used two modern 1000 Genomes Project populations to scale the LD per bin. The LD between two randomly chosen PELs (modern-day Peruvian individuals) was set to 1 and the LD between two randomly chosen TSIs was set to 0. This approach is used to obtain a relative scale for the ancient populations, and we caution against a direct interpretation of the differences to modern populations because technical differences in the modern data (e.g., SNP calling or imputation) may have substantial effects.
We are using MSMC’s implementation of PSMC’ [44] to infer effective population sizes over time from single high-coverage genomes. We restrict this analysis to UDG-treated individuals (SF12, Loschbour, Stuttgart, Ust-Ishim) as postmortem damage would cause an excess of false heterozygous transition sites. Input files were prepared using scripts provided with the release of MSMC (https://github.com/stschiff/msmc-tools) and MSMC was run with the nondefault parameters--fixedRecombination and -r 0.88 in order to set the ratio of recombination to mutation rate to a realistic level for humans. We also estimate effective population size for six high-coverage modern genomes [84] (Fig 3C). We plot the effective population size assuming a mutation rate of 1.25x10e-8 and a generation time of 30 years. The curves for ancient individuals were shifted based on their average C14 date. Additionally, we used multihetsep_bootstrap.py to generate 100 bootstraps per individual. The results are shown in S4 Fig.
We scanned the genomes for SNPs with similar allele frequencies in Mesolithic and modern-day northern Europeans and contrasted it to a modern-day population from southern latitudes. Pooling all Mesolithic Scandinavians together, we obtain an allele frequency estimate for SHGs, which is compared to FINs and TSIs from the 1000 Genomes Project [32]. We use the Finnish population as representatives of modern-day northern Europeans (this sample contains the largest number of sequenced genomes from a northern European population). Tuscans are used as an alternative population, who also trace some ancestry to Mesolithic populations, but who do not trace their ancestry to groups that lived at northern latitudes in the last 7,000–9,000 years. Our approach is similar to PBS [85] and inspired by DAnc [46]. For each SNP, we calculated the statistic Dsel, comparing the allele frequencies between one ancestral and two modern populations:
Dsel=|DAFSHG−DAFTSI|−|DAFSHG−DAFFIN|
This scan was performed on all transversion SNPs extracted from the 1000 Genomes Project data. Only sites with a high-confidence ancestral allele in the human ancestor (as used by the 1000 Genomes Project [32]) and with coverage for at least six ancient Scandinavians were included in the computation. More information can be found in S9 Text.
|
10.1371/journal.ppat.1001101 | Bunyaviridae RNA Polymerases (L-Protein) Have an N-Terminal, Influenza-Like Endonuclease Domain, Essential for Viral Cap-Dependent Transcription | Bunyaviruses are a large family of segmented RNA viruses which, like influenza virus, use a cap-snatching mechanism for transcription whereby short capped primers derived by endonucleolytic cleavage of host mRNAs are used by the viral RNA-dependent RNA polymerase (L-protein) to transcribe viral mRNAs. It was recently shown that the cap-snatching endonuclease of influenza virus resides in a discrete N-terminal domain of the PA polymerase subunit. Here we structurally and functionally characterize a similar endonuclease in La Crosse orthobunyavirus (LACV) L-protein. We expressed N-terminal fragments of the LACV L-protein and found that residues 1-180 have metal binding and divalent cation dependent nuclease activity analogous to that of influenza virus endonuclease. The 2.2 Å resolution X-ray crystal structure of the domain confirms that LACV and influenza endonucleases have similar overall folds and identical two metal binding active sites. The in vitro activity of the LACV endonuclease could be abolished by point mutations in the active site or by binding 2,4-dioxo-4-phenylbutanoic acid (DPBA), a known influenza virus endonuclease inhibitor. A crystal structure with bound DPBA shows the inhibitor chelating two active site manganese ions. The essential role of this endonuclease in cap-dependent transcription was demonstrated by the loss of transcriptional activity in a RNP reconstitution system in cells upon making the same point mutations in the context of the full-length LACV L-protein. Using structure based sequence alignments we show that a similar endonuclease almost certainly exists at the N-terminus of L-proteins or PA polymerase subunits of essentially all known negative strand and cap-snatching segmented RNA viruses including arenaviruses (2 segments), bunyaviruses (3 segments), tenuiviruses (4–6 segments), and orthomyxoviruses (6–8 segments). This correspondence, together with the well-known mapping of the conserved polymerase motifs to the central regions of the L-protein and influenza PB1 subunit, suggests that L-proteins might be architecturally, and functionally equivalent to a concatemer of the three orthomyxovirus polymerase subunits in the order PA-PB1-PB2. Furthermore, our structure of a known influenza endonuclease inhibitor bound to LACV endonuclease suggests that compounds targeting a potentially broad spectrum of segmented RNA viruses, several of which are serious or emerging human, animal and plant pathogens, could be developed using structure-based optimisation.
| Bunyaviruses are a large family of RNA viruses that include serious human, animal and plant pathogens. The viral RNA-dependent RNA polymerase (L-protein) is responsible for replication and transcription of the viral RNA, but apart from its central polymerase domain, it is poorly characterized. Like influenza virus polymerase, bunyavirus L-proteins employ a cap-snatching mechanism to transcribe viral mRNAs, by which host mRNAs are endonucleolytically cleaved as a source of short capped primers. Influenza polymerase endonuclease has recently been located at the PA subunit N-terminus. Here we show biochemically and by crystal structure determination that a similar two-manganese dependent nuclease exists at the N-terminus of La Crosse orthobunyavirus L-protein, whose function is required for cap-dependent transcription. By sequence analysis we show that similar endonuclease signature motifs exist in almost all known segmented RNA, cap-snatching viruses including arenaviruses, bunyaviruses, tenuiviruses and orthomyxoviruses. This suggests that the polymerases of these viruses might share a conserved global architecture with the L-protein being equivalent to a concatenation of the orthomxyovirus PA-PB1-PB2 subunits. We also propose that broad spectrum drugs targeting the endonuclease domain of such viruses could be developed, as exemplified by our structure of the LACV endonuclease complexed with a known influenza endonuclease inhibitor.
| Bunyaviridae is the largest single family of mostly animal viruses comprising more than 300 species, divided into five genera: Orthobunyavirus, Phlebovirus, Nairovirus, Hantavirus and Tospovirus, the latter infecting plants. The viruses are mainly insect transmitted except Hantaviruses which are rodent borne. They possess a tri-partite negative sense RNA genome, the segments being designated according to size as L, M and S. The L segment encodes a single protein, the RNA-dependent RNA polymerase (polymerase or L protein) which ranges according to genus from 240–460 KDa; the M segment encodes two glycoproteins (Gn, Gc) and in some cases a non-structural protein (NSm) and the S segment encodes the nucleocapsid protein (N) and generally a non-structural protein (NSs). In common with other negative strand RNA viruses, the RNA genome is coated with N protein forming ribonucleoprotein complexes (RNPs) which also contain the polymerase. Bunyavirus particles are generally spherical with the glycoproteins embedded in a membrane envelope which surrounds the RNPs. Replication occurs in the cytoplasm, unlike influenza virus, a negative strand segmented RNA virus of the orthomyxovirus family, which replicates in the nucleus. Bunyaviruses are globally widespread although individual species may be locally restricted by the specificity for particular insect species. Several bunyaviruses are important or emerging human or plant pathogens including La Crosse orthobunyavirus (childhood encephalitis), Hantaan virus (hemorrhagic fever with renal syndrome), Rift Valley fever phlebovirus, tomato spotted wilt tospovirus and Crimean-Congo (hemorrhagic fever) nairovirus.
Bunyaviridae polymerases share with those of Orthomyxoviridae (e.g. influenza viruses) use of the mechanism of ‘cap-snatching’ for viral mRNA transcription, since, unlike the polymerases from non-segmented negative strand RNA viruses, they do not possess a capping activity. Cap-snatching involves binding of host capped mRNAs to the RNPs, cleavage of these RNAs close to the 5′ cap by a viral endonuclease activity and use of the short capped fragments as primers for viral mRNA transcription. This mechanism was first demonstrated for influenza virus polymerase [1]. An additional 11–15 nucleotides, heterogeneous in sequence, at the 5′ end of the viral mRNA prior to the start of the viral transcribed sequence was observed for snowshoe hare virus [2] and subsequently it was shown that La Crosse virions contain a primer-stimulated RNA polymerase and a methylated cap-dependent endonuclease [3], analogous to the situation found for influenza virus. Subsequently it has been shown that cap-snatching is employed by representative viruses of all five genera of Bunyaviridae [4], [5], [6]. Arenaviridae, another family of segmented RNA viruses, are also proposed to have a cap-snatching activity [7].
Although it is well known that the bunya- and arenavirus L-proteins contain in their central region the six polymerase motifs (designated preA, A–E) characteristic of negative-strand RNA viruses [8], [9], [10], the rest of the large protein is completely uncharacterised functionally and structurally, partly due to its lack of sequence homology with other proteins. Recently, crystallographic studies of functional domains of influenza virus polymerase, which is likely to be evolutionary related to the bunyavirus and arenavirus L-protein [8], have precisely defined the location and atomic structure of the two key domains for cap-snatching [11]. The mRNA cap-binding domain is located in the central region of the PB2 subunit [12], whereas the endonuclease activity resides in the N-terminal region of the PA subunit [13], [14]. We therefore asked the question whether this new structural information could aid localisation of domains relevant to cap-snatching in the bunyavirus L-protein? The influenza virus endonuclease domain has a core fold and divalent cation binding residues characteristic of the PD-(D/E)xK nuclease superfamily [15]. Unusually, it has a histidine as one of the metal ligands, which leads to a strong manganese preference for activity [13]. Surprisingly, a highly conserved motif (H....PD...D/E...K) at the extreme N-terminal region of diverse bunyavirus L-proteins with very similar features as now recognised to be important in the influenza PA N-terminal domain, was reported some time ago [8], [16] (Figure 1). This strongly suggested the presence of an endonuclease at the N-terminal of bunyavirus L-proteins. To investigate this further, we used the fact that the influenza endonuclease domain is about 200 residues [13] and made a synthetic gene comprising the first 250 residues of the La Crosse orthobunyavirus (LACV) L protein. Here we present biochemical and structural data that clearly show that the LACV L protein has a functional, manganese-dependent N-terminal endonuclease domain that indeed has very similar characteristics to that of influenza virus endonuclease. We also show that single point mutations that disable the nuclease activity in vitro, when introduced into the full-length L-protein, eliminate cap-dependent transcription in a LACV RNP reconstitution assay in cells. By sequence analysis we extend our results to show that all Bunyaviridae most likely possess such an endonuclease as well as members of other segmented RNA virus families, such as the bi-segmented Arenaviridae and four to six segmented Tenuiviruses. Implications for the evolution of segmented RNA viral polymerases are discussed as well as the prospects for a broad spectrum anti-viral targeting this endonuclease.
The original 1–250 residue construct of LACV L-protein (LC250) was truncated on the basis of partial proteolysis with papain in order to identify a minimal active and stable fragment that was well expressed and soluble. Papain resistant constructs with C-terminal residue 176, 180, 183, 186 and 190 were produced. The protein encompassing residues 1–180 (LC180) was biochemically characterised and found to be active as a nuclease (see below). The protein encompassing residues 1–183 (LC183) yielded hexagonal crystals which diffract to 2.1 Å resolution, with four molecules per asymmetric unit. The crystal structure was solved by the single anomalous dispersion (SAD) method using seleno-methionine substituted protein. A native data set was refined to an R-factor/R-free of 0.185/0.223 at 2.2 Å. This structure shows clearly one manganese ion bound with octahedral co-ordination in the active site cavity (designated site 1) (Supplementary Figure S1). A second structure, at 2.3 Å resolution (R-factor/R-free = 0.177/0.216), was obtained after soaking the crystals with the diketo acid inhibitor 2,4-dioxo-4-phenylbutanoic acid (DPBA). This is a member of the family of 4-substituted 2,4-dioxobutanoic acids which are known inhibitors of influenza virus endonuclease ([13], [17]). This structure clearly shows in addition to the manganese ion in site 1, a second in an adjacent site 2, with the inhibitor co-ordinating the two ions. The two ions are separated by 3.8 Å and have overlapping octahedral co-ordination (Supplementary Figure S1). In both cases, the identity of the manganese ions was indicated by anomalous scattering (Supplementary Figure S1).
The crystal structure of LC183 and its comparison with the N-terminal endonuclease domain of influenza virus polymerase PA subunit (PA-Nter, PDB entry 2W69 [13]) is shown in Figure 2. The secondary structure of LC183, together with a structural alignment of the N-terminal regions of selected orthobunya and tospoviruses L-proteins, is displayed in Figure 1. Comparison of Figures 2a and 2b shows that LC183 has a very similar alpha-beta topology to PA-Nter, although the helices are of significantly different lengths. Notably, the different position of PA-Nter helix αa and the increased length of helix αb gives LC183 a more slender, elongated shape with a more exposed active site that actually lies in a groove between two lobes (Supplementary Figure S2). Focussing in on the active site region, based around a four-stranded anti-parallel beta sheet, the similarity in structure is even more striking (Figure 3), despite essentially no sequence homology. As expected from the initial sequence analysis, LC183 has exactly the same core, cation-binding fold as found in PA-Nter and more generally in the PD-(D/E)xK nuclease superfamily [15]. This core region comprises 55 residues which can be superposed with a root-mean-square deviation of carbon alpha positions of 1.36 Å. Indeed there is a one-to one mapping between the ligands of the two metal binding sites: site 1 has ligands His34, Asp79, Asp92 and the carbonyl-oxygen of Tyr93 in LC183 corresponding to His41, Asp108, Glu119 and Ile120 in PA-Nter; site 2 has ligands Asp52 and Asp79 in LC183 corresponding to Glu80 and Asp108 in PA-Nter. Interestingly the putative catalytic lysine, characteristic of the PD-(D/E)xK nuclease superfamily, is likely to be Lys94 in LC183 (for confirmation, see below). As in EcoRV restriction enzyme (see [13] for a comparison of PA-Nter with EcoRV), this residue emerges from the central β-strand (βb) of the core fold rather than from helix αd as in the case of PA-Nter (Figure 3). Finally there is a clear correspondence between Lys108 and Lys137 in respectively LC183 and PA-Nter, both emerging from helix αd; in both cases this basic residue is in a position to potentially interact with a nucleic acid substrate. These similarities strongly suggest that LC183 will have a similar two-metal dependent nuclease activity to that of PA-Nter [13].
The inhibitor DPBA binds tightly to the two metal ions in the active site with three of its oxygen atoms replacing three water molecules in the two metal ion co-ordination (Figure 3c, Supplementary Figure S1). The phenyl group of the inhibitor is less well-defined in the electron density indicative of some residual rotational flexibility. This is indicative of the fact that no direct interactions are made between the DPBA and residues of the protein.
Despite the overall high degree of structural similarity of LC183 and PA-Nter, there are some significant differences. In the case of LC183, Asp52, one of the acidic ligands of cation site 2, is on a flexible loop. Indeed in the native structure, this loop is in an open conformation with Asp52 turned away from the active site and consequently only the manganese ion bound in site 1 is present (Figure 3c). In the inhibitor bound structure, the loop is in a closed conformation and Asp52 contributes to binding the second manganese. This suggests that there is preferential and tighter metal binding to site 1, consistent with it having four protein and two water ligands, and weaker metal binding to site 2, which has only two protein oxygen ligands. Furthermore metal binding to site 2 requires closure of the Asp52 loop and may be co-operatively dependent on binding of a nucleic acid substrate or metal binding inhibitor such as DPBA. In contrast, in PA-Nter there is no evidence of flexibility of the corresponding residue Glu80 and PA-Nter can in fact bind a single magnesium atom in site 2 only in the absence of manganese ions [14], [18]. Also, as mentioned above, in PA-Nter, helix α2 and the following loop are positioned to restrict substrate access to the active site cavity, whereas in LC183 the active site opens into a channel which could allow larger, more structured substrates to be cleaved (Supplementary Figure S2). These differences might account for some of the small discrepancies observed in nuclease activity between the two enzymes (see below).
Biochemical characterisation of the nuclease activity of LC180 was investigated using RNA and DNA digestion assays in the presence of a variety of divalent metal ions. Because of the structural similarity of the active sites of LC180 and PA-Nter, the experiments were guided by our previous work on influenza PA-Nter and used two of the same RNAs, a single-stranded, unstructured 51 nucleotide (nt) U-rich RNA and a highly structured 110 nt RNA, SRP Alu RNA as well as ssDNA [13]. Using 2 mM metal ions, Figure 4a shows that LC180 fully digests the U-rich RNA and partially digests the Alu RNA only in the presence of manganese, cobalt, zinc and nickel ions, with the preference Mn>Co≫Zn>Ni and not in the presence of magnesium, calcium or iron. Manganese dependent ssDNA endonuclease activity was observed for LC180 using a circular ssDNA, as for PA-Nter (Figure 4c). Since no digestion of RNAs was observed with 2 mM magnesium ions, increasing amounts of magnesium were tested. Weak nuclease digestion of the U-rich RNA was only observed at very high magnesium concentrations above 12.5 mM (Supplementary Figure S3). We next tested inhibition of the manganese dependent nuclease activity by the diketo acid DPBA. As in the case of PA-Nter, DPBA inhibits digestion of both test RNAs with an estimated IC50 of between 25–50 µM (Figure 5).
In parallel with these nuclease activity tests we measured the metal ion and inhibitor dependent thermal stability of LC180 by a Thermofluor assay [19], again in analogy to previous experiments described for PA-Nter [13]. The metal-free domain has an apparent melting temperature of 52.2 (±0.7)°C. Thermal stability is enhanced by 9°C upon addition of 2 mM manganese, presumably due to metal binding, with smaller increases with magnesium, calcium and cobalt (Figure 4b). A supershift of about 5°C in thermal stability is observed when DPBA is added to LC180 that is pre-bound with manganese and this supershift correlates with the nuclease inhibition affect of DPBA (Figure 5b), strongly suggesting that DPBA binds to the metal ions in the active site, as indeed observed in the crystal structure (Figure 3c).
The relative activity of influenza virus endonuclease (PA-Nter) and LC180 was compared under the same experimental conditions for both test RNAs with 2 mM of manganese, magnesium and calcium. Both enzymes are inactive with calcium, LC180 is more active with manganese and PA-Nter is active with 2 mM magnesium, whereas LC180 is not (Supplementary Figure S4). These experiments highlight three differences between LACV and influenza endonucleases. Firstly, influenza endonuclease is active in the presence of magnesium whereas LACV is not, secondly LACV is more active against the largely double stranded SRP Alu RNA and thirdly, LC180 is intrinsically more thermally stable with an apparent melting temperature of 52°C compared to 44°C for PA-Nter [13]. The more efficient activity against structured RNA could be due to the greater accessibility of the active site for LC183 as mentioned above (Supplementary Figure S2). It remains to be investigated whether there are any sequence preferences in the cleavage site favoured by the LACV endonuclease.
We made a series of alanine point mutants of key conserved residues in the active site of LC180 in order to assess their importance for activity. These included the ligands of metal 1 (His34, Asp92 and Asp79) and of metal 2 (Asp79 and Asp52), the putative catalytic lysine 94 and a second lysine (Lys108) close to the active site that is also highly conserved in bunya viruses and in PA-Nter. As a negative control, we also mutated Glu48, again conserved in all orthobunya viruses, which was not predicted from the structure to be directly involved in the nuclease activity. All mutant proteins were well expressed as for wild-type and purified as folded proteins as judged by behaviour on gel-filtration and in thermal stability assays (Figure 6). Only the H34A mutated protein was found to be somewhat less temperature stable, probably due to a loss of charge complementation in the highly acidic active site. A H34K mutant was made and assayed instead. The results of nuclease assays with these mutant proteins with the two RNAs and of thermal stability assays are shown in Figure 6a and 6b respectively. Mutations of any single of the four metal binding ligands (H34K, D52A, D79A and D92A) leads to elimination of nuclease activity, as does mutation of the catalytic lysine (K94A). The mutation E48A has no effect on activity and the mutation K108A leads to a reduction in activity, possibly because loss of the positively charged side-chain reduces substrate RNA binding. As described above, the wild-type protein shows significantly higher temperature stability upon binding of 2 mM manganese which is reinforced by subsequent binding of the inhibitor DPBA (Figure 4, Figure 5). Essentially the same pattern is shown by the mutants E48A (negative control, which however has somewhat reduced protein stability), K108A and K94A. This is consistent with the structural information which shows that none of these residues are directly involved in metal ligation or inhibitor binding. The D52A binding retains the manganese effect but has no inhibitor effect. Since D52 only binds metal 2, and as our structures show, a single manganese can bind strongly to metal site 1 (with the D52 loop in an open conformation, Figure 3c), we interpret this to imply that a single manganese ion bound in site 1 is sufficient to give the enhanced stability effect, whereas one ion is not sufficient to bind the inhibitor DPBA. The mutant D79A has no enhanced stability (in fact slightly reduced stability) in either the presence of manganese ions, with or without inhibitor, consistent with the fact that it binds simultaneously both metal ions. The H34K mutant has slightly increased stability compared to wild-type, probably because the lysine side-chain charge compensates the acidic active site better than the histidine, but there is no effect of manganese or inhibitor. This mutation thus almost certainly prevents any metal binding. Finally the mutation D92A shows a very modest manganese and inhibitor effect. It is thus possible that this mutant has still some low affinity for both metals. In summary these mutational studies show that the nuclease activity of LC180 depends critically on an intact binding site for two metal ions, preferably manganese, as well as the presence of the catalytic Lys94. Furthermore, binding of only one manganese ion is sufficient to lead to enhanced thermal stability, whereas both metal ions are required for DPBA binding. These results are fully consistent with the crystal structures described above.
To quantify the thermodynamics of binding of manganese to LC180 we used isothermal titration calorimetry in which manganese ions were titrated into wild-type or D52A mutant LC180 (see methods and Supplementary Figure S6). For wild-type protein the ITC data were fitted with a model comprising two independent sites yielding Kd's of 7.20 (±1.73) and 159.0 (±42.9) µM, although in the experiment saturation of the weaker binding site was not achieved. For the D52A mutant the ITC data were satisfactorily fitted with a model comprising a single site giving a Kd of 21.0 (±2.3) µM, with saturation of the single site being achieved. More complete results for the thermodynamic parameters of manganese binding are given in Supplementary Table S1. Once again these results are fully consistent with our structural and thermal stability experiments with the interpretation that the strongly bound ion for the wild-type and the single site for the D52A mutant (which have comparable affinities) corresponds to metal site 1 and the more weakly bound site for the wild-type corresponds to metal site 2. When magnesium was substituted for manganese no binding was detected by ITC. An analogous mutational and quantitative metal binding analysis has recently been performed for influenza virus endonuclease [20], with slight differences in behaviour being observed, as mentioned above.
To test the effect of the nuclease inactivating mutants in the context of the full-length LACV L-protein we used a previously described in vivo RNP reconstitution system in which a Renilla Luciferase (REN-Luc) reporter gene is used as a readout of cap-dependent transcription by the viral polymerase [21] (For a schematic outline of this assay see Supplementary Figure S7). From the in vitro work we know that the mutations do not disrupt the folding of the endonuclease domain and therefore presumably not of the full-length L-protein. Moreover, expression levels of full-length wild-type and mutant L constructs are comparable as detected by immunofluoresence (Supplementary Figure S8a). The transcription assay results with the various mutants (Figure 6c) parallel very closely the in vitro nuclease activity of the isolated LC180 domain mutants. Only the wild-type, negative control (E48A) and K108A (slightly reduced activity) L proteins give rise to significant REN-Luc production. To detect whether these active mutants are indeed producing capped mRNAs, we co-expressed them with the polio virus 2Apro protein. This protease specifically abrogates cap-dependent mRNA translation by cleaving eukaryotic initiation factor (eIF)4G [22]. The T7-driven expression constructs for the LACV L and N proteins, as well as the firefly luciferase (FF-Luc) transfection control escape this inhibition, since their translation is mediated by a viral internal ribosome entry site (IRES). As shown in Supplementary Figure S8b the Ren reporter activity of all active LACV L variants is drastically reduced upon co-expression of 2Apro, whereas the 2Apro mutant G60A, which has lost eIF4G cleaving activity [22], had no such effect. Moreover, IRES-driven FF-Luc expression was not affected by 2Apro, as expected (Supplementary Figure S8c). Thus, the specific sensitivity of L-protein driven Ren activity to the polio virus 2Apro indicates that wt L and both the E48A and the K108A mutant transcribe capped mRNAs.
Taken together, these results show that cap-dependent transcription is absolutely dependent on a functional two manganese-dependent nuclease activity at the N-terminus of the LACV L-protein, strongly suggesting that this domain is the cap-snatching endonuclease of the viral polymerase.
Cap-snatching as a method of priming transcription is uniquely restricted to segmented negative strand viruses, notably orthomyxoviruses (influenza), bunyaviruses and arenaviruses. The recent structural characterisation of two functional domains relevant for cap-snatching by influenza polymerase, the cap-binding domain and the endonuclease, in respectively the PB2 and PA polymerase subunits, raise the question as to whether similar domains exist in the L-protein (polymerase) of bunya- and arenaviruses. The work presented here shows unequivocally that the extreme N-terminal 200 residues of LACV has a cap-snatching endonuclease activity with very close structural and biochemical features to that of the N-terminal domain of the influenza virus polymerase PA subunit. We do not yet know the context of the bunyavirus N-terminal endonuclease within the 3-dimensional structure of the complete polymerase. However it is likely that there is a cap-binding domain and probably other RNA binding domains within the polymerase (this is certainly true for influenza polymerase) that enhance affinity and provide specificity for capped cellular mRNAs. Also it is possible that, as with influenza virus, there are allosteric effects that activate or make accessible the endonuclease active site only upon cap-binding.
We next examined whether the endonuclease signature could be identified in the L-protein of other segmented RNA viruses. Sequence analysis gives strong evidence that a homologous endonuclease domain exists at the N-terminus of the L-protein of four Bunyaviridae genera, orthobunya-, tospo, phlebo and hantaviruses, as well as tenuiviruses (which have four to six genome segments, [23], http://www.ncbi.nlm.nih.gov/ICTVdb/ICTVdB/00.069.0.01. Tenuivirus) and orthomyxovirus (Figure 7). In each case, the key metal binding and catalytic lysine residues can be identified. The sequence analysis shows that there are two sub-groups of these enzymes, with slightly different endonuclease signatures. Orthobunya- and Tospoviruses have the motif H....D...PD....DxK.....T, whereas Phlebo- and Hantaviruses have the motif H....E...PD....ExT.....K (although in Phleboviruses the first E is replaced by a D). The Hantavirus motif is identical to that found in orthomyxoviruses (Figure 7). The first version has a preference for aspartates and the catalytic lysine emerges from beta-strand βb, whereas the second version has a preference for glutamates and the catalytic lysine emerges from alpha helix αd (see Figure 3ab). Interestingly, the catalytic lysine interchanges with an absolutely conserved threonine at the two alternative positions (Figure 7). Nairoviruses are not included in this alignment as the location of the endonuclease is less certain. This genus of Bunyaviridae, which includes Crimean-Congo hemorrhagic fever virus, has an unusually long L-protein (about 4000 residues, compared to 2100–2900 for most other bunyaviruses). The N-terminal half of nairovirus L-proteins (i.e. prior to the polymerase motifs which start around residue 2050) contains a putative ovarian tumour (OTU)-like cysteine protease at the beginning [24], [25] as well as other predicted motifs and domains [10]. A putative endonuclease motif of the Phlebo/Hanta/Orthomyxo type exists in the residue range 630–710 (H(632)...PD(672)....E(686)F....K(699), numbering for Crimean-Congo virus) [10], but this needs to be confirmed by structural and functional data. It is interesting to note that the rice stripe tenuivirus also contains a predicted N-terminal OTU-like protease before the endonuclease motif [26]. It has been suggested that the protease might release the viral polymerase and one or more additional proteins by autoproteolytic cleavage and/or have de-ubiquitination activity [26]. Indeed de-ubiquitination activity of Crimean-Congo virus OTU domain has been shown to inhibit Ub- and ISG15-dependent antiviral pathways [27]. Arenavirus L-proteins have a highly conserved N-terminal region of about 200 residues that contains the absolutely conserved sequence of residues PD(89)...E(102)xF....K(122)L (alignment not shown, numbering for Lassa virus). This closely resembles the Phlebo/Hanta/Orthomyxo endonuclease motif, although the histidine is clearly lacking. Very recently, systematic alanine mutation of conserved charged residues in Lassa virus L-protein outside the polymerase motifs have been performed and the effect on transcription and replication have been tested in a RNP reconstitution system [28]. Seven charged residues in the N-terminal region, including Asp89, Glu102 and Lys122 and Asp129, were selectively important for mRNA synthesis but did not affect genome replication. The authors concluded from these results, combined with sequence similarities to type II endonucleases and influenza virus endonuclease, that this region of the L-protein was likely to be the cap-snatching endonuclease of arenaviruses, in full agreement with our analysis. Finally, the endonuclease signature is also clearly present in the L-proteins of two related but unclassified bunyaviruses (proposed to be called emaraviruses) which have four rather than the usual three genome segments, European mountain ash ringspot disease (Acc. No. YP003104764, [29]) and fig mosaic virus (Acc. No. CAQ03479, [30]). Both have the motif RH(105)D...PD(144)...E(158)xK(160) (numbering for mountain ash ringspot disease virus) and are thus most closely related to the Orthobunya and Tospoviruses, All these observations are summarised in Figure 8 which shows a schematic diagram of the architecture of polymerases from negative strand segmented RNA viruses.
It is well known that the 6 motifs characteristic of negative strand RNA-dependent RNA polymerases (pre-motif A and motifs A–E) are present in the central region of bunya and arenavirus L-proteins and in the PB1 subunits of orthomyxoviruses [8], [9], [10]. The current work shows that the extreme N-terminal region of bunya-, tenui- and arenavirus L-proteins functionally corresponds to the N-terminal region of the PA subunit of orthomyxoviruses. Given that the three influenza A polymerase subunits total 2252 residues, very similar to the size of many bunyavirus complete L-proteins and all these viral enzymes have common mechanisms of transcription (cap-snatching) and replication, a natural hypothesis that follows is that the L-proteins might be architecturally, structurally and functionally equivalent to a concatemer of the three influenza polymerase subunits in the order PA-PB1-PB2 (Figure 8). Some indirect support for the functional concatenation of the influenza polymerase subunits comes from the fact that the inter-subunit interactions are dominated by contacts between the C and N-terminal extremities of respectively PA and PB1 and PB1 and PB2 as visualised by recent crystal structures (reviewed in [11]). The most significant implication of this hypothesis is that the C-terminal third of the L-protein might be structurally and functionally equivalent to PB2, which contains the cap-binding domain required for cap-snatching. Unfortunately, this region of the L-protein is the least well conserved and there are no obvious cross-genera conserved motifs that could point to a putative cap-binding site similar to that described for influenza A PB2 subunit [12]. This is perhaps not surprising as the PB2-like subunits of, for instance, salmon anaemia and Quaranfil viruses, two non-influenza orthomyxoviruses, are highly diverged from influenza [31], [32], even though both these viruses appear to possess an endonuclease at the N-terminus of the PA subunit (Figure 7). Furthermore the fact that the distance of endonucleolytic cleavage from the 5′ cap is rather variable amongst cap-snatching viruses [6] suggests that the location of the cap-binding domain might vary. In fact, there is no clear proof that any L-protein directly binds capped RNAs and even some evidence that in hantaviruses the viral N-protein may play this role [33]. Clearly more experimental work is required to elucidate the complete mechanism of cap-snatching in bunya-, tenui- and arenaviruses and to validate or otherwise the hypothesis that L-proteins are architecturally equivalent to the concatenation of PA-PB1-PB2.
Finally it is important to note that for nearly two decades, influenza virus endonuclease has been targeted for anti-viral drug discovery and a number of specific endonuclease inhibitors have been described [17], [34], [35], [36], [37]. Most of these compounds implicitly target the two metal binding site of the endonuclease, which is also the target for many HIV integrase inhibitors [38] including the currently approved raltegravir [39], [40]. The recent structure determination of the endonuclease of influenza virus polymerase [13], [14] gives new impetus to structure-based optimisation of these inhibitors. The results described here show that bunyaviruses and arenaviruses, amongst which are several dangerous and emerging pathogens, contain a very similar endonuclease to influenza virus, which is also therefore a good target for anti-viral drug design. Indeed, the close similarities between influenza and bunyavirus endonucleases suggests that compounds targeting a broad spectrum of segmented negative strand RNA viruses could be envisaged. Our structure of DPBA bound to LACV endonuclease shows that this is indeed the case, although this compound is of low potency [17]. In addition this structure provides the first concrete proof that these compounds do indeed chelate the two divalent cations in the endonuclease active site.
The coding sequence of the N-terminal 250 residues of LACV-L (LC250) (UniProt accession code A5HC98) was optimised for expression in E. coli and synthesized (Geneart). A histidine tag followed by a linker and a TEV cleavage site (MGHHHHHHDYDIPTTENLYFQG) was added to the amino terminus of all protein constructs. All protein variants were amplified by PCR and cloned into a pET9a (Novagen) modified vector between NdeI 5′ and NotI 3′ sites for expression in E. coli. Mutagenesis of the proteins expressed in E. coli was performed on LC180. Mutant constructs were obtained by site directed mutagenesis using overlapping oligonucleotides and Pfu or KOD (Novagen) DNA Polymerases. The constructs pTM-LACV-L, pTM-LACV-N, pLACV-vRen, pCAGGs-T7 and pTM-FF-Luc used in the RNP reconstitution have been described previously [21]. The pTM1-based expression constructs for poliovirus 2APro wt and G60R mutant were kindly provided by Luis Carrasco, Universidad Autónoma de Madrid, Spain [22]. Mutagenesis of the cDNA for the RNP reconstitution experiments was performed by generating mutant DNA fragments by PCR and insertion into the KpnI/BmtI restriction sites of the pTM-LACV-L vector. In all cases the correctness of the DNA constructs were confirmed by DNA sequencing.
Proteins were expressed in Escherichia coli strain BL21 (DE3) in LB media with 25 µM kanamycin at 18°C overnight after induction with 0.2 mM of IPTG. Labelled protein was obtained by expressing LC183 protein in E. coli with M9 minimal medium and 50 mg/L of seleno-methionine. The cells were disrupted by sonication on ice for 3 minutes in lysis buffer (20 mM Tris-HCl pH 7.6, 150 mM NaCl, 2.5 mM β-mercapto-ethanol) with EDTA-free protease inhibitor cocktail (Roche). The protein from the soluble fraction was loaded onto a 5 ml Nickel column, washed with 10 volumes of lysis buffer with 50 mM imidazol and eluted with 5 volumes of 400 mM imidazol. The eluted protein was cleaved with histidine tagged TEV protease overnight at 4°C in dialysis against lysis buffer. After TEV cleavage all proteins have an additional glycine at the N-terminus. A second nickel column step was performed to remove unwanted material. The resulting untagged proteins were concentrated and purified by gel filtration chromatography using a SD75 column (Pharmacia) with lysis buffer for in vitro experiments or 20 mM HEPES pH 7.6, 150 mM NaCl, 2.5 mM β-mercapto-ethanol for crystallization trials. Influenza A/H3N2 endonuclease (PA 1–209) was obtained as described [13]. Purified LC proteins are contaminated with a small percentage of a degradation fragment of size LC163.
The length of the proteolytically stable amino terminal domain was defined from the LC250 purified protein by limited papain digestion with 1∶500 (w∶w) papain: protein ratio. Products were characterized by N-terminal sequencing and mass spectrometry. The resulting papain resistant fragments had molecular weights between 20.7 and 21.2 KDa corresponding to the first 175–178 residues of the LACV-L protein. Proteins LC176, 180, 183, 186 and 190 were subsequently produced. Finally, the protein construct LC180 was used for all in vitro biochemical experiments and LC183 for structural studies.
The influence of metal ion and DPBA binding on protein stability was measured by Thermofluor assays [19] at a protein concentration of 25 µM in lysis buffer and 2 mM concentration of various metal ions. For nuclease activity experiments, 12 µM of LC180 wild type and mutant proteins were incubated with 12 µM of Alu RNA (110 nucleotides of the Alu domain of Pyrococcus horikoshii SRP RNA) or 15 µM of 51 nucleotides U-rich RNA (5′-GGCCAUCCUGU7CCCU11CU19-3′) [13] at 37°C in the same buffer. The reaction was stopped by adding EGTA at a final concentration of 12 mM. Divalent cations were added to 2 mM final concentration. The reaction products were loaded onto 8 M urea, 15% acrylamide, Tris-borate gels and stained with methylene blue.
ITC experiments were performed using a high-precision VP-ITC titration calorimetric system (Microcal Inc., Northampton, MA). Binding experiments were performed with 60 µM of freshly purified LC180 protein at 25 C in the same buffer used for the nuclease activity assays. Titrations were made by injecting 15 µl of 1.8 mM or 3 mM MnCl2 into the LC180 D52A or wt respectively. For data analysis the heat produced by the metal ion dilution into the buffer was subtracted from the heat obtained in the presence of protein. The same procedure was performed with up to 12 mM of MgCl2 but gave no interaction signal. The binding isotherms were analyzed by non-linear least squares fitting (Microcal Origin software) using models corresponding to a single site or two independent sites for the D52A and the wt respectively. Thermodynamic values given are the average and standard deviation of at least two experiments.
Proteins LC176, 180, 183, 186 and 190 were expressed and tested for crystallization using a Cartesian nanovolume robotic system for screening. Only LC183 and LC186 crystallised and LC183 was used for all subsequent work. Crystals were obtained by mixing 1∶1 ratio protein: reservoir solution of 15–20 mg/ml LC183 protein in 20 mM HEPES pH 7.5, 150 mM NaCl, 5 mM MnCl2 and 2.5 mM β-mercapto-ethanol, and a reservoir composition of 3.4 M sodium formate, 0.1 M Tris-HCl at pH 8. The seleno-methionine LC183 crystals were obtained with a reservoir composition of 3.6 M Na-formate, 0.1 M HEPES pH 7. The dataset of the inhibitor-endonuclease complex was obtained after an overnight soaking of native crystals into reservoir buffer with 5 mM MnCl2, 10 mM MgCl2 and 5 mM of DPBA. The crystals were frozen in liquid nitrogen in the reservoir buffer with 30% glycerol for the selenomethionine labelled protein and with 30% glycerol, 5 mM MnCl2, 10 mM MgCl2 and 5 mM of DPBA for the inhibitor complex.
Crystals are of space-group P6122 with four molecules in the asymmetric unit. Selenomethionine derivative data were collected on a 180×160×140 µm3 crystal to 2.1 Å resolution on beamline ID29 at the European Synchrotron Radiation Facility (ESRF) at the selenium edge (X-ray wavelength 0.979 Å) for experimental phasing. Native and DPBA data were collected to 2.2 Å resolution on ID29 with wavelengths of 0.954 Å and 0.976 Å respectively. Data were processed and scaled with the XDS package [41] and subsequent analysis performed with the CCP4i package. Statistics of data collection and refinement are given in Table 1. The structure solution was obtained by the SAD method using autoSHARP [42] which found 16 anomalous sites, four (including a manganese site) for each of the four chains in the asymmetric unit. The resultant map was excellent and could be largely built automatically by ARP/wARP [43]. Refinement was performed with REFMAC [44] without applying non-crystallographic symmetry restraints. Extra density was observed for a single Mn2+ ion in the active site of each of the four molecules in the asymmetric unit as confirmed by strong anomalous scattering, even though the X-ray energy was well away from any manganese edge (Supplementary Figure S1). The loop containing Asp52 is either in the open position or partially open and intermediate. The structure of the complex with the inhibitor was solved by molecular replacement using PHASER [45] and the previously obtained model. Extra density was observed for a second Mn2+ and for the DPBA (Supplementary Figure S1). The loop containing Asp52 is in the closed position.
Sub-confluent monolayers of Huh7 cells in 12-well plates were transfected with 0.25 µg each of pLACV-vREN and pCAGGs-T7, 0.4 µg of pTM-LACV L (wild type or mutants) and pTM-LACV N, and 0.1 µg of pTM-FF-Luc using Nanofectin transfection reagent (PAA). In the negative control, the LACV-L expression plasmid was omitted from the transfection mix. An additional 0.2 µg of empty vector pTM1, or expression constructs pTM1-2APro or pTM1-2APro(G60R) were transfected in some experiments, as indicated. After transfection, cells were incubated for 24 h and lysed in 100 µl Dual Luciferase Passive Lysis Buffer (Promega). An aliquot of 20 µl of the lysate was assayed for FF-Luc and Ren-Luc activities as described by the manufacturer (Promega).
Structure figures were drawn with Molscript [46] or Bobscript [47] and rendered with Raster3d [48]. Sequence alignments were performed with ClustalW [49] and drawn with ESPript (http://espript.ibcp.fr/ESPript/cgi-bin/ESPript.cgi) [50]. Molprobity was used to analyse the quality of the structures (http://molprobity.biochem.duke.edu/).
The native structure of LC183 has wwPDB ID 2xi5 for the coordinate entry and r2xi5sf for the structure factors. The DPBA-bound form of LC183 has wwPDB ID 2xi7 for the coordinate entry and r2xi7sf for the structure factors.
|
10.1371/journal.pcbi.1003070 | Hybrid Equation/Agent-Based Model of Ischemia-Induced Hyperemia and Pressure Ulcer Formation Predicts Greater Propensity to Ulcerate in Subjects with Spinal Cord Injury | Pressure ulcers are costly and life-threatening complications for people with spinal cord injury (SCI). People with SCI also exhibit differential blood flow properties in non-ulcerated skin. We hypothesized that a computer simulation of the pressure ulcer formation process, informed by data regarding skin blood flow and reactive hyperemia in response to pressure, could provide insights into the pathogenesis and effective treatment of post-SCI pressure ulcers. Agent-Based Models (ABM) are useful in settings such as pressure ulcers, in which spatial realism is important. Ordinary Differential Equation-based (ODE) models are useful when modeling physiological phenomena such as reactive hyperemia. Accordingly, we constructed a hybrid model that combines ODEs related to blood flow along with an ABM of skin injury, inflammation, and ulcer formation. The relationship between pressure and the course of ulcer formation, as well as several other important characteristic patterns of pressure ulcer formation, was demonstrated in this model. The ODE portion of this model was calibrated to data related to blood flow following experimental pressure responses in non-injured human subjects or to data from people with SCI. This model predicted a higher propensity to form ulcers in response to pressure in people with SCI vs. non-injured control subjects, and thus may serve as novel diagnostic platform for post-SCI ulcer formation.
| Pressure ulcers are costly and life-threatening complications for people with spinal cord injury (SCI). To gain insight into the pathogenesis and effective treatment of post-SCI pressure ulcers, we constructed a computer simulation in a hybrid modeling platform which combines both equation- and agent-based models. The model was calibrated using skin blood flow data and reactive hyperemia in response to pressure and predicted a higher propensity to form ulcers in response to pressure in people with SCI vs. non-injured control subjects. The methodology we present in the paper may eventually be used as a novel platform to study post-SCI ulcer formation, as well as serving as a framework for other biological contexts in which agent-based models and mathematical equations can be integrated.
| In the United States, it is estimated that approximately 250,000 people live with spinal cord injury (SCI). Approximately 12,000 new cases occur each year [1], with total direct costs for treating all cases of SCI exceeding $7 billion annually [2], [3]. Pressure ulcers are common, costly and life-threatening complications for people with SCI. The prevalence of pressure ulcers in people with SCI is estimated to range from 8% to as high as 33% [4]. Post-SCI pressure ulcers are caused by a combination of impaired sensation, reduced mobility, muscle atrophy, as well as reduced vascularity and perfusion [5]. The current consensus is that pressure alone or pressure in combination with shear force cause localized injury to the skin and/or underlying tissue, usually over a bony prominence [6]. Several pathways have been identified for pressure/shear-induced ulceration, the major one being tissue ischemia.
Prolonged tissue ischemia may cause inflammation, necrosis, and the eventual formation of a pressure ulcer [7], [8]. Tissue inflammation is the common physiological reaction caused by tissue ischemia before necrosis occurs. We have focused our attention on this complex biological process. Inflammation is a central, modulating process in many complex diseases (e.g. sepsis, infectious disease, trauma, and wound healing), and is a central driver of the physiology of people with SCI [9]–[12]. However, inflammation is not an inherently detrimental process: properly regulated inflammation is required for successful immune response and wound healing [9], [13], [14]. Inflammation is a prototypical complex, nonlinear biological process that has defied reductionist, linear approaches [15]–[18]. Dynamic computational simulations, including ordinary differential equation (ODE)- and agent-based models (ABM), have been employed to gain insights into inflammation. These simulations have been useful in integrating mechanistic information and predicting qualitative and quantitative aspects of the inflammatory/wound healing response [19]–[22]. The purpose of the present study was to integrate blood flow data and the process of skin injury, inflammation, and healing using a hybrid model that combines ABM and ODE into a single computational model.
Agent-based modeling is an object-oriented, rule-based, discrete-event method of constructing computational models, and this technique can be used to model complex biological systems in which the behavior of individual components/agents, as well as pattern formation and spatial considerations are important [23]. Systems of ODE are well-suited for describing processes (or physiological responses) that can be approximated as well-mixed systems [22]–[26]. Modeling with differential equations (ordinary or partial) is the most widely used method of mathematical modeling. The main advantage of this approach is that there is a well-developed mathematical theory of differential equations which helps to analyze such equations and in some cases completely solve them [23], [25], [26]. To model a complex biological system as an ABM, the system is divided into small computational units (“agents”), with each agent obeying a set of rules that define the behavior of this agent. These simple rules, performed stochastically by agents in the model, lead to a complex, often emergent behavior of the system as a whole. In many cases, agents need only local information on the state of the system, rather than being affected by the global system state. As such, ABM's are particularly well suited to representing the transition between mechanisms at one scale of organization to behavior observed at another. The object-rule emphasis of an ABM greatly simplifies the process of model construction without loss of important features in the system, and also allows for modeling biological processes that are known to have both local and global features [23].
Our primary goal in this study was to gain translationally-useful insights into post-SCI pressure ulcer formation using dynamic, mechanistic computational modeling. However, several issues exist with the use of either ABM's or ODE's alone in modeling the pressure ulcer formation. It is difficult to analyze the output of ABM's in order to derive insights into qualitative regimes or primary drivers of outcome. In addition, simulating ABM's is more computationally intensive than simulating ODE-based models. On the other hand, real-life systems are often too complex to be modeled using only ODE, and the corresponding equation-based models may become too complicated to carry out practically useful results. Hybrid modeling is an emerging technique that involves combining diverse types of computational models into a single simulation [27]–[29]. In this approach, ODE can be used to define certain agent rules (low-level details), and ABM to describe the behavior of the high-level components of our system. In the present study, we utilized ODE to model properties tissue ischemia, and an ABM to model the stochastic, pressure-driven ulcer formation behavior in people with and without SCI. Using this approach, we find that a model calibrated with blood flow data predicted a higher propensity to form ulcers in response to pressure in SCI patients vs. non-injured control subjects.
The skin blood flow data used for computing the parameters of the differential equation model were collected from 12 adults (six with SCI and six without). This study was approved by the University of Pittsburgh Institutional Review Board (IRB# PRO08060015), and was carried out after obtaining informed consent from the participants. The age range of the subjects recruited for this study was 20–50 years old. The actual age in each group was: subjects with spinal cord injury (26, 27, 35, 35, 43, 48 years old); subjects without any neurological deficits (21, 25, 29, 35, 36, 44 years old). There was no statistically significant difference in age between the two cohorts of subjects (data not shown). For people with SCI, only those with ASIA [30], a scale for classification of spinal injury, grade A and B, one-year post-injury and non-ambulatory are recruited. The reactive hyperemic response was induced with 60 mmHg of pressure for 20 min on the sacral skin, with the participants lying on their stomach on a mat table. A laser Doppler probe was located at the center of the indenter to collect the skin blood flow. Instrumentation details are published previously [31]. A sample blood flow data collected in the experiment is demonstrated in Figure S1. The raw blood flow data of all tested subjects are provided in Dataset S1 and the plots of these data are shown in Dataset S2.
The hybrid model utilized in our study is comprised of an ABM of skin/muscle injury, inflammation, and ulcer formation along with an ODE model of blood flow and reactive hyperemia. The ABM portion of the model comprises interactions among oxygen, pro-inflammatory elements, anti-inflammatory elements, and skin damage, with realistic predictions of the pattern, size, and progression of pressure ulcers. All rules of this ABM were generated based on literature reviews and previously-described ABM's of diabetic foot ulcer formation [21] and simplified pressure ulcer formation [32]. The ODE portion of the model simulates the ischemia-induced reactive hyperemic response, and is derived from a previous circuit model [33]. Figure 1 shows the model representation of the pressure ulcer formation. Figures 2A&B depict the model components and their interactions within the hybrid model, with the solid rectangles, ellipses and arrows representing the components of the ABM portion and the dashed ellipse and arrows representing the components of the ODE portion of the model.
The ABM portion of the model is based on our previously-developed models [21], [32]. This ABM is a simplified model that simulates inflammation and reactive hyperemic response (as the result of applied pressure) in a small segment of tissue (epithelial cells in the model). We implemented this ABM in SPARK (Simple Platform for Agent-based Representation of Knowledge; freely downloadable at http://www.pitt.edu/~cirm/spark) [34], following an extensive process of literature search and creation of graphical diagrams that incorporate known biological influences [20], [35], [36]. From such diagrams and based on our prior work on modeling of the formation of diabetic foot ulcers [21], we constructed rules by which individual agents (e.g. cells or cytokines) interact with each other and bring about biological effects. The ABM portion of the model consists of key cells and diffusible inflammatory signals assumed to be involved in the process of formation of a pressure ulcer. A similarly parsimonious approach was used to construct the rules and relationship among agents, with the goal of generating a high-level view of the process of pressure ulcer formation. The components and inter-relationships among the agents and variables of the pressure ulcer ABM are presented in Figure 2. Importantly, our model adheres to our prior work on the importance of the positive feedback loop of tissue damage/dysfunction→inflammation→tissue damage/dysfunction [22], [25].
The main components of the ABM portion of the model are: structural/functional skin cell (nominally epithelial cells); inflammatory cells (nominally macrophages); blood vessels; an aggregate pro-inflammatory cytokine agent (nominally TNF-α); an aggregate anti-inflammatory/pro-healing cytokine (nominally TGF-β1); and oxygen.
These agents interact according to the following rules. Epithelial cells are damaged by applied pressure. A damaged epithelial cell produces TNF-α. Epithelial cells also are damaged by excessive amount of TNF-α. A severely damaged epithelial cell dies. An epithelial cell can be healed by TGF-β1, and the healing rate is proportional to the amount of oxygen at the position of the epithelial cell.
Macrophages are attracted by TNF-α, and they also produce TNF-α and TGF-β1. Each macrophage has a fixed lifespan (measured in simulation steps) and a macrophage dies after several simulation steps.
Blood vessels create new macrophages and release oxygen. The rate of macrophage production and oxygen release depends on the amount of blood flowing through a blood vessel. The ODE portion of the model (see below) is incorporated into blood vessel rules, which specify how the oxygen is produced. Blood flow depends on the pressure applied on a blood vessel. A blood vessel dies if the surrounding epithelial cells die.
There are also global model rules which specify how oxygen, TNF-α, and TGF-β1 diffuse and evaporate.
Physical pressure in ABM portion of the model is applied periodically. More specifically, the pressure is applied for a fixed period of time. The pressure is then released for the same amount of time, and the process repeats. A specific model parameter (called Pressure Interval) specifies the pressure time interval.
A detailed description of ABM rules and parameters is given in Text S1.
Ischemia-induced hyperemia (the reactive hyperemic response) is a sudden increase in skin blood flow following tissue ischemia [37]. Hyperemia is a normal physiological response that can be easily induced with non-damaging ischemic events, and it has been used in numerous fields to examine endothelial function [38] and vascular activity [39]. We incorporated an ODE model of reactive hyperemia into the pre-existing ABM of ulceration in order to link measurable parameters of reactive hyperemia to the process of ulceration induced by repeated cycles of pressure and ischemia/reperfusion. To do so, we adopted the ODE-based circuit model of de Mul et al [40]. These authors suggested that the reactive hyperemic response could be modeled as the circuit shown at Figure 3, with R (resistance) representing vascular resistance, C (capacitance) representing vessel compliance, V(t) representing the input blood flow pressure, and I (current) representing blood flow. I2(t) represents the skin blood flow (specifically, reactive hyperemia) as measured using a laser Doppler flowmetry system.
The ODE system derived from the circuit model has the following formNote that here we have only two differential equations for I1(t) and I2(t). I3(t), I4(t), I5(t), and I6(t) can be algebraically eliminated.
We are interested in modeling a situation when an occlusion occurs in the input blood flow due to application of an external pressure. De Mul et al [40] model such a situation by considering the following stepwise input blood flow functionHere V0 is the aortic pressure. Based on this expression of V(t), an explicit solution for I2(t) can be derived with initial conditions I1(0) = I2(0) = 0. This solution has the following formHere I2,rest, a, b, p1, and p2 are constants expressed in terms of R1, R2, R3, R4, C1, C2, V0.
We used this explicit solution for I2(t) for finding parameter values of the circuit model (the ODE portion of the model) based on available blood flow experimental data. In our agent-based simulations, the input blood pressure was a periodic function. In order to obtain the blood flow in these simulations, we used the ODE explicitly in our ABM.
The main components of SPARK models are Space, Data Layers, Agents, and the Observer [34]. Space is analogous to the physical space, and provides a context within which the model evolves. Data Layers provide a convenient way of tracking variables in space. Data layers update in time simultaneously at all positions. This is a computationally efficient way of handling processes such as diffusion and evaporation without employing an agent at each position to carry out the calculation. Agents can move, perform functions, interact with each other, and also interact with the space they occupy. Each agent has a set of behaviors and rules of action. The Observer contains information about space and all agents in the model. We extended SPARK with a feature for simple incorporation of ODE into an ABM. Epithelial cells, blood vessels, and macrophages were implemented as agents in SPARK. Oxygen, TNF-α, and TGF-β1 were implemented as data layers in SPARK. Pressure was implemented as a global model variable that periodically changes during the model simulation process.
The ODE portion of the model is integrated into the code of blood vessel agents. The following example shows how ODE's were added into SPARK-PL code:
equations
[
I4 = (V - R1 * I1)/R4
I3 = (V - R1 * I1 - R2 * I2)/R3
I5 = I1 - I2 - I4
I6 = I2 - I3
Dt I1 = (dV - I5/C1)/R1
Dt I2 = (I5/C1 - I6/C2)/R2
]
All variables in the example above are local variables of a blood vessel agent. Equations describe the evaluation of these variables in time. Each time step, the equation is integrated on the interval [t1, t1+dt], where t1 is the current simulation time and dt is the global parameter which specifies the time step. The output values of the equations are used in other rules defined for a blood vessel agent.
V represents the input blood pressure which is a periodic function in our simulations which depends on three parameters:Here, Vmax and Vmin represent maximal and minimal blood pressures respectively; Tp is the pressure interval parameter of the model; k = 0,1,2, etc; t is the number of simulation ticks. In other words, we set V = Vmin when the external pressure is applied and V = Vmax when the external pressure is released. The SPARK source codes of this hybrid model are provided in Dataset S3.
The ODE-based portion of the model was fit to data on blood flow for two different groups of subjects: a control group (CTRL) and an SCI group, as follows. We initially fixed parameters of the agent-based portion of the model. We chose these parameters based on a literature search. Only the approximate scale of parameters could be selected in this fashion, since our ABM is a simple, lumped-parameter model. With this set of parameter values, the ABM produces qualitative behavior commensurate with normal inflammation and wound healing [21].
Raw blood flow data was filtered with low pass filters. The filtered data were averaged over all six subjects in each group. Figures 4A and 4B depict the averaged reactive hyperemia blood flow data in people with and without SCI, respectively. We note that Figure 4A tend to oscillate more than Figure 4B. Depending on the level, and severity of injury, the reactive hyperemic response as measured with skin blood flow varied in people with SCI as compared to people without any neurological deficits. One main difference was the rate of increase and decrease in the skin blood flow of the reactive hyperemic response [41], in other words, one subject's peak blood flow may occur at 0.5 minute, and the other one may occur at 2.0 minute. With this variation, the blood flow oscillates more in Figures 4A as compared to Figures 4B. Another possible explanation is that, the skin blood flow as measured with the laser Doppler flowmetry system does oscillate naturally. When the skin blood flow signal was computed with Fourier transform, previous studies have identified that different frequency bands represent different physiological control mechanism of the blood flow [42]. Therefore the oscillation of skin blood flow is inevitable. We also note that the data in our simulation focused on the first 4 minutes. The interesting portion of the experimental data is the time when the peak blood flow occurs. We obtained approximately 10 minutes of raw data after releasing the pressure. The important information includes the time of the peak and the rate of decrease after the peak; both these values can be extracted from first 4 minutes after the pressure is released for all recorded data. We believe that it is simpler and more reliable to fit the ODE parameters based on the most important part of the experimental data (i.e. the first 4 minutes), since the rest of the data do not contain any important information for model fitting.
We then calibrated the ODE portion of the model based on the averaged data. Calibration was done using the following error function which measures the distance between actual (averaged) data and simulated results:Here i is the group index, i.e., i is either CTRL or SCI. Ei (p) is the error for the i-th group; y(p,k) is the value of the model function evaluated at the point k with the parameter vector p. Mi (k) is the averaged i-th data at the point k.
Calibration was performed using Matlab R2011 (The Mathworks, Inc., Natick, MA, USA). We used the explicit expression of I2(t) for finding best-fit parameters. The values of Vmax were assumed to be 85 mmHg for the control group and 75 mmHg for the SCI group, the same pressure values as in the experiments. For all other parameters, we defined possible lower and upper bounds. For the control group, we set 200 as the upper bound of all parameters, and 0.01 as the lower bound for all parameters except R4, for which we chose 190 as the lower bound since it is assumed that R4>>R1, R2, R3 [40]. Then we randomly selected 1000 points in the space of parameters and ran the standard Matlab minimization function fminsearch for all these initial points, and picked the best fit results. The search of best-fit parameters for the SCI group was carried out in a similar way. The only differences were that the value of Vmax = 75, and in addition we changed the upper bounds of C1 and C2 and set them equal to the best-fit values of C1 and C2 for the control group. This change was made to reflect the fact that C1,2SCI<C1,2CTRL [43].
Figures 5A and 5B show the best-fit simulation results, which minimize the error function Ei (p) in data from people with and without SCI, respectively. Table 1 lists the values of the best-fit parameters for both group with the ratios calculated in the Figure 6 to show the significant change of parameters for people with and without SCI. The results show that vascular resistance (R1) is significantly increased and that blood vessel compliance (C1, C2) is decreased in the SCI group by comparing with the control group.
We next sought to determine the behavior of our simulation under a more clinically realistic setting, in which pressure to tissues alternates with periods of pressure relief. We also sought to determine if, once partially calibrated with blood flow data from control vs. SCI subjects, our model would predict differential propensity to ulcerate between these two groups of patients. We simulated the application of medium-scale pressure on the skin with different frequencies, first applying a pressure on the skin for a given period of time (pressure interval), releasing the pressure for the same amount of time, and then repeating the process. Using the parameters obtained as described above, we ran the model simulations for both groups and compared the outcome. We ran the model for 2000 steps with various values of the pressure interval parameter. All other ABM parameters were fixed. We assumed Vmax = V0 (i.e., Vmax = 85 for the control group and Vmax = 75 for the SCI group) and Vmin = 40 for both groups.
We initially examined the minimal value of the pressure interval that would be predicted to result in substantial tissue damage (death of some epithelial cell agents). Figures 7A and 7B show the SPARK simulation results for control and SCI subjects. Green squares represent healthy epithelial cells, red squares represent damaged epithelial cells, red circles represent blood vessels, and blue circles represent macrophage. For the control group, the minimal value of the pressure interval was 205–210 simulation ticks (Figure 7A); in contrast, for the SCI group, the minimal value was 105–110 simulation ticks (Figure 7B). We also performed subject-specific fitting of the ODE parameters and measured the minimal value of the pressure interval resulting in substantial tissue damage for each subject. The results are given in Table 2. The average subject-specific value of the minimal pressure interval was 207 for control subjects and 168 for SCI subjects. These results agree qualitatively with our findings for the averaged data presented above: the minimal pressure interval is larger for the control group.
We next examined the predicted effect of turning frequency on control and SCI subjects. Figures 8A and 8B show how the predicted health of epithelial cells progresses over time for simulations of the control and SCI groups, respectively, over varying pressure on/off cycles. Increasing the frequency (or applying pressure for a short period of time and then subsequently relieving this pressure), we obtained an outcome in which a pressure ulcer did not form: when the simulated pressure is applied, the tissue is damaged somewhat, but when the pressure is relieved tissue health is restored. Also, simulated damage/dysfunction was predicted to increase more rapidly in the SCI group vs. the control group when the pressure interval was increased.
The components of the inflammatory response are time-driven, highly interconnected, and interact in a nonlinear fashion [15]–[18], [44]. The systems biology community has integrated mathematical and simulation technologies to understand complex biological processes [45]. More recently, we have suggested translational systems biology as a framework in which computational simulations are designed to facilitate in silico clinical trials, simulations are appropriate for in vivo and specifically clinical validation, and mechanistic simulations of whole-organism responses could guide rational therapeutic approaches [25].
Agent-based models have emerged as a useful complement to ODE-based models for elucidating complex biological systems, including inflammation, wound healing, angiogenesis, and cancer [19], [21], [23], [36], [46]–[49]. In the present study, we utilized a hybrid modeling approach that combines the both features of ODE and agent-based models. Using this approach, we integrate data regarding blood flow properties in SCI patients and compare them to data from control subjects. Our analysis suggests that, based on an abstraction of these blood flow properties and a stochastic model of tissue inflammation and ulcer formation, and in agreement with the literature [50], SCI patients are predicted to be more prone to ulceration. Our study, along with prior work [28], [51], [52], suggests that such hybrid modeling methodology could have a wide application in modeling complex, multiscale biological systems.
Despite the lack of sensation and motor function after SCI, several physiological changes at the chronic stage of SCI (more than 12 months since injury) increase a person's susceptibility to develop pressure ulcers, including changes in body composition (increased proportion of fatty tissue) and vascularity [5]. The linkage between changes in vascularity, epithelial function and pressure ulcer formation in people with SCI is not fully explored. Therefore, this pilot hybrid model was aimed at simulating pressure ulcer development by including a key vascular response (reactive hyperemia) observed in human subjects.
The goal of our previous research was to find the optimal turning frequency for patients with SCI [32]. The goal of the present model is the improvement of our previous model by coupling an ODE model of the reactive hyperemic response observed experimentally to an ABM based on rules derived from the literature. This model was capable of simulating the intensity in epithelial cell damage as a function of changes of amount and duration of localized pressure on the skin of people with and without SCI.
Results from the best-fit parameters of the circuit model set showed differences in vascular resistance (R1) and blood vessel compliance (C1, C2) between the two groups. The arterial resistance was bigger while the capillary resistance was smaller, respectively, in subjects with SCI as compared to controls. Changes in vascularity in people with SCI may be caused by denervation of sympathetic nervous system [53] as well as physical inactivity [54]. Our finding of increased vascular resistance in the arterial system was consistent with previous studies. With the loss of supraspinal control of the vascular system after high level of injury, people with SCI were reported to have increased vascular resistance in order to maintain the vascular tone by compensating for the loss of supraspinal sympathetic control [55]. Additionally, the increased vascular resistance may result from preservation of α-adrenergic tone. The increased vascular resistance could also result from vascular adaptation to deconditioning with the loss of motor function [56]. One prior study found that there was an increased activation of the receptor of the endothelin-1, which increases the vascular tone [56]. The results of decreased vascular resistance in the capillary system were not consistent with observations regarding vascular resistance in the arterial system. The capillary resistance was not investigated in previous studies; thus, our findings regarding vascular resistance in the arterial system may not be generalized to the capillary resistance, since the vascular resistance was measured with venous occlusion plethysmography in previous studies and the measurement was not directly on capillary blood flow. In addition, the measurement of reactive hyperemia in our study was at the lower back using an indenter, whereas the aforementioned previous studies measured this response at lower limbs with cuff. Future study on structural changes in capillary system and vascularity of the microcirculation might be beneficial in understanding the linkage to ulceration.
Results from the analysis of the best-fit parameters of the circuit model set also showed that the vessel compliance is smaller in people with SCI as compared to the controls. De Groot et al. found that the femoral artery compliance is smaller in individuals with SCI [43], and they suggested that this physiological change may be due to inactivity of the muscle since arterial compliance could be enhanced with functional electrical stimulation.
Our model validation studies suggest that the minimal amount of repeated pressure required to cause endothelial cell damage would be smaller in subjects with SCI. People with SCI are susceptible to ulcer formation, and there are several physiological changes that may contribute to the susceptibility of pressure ulcer development in this population. For example, people with complete SCI had decreased cross-sectional area of muscle fibers [57] and increased fat mass in lower limbs [58]. A recent study from Linder-Ganz et al. directly pointed out the relationship between physiological changes after injury and the pressure ulcer formation by using finite element model. They found that with the use of the same seat cushion, people with SCI had greater deep muscle stress as compared to controls [59]. To date, there is no study that investigated the direct linkage between changes in vascularity and ulcer formation in people with SCI. We were not aware of the underlying mechanism of changes in vascularity and the ulcer formation. However, from the rules and results of our model, it is indicated that changes in vascularity may play a role in decreased tolerance of pressure and endothelial function that leads to more severe damage with the same amount and duration of pressure.
There are several limitations of this study. This study only recruited limited numbers of subjects (six CTRL and six SCI), and people with SCI and controls were not matched for comparison. If additional subjects were used for the model calibration, the conclusion could be reached at a higher level degree of confidence. Though the ages of the subjects in the cohorts were not identical, there was no statistically significant difference with regard to age between the two groups of patients. In addition, previous studies [60], [61] found that the reactive hyperemic response was not different between healthy elderly population and healthy adults; these authors only found an impaired reactive hyperemic response among individuals in a hospitalized elderly population. Since there was no statistically significant difference in age between non-injured and SCI-injured subjects in our studies, and since all subjects recruited in our studies were healthy and not hospitalized during the time of the study, age is unlikely to be a significant factor in our data analysis. This is a pilot study developing this hybrid model of ulcer formation with different input of people with and without SCI. For a more realistic simulation, the ABM portion of the model could be expanded by incorporating additional physical and biological components, such as shear force and reperfusion injury, which may contribute to the formation of the pressure ulcer. Nevertheless, in this work, we present a first attempt to construct a biological model in a single computational platform where mathematical and agent-based models work in a seamless manner, and the result of the model reveals useful insight into the ulceration in people with and without SCI.
In conclusion, we used a hybrid approach combining ordinary differential equations related to blood flow along with an agent-based model of skin injury and subsequent inflammation in a single modeling platform, in order to investigate pathogenesis difference between people with SCI and without SCI in the process of ulcer formation. Our current finding suggests that people with SCI have higher propensity to form ulcers in response to pressure than non-injured control subjects.
|
10.1371/journal.pntd.0000322 | Tribendimidine and Albendazole for Treating Soil-Transmitted Helminths, Strongyloides stercoralis and Taenia spp.: Open-Label Randomized Trial | Tribendimidine is an anthelminthic drug with a broad spectrum of activity. In 2004 the drug was approved by Chinese authorities for human use. The efficacy of tribendimidine against soil-transmitted helminths (Ascaris lumbricoides, hookworm, and Trichuris trichiura) has been established, and new laboratory investigations point to activity against cestodes and Strongyloides ratti.
In an open-label randomized trial, the safety and efficacy of a single oral dose of albendazole or tribendimidine (both drugs administered at 200 mg for 5- to 14-year-old children, and 400 mg for individuals ≥15 years) against soil-transmitted helminths, Strongyloides stercoralis, and Taenia spp. were assessed in a village in Yunnan province, People's Republic of China. The analysis was on a per-protocol basis and the trial is registered with controlled-trials.com (number ISRCTN01779485). Both albendazole and tribendimidine were highly efficacious against A. lumbricoides and, moderately, against hookworm. The efficacy against T. trichiura was low. Among 57 individuals who received tribendimidine, the prevalence of S. stercoralis was reduced from 19.3% to 8.8% (observed cure rate 54.5%, p = 0.107), and that of Taenia spp. from 26.3% to 8.8% (observed cure rate 66.7%, p = 0.014). Similar prevalence reductions were noted among the 66 albendazole recipients. Taking into account “new” infections discovered at treatment evaluation, which were most likely missed pre-treatment due to the lack of sensitivity of available diagnostic approaches, the difference between the drug-specific net Taenia spp. cure rates was highly significant in favor of tribendimidine (p = 0.001). No significant adverse events of either drug were observed.
Our results suggest that single-dose oral tribendimidine can be employed in settings with extensive intestinal polyparasitism, and its efficacy against A. lumbricoides and hookworm was confirmed. The promising results obtained with tribendimidine against S. stercoralis and Taenia spp. warrant further investigations. In a next step, multiple-dose schedules should be evaluated.
| More than a billion people are infected with intestinal worms and, in the developing world, many individuals harbor several kinds of worms concurrently. There are only a handful of drugs available for treatment, and drug efficacy varies according to the worm species. We compared the efficacy of a single oral dose of tribendimidine, a new broad-spectrum worm drug from China, with the standard drug albendazole for the treatment of hookworm, large roundworm (Ascaris lumbricoides), whipworm (Trichuris trichiura) and, for the first time, Strongyloides stercoralis and tapeworm (Taenia spp.). Our single-blind randomized trial was conducted in a village in Yunnan province, southwest China. Both drugs showed high efficacy against A. lumbricoides and a moderate efficacy against hookworm. Among 57 tribendimidine recipients, the prevalence of S. stercoralis was reduced from 19.3% to 8.8%, and that of Taenia spp. from 26.3% to 8.8%. Similar prevalence reductions were noted among the 66 albendazole recipients. Taking into account additional infections only discovered at treatment evaluation, the difference between the drug-specific Taenia spp. net cure rates was highly significant in favor of tribendimidine. In view of our promising results, multiple-dose schedules with tribendimidine against S. stercoralis and Taenia spp. should be evaluated next.
| There is a growing awareness of the intolerable burden due to the so-called neglected tropical diseases [1]. Hence, new initiatives are underway for their control [2]. For helminth infections, the mainstay of control in high-burden areas rests on regular administration of anthelminthic drugs [1]–[5]. However, only a handful of drugs that have been developed many years ago are available [6], and there is considerable concern that resistance might develop, e.g., following repeated exposure of helminths to sub-curative doses. Reduced efficacy of common anthelminthic drugs is a major problem in veterinary medicine already, but for humans, it is of no clinical relevance thus far [7]. Another issue is that in a world where intestinal polyparasitism is common, yet neglected [8]–[12], and frequently used treatment regimens are only effective against a limited number of helminths [6], some species which are not effectively controlled by common drugs might increase in relative frequency, e.g., Strongyloides stercoralis and Taenia spp.
High prevalences of soil-transmitted helminths (Ascaris lumbricoides, hookworm, and Trichuris trichiura) and, consequently, a high level of intestinal multiparasitism, have recently been described from Nongyang, a settlement in Manguo administrative village, located in southwest Yunnan province, People's Republic of China [13]. Whilst S. stercoralis is endemic in the People's Republic of China [14], the local epidemiology of this parasite is not well understood. A prevalence of ∼15% was found in Nongyang [15]. Taeniasis and infections with the larval stage of Taenia solium that causes cysticercosis, have been documented throughout the People's Republic of China. Most infections occur in counties inhabited by non-Han nationalities whose traditional diets include the consumption of raw or undercooked meat [14],[16]. However, there is a paucity of epidemiologic data for taeniasis from the People's Republic of China, at least in the English literature [17]. In a recent cross-sectional survey, we found an egg-prevalence of Taenia spp. of 3.5% among 3220 individuals in Eryuan county, northwest Yunnan province [18]. In Nongyang, the prevalence of Taenia spp. was 5.1% [13].
The benzimidazoles, i.e., albendazole and mebendazole, are the most widely used drugs for the control of soil-transmitted helminthiasis [1],[19]. Both drugs have some effect against S. stercoralis and Taenia spp., but triple doses are recommended to achieve high cure rates [6],[20]. The drug of choice for treating S. stercoralis is ivermectin [21],[22]. Praziquantel and niclosamide are the recommended drugs against Taenia spp. [23]–[26]. Large-scale administration of ivermectin is the cornerstone of control programs targeting filarial infections, most notably onchocerciasis [27]. However, Onchocerca volvulus is not endemic in the People's Republic of China. Additionally, ivermectin is highly efficacious against A. lumbricoides, shows some activity against T. trichiura, but fails to cure hookworm infections. Since hookworms are common in the People's Republic of China, ivermectin is not commonly used for soil-transmitted helminth control in this country. These issues might explain why ivermectin is registered for human use in the People's Republic of China, but not readily available. Ivermectin, however, is produced at large scale for veterinary medicine, the bulk of which is exported.
Tribendimidine is an anthelminthic drug that has been registered in the People's Republic of China for use in humans [6],[28]. Tribendimidine is a symmetrical diamidine derivative of amidantel [28], its CAS registration number is 115103-15-6. Used at the current standard dose of 200 mg for children aged 5 to 14 years, and 400 mg for individuals aged ≥15 years, tribendimidine is safe and efficacious against A. lumbricoides, hookworm, and Enterobius vermicularis [28]. It also shows some activity against T. trichiura [29],[30], cestodes [28], and some trematodes [31],[32]. New research revealed in vitro and in vivo activity of tribendimidine against Strongyloides ratti [33].
The objective of this study was to assess the safety and efficacy of single-dose oral tribendimidine for treating intestinal helminth infections in a rural setting where polyparasitism is common, with a focus on the effects on S. stercoralis and Taenia spp. Comparison is made with single-dose oral albendazole as half of the participants were administered either drug. The primary outcome measures were the reduction of the infection prevalence of intestinal helminths, and the frequency and severity of adverse events. The secondary outcome measure was the Kato-Katz-derived egg count reduction of common soil-transmitted helminths. Multiple stool samples were collected before and after drug administration and examined by different diagnostic tools to enhance the diagnostic sensitivity.
The study was carried out in Nanweng, a village in Menghai county, Xishuangbanna prefecture, Yunnan province, People's Republic of China, from May to July 2007. Nanweng has 81 resident families and is situated on the slope of a mountain, 1650 m above sea level at 21.77 N latitude and 100.40 E longitude. The village is exclusively inhabited by the Bulang ethnic group. Its economic basis is provided by the surrounding tea plantations and the more distant, partially irrigated rice and other crop fields. Untreated tap water is delivered to every house but no sanitation facilities are available in the entire village.
The leader of the Menghai-based county Center for Disease Control and Prevention (CDC) briefed the village authorities about the study. Village leaders then informed the residents who were all invited to participate. Over a 3-week period, 20–30 families were enrolled weekly, and household as well as individual questionnaires that have been used before [15],[18] were administered. Children below the age of 15 years were assisted by their parents or legal guardians to answer the questions, and infants younger than 2 years were excluded from the study. Participants were asked to provide a large stool sample in pre-labeled collection containers. Filled containers were collected daily and exchanged by empty ones with the goal to obtain 3 stool samples per participant. The evaluation of the treatment efficacy commenced 2 weeks post-treatment and followed the same field procedures. Stool samples were collected over a 2-week period, again aiming to obtain 3 samples per study participant.
Stool samples were collected in the village between 07:00 and 09:00 a.m., transferred to the laboratory in Menghai city, and processed within a maximum of 6 hours of receipt. First, ∼10 g of stool were placed on a gauze which was embedded on a wire mesh in a glass funnel equipped with a sealable rubber tube, the so-called Baermann device [33]. The funnel was then filled with de-ionized water and illuminated from below with an incandescent bulb. Second, for the Koga agar plate test [34], ∼2 g of stool were placed in the centre of a 9 cm Petri dish with freshly prepared nutrient agar. Third, a single 41.7 mg Kato-Katz thick smear [35] was prepared on a microscope slide and helminth eggs were enumerated after a clearing time of 30–60 min.
The lowest 45 ml of the liquid in the Baermann funnel were drained after 2 h, centrifuged, and the sediment was examined for S. stercoralis larvae at low magnification (40×). Koga agar plates were inspected for helminth larvae at similar magnification after a 2-day incubation period at 28°C in a humid chamber. Subsequently, all plates were rinsed with 12 ml sodium acetate–acetic acid–formaline (SAF) solution [36], gently scraped, and the eluent was centrifuged. The sediment was examined for helminth larvae, i.e., S. stercoralis and hookworms. Helminth larvae were differentiated based on established anatomical criteria, i.e., considering the buccal cavity and genital primordium of first-stage (L1) S. stercoralis larvae (buccal cavity: short; genital primordium: prominent), and the tip of the tail of third-stage (L3) S. stercoralis larvae (tip of the tail: cut). Samples were considered positive if larvae were detected at any stage and by any test. Helminth eggs in the Baermann and Koga sediment were also noted.
Study participants who had submitted at least 1 stool sample were listed according to their identification number in ascending order. Subsequently, a random sequence of 0's and 1's was generated in Excel (Microsoft Corp.) by the study coordinator, and aligned with the list of eligible study participants. Individuals matched with a “0” were assigned albendazole, whereas those with a “1” were assigned tribendimidine. Albendazole and tribendimidine were purchased from Shanxi Hanwang Medicine Co. Ltd. (Han Zhong, People's Republic of China) and Shandong Xinhua Pharmaceutical Co. Ltd. (Zibo, People's Republic of China), respectively. Drugs were administered as single 200 mg oral dose for children aged 5 to 14 years, and 400 mg for individuals ≥15 years of age. Drugs at the appropriate dosage according to participants' age were packed in identical envelopes labeled with the participant's name only. Hence, our study was an open-label trial, i.e., participants did not know which drug they received. Participants were asked to avoid alcohol consumption on the day of drug administration. After dinner time, teams of fieldworkers visited the village, met the participants at home and asked them about signs of acute or chronic illness, and alcohol consumption. Women aged ≥14 years were asked about pregnancy. Drugs were handed out together with fresh bottled water to healthy, non-drunk and non-pregnant participants, and drug intake was observed. Those treated were asked to report any potential drug-related signs or symptoms including sleeping troubles to the accompanying medical doctor.
The study was approved by the institutional research commission of the Swiss Tropical Institute (Basel, Switzerland). Ethical clearance was obtained from the Ethics Committee of Basel (EKBB), Switzerland (reference no. 149/07) and the Ethics Committee of the National Institute of Parasitic Diseases, Chinese Center for Disease Control and Prevention (Shanghai, People's Republic of China). The trial is registered with controlled-trials.com under the registration number ISRCTN01779485. The study procedures, potential risks, and benefits were explained to the village leaders. After their consent to perform the study, field workers visited the homes of the selected families where detailed information was provided to all potential participants, and questions were answered. Emphasis was placed on voluntary participation and the option to quit the study at any moment without further obligation. Written confirmation that full information had been provided and individual participation was voluntary (informed consent) was obtained from the head of each participating household or a literary substitute (adult child or relative), and this procedure was approved by the above-mentioned ethical committees.
A single 200 mg (for children aged 5 to 14 years), or 400 mg (individuals aged ≥15 years) oral dose of albendazole was offered to those participants who were not eligible for randomization because they had failed to provide any stool sample. The assessment of their health status and the treatment procedures followed the same protocol as for study participants, but the treatment outcome was not assessed.
Locally-used remedies for Taenia spp. infection and ivermectin for treating S. stercoralis at the standard dose (200 µg/kg) were provided at the end-of-study follow-up. Finally, albendazole was provided to the village authorities for later distribution to untreated inhabitants and participants who still harbored active infections.
The target sample size was 130 individuals, based on the following assumptions: Prevalence of S. stercoralis: 20%; efficacy of tribendimidine and albendazole against S. stercoralis: 85% (similar to ivermectin at standard dosage [37]) and 0%, respectively, with a confidence level of 95% and a power of 80%.
The questionnaire data were double-entered and cross-checked in EpiData version 3.1 (EpiData Association; Odense, Denmark). The laboratory data were examined for internal consistency, and merged with the questionnaire data. Statistical analysis was done with STATA version 9.2 (StataCorp; College Station, USA). Our final study cohort consisted of individuals aged ≥5 years who had submitted at least 2 stool samples at baseline, did not suffer from any chronic or acute illness, had not drunk alcohol on the day of drug administration, for women were not pregnant, had taken the randomly assigned drug, and had again at least 2 stool samples examined at the end-of-study survey. Thus, statistical analyses only considered treated individuals with complete diagnostic data records, stratified by drug (see Figure 1, end points of the right arm). Multiple stool readings at baseline and follow-up were required in order to boost diagnostic sensitivity [13],[15].
The infection status was determined based on the pooled results from the different diagnostic methods (i.e., soil-transmitted helminths and Taenia spp.: all tests; S. stercoralis: Baermann and Koga agar plate tests). Pearson's χ2-test and Fisher's exact test, as appropriate, were used to assess the association between infection and demographic variables. Treatment outcomes by drug and the differences between albendazole and tribendimidine were explored by calculating drug-specific prevalence reductions, and analyzing the difference between the observed cure rates (2-sample test of proportions). The infection intensity for the common soil-transmitted helminths was determined based on the quantitative Kato-Katz thick smear readings, multiplied by a factor of 24. Infection intensities for individual study participants were obtained by calculating the arithmetic mean of the available egg counts. The arithmetic mean of the averaged egg counts among the infected yielded a summery measure of infection intensity among the infected population. The individual infection intensities were subsequently stratified according to the classification proposed by the World Health Organization (WHO) [38]. The effects of albendazole and tribendimidine on the infection intensities in the respective groups were analyzed using a paired t-test, and drug-specific infection intensity reductions were compared by a 2-sample test of proportions.
We counted 294 family members, aged above 2 years, in the 81 resident families. Another 60 individuals were recorded in the village registry but they had either left the village, were younger than 2 years, or refused to answer the questionnaire. Figure 1 shows that 106 (36%) of the eligible participants provided none (n = 57), or only a single (n = 49) stool sample of sufficient quantity to perform all diagnostic tests. The remaining 188 individuals (64%) were 5 to 87-year-old and among them, 17 could not be treated due to pregnancy (n = 8), ill health (n = 5) or other reasons (n = 4). The randomization of the 171 participants with complete parasitological baseline data who were eligible for treatment resulted in equal-sized groups for single-dose oral tribendimidine or albendazole administration, but stool sample submission at baseline and actual treatment rates were somewhat lower among tribendimidine recipients.
Of the 91 participants treated with albendazole, 2 to 3 sufficiently large stool samples were available from 66 individuals (73%) at the end-of-study follow-up. A similar stool sample submission rate was observed for the 80 tribendimidine recipients (71%). The final cohort consisted of 123 individuals who received either albendazole (n = 66) or tribendimidine (n = 57), and submitted at least 4 (2 pre- plus 2 post-treatment) sufficiently large stool samples to carry out the full range of diagnostic tests. While drop-out rates were similar for males and females (42.2% versus 41.5%), full participation ranged from 21% for 10 to 14-year-olds to 58% for those aged ≥40 years (data not shown).
Considering the joint results of the 2 to 3 Kato-Katz, Koga agar plate and Baermann tests, our final cohort showed high prevalences of T. trichiura (87.8%), hookworm (74.8%), and A. lumbricoides (72.4%). The prevalences of Taenia spp. and S. stercoralis were 26.0% and 17.9%, respectively.
Table 1 shows that Taenia spp. infections were significantly more prevalent among males (33.9%) than females (18.0%, χ2 = 4.01, degrees of freedom (df) = 1, p = 0.045) and increased with age, albeit not significantly (p = 0.163). An increase with age was also noted in the prevalence of S. stercoralis infections. No S. stercoralis were diagnosed among pre-school children and students as opposed to farmers (p = 0.121); Taenia spp. was also more common among farmers (27.3%) than pre-school children and students (8.3%, p = 0.292). Illiterates had a lower A. lumbricoides prevalence than those with formal education (50.0% vs. 79.2%, χ2 = 9.32, df = 1, p = 0.002), reflecting the generally lower prevalence among older age groups. Table 1 also summarizes the infection intensities for the common soil-transmitted helminths according to the Kato-Katz results. While most hookworm and T. trichiura infections were of light intensity, a considerable number of moderate and a few heavy infections were noted for A. lumbricoides.
As detailed in Table 2, single-dose oral albendazole and tribendimidine significantly reduced the prevalence of A. lumbricoides (albendazole: from 75.8% to 0; tribendimidine: from 68.4% to 5.3%; both p<0.001), and hookworm (albendazole: from 69.7% to 21.2%; tribendimidine: from 80.7% to 38.6%; both p<0.001). Whilst the difference between the drug-specific cure rates against A. lumbricoides showed borderline significance in favor of albendazole, there was no difference in the efficacy of the two drugs against hookworm. Although neither albendazole nor tribendimidine resulted in a significant reduction in the prevalence of T. trichiura, single-dose oral albendazole was significantly more efficacious than tribendimidine in curing T. trichiura (p = 0.014). Single-dose albendazole and tribendimidine significantly reduced the arithmetic mean egg counts among those who were infected at baseline (all p<0.05) except for tribendimidine administered to individuals with T. trichiura (p = 0.136). Both drugs decreased the mean egg counts of A. lumbricoides and hookworm to a similar extent, but mean T. trichiura egg counts declined significantly more following albendazole than tribendimidine treatment (p = 0.005).
Single-dose albendazole and tribendimidine resulted in prevalence reductions of S. stercoralis of 6.1% and 10.5%, respectively, which was not statistically significant (both p>0.05). The prevalence of Taenia spp. was reduced from 25.8% to 10.6% after albendazole administration (p = 0.024), whilst among the tribendimidine recipients, the prevalence was lowered from 26.3% to 8.8% (p = 0.014). Table 2 also shows that after administration of albendazole, S. stercoralis larvae could still be found among 7 of the previously 11 infected individuals (observed cure rate: 36.4%). Among those treated with tribendimidine, 6 out of 11 individuals were larvae-free (observed cure rate: 54.5%; difference: 18.1%, p = 0.394). The baseline Taenia spp. prevalence of 25.8% and 26.3% among those given albendazole and tribendimidine was reduced by 58.8% and 66.7%, respectively (difference: 7.9%, p = 0.645).
Taking into account S. stercoralis and Taenia spp. infections that had only been recognized at treatment evaluation (these infections were most likely missed pre-treatment due to the lack of diagnostic sensitivity of the available tests), the efficacy of the drugs was lower (Table 3). In both treatment groups, S. stercoralis was diagnosed in 2 individuals previously declared uninfected. Hence, the net cure rate was 18.2% for albendazole and 36.4% for tribendimidine recipients (difference: 18.2%, p = 0.338). The number of “new” Taenia spp. infections in the albendazole group diagnosed during follow-up equaled the number of recoveries (n = 10), resulting in a zero overall cure rate. Among tribendimidine recipients, only 2 additional Taenia spp. infections were found; the prevalence reduction showed borderline significance (−14.0%, p = 0.058). The net cure rate of Taenia spp. in the tribendimidine recipients was 53.3%, significantly different from the albendazole group (difference: 53.3%, p = 0.001).
No adverse events were mentioned by participants treated with single-dose oral albendazole. In the tribendimidine group, an 87-year-old woman reported mild sleeping disorders, headache, dizziness, and gastrointestinal symptoms, including a single episode of vomiting. This subject had a light infection with T. trichiura and hookworm at baseline. Upon treatment evaluation, the participant did not submit any further stool samples, and hence was excluded from the final cohort.
To our knowledge, this is the first investigation assessing the safety and efficacy of single-dose tribendimidine for treating S. stercoralis and Taenia spp. infections, and the first clinically-monitored use of tribendimidine in a setting with high rates of intestinal multiparasitism. Indeed, infections with one of the three main soil-transmitted helminths were found in 72.4–87.8% of the study participants, and only 4 of the 123 individuals in our final cohort (3.3%) harbored none of these helminths. The prevalence of S. stercoralis and Taenia spp. at baseline was 17.9% and 26.0%, respectively. Our study was an open-label trial, comparing a single oral dose of albendazole with that of tribendimidine, with both drugs administered at either 200 mg or 400 mg according to participants' age.
The final study cohort comprised less than 50% of those initially contacted. Whilst the cohort had a similar sex distribution than the total population of Nanweng village, it was considerably biased toward older age groups. We screened multiple stool samples for intestinal helminths and randomly assigned the participants to either albendazole or tribendimidine. Treatment outcome was assessed 2 to 4 weeks after dosing using multiple stool samples and a diversity of diagnostic approaches. The prevalence of S. stercoralis was not significantly reduced by either albendazole or tribendimidine, and no significant drug-specific difference was observed. Among individuals infected with Taenia spp. at baseline, the observed cure rates of 58.8% for albendazole and 66.7% for tribendimidine showed statistical significance (both p<0.05). During follow-up, however, additional infections were found, mainly Taenia spp. among those who had received albendazole, and the difference between the drug-specific overall cure rates was 53.3% (p = 0.001). The observed cestocidal effect of tribendimidine, for the time being, should rather be regarded as an indication of a possible activity than as a proof-of-concept. This is due to the obvious diagnostic challenges encountered in field-based clinical studies on large cestodes.
In the current trial, albendazole generally performed slightly better than tribendimidine in curing common soil-transmitted helminth infections. The most notable difference was seen with T. trichiura, confirming earlier observations that single-dose albendazole is somewhat more efficacious than single-dose tribendimidine against this nematode [6]. The observed cure rates against hookworm and T. trichiura following single-dose albendazole are rather low compared to the results of a recent meta-analysis of this and other WHO-recommended anthelminthics commonly used against common soil-transmitted helminth infections [39]. We speculate that this observation is rather reflecting the rigorous diagnostic approach employed than an unusually low susceptibility of local hookworm and T. trichiura to albendazole. For example, hookworm infections could not only be detected by the widely used Kato-Katz technique, but also by the Koga agar plate method. However, the low cure rates observed in this study should also be seen as a warning sign and call for monitoring of drug efficacy and the potential emergence of drug resistance [38].
The inclusion of only 123 individuals who met our sample submission requirements into the final study cohort reduced the reported compliance rate but increased the reliability of the results due to the increased overall sensitivity of the employed diagnostic methods [15]. Indeed, a lower prevalence was found among those 175 participants who had at least 1 stool sample analyzed, but the drug-specific efficacies were similar (data not shown).
The discovery of notable numbers of infections among those who were deemed negative before treatment can be explained by at least 2 mechanisms, or a combination thereof. First, it is conceivable that the baseline evaluation fell within the prepatent period of recent infections. Second, it is well known that parasitological diagnosis of both S. stercoralis [40] and Taenia spp. [25],[41] lacks sensitivity. For S. stercoralis, the main remedy for this challenge is screening of multiple stool specimens [21], whilst for Taenia spp., sensitive coproantigen enzyme-linked immunosorbent assay (ELISA) tests provide valuable alternatives [23],[25],[26],[41]. The current diagnostic ‘gold’ standard to confirm treatment success in taeniasis is the recovery of the tapeworm scolex. Alternatively, the absence of proglottids from stools and underwear over a period of 3 months also provides solid proof of cure. However, such extensive observation is usually only feasible in hospital settings. Re-infection after treatment can almost certainly be excluded for Taenia spp., and it is rather unlikely for S. stercoralis. We speculate that in our study, the limited sensitivity of the diagnostic tools was more significant since the 3 to 5-week period between baseline and follow-up investigation is rather short for any notable level of re-infections.
We are confident that our results are valid despite the imperfect sensitivity of the employed diagnostic tests, not least due to our rigorous sampling effort. This assumption is supported by the following observations. For S. stercoralis, the numbers of “new” infections at follow-up was similar in both treatment groups (both n = 2), thus reducing the observed cure rate but not affecting the overall conclusion that both drugs exhibit some effect at the employed dosage. A mathematical model [42] for the prediction of “true” prevalence further suggested an underestimation of the S. stercoralis prevalence by the employed procedures within the range actually observed in the present trial [15]. After a study involving extensive stool sample collection and analysis by the Baermann technique, Dreyer and co-workers [43] suggested that at least 4 stool samples need to be collected to accurately assess the S. stercoralis infection status, and that only those with at least 2 positive test results should be included in clinical drug trials. In our study, we only included those individuals who had at least 2 stool specimens examined with 2 different diagnostic approaches. Thus, at least 4, and ideally 6, results were available to judge the infection status of the participants both before and after drug administration. Among the 22 S. stercoralis positives at baseline, 6, 6, 3, 5, 1 and 1 individuals had 1, 2, 3, 4, 5 and 6 positive test results, respectively. The 10 arguably cured individuals had 1 (n = 3), 2 (n = 3), 3 (n = 2) and 4 (n = 2) positive baseline test results. Our findings indicate that participants with only 1 or 2 positive tests at baseline were not more likely to be considered cured at treatment evaluation than those with multiple positive tests. However, infections were still found in all participants with 5 or 6 positive tests at baseline. The four individuals who were only found to be infected at treatment evaluation then had 1, 1, 2 and 3 positive test results. Finally, the observed activity of albendazole against Taenia spp. among those who were found to be infected at baseline has to be put into perspective with the high number of “newly” detected infections at treatment evaluation. Among tribendimidine recipients, only few additional Taenia spp. infections were found, indeed indicating that single-dose tribendimidine, but not albendazole, might have some effect against Taenia spp.
Unfortunately, the eggs of T. saginata, T. solium, and T. asiatica cannot be readily distinguished microscopically [23],[25]. Hence, we are not in a position to determine their relative frequency in our study population. However, the reported and observed diets suggest that the locally dominant species is T. solium or possibly T. asiatica since Bulang favor raw pork over raw beef.
As a next step, the efficacy of multiple-dose tribendimidine could be assessed as our results indicate some, albeit currently unsatisfactory effect of this drug against S. stercoralis and Taenia spp. In future studies with a focus on these 2 parasites rather than the common soil-transmitted helminths, the reference drug should be praziquantel or niclosamide for Taenia spp., and ivermectin for S. stercoralis. Alternatively, triple-dose albendazole might be used as reference treatment [20]. When further investigating the efficacy of tribendimidine against large cestodes, including Taenia spp. in humans, we propose to treat a small group of confirmed taeniasis cases who agreed to submit multiple stool samples and observe proglottids in their stools and underwear over extended time periods.
Infections with A. lumbricoides and hookworm are the main targets for single-dose mass chemotherapy using albendazole or mebendazole. Discussions are underway in the People's Republic of China for the larger-scale use of tribendimidine. Efficacy of the latter drug on other intestinal parasites would be of considerable public health significance, which is explained by the geographic overlap of different helminth infections, including S. stercoralis and Taenia spp. Treatment of individuals with multiple species parasite infections, including S. stercoralis and Taenia spp., is likely to occur. Hence, there is a pressing need to determine the most efficacious tribendimidine treatment regimen for S. stercoralis and Taenia spp. since the exposure of the parasites to sub-curative doses exacerbates the risk of resistance development. Therefore, pharmacovigilance needs to also cover non-target parasites to assure timely detection of emerging resistance.
|
10.1371/journal.pcbi.1007263 | Identifying nonlinear dynamical systems via generative recurrent neural networks with applications to fMRI | A major tenet in theoretical neuroscience is that cognitive and behavioral processes are ultimately implemented in terms of the neural system dynamics. Accordingly, a major aim for the analysis of neurophysiological measurements should lie in the identification of the computational dynamics underlying task processing. Here we advance a state space model (SSM) based on generative piecewise-linear recurrent neural networks (PLRNN) to assess dynamics from neuroimaging data. In contrast to many other nonlinear time series models which have been proposed for reconstructing latent dynamics, our model is easily interpretable in neural terms, amenable to systematic dynamical systems analysis of the resulting set of equations, and can straightforwardly be transformed into an equivalent continuous-time dynamical system. The major contributions of this paper are the introduction of a new observation model suitable for functional magnetic resonance imaging (fMRI) coupled to the latent PLRNN, an efficient stepwise training procedure that forces the latent model to capture the ‘true’ underlying dynamics rather than just fitting (or predicting) the observations, and of an empirical measure based on the Kullback-Leibler divergence to evaluate from empirical time series how well this goal of approximating the underlying dynamics has been achieved. We validate and illustrate the power of our approach on simulated ‘ground-truth’ dynamical systems as well as on experimental fMRI time series, and demonstrate that the learnt dynamics harbors task-related nonlinear structure that a linear dynamical model fails to capture. Given that fMRI is one of the most common techniques for measuring brain activity non-invasively in human subjects, this approach may provide a novel step toward analyzing aberrant (nonlinear) dynamics for clinical assessment or neuroscientific research.
| Computational processes in the brain are often assumed to be implemented in terms of nonlinear neural network dynamics. However, experimentally we usually do not have direct access to this underlying dynamical process that generated the observed time series, but have to infer it from a sample of noisy and mixed measurements like fMRI data. Here we combine a dynamically universal recurrent neural network (RNN) model for approximating the unknown system dynamics with an observation model that links this dynamics to experimental measurements, taking fMRI data as an example. We develop a new stepwise optimization algorithm, within the statistical framework of state space models, that forces the latent RNN model toward the true data-generating dynamical process, and demonstrate its power on benchmark systems like the chaotic Lorenz attractor. We also introduce a novel, fast-to-compute measure for assessing how well this worked out in any empirical situation for which the ground truth dynamical system is not known. RNN models trained on human fMRI data this way can generate new data with the same temporal structure and properties, and exhibit interesting nonlinear dynamical phenomena related to experimental task conditions and behavioral performance. This approach can easily be generalized to many other recording modalities.
| A central tenet in computational neuroscience is that computational processes in the brain are ultimately implemented through (stochastic) nonlinear neural system dynamics [1–3]. This idea reaches from Hopfield’s [4] early proposal on memory patterns as fixed point attractors in recurrent neural networks, working memory as rate attractors [5,6], decision making as stochastic transitions among competing attractor states [7], motor or thought sequences as limit cycles or heteroclinic chains of saddle nodes [8,9], to the role of line attractors in parametric working memory [10–12], neural integration [13], interval timing [14], and more recent thoughts on the role of transient dynamics in cognitive processing [15]. To test and further develop such theories, methods for directly assessing system dynamics from neural measurements would be of great value.
Traditionally, mostly linear approaches like linear (Gaussian or Gaussian-Poisson) state space models [16–19], Gaussian Process Factor Analysis [GPFA; 20], Dynamic Causal Modeling [DCM; 21], or (nonlinear, but model-free) delay embedding techniques [22,23], have been used for reconstructing state space trajectories from experimental recordings. While these are powerful visualization tools that may also give some insight into system parameters, like connectivity [21], linear dynamical systems (DS) are inherently very limited with regards to the range of dynamical phenomena they can produce [e.g. 24]. The representation of smoothed trajectories in the latent space may still inform the researcher about interesting aspects of the dynamics, but the inferred latent model on its own is not powerful enough to reproduce many interesting and computationally important phenomena like multi-stability, complex limit cycles, or chaos [24,25]. More formally, given experimental observations X = {xt} supposedly generated by some underlying latent dynamical process Z = {zt} (Fig 1), linear state space models may yield a useful approximation to the posterior p(Z|X), but–due to their linear limitations–they will not produce an adequate mathematical model of the prior dynamics p(Z) that could generate the actual observations via p(X|Z).
In contrast, recurrent neural networks (RNNs) represent a class of nonlinear DS models which are universal in the sense that they can approximate arbitrarily closely the flow of any other dynamical system [26–28]. Hence, RNNs are, in theory, sufficiently powerful to emulate any type of brain dynamics. Based on previous work embedding RNNs into a statistical inference framework [29,30], we have recently developed a nonlinear state space model utilizing piecewise-linear RNNs (PLRNNs) for the latent dynamical process [31]. In state space models, similar to sequential variational auto-encoders (VAE) [32–34], one attempts to infer the system parameters θ by maximizing a lower bound on the log-likelihood log p(X|θ). In contrast to many other RNN-based approaches [30,35], including current sequential VAE implementations [36], our method returns a set of neuronally interpretable and partly analytically tractable dynamical equations that could be used to gain further insight into the generating system.
The present work further advances this powerful methodology along three major directions: First, we develop a stepwise initialization and training scheme that forces the latent PLRNN model toward the correct underlying dynamics: Good prediction of the time series observations and informative smooth latent trajectories may be achieved even without recreating a sufficiently good approximation to the underlying DS (as evidenced by the success of linear state space models). Through a kind of annealing protocol that places increasingly more burden of predicting the observations onto the latent process model, we enforce the correct dynamics. Second, we show that a Kullback-Leibler divergence defined across state space between the prior generative model dynamics p(Z) (independent of the observations) and the inferred latent states given the observations, p(Z|X), provides a good measure for how well the reconstructed DS (emulated by the PLRNN) can be expected to have captured the correct underlying system. Hence, our approach, rather than just inferring the latent space underlying the observations, attempts to force the system to capture the correct dynamics in its governing equations, and provides a quantitative sense of how well this worked for any empirically observed system for which the ground truth is not known. Third, given that fMRI is likely the most important non-invasive technique for gaining insight into human brain function in healthy subjects and psychiatric illness, we provide an observation (‘decoder’) model for the PLRNN that takes the hemodynamic response filtering into account.
We start by introducing our nonlinear state space model (SSM) and statistical inference framework [originally developed in 31]. Within a SSM, one aims to predict observed experimental time series xt∈RN from a set of latent variables zt∈RM (where usually M≠N) and their temporal dynamics. Here we use a piecewise-linear (or, strictly, piecewise-affine) recurrent neural network (PLRNN) (i.e., a RNN composed of rectified-linear units [ReLUs]) for modeling the unknown latent dynamics:
zt=Azt−1+Wφ(zt−1)+h+Cst+εt,εt∼N(0,Σ)
(1)
z1∼N(μ0+Cs1,Σ)
where zt is the latent state vector at time t = 1…T, A∈RMxM is a diagonal matrix of (linear) auto-regression weights, W∈RMxM an off-diagonal matrix of connection weights, and φ(zt) = max(zt,0) is an (element-wise) ReLU transfer function. st∈RK denotes time-dependent external inputs that influence latent states through coefficient matrix C∈RMxK, and εt is a Gaussian white noise process with diagonal covariance matrix Σ. (The basic model was modified from Durstewitz [31] to enable efficient estimation of bias parameters h and speeding up the inference algorithm by orders of magnitude.) The diagonal and off-diagonal structure of A and W, respectively, help to ensure that system parameters remain identifiable. Although here we advance model (Eq 1) mainly as a tool for approximating unknown dynamical systems, it may be interpreted as a neural rate model [e.g. 37,38], with A the units’ passive time constants, W the synaptic coupling matrix, and φ(z) a current/voltage to spike rate transfer function which for cortical pyramidal cells is often non-saturating and close to a ReLU within the physiologically relevant regime [e.g. 39].
The observed time series are generated from the ReLU-transformed latent states (Eq 1) through a linear-Gaussian model:
xt=Bφ(zt)+ηt,ηt∼N(0,Γ)
(2)
where xt are the observed N-dimensional measurements at time t generated from zt, B∈RNxM is a matrix of regression weights (factor loadings), and ηt denotes a Gaussian white observation noise process with diagonal covariance matrix Γ.
Thus, the model is specified by the set of parameters θ = {μ0,A,W,C,h,B,Γ,Σ}, and we are interested in recovering θ as well as the posterior distribution p(Z|X) over the latent state path Z = {z1:T} from the experimentally observed time series X = {x1:T} and experimental inputs S = {s1:T}. In the following, we will sometimes use the notation θlat = {μ0,A,W,C,h,Σ} and θobs = {B,Γ} to exclusively refer to parameters in the evolution or observation equation, respectively.
An appealing feature of the SSM framework is that different measurement modalities and properties can be accommodated by connecting different observation models to the same latent model. In order to apply our model to fMRI time series, we need only to adapt observation Eq 2 to meet the distributional assumptions and temporal filtering of the blood-oxygen-level dependent (BOLD) signal, while retaining process Eq 1 with its universal approximation capabilities. In contrast to electrophysiological measurements, BOLD time-series are a strongly filtered, highly smoothed version of some underlying neural process, only accessible through the hemodynamic response function (HRF) [e.g. 40]. Hence, we modified the observation model (Eq 2) such that the observed BOLD signal is generated from the latent states (Eq 1) through a linear-Gaussian model with HRF convolution:
xt=B(hrf*zτ:t)+Jrt+ηt,ηt∼N(0,Γ)
(3)
where xt are the observed BOLD responses in N voxels at time t generated from zτ:t (concatenated into one vector and convolved with the hemodynamic response function). We also added nuisance predictors rt∈RP, which account for artifacts caused, e.g., by movements. J∈RNxP is the coefficient matrix of these nuisance variables, and B,Γ and ηt are the same as in Eq 2. Hence, the observation model takes the typical form of a General Linear Model for BOLD signal analysis as, e.g., implemented in the statistical parametric mapping (SPM) framework [40]. Note that while nuisance variables are assumed to directly blur into the observed signals (they do not affect the neural dynamics but rather the recording process), external stimuli presented to the subjects are, in contrast, assumed to exert their effects through the underlying neuronal dynamics (Eq 1). Thus, the fMRI PLRNN-SSM (termed ‘PLRNN-BOLD-SSM’) is now specified by the set of parameters θ = {μ0,A,W,C,h,B,J,Γ,Σ}. Model inference is performed through a type of Expectation-Maximization (EM) algorithm (see Methods and full derivations in supporting file S1 Text).
One complication here is that the observations in Eq 3 do not just depend on the current state zt as in a conventional SSM, but on a set of states zτ:t across several previous time steps. This severely complicates standard solution techniques for the E-step like extended or unscented Kalman filtering [41]. Our E-step procedure [cf. 31], however, combines a global Laplace approximation with an efficient iterative (fixed point-type) mode search algorithm that exploits the sparse, block-banded structure of the involved covariance (inverse Hessian) matrices, which is more easily adapted for the current situation with longer-term temporal dependencies (see Methods sect. ‘Model specification and inference’ & S1 Text for further details).
The EM-algorithm aims to compute (in the linear case) or approximate the posterior distribution p(Z|X) of the latent states given the observations in the E-step, in order to maximize the expected joint log-likelihood Eq(Z|X)[log pθ(Z,X)] with respect to the unknown model parameters θ under this approximate posterior q(Z|X)≈p(Z|X) in the M-step (by doing so, a lower bound of the log-likelihood log p(X|θ)≥Eq[log p(Z,X)]−Eq[log q(Z|X)] is maximized, see Methods sect. ‘Parameter estimation’ & S1 Text). This does not by itself guarantee that the latent system on its own, as represented by the prior distribution pθlat(Z), provides a good incarnation of the true but unobserved DS that generated the observations X. As for any nonlinear neural network model, the log-likelihood landscape for our model is complicated and usually contains many local modes, very flat and saddle regions [42–45]. Since Eq[log p(Z,X)] = Eq[log p(X|Z)]+ Eq[log p(Z)], with the expectation taken across q(Z|X)≈p(Z|X)∝p(X|Z)p(Z), the inference procedure may easily get stuck in local maxima in which high likelihood values are attained by finding parameter and state configurations which overemphasize fitting the observations, p(X|Z), rather than capturing the underlying dynamics in p(Z) (Eq 1; see Methods for more details). To address this issue, we here propose a step-wise training by annealing protocol (termed ‘PLRNN-SSM-anneal’, Algorithm-1 in Methods) which systematically varies the trade-off between fitting the observations (maximizing p(X|Z); Eqs 2 and 3) as compared to fitting the dynamics (p(Z); Eq 1) in successive optimization steps [see also 46]. In brief, while early steps of the training scheme prioritize the fit to the observed measurements through the observation (or ‘decoder’) model p(X|Z) (Eqs 2 and 3), subsequent annealing steps shift the burden of reproducing the observations onto the latent model p(Z) (Eq 1) by, at some point, fixing the observation parameters θobs, and then enforcing the temporal consistency within the latent model equations (as demanded by Eq 1) by gradually boosting the contribution of this term to the log-likelihood (see Methods).
We examined the performance of this annealing protocol in terms of how well the inferred model was capable of recovering the true underlying dynamics of the Lorenz system. This 3-dimensional benchmark system (equations and parameter values used given in Fig 4 legend), conceived by Edward Lorenz in 1963 to describe atmospheric convection [47], exhibits chaotic behavior in certain regimes (see, e.g., Fig 4A). We measured the quality of DS reconstruction by the Kullback-Leibler divergence KLx(ptrue(x),pgen(x|z)) between the spatial probability distributions ptrue(x) over observed system states in x-space from trajectories produced by the (true) Lorenz system and pgen(x|z) from trajectories generated by the trained PLRNN-SSM (KLx, in the following refers to this divergence evaluated in observation space, see (Eq 9) in Methods, where KL˜x denotes a normalized version of this measure; see Fig 1 and Methods sect. ‘Reconstruction of benchmark dynamical systems’ for details). Hence, importantly, our measure compares the dynamical behavior in state space, i.e. focuses on the agreement between attractor (or, more generally trajectory) geometries, similar in spirit to the delay embedding theorems (which ensure topological equivalence) [48–50], instead of comparing the fit directly on the time series themselves which can be highly misleading for chaotic systems because of the exponential divergence of nearby trajectories [e.g. 51], as illustrated in Fig 2A. Note that for a (deterministic, autonomous) dynamical system the flow at each point in state space is uniquely determined [e.g. 24] and induces a specific spatial distribution of states, in this sense translates aspects of the temporal dynamics into a specific spatial geometry. Fig 2B gives examples where our measure KL˜x correctly indicates whether the Lorenz attractor geometry was properly mapped by a trained PLRNN, while a direct evaluation of the time series fit (incorrectly) indicated the contrary.
For evaluating our specific training protocol (termed ‘PLRNN-SSM-anneal’, Algorithm-1 in Methods), trajectories of length T = 1000 were drawn with process noise (σ2 = .3) from the Lorenz system and handed to the inference algorithm (for statistics, a total of 100 such trajectories were simulated and model fits carried out on each, and a range of different numbers of latent states, M = {8, 10, 12, 14}, was explored). Models were trained through ‘PLRNN-SSM-anneal’ and compared to models trained from random initial conditions (termed ‘PLRNN-SSM-random’) in which parameters were randomly initialized (see Fig 3).
In general, the PLRNN-SSM-anneal protocol significantly decreased the normalized KL divergence KL˜x (Eq 9) and increased the joint log-likelihood when compared to the PLRNN-SSM-random initialization scheme (see Fig 3A and 3B, independent t-test on KL˜x: t(686) = -16.3, p < .001, and on the expected joint log-likelihood: t(640) = 11.32, p < .001). More importantly though, the PLRNN-SSM-anneal protocol produced more estimates for which KL˜x was in a regime in which the chaotic attractor could be well reconstructed (see Fig 4, grey shaded area indicates KLx values for which the chaotic attractor was reproduced). Furthermore, the expected joint log-likelihood increased (Fig 3D) while KLx decreased (Fig 3C) over the distinct training steps of the PLRNN-SSM-anneal protocol, indicating that each step further enhances the solution quality. KLx and the normalized log-likelihood were, however, only moderately correlated (r = -.27, p < .001), as expected based on the formal considerations above (sect. ‘Stepwise initialization and training protocol’).
After establishing an efficient training procedure designed to enforce recovery of the underlying DS by the prior model (Eq 1), we more formally evaluated dynamical reconstructions on the chaotic Lorenz system and on the van der Pol (vdP) nonlinear oscillator. The vdP oscillator with nonlinear dampening is a simple 2-dimensional model for electrical circuits consisting of vacuum tubes [52] (equations given in Fig 4). Fig 4 illustrates its flow field in the plane, together with several trajectories converging to the system’s limit cycle (note that training was always performed on samples of the time series, not on the generally unknown flow field!).
As for the Lorenz system, we drew 100 time series samples of length T = 1000 with process noise (σ2 = .1) using Runge-Kutta numerical integration, and handed each of those over to a separate PLRNN-SSM inference run, testing with a range M = {8, 10, 12, 14} of latent states (see below and Discussion for how to determine a suitable latent space dimensionality M). As above, reconstruction performance was assessed in terms of the (normalized) KL divergence KL˜x (Eq 9) between the distributions over true and generated states in state space. In addition, for the chaotic attractor, the absolute difference between Lyapunov exponents [e.g. 50] from the true vs. the PLRNN-SSM-generated trajectories was assessed, as another measure of how well hallmark dynamical characteristics of the chaotic Lorenz system had been captured. For the vdP (non-chaotic) oscillator, we instead assessed the correlation between the power spectrum of the true and the generated trajectories (see Methods sect. ‘Reconstruction of benchmark dynamical systems’).
Overall, our PLRNN-SSM-anneal algorithm managed to recover the nonlinear dynamics of these two benchmark systems (see Fig 4). The inferred PLRNN-SSM equations reproduced the ‘butterfly’ structure of the somewhat challenging chaotic attractor very well (Fig 4D). The KL˜x measure effectively captured this reconstruction quality, with PLRNN reconstructions achieving values below KL˜x≈.4 agreeing well with the Lorenz attractor’s ‘butterfly’ structure as assessed by visual inspection (see Fig 4B). At the same time, for this range of KL˜x values the deviation between Lyapunov exponents of the true and generated Lorenz system was generally very low (see Fig 4C, grey shaded area). If we accept this value as an indicator for successful reconstruction, our algorithm was successful in 15%, 24%, 35%, and 28% of all samples for M = 8, 10, 12, and 14 states, respectively. Note that our algorithm had access only to rather short time series of T = 1000, to create a situation comparable to that for fMRI data. When examining the dependence of KL˜x on the number of latent states across a larger range in more detail, M ≈ 16 turned out to be optimal for this setting (S1 Fig), as for M > 16 no further decrease in KL˜x (hence no further improvement in approximating the true attractor geometry) was observed.
Importantly and in contrast to most previous studies, note we requested full independent generation of the original attractor object from the once trained PLRNN. That is, we neither ‘just’ evaluated the posterior p(Z|X) conditioned on the actual observations (as e.g. in [53], or [36]) , nor did we ‘just’ assess predictions a couple of time steps ahead (as, e.g., in [31]), but rather defined a much more ambitious goal for our algorithm.
For the vdP system, our inference procedure yielded agreeable results in 20%, 31%, 25%, and 35% of all samples for M = 8, 10, 12, and 14 states, respectively (grey shaded area in Fig 4F), with M = 14 about optimal for this setting (S1 Fig). Furthermore, around 50% of all estimates generated stable limit cycles and hence a topologically equivalent attractor object in state space, although these limit cycles varied a lot in frequency and amplitude compared to the true oscillator. Like for the Lorenz system, the KL˜x measure generally served as a good indicator of reconstruction quality (see Fig 4H), particularly when combined with the power spectrum correlation (Fig 4G), although low KL˜x values did not always guarantee and high values did not exclude the retrieval of a stable limit cycle.
As noted in the Introduction, a linear dynamical system (LDS) is inherently (mathematically) incapable of producing more complex dynamical phenomena like limit cycles or chaos. To explicitly illustrate this, we ran the same training procedure (Algorithm-1) on a linear state space model (LDS-SSM) which we created by simply swapping the ReLU nonlinearity φ(z) = max(z,0) with the linear function φ(z) = z in Eq 1 and 2. As expected, this had a dramatic effect on the system’s capability to capture the true underlying dynamics, with KL˜x close to 1 in most cases for both the Lorenz (Fig 4B and 4C) and the vdP (Fig 4F and 4G) equations. Even for the simpler (but nonlinear) oscillatory vdP system, LDS-SSM would at most produce damped (and linear, harmonic) oscillations which decay to a fixed point over time (Fig 5A).
We next tested our PLRNN inference scheme, with a modified observation model that takes the hemodynamic response filtering into account (PLRNN-BOLD-SSM; see sect. ‘Observation model for BOLD time series’), on a previously published experimental fMRI data set [54]. In brief, the experimental paradigm assessed three cognitive tasks presented within repeated blocks, two variants of the well-established working memory (WM) n-back task: a 1-back continuous delayed response task (CDRT), a 1-back continuous matching task (CMT), and a (0-back control) choice reaction task (CRT). Exact details on the experimental paradigm, fMRI data acquisition, preprocessing, and sample information can be found in [54]. From these data obtained from 26 subjects, we preselected as time series the first principle component from each of 10 bilateral regions identified as relevant to the n-back task in a previous meta-analysis [55]. These time series along with the individual movement vectors obtained from the SPM realignment procedure (see also Methods sect. ‘Data acquisition and preprocessing’) were given to the inference algorithm for each subject: Models with M = {1,…,10} latent states were inferred twice: once explicitly including, and once excluding external (experimental) inputs (i.e., in the latter analysis, the model had to account for fluctuations in the BOLD signal all by itself, without information about changes in the environment).
For experimentally observed time series, unlike for the benchmark systems, we do not know the ground truth (i.e., the true data generating process), and generally do not have access to the complete true state space either (but only to some possibly incomplete, nonlinear projection of it). Thus, we cannot determine the agreement between generated and true distributions directly in the space of observables, as we could for the benchmark systems. Therefore we use a proxy: If the prior dynamics is close to the true system which generated the experimental observations, and those represent the true dynamics well (at the very least, they are the best information we have), then the distribution of latent states constrained by the data, i.e. p(Z|X), should be a good representative of the distribution over latent states generated by the prior model on its own, i.e. p(Z). Hence, our proxy for the reconstruction quality is the KL divergence KLz(pinf(z|x),pgen(z)) (KLz for short, or, when normalized, KL˜z; see (Eq 11) in Methods) between the posterior (inferred) distribution pinf(z|x) over latent states z conditioned on the experimental data x, and the spatial distribution pgen(z) over latent states as generated by the model’s prior (governing the free-running model dynamics; we use capital letters, Z, and lowercase letters, z, to distinguish between full trajectories and single vector points in state space, respectively). Note that the latent space defines a complete state space as we have that complete model available (also note that our measure, as before, assesses the agreement in state space, not the agreement between time series).
For the benchmark systems, our proposed proxy KLz was well correlated with the KL divergence KLx assessed directly in the complete observation space, i.e., between true and generated distributions (Fig 6A, r = .72 on a logarithmic scale, p < .001; likewise, KLz(pinf(z|x),pgen(z)) and KLz(pgen(z),pinf(z|x)) were generally correlated highly; r>.9, p < .001). Moreover, although especially for chaotic systems we would not necessarily expect a good fit between observed or inferred and generated time series [c.f. 51], KL˜z on the latent space turned out to be significantly related to the correlation between inferred and generated latent state series in our case (on a logarithmic scale, see Fig 6B). That is, lower KL˜z values were associated with a better match of inferred and generated state trajectories.
This tight relation was particularly pronounced in models including external inputs (Fig 6B blue, top). This is expected, as in this case the internal dynamics are reset or partly driven by the external inputs, which will therefore induce correlations between directly inferred and freely generated trajectories. Thus, overall, KLz was slightly lower for models including external inputs as compared to autonomous models (see also Fig 6C). One simple but important conclusion from this is that knowledge about additional external inputs and the experimental task structure may (strongly) help to recover the true underlying DS. This was also evident in the mean squared error on n-step ahead predictions of generated as compared to true data (Fig 6D), i.e. when comparing predicted observations from the PLRNN-BOLD-SSM run freely for n time steps to the true observations (once again we stress, however, that a measure evaluated directly on the time series may not necessarily give a good intuition about whether the underlying DS has been captured well; see also Fig 2). Accuracy of n-step-ahead predictions also generally improved with increasing number of latent state dimensions, that is, adding latent states to the model appeared to enhance the dynamical reconstruction within the range studied here.
In contrast to the PLRNN-BOLD-SSM, the performance of the LDS-SSM with the same BOLD observation model (termed LDS-BOLD-SSM), and trained according to the same protocol (Algorithm-1, see also previous section), quickly decayed after about only three prediction time steps (Fig 6D), clearly below the prediction accuracy achieved by the PLRNN-BOLD-SSM for which the decay was much more linear. Interestingly, this comparatively sharp drop in prediction accuracy for the LDS-BOLD-SSM, unlike the PLRNN-BOLD-SSM, was accompanied by a similarly sharp rise in the discrepancy between generated and inferred latent state trajectories (Fig 6E), which was not apparent for the PLRNN-BOLD-SSM. This suggests that the rise in LDS-BOLD-SSM prediction errors is directly related to the model’s inability to capture the underlying system in its generative dynamics (while the inferred latent states may still provide reasonable fits), and–moreover–that the agreement between inferred and generated latent states is indeed a good indicator of how well this goal of reconstructing dynamics has been achieved. The linear model’s failure to capture the underlying dynamics was also evident from the fact that its generated trajectories often quickly converged to fixed points (Fig 7C), while the trained PLRNNs often mimicked the oscillatory activity found in the real data in their generative behavior (Fig 7B, see also S1 Video).
Moreover, we observed that a PLRNN-BOLD model fit directly to the observations (as one would, e.g., do for an ARMA model; see Methods), i.e. essentially lacking latent states, was much worse in forecasting the time series than either the PLRNN-BOLD-SSM or the LDS-BOLD-SSM, with predictions errors on average above 3.28 even for just a single time step ahead, either when external inputs were absent (MSE > 2.79 for 1-step) or present (MSE > 3.77 for 1-step), as compared to the results for the latent variable models in Fig 6D. On top, they produced a large number of globally unstable solutions (35% and 46%, respectively). This suggests that the latent state structure is absolutely necessary for reconstructing the dynamics, perhaps not surprisingly so given that the whole motivation behind delay embedding techniques in nonlinear dynamics is that the true attractor geometries are almost never accessible directly in the observation space [50].
To ensure that the retrieved dynamics did not simply capture data variation related to background fluctuations in blood flow (or other systematic effects of no interest), we examined whether the generated trajectories carried task-specific information. For this purpose, we assessed how well we could classify the three experimental tasks (which demanded distinct cognitive processes) via linear discriminant analysis (LDA) based on the generated (through the prior model) latent state trajectories. (We exclusively focused on classifying task phases, as these were pseudo-randomized across subjects, while ‘resting’ and ‘instruction’ phases occurred at fixed times, and we wanted to prevent significant classification differences which may occur either due to a fixed temporal order, or due to differences in presentation of experimental inputs during resting/instruction vs. proper task phases.) Fig 7A shows the relative classification error obtained when classifying the three tasks by the generated trajectories (bottom) as compared to that from the directly inferred trajectories (top), and to bootstrap permutations of these trajectories (black solid lines).
Overall, for M>2 latent states, generated trajectories significantly reduced the relative classification error, even in the absence of any external stimulus information, suggesting that distinct cognitive processes were associated with distinct regions in the latent space, and that this cognitive aspect was captured by the PLRNN-BOLD-SSM prior model (see also Fig 7B for an example of a generated state space for a sample subject, and Fig 8). As observed for the ahead-prediction error above, performance improved with increasing latent state dimensionality. While adding dimensions will boost LDA classifications in general, as it becomes easier to find well separating linear discriminant surfaces in higher dimensions, we did not observe as strong a reduction in classification error for the permutation bootstraps, suggesting that at least part of the observed improvement was related to better reconstruction of the underlying dynamics. Of note, models which included external inputs enabled almost perfect classifications with as few as M = 8 states. These results are not solely attributable to the model receiving external inputs, as these did not differentiate between cognitive tasks (i.e., number and type of inputs were the same for all tasks, see Methods sect. ‘Experimental paradigm’).
This is further supported by the observation that the LDS-BOLD-SSM produced much higher classification errors than the PLRNN-BOLD-SSM when either external inputs were present or absent (Fig 7A, dashed lines). Hence, not only does the LDS fail to capture the underlying dynamics and fares worse in ahead predictions (cf. Fig 6D and 6E), but it also seems to contain less information about the actual task structure, even in the inferred trajectories. This was particularly evident in the situation where trajectories were simulated (generated) and information about external stimuli was not provided to the models, where LDS-BOLD-SSM-based classification performance was close to chance level across all latent state dimensionalities (Fig 7A bottom, red dashed line), consistent with the fact that simulated LDS quickly converged to fixed points (cf. Fig 7C).
Lastly, we observed that trained PLRNN-BOLD-SSMs in many cases produced interesting nonlinear dynamics, including stable limit cycles, chaotic attractors, and multi-stability between various attractor objects (Fig 9). This indicates that the fMRI data may indeed harbor interesting dynamical structure that one would not have been able to reveal with linear state space models like classical DCMs, at least not within the retrieved system of equations (as argued above, the inferred posterior p(Z|X) may still reflect this structure, but the model itself would not reproduce it).
Furthermore, some of this structure clearly appeared to be linked to task properties: A power spectral analysis of time series generated by the trained PLRNNs revealed that the oscillations exhibited by these models had dominant periods in the same range as the durations of different task phases, as well as periods on the order of the duration of all three different tasks which were delivered in a repetitive manner (Fig 10A). Hence the PLRNN-BOLD-SSM has captured the periodic nature of the experimental design and associated cognitive demands within its limit cycle behavior, even when it was provided with no other source of information than the recorded BOLD activity itself (Fig 10A, left). Moreover, it appeared that the total number of stable objects and unstable fixed points in state space was related to task performance, with better performance (in terms of % correct choices) associated with a larger difference in the number of unstable relative to that of stable objects in the CMT (Fig 10B). From a dynamical systems perspective, one may speculate that these changes in state space structure are associated with a richer and more complex system dynamics [e.g. 8,9,56], which in turn may imply better and more flexible cognitive performance (note that by ‘unstable objects’ we are referring to unstable fixed points of the system dynamics, not to single latent states; unstable fixed points are as physiological as stable fixed points, only that they are hardly accessible experimentally since activity diverges from them, while our method by inferring the generating equations makes them ‘visible’).
While these observations serve to illustrate the new possibilities for analyzing links between system dynamics and computational properties provided by our approach, and the new types of questions about neural systems one may be able to ask, we caution that more detailed analyses (and possibly purpose-tailored task designs), beyond the scope of the present study, would be required to establish a stronger link. For instance, unstable limit cycles or chaotic objects were not considered here (for reasons of computational tractability), ceiling effects in percent of correct choices, and an increase in the proportion of globally unstable system estimates for M>9 (partly possibly due to the limited length of the time series) made a more systematic evaluation difficult in the present experimental data set.
Theories about neural computation and information processing are often formulated in terms of nonlinear DS models, i.e. in terms of attractor states, transitions among these, or transient dynamics still under the influence of attractors or other salient geometrical properties of the state space [4,9,57]. Given the success of DS theory in neuroscience, and the recent surge in interest in reconstructing trajectory flows and state spaces from experimental recordings [23,58–61], methodological tools which would return not only state space representations, but actually a model of the governing equations, would be of great benefit. Here we suggested a novel algorithm within an SSM framework that specifically forces the latent model, represented by a PLRNN, to capture the underlying dynamics in its intrinsic behavior, such that it can produce on its own time series of ‘fake observations’ that closely match the real ones (see also S1 Video). We also evaluated a measure, the KL divergence defined across state space (not time) between the inferred (posterior) and intrinsically generated (prior) distribution of latent states, which would give us a quantitative sense of how well the underlying state space geometry has been captured in empirical situations where no ground truth is available. Finally, given that fMRI is the most common non-invasive technique to study human cognition in health and psychiatric illness, we derived a new observation model specifically for fMRI data that takes the HRF into account. Using this, we demonstrated that our approach could recover nonlinear dynamics and trajectory flows from human fMRI recordings that were related to task structure and behavioral performance in a working memory paradigm. This, to our knowledge, has not been shown before.
Our major goal here was to establish an efficient methodological approach for recovering dynamical systems from empirical data in a truly generative sense, i.e. such that the trained models exhibit an intrinsic, standalone dynamics that mimics the underlying dynamics of the unknown real system, and to provide a specific measure based on attractor geometries for how well this aim has been achieved. We chose RNNs for the latent model because they are universal approximators of dynamical systems [26–28] and can emulate any Turing machine [62]. Just like the computations performed by a Turing machine can be implemented in many different substrates and algorithmic environments [see, e.g., discussion in 63], the same nonlinear dynamical system and behavior can be implemented in numerous different ways [e.g. 62]. Note, for instance, that the PLRNN can reproduce the chaotic Lorenz attractor although its set of equations is quite different from the original Lorenz equations. Hence, from a pure dynamical systems perspective, the functional form of the nonlinear model, and how close it is to biology, may be largely irrelevant as long as it is powerful enough to approximate any kind of dynamics sufficiently well, i.e. has the required representational expressiveness.
Nevertheless, we would like to repeat that our PLRNN does in fact have the mathematical form of a typical neural rate model as indicated in the first Results section [e.g. 37,38], and that its ReLU nonlinearity compares quite well to I/O functions of cortical pyramidal cells within the physiologically relevant regime [39,64,65], making the model neuronally directly interpretable in principle.
The major reason for settling on a ReLU nonlinearity was, however, that it allows for highly efficient optimization approaches, which also made ReLUs the de-facto standard in modern deep learning applications [44]. In our case, the ReLUs are centerpiece to an efficient fixed-point-iteration-type algorithm for the E-step and enable to compute most expectations analytically and fast (see Methods ‘State Estimation’). We believe that this efficiency of optimization, assuring that, in probability, we achieve better approximations to the underlying (biological or physical) system, is more important for capturing biology than the precise functional form of the latent model.
Although this was not a goal here, we further would like to point out that of course also task-specific coupling matrices W could be estimated, with subsets of latent states strictly assigned to only certain brain regions (via restrictions on B, Eqs 2 and 3). From a DS perspective, however, one might rather want to think about the same DS (with same parameters) producing different types of tasks (e.g., [38]), 2019), where the different tasks are more reflected by different local dynamics in possibly different regions of state space (cf. Fig 7B) rather than by differences in coupling parameters.
Finally, so far we have touched only briefly on the important question of how to determine the latent space dimensionality M in any practical setting. In our presentation we have deliberately explored a larger range of M values for testing and illustrating our algorithm, and mostly demonstrated that results were consistent across this larger range. While one may hope that reconstructing the underlying dynamical system involves a dimensionality reduction (M < N), i.e. that the effective dynamics lives in a lower-dimensional space than occupied by the observed measurements, the delay embedding theorems [48,49] as well as the universal approximation theorems for RNN [26,27] imply that we may instead have to move to (much) higher-dimensional spaces for achieving a good approximation to the underlying system and disentanglement of trajectories (an RNN approximates the underlying system through a type of basis expansion, and for, e.g., the Lorenz attractor, a set of just M = 3 piecewise linear functions cannot be expected to yield a reasonable representation). This implies that M should not be too low, but on other hand, for obtaining a well tractable and parsimonious system, we would not want to increase the latent space dimensionality more than absolutely necessary. Based on S1 Fig we had suggested that M ≈14 and 16 may be optimal for the vdP and Lorenz systems, respectively, based on the observation that from these points onwards no further improvement in geometry reconstruction according to KL˜x was observed. For Fig 10B, which analyzes the number of stable and unstable dynamical objects, M≈9 was chosen based on the fact that the number of retrieved dynamical objects roughly plateau-ed at this level (S2A Fig). Moreover, the finite length of the time series (which are very short in fMRI) will also place an upper bound on the system size for which reliable estimates could still be achieved. In our case, for M >9 we obtained more globally unstable model estimates which curtails the possibilities for analysis. More generally, in practice, one could try to devise a type of cross-validation procedure [25,66,67] based on KL˜z, but cross-validation for latent-variable time series models is notoriously difficult [68] and for M≥4 a clear dip in the KL˜z curve (see Fig 6C, bottom) was hard to discern in our case. Hence, beyond the empirical guidelines given here, this certainly remains a topic for future investigation.
The ‘classical’ technique for reconstructing attractor dynamics from experimental time series is delay embedding, based on the delay embedding theorems by Takens [48] and Sauer et al. [49]. It has been used to disentangle task-related trajectory flows and attractor-like properties in experimentally assessed neuronal time series [22,23]. However, as a completely non-parametric technique, delay embedding will not give a complete picture of the system’s flow field, nor access to the governing equations. Linear dynamical systems, coupled to Gaussian or Poisson observation equations [16,18,19], and related approaches like GPFA [20], are quite popular in neurophysiology for obtaining smoothed trajectories and state spaces, but–due to their linear latent dynamics–are inherently unsuitable for reconstructing the underlying DS itself in most cases (as explained above, they may still yield a good approximation to the posterior p(Z|X), thus still useful, but they would fail to capture the generative dynamics itself as explicitly shown in Fig 5 and Fig 7). In consequence, unlike the PLRNN-based models, LDS models were not able to pick up the nonlinear structure present in the BOLD signals in their generative dynamics (but mostly converged to simple fixed points), and probably as a result thereof produced worse forward predictions and contained less information about the cognitive tasks than the PLRNN.
To our knowledge, Roweis and Ghahramani [30], and somewhat later Yu et al. [29], were among the first to suggest an RNN for the latent model in order to reconstruct dynamics. These earlier contributions still focused more on in the inferred space p(Z|X), rather than on the fully generative capabilities of their models (at least were these not systematically analyzed), perhaps partly due to the fact that numerically less stable and efficient inference methods like the extended Kalman filter were employed at the time. Very recent work by Zhao and Park [35] built on the radial basis function networks suggested by Roweis and Ghahramani [30] for the latent model, and combined it with variational inference. They showed ahead predictions of their model for up to 1000 time steps. Similarly, Pandarinath et al. [36] recently proposed a sequential variational auto-encoder framework for inferring dynamics from neural recordings (although here as well the focus was more on the posterior encoding in the latent states, and on inference of initial conditions and perturbations). Both these models, however, are fairly complex and not directly interpretable in neural terms, and, moreover, hard to analyze with respect to their intrinsic dynamics.
The PLRNN framework offers several distinct advantages compared to other approaches: The equations have a fairly direct neural interpretation [31], in fact have the general form of neural rate equations that have been used to model various neural and cognitive phenomena [37,38], and–due to their piecewise-linear structure–can also be easily translated into an equivalent continuous-time neural rate model [see 69]. Dynamical phenomena can be analyzed more easily in PLRNNs than in other frameworks, e.g. fixed points and their stability can be determined analytically [31]. Furthermore, ReLU-type activation functions appear to be a quite good approximation to the I/O-functions of many neocortical cell types [39,64], and, besides, are almost the default now in deep networks due to their favorable properties in optimization [44], a feature our iterative state inference algorithm exploits as well. Finally, in contrast to most previous approaches, here we demonstrated that the prior PLRNN model on its own, after training, can produce the same attractor dynamics in state space as the true DS.
In the physics literature, several other methods based on reservoir computing [70], RNNs formed from feedforward networks trained directly on the flow field [see also 26,28], or LASSO regression combined with polynomial basis expansions [71], have recently been discussed for identifying DS. Process noise is usually not included in these models, i.e. the latent dynamics is deterministic, which entails the risk that noise in the process is wrongly attributed to deterministic aspects of the dynamics. While some of these methods required hundreds of hidden states and millions of samples to reconstruct the van der Pol or Lorenz attractors [28], we found that as few as just eight latent states and a single time series of length 1000, within the range of typical fMRI data, can be sufficient for the PLRNN-SSM to rebuild the chaotic Lorenz attractor, another tremendous advantage in empirical settings.
In this contribution, we have derived a new observation model for fMRI that accounts for the HRF filtering of the BOLD signal. The HRF implies that current observations do not depend only on the system’s current state (the common assumption in SSMs), but on a sequence of previous states, a situation handled relatively seamlessly by our PLRNN-SSM inference algorithm. fMRI is still the most common recording technique for monitoring brain function during cognitive and emotional processing in healthy and psychiatric subjects. Huge data bases have been compiled in large cohort studies over the past decade or so (e.g., the German National Cohort Study initiated by the Helmholtz association: https://www.helmholtz.de/en/research_infrastructures/national_cohort_study/; see also Collins and Varmus [72]) as a reference for monitoring and assessing neurological and psychiatric dysfunction. Although other noninvasive recording techniques with finer temporal resolution, like MEG/ EEG, may be more suitable for addressing questions about the DS basis of cognition, clinical research cannot afford to discard this large body of medically relevant data.
On the other hand, important hypotheses about the neural underpinnings of psychiatric conditions like schizophrenia, attention deficit hyperactivity disorder, or depression, have been formulated in terms of altered system dynamics [see 73 for a recent review]. For instance, based on physiological single unit and synapse data combined with biophysical network models on dopamine modulation in prefrontal cortex, it has been suggested that a dysregulated dopamine system by overly ‘deepening’ cortical attractor landscapes may inhibit transitions among states, and thereby cause some of the (cognitive) symptoms in schizophrenia [74]. This proposal has been supported by a number of neurophysiological and neuropsychological observations [e.g. 23,75], but a direct experimental evaluation of the specific changes in attractor basins in schizophrenia is still lacking. Tools like the one proposed here could be applied to directly test these types of hypotheses in human subjects recorded with fMRI. More generally, however, an extensive literature suggests that dynamical properties assessed from fMRI predict psychopathological conditions [e.g. 76,77,78], where the methodological framework proposed here could help to better understand the underlying dynamics and define targets for intervention (e.g. in the context of neurofeedback).
Beyond fMRI, most neuroimaging techniques, including, e.g., calcium imaging [79] or imaging by voltage-sensitive dyes [80] in neural tissue, involve some form of filtering that has to be taken into account when the goal is to capture underlying dynamical processes (like neural interactions) that evolve at a faster time scale. Through introduction of a filtering observation model (Eq 3), the present paper establishes a framework for inferring nonlinear dynamics in such situations where the measurement technique involves low- or band-pass-filtering of the process of interest. More generally, while we chose fMRI data here as our applicational example, we emphasize that our methodological framework is generic and could ultimately be applied to any other recording modality, like EEG, MEG, multiple single-unit data, or time series from mobile sensors, ecological momentary assessments [81], or electronic health records, for instance, by simply swapping the observation model (Eqs 2 and 3).
There is room for improvement in both our training algorithm and the measures used to evaluate its success in empirical situations. Our stepwise training algorithm was devised based on an intuitive heuristic, namely that by shifting the workload for fitting the observations onto the latent model and gradually increasing the requirements for its temporal consistency, a better representation of the unobserved system dynamics could be achieved. We could show that this was indeed the case when compared to a bootstrap (random) sample of models trained in the ‘standard’ way, and that our procedure seemed to work in general, but a more systematic theoretical derivation and testing of alternative schemes and explicitly designed optimization criteria (directly utilizing Eq 10, or combining our geometric measure with a time series measure) would certainly be desirable in future work.
We also find it important that in testing the performance of different reconstruction algorithms not only ‘good examples’ that prove the basic concept (‘my algorithm works’) are shown, but a more thorough quantitative statistical evaluation of precisely how well it performed in what percentage of cases is provided, like the one attempted here (Fig 4). For applications to empirical data, for which we do not know the ground truth, an open issue is how we could best quantify how much confidence we could have in the reconstructed stochastic equations of motion. Cross-validation and out-of-sample prediction errors provide a guidance, but for DS it is less clear in terms of what these should be measured: It is known that for nonlinear systems with complex or chaotic dynamics standard squared-error or likelihood-based measures evaluated along time series are not too useful [e.g. 51], since miniscule differences in initial conditions or noise perturbations may cause quick decorrelation of trajectories even if they come from the very same DS. We therefore decided to compare true and simulated data in terms of probability distributions across state space, arguing that if the observations come from the same attractor or system dynamics they should fill roughly the same volume of state space–this is more along the lines of a DS view which compares dynamical objects in terms of their geometrical or topological equivalence in state space [48–50,82], rather than the literal overlap among time series. Another corollary of this view is that to establish the equivalence between two DS, it is neither sufficient nor potentially even useful to predict observations just a couple of time steps ahead: In a chaotic noisy system, the prediction horizon is inherently limited to begin with (because of exponential divergence of trajectories). One also has to demonstrate that the ‘general type’ of long-term behavior in the limit is the same (e.g. a limit cycle of a certain periodicity and order), potentially in combination with other measures that quantify temporal aspects in the form of summary statistics (e.g., power spectrum). Here we therefore suggested to evaluate performance in terms of completely newly generated (‘faked’) trajectories that the trained system produces when no longer guided by the actual observations (i.e., the prior pgen(Z) rather than the posterior pinf(Z|X)).
Especially in fMRI, however, the data space is often very high-dimensional (>103) while at the same time often only a single time series sample of limited length (T≤1000) is available, i.e. the x-space is very sparse. In these cases we cannot obtain a good approximation of the distribution p(x), as we could for the benchmarks, and hence our original measure is not directly applicable. Hence we reverted to performing the comparison in latent space, between two distributions we do have in principle available, the one constrained by the observations, pinf(z|x), and the other, pgen(z), obtained from the completely freely running (simulated) system. We argued that if our actual observations X reflect the true dynamics well, then states obtained under pinf(z|x) should be highly likely a priori, i.e. under pgen(z), and hence these distributions should highly overlap. As direct sampling from pinf(z|x) is difficult and time-consuming, due to degeneracy problems, and the latent space dimensionality may also be prohibitively high, we approximated it by a mixture of Gaussians, which is a reasonable assumption for our ReLU-based RNN model and allows for an efficient analytical approximation to KLz [83]. More generally, if we are only interested in topological equivalence [48,49], we may also want to accept translations, rotations, rescaling, and potentially other deformations of the true state space that do not change topological aspects. Procrustes analysis [84] could be performed to (partly) allow for such transformations (on the other hand, since pgen(Z) and pinf(Z|X) come from the same underlying model, in our specific case such transformations may neither be necessary, nor necessarily desired).
The formulation of the state space model for BOLD time series (PLRNN-BOLD-SSM) is given in the Results section. To infer the parameters and latent variables of the model, we used Expectation-Maximization (EM) [41,85]. The EM algorithm maximizes a lower bound L(θ,q) (also called the evidence lower bound, ELBO) of the log-likelihood log p(X|θ) given by (see S1 Text sect. ‘PLRNN-BOLD-SSM model inference’ for full details):
logp(X|θ)≥Eq[logp(X,Z|θ)]+H(q(Z|X))=logp(X|θ)−KL(q(Z|X),pθ(Z|X))=:L(θ,q),
(4)
with q(Z|X) some proposal density over latent states, and KL(q(Z|X), p(Z|X)) the Kullback-Leibler divergence between proposal density q(Z|X) and true posterior p(Z|X). This expression can be derived by, e.g., using Jensen's inequality [e.g. 30]. From this we see that the bound becomes exact when proposal density q(Z|X) exactly matches the true posterior density p(Z|X) (defined through the latent state model here) which we aim to determine in the E-step (in contrast to variational inference where we assume q(Z|X) to come from some parameterized family of density functions, in EM we usually try to compute [in the linear case] or approximate p(Z|X) directly).
We introduce here an efficient approach for pushing the latent model to capture the underlying DS that generated the observations. Our approach rests on a step-wise procedure in which we gradually increase the importance of fitting the latent state dynamics as compared to fitting the observations. Since the latent state process and the observation process account for additive terms in the joint log-likelihood (Eq 5), the tradeoff between fitting the dynamics and fitting the observations is regulated by the ratio of the two covariance matrices Σ and Γ (Eqs 1–3 and 5). Hence, the idea of our training scheme is to begin with fitting the observation model and putting milder constraints on the latent process, using a linear latent model for initialization in a first step [or even factor analysis which places no constraints on the temporal relationship among latent states; cf. 30], and then gradually decreasing “Σ:Γ” during training to enforce the temporal consistency of the latent model. Furthermore, one may force all burden of fitting the observations completely onto the latent model by fixing θobs from some step onwards. The complete training protocol is outlined in Algorithm-1. For inferring a linear model (LDS-SSM, LDS-BOLD-SSM), the exact same algorithm was used with φ(z) = max(z,0) just replaced by φ(z) = z in Eqs 1 and 2.
0) Draw initial parameter estimates θ(0)~p(θ) from some suitable prior, constraint to max|eig(A+W)|<1 for biasing toward stable models [see also 18].
1) Fix Σ = I and run linear dynamical system (LDS) SSM for initialization → θ(1)
2) Fix Σ = I and run PLRNN-SSM inference → θ(2)
3) for i = 1:3
- Fix Σ = diag(10−i), B = B(2); fix Γ = Γ(2) (for fMRI data)
- Initialize PLRNN-SSM training with previous estimate θ(i+1)
- Run PLRNN-SSM inference → θ(i+2)
4) Re-estimate state covariance matrix Var(zt|x1:T) with Σ = I fixed.
We evaluated the performance of our PLRNN-SSM approach (and an LDS-SSM for comparison), on two popular benchmark DS, the Lorenz equations and the van der Pol nonlinear oscillator (vdP). Within some parameter range, the 3-dimensional Lorenz system exhibits a chaotic attractor and the 2-dimensional vdP-system exhibits a limit cycle (see Fig 4 for parameter settings used, system equations, and sample trajectories of the systems). We were interested in solutions where the true system dynamics is not just reflected in the directly inferred posterior distribution p(Z|X) over the PLRNN states {z1:T} given the actual observations {x1:T}, but also in the model’s generative or prior distribution p(Z), i.e. whether the once estimated PLRNN when run on its own would produce similar trajectories with the same dynamical properties as the ground truth system.
For evaluation, n = 100 samples of (standardized) trajectories of length T = 1000 were drawn from the ground truth systems using Runge-Kutta numerical integration and random initial conditions. PLRNN-SSMs were trained on these sample sets as described above for M = 5…20 latent states, using Eq 2 for the observations (see also Fig 1). To probe our stepwise training protocol (Algorithm-1), PLRNN-SSM training under this protocol (termed ‘PLRNN-SSM-anneal’) was compared to simple EM training of the PLRNN-SSM started from random initializations of parameters (termed ‘PLRNN-SSM-random’; essentially just step 1 of Algorithm-1 with Σ directly fixed to 10−3) for M = {8, 10, 12, 14}.
To quantify how well the true system dynamics was captured by the ‘free-running’ PLRNN (after training, but unconstrained by the observations), we used the Kullback-Leibler divergence defined across state space, i.e. integrating across space, not across time. Similar in spirit to the criteria defined for the classical delay embedding theorems [48–50], our measure therefore assessed the agreement between the original and reconstructed attractor geometries. Integrating across time (i.e., computing divergence between time series) is problematic for nonlinear DS, since two time series from the very same chaotic DS usually cannot be expected to overlap very well with even miniscule differences in initial conditions [cf. 51]. For the ground truth benchmark systems, for which we have access to the true distribution ptrue(x) and the complete state space, this KL divergence can be computed directly in observation space and was defined as
KLx(ptrue(x),pgen(x|z))≔∫x∈RNptrue(x)logptrue(x)pgen(x|z)dx,
(8)
where the integration is performed across x-space, and pgen(x|z) is the distribution across observations generated from PLRNN simulations (i.e., after PLRNN-SSM training, but discarding the original set of time series observations Xobs = {x1:T} used for training). Hence, this measure assesses whether PLRNN-SSM-simulated trajectories in the limit fill the same volume of state space as the true DS trajectories, and in this sense whether the systems’ attractor objects are topologically and geometrically ‘equivalent’. (As a terminological remark, in the machine learning literature pgen(x|z) is often called the ‘generative’ or ‘decoding’ model, while p(z|x) or q(z|x) is sometimes referred to as the ‘encoder’ or ‘recognition’ model [e.g. 32,90]. Here we will, more generally, refer with pgen(z) to the (prior) distribution of latent states generated by the PLRNN independent of the training observations Xobs = {x1:T}, and with pgen(x|z) to the distribution of simulated observations produced from samples zgen~pgen(z) according to the observation model [Eq 2]).
Practically, we discretized the x-space into K bins of width Δx and evaluated the probabilities ‘empirically’ as relative frequencies p^(k)=n(k)T by filling the space with trajectories (T = 100,000) sampled from the true DS and trained PLRNNs (here we used Δx = 1 across a range xn∈[−4 4] for standardized variables, but smaller bin sizes yielded qualitatively similar results, see S4 Fig). To avoid p^k(x|z)=0 for the generative model, where the KL divergence is not defined, we further adjusted this relative frequency to p^(k)=n(k)+αT+αK, with α = 10−6, also known as Laplace or additive smoothing [91] such that Eq 8 becomes
KLx(ptrue(x),pgen(x|z))≈∑k=1Kp^true(k)(x)log(p^true(k)(x)p^gen(k)(x|z)).
(9)
Lastly, to obtain an interpretable measure between 0 and 1, we normalized the KL divergence (termed KL˜x) by dividing it by the expected maximum deviation. KL˜x and the expected joint log-likelihood were compared between PLRNN-SSM-anneal and PLRNN-SSM-random via independent t-tests. For these analyses, all unstable system estimates were removed (≈14%). Furthermore, strong outliers with joint log-likelihood values < -1000 (which occurred only for PLRNN-SSM-random in ≈3.8% of cases) were removed.
A standard measure of chaoticity in nonlinear DS is the maximal Lyapunov exponent [24]. We thus also assessed how well our KL measure correlated with the deviation in Lyapunov exponents between true and estimated systems. The Lyapunov exponent was assessed numerically by a linear regression fit to the initial slope of the log-Euclidean distance log dΔt(X(1),X(2)) between initially close (d0<10−10) trajectories X(1) and X(2) as a function of time lag Δt, up to the point in the curve where a plateau indicating the full extent of the attractor object has been reached. For the van der Pol nonlinear (non-chaotic) oscillator, the agreement in the power spectra between the true and generated systems is more informative as a measure of how well the system dynamics has been captured (the maximum Lyapunov exponent for a stable limit cycle is 0), which was simply assessed by the average Pearson correlation.
|
10.1371/journal.pbio.0050254 | The Diploid Genome Sequence of an Individual Human | Presented here is a genome sequence of an individual human. It was produced from ∼32 million random DNA fragments, sequenced by Sanger dideoxy technology and assembled into 4,528 scaffolds, comprising 2,810 million bases (Mb) of contiguous sequence with approximately 7.5-fold coverage for any given region. We developed a modified version of the Celera assembler to facilitate the identification and comparison of alternate alleles within this individual diploid genome. Comparison of this genome and the National Center for Biotechnology Information human reference assembly revealed more than 4.1 million DNA variants, encompassing 12.3 Mb. These variants (of which 1,288,319 were novel) included 3,213,401 single nucleotide polymorphisms (SNPs), 53,823 block substitutions (2–206 bp), 292,102 heterozygous insertion/deletion events (indels)(1–571 bp), 559,473 homozygous indels (1–82,711 bp), 90 inversions, as well as numerous segmental duplications and copy number variation regions. Non-SNP DNA variation accounts for 22% of all events identified in the donor, however they involve 74% of all variant bases. This suggests an important role for non-SNP genetic alterations in defining the diploid genome structure. Moreover, 44% of genes were heterozygous for one or more variants. Using a novel haplotype assembly strategy, we were able to span 1.5 Gb of genome sequence in segments >200 kb, providing further precision to the diploid nature of the genome. These data depict a definitive molecular portrait of a diploid human genome that provides a starting point for future genome comparisons and enables an era of individualized genomic information.
| We have generated an independently assembled diploid human genomic DNA sequence from both chromosomes of a single individual (J. Craig Venter). Our approach, based on whole-genome shotgun sequencing and using enhanced genome assembly strategies and software, generated an assembled genome over half of which is represented in large diploid segments (>200 kilobases), enabling study of the diploid genome. Comparison with previous reference human genome sequences, which were composites comprising multiple humans, revealed that the majority of genomic alterations are the well-studied class of variants based on single nucleotides (SNPs). However, the results also reveal that lesser-studied genomic variants, insertions and deletions, while comprising a minority (22%) of genomic variation events, actually account for almost 74% of variant nucleotides. Inclusion of insertion and deletion genetic variation into our estimates of interchromosomal difference reveals that only 99.5% similarity exists between the two chromosomal copies of an individual and that genetic variation between two individuals is as much as five times higher than previously estimated. The existence of a well-characterized diploid human genome sequence provides a starting point for future individual genome comparisons and enables the emerging era of individualized genomic information.
| Each of our genomes is typically composed of DNA packaged into two sets of 23 chromosomes; one set inherited from each parent whose own DNA is a mosaic of preceding ancestors. As such, the human genome functions as a diploid entity with phenotypes arising due to the sometimes complex interplay of alleles of genes and/or their noncoding functional regulatory elements.
The diploid nature of the human genome was first observed as unbanded and banded chromosomes over 40 years ago [1–4] , and karyotyping still predominates in clinical laboratories as the standard for global genome interrogation. With the advent of molecular biology, other techniques such as chromosomal fluorescence in situ hybridization (FISH) and microarray-based genetic analysis [5,6] provided incremental increases in the resolution of genome analysis. Notwithstanding these approaches, we suspect that only a small proportion of genetic variation is captured for any sample in any one set of experiments.
Over the past decade, with the development of high-throughput DNA sequencing protocols and advanced computational analysis methods, it has been possible to generate assemblies of sequences encompassing the majority of the human genome [7–9]. Two versions of the human genome currently available are products of the Human Genome Sequencing Consortium [9] and Celera Genomics [7], derived from clone-based and random whole genome shotgun sequencing strategies, respectively. The Human Genome Sequencing Consortium assembly is a composite derived from haploids of numerous donors, whereas the Celera version of the genome is a consensus sequence derived from five individuals. Both versions almost exclusively report DNA variation in the form of single nucleotide polymorphisms (SNPs). However smaller-scale (<100 bp) insertion/deletion sequences (indels) or large-scale structural variants [10–15] also contribute to human biology and disease [16–18] and warrant an extensive survey.
The ongoing analyses of these DNA sequence resources have offered an unprecedented glimpse into the genetic contribution to human biology. The simplification of our collective genetic ancestry to a linear sequence of nucleotide bases has permitted the identification of functional sequences to be made primarily through sequence-based searching alignment tools. This revealed an unexpected paucity of protein coding genes (20,000–25,000) residing in less than 2% of the DNA examined, suggesting that alternative transcription and splicing of genes are equally important in development and differentiation [19,20]. The sequencing of DNA of various eukaryotic genomes, such as for murine [21,105] and primate [22,23] as well as many others, has enabled a comparative genomics strategy to refine the identification of orthologous genes. These genomic datasets have also enabled the identification of additional functional sequence such as cis-regulatory DNA [24–29] as well as both noncoding and microRNA [30–34] .
Building on the existing genome assemblies, numerous initiatives have explored variation at the population level, in particular to generate markers and maps as a means of understanding how sequence variation evolves and can contribute to phenotype. The initial drafts of the two human genomes provided an excess of 2.4 million SNPs [7,8] providing a platform for the initial phase of the HapMap project [35]. This ambitious project initially catalogued genetic variation at more than 1.2 million loci in 269 humans of four ethnicities, enabling a definition of common haplotypes and resulting in tag SNP sets for these populations. The use of these data has already allowed the mapping and identification of susceptibility genes and loci involved in complex diseases such as asthma [36], age related macular degeneration [37], and type II diabetes [38]. Notwithstanding, there are limitations with current SNP-based genome-wide association studies, because they rely on reconstructing haplotypes based on population data and can be uninformative or misleading in regions of low linkage disequilibrium (LD). Further, association studies have been designed to detect common disease variants and are not optimized to detect rare etiological variants [39].
The ability to generate a diploid genome structure via haplotype phasing for the HapMap samples is limited by the SNPs that were genotyped and their spacing. By using LD measures, it was possible to identify diploid blocks of DNA averaging 16.3 kb for Caucasians (CEU), 7.3 kb for Yorubans (YRI), and 13.2 kb for grouped Han Chinese and Japanese (CHB+JPT) [35]. However, LD varies across the genome, and regions of low LD, i.e., high recombination, cannot be represented by haplotype blocks. Furthermore, these diploid blocks are incomplete because there may be unknown variants between the SNP loci sampled. These results do not permit a comprehensive definition of the sequence present at each allele nor the information that produces the relevant allelic combinations, which are essential in identifying the differences of biological information encoded by the diploid state. The ability to perform, in a practical manner, whole-genome sequencing in large disease populations would enable the construction of haplotypes from individuals' genomes, thus phasing all variant types throughout the genome without assumptions about population history. Clearly, to enable the forthcoming field of individualized genomic medicine, it is important to represent and understand the entire diploid genetic component of humans, including all forms of genetic variation in nucleotide sequences, as well as epigenetic effects.
To understand fully the nature of genetic variation in development and disease, indeed the ideal experiment would be to generate complete diploid genome sequences from numerous controls and cases. Here we report our endeavor to fully sequence a diploid human genome. We used an experimental design based on very high quality Sanger-based whole-genome shotgun sequencing, allowing us to maximize coverage of the genome and to catalogue the vast majority of variation within it. We discovered some 4.1 million variants in this genome, 30% of which were not described previously, furthering our understanding of genetic individuality. These variants include SNPs, indels, inversions, segmental duplications, and more complex forms of DNA variation. We used the variant set coupled with the sequence read information and mate pairs to build long-range haplotypes, the boundaries of which provide coverage of 11,250 genes (58% of all genes). In this manner we achieved our goal of the construction of a diploid genome, which we hope will serve as a basis for future comparison as more individual genomes are produced.
The individual whose genome is described in this report is J. Craig Venter, who was born on 14 October 1946, a self-identified Caucasian male. The DNA donor gave full consent to provide his DNA for study via sequencing methods and to disclose publicly his genomic data in totality. The collection of DNA from blood with attendant personal, medical, and phenotypic trait data was performed on an ongoing basis. Ethical review of the study protocol was performed annually. Additionally, we provide here an initial foray into individualized genomics by correlating genotype with family history and phenotype; however, a more extensive analysis will be presented elsewhere.
The donor's three-generation pedigree is shown in Figure 1A. The donor has three siblings and one biological son, his father died at age 59 of sudden cardiac arrest. There are documented cases of family members with chronic disease including hypertension and ovarian and skin cancer. According to the genealogical record, the donor's ancestors can be traced back to 1821 (paternal) and the 1700s (maternal) in England. Genotyping and cluster analysis of 750 unique SNP loci discovered through this project support that the donor is indeed 99.5% similar to individuals of European descent (Figure 1B), consistent with self-reporting. This is further corroborated by an extensive five-generation family history provided by the donor (unpublished data). Cytogenetic analysis through G-banded karyotyping and spectral karyotypic chromosome imaging reveals no obvious chromosomal abnormalities (Figure 2) that need to be considered in interpretation of genome assembly results or phenotypic association analyses.
The assembly, herein referred to as HuRef, was derived of approximately 32 million sequence reads (Table S1) generated by a random shotgun sequencing approach using the open-source Celera Assembler. The approach used is similar in many respects to the whole-genome shotgun assembly (WGSA) reported previously [40], but there are three major differences: (i) HuRef was assembled entirely from shotgun reads from a single individual, whereas WGSA was based on shotgun reads from five individuals [7,40,41], albeit the majority of reads were from the same individual as HuRef; (ii) the approximate depth of sequence coverage for HuRef was 7.5 versus 5.3 for WGSA, although the clone coverage was about the same for both (Table 1) [7,40]; and (iii) the release of Celera Assembler as an open-source project has allowed us and others to continue to improve the assembly algorithms. As a consequence, we made modifications for the specification of consensus sequence differences found at distinct alleles. The multiple sequence alignment methodology was improved and reads were grouped by allele, thus allowing the determination of alternate consensus sequences at variant sites (see Materials and Methods).
HuRef is a high-quality draft genome sequence as evidenced from the contiguity statistics (Table 2). Improving the assembly algorithms and increasing the sequencing depth of coverage (compared to WGSA) resulted in a 68% decrease in the number of gaps within scaffolds from 206,552 (WGSA) to 66,815 (HuRef) as previously predicted [40]. We also observed a more than 4-fold increase in the N50 contig size (the length such that 50% of all base pairs are contained in contigs of the given length or larger) to 106 kb (HuRef) from 23 kb (WGSA). We used a fairly standard, but arbitrary, cutoff of 3,000 bp (similar to what was used for WGSA) to distinguish between scaffolds that were part of the HuRef assembly proper versus partially assembled and poorly incorporated sequence (see Materials and Methods). This resulted in 4,528 scaffolds (containing 2,810 Mb) of which 553 scaffolds were at least 100 kb in size (containing 2,780 Mb), whereas WGSA had 4,940 scaffolds (containing 2,696 Mb) of which 330 scaffolds were at least 100 kb (containing 2,669 Mb). The scaffold lengths for HuRef (N50 = 19.5 Mb) were somewhat shorter than WGSA (N50 = 29 Mb) primarily due to the difference in insert size for bacterial artificial chromosome (BAC) end mate pairs—HuRef 91 kb versus WGSA > 150 kb (Table 2) [41]. We determined that 144 of the 553 large HuRef scaffolds could be joined by two or more of the WGSA BAC mate pairs, and 98 more by a single WGSA BAC mate pair (see Materials and Methods), suggesting that use of large insert BAC libraries (>150 kb) would generate larger scaffolds.
Genomic variation was observed by two approaches. First, we identified heterozygous alleles within the HuRef sequence. This variation represents differences in the maternal and paternal chromosomes. In addition, a comparison between HuRef and the National Center for Biotechnology Information (NCBI) version 36 human genome reference assembly, herein referred to as a one-to-one mapping, also served as a source for the identification of genomic variation. These comparisons identified a large number of putative SNPs as well as small, medium, and large insertion/deletion events and some major rearrangements described below. For the most part, the one-to-one mapping showed that both sequences are highly congruent with very large regions of contiguous alignment of high fidelity thus enabling the facile detection of DNA variation (Table S2).
The one-to-one mapping to NCBI version 36 (hereafter NCBI) was also used to organize HuRef scaffolds into chromosomes. HuRef scaffolds were only mapped to HuRef chromosomes if they had at least 3,000 bp that mapped and the scaffold was mostly not contained within a larger scaffold. With the exception of 12 chimeric joins, all scaffolds were placed in their entirety with no rearrangement onto HuRef chromosomes. The 12 chimeric regions represent the misjoining of a small number of chimeric scaffold/contigs by the Celera Assembly [40], as detected with mate pair patterns [7,42], and are also apparent by comparison to another assembly (Materials and Methods). The 12 chimeric joins in the HuRef scaffolds were split when these scaffolds were assigned to build HuRef chromosomes. Inversions and translocations within the nonchimeric scaffolds relative to NCBI are thus maintained within the HuRef chromosomes. The final set of 24 HuRef chromosomes were thus assembled from 1,408 HuRef assembly scaffolds and contain 2,782 Mb of ordered and oriented sequence.
The NCBI autosomes are on average 98.3% and 97.1% represented by runs and matches, respectively, in the one-to-one mapping to HuRef scaffolds (Table S3). A match is a maximal high-identity local alignment, usually terminated by indels or sequence gaps in one of the assemblies. Runs may include indels and are monotonically increasing or decreasing sets of matches (linear segments of a match dot plot) with no intervening matches from other runs on either axis.
The Y chromosome is 59% covered by the one-to-one mapping due to difficulties when producing comparison between repeat rich chromosomes. In addition, the Y chromosome is more poorly covered because of the difficulties in assembling complex regions with sequencing depth of coverage only half that of the autosomal portion of the genome. The X chromosome coverage with HuRef scaffolds is at 95.2%, which is typical of the coverage level of autosomes (mean 98.3% using runs). However it is clear that the X chromosome has more gaps, as evidenced by the coverage with matches (89.4%) compared with the mean coverage of autosomes using matches (97.1%). The overall effects of lower sequence coverage on chromosomes X and Y are clearly evident as a sharp increase in number of gaps per unit length and shorter scaffolds compared to the autosomes (Figure 3). Similarity between the sex chromosomes is another source of assembly and mapping difficulties. For example, there is a 1.5-Mb scaffold that maps equally well to identical regions of the X and Y chromosomes and therefore cannot be uniquely mapped to either (see Materials and Methods and Figure 3). From our one-to-one mapping data, we are also able to detect the enrichment of large segmental duplications [10] on Chromosomes 9, 16, and 22, resulting in reduced coverage based on difficulties in assembly and mapping (Table S3).
Since NCBI, WGSA, and HuRef are all incomplete assemblies with sequence anomalies, assembly-to-assembly mappings also reflect issues of completeness and correctness. We compared three sets of chromosome sequences to evaluate this issue (see Materials and Methods): NCBI with the exclusion of the small amount of unplaced sequences, HuRef, and WGSA (Table S2) were thus compared in a pairwise manner. The comparison of WGSA and HuRef revealed 83 Mb more sequence in HuRef in matched segments of these genomes. This sequence is predominantly from HuRef that fills gaps in WGSA. Comparisons of HuRef and WGSA to NCBI showed the considerable improvement of HuRef over WGSA. Correspondingly, in HuRef there are approximately 120 Mb of additional aligned sequence, composed of 47 Mb of HuRef sequence that aligns to NCBI that was not aligned in WGSA and 73 Mb within aligned regions that fill gaps in WGSA. This comparison also showed an improvement factor of two in rearrangement differences (order and orientation) from WGSA to HuRef when mapped to the NCBI reference genome at small (<5 kb), medium (5–50 kb), and large (>50 kb) levels of resolution (Table S2). HuRef includes 9 Mb of unmatched sequence that fill gaps in NCBI or are identified as indel variants. An additional 14 Mb of HuRef chromosome sequence outside of aligned regions with NCBI represents previously unknown human genome sequence. The large regions of novel HuRef sequence are identified to be either: (a) gap filling or insertions, (b) unaligned NCBI chromosome regions, or (c) large scaffolds not mapped to NCBI chromosomes. Some of these were investigated using FISH analysis and are discussed below. Although we were able to organize HuRef scaffolds into HuRef chromosome sequence, all of the subsequent analyses in this report were accomplished using HuRef scaffold sequences.
To examine sequence diversity in the genome, we estimated nucleotide diversity using the population mutation parameter θ [43]. This measure is corrected for sample size and the length of the region surveyed. In the case of a single genome with two chromosomes, θ simplifies to the number of heterozygote variants divided by the number of base pairs (see Materials and Methods). We define θSNP as the nucleotide diversity for SNPs (number of heterozygous SNPs/number of base pairs) and θindel as the diversity for indels (number of heterozygous indels/number of base pairs) [44]. For both θSNP and θindel, the 95% confidence interval would be [0, 3θ] due to the small number of chromosomes (n = 2) being sampled (see Materials and Methods).
Across all autosomal chromosomes, the observed diversity values for SNPs and indels are 6.15 × 10−4 and 0.84 × 10−4 respectively. When restricted to coding regions only, θSNP = 3.59 × 10−4 and θindel = 0.07 × 10−4, indicating that 42% of SNPs and 91% of indels have been eliminated by selection in coding regions. The strong selection against coding indels is not surprising, because most will introduce a frameshift and produce a nonfunctional protein. Our observed θSNP falls within the range of 5.4 × 10−4 to 8.3 × 10−4 that has been previously reported by other groups [44–47].
Our observed θindel (0.84 × 10−4) is approximately 2-fold higher than the diversity value of 0.41 × 10−4 that was reported from SeattleSNPs (http://pga.gs.washington.edu), which was derived from directed resequencing of 330 genes in 23 individuals of European descent [44]. The values of θindel in repetitive sequence regions are 1.2 × 10−4 for regions identified by RepeatMasker (http://www.repeatmasker.org) and 4.9 × 10−4 for regions identified by TandemRepeatFinder [48], respectively. Thus, the indel diversity in repetitive regions is between 1.4 and 5.8 times higher than the genome-wide rate. This suggests that the high value of θindel over all loci is likely mediated by the abundance of indels in repetitive sequence. It is also possible that repetitive regions in genic sequence are under stronger selective pressure and therefore have lower indel diversity. These are precisely the regions that have been targeted in previous resequencing projects [44] from which indel diversity values have been determined. Additionally, repetitive regions also have more erroneous variant calls due to technical difficulties in sequencing and assembly of these types of regions. Therefore, our estimate for θindel is likely a combination of both a true higher mutation rate in repetitive regions and sequencing errors.
Values of θindel are consistent among the chromosomes (Figure 5). Chromosomes with high θindel values also have a larger fraction of tandem repeats. For example, Chromosome 19 has the highest θindel (1.1 × 10−4 compared with the chromosomal average of 0.86 × 10−4), and it also has the highest proportion of tandem repeats (13% compared with the chromosomal average of 7%). The fraction of tandem repeats of a chromosome is positively correlated with the value of θindel for each chromosome (r = 0.73), so that the diversity of indels is associated with the underlying sequence composition.
The SNP variants identified in the HuRef genome include a larger-than-expected number of homozygous variants than those commonly observed in population-based studies (compare ratios of heterozygous SNP:homozygous SNP in Table 5). Our homozygous variants are detected as differences between the HuRef genome and the NCBI genome. One common interpretation of a homozygous variant is that given a common allele A and a rare allele B, the homozygous SNP is BB. However, because not all variant frequencies are known, we cannot determine if a position may carry the minor B allele in homozygous form. We analyzed ENCODE data using this definition and found the ratio of heterozygous SNPs to homozygous SNPs is 4.9 in an individual [49]. For our dataset, the observed ratio of heterozygous to homozygous SNP, where our “homozygous” SNPs are detected as bases differing from the NCBI human genome, is 1.2. To resolve this discrepancy, we examined the homozygous positions in the HuRef assembly and found that the increased frequency of homozygous SNPs results from the presence of minor alleles (BB) in the NCBI genome assembly. We observed that 75% of the homozygous positions in HuRef also had a SNP identified by the ENCODE [49]. A comparison of the alleles at these positions revealed that in 56% of the instances the HuRef genome had the more common allele, whereas the NCBI genome contained the minor allele. The remaining homozygous SNPs tended to be common minor alleles (76% had minor allele frequency [MAF] ≥ 0.30), consistent with their observation in homozygous form in the HuRef genome. Therefore, we confirmed that a large fraction of homozygous alleles from HuRef are real, and that differences between the HuRef and NCBI assemblies are due to NCBI containing the minor allele at a given SNP position, or HuRef containing a common SNP in homozygous form.
We also modeled the inter/intraindividual genome comparison using directed resequencing data from SeattleSNPs data (see Materials and Methods) to determine if our variant detection frequencies were commonly found for different types of variants. By sampling and comparing the genotypes of two individuals from the SeattleSNPs data, we were able to simulate the conditions for calling “heterozygous” and “homozygous” variants as we have defined them in an independently generated set (Table 5). The ratio of heterozygous variants to homozygous variants from the modeled SeattleSNPs is lower in the HuRef genome compared with the SeattleSNPs data. This suggests that there are an overabundance of homozygous variants and/or an under-representation of heterozygous variants, and this trend is more pronounced for indels compared to SNPs. A possible explanation for this is that homozygous genotypes are actually heterozygous and the second allele is missed due to low sequence coverage. Our attempts to explain this phenomenon using statistical modeling did support our hypothesis that low sequence coverage resulted in excess homozygous over heterozygous variant calls. Indeed, our modeling provided us with a bound on the missed heterozygous calls for both indels (described below) and SNPs (see section below titled: Experimental Validation of SNP Variants).
In an attempt to explain the discrepancy in the heterozygous to homozygous indel ratio (Table 5), we modeled the rate of identification of true heterozygous variants given the depth of coverage of HuRef sequencing reads and the various variant filtering criteria. This enabled us to determine that between 44% and 52% of the time, heterozygous indels will be missed due to insufficient read coverage at 7.5-fold redundancy and these indels be erroneously called homozygous. Therefore, the projection for the true number of homozygous indels is between 418,731 and 459,639, a reduction of 17%–25% from the original number of 559,473 homozygous indels, and the corresponding ratio of heterozygous to homozygous indels is between 1:1 and 1.3:1. Furthermore, our modeling also allowed us to determine that approximately 20× sequence coverage would be required to detect a heterozygous variant with 99% probability in unique sequence given our current filtering criteria of random shotgun sequence reads.
Another further explanation for the overabundance of homozygous indels is the error-prone nature of repeat regions. Using a subset of genes (55) completely sequenced by SeattleSNPs, we found that 28% of the potential 92 HuRef homozygous indels overlap with indels in these genes, as opposed to 75% confirmation rate for homozygous SNPs described earlier. When one categorizes the repeat status of a homozygous indel, a higher confirmation rate (46%) is seen for indels excluded from regions identified by RepeatMasker or TandemRepeatFinder. The confirmation rate for an indel in a transposon or tandem repeat region is much lower at 16%. Therefore, indels in nonrepetitive loci have a higher probability of authenticity than indels in repeat regions.
The ratio of SNPs to indels is lower in the HuRef assembly than what is observed by the SeattleSNPs data (Table 5), indicating that relatively fewer SNPs or relatively more indels are called. This is likely due to relatively more indels being identified, as discussed above. We note that a large fraction of indels occur in repeat sequence (Table 6), which has higher indel frequency as well as higher incidence of sequencing error. Moreover, SeattleSNPs resequencing data is focused on variant discovery in genic regions, which may not reflect genome-wide indel rates.
We identified in the HuRef assembly 263,923 heterozygous indels spanning 635,314 bp, with size ranges from 1 to 321 bp. The characteristics of the indels we detected, their distribution of sizes <5 bp, and the inverse relationship of the number of indels to length are similar to previous observations [50,51] (Figure 6A and 6B). As noted previously (Table 6), there are 2-fold more homozygous indels (559,473) than heterozygous indels, and these span 5.9 Mb and range from 1 to 82,771 bp in length. We observe that genome-wide, even-length indels are more frequent than odd-length indels (Figure 6C and 6D, χ2 = 12.4; p < 0.001, see Materials and Methods). One possible explanation for these results is that tandem repeats often have motif sizes that occur in even numbers, such as through the expansion of dinucleotide repeats. In fact, based on RepeatMasker, the majority of simple repeats are composed of even-numbered–sized motifs rather than odd-numbered–sized motifs (73%). Furthermore, of the heterozygous indels that occur in simple repeats identified by RepeatMasker, 79% occur in even-numbered bp repeats. This suggests that the preponderance of even-base–sized indels likely results from the inherent composition of simple repeats.
There are 6,535 homozygous indels that are at least 100 bases in length for which both flanks of the indel can be located precisely on HuRef and NCBI assemblies. These comprise 3,431 insertions uniquely occurring on HuRef, totaling 2.13 Mb, and 3,104 deletions, totaling 1.82 Mb, found only on NCBI (Figure 7). These homozygous indels have a higher representation of repetitive elements (66%–67%) than the overall HuRef and NCBI assemblies (each 49%). This enrichment derives mainly from a higher relative content of short interspersed nuclear elements (SINEs), simple repeats, and unclassified SVAs (Table 7). For 657 (19% of the total) insertions with a minimum length of 100 bp, at least 50% of the segment length (mean = 95%) is composed of a single SINE insertion. Most of these SINE insertions (88%) belong to the youngest Alu family (AluY), for which insertion polymorphisms are well documented in the human genome [52,53]. Similarly, for 26% of deletions at least 100 bp in length, an average of 95% of the segment consists of a single SINE element, and 92% of these elements are classified as AluY. Interestingly, the combined total of 1,316 AluY insertions that differ between HuRef and NCBI include 703 (53%) that are not currently identified in the most comprehensive database of human bimorphic SINE insertions, the database of retrotransposon insertion polymorphisms in human (dbRIP;1625 loci; http://falcon.roswellpark.org:9090/) (Table S4) [54].
To evaluate the accuracy and validity of SNP calling from the sequencing reads, the donor DNA was interrogated using hybridization-based SNP microarrays: the Affymetrix Mapping 500K Array Set, which targets 500,566 SNP markers, and the Illumina HumanHap650Y Genotyping BeadChip, which targets 655,362 SNPs. The Affymetrix array experiment was performed twice to provide a technical replicate for genotyping error estimation, and 0.12% of genotype calls were discordant. Of the 92,144 assays with an annotation in dbSNP that overlap between the two different platforms, 99.87% were concordant (0.13% discordant). Thus, the discordance rate between platforms was similar to that between Affymetrix technical replicates. Genotype calls that were discordant between technical replicates or between the Affymetrix and Illumina platforms were excluded from further analysis. This resulted in 1,029,688 nonredundant SNP calls from the two genotyping platforms, which were then compared to the HuRef assembly and to the single nucleotide variants extracted from the sequencing data. Of these, 943,531 genotypes (91.63%) were concordant between the genotyping platforms and the HuRef assembly (Table 8). Of the 86,157 discordant genotype calls, the vast majority (83.9%) were identified as heterozygous in the merged genotyping platform data, but called as homozygous in the HuRef assembly (Table 9). This is consistent with a predictable effect of finite sequence coverage in the HuRef dataset: assuming uniform random sampling of both haplotypes, 21.6% of true heterozygous SNPs are expected to be missed given 7.5× coverage of the diploid genome and the requirements for calling a heterozygous SNP (i.e., at least two instances of each allele and ≥20% of reads confirming the minor allele). This is close to the observed false-negative error of 24.6% (Table 9 and Figure 8). Consistent with this explanation, the level of coverage is significantly lower for the missed heterozygous SNPs than for the heterozygous SNPs detected in the HuRef assembly (average read depth 5.2 and 8.8, respectively) (Figure 9).
Another possible form of error would be to erroneously call a truly homozygous position a heterozygous variant. Of the 65,337 homozygote calls that were concordant between the Affymetrix and Illumina platforms, none were called as heterozygous in the HuRef assembly. Therefore, the upper bound for the false-positive rate is 0.0046% (one-tailed 95% confidence interval), and one would expect false-positive heterozygote calls approximately once every 22 kb from the upper bound of this confidence interval. However, this estimate may be lower than the genome-wide false-positive error, because it is based on the positions chosen by the microarray platforms, which tend to be biased away from repetitive, duplicated, and homopolymeric regions. Approximately three-quarters of the novel heterozygous SNPs (73%) and novel heterozygous indels (75%) are in a region identified by RepeatMasker, TandemRepeatFinder, or a segmental duplication. Therefore, approximately three-quarters of the novel heterozygous variants are in regions that are most likely underrepresented in the microarrays. Consequently, we cannot readily extrapolate the false-positive error determined from the microarrays to be the discovery rate of the HuRef variant set. The repetitive regions are likely to have a higher false-positive rate due to sequencing error and misassembly. Further, they are not represented in the current estimate of the false-positive rate. However, they also exhibit a higher rate of authentic variation.
Homozygous and heterozygous insertions and deletions identified in the HuRef assembly were computationally validated by comparison to previously published datasets. As indicated in Figure 4, the homozygous insertion and deletions variants are operationally defined as either inserted or deleted sequence in the HuRef genome respectively since there is no other read evidence for heterozygosity. The homozygous nature of these variants does not imply any notion of ancestral allele. The largest set of indel variants that has been published is based on mapping of trace reads to the NCBI human genome reference assembly [55]. This approach can be used to identify deletions of any size and insertions that are small enough to be spanned by sequence reads. In this analysis, the 216,179 deletions and 177,320 insertions from Mills et al. [55] were compared to the insertions and deletions identified from the HuRef assembly. Based on this analysis, we found support for 37,893 homozygous deletions and 46,043 homozygous insertions that overlapped between the two datasets (Table 11). Comparison with the heterozygous deletions and insertions from the HuRef assembly yielded support for 9,431 deletions and 7,738 insertions, respectively (Table 10). These values represent a lower limit due to possible alignment issues in regions with tandem repeats. This dataset produced the largest overlap with the HuRef variant set compared to all others discussed below. However the Mills et al. published dataset used reads from the NCBI TraceArchive that we also used during assembly (i.e., Celera reads, donor HuBB). This suggests that essentially the same dataset used by two different groups produced an overlapping result by using different methods. As a consequence, we cannot determine which part of the overlapped variants with the Mills et al. data came from non-Celera sources, and therefore we cannot comment on novelty or polymorphic supporting evidence for HuRef variants.
Next, the HuRef homozygous deletions were compared to three other sets of previously identified deletion polymorphisms [56–58]. However, the overlap with these datasets was minimal, possibly due to the larger size of these variants (Table 11). Finally, the set of HuRef homozygous insertions was compared to those variants identified in an assembly comparison approach [59], and support was found for additional 243 insertion variants.
We sought further evidence in support of the longest indels identified by the one-to-one HuRef–NCBI mapping. We focused on the 20 longest insertions (9–83 kb) and the 20 longest deletions (7–20 kb) and examined the presence of these large indels in the genomes of eight other individuals by identifying fosmid clones that map to these 40 loci (Table S5). The fosmid mapping provided support for all 20 insertions, and 17 of 20 deletions. The lack of support for two of the deletions (Unique Identifiers 1104685056026, 1104685093410) is likely due to their location at the ends of HuRef scaffolds, which greatly reduces the possibility of mapping fosmids that span the insertion site. Support from multiple fosmids provides the strongest evidence for variation in indels between individuals. For example, the presence of a 24 kb insertion on Chromosome 22 (Unique Identifier 1104685552590) is supported by 13–17 fosmids in three individuals (with no evidence for absence), whereas its absence is supported by 19 fosmids in another individual (with no evidence for presence). These data suggest that the majority of large indels defined by the one-to-one HuRef–NCBI mapping are genuine variations among human genomes.
We selected 19 non-genic heterozygous indels in a nonrandom manner, ranging in length from 1 to 16 bp, for experimental validation using PCR coupled with PAGE detection of allelic forms. We ensured that the read depth coverage was in an acceptable range (not greater than 15 reads), suggesting that these loci were not in segmental duplications and would therefore not produce spurious PCR amplification. Three Coriell DNA samples and HuRef donor DNA were examined, and 15 out of 19 PCR assays assessed generated results consistent with the positive and negative controls. The indel lengths that yielded experimental data ranged from 1 to 8 bp in length. In four out of 15 indels, the heterozygote variant was identified in all four DNA samples, and in three out of 15, it was only found the HuRef donor DNA. For the remaining eight out of 15 cases, the indels were differentially observed among the four DNA samples (Figure S1).
We selected 51 putative homozygous HuRef insertions in a nonrandom manner for validation in 93 Coriell DNA samples based on their proximity to annotated genes, their size range of 100–1,000 bp, the absence of transposon repeat or tandem repeat sequence, uniqueness in the HuRef genome, and the absence of any similarity to chimpanzee sequence. The experimental results (Table S6) indicated that for 43 of 51 insertions (84%), we were able to generate specific PCR products for which the size of PCR products were as predicted and fell within the detectable range of the gel. For 84% of these 43 cases, insertions were identified in HuRef and additional DNA samples, and most follow Hardy-Weinberg equilibrium in CEU samples. Approximately 7% of the insertions tested (3 of 43) were false positives, because the HuRef donor DNA and all the 93 Coriell DNAs were homozygous for no insertion. In four insertions (9%), all of the tested Coriell samples displayed normal Hardy-Weinberg equilibrium; however, the insertion was absent in the HuRef sample. The inability to observe the insertion in the HuRef sample in these instances might be due to allelic dropout in the PCR process for the HuRef sample. This could be caused by specific SNPs at the primer annealing sites that were not accounted for during the primer design process.
In 22 (61%) confirmed experiments, the HuRef donor bears homozygous insertions in agreement with our computational analyses. There are four insertions in this set, among the 22, where the HuRef donor and all 93 Coriell DNA donors tested were homozygous for insertions. This suggests that these sequences were either not assembled in the NCBI human genome assembly or that the NCBI donor DNA sequenced had a rare deletion in these regions.
For the remaining 14 insertions (39%), the HuRef donor was heterozygous for the insertion instead of homozygous as was predicted by our indel detection pipeline. We searched for these alternative shorter alleles in the HuRef assembly and observed that two of the alternative alleles matched degenerate scaffolds and two matched singleton unassembled reads. These are sequence elements that are typically small or unassembled elements respectively, signifying that the assembly process selected one allele.
We note that many of the insertions tested (84%) are polymorphic in the Coriell panel tested, and although many are intronic, there are instances of UTR and exonic insertions whose impact on function may be more directly ascertained.
It has previously been shown that extended regions of high sequence identity complicate de novo genome assembly [10,60,61]. An analysis was undertaken to assess how well the segmental duplications (identified as regions of >5 kb with >90% sequence identity) annotated in the NCBI assembly are represented in the HuRef genome sequence. We analyzed the NCBI sequence (90.1 Mb) external to the one-to-one mapping with the NCBI assembly for segmental duplication content by comparison to the Human Segmental Duplication Database (http://projects.tcag.ca/humandup/) [61]. More than 70% of these nucleotides (63.6 Mb) are contained within segmental duplications, compared with 5.14% across the entire NCBI assembly. This suggests that the regions of the NCBI assembly that are not aligned to HuRef likely result from the absence of assembled segmental duplication regions in HuRef. This is further supported by the fact that only 57.2% of all regions annotated as segmental duplications in NCBI are present in HuRef. Clearly, these are some of the most difficult regions of the genome to represent accurately with a random shotgun approach and de novo assembly. However, it is also important to note that at least 25% of segmental duplication regions differ in copy number between individuals [62], and the annotation of such sequences will certainly differ between independent genomes.
Copy number variants (CNVs) have been identified to be a common feature in the human genome [11,15,62–64]. However, such variants can be difficult to identify and assemble from sequence data alone, because they are often associated with the repetition of large segments of identical or nearly identical sequences. We tested for CNVs experimentally to compare against those annotated computationally, and also to discover others not represented in the HuRef assembly. We used comparative genomic hybridization (CGH) with the Agilent 244K array and Nimblegen 385K array, as well as comparative intensity data from the Affymetrix and Illumina SNP genotyping platforms (using three analysis tools for Affymetrix and one for Illumina). In total, 62 CNVs (32 losses and 30 gains) were identified from these experiments (Table S7). It is noteworthy that the Agilent and Nimblegen CGH experiments, as well as the analysis of Affymetrix data using the GEMCA algorithm, were run against a single reference sample (NA10851). Therefore, a subset of the regions reported as variant may reflect the reference sample rather than the HuRef donor, even though all previously identified variants in the reference sample [62] were removed from the final list of CNV calls in the present study. The majority of the variant regions were detected by only one platform, reflecting the difference in probe coverage and sensitivity among various approaches [12,62]. As an independent form of validation, the CNVs detected here were compared to those reported in the Database of Genomic Variants (DGV) [63], and 54 of the variants (87%) have been described previously (with the thresholds used for these analyses we expect approximately 5% of calls to be false positive). A summary of the genomic features overlapped by these CNVs is presented in Table 12. Approximately 55% of the CNVs overlap with annotated segmental duplications, which is slightly higher than reported in previous studies [63,64]. The CNVs also overlap 95 RefSeq genes, seven of which are described in the Online Mendelian Inheritance in Man database (OMIM) as linked to a specific phenotype (Table S7). These include blood group determinants such as RHD and XG, as well as a gain overlapping the coagulation factor VIII gene.
Numerous HuRef sequences that span the entire or partial scaffolds did not have a matching sequence in the NCBI genome. Some had putative chromosomal location assignments (e.g., sequences extending into NCBI gaps), whereas others were unanchored scaffolds with no mapping information. We selected sequences >40 kb in length with no match to the NCBI genome and identified fosmids (derived from the Coriell DNA NA18552) mapping to these sequences based clone end-sequence data. The fosmids were then used as FISH probes with the aim of confirming annotated locations for anchored sequences and assigning chromosomal locations to unanchored scaffolds. Fosmids were hybridized to metaphase spreads from two different cells lines. At least 10 metaphases were scored for each probe, and a differentially labeled control fosmid was included for each hybridization. For 23 regions, there was no mapping information available from mate-pair data or the one-to-one mapping comparison. Of the remaining 26 regions, 24 had a specific chromosomal location assigned at the nucleotide level (Figure 10A and 10B), whereas two regions were assigned to specific chromosomes but lacked detailed mapping information. The results of the FISH experiments are outlined in Table S8. Of the 23 regions with no prior mapping information, 13 gave a single primary mapping location (Figure 10C). The majority of the remaining 10 regions located to multiple centromeric regions (Figure 10D), suggesting that there are large euchromatic-like sequences present as low-copy repeats in the current centromeric assembly gaps. For the 26 regions with mapping information, the expected signal was observed for 22 (85%). However, in six of these hybridizations, there were additional signals of equal intensity at other locations. Ten of the scaffolds chosen for FISH extend into contig or clone gaps in the current reference assembly. Of these 10 regions, the expected localization was corroborated for seven. The combined data indicate that the HuRef assembly contributes significant amounts of novel sequence important for generating more complete reference assemblies.
Haplotypes have more power than individual variants in the context of association studies and predicting disease risk [65–67] and also permit the selection of reduced sets of “tagging” SNPs, where linkage disequilibrium is strong enough to make groups of SNPs largely redundant [68,69]. The potential for shotgun sequences from a single individual to be used to separate haplotypes has been examined previously [70,71]. For a given polymorphic site, sequencing reads spanning that variant can be separated based on the allele they contain. For data from a single individual, this amounts to separation based on chromosome of origin. When two or more variant positions are spanned by a single read, or occur on paired reads derived from the same shotgun clone, alleles can be linked to identify larger haplotypes. This is sometimes known as “haplotype assembly.” When single shotgun reads are considered, the problem is computationally tractable [70,71] but the resulting partial haplotypes would be quite short with reads produced by existing sequencing technology, given the observed density of polymorphisms in the human genome (R. Lippert, personal communication). Mate pairing has the potential to increase the degree of “haplotype assembly,” but finding the optimal solution in the presence of errors in the data has been shown to be computationally intractable [71]. Nevertheless, we show that the character and quality of the data is such that heuristic solutions, while not guaranteed to find the best possible solution, can provide long, high-quality phasing of heterozygous variants.
The set of autosomal heterozygous variants described above (n = 1,856,446) was used for haplotype assembly. The average separation of these variants on the genome was ∼1500 bp (twice the average read length). Fewer than 50% of variants could be placed in “chains” of six or more variants where successive variants were within 1 kb of one another. Consequently, single reads cannot connect these variants into large haplotypes. However, the effect of mate pairing is substantially greater than would be observed simply by doubling the length of a read, as shown in Figure 11: variants are linked to an average of 8.7 other variants.
Using this dataset, haplotype assembly was performed as described in Materials and Methods. Half of the variants were assembled into haplotypes of at least 401 variants, and haplotypes spanning >200 kb cover 1.5 Gb of genome sequence. The full distributions of haplotype sizes, both in terms of bases spanned and in terms of numbers of variants per haplotype, are shown in Figure 12. Although haplotypes inferred in this fashion are not necessarily composed of continuous variants, haplotypes do in fact contain 91% of the variants they span. More than 75% of the total autosomal chromosome length is in haplotypes spanning at least four variants, and 89% of the variants are in haplotypes that include at least four heterozygous HapMap (phase I) variants.
Both internal consistency checks and comparison to HapMap data indicate that the HuRef haplotypes are highly accurate. Comparing individual clones against the haplotypes to which they are assigned, 97.4% of variant calls were consistent with the assigned haplotype. Moreover, the HuRef haplotypes were strongly consistent with those inferred as part of the HapMap project [35]. Where a pair of variants is in strong LD according to the HapMap haplotypes, the correct phasing of the HuRef data would be expected to match the more frequent phasing in the HapMap set in most cases. Exceptions would require a rare recombination event, convergent mutation in the HuRef genome, or an error in the HapMap phasing in multiple individuals.
We accessed the 120 phased CEU haplotypes from HapMap and identified the subset of heterozygous HuRef SNP variants that also coincided with the HapMap data. For adjacent pairs of such variants that were in strong LD (r2 ≥ 0.9; n = 197,035), fewer than 1 in 40 of the HuRef-inferred haplotypes conflicted with the preferred HapMap phasing. Figure 13 shows more generally the consistency of HuRef haplotypes with the HapMap population data as a function of r2 and D′. Because the inference of HuRef haplotypes is completely independent of the data and methods used to infer HapMap haplotypes, this is a remarkable confirmation of the HuRef haplotypes.
The restriction to variants in strong LD has no clear selection bias with respect to our inferred haplotypes. On the other hand, it provides only weaker confirmation for the HapMap phasing, since it is restricted to the easiest cases for phasing using population data—namely only those pairs of variants in strong linkage disequilibrium.
The lengths and densities of the inferred HuRef haplotypes described above are possible due to the use of paired end reads from a variety of insert sizes. Given the relatively simple means that were used for separating haplotypes, the high accuracy of phasing is likewise due to the quality of the underlying sequence data, the genome assembly, and the set of identified variants. The rate of conflict with HapMap with regard to variants in high LD can be further decreased by filtering the variants more aggressively (particularly excluding indels; unpublished data), although at the expense of decreasing haplotype size and density. It is also possible to improve the consistency measures described above by using more sophisticated methods for haplotype separation. One possibility we have explored is to use the solutions described above as a starting point in a Markov chain Monte Carlo (MCMC) algorithm. This produces solutions for which the fraction of high LD conflicts with HapMap is reduced by ∼30%. This approach has other advantages as well: MCMC sampling provides a natural way to assess the confidence of a partial haplotype assignment. Assessment of this and other measures of confidence is a topic for future investigation.
We used the generated haplotypes to view how well they span the current gene annotation. We were able to identify 84% (19,407 out of 23,224 protein coding genes) of Ensembl version 41 genes partially contained within a haplotype block and 58% of protein coding genes completely contained within a haplotype block. We note that in population-based haplotypes, denser sampling of SNPs in regions of low LD leads to reduction in the size of the average haplotype block [72]. In contrast to this finding, detection of additional true heterozygous variants through personal sequencing, regardless of LD, would lead to larger partial haplotypes, because additional variants increase the density of variants and thus their linkage to one another.
The sequencing, assembly, and cataloguing of the variant set and the corresponding haplotypes of the HuRef donor provided unprecedented opportunity to study gene-based variation using the vast body of scientific literature and extensively curated databases like OMIM [73] and Human Genetic Mutation Database (HGMD, [18]). A preliminary assessment indicates that 857 OMIM genes have at least one heterozygous variant in the coding or UTR regions, and 314 OMIM genes have at least one nonsynonymous SNP (Figure 14A). Overall, we observed 11,718 heterozygous and 9,434 homozygous coding SNPs and 236 heterozygous and 627 homozygous coding indels (Figure 14B). In addition, 4,107 genes have 6,114 nonsynonymous SNPs indicating that at least 17% (4,107/23,224) of genes encode differential proteins. The nonsynonymous SNPs define a lower limit of a potentially impacted proteome, because 44% of genes (10,208/23,224) have at least one heterozygous variant in the UTR or coding region and these variants could also affect protein function or expression. Therefore, almost half of the genes could have differential states in this diploid human genome, and this estimate does not include variation in nonexonic regions involved in gene regulation such as promoters and enhancers.
Understanding potential genotype-to-phenotype relationships will require many more extensive population-based studies. However, the complexities of assessing genotype–phenotype relationships begin to emerge even from a very preliminary glimpse of an individual human genome (Table 13). For Mendelian conditions such as Huntington disease (HD), the predictive nature of the genomic sequence is more definitive. Our data reveal the donor to be heterozygous (CAG)18/(CAG)17 in the polymorphic trinucleotide repeat located in the HD gene (HD affected individuals have more than 29 CAG repeats) [74]. The genotype matches the phenotype in this case, since the donor does not have a family history of Huntington disease and shows no sign of disease symptoms, even though he is well past the average onset age. The HuRef donor's predisposition status for multifactorial diseases is, as expected, more complicated. For example, the donor has a family history of cardiovascular disease prompting us to consider potentially associated alleles. The HuRef donor is heterozygous for variants in the KL gene; F352V (r9536314) and C370S (rs9527025). It has previously been observed that these heterozygous alleles present a lower risk for coronary artery disease [75]. However, the donor is also homozygous for the 5A/5A in rs3025058 in the promoter of the matrix metalloproteinase-3 (MMP3) [76]. This genotype is associated with higher intra-arterial levels of stromelysin and has a higher risk of acute myocardial infarction. This observation highlights the forthcoming challenge toward assessing the effects of the complex interactions in the multitude of genes that drive the development and progression of phenotypes. On occasion, these variant alleles may provide either protective or deleterious effects, and the ascertainment of resulting phenotypes are based on probabilities and would need to account for impinging environmental effects.
In our preliminary analysis of the HuRef genome, we also identified some genetic changes related to known disease risks for the donor. For example, approximately 50% of the Caucasian population is heterozygous for the GSTM1 gene, where the null mutation can increase susceptibly to environmental toxins and carcinogens [77–79]. The HuRef assembly identifies the donor to be heterozygous for the GSTM1 gene. Currently, it is not possible without further testing (including somatic analysis) and comparison against larger datasets to determine if this variant contributes to the reported health status events experienced by the donor, such as skin cancer.
We also found some novel changes in the HuRef genome for which the biological consequences are as yet unknown. For example, we found a 4-bp novel heterozygous deletion in Acyl-CoA Oxidase 2 (ACOX2) causing a protein truncation. ACOX2 encodes an enzyme activity found in peroxisomes and associates intimately with lipid metabolism and further was found to be absent from livers of patients with Zellweger syndrome [80]. The deletion identified would likely abolish peroxisome targeting, but the biological function of the mutation remains to be tested.
We have also been able to detect inconsistencies between detected genotypes in the donor's DNA and the expected phenotype based on the literature given the known phenotype of the HuRef donor. For example, the donor's LCT genotype should confer adult lactose tolerance according to published literature [81], but this does not match with the self-reported phenotype of the donor's lactose intolerance. Apparent inconsistencies of this nature may be explained by considering the modifying effect of other genes and their products, as well as environmental interactions.
We describe the sequencing, de novo assembly, and preliminary analysis of an individual diploid human genome. In the course of our study, we have developed an experimental framework that can serve as a model for the emerging field of en masse personalized genomics [82]. The components of our strategy involve: (i) sample consent and assessment, (ii) genome sequencing, (iii), genome assembly, (iv) comparative (one-to-one) mapping, (v) DNA variation detection and filtering, (vi) haplotype assembly, and (vii) annotation and interpretation of the data. We were able to construct a genome-wide representation of all DNA variants and haplotype blocks in the context of gene annotations and repeat structure identified in the HuRef donor. This provides a unique glimpse into the diploid genome of an individual human (Poster S1).
The most significant technical challenge has been to develop an assembly process (points ii–v) that faithfully maintains the integrity of the allelic contribution from an underlying set of reads originating from a diploid DNA source. As far as we know, the approach we developed is unique and is central to the identification of the large number of indels less than 400 bp in length. We attempted de novo recruitment of sequence reads to the NCBI human reference genome, using mate pairing and clone insert size to guide the accurate placement of reads [83]. Although this approach can produce useful results, it does limit variant detection to completed regions of the reference genome and, like genome assembly, can be confounded by segmentally duplicated regions.
The genome assembly approach with allelic separation allows the detection of heterozygous variants present in the individual genome with no further comparison. The one-to-one mapping of our HuRef assembly against a nearly completed reference genome permits the detection of the remaining variants. These variants arise from sequence differences found within and also outside the mapped regions, where the precision of the compared regions is being provided by the genome-to-genome comparison [59]. The ability to provide a highly confident set of DNA variants is challenging, because more than half of the variants are a single base in length but include both SNPs and indels. A filtering approach was used that accounts for the positional error profile in a Sanger sequenced electropherogram in relation to the called variant. Additional filtering considerations necessitated minimal requirements for read coverage and for the proportional representation of each allele. The filtering approaches were empirical and used the large amounts of previously described data on human variation (dbSNP). The utility of using paired-end random shotgun reads and the variant set defined on the reads via the assembly enabled the construction of long-range haplotypes. The haplotypes are remarkably well constructed given that the density of the variant map is comparable to those used in other studies [35], reflecting the utility of underlying sequence reads beyond just genome assembly. To understand how an individual genome translates into an individual transcriptome and ultimately a functional proteome, it is important to define the segregation of variants among each chromosomal copy.
While several new approaches for DNA sequencing are available or being developed [84–86], we chose to use proven Sanger sequencing technology for this HuRef project. The choice was obviously motivated in part for historical reasons [7], but not solely. We attached a high importance to generating a de novo assembly including maximizing coverage and sensitivity for detecting variation. We further anticipated that long read lengths (in excess of 800 nucleotides), compatibility with paired-end shotgun clone sequencing, and well-developed parameters for assessing sequencing accuracy would be required. High sequence accuracy is essential to avoid calling large numbers of false-positive variants on a genome-wide scale. Long paired-end reads are especially useful for achieving the best possible assembly characteristics in whole-genome shotgun sequencing and for providing sufficient linkage of variants to determine large haplotypes.
We have been able to categorize a significant amount of DNA variation in the genome of a single human. Of great interest is the fact that 44% of annotated genes have at least one, and often more, alterations within them. The vast majority—3,213,401 events (78%) of the 4.1 million variants detected in the HuRef donor—are SNPs. However, the remaining 22% of non-SNP variants constitute the vast majority, about 9 Mb or 74%, of variant bases in the donor. Using microarray-based methods, we also detected another 62 copy number variable regions in HuRef, estimated to add some 10 Mb of additional heterogeneity. Given these potential sources of measured DNA variation, we can, for the first time, make a conservative estimate that a minimum of 0.5% variation exists between two haploid genomes (all heterozygous bases, i.e., SNP, multi-nucleotide polymorphisms [MNP], indels, [complex variants + putative alternate alleles + CNV]/genome size; [2,894,929 + 939,799 + 10,000,000]/2,809,547,336) namely those that make up the diploid DNA of the HuRef assembly. We also note that there will be significantly more DNA variation discovered in heterochromatic regions of the genome [87], which largely escaped our analysis in this study.
We had mixed success when attempting to find support for the experimentally determined CNVs in the HuRef assembly itself or the data from which it was derived. More than 50% of the CNVs overlapped segmental duplications, and these regions are underrepresented in HuRef, which complicated the analysis. We attempted to map the sequence reads onto the NCBI human genome and then identify CNVs by detecting regions with significant changes in read depth. However, we found significant local fluctuations in read depth across the genome, limiting the ability for comparison and suggesting that a higher coverage of reads may be required to use this approach effectively.
As we have emphasized throughout, a major difference of the genomic assembly we have described is our approach to maintain, wherever possible, the diploid nature of the genome. This is in contrast to both the NCBI and WGSA genomes, which are each consensus sequences and, therefore, a mosaic of haplotypes that do not accurately display the relationships of variants on either of the autosomal pairs. For BAC-based genome assemblies such as the NCBI genome assembly, the mosaic fragments are generally genomic clone size (e.g., cosmid, PAC, BAC), with each clone providing contiguous sequence for only one of the two haplotypes at any given locus. Moreover, there are substantial differences in the clone composition of different chromosomes due to the historical and hierarchical mapping and sequencing strategies used to generate the NCBI reference assemblies [7,8].
In contrast, for WGSA, the reads that underlie most of the consensus sequence are derived from both haplotypes. This can result in very short-range mosaicism, where the consensus of clustered allelic differences does not actually exist in any of the underlying reads. To address this issue, the Celera assembler was modified to consider all variable bases within a given window and to group the sequence forms supporting each allele before incorporation into a consensus sequence (see Materials and Methods). In our experience, this reduces the incidence of local mosaicism, although, between windows, the consensus sequence remains a composite of haplotypes. Efforts to build haplotypes from the genome assembly (Haplotype Assembly) will likely lead to future modification of the assembler, allowing it to output longer consensus sequences for both haplotypes at many loci. Clearly, a single consensus sequence for a diploid genome, whether derived from BACs or WGS, has limitations for describing allelic variants (and specific combinations of variants) within the genome of an individual.
Partial haplotypes can be inferred for an individual from laboratory genotype data (e.g., from SNP microarrays) in conjunction with population data or genotypes of family members. However, at least in the absence of sets of related individuals (e.g., family trios), it is difficult to determine haplotypes from genotype data across regions of low LD. We have shown that sequencing with a paired-end sequencing strategy can provide highly accurate haplotype reconstruction that does not share these limitations. The assembled haplotypes are substantially larger than the blocks of SNPs in strong LD within the various populations investigated by the HapMap project. In addition to being larger, haplotypes inferred in our approach can link variants even where LD in a population is weak, and they are not restricted to those variants that have been studied in large population samples (e.g., HapMap variants). We note that in addition to the implications for human genetics, this approach could be applied to separating haplotypes of any organism of interest—without the requirement for a previous reference genome, family data, or population data—so long as polymorphism rates are high enough for an acceptable fraction of reads or mate pairs to link variants.
There are several avenues for extending our inference of haplotypes. As noted, although the naive heuristics used here give highly useful results, other approaches may give even more accurate results, as we have observed with an MCMC algorithm. There are various natural measures of confidence that can be applied to the phasing of two or more variants, including the minimum number of clones that would have to be ignored to unlink two variants, or a measure of the degrees of separation between two variants. The analysis presented here provides phasing only for sites deemed heterozygous, but data from apparently homozygous sites can be phased as well, so we can tell with confidence whether a given site is truly homozygous (i.e., the same allele is present in both haplotypes) or whether the allele at one or even both haplotypes cannot be determined, as occurs as much as 20% of the time with the current dataset. Lastly, it should be possible to combine our approach with typical genotype phasing approaches to infer even larger haplotypes.
Our project developed over a 10-year period and the decisions regarding sample selection, techniques used, and methods of analysis were critical to the current and continued success of the project. We anticipated that beyond mere curiosity, there would be very pragmatic reasons to use a donor sample from a known consented individual. First and foremost, as we show in a preliminary analysis, genome-based correlations to phenotype can be performed. Due to the still rudimentary state of the genotype-phenotype databases it can be argued that at the present time, DNA sequence comparisons do not reveal much more information than a proper family history. Even when a disease, predisposition, or phenotypically-relevant allele is found, further familial sampling will usually be required to determine the relevance. Eventually, however, populations of genomes will be sequenced, and at some point, a critical mass will dramatically change the value of any individual initiative providing the potential for proactive rather than reactive personal health care. In a simple analogy, absent of family history, genealogical studies can now be quite accurate in reconstructing ancestral history based purely on marker-frequency comparisons to databases. Here, with a near-unlimited amount of variation data available from the HuRef assembly, we can reconstruct the chromosome Y ethno-geneographic lineage (Figure 15), which is not only consistent with, but better defines the self-reported family tree data (Figure 1A and unpublished data).
There are always issues regarding the generation and study of genetic data and these may amplify as we move from what are now primarily gene-centric studies to the new era where genome sequences become a standard form of personal information. For example, there are often concerns that individuals should not be informed of their predisposition (or fate) if there is nothing they can do about it. It is possible, however, that many of the concerns for predictive medical information will fall by the wayside as more prevention strategies, treatment options, and indeed cures become realistic. Indeed we believe that as more individuals put their genomic profiles into the public realm, effective research will be facilitated, and strategies to mitigate the untoward effects of certain genes will emerge. The cycle, in fact, should become self-propelling, and reasons to know will soon outweigh reasons to remain uninformed.
Ultimately, as more entire genome sequences and their associated personal characteristics become available, they will facilitate a new era of research into the basis of individuality. The opportunity for a better understanding of the complex interactions among genes, and between these genes and their host's personal environment will be possible using these datasets composed of many genomes. Eventually, there may be true insight into the relationships between nature and nurture, and the individual will then benefit from the contributions of the community as a whole.
We used the assembled chromosome sequence of the human genome available as NCBI version 36. The gene annotation of this genome was provided by Ensembl (http://www.ensembl.org) version 41, which incorporates dbSNP version 126. Haplotype map data was obtained from http://www.hapmap.org, Release version 21a. Celera-generated chromatograms for the HuBB individual [7] were obtained from the NCBI trace archive. These included reads from two tissues sources: blood and sperm. Sequence reads were generated from these traces using Phred version 020425.c [88] and a modified version of Paracel TraceTuner (http://sourceforge.net/projects/tracetuner/). This reprocessing significantly improved accuracy and quality in the 5′ portion of the reads, increasing their usable length by 7%, and reducing variants encoding spurious protein truncations, as well as reducing apparent heterozygous variants in the assembly.
200-μl aliquots of thawed, whole blood were processed using the MagAttract DNA Blood Mini M48 Kit and the MagAttract DNA Blood >200 μl Blood protocol on the BioRobot M48 Workstation running the GenoM-48 QIAsoft software (version 2.0) (Qiagen; http://www.qiagen.com). Tris:EDTA (10:0.1) was used for the final 200 μl elution step. A260/A280 readings (SPECTRAmax Plus spectrophotometer (Molecular Devices; http://www.moleculardevices.com) or an ND-1000 spectrophotometer (NanoDrop Technologies; http://www.nanodrop.com), and gel images were used to quantify the DNA and to confirm that high-quality, high–molecular weight DNA was available for downstream processing. 1.0 μl of extracted DNA was run on a 0.8% agarose gel containing ethidium bromide, for 4 h at 60 V and imaged using Gel Doc and Quantity One Software (Bio-Rad Laboratories; http://www.bio-rad.com).
Phytohemagglutin-stimulated lymphocytes from peripheral blood were cultured for 72 h with thymidine synchronization. G-banding analysis was performed on metaphase spreads from peripheral blood lymphocytes using standard cytogenetic techniques.
Spectral karyotyping was performed on metaphase spreads from cultured lymphocytes. SkyPaint probes were used according to manufacturer's instructions (Applied Spectral Imaging; http://www.spectral-imaging.com). Metaphases were viewed with a Zeiss epifluorescence microscope and spectral images were acquired with an SD300 SpectraCube system and analyzed using SkyView software 1.6.2 (Applied Spectral Imaging).
Plasmid and Fosmid Library Construction. We nebulized genomic DNA to produce random fragments with a distribution of approximately 1–25 kb, end-polished these with consecutive BAL31 nuclease and T4 DNA polymerase treatments, and size selected using gel electrophoresis on 1% low–melting-point agarose. After ligation to BstXI adapters, we purified DNA by three rounds of gel electrophoresis to remove excess adapters, inserted fragments into BstXI-linearized medium-copy pBR322 plasmid vectors, and inserted the resulting library into GC10 cells by electroporation. To ensure that plasmid libraries contained few clones without inserts and no clones with chimeric inserts, we used vectors (pHOS) that include several features: (i) the sequencing primer sites immediately flank the BstXI cloning site to avoid sequencing of vector DNA, (ii) there are no strong promoters oriented toward the cloning site, and (iii) BstXI sites for cloning facilitate a high frequency of single inserts and rare no-insert clones. Sequencing from both ends of cloned inserts produced pairs of linked sequences of ∼800 bp each. We constructed fosmid libraries with approximately 30 μg of DNA that was sheared using bead beating and repaired by filling with dNTPs. We used a pulsed-field electrophoresis system to select for 39–40 kb fragments, which we ligated to the blunt-ended pCC1FOS vector.
Clone Picking and Inoculation. Libraries were propagated on large-format (16 × 16 cm) diffusion plates and colonies were picked for template preparation using a Q-bot or Q-Pix colony-picking robots (Genetix; http://www.genetix.com) and inoculated into 384-well blocks.
DNA Template Preparation. We prepared plasmid DNA using a robotic workstation custom built by Thermo CRS, based on the alkaline lysis miniprep [89], modified for high-throughput processing in 384-well plates. The typical yield of plasmid DNA from this method was approximately 600–800 ng per clone, providing sufficient DNA for at least four sequencing reactions per template.
Sequencing Reactions. Sequencing protocols were based on the di-deoxy sequencing method [90]. Two 384-well cycle-sequencing reaction plates were prepared from each plate of plasmid template DNA for opposite-end, paired-sequence reads. Sequencing reactions were completed using Big Dye Terminator (BDT) chemistry version 3.1 Cycle Sequencing Ready Reaction Kits (Applied Biosystems) and standard M13 forward and reverse primers. Reaction mixtures, thermal cycling profiles, and electrophoresis conditions were optimized to reduce volume and extend read lengths. Sequencing reactions were set-up by the Biomek FX (Beckman Coulter; http://www.beckmancoulter.com) pipetting workstations. Templates were combined with 5-μl reaction mixes consisting of deoxy- and fluorescently labeled dideoxynucleotides, DNA polymerase, sequencing primers, and reaction buffer. Bar coding and tracking promoted error-free transfer. Amplified reaction products were transferred to a 3730xl DNA Analyzer (Applied Biosystems).
The Celera Assembler Software (https://sourceforge.net/projects/wgs-assembler/) [7,40,91] generated contiguous sequences (contigs) that could be linked via mate-pair information into scaffolds. It has a phase for splitting initial apparently chimeric contigs (referred to as unitigs), but this process is not repeated for the final set of contigs and scaffolds as with some other assemblers (Arachne 2 [92]). This leaves a small number of chimeric scaffolds, which can be detected and split as described below. All assemblers fail to discriminate alternate alleles in polymorphic regions from distinct regions of the genome. These polymorphic regions, containing highly repetitive sequence with short unique anchoring sequence and simple algorithmic failures, result in a number of small scaffolds that are highly redundant. Although there are valuable data in these small scaffolds, they are usually not treated as part of the assembled sequence.
For this project we made specific modifications to the Celera Assembler to enable the grouping of reads into separate alleles when heterozygous variants were encountered. Instead of taking a column-by-column approach to determine the consensus sequence from a set of aligned reads, the region of variation was considered as a whole, defined as that between at least 11 bp nonvariant columns. In practice, variant regions would most frequently be single columns (SNPs), but the new algorithm only applied to longer regions. The reads spanning a variant region were split between alleles. An allele, for this purpose, was one or more spanning reads sharing an identical sequence for the variant region, and was considered confirmed if represented by two or more reads. Each allele was assigned a score equal to the sum of average quality values for the spanning portions of its reads. The highest-scoring confirmed allele was used for the consensus sequence. Alternate confirmed allele sequences were reported separately. As expected, there were usually two confirmed alleles in each region of sequence variation. Regions with more than two apparent confirmed alleles represented either collapsed repetitive sequence or a group of reads with systematic base calling error, rather than true genetic variation.
The set of The Institute for Genome Research (TIGR) BAC ends [41] used in the WGSA [40] assembly were aligned to the 553 HuRef scaffolds of at least 100 kb in length. We kept BAC ends that mapped uniquely to a single scaffold and near the end of a scaffold, such that their mate was likely to reside outside of the scaffold. Mate pairs were kept if both BAC ends passed the above criterion, and these indicated a possible joining of two scaffolds in a certain orientation. There were 144 consistent scaffold joins with at least two supporting mate pairs and 98 with one supporting mate pair. Using these scaffold joins would result in 409 or 311 scaffolds, respectively, of at least 100 kb, with a concomitant increase in the scaffold N50 length.
We used open-source software (http://sourceforge.net/projects/kmer/) [40,93,94] to generate a one-to-one comparison between HuRef and NCBI human genome reference assembly. For sequences that do not contain very large, nearly identical duplications, this mapping is accurate [93]. Nearly identical duplicated regions tend to be underrepresented in whole-genome shotgun assemblies such as HuRef [10]. Segments that are duplicated in one sequence but not the other (for instance when failing to merge overlapping contigs) cannot be fully included in any one-to-one mapping. For example the first few megabases of NCBI version 36 Chromosomes X and Y are identical; therefore, a 1.5-Mb scaffold from HuRef that maps to both of these regions is not part of the one-to-one mapping. Tandem repeats with variable unit copy number are also problematic for a one-to-one mapping.
For each one-to-one mapping we determined three levels: matches, runs, and clumps. A match is a maximal high-identity local alignment, usually terminated by indels or sequence gaps in one of the assemblies. Runs may include indels, and are monotonically increasing or decreasing sets of matches (linear segments of a match dot plot) with no intervening matches from other runs on either axis. Clumps are similar to runs but allow small intervening matches/runs (such as small inversions) to be skipped over. The total number of base pairs in matches is a measurement of how much of the sequence is shared between assemblies. Within a run, the number of base pairs in each assembly is different, because indels are allowed among matches in the run. These could be gaps that are filled in one assembly but not the other, polymorphic insertions or deletions, or artifactual sequence. Runs span regions in both assemblies that have no rearrangements with respect to each other, providing a direct measure of the order and orientation differences between a pair of assemblies. Clumps provide a similar measure of rearrangement but allow for small differences that may be due to noise or polymorphic inversions. Remaining sequence may be unique to one assembly or the other, but some will also be large repetitive regions without good one-to-one mapping but present in some copy number in both assemblies. Apparently unique sequence may also represent some form of contaminant.
We determined an initial set of potentially chimeric scaffolds by finding those that contained more than one clump of at least 5,000 bp relative to NCBI version 36. By mapping all HuRef and Coriell fosmid mate pairs to NCBI human reference genome and to HuRef, we assessed whether mate pair constraints were violated at the potentially chimeric junctions. Accordingly, we split 12 scaffolds.
DNA variants were characterized by alignment of sequencing reads in the HuRef assembly and by comparison of regions of difference in the one-to-one HuRef to NCBI reference genome map. The contribution of each sequence read to a single position in the HuRef consensus was evaluated both during and after the assembly process to identify positions that contain more than one allele. This process identified heterozygous SNPs and indel polymorphisms, and typically two or more reads were required for the initial identification of an alternate allele. Homozygous SNPs and MNPs were identified when (respectively) single or multiple contiguous loci differed in the one-to-one mapping, and all underlying HuRef reads supported one allele. Finally, homozygous insertion or deletion loci were identified where the HuRef assembly had or lacked sequence relative to the NCBI assembly, respectively. These were commonly referred to as homozygous indels unless it was relevant for analysis purposes, computational or experimental, to refer to a homozygous insertion or deletion as a way of indicating presence or absence of the sequence, respectively, in the HuRef assembly.
DNA variations were identified by examining the base changes within the HuRef assembly multialignment and between the HuRef assembly and the NCBI reference human genome. 5,061,599 SNPs and heterozygous variants were identified initially, after which filters were applied to eliminate erroneous calls. For a potential SNP, each read supporting that SNP was considered, and if the QV was <15 at the putative SNPs position in the read, then the read was considered invalid and was discarded as evidence for that particular variant. We also observed that deletions were overcalled at the beginnings and ends of reads, and insertions were overcalled at the ends of reads (Figure S2). By using the relative positions in the read where overcalling was detected, we were able to invalidate reads contributing to indel variant calls. We further observed that the relative read positions at which overcalling occurred was dependant on whether the read source was produced at Celera or The J. Craig Venter Institute (JCVI). Thus, any Celera read containing a putative deletion at a relative read position ≤0.18 or ≥0.76 was considered invalid for that particular deletion. Correspondingly, any JCVI read containing a putative deletion, at the relative read position ≤0.07 or ≥0.81 was deemed invalid in contributing to that particular variant call. Any Celera read was deemed invalid if it contained an insertion at a relative read position ≥0.70, and any JCVI read with an insertion at relative read position ≥0.77 was discarded as evidence. These thresholds were determined by plotting the frequency of insertions and deletions with respect to read position, and choosing the value where the call frequency was twice that of baseline (Figure S2).
Subsequent to the quality value and read location filtering the remaining variants were inspected for the percentage, number, and directionality of reads supporting the alternate alleles. Additionally these variants were inspected for the total number of reads in their assembled locus and the repeat sequence status (transposon and tandem repeat). Transposon repeats were identified using the RepeatMasker program (http://www.repeatmasker.org), and tandem repeats were identified using the Tandem RepeatFinder program [48]. The distribution of the percentage of reads containing the minor allele for heterozygous SNP and indels in Figure S3 shows that a large fraction of those putative variants that are found in dbSNP version 126 have a “minor allele frequency” (fraction of reads supporting the allele with fewer reads) of at least 20% and 25% for SNPs and indels, respectively. Therefore, we decided to apply the following filters separately to the QV and read location filtered variants, calculating at each filter step the fraction of passing variants that could be found in dbSNP. The filters applied to allow variants to be counted as bona-fide were: (i) 20% reads support minor allele for heterozygous SNP and 25% reads support minor allele for heterozygous indels, and (ii) two or more reads supporting the variant. The results of this analysis are presented in Table 3 and discussed in the Results section.
Manual inspection showed that some neighboring variants identified within the one-to-one mapping of HuRef to the NCBI genome reference would be more precisely represented as one larger variant after realignment. To address these regions of clustered variants, we identified these problematic regions by clustering SNPs within 2 bp of each other or any non-SNP variants with 10 bp of another variant. For these variable regions, we recalled the variant(s) using the variant calling algorithm developed as part of the consensus sequence generation found in the Celera assembler.
Homozygous insertion/deletions were filtered in the same manner as SNPs and heterozygous variants. All variants that were not confirmed by two or more reads were eliminated, as were those that did not fulfill minimal requirements of at least one spanning mate pair, and that the inserted sequence on the HuRef assembly or deleted sequence on the NCBI assembly not contain any ambiguous bases
We estimate the population mutation parameter (θ) [43] as:
where K is the number of variants identified, L is the number of base pairs, and n is the number of alleles. For indels, K is the number of indel events. In the case of a single diploid genome, n = 2, so a and b reduce to 1. Then θ = K/L, which is simply the number of heterozygous variants divided by the length sequenced. The standard deviation of θ reduces to θ:
Thus, the 95% confidence interval for θ is [0, θ+2θ] or [0, 3θ].
Two individuals of European ancestry were randomly selected from the SeattleSNPs data (http://pga.gs.washington.edu/) [95]. For the first individual, we constructed a haploid representation (without phasing) by randomly choosing one allele at each variant position. This reconstructed sequence is analogous to the NCBI genome sequence that we used to call HuRef homozygous variants. For the second individual, all variant positions were examined and scored. If the second individual was heterozygous at a position, then the heterozygous count was incremented by one. If the second individual had a homozygous genotype that did not match the allele seen in the reconstructed sequence then the homozygous variant was incremented by one. The second individual is analogous to the HuRef assembly sequence, and this procedure mimics our variant-calling algorithm and our definitions for heterozygous and homozygous variants. One caveat is that the NCBI human genome sequence, while only being one sequence, represents multiple individuals, and thus possibly contains more rare alleles in its sequence.
We developed a statistical model based on our assembly read coverage in the single diploid genome and on the filtering criteria used for calling high confidence variants. We assumed that chromosomes containing each of the two alleles are equally likely to be sampled and that allele loci are independent. At a given heterozygous locus, the probability of observing both alleles in at least x reads follows the binomial distribution with p = 0.50 and n = depth of coverage, where x is defined by the filtering criteria. To calculate the false-negative rate genome wide, a Poisson distribution is also incorporated to estimate sequence depth at different loci, where λ is set to the genome sequence coverage (7.5 for SNPs, 5.5 for insertions, 4.9 for deletions, after read filtering is taken into account).
A number of heterozygous indels between 1 and 20 bp were manually selected for experimental validation by verifying trace quality in the region of the indel, read coverage depth, and repeat sequence status. In order to detect heterozygous indels from the HuRef assembly, we ran PCR-amplified genomic DNA on PAGE to look for homoduplex and heteroduplex bands. Large insertions and deletions were also recognized by this process.
Primers were designed by centering the targeted indel to produce amplicons 150–250 bp in length with the melting temperatures of these amplicons ranging between 70 °C and 86 °C. PCR for polymorphism analysis was carried out in 10-μl volume reactions containing 30 ng of purified genomic DNA, 1× PCR buffer, 20 μM deoxynucleoside triphosphates, 2 mM MgCl2, 8% glycerol, 0.18 μM primers, and 0.0375 U AmpliTaq Gold DNA polymerase. Post-amplification treatment of each sample involved digestion with shrimp alkaline phosphatase (0.5 U) and exonuclease I (1.76 U) for 45 min at 37 °C, 15 min at 50 °C, with heat inactivation for 15 min at 72 °C.
PAGE was carried out at room temperature for 4 h at 650 V (constant) in a standard vertical gel measuring 1 mm thick, 20 cm wide, and 30 cm long (apparatus Model SG-400–20, CBS Scientific Company Inc, http://www.cbssci.com). The native gel consisted of 10% acrylamide with the 40% acrylamide stock solution having an acrylamide/ N,N′-methylenebisacrylamide ratio of 29:1. The running buffer consisted of 1× TBE. A loading dye consisting of 2× BlueJuice (Invitrogen) was added to each amplified sample and 5 μl was loaded per gel lane. After electrophoresis, the DNA bands were visualized by staining with a 1:10,000 dilution of SYBR Gold (Invitrogen).
Fifty-one apparent homozygous insertions in the HuRef assembly were selected based on assembly structure (appropriate read depth coverage and supporting mate pair evidence), their proximity to annotated genes, and their size. The insertion sequences were from 100 to 1,200 bp with few repeat sequences, and no detectable alignments to human (NCBI 36) or chimpanzee [22] genomes. We tested 93 Coriell DNA donors in addition to the HuRef DNA sample: 21 samples of European origin (CEU - NA06985, NA07056, NA11832, NA11839, NA11840, NA11881, NA11882, NA11992, NA11993, NA11994, NA11995, NA12057, NA12156, NA12239, NA12750, NA12751, NA12813, NA12814, NA12815, NA12891, NA12892), 12 Han Chinese samples (NA18524, NA18526, NA18537,NA18545, NA18552, NA18562, NA18566, NA18572, NA18577, NA18609, NA18621, NA18635), 11 Japanese (Tokyo) samples (NA18940, NA18942, NA18945, NA18949, NA18953, NA18961, NA18964, NA18967, NA18981, NA18994, NA18998), 22 samples of Hispanic origin (NA17438, NA17439, NA17440, NA17441, NA17442, NA17443, NA17444, NA17445, NA17446, NA17448, NA17449, NA17450, NA17451, NA17452, NA17453, NA17454, NA17456, NA17457, NA17458, NA17459, NA17460, NA17461, 15 samples of African American origin (NA17101, NA17102, NA17103, NA17104, NA17105, NA17106, NA17107, NA17108, NA17109, NA17110, NA17111, NA17112, NA17113, NA17114, NA17115) and 12 samples of Yoruban origin (NA18502, NA18504, NA18855, NA18870, NA19137, NA19144, NA19153, NA19200, NA19201, NA19203, NA19223, NA19238). A 200-bp amplicon was designed for each insertion. By design, a homozygous insertion sequence yielded a single high–molecular weight band of (200 bp + the insertion size) on the agarose gel. Absence of the insertion would be detected as a single low molecular band of 200 bp alone and a heterozygous indel would be detected as presence of both bands.
The amplicons were classified according to theoretical melting temperatures (Tm). Standard GC content and high GC content amplicons (82 °C < Tm < 87 °C) were processed separately in the laboratory using optimized high-throughput PCR protocols enabling all amplifications to be performed in 384-well plates in a volume of 10 μl. The standard GC content PCR protocol was composed of 3.0 μl of 0.4 μM mixed forward and reverse primers, 3.0 μl of DNA (1.67 ng/μl) and 0.05 μl (0.25 Us) of AmpliTaq Gold DNA polymerase (Applied Biosystems). The high-GC PCR protocol comprised 3.0 μl of 1.2 μM mixed forward and reverse primers, 3.0 μl of DNA (10.0 ng/μl), and 0.075 μl (0.375 U) of AmpliTaq Gold DNA polymerase (Applied Biosystems). PCR was set up using a Biomek FX (Beckman Coulter) pipetting robot and a Pixsys 4200 (Cartesian Technologies; http://www.cartesiantech.com/) nanoliter dispenser. All PCR amplifications were performed on dual 384-well GeneAmp PCR System 9700 thermal cyclers (Applied Biosystems) under the following program: 96 °C for 5 min (1×); 94 °C for 30 s, 60 °C for 45 s, 72 °C for 45 s (40×); 72 °C for 10 min (1×); and a 10 °C final hold.
2.0 μl of PCR product was combined with 5.0 μl of diluted loading dye (Invitrogen) and run on a 2.0% agarose gel, containing ethidium bromide. Gels were run for 45 min at 90 V and imaged using a Gel Doc and Quantity One Software (Bio-Rad Laboratories). Gel images were manually evaluated for the presence or absence of expected products.
Segments of the human genome that were found exclusively in either HuRef or NCBI version 36 represent potential misassemblies or genuine variations. In order to distinguish between these possibilities, we attempted to confirm the existence of the largest one-to-one HuRef–NCBI indels in a collection of fosmid clones, derived from eight individuals (see Table S5 legend). Fosmid end reads were downloaded from the Trace Archive, and mapped to HuRef and NCBI human reference genome using Snapper (http://sourceforge.net/projects/kmer/). To avoid short allelic variants of single loci, the HuRef assembly included only scaffolds that spanned at least 30 kb. The initial alignments required a unique best score with at least 90% nucleotide identity for at least 25% of the read length. Pairs of end read alignments were then filtered sequentially to retain only those that mapped to the same scaffold (HuRef genome) or chromosome (NCBI reference genome), in a tail-to-tail orientation, and within three standard deviations of the mean insert length. First, regions of the HuRef genome that failed to map to NCBI reference genome in the one-to-one mapping and were spanned to an average depth of 10x by fosmids that failed to map to the NCBI reference genome were identified as potentially novel segments. Their sequences were aligned to NCBI using ncbi-blastn (-W 100), and novelty was defined by the absence of nucleotide identity (≥98%) for lengths of ≥1 kb in spans of at least 35 kb. Second, the mapping coordinates of clones that mapped discordantly to either HuRef or NCBI were intersected with the 40 largest one-to-one HuRef-NCBI–derived indels to identify fosmid clones that support the existence of these indels in other human genomes. To define inserted DNA, we required one fosmid end to map within the insert exclusive to one assembly, the other to map within flanking sequence common to both assemblies, and inconsistent mapping to the genome assembly that lacked the insertion. Defining absence of inserted DNA required the fosmid mapping to span the putative insertion point in the assembly that lacked the insertion, and inconsistent mapping to the assembly that contains the insertion.
Haplotypes of heterozygous variants were inferred using a greedy heuristic with iterative refinement of the initial solution.
Data Encoding. An SNP matrix (rows = reads or mate pairs, columns = variants) was constructed as follows: for each variant location, reads whose sequence matched the consensus sequence were assigned state “0,” while reads not matching the consensus were assigned state “1.” A pair of mated reads was merged into a single row only if they were in the same scaffold, with the expected orientation and separated by the expected distance (± 3 SD). Thus, a row in the matrix correspond to one of the following: (i) a pair of mated reads with consistent placements and (ii) a single unmated read or single mated read whose mate is not consistently placed.
Initial Haplotype Construction. Initial partial haplotypes were constructed by repeating the following sequence of steps until all rows were assigned. From the remaining set of unassigned rows (initially all), choose the row with fewest missing elements. Use this row to seed a partial haplotype pair (i.e., assign the row to one haplotype, which is initialized with the non-missing states from this row, and initialize the other haplotype with the complementary states). Until no more rows share non-missing information, identify the row that has the strongest signal (i.e., number of columns indicating one haplotype minus number of columns indicating the other haplotype is maximal), and assign that row to the indicated haplotype, extending the haplotypes to include any additional columns that are non-missing for that row. When no unassigned rows overlap the current haplotypes, consider this pair of partial haplotypes final and go back to the beginning.
Iterative Haplotype Refinement. When all rows have been assigned to partial haplotypes, each haplotype pair and the rows it includes can be refined iteratively, repeating the following two steps until no changes result. First, for each column (variant position) in the haplotypes, determine by majority rule the state assignment of each haplotype. Second, for each row (read or mate pair), determine the haplotype assignment by majority rule.
Measurement of Haplotype Sizes. For each pair of partial haplotypes, two measures of size are natural: the number of variants that are phased and the distance in bp from the first to the last variant. In addition to the average of such values, the N50 statistic indicates a haplotype size that encompasses at least half of the variants.
Comparison of Phasing to HapMap. Consistency of HuRef haplotypes with HapMap haplotypes was assessed as follows. Within each partial HuRef haplotype, variants that were present in Phase I HapMap data were identified (henceforth “HapMap variants”). For each pair of HapMap variants that were adjacent in a HuRef haplotype, two measures were determined. The first was the degree of LD between the paired variants from the HapMap CEU panel. The second was the conditional probability of observing the HuRef haplotype in the CEU panel given the observed genotypes. When r2 ≥ 0.9 and the conditional probability was <0.5, this was considered a clear conflict of HuRef and HapMap haplotypes.
The HuRef sample was genotyped in duplicate on each of the GeneChip Human (500K) Mapping NspI and StyI Array Sets (Affymetrix; http://www.affymetrix.com), according to the manufacturer's instructions and as described previously [96]. Each array contains an average of 250,000 SNP markers. The arrays were scanned using the Gene Chip Scanner 3000 7G and Gene Chip Operating System. The call rate was >96% for all four all hybridizations; 0.1% discordant genotype calls between the technical replicates were excluded from further analysis.
The NspI and StyI array scans were analyzed for copy number variation using a combination of DNA Chip Analyzer (dChip) [97], Copy Number Analysis for GeneChip (CNAG) [98], and Genotyping Microarray-based CNV Analysis (GEMCA) [99]. Analysis with dChip (http://www.dchip.org) was performed using a Hidden Markov Model (HMM) as previously described [100], and a set of 50 samples run in the same facility were used as reference. For analyses with CNAG version 2.0 (http://www.genome.umin.jp), the copy number changes were inferred using a HMM built into CNAG [98]. GEMCA analysis was performed essentially as described [99], except that we used one designated DNA samples (NA10851) as reference for pair-wise comparison. This sample has been screened for CNVs in a previous study [62] and the CNVs known to be present in the reference genome were excluded.
The HuRef sample was genotyped using the Sentrix HumanHap650Y Genotyping BeadChip according to the manufacturer's instructions. All chips were scanned using the Sentrix Bead-Array reader and the Sentrix Beadscan software application. The results from the BeadChip were analyzed for CNV content using QuantiSNP as previously described [101].
The Agilent human genome CGH array contains 244,000 60mer probes on a single slide. The experiment was run using 2.5 μg of genomic DNA for Cy3/Cy5 labeling for each hybridization, with a standard dye-swap experimental design. DNA sample NA10851 was used as a reference. The slides were scanned at 5-μm resolution using the Agilent G2565 Microarray Scanner System (Agilent Technologies; http://www.agilent.com). Feature extraction was performed using Feature Extraction v9.1 and results were analyzed using CGH Analytics v3.4.27.
CGH was performed using the Nimblegen human genome CGH array. The array contains 385,000 isothermal probes yielding a median spacing of 6 kb across the human genome. The experiment was performed as previously described [102] with a standard dye-swap experimental design. Results were analyzed using the CNVfinder algorithm [103]. One of the dye-swap experiments did not meet the quality control cut-offs, and because of this, the Nimblegen CNV calls were only employed for confirmation of CNV identified by the other platforms, and not used for identification of additional CNVs
FISH analysis was performed to find the location of DNA segments present in the HuRef DNA but either missing or represented by gaps in HuRef assembly. The FISH analysis was performed as previously described [104]. Initially, fosmids representing 107 different regions were chosen and end-sequenced to confirm that they mapped to the intended scaffolds. After excluding fosmids for which the original mapping was erroneous or uncertain, 88 fosmids remained. The entire sequence for each fosmid was then computationally excised from the scaffolds sequence and analyzed for repeat content using RepeatMasker. Fosmids with more than 6 kb (∼17%) satellite repeat content were excluded from further analysis. All fosmids that passed these filtering criteria were analyzed on metaphase spreads from two different cell lines (GM10851 and GM15510) to determine the chromosomal location of the fosmid probe. At least 10 metaphases were scored for each probe, all in duplicate by two experienced cytogeneticists.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession number for sequences discussed in the paper are: AADD00000000 (WGSA) and ABBA01000000 (the consensus sequences of both HuRef scaffolds and chromosomes).
|
10.1371/journal.pgen.1004727 | Unifying Genetic Canalization, Genetic Constraint, and Genotype-by-Environment Interaction: QTL by Genomic Background by Environment Interaction of Flowering Time in Boechera stricta | Natural populations exhibit substantial variation in quantitative traits. A quantitative trait is typically defined by its mean and variance, and to date most genetic mapping studies focus on loci altering trait means but not (co)variances. For single traits, the control of trait variance across genetic backgrounds is referred to as genetic canalization. With multiple traits, the genetic covariance among different traits in the same environment indicates the magnitude of potential genetic constraint, while genotype-by-environment interaction (GxE) concerns the same trait across different environments. While some have suggested that these three attributes of quantitative traits are different views of similar concepts, it is not yet clear, however, whether they have the same underlying genetic mechanism. Here, we detect quantitative trait loci (QTL) influencing the (co)variance of phenological traits in six distinct environments in Boechera stricta, a close relative of Arabidopsis. We identified nFT as the QTL altering the magnitude of phenological trait canalization, genetic constraint, and GxE. Both the magnitude and direction of nFT's canalization effects depend on the environment, and to our knowledge, this reversibility of canalization across environments has not been reported previously. nFT's effects on trait covariance structure (genetic constraint and GxE) likely result from the variable and reversible canalization effects across different traits and environments, which can be explained by the interaction among nFT, genomic backgrounds, and environmental stimuli. This view is supported by experiments demonstrating significant nFT by genomic background epistatic interactions affecting phenological traits and expression of the candidate gene, FT. In contrast to the well-known canalization gene Hsp90, the case of nFT may exemplify an alternative mechanism: Our results suggest that (at least in traits with major signal integrators such as flowering time) genetic canalization, genetic constraint, and GxE may have related genetic mechanisms resulting from interactions among major QTL, genomic backgrounds, and environments.
| Biological traits often display large amounts of genetic variability as well as genetic correlations among traits. This variability provides the raw material for evolutionary change and may alter the direction of trait evolution under selection. Despite this importance, it is unclear whether the genetic controls of variability in single traits and relationships among multiple traits have related mechanisms. Using the flowering time of a plant species as model, here we performed genetic mapping and identified a locus altering single-trait variability and multi-trait relationships. The effect likely results from the distinct thresholds required by its different alleles to trigger flowering, which can be explained by the interaction among this major locus, the variable genomic backgrounds, and the distinct environments. This view is supported by experiments showing epistatic effects of this major locus on flowering time and expression pattern of the candidate gene. Together, our results show that, at least for traits with major signal integrator genes such as flowering time, the genetic control of single-trait variability and multi-trait relationships may have a common underlying mechanism that may be generalizable to other genes or pathways, mediated by interaction among major loci, genomic backgrounds, and surrounding environments.
| Elucidating the genetic basis of quantitative traits enables analysis of the processes that shape trait evolution [1]. Two properties that characterize quantitative traits in a population are mean and (co)variance. Most genetic mapping studies focus on loci that influence trait means, while the genomic regions capable of altering trait (co)variances, in contrast, receive relatively little attention despite the crucial role of trait (co)variances in determining the potential for traits to respond to selection [but see 2]. Although researchers have long recognized the importance of genetic control over phenotypic variance [3]–[6], only recently have analyses begun to map genomic regions or genes responsible for trait (co)variance in organisms such as Arabidopsis thaliana [7]–[9], yeast [10], or mouse [11], [12]. In humans, recent studies also have identified variance-controlling loci with biomedical implications [13]–[16].
For a quantitative trait, the phenotypic variation, VP, can be decomposed as VP = VE+VG+VGxE+Ve, where VE is the major environmental variance from different growth chambers or experimental gardens, VG is the genetic variance, VGxE stands for the genotype-by-environment interaction, and Ve is the stochastic noise caused by micro-environmental differences or other factors. For experiments across several different major environments, ‘plasticity’ consists of VE and VGxE. Within the same major environment (where VE and VGxE can be ignored) canalized traits exhibit little phenotypic variation (VP) [6], and causes of phenotypic canalization can be environmental or genetic. That is, one genotype produces a constant phenotype despite environmental variation (environmental canalization reducing Ve), or genetically distinct individuals produce similar phenotypes despite different genetic backgrounds (genetic canalization reducing VG) [6], [17], [18]. Most mapping studies of canalization loci focused on environmental canalization, where the stochastic noise (Ve or similar measures) of inbred lines was often used as a quantitative trait in standard mapping algorithms [7], [9]–[12]; only in a few cases are we aware of attempts to specifically map loci controlling genetic canalization [8], [12]. While Shen et al. [8] identified loci whose alleles differ in the genetic variance among inbred lines (VG), Fraser and Schadt [12] employed separate approaches to map loci controlling genetic canalization and environmental canalization. Here similar to both studies [8], [12], we focus on genetic canalization, which epistatically reduces genetic variance conferred by other genes [4], [6], [17] without changing the extent of molecular polymorphism in the genome. Beyond single traits, in this study we further consider the genetic control of the relationship among multiple traits.
The relationship among traits is often expressed in the form of their genetic variance-covariance matrix, the G matrix [19]. The G matrix of different traits in the same environment indicates the magnitude of potential genetic constraints, whereas the G matrix of the same traits in different environments is algebraically related to the genotype-by-environment interaction component of plasticity (VGxE, sometimes referred to as GxE below) [20]–[22]. While the term ‘genetic constraint’ may be used more broadly for many different combinations of traits (such as the same trait in different ages) [23], here we use this term only to describe the relationship among different traits in the same environment or age. Genetics may control or alter the magnitude of single-trait genetic variance and the size, shape, or orientation of multi-trait genetic covariance structure (genetic constraint or the genotype-by-environment component of plasticity). While to date several studies are available for the genetic mapping of single-trait genetic canalization, few ecologically or evolutionarily important loci controlling trait covariance structure have been investigated, and it is unclear whether these three attributes of quantitative traits (genetic canalization, genetic constraint, and GxE) have related underlying genetic mechanisms.
Because plants often synchronize their phenology with specific environmental conditions [24], flowering time provides a useful model to investigate these attributes. We examined Boechera stricta, an ecological model organism closely related to Arabidopsis [25], [26]. Previously we showed that flowering time is under strong selection in nature [25], and a large-effect phenology QTL (nFT) exhibits trade-offs in flowering probability and fitness across different natural environments [27], [28], providing a good model to investigate the interaction between genetics and environments. Here, using phenological traits in this species, we provide a unifying framework by identifying the same QTL that alters the magnitude of genetic canalization, genetic constraint, and GxE, and proposing that the genetic mechanism likely involves major QTL by genomic background by environment interaction effects. We further test the proposed epistasis effect between major QTL and genomic background on plant phenological traits and candidate gene expression.
Using recombinant inbred lines (RIL) between B. stricta parents from Colorado and Montana, we performed mapping for loci that affect the among-RIL genetic variance (i.e., genetic canalization) of phenological traits (flowering time and plant size [number of leaves] at flowering) in six different growth chamber environments. We identified three quantitative trait loci (QTL) under the stringent genome-wide significance threshold 0.01 (i.e., the observed Brown-Forsythe value is larger than the genome-wide maximum value from at least 990/1000 permutations) (Table 1, Figure S1). The significance threshold of 0.05 had additional QTL identified, but in this study we only focused on QTL with stronger effects. Two QTL (BST031941 and Bst004238) only had canalization effects on flowering time in one environment (Figure S1, 16 hour days, 25°C, 4 week vernalization; i.e., long days, elevated temperature, short winter). The effects of these two QTL were opposite (Figure S2): the Colorado genotype at BST031941 on linkage group 2 reduced among-RIL genetic variance of (canalized) flowering time, and the Montana genotype at Bst004238 on linkage group 7 caused flowering time canalization.
The third QTL, nFT, had widespread canalization effects in multiple environments. This QTL is syntenic with the region containing flowering time gene FT (AT1G65480, hence the name “near FT”) in Arabidopsis thaliana and is a major QTL influencing Boechera stricta phenology, life history and fitness in a broad range of environments [25], [27], [28]. The draft genome assembly from Joint Genome Institute also indicated that the FT gene locates within this region. The canalization effect of nFT was environment-dependent: under the genome-wide significance threshold of 0.01, its effect was significant for flowering time in four environments (both vernalization lengths in 12 hour 18°C and 16 hour 25°C treatments, Figure 1) and number of leaves at flowering in two environments (both vernalization lengths in 12 hour 18°C, Figure S3). For flowering time, ambient environment significantly influenced the canalization effect, but the duration of vernalization did not (Figure 1). Interestingly, the magnitude and direction of nFT's canalization effect depended on the environment (Figure 2). The Montana genotype reduced the among-RIL genetic variance of (i.e., canalized) phenological traits at 12 hour days and 18°C, but the Colorado genotype had this canalization effect in the 16 hour days, 25°C treatment. Although it is known that the existence and magnitude of genetic canalization vary among environments [29], as far as we know, our study may be the first to show that the direction of canalization effect can be reversed between environments (Figure 2).
Based on these observations, we proposed a ‘threshold hypothesis’ to explain the environment-dependent magnitude and direction of nFT's canalization effect (Figure 3), hypothesizing that the FT gene may be the causal locus within the nFT QTL: In growth chambers, the Montana genotype of nFT accelerated flowering in all environments [25]. In addition to nFT, these recombinant inbred lines also segregate for other flowering time genes in the genome [25]. Since the FT gene within the nFT QTL serves as a hub, which promotes flowering after integrating signals from multiple upstream pathways in the Arabidopsis flowering time network [30], the environment-dependent canalization effect of nFT may be created by the interaction among these factors: 1) different growth chamber environments, 2) distinct genomic backgrounds among RILs created by segregating genotypes of other flowering genes, and 3) the different input-signal threshold needed for either FT genotype to trigger flowering (Figures 1 and 3). For plants in the flowering-promoting environment under long days and cool temperature (16 hour days 18°C, mean flowering time at 4 week vernalization = 141 days, and at 6 week vernalization = 122 days), nFT did not confer a canalization effect because all genomic backgrounds created high input signals for both FT genotypes to initiate flowering. Plants took longer to flower under short days and cool temperatures (12 hour days 18°C, mean flowering time = 147 or 139 days, at 4 or 6 week vernalization, respectively). In these slightly flowering-inhibiting environments, most genomic backgrounds generated lower input signals, which were still enough for the low-threshold Montana genotype to express (predominantly flowering in the first season, within 180 days), whereas a portion of families with the Colorado genotype did not flower until the second growing season because many genomic backgrounds did not generate enough input signals for the high-threshold Colorado genotype (Figures 1 and 3). This resulted in higher variance among nFT Colorado homozygotes, such that Montana was the canalization genotype in these environments. Finally, flowering was significantly delayed at elevated temperatures (16 hour days 25°C, mean flowering time at 4 week vernalization = 229 days, at 6 week vernalization = 190 days). In these flowering-inhibiting environments where most genomic backgrounds generated low input signal, some genomic backgrounds still had enough signal for the Montana genotype to flower in the first season, but some families with Montana genotype flowered only in the second season (Figure 1 and Figure 3). The majority of families homozygous for the Colorado genotype, however, did not flower until the second season due to nFT Colorado genotype's high threshold. Therefore, the Colorado genotype canalized the onset of reproduction in this environment.
While genetic canalization concerns the variance of single traits, the covariance structure of multiple traits may also be affected by genetic elements that canalized some traits but not others. The environment- and trait-dependent canalization effect of nFT therefore suggests that it may affect trait covariances (the G matrix; the pairwise genetic correlations were reported in Table S1). Indeed, we identified nFT as the strongest QTL altering the size (Box's M [31]) and orientation (the angle between the first principal component Gmax) of G matrices in different trait-by-environment combinations: flowering time in six environments, leaf number at flowering in six environments, and the combination of all 12 traits (Figure S4). The Krzanowski index method (see Materials and Methods) [32], [33], on the other had, only identified significant nFT effect on the covariance structure of flowering time between the six environments but not for plant leaf number at flowering. In contrast to Gmax angle, the Krzanowski method is likely conservative because it compares the angular difference between subspaces formed by many dimensions, some of which may explain little variation. Figure 4 indicated the difference in size, shape, and orientation between the G matrices from the two nFT genotypes, and the respective trait loadings on each axis were reported in Table S2. Another QTL, BST031941, was also mapped by the Box's M method for the six-flowering-time data set (Figures S4 and S5).
We next mapped QTL controlling the G matrix between flowering time and leaf number in the same environment (indicating potential genetic constraint) and between the same traits in different environments (indicating the magnitude of the genotype-by-environment interaction component of plasticity). Again, nFT and BST031941 were the only QTL influencing most trait pairs. Similar to the canalization result for univariate traits (above), nFT altered the genetic covariance structure between flowering time and leaf number only in both vernalization lengths of two ambient environments: 12 hour days, 18°C and 16 hour days, 25°C, but not 16 hour days, 18°C (Figure 5). The nFT effects, however, were not strong (Figure 5) and only significantly influenced the relative size (Box's M) but not the orientation (Gmax angle) between two G matrices in each case. BST031941, on the other hand, had effects only in environment 16 hour days, 25°C, 4 week vernalization and altered both the size and orientation of the G matrix (Figure S6), consistent with the previous observation that its canalization effect on flowering time only existed in this environment (Figure S1 and S2).
Finally, nFT influenced the magnitude of GxE (genotype by environment) interactions (Figure 6). For many trait pairs, nFT had significant effects on both the size and orientation of the covariance matrices, especially in the comparison between chambers 12 hour days 18°C and 16 hour days 25°C, where nFT had significant canalization effect with different directions on univariate traits (Figure 1 and S3). The BST031941 QTL had a significant effect on GxE for flowering time only in the comparison between 16 hour days, 25°C, 4 week vernalization and all other environments (Figure S7). In addition, the G matrices in each comparison differed in both size and orientation. This again is consistent with BST031941's environment-dependent effect described earlier.
Since nFT influenced the genetic variance and covariance of phenological traits, we hypothesized that this QTL might interact epistatically with other flowering time genes in the genome and alter the magnitude of their effects, as predicted by our ‘threshold hypothesis’. Our previous study did not identify significant epistatic QTL in these growth chambers [25], nor did our re-analyses identify significant interaction effect between nFT and other QTL (Table S3). These results, however, did not contradict our prediction, since the threshold hypothesis emphasized the interaction between major QTL (nFT) and the cumulative effect from upstream genes (the genomic background), and epistasis between nFT and individual upstream QTL may be too weak to detect in the previous experiment.
To test the epistatic effect between nFT and genomic background, we conducted greenhouse experiments using several heterogeneous inbred families (HIF). Each family had almost identical genomic background and segregated for the two nFT genotypes, and the use of several such families allowed the test for genomic background by QTL epistatic effects (Figure S8). Univariate analyses revealed nFT by genomic background (HIF) interactions for all three traits in the greenhouse environment (flowering time, leaf number and height at flowering, Table 2 and Figure 7 A to C). These effects remained significant after sequential Bonferroni correction. Previously, we detected significant nFT effects on mean flowering time in F6 recombinant inbred lines [25]. Here, analyses in the HIF experiments showed significant epistasis but not main nFT effects on flowering time. This difference likely reflects the complex effects of nFT that depend on genetic background (see Discussion) and is consistent with the idea that the additive effect of a QTL depends on other epistatic genes in the genome [34]. There was, however, a main effect of nFT QTL on height at flowering: Montana homozygotes at nFT are shorter at flowering than Colorado homozygotes.
In multivariate analysis (MANOVA) simultaneously treating all three traits as response variables, there were significant main effects for nFT (P = 0.024) and genomic background (P<0.001), as well as nFT by genomic background interaction effect (P = 0.005). Figure 7 D to F show nFT by genomic background reaction norms for each pair of traits, demonstrating this epistatic effect.
While our ‘threshold hypothesis’ focused on the effect of the FT gene, the HIF phenotypic experiments only showed effects of the nFT QTL. To further test this prediction that the FT gene interacts epistatically with other flowering genes in the genome, we used the same HIF experimental design (Figure S8) to test the nFT by HIF (genomic background) interaction effect on expression of the FT transcript, using two HIF backgrounds. This interaction effect was highly significant (Table 2). While the Montana genotype had low expression in both genomic backgrounds, the Colorado genotype had significantly higher expression in HIF 98A than in HIF 89A (Figure 8). In HIF 89A, both nFT genotypes conferred low FT gene expression, while in HIF 98A the Colorado nFT genotype conferred higher FT gene expression than the Montana genotype (Figure 8). This observation is consistent with the threshold hypothesis that the genetic variation of other flowering genes in the HIF 89A background does not generate enough input signals for either FT genotype to express. On the other hand, in the greenhouse the HIF 98A genomic background generated enough input signals for the Colorado FT genotype to express but not enough for the Montana genotype. In addition, we found a highly significant association between flowering phenotype (presence of visible flowering buds) and quantitative expression of the FT transcript in individual plants (F1,17 = 9.72, P value = 0.006). This supports that complex trait variation at the nFT QTL may be functionally mediated by the FT locus itself.
The genetic basis and evolutionary history of canalization in one trait, genetic constraint among multiple traits, and genotype-by-environment interaction (GxE) across different environments are active foci for research in evolutionary genetics [2]. While many genetic mapping studies are available for single-trait canalization, the majority concerned environmental canalization [7], [9]–[12], and only in a few studies are we aware of mapping for genetic canalization [8], [12]. Here we used similar approaches to previous studies for single-trait genetic canalization [8], [12], but extended our analysis beyond single traits and focused on the QTL altering trait covariance structures. We provide a unifying framework for these mechanisms in the flowering time of B. stricta by showing how one QTL (nFT, which contains the floral integrator gene FT) influences all three attributes for phenological traits of Boechera stricta. nFT influences single-trait genetic canalization, and the magnitude and direction of its effects can be reversed across environments. nFT's environmentally dependent and reversible effect in one trait, when extending to multiple traits, creates the distinct patterns of covariance structure among different traits in the same environment (altering the magnitude of genetic constraint) or among the same traits in different environments (altering the magnitude of GxE). We propose a ‘threshold hypothesis’ (see Results and Figure 3) to explain this pattern, and the hypothesis is supported by the significant nFT by genomic background epistatic effects on phenological traits and transcriptional variation of the candidate FT locus in the heterogeneous inbred family (HIF) experiment, echoing studies showing the prevalence of epistasis in both trait and molecular evolution [34]–[41].
In this study we proposed a ‘threshold hypothesis’ to explain the environment-dependent genetic canalization effects of nFT on flowering time, and the concept is illustrated in Figure 3. This hypothesis focuses on three components of flowering time regulation: the downstream floral signal integrator gene FT, the genomic backgrounds with different combinations of polymorphic upstream genes (each generating small input signals to FT), and multiple growth chamber environments that also generate different levels of input signals. In this hypothesis, the activation of FT expression (which then triggers flowering) depends on whether the upstream input signals, which vary with different combinations of genomic backgrounds and environmental stimuli, exceed the genotype-specific threshold of FT. The interaction among major threshold gene, genomic backgrounds, and environmental stimuli therefore triggers the environment-dependent canalization effect of flowering. The threshold model may also help explain how discrete and large-effect phenotypic changes may be caused by continuous and small-effect upstream genetic mechanisms.
In this simple threshold hypothesis, the effect of the genomic background on the initiation of FT expression is binary: input signals generated by different backgrounds or environments are either below or above the threshold for FT alleles to express. In a previous study of Arabidopsis thaliana, Welch et al. [42] used sigmoid functions to model genes in the flowering time pathway. These sigmoid functions model the relationship between quantitative upstream input signals and the response of downstream genes, and these models can be viewed as the quantitative generalization of our threshold hypothesis (see details in Figure S9). While in Welch et al. [42] the wild-type allele of each gene was assigned a sigmoid function and knockouts always have zero output, in our case the two FT alleles with different thresholds simply have different response curves (Figure S9). The threshold hypothesis is therefore supported by previous studies and may be applied to genes or networks that can be modeled with continuous functions.
Our threshold hypothesis also echoes the recent idea that cryptic genetic variation (CVG) is not ‘mechanistically special and mysterious’ [43], but instead that genetic effects are conditional on other genes (epistasis or dominance) or on the environment (GxE), and CVG can be viewed as ‘conditionally neutral genetic variation’ [43]. It is worth emphasizing that the mechanism underlying genetic canalization effects of nFT and the well-known heat shock protein 90 (Hsp90) [29], [44] may be different. Hsp90 has a specific function as a protein chaperone [45]–[47], whose expression pattern may be independent of the proteins it helps folding. On the other hand, expression of the FT gene is trigged by signals from upstream genes. As a consequence, we predict that distinct input-signal requirements for FT's different genotypes may generate QTL by genomic background interaction effects on phenological traits and on its own expression pattern, which is supported by our HIF experiments. Therefore, in contrast to Hsp90, whose specific function may not easily be extrapolated to most genes or pathways, our threshold hypothesis and the case of the FT gene in the flowering time pathway may represent an alternative mechanism for other cases of canalization, genetic constraint, and genotype-by-environment interaction in traits with major signal integrators such as flowering time.
In fact, a similar idea has been proposed and tested previously [48]–[50]. Siegal and Bergman [48] modeled the effect of biological networks on trait canalization and showed that canalization is ‘an inevitable consequence of complex developmental-genetic processes’ that does not require the force of stabilizing selection or a specific molecular function such as Hsp90. Further studies supported this model by showing that gene knock-outs can affect both genetic and environmental canalization of yeast gene expression [49] as well as phenotypes [50]. More importantly, they showed that ‘capacitors’ (canalization genes) are more likely to be network hubs [50], like the FT gene in this study. While these previous studies focus on artificial gene knockouts, here we provide an example of trait canalization from natural variation for ecologically important traits.
Different from other modeling studies, in this study our example of the threshold hypothesis separates the growth chambers into three categories (flowering promoting, slightly inhibitory, and strongly inhibitory) instead of by specific environmental factors (day length, ambient temperature, or vernalization length). For flowering time, pathways responding to different environmental signals converge at floral signal integrator genes such as FT, the current focus of the threshold hypothesis, and therefore it is more straightforward to classify environments based on their effect on flowering promotion (the horizontal axis in Figure 3). This is analogous to studies that classify research sites by their effect on plant growth or crop yield [51] and allows modeling flowering time variation without addition information on the expression of upstream genes. We recognize that the threshold hypothesis represents a simplification of the underlying continuous biological process into a qualitative factor (whether plants flower in the first season or not), and future work may explicitly consider the effect of individual environmental factors to allow detailed quantitative modeling of flowering time.
The link between epistasis and genetic canalization of flowering time has been established in the model plant Arabidopsis thaliana. Stinchcombe et al. [52] observed that the latitudinal cline in flowering time only exists in genotypes with the wild-type functional allele of FRIGIDA but not for the deletion allele (FRIΔ), and the flowering time of accessions bearing the deletion allele is canalized. This canalization effect is caused by epistasis between FRIGIDA and FLC [37], where the canalizing FRIΔ allele suppressed the effect between different alleles of FLC. Similarly, in our study we identified nFT as the canalization locus and demonstrated the epistatic effects in the HIF experiment, where the different nFT genotypes alter the phenotypic effect of different genomic backgrounds (specific allelic combinations of other flowering genes). In addition, as in other cases of canalization genes [29], we also find that nFT's canalization effect depends on the environment.
The adaptive value of canalization has received considerable attention [5], [6]. Theoretical analyses have modeled the conditions under which canalization will evolve [17], [53], [54] (but see [48]–[50]), and empirical studies in Drosophila have shown that traits with higher fitness effects have greater canalization [55], [56]. Rutherford et al. [29] suggested that the wild-type canalizing allele of Hsp90 gene in Drosophila may be favored because it buffers against potentially unfit background genetic variation. On the other hand, the non-canalizing allele might be favored when a trait is under directional or disruptive selection. Under both scenarios, however, the allelic polymorphism in canalization genes will be eliminated if trait canalization is universally favored or disfavored by natural selection. Our results provide a mechanism where the polymorphism of canalization genes can be maintained. Because which nFT genotype has the canalization effect depends on specific environments, different genotypes may be favored under distinct environments even with consistent natural selection for or against trait canalization, beyond mutation-selection balance [17]. For example, if natural selection favors flowering time canalization, in populations under 12 hour days 18°C the Montana genotype would be favored, and the Colorado genotype would be favored under 16 hour days 25°C because they are the canalization genotypes in these respective environments.
The environmental dependence of nFT canalization may also maintain the molecular polymorphism of other flowering genes in the genome. Canalization may change the selective influence on other genes by suppressing the genetic variation expressed in traits [18], [57]. In the case of heat shock protein 90 (Hsp90), the wild type allele buffers the phenotypic effect of potentially deleterious mutations in the genome, thereby reducing the force of purifying selection on these mutations [29], [44]. The accumulation of cryptic molecular variation in the genome may provide additional evolutionary potential [18] once the canalization effect is disrupted. The case of Boechera stricta flowering time, however, may be more complex due to the nFT by environment interaction effect on canalization. Even if the selection force on flowering time is identical across populations, other flowering loci in the genome may still be subject to different types and magnitude of natural selection depending on local environment and nFT genotype. Such variation in selection may in turn influence the molecular polymorphism of other flowering time genes.
In this study, flowering time was defined as the days elapsed since initial vernalization, excluding the duration of the second vernalization for plants that did not flower in the first growing season. Since the time of vernalization simulated ‘winter’ conditions with little plant growth, this approach quantifies the number of growing-season days the plants experienced before first flowering. Our results and the threshold hypothesis suggest that the canalization effect primarily reflects whether plants flowered in the first growing season, and therefore this approach captures both the variance between and within seasons. In addition to flowering time, the nFT locus also influenced the (co)variance of ‘leaf number when flowering’, a well-defined quantitative trait often used as an indicator of the reproductive phenology. We therefore think these phenological traits reflect important underlying biological processes.
The covariance structure among multiple traits (G matrix) indicates the magnitude of potential genetic constraints and can have profound effect on the magnitude and direction of trait evolution under selection [3], [58]–[60]. While many methodological and empirical studies have compared the G matrix evolution among evolutionary lineages [e.g., 3], only a few studies have investigated the genetic mechanism or identified the genomic regions responsible for this G matrix difference. In Arabidopsis, Stinchcombe et al. [61] found that two alleles of the ERECTA gene confer different structure of the G matrix among four traits. In mice, Wolf et al. [62] also identified significant epistatic pleiotropy effects of QTL on the covariance between traits. Both examples investigated the change of covariance structure caused by candidate loci, and to our knowledge, our study may be one of the first attempts at genome-wide mapping of QTL altering the covariance structure among multiple traits.
In addition to canalization in single traits, we have identified nFT as the major QTL altering the genetic covariance structure in both multivariate trait combinations: 1) between different traits in the same environment (indicating the magnitude of genetic constraint), and 2) the same trait between different environments (indicating the magnitude of GxE, the genotype-by-environment interaction component of plasticity). This supports the idea that genetic constraint and plasticity are related concepts [63] and can be connected by trait- or environment-dependent canalization: When one trait is canalized but another is not, the magnitudes of genetic constraint or GxE are altered [55]. For example, considering the relationship between flowering time and leaf number at flowering under 16 hour days, 25°C, 4 week vernalization, the nFT Colorado genotypes canalize flowering time, but there is no nFT canalization effect for leaf number (Figure 5). This changes the orientation of covariance structure by ∼35°. For GxE, the nFT Montana genotype canalizes flowering time in 12 hour days, 18°C, 4 week vernalization, and the Colorado genotype canalizes flowering time in 16 hour days, 25°C, 4 week vernalization. This leads to a significant difference in the orientation of the G matrices for the two nFT genotypes (∼59°, Figure 6, row 1, column 5). Therefore, it is helpful to view the multi-trait covariance structure as a multivariate extension of single-trait variance, and our result supports the idea that the change of magnitude in genetic constraint or GxE may be a consequence of trait- or environment-dependent canalization effects [64]. Our results also suggest that, in a threshold-like gene regulation system, the change of trait genetic covariance structure may be achieved simply by the shift of the gene activation threshold (Figures 3 and S9).
As described in the Introduction, plasticity consists of VE and VGxE. In this study we showed that the nFT QTL altered the magnitude of VGxE. It is also possible that genetic mechanisms altering VE exist. With this genetic mechanism (here termed macro-environmental canalization, as opposed to micro-environmental canalization which concerns Ve, the variance from stochastic noise), one genotype of the QTL may exhibit similar phenotypes across diverse environments while the other genotype has varying phenotypes. While we did not specifically map for this macro-environmental canalization, the genetic mechanism altering VE may also be realized from the threshold hypothesis (Figure 3): consider the flowering time under two hypothetical environments, one strongly promotes flowering (‘ENV1’ hereafter: all genomic backgrounds generate signal above the Colorado threshold – all dots are white in Figure 3) and the other has effect between ‘slightly inhibiting’ and ‘strongly inhibiting’ (‘ENV2’ hereafter: all genomic backgrounds generate input signal above Montana but below the Colorado threshold – all dots are grey in Figure 3). Given the low threshold of the FT Montana genotype, all genomic backgrounds flower in both environments, and the VE for the Montana genotype is low. For the Colorado genotype, however, all genomic backgrounds flower under ENV1 but none flowers in ENV2, and VE is large for the Colorado genotype. Therefore, although in terms of upstream input signals the VE stays the same regardless of FT genotypes, for flowering time FT may control macro-environmental canalization due to the threshold-type reaction to upstream signals. This also suggests that environment-dependent genetic canalization, macro-environmental canalization, and the alteration of the magnitude in GxE may represent different viewpoints of the same concept.
Results from the HIF experiment show strong nFT by genomic background epistatic effects on phenological traits and on expression of the FT locus, demonstrating: 1) nFT's role as an epistatic modifier of other flowering genes and 2) the effect of other flowering genes (different HIF genomic background) on gene expression of the FT locus itself, a network hub integrating signals from upstream genes in the flowering time pathway. These observations are consistent with the ‘threshold hypothesis’ illustrating how flowering pathway function can generate epistasis between FT and other flowering genes (Figure 3) and also echo studies with gene by genomic background epistatic effects in the flowering time pathway of Arabidopsis [65]. Unlike the strong and direct epistatic relationship between Arabidopsis flowering genes FRIGIDA (FRI) and FLOWERING LOCUS C (FLC) [37], [66], FT responds to the combined effect of multiple upstream pathways, and the one-to-one epistasis between nFT and individual flowering time loci may be too weak to be detected by our previous [25] or current study (Table S3). Our novel algorithm to map (co)variance QTL therefore serves as a valuable alternative to standard pairwise searches for epistasis, paralleling recent developments in human genetics [15], [67]. In the flowering time of B. stricta, this epistatic relationship between nFT and genomic background has been supported by our HIF experiment.
In this study we employed a linkage mapping approach to map (co)variance QTL, and the issue of sample size and statistical power may be a limitation [68]–[70]. We recognize that the experimental design may not have sufficient power to detect all QTL, especially those with minor effects. This limitation, however, does not affect the significant functional variation at nFT. Although it might be possible that the nFT QTL's effect on trait mean, variance, and covariance structure are effects of several closely linked genes, our threshold hypothesis and following HIF experiments both suggest that the FT gene may exhibit these pleiotropic effects: being a floral signal integrator, the two FT alleles may influence trait means due to different thresholds for activation by upstream signals (which predicts and is supported by our results that the two alleles vary in gene regulation patterns instead of amino acid substitutions). Such threshold differences may interact with various genomic backgrounds or environmental stimuli and thus alter the pattern of trait (co)variation (see Results and Figure 3). Further, it is not uncommon that genes or QTL can simultaneously control trait means and (co)variances, as previous studies mapping canalization loci have identified QTL or genes known to control trait means [7]–[10], and a recent study has shown strong genetic correlation between developmental instability (environmental canalization) and phenotypic plasticity [71]. Taken together, these results suggest that, at least in traits with major signal integrators such as flowering time, the control of trait means, (co)variances, and genotype-by-environment interaction may have a similar genetic basis.
All RIL data were obtained from our previous study [25]. Briefly, a cross was made between one genotype from Montana and one from Colorado [72], and F6 RIL were generated through self-pollination and single seed decent. From each family, one F6 individual was genotyped at 164 polymorphic molecular markers, with an average spacing of 5.5 cM between neighboring markers. The F6 RILs were predominantly homozygous (95.9%). Heterozygous genotype calls in any marker of any family were treated as missing data.
In the previous study, we measured flowering time and leaf number at flowering (N = 5 individuals/RIL/treatment and N = 35 individuals/parental line/treatment) in six distinct environments, composed of two vernalization lengths (four or six weeks) at 4°C and three growth conditions (12 hour days 18°C, 16 hour days 18°C, and 16 hour days 25°C). In this study, we analyze family mean trait values for the 178 RIL and 2 parental lines obtained from the previous study [25]. The growth chamber experiments consisted of two growing seasons. Individuals that had not flowered within 180 days after the first vernalization were subject to another 6-week vernalization. In addition, during the second growing season, plants from the 16 hour days 25°C chambers were moved to 16 hour days 18°C [25]. Two traits from each growing condition were used in this study: flowering time and leaf number at the time of first flowering. Flowering time is defined as the number of elapsed days since the end of the first vernalization, and the 6-week period of the second vernalization was excluded from the flowering time estimation. All traits were standardized to a mean of zero and standard deviation of one before further analysis.
For each genetic marker, we used the Brown-Forsythe test, a modification of Levene's test based on median, to estimate the difference in trait variance between the Colorado homozygote and the Montana homozygote at markers across the genome. This approach is similar to Shen et al. [8] and can detect QTL responsible for genetic canalization. We determined the statistical significance by the genome-wide permutation method of Churchill and Doerge [73]. One thousand permuted datasets were generated by randomizing trait values with respect to marker genotypes. The marker-trait relationship was randomized, but the genotype vector and the trait vector for each individual were not altered. From each permuted data set, the Brown-Forsythe statistic was calculated at each genetic marker, and the genome-wide maximum Brown-Forsythe value was recorded, providing a genome-wide null probability distribution. The P-value of the Brown-Forsythe statistics for each marker in the observed data was obtained by comparing this value to the null distribution of Brown-Forsythe values from the 1,000 randomized datasets. Our genome-wide permutation procedure provides a straightforward control for multiple tests across all markers and is also robust to violations of the assumption of multivariate normality. The estimation of variance and covariance, however, may be limited by small sample size, perhaps resulting from missing data or segregation distortion. To prevent possible bias, we therefore excluded six markers with minor allele frequency less than 0.33. All computations were performed in R (http://www.r-project.org/) using scripts available upon request from CL.
Considering the effect of a QTL, the total trait variation (Vp) can be decomposed into:where Vm is the variation explained by the difference in mean of the two homozygous genotypes, Vv is the variation explained by the difference in variance of the two genotypes, and Vr is residual variance arising from other sources [8]. For each significant QTL, we calculated the proportion of variation explained by Vm and Vv, following previously published equations designed for populations with two homozygous genotypes in each SNP [8]:where p and q are the genotype frequencies of the Colorado and Montana homozygotes, respectively. μCO and μMT are the mean, and σCO and σMT represent the standard deviation of the Colorado and Montana homozygotes.
Here we aim to map QTL altering the covariance structure of three groups of traits: 1) flowering time and plant size at flowering (number of leaves) in all six environments (G matrix with 12 traits); 2) flowering time in all environments (G matrix of six traits); 3) plant leaf number at flowering in all environments (G matrix with six traits). We further mapped QTL changing the covariance structure between flowering time and plant size at flowering separately for each environment (representing the magnitude of genetic constraint) and between the same trait in pairs of different environments (representing the magnitude of the genotype-by-environment interaction component of plasticity, GxE). Among the multiple ways to model plasticity (reviewed in [20], [74], here following Falconer [21]), we treat the same trait in distinct environments as separate traits and model their covariance structure. We choose this definition because this view generalizes both GxE and genetic constraint into the relationship among traits, allowing the use of established methods for G matrix comparisons.
For each molecular marker, we separated the data into two groups of homozygous genotypes. Two separate (co)variance matrices (G matrices) were estimated, and we assessed the QTL effect by comparing the G matrices via three methods: 1) Box's M statistics; 2) the angle between Gmax; 3) the Krzanowski index. Statistical significance is determined by the genome-wide permutation algorithm described above. We acknowledge that the covariance matrix estimated from family means may not be identical to the genetic covariance matrix estimated from individual-level mixed-model MANOVA. This simplification, however, was necessary to ensure computational feasibility since the G matrices needed to be calculated twice (one for each homozygous genotype of a marker) for ∼160 markers for each of the 1,000 permuted data sets.
Box's M statistic [31] compares the difference between the trace of multiple covariance matrices and the trace of their pooled covariance matrix:where g is the number of matrices to be compared (two in our case), Vi is the degrees of freedom, and Si is the trace of the i-th matrix. N, the overall degrees of freedom, is the sum of all Vi values. The trace, S, of the pooled covariance matrix is:Since the trace of a covariance matrix is the sum of its diagonal elements and is equal to the sum of eigenvalues from its principal components, the Box's M value could be interpreted as the difference between the multivariate volumes occupied by different covariance matrices. In our case, the Box's M method compares the overall size of G matrices from the two genotypes at each genetic marker. Traditionally the significance is determined by an F-test and is sensitive to deviations from multivariate normality, but our genome-wide permutation procedure alleviates this parametric distributional requirement.
Two covariance matrices could differ not only in size but also in their orientation. We used two methods to compare the orientation between G matrices [32], [75]. To estimate the radian angle between the respective Gmax (first principal component), we first calculated the angle between the first eigenvector, u and v, of the two G matrices respectively:where the dot symbol calculates the dot product between vectors. Since eigenvectors are directional, θ may be larger than π/2 (90 degrees, the maximum possible angular difference between two non-directional axes). Therefore if θ is larger than π/2, the radian angle between Gmax is calculated as π - θ, otherwise the angle equals θ. This method estimates the angular difference between the respective axes with most variation in each G matrix.
The Gmax method, however, has a caveat that when G matrices have many dimensions (traits), other principal components may carry substantial amounts of variation, and comparing Gmax may not be sufficient [32]. Therefore, we employed the method of Krzanowski [33], which has also been used in recent studies [32]. In brief, this method compares the k-dimensional subspace between two G matrices, where k is less than or equal to half of the dimension of the original G matrix. For example, in a data set with 10 traits, we estimated the degree of similarity between two subspaces formed by the first five eigenvectors of two G matrices. Similar to the Gmax method above, each eigenvector w that will be used in the analysis was first standardized by the square root of its dot product:The ‘matrix of similarity’ (S) then is calculated as:where A and B are matrices containing the first five standardized eigenvectors of the two respective G matrices, and superscript T denotes matrix transpose. As in other studies [32], we used the sum of eigenvalues of this S matrix (the Krzanowski index) as a measure of overall similarity between the two subspaces. This index ranges from 0 to 5 in our example of 10 traits, with 0 signifying non-overlap and 5 indicating total overlap between subspaces. For our purpose of mapping QTL whose different genotypes confer the most dissimilar G matrices, we compared the negative Krzanowski index of each marker to the 1,000 maximum negative Krzanowski indexes from permutation.
In summary, while the Gmax and Krzanowski methods compare the orientation of linear relationship among traits, Box's M tests the dispersion of points from this linear relationship. All mapping algorithms were written in R (http://www.r-project.org/). When only two traits are involved, the angle between Gmax captures all the difference in orientation between G matrices, and therefore the Krzanowski method is not necessary.
To test the existence of epistasis as predicted by the threshold hypothesis, we performed analysis of variance for the interaction effect between nFT and other flowering time QTL identified in the same growth chambers from our previous study [25]. The epistatic effects of other QTL were tested separately, using flowering time as response and nFT, the other QTL, and their interaction as fixed-effect predictor variables in each model.
We generated four heterogeneous inbred families (HIFs, Figure S8) to test the epistatic effect between nFT and genomic background (the cumulative effect of other genes in the genome). Based on the genotype data in the F6 generation [25] we identified four F5 parents heterozygous at nFT and mostly homozygous at other markers. From each F5 parent we planted approximately 250 seeds, which are self-full siblings of the original genotyped F6 individual in Anderson et al. [25] (N = 1097 plants total from four F5 parents). We genotyped the microsatellite marker C02 [∼5 cM from the FT gene,25] in all plants and collected seeds from those that were homozygous at C02. Seeds from the same F6 individual (a ‘family’ hereafter) have virtually identical genomic composition. All plants within the same HIF are nearly identical in other genomic regions but segregate for two nFT homozygous genotypes. With four HIFs that are different in genomic background, the interaction between nFT genotype and HIF (genomic background) provides a statistical test for epistatic effects on phenology.
The HIF experiment was conducted in the Duke University Greenhouse rather than in multiple growth chamber environments as in the RIL experiment. To provide independent replication for each homozygous nFT genotype, we selected at least 20 homozygous families from each HIF (24 families from HIF 3A; 22 families from HIF 89A; 23 families from HIF 98A; and 20 families from HIF 105A), for a total of 89 families. In November 2011 we place 10–15 seeds from each of the 89 families on moist filter paper in petri dishes in dark conditions at ambient temperature for 3 weeks until germination. As in other B. stricta greenhouse experiments [76], seedlings were then planted in Ray Leach SC10 ‘Cone-tainers’ (21 cm in depth and 3.8 cm in diameter, Stuewe & Sons Inc., Tangent, OR, USA), with the lower 80% of each Cone-tainer filled with Fafard 4P Mix soil (Conrad Fafard, Agawam, MA, USA) and top 20% with Sunshine MVP soil (Sun Gro Horticulture, Vancouver, BC, Canada). Greenhouse conditions were as follows: 16-hour days (6 AM to 10 PM), diurnal temperature of 18–21°C, and nocturnal temperature of 13–16°C. We used a random number generator to assign seedlings to distinct positions in 9 blocks, each containing 91–96 plants. Each block included individuals from all HIFs and most families from each HIF (In some cases, a family did not have enough siblings to be represented in each block). The blocks were rotated around a greenhouse bench once a week to minimize the effects of environmental gradients in the greenhouse.
In January 2012, all rosettes were vernalized at 4°C for 8 weeks. Plants were removed from vernalization on 29 February 2012, at which point we monitored them 7 days/week and recorded the date of first flowering as well as the number of leaves and plant height at first flowering. By April 23, 2012, we had collected phenological data from 8–10 full siblings per family (N = 785 F7 individuals flowered successfully). No individuals flowered after that date. Relevant data are available in Dataset S1.
Statistical analysis was performed with REML mixed-model ANOVA (Proc Mixed, SAS 9.3, SAS, Cary, NC). We first conducted a multivariate ANOVA (MANOVA) to address how the three response variables (day of first flowering, plant height and number of leaves at flowering) varied with HIF (3A/89A/98A/105A), nFT genotype (Montana/Colorado homozygote), and nFT by HIF interaction (all are fixed effects). We incorporated ‘family’ (nested within nFT homozygote, cross-classified with HIF) and block as random effects. We then conducted univariate ANOVA for each response variable with the same statistical model.
The contrasting canalization effect of the nFT locus in different environments suggests a three way interaction of the major QTL (nFT) by genomic background (the combination of other flowering-related genes) by environment conditions, and the interaction between nFT and genomic background on phenological traits is tested in the HIF experiment. The nFT locus contains the ortholog of the FT gene in Arabidopsis (AT1G65480). FT serves as a major hub for integrating upstream signals of flowering, and its expression often correlates with the onset of flowering [30]. If the variation in the FT gene in Boechera stricta is responsible for the differential canalization effect of the nFT QTL in our Montana by Colorado cross, its expression pattern should vary depending on the nFT genotype and genomic background. We therefore test the expression pattern of FT in the same HIF experimental design.
Two HIFs (HIF 89A and HIF 98A) were used in this experiment. Within each HIF we obtained five families from each homozygous nFT genotype for a total of 20 families. Forty experimental plants (two individuals from each family) were completely randomized, and all planting procedures and greenhouse environmental settings were as above. Rosettes were grown in the Duke greenhouse for 12 weeks and stratified at 4°C for 8 weeks. In Arabidopsis, FT mainly expresses in leaves, where protein translation happens, and the proteins are transferred to floral meristems [77]. We therefore collected one young leaf from each plant four weeks after vernalization ended. FT in Arabidopsis exhibits circadian rhythm in gene expression, and under 16-hour days, its maximum expression is in the end of daytime [78]–[80]. We therefore collected leaves from all 40 experimental plants around 10 pm, when the 16-hour Duke greenhouse days end. Leaves were immediately flash frozen in liquid nitrogen and stored at −80°C. RNA was extracted with Sigma Spectrum Plant Total RNA Kit, and cDNA was synthesized with Thermo Scientific DyNAmo cDNA Synthesis Kit. Two samples failed during the RNA extraction and cDNA synthesis steps, leaving 38 samples in total.
Our partial genomic sequencing shows that there may be more than one FT gene copy in Boechera stricta (Joint Genome Institute and Mitchell-Olds lab, unpublished). Therefore, we cloned and sequenced FT full-length coding sequences from both parents. Only one copy is expressed, and both parents have the same expressing copy with identical coding region sequences (KJ576855 and KJ576856 in GenBank, where the Montana genotype is denoted as ‘LTM’ and Colorado genotype as ‘SAD12’). All primer sequences are available in Table S4. FT gene expression was measured by quantitative PCR (qPCR) with Thermo Scientific DyNAmo SYBR Green qPCR Kits. Following previous experiments [81], the ACTIN2 gene (ACT2) is used as reference gene, and FT expression level for each of the 38 samples was calculated as:where CtACT2 is the Ct value in qPCR of the reference gene ACTIN2, and CtFT is the Ct value of FT. Since within each sample the Ct value of FT is always larger (i.e., the signal is lower) than ACT2, ΔCt is always negative, and larger ΔCt represents higher FT gene expression. The relative qPCR signal of FT to ACT2 can be calculated as 2ΔCt. This 2ΔCt value, however, has a skewed distribution among samples. Since log transformation of 2ΔCt yields a value that is proportional to ΔCt, we used the original ΔCt as the response variable for statistical analysis. Relevant data are available in Dataset S2.
Statistical analysis was performed as in the HIF phenotypic experiment, where nFT genotype, HIF, and nFT by HIF interaction were treated as fixed effects, and family was treated as a random effect nested within nFT and HIF. All 40 plants were grown in the same block, so no block effect exists for this experiment. To further test if FT expression in Boechera stricta is related with flowering, we recorded whether each of the 40 experimental plants had visible flowering buds during the time of leaf-tissue collection. The analysis incorporates ΔCt as the response variable, and the phenological indicator ‘whether a plant has visible bud’ as a fixed-effect categorical predictor, and family as random-effect predictor variable.
|
10.1371/journal.ppat.1004207 | Salmonella enterica Serovar Typhi Conceals the Invasion-Associated Type Three Secretion System from the Innate Immune System by Gene Regulation | Delivery of microbial products into the mammalian cell cytosol by bacterial secretion systems is a strong stimulus for triggering pro-inflammatory host responses. Here we show that Salmonella enterica serovar Typhi (S. Typhi), the causative agent of typhoid fever, tightly regulates expression of the invasion-associated type III secretion system (T3SS-1) and thus fails to activate these innate immune signaling pathways. The S. Typhi regulatory protein TviA rapidly repressed T3SS-1 expression, thereby preventing RAC1-dependent, RIP2-dependent activation of NF-κB in epithelial cells. Heterologous expression of TviA in S. enterica serovar Typhimurium (S. Typhimurium) suppressed T3SS-1-dependent inflammatory responses generated early after infection in animal models of gastroenteritis. These results suggest that S. Typhi reduces intestinal inflammation by limiting the induction of pathogen-induced processes through regulation of virulence gene expression.
| Bacterial pathogens translocate effector proteins into the cytoplasm of host cells to manipulate the mammalian host. These processes, e.g. the stimulation of small regulatory GTPases, activate the innate immune system and induce pro-inflammatory responses aimed at clearing invading microbes from the infected tissue. Here we show that strict regulation of virulence gene expression can be used as a strategy to limit the induction of inflammatory responses while retaining the ability to manipulate the host. Upon entry into host tissue, Salmonella enterica serovar Typhi, the causative agent of typhoid fever, rapidly represses expression of a virulence factor required for entering tissue to avoid detection by the host innate immune surveillance. This tight control of virulence gene expression enables the pathogen to deploy a virulence factor for epithelial invasion, while preventing the subsequent generation of pro-inflammatory responses in host cells. We conclude that regulation of virulence gene expression contributes to innate immune evasion during typhoid fever by concealing a pattern of pathogenesis.
| One function of the innate immune system in the intestinal tract is to generate temporary inflammatory responses against invasive enteric pathogens while avoiding detrimental overreaction against harmless commensal bacteria under homeostatic conditions. In contrast to commensal microbes, pathogenic microbes express an array of virulence factors to manipulate host cell functions. Pathogen-induced processes, also known as patterns of pathogenesis [1], activate specific pathways of the innate immune system, enabling the host to distinguish virulent microbes from ones with lower disease-causing potential. By detecting pathogen-induced processes the host can escalate innate immune responses to levels that are appropriate to the threat [2].
Salmonella enterica serovar Typhimurium (S. Typhimurium), an invasive enteric pathogen associated with human gastroenteritis, triggers acute intestinal inflammation in the terminal ileum and colon, thereby producing symptoms of diarrhea and abdominal pain within less than one day after ingestion [3]. The inflammatory infiltrate in the affected intestinal tissue is dominated by neutrophils [4], [5]. Similarly, neutrophils are the primary cell type in the stool during acute illness [6]–[8]. In contrast, individuals infected with serovar Typhi (S. Typhi) develop a febrile illness (typhoid fever) with systemic dissemination of the organism. In contrast to Salmonella-induced gastroenteritis, only a third of patients develop diarrhea that is characterized by a dominance of mononuclear cells in the stool [6]. The dominant cell type in intestinal infiltrates is mononuclear, while neutrophils are infrequent [9]–[11]. Unlike S. Typhi, interaction of S. Typhimurium with intestinal model epithelia induces hepoxilin A3-dependent transmigration of neutrophils [12]. Moreover, infection of human colonic tissue explants with S. Typhimurium results in the increased production of the neutrophil-attracting chemokine IL-8, while S. Typhi does not elicit this response [13]. These observations suggest that invasion of the intestinal mucosa by S. Typhimurium is accompanied by a rapid escalation of host responses leading to acute, purulent inflammation, while S. Typhi elicits little intestinal inflammation during early stages of infection, however the molecular mechanisms underlying these apparent differences are poorly defined.
One pathogen-induced processes that triggers pro-inflammatory immune responses is the transfer of bacterial molecules into the host cell cytosol by secretion systems. The invasion-associated type III secretion system (T3SS-1) expressed by all Salmonella serovars and delivers effector proteins into the cytosol of epithelial cells [14]. A subset of these translocated effector proteins activate Rho-family GTPases [15]–[18], thereby triggering alterations in the host cell cytoskeleton that result in bacterial invasion of epithelial cells [19]. Excessive stimulation of Rho-family GTPases activates the transcription factor nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) and promotes the subsequent release of proinflammatory cytokines and chemokines [15], [20], [21]. In a bovine model of S. Typhimurium-induced gastroenteritis, the rapid induction of intestinal inflammation and diarrhea requires the T3SS-1 apparatus as well as the effector proteins SipA, SopA, SopB, SopD, and SopE2 [22]–[24]. Similarly, in a murine model of Salmonella induced colitis, SipA, SopE and SopE2 can independently induce intestinal inflammation [25] and mutants lacking a functional T3SS-1 are unable to initiate neutrophil recruitment to the intestinal mucosa during early infection [25], [26]. These findings indicate that T3SS-1-mediated effector translocation induces innate immune responses during S. Typhimurium-induced colitis.
Similar to S. Typhimurium, invasion of cultured intestinal epithelial cells by S. Typhi is mediated by the T3SS-1 [27]. Replacement of S. Typhimurium T3SS-1 effector proteins with their S. Typhi orthologues does not attenuate inflammatory responses elicited by S. Typhimurium in the intestinal mucosa of calves [28], demonstrating that S. Typhi T3SS-1 effector proteins can exhibit intrinsic pro-inflammatory properties in vivo. Thus, the molecular basis for the absence of T3SS-1-dependent innate immune responses early during S. Typhi infection remains unclear.
To study the induction of pro-inflammatory signaling pathways upon infection with S. Typhimurium and S. Typhi, we employed a human epithelial cell line permanently transfected with a NF-κB-dependent luciferase reporter (HeLa 57A) [29]. Infection with the S. Typhimurium wild-type strain SL1344 resulted in a significant increase (7-fold; P<0.01) in luciferase activity compared to mock-infected cells (Fig. 1A), while a derivative of S. Typhimurium SL1344 carrying a mutation in the T3SS-1 apparatus gene invA (SW767) did not elicit NF-κB signaling [20], [30]. In contrast to the S. Typhimurium wild type, the S. Typhi wild-type strain Ty2 failed to trigger NF-κB activation (Fig. 1A), suggesting that S. Typhi is a poor activator of T3SS-1-dependent inflammatory processes in human epithelial cells.
The T3SS-1 mediates invasion of non-phagocytic cells. S. Typhi has been reported to differ from S. Typhimurium with regards to invasion of human epithelial cells [31]–[33], thus raising the possibility that the observed differential activation of the NF-κB signaling pathway could be due to varying degrees of invasiveness. To test this hypothesis, HeLa cells were infected with S. Typhimurium and S. Typhi strains and a gentamicin protection assay was performed (Fig. 1B). The S. Typhimurium wild type SL1344 and the S. Typhi wild type Ty2 were recovered in similar numbers, while the respective isogenic invA-deficient mutants displayed significantly reduced invasiveness. T3SS-1 activity has to two functional consequences: manipulation of host signaling pathways and subsequent bacterial uptake. To discern between effects mediated directly by the T3SS-1 or indirectly by increasing the intracellular bacterial load, we next sought to reinstate invasiveness of the S. Typhimurium invA mutant without restoring T3SS-1 function. Expression of the Yersinia pseudotuberculosis invasin, encoded by the plasmid pRI203, raised invasiveness of the S. Typhimurium invA mutant comparable to the wild type strain (Fig. 1B), but failed to restore the ability to induce NF-κB activation in epithelial cells (Fig. 1C) [34]. Taken together, these observations indicate that immune evasion by S. Typhi did not directly correlate with the intracellular bacterial load or invasiveness.
Despite causing disparate disease entities, the genomes of S. Typhimurium and S. Typhi display remarkable similarity. Chromosomal DNA sequences of both serovars are highly syntenic, with mostly minor inversions, deletions and insertions [35], [36]. One DNA region that is present in S. Typhi but absent from S. Typhimurium is the Salmonella pathogenicity island 7 (SPI-7). Situated within SPI-7 is the viaB locus, an operon encoding regulatory (tviA), biosynthesis (tviBCDE), and export (vexABCDE) genes involved in the production of the virulence (Vi) capsular polysaccharide of S. Typhi [37] (Fig. S1A). The viaB locus has been shown to suppress Toll-like receptor (TLR) signaling pathways [13], [38], [39]. We therefore explored the contribution of the viaB locus on diminishing NF-κB activation in epithelial cells (Fig. 1D and S1B). Deletion of the entire viaB locus in S. Typhi (ΔviaB mutant; SW347) markedly increased the ability to activate NF-κB in epithelial HeLa cells (P<0.001). Akin to the findings with S. Typhimurium, NF-κB signaling induced by the S. Typhi viaB mutant was independent of invasiveness (Fig. S1C) but required a functional T3SS-1 since inactivation of invA in the viaB mutant background (ΔviaB invA mutant, STY4) completely abolished luciferase activity (P<0.001) (Fig. S1D). These results supported the idea that the viaB locus attenuates T3SS-1-induced, pro-inflammatory signaling pathways in human epithelial cells.
The viaB locus has been shown to alter interaction of S. Typhi with host cells through multiple distinct mechanisms (reviewed in [40]). The Vi capsular polysaccharide prevents complement deposition, phagocytosis, and TLR4 activation, while the regulatory protein TviA is known to dampen TLR5 signaling. We therefore wanted to discern whether the absence of NF-κB signaling in human epithelial cells is due to the production of the Vi capsule or due to altered gene expression mediated by TviA. To this end, the tviA gene cloned into a low copy number plasmid (pTVIA1) was introduced into a S. Typhi viaB mutant (STY2). Expression of tviA under control of the native promoter significantly lowered NF-κB activation (P<0.01) in comparison to cells infected with the S. Typhi viaB mutant carrying the empty vector control (pWSK29). Remarkably, expression of tviA reduced inflammatory responses to levels comparable to the S. Typhi wild-type strain (Fig. 1D and S1B), suggesting that the regulatory protein TviA is involved in dampening inflammatory responses in cultured human epithelial cells.
We had recently demonstrated that a S. Typhimurium strain carrying the S. Typhi viaB locus on a plasmid elicits less mucosal inflammation in a bovine ligated ileal loop model than the isogenic S. Typhimurium wild type ATCC14028 [38], raising the possibility that TviA might be involved in suppressing inflammatory responses in vivo. To delineate the relative contribution of the Vi capsule and the regulator TviA to reducing inflammatory responses in the bovine ligated ileal loop model [23], we repeated these studies with derivatives of S. Typhimurium strain ATCC 14028 in which the phoN gene in the chromosome had been replaced with the entire S. Typhi viaB locus (phoN::viaB mutant, TH170) or the tviA gene only (phoN::tviA mutant, SW474). In these strains, transcription of tviA and the downstream genes is solely controlled by the native S. Typhi promoter [41], [42]. This strategy was chosen to ensure that attenuation of intestinal inflammation in this model was not caused by introduction of the viaB locus on a multi-copy plasmid [38].
We compared the phoN::viaB mutant and the phoN::tviA mutant to a strain carrying an antibiotic resistance gene inserted chromosomally in the phoN gene (phoN mutant, AJB715). The phoN::viaB mutant, the phoN::tviA mutant, and the isogenic phoN mutant were recovered in equal numbers from gentamycin-treated tissue samples five hours after inoculation (Fig. 2A), suggesting that neither the tviA gene nor the entire viaB locus interfered with tissue invasion. Consistent with our previous observations [38], the phoN::viaB mutant elicited less fluid accumulation (Fig. 2B) and less pathological changes in the mucosa (Fig. 2C and D) than the isogenic phoN mutant. Remarkably, expression of tviA alone (phoN::tviA mutant) significantly reduced fluid accumulation and inflammation compared to the phoN mutant (P<0.01). The responses elicited by the phoN::tviA mutant and the phoN::viaB mutant were indistinguishable, suggesting that the viaB-mediated attenuation of inflammatory responses five hours after inoculation of bovine ligated ileal loops with S. Typhimurium was mostly attributable to the action of the TviA regulatory protein. Taken together, these data suggested that gene regulation mediated by TviA could dampen inflammatory processes in vivo.
A functional T3SS-1 is required for the induction of intestinal host responses in cattle [22], [24], [43]. A S. Typhimurium strain carrying a mutation in the T3SS-1 apparatus gene invA (invA phoN mutant, SW737) was significantly less invasive than a phoN mutant (Fig. 2A) (P<0.05). Interestingly, inactivation of invA (invA phoN mutant) reduced fluid accumulation (Fig. 2B) and intestinal inflammation (Fig. 2C and D) by a magnitude that was similar to that observed for the phoN::tviA mutant. This finding was consistent with the idea that TviA reduces T3SS-1-dependent host responses in vivo, prompting us to further investigate the mechanism by which TviA inhibits T3SS-1 gene expression.
TviA is a key activator of the tviBCDEvexABCDE operon but can also control transcription of genes outside its own operon (Fig. S2A). Expression of TviA results in diminished motility and flagellin secretion due to downregulation of the flagellar regulon by repressing transcription of the flhDC genes [42], [44]. FlhDC, the master regulator of flagellar gene expression, activates transcription of class II flagellar genes, such as fliA and fliZ [45], [46]. FliA is a positive regulator of class III flagellar genes, including flagellin [45], [47]. To determine whether reduced motility or diminished flagellin production could account for the TviA-dependent reduction in NF-κB activation, we inactivated the fliC gene encoding the sole flagellin of the monophasic serovar Typhi, thereby rendering strains carrying these mutations aflagellate and non-motile. Deletion of the entire viaB operon (ΔviaB ΔfliC mutant, SW483) in the fliC background (ΔfliC mutant, SW359) significantly increased NF-κB signaling in infected HeLa and HEK293 epithelial cells (Fig. S2B and S2C). Expression of TviA from a plasmid (pTVIA1) in a viaB fliC mutant reduced luciferase activity to levels comparable to the fliC mutant (Fig. S2B and S2C), demonstrating that TviA-dependent repression of NF-κB activation was flagellin-independent.
Gene expression profiling experiments suggest that TviA affects transcription of T3SS-1 genes through the following signaling cascade [42]: By repressing transcription of flhDC, TviA downregulates expression of FliZ. The regulatory protein FliZ is an activator of hilA [48]–[50], the master regulator of T3SS-1 genes [51], [52], thus placing T3SS-1 gene expression under negative control of TviA (Fig. S2A). We therefore analyzed the effect of TviA on the transcription of a subset of regulatory, structural, and effector proteins in S. Typhi (Fig. S3). Consistent with previous findings, deletion of the Vi capsule biosynthesis genes alone (ΔtviB-vexE mutant, SW74) did not alter transcription of T3SS-1 genes [42], [44]. In contrast, concomitant deletion of tviA and capsule biosynthesis genes (ΔviaB mutant, SW347) significantly enhanced transcription of the regulatory genes flhD, hilA, and invF, the structural component gene prgH, as well as the effector genes sipA and sopE (Fig. S3).
We next determined which T3SS-1 effector proteins contributed to pro-inflammatory responses elicited by S. Typhimurium and S. Typhi. Previous work has demonstrated that SopE, SopE2, SopB, and SipA contribute to NF-κB activation in epithelial cells [15]–[17], [53]. The bacteriophage-encoded sopE gene is present in S. Typhi Ty2 but absent from S. Typhimurium strain ATCC 14028. To better model the contribution of TviA on attenuating T3SS-1-induced host responses, we chose to continue our studies using the S. Typhimurium strain SL1344, an isolate that carries the sopE gene. Consistent with previous reports [15]–[17], [53], we found that simultaneous inactivation of sopE, sopE2, sopB, and sipA (sopE sopE2 sopB sipA mutant, SW868) reduced the ability of the S. Typhimurium strain SL1344 to induce NF-κB activation to levels observed in an isogenic S. Typhimurium strain unable to translocate effector proteins (invA mutant; SW767) (Fig. S4). A S. Typhimurium strain only expressing SopE (sopE2 sopB sipA mutant, SW867) elicited considerable NF-κB activation. A moderate NF-κB activation was also observed with S. Typhimurium strains only expressing SopB (sopE sopE2 sipA mutant, SW972) or only expressing SipA (sopE sopE2 sopB mutant, SW940) (Fig. S4). Essentially no response was observed in cells infected with a SL1344 derivative that only expressed SopE2 (sopE sopA sopB mutant, SW973). Collectively, these data suggested that SopE was the most potent inducer of pro-inflammatory responses in this tissue culture model, while the contributions of SopB and SipA were more modest.
We next determined the potential contribution of the S. Typhi orthologues of these effectors to the induction of NF-κB signaling in the absence of the tviA gene (S. Typhi ΔviaB mutant, SW347) (Fig. 3). The sopE2 gene is a pseudogene in S. Typhi Ty2 and was not further analyzed. Concomitant inactivation of sopE, sipA and sopB in the S. Typhi viaB mutant (sopB sipA sopE ΔviaB mutant, SW1217) completely abolished NF-κB-driven luciferase activity (Fig. 3). This indicated that, akin to the findings with the S. Typhimurium strain SL1344, SopE, SipA, and SopB are critical for the induction of inflammatory responses in epithelial cells upon infection with S. Typhi. A S. Typhi viaB sopB sipA mutant (SW1211) elicited pronounced NF-κB activation, but a more modest NF-κB activation was also observed with the S. Typhi viaB sopE sipA mutant (SW1214) and the viaB sopE sopB mutant (SW1216) (Fig. 3). These data suggested that SopE was the most potent inducer of pro-inflammatory responses in S. Typhi strains lacking the tviA gene while SopB and SipA contributed moderately. In contrast, diminished NF-κB activation was observed with S. Typhi tviB-vexE mutant (carrying the tviA gene) and its derivatives (Fig. 3). This intricate comparison between derivatives of the viaB mutant and the tviB-vexE mutant allowed us to preclude any confounding effects expression of the Vi antigen might have on gene regulation: both the viaB mutant and the tviB-vexE mutant are non-encapsulated and only differ in their capability of expressing tviA. In contrast, a simple tviA mutant would exhibit a pleiotropic effect, i.e. it would lack the regulatory TviA protein but at the same time exhibit virtually no production of the Vi antigen [37]. Collectively, these data suggested that TviA-mediated gene regulation reduced T3SS-1 effector-triggered NF-κB activation.
Since SopE triggered the most pronounced host responses in the absence of tviA, we focused our further analysis on this signaling pathway. Mechanistic studies in cultured epithelial cells have revealed that the bacterial guanine nucleotide exchange factor (GEF) SopE activates the Rho-family GTPase Ras-related C3 botulinum toxin substrate 1 (RAC1) [15]. Excessive stimulation of RAC1 by bacterial effectors is sensed by the nucleotide-binding oligomerization domain-containing protein 1 (NOD1) [30]. Activation of NOD1 leads to phosphorylation of the receptor-interacting serine/threonine-protein kinase 2 (RIP2) and activation of NF-κB signaling in epithelial cells [15], [20], [30], [54]. The NOD1/2 signaling pathway in HeLa cells can also be triggered by SipA [34], although this pathway plays a lesser role in the SopE-encoding strain SL1344 (Fig. S4). Taken together, these findings raised the possibility that TviA-mediated downregulation of SopE allows S. Typhi to abate immune recognition by the RAC1-NOD1/2-RIP2 signaling pathway. To test this hypothesis, we abrogated RAC1 and RIP2 signaling by either ectopically expressing a dominant negative form of RAC1 (RAC1-DN) [30], [55] or by treating cells with the RIP2 inhibitor (SB203580) (Fig. 4). Consistent with previous reports, ectopic expression of a GFP-SopE fusion protein alone was sufficient to induce NF-κB activation while no upregulation of this signaling pathway was observed with a GFP-SopE construct lacking GEF activity (GFP-SopE G168A) [30], [56]. Simultaneous expression of the GFP-SopE fusion protein and a RAC1-DN construct abrogated NF-κB signaling (Fig. 4A).
Infection of HeLa cells with the S. Typhi wild type or the T3SS-1-deficient viaB invA mutant did not result in a statistically significant increase in NF-κB activation and abrogation of RAC1 or RIP2 signaling did not further impact signaling (Fig. 4A, B, and C). In marked contrast, infection with the S. Typhi viaB mutant led to a substantial upregulation of NF-κB-driven responses. Abrogation of RAC1 or RIP2 activity significantly blunted the induction of NF-κB responses in cells infected with the viaB mutant. Moreover, NF-κB activation in cells infected with a viaB sopB sipA mutant was inhibited when cells were transfected with a plasmid construct encoding RAC1-DN (Fig. 4C), suggesting that SopE, translocated into host cells in the absence of TviA, could activate NF-κB signaling in a RAC1-dependent manner. Treatment with the RIP2 inhibitor did not impact T3SS-1-mediated invasion of S. Typhi strains towards epithelial cells (Fig. S5), excluding the possibility that the RIP2 inhibitor inadvertently interfered with the function of the T3SS-1 machinery. Collectively, these data supported the idea that TviA restricts activation of the RAC1-NOD1/2-RIP2 signaling pathway in S. Typhi-infected epithelial cells.
In addition to repressing T3SS-1 genes, TviA also suppresses flagella expression (Fig. S2A) [57]. Flagellin is known to induce pro-inflammatory responses by activating TLR5 [58] and the NLRC4- (nucleotide-binding oligomerization domain [NOD]-like receptor [NLR] family caspase-associated recruitment domain [CARD]-containing protein 4-) inflammasome [59], [60]. While our initial experiments in the bovine ligated ileal loop model suggest that TviA could mitigate mucosal inflammation (Fig. 2), it is conceivable that TviA-mediated gene regulation of flagellar biosynthesis could have affected flagellin-dependent innate immune pathways. To better study consequences of the expression of TviA on the RAC1-NOD1/2-RIP2 signaling pathway in an animal model, we therefore generated a phoN::tviA mutant in the S. Typhimurium SL1344 background (SW760). Akin to the findings with S. Typhi, expression of TviA in S. Typhimurium reduced transcription of T3SS-1 genes (Fig. S3) and the phoN::tviA mutant elicited significantly less (P<0.05) NF-κB activation than the phoN control strain (Fig. 5A and S6). We next introduced the tviA gene into SL1344 derivatives that only expressed the most potent inducers of the NF-κB pathway, SipA (sopE sopE2 sopB phoN::tviA mutant; SW809) and SopE (sopE2 sopB sipA phoN::tviA mutant; SW807). Upon infection of HeLa cells (Fig. 5B), strains carrying the phoN::tviA insertion elicited significantly less luciferase activity than the respective phoN mutants (P<0.01), indicating that TviA is able to reduce the NF-κB activation elicited by the S. Typhimurium orthologues of SopE and SipA. Inhibition of RIP2 significantly reduced NF-κB activation levels induced by the wild-type strain or the phoN mutant (Fig. 5C). The modest response induced by the phoN::tviA mutant was further blunted by inhibition of RIP2 signaling (P<0.05) (Fig. 5C), suggesting that TviA-mediated regulation of T3SS-1 is partially able to avoid induction of the NOD1/2-RIP2 pathway in vitro.
To exclude any effects of TviA on flagellin-dependent pathways, we introduced the phoN::tviA mutation into a non-motile S. Typhimurium strain lacking phase 1 and 2 flagellins, FliC and FljB (fliC fljB mutant, SW762) (Fig. 6A). Both the fliC fljB mutant and the fliC fljB phoN mutant (SW793) elicited significant levels of NF-κB activation in cultured epithelial cells (Fig. 6A), while this response was greatly reduced in cells infected with the fliC fljB phoN::tviA mutant (SW764). Inactivation of the essential T3SS-1 gene invA completely abolished the ability to induce NF-κB signaling (Fig. 6A).
To directly assess the ability of TviA to impede inflammatory processes in the intestinal mucosa, we used the Streptomycin pre-treated mouse model [61]. In this model, detection of cytosolic access by the S. Typhimurium T3SS-1 through the NOD1/2 signaling pathway contributes to intestinal inflammation early during infection [30], [34], [62]. Compared to mock infected mice, transcript levels of the pro-inflammatory genes Nos2, encoding inducible nitric oxide synthase (iNOS), and Tnfa, encoding tumor necrosis factor (TNF)-α, were significantly (P<0.05) elevated in the cecal mucosa at 12 hours after infection with a non-flagellated S. Typhimurium phoN fliC fljB mutant (SW793) (Fig. 6B and C). Introduction of the S. Typhi tviA gene into a S. Typhimurium fliC fljB mutant (phoN::tviA fliC fljB mutant, SW764) significantly (P<0.05) reduced pro-inflammatory gene expression (Fig. 6B and 6C), but not bacterial numbers recovered from intestinal contents or Peyer's patches (Fig. S7). Inflammatory responses observed in the cecal mucosa at this early time point were T3SS-1-dependent, because introduction of a mutation in invA abrogated the ability of S. Typhimurium to elicit pro-inflammatory gene expression. Collectively, these data suggested that TviA represses T3SS-1-dependent, early inflammatory responses in vivo through a flagellin-independent mechanism.
S. Typhi invades the intestinal mucosa without triggering the massive neutrophil influx observed during gastroenteritis caused by non-typhoidal serovars. Here we show that one mechanism for attenuating host responses is a TviA-mediated repression of T3SS-1, a virulence factor known to induce potent inflammatory host responses. Effector molecules translocated by the T3SS-1 into the host cell cytosol activate Rho-family GTPases [15]–[17]. The activation of Rho-family GTPases is a pathogen-induced process that is sensed by NOD1 [21], [30], which ultimately results in the activation of pro-inflammatory responses in vitro [15], [20], [54] and in vivo [30], [62]. However, S. Typhi requires a functional T3SS-1 to invade the intestinal epithelium during infection [27]. Our data suggest that S. Typhi might have evolved to invade the intestinal epithelium without inducing a potent antibacterial inflammatory response by regulating T3SS-1 expression in a TviA-dependent manner. Osmoregulation prevents expression of TviA in the intestinal lumen, which renders S. Typhi invasive [42] (Fig. 2A). However, TviA expression is rapidly upregulated upon entry into tissue [63], resulting in repression of T3SS-1 and flagella expression while biosynthesis of the Vi capsule is induced [42], [44]. Here we show that TviA prevented NF-κB activation in epithelial cells by reducing T3SS-1-dependent activation of RAC1. Furthermore, the TviA-mediated reduction of T3SS-1-dependent inflammatory responses elicited at early time points in animal models was independent of flagella and the Vi capsule. These data support the hypothesis that TviA attenuates inflammation because it rapidly turns off T3SS-1 expression upon entry into tissue, thereby concealing a pathogen-induced process from the host.
Bovine ligated ileal loops are suited to model the initial 12 hours of host pathogen interaction, a time period during which inflammatory responses are largely T3SS-1-dependent [22], [23], [64]. Similarly, in the mouse colitis model, inflammatory responses elicited in the cecum at early time points (i.e. during the first 2 days) after infection are largely T3SS-1-dependent [61], [65], [66]. However, mechanisms independent of T3SS-1 are responsible for cecal inflammation observed at later time points (i.e. at days 4 and 5 after infection) in the mouse colitis model [65]. Expression of the S. Typhi Vi capsular polysaccharide in S. Typhimurium leads to an attenuation of these T3SS-1-independent inflammatory responses in the mouse colitis model [41], by reducing complement activation and TLR4 signaling [67], [68]. Thus the viaB locus reduces intestinal inflammation by multiple different mechanisms (Fig. S2A). A TviA-mediated repression of T3SS-1 reduces early inflammatory responses while the Vi capsular polysaccharide attenuates responses generated through T3SS-1-independent mechanisms at later time points. It is tempting to speculate that the result of these immune evasion mechanisms is a reduction in the intestinal inflammatory response that could contribute to differences in disease symptoms caused by typhoidal and non-typhoidal serotypes.
This study was performed 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 on mouse experiments was approved by the Institutional Animal Care and Use Committee of the University of California, Davis (Permit Number: 16179). The protocol on calf experiments was approved by the Institutional Committee at the Universidade Federal de Minas Gerais, Brazil (Permit Number: CETEA 197/2008).
The bacterial strains, including relevant properties, are listed in table 1. Unless noted otherwise, bacteria were aerobically grown at 37°C in Luria-Bertani (LB) broth (10 g/l tryptone, 5 g/l yeast extract, 10 g/l NaCl) or LB agar (15 g/l agar). To induce expression of tviA and Vi capsule biosynthesis genes, an overnight culture in LB broth was diluted 1∶50 in tryptone yeast extract (TYE) broth (10 g/l tryptone, 5 g/l yeast extract) or Dulbecco's modified Eagle's medium (DMEM) as indicated and incubated aerobically at 37°C for 3 h. When appropriate, antibiotics were added to LB broth cultures or LB agar plates at the following concentrations: carbenicillin (0.1 mg/ml), chloramphenicol (0.03 mg/ml), kanamycin (0.05 mg/ml), nalidixic acid (0.05 mg/ml), and tetracycline (0.01 mg/ml).
Standard cloning techniques were performed to generate the plasmids listed in table 2. Cloning vectors and ori(R6K)-based suicide plasmids were routinely maintained in E. coli TOP10 and DH5α λpir, respectively.
An internal fragment of the phoN coding sequence was PCR amplified from the S. Typhimurium IR715 chromosome using the primers listed in table 3, subcloned into pCR2.1 (TOPO TA cloning kit, Life Technologies), and cloned into pEP185.2 utilizing the unique XbaI and SacI restriction sites to give rise to pSW208. To generate pSW233, pSW28 was digested with EcoRI and the DNA fragment comprising the joint upstream- and downstream regions of the tviB and vexE genes, respectively, was cloned into the EcoRI site of pRDH10.
Plasmids were introduced into S17-1 λpir and conjugation performed as described previously [57]. The unmarked S. Typhi ΔtviB-vexE mutant SW904 was constructed by inserting the plasmid pSW233 into the STY2 mutant chromosome, selecting for single crossover events (creating merodiploids) on LB agar plates containing Cm and Kan. Sucrose selection was performed as described previously [69] to select for a second crossover event, thus effectively deleting the tviBCDEvexABCDE genes, yielding SW904. The deletion was confirmed by PCR. To facilitate transduction of the unmarked ΔsopE mutation, pSW245 was introduced in this locus in the SW976 chromosome by conjugation with S17-1 λpir as the donor strain, creating SW977 as an intermediate.
Phage P22 HT int-105 was utilized for generalized phage transduction in S. Typhimurium as described previously [70]. For S. Typhi recipients, a similar protocol was followed except the multiplicity of infection (MOI) was increased to 100.
A phage lysate of SW399 was used to transduce the invA::pSW127 mutation into SW483 and SW74, thus generating the S. Typhi ΔviaB ΔfliC invA mutant (SW398) and the ΔtviB-vexE invA mutant (SW611). SW1207 and SW1208 were created by transducing the sopB::MudJ mutation from SW798 into the ΔviaB mutant (SW347) and the ΔtviB-vexE mutant (SW904), respectively. The S. Typhi ΔviaB ΔsopE (SW1209), ΔtviB-vexE ΔsopE (SW1210), ΔviaB sopB::MudJ ΔsopE (SW1216), and ΔtviB-vexE sopB::MudJ ΔsopE (SW1213) mutants were constructed by transducing the ΔsopE::pSW245 mutation from SW977 into SW347, SW904, SW1207, and SW1208, respectively. Subsequent sucrose selection allowed selecting for mutants that had lost the plasmid by allelic exchange and generated a clean ΔsopE mutation, thus creating SW1209, SW1211, SW1216, and SW1213, respectively. Similarly, a P22 lysate of SW839 was used to transduce the ΔsipA::pSW244 mutation (SW839) into SW1207, SW1208, SW1209, and SW1210. The intermediates were subjected to sucrose selection, thus creating the clean ΔsipA mutation of strains SW1211, SW1212, SW1214, and SW1215, respectively. The ΔviaB sopB::MudJ ΔsipA ΔsopE mutant (SW1217) was generated through transduction of the ΔsopE::pSW245 mutation from SW977 into the SW1211 chromosome and sucrose selection.
The S. Typhimurium SL1344 derivatives SW759 and SW760 were established by transducing the phoN::CmR and phoN::tviA-CmR mutations from SW284 and SW474 into the SL1344 wild type. Transduction of the ΔfliC::pSPN29 from SPN305 into the SL1344 wild type and subsequent sucrose selection gave rise to the SL1344 fliC deletion mutant SW761. Subsequent introduction of the fljB5001::MudJ into this strain led to the SL1344 ΔfliC fljB5001::MudJ mutant (SW762). To construct SW764 and SW793, the phoN::tviA-CmR (SW474) and phoN::pSW208 (SW751) mutations were transduced separately into SW762. Invasion-deficient derivatives of these strains were generated by transducing the invA::TetR mutation from SW562 into SW764, SW793, and SL1344, thus creating strains SW766, SW794, and SW767, respectively. SW806, SW807, SW808, and SW809 were generated by transducing the phoN::Tn10dCm (CS019) or phoN::tviA-CmR (SW474) into SW867 or SW940. The ΔsipA::pSW244 mutation (SW974) was moved into the SL1344 wild type to create SW839. SW940 was established by transduction of the sopB::MudJ mutation (SW798) into SW976 and subsequent introduction of the sopE2::pSB1039 mutation (SW800). A P22 phage lysate of SW800 was used to create SW972 using SW1009 as the recipient strain. The phoN::KanR mutation from AJB715 was transduced into SW562 to give rise to the phoN::KanR invA::TetR mutant (SW737).
HeLa 57A cells [29], [34] were generously provided by R. T. Hay (the Wellcome Trust Centre for Gene Regulation and Expression, College of Life Sciences, University of Dundee, United Kingdom). HEK-293 cells were obtained from ATCC (ATCC CRL-1573). Both cells lines were routinely cultured at 37°C in a 5% CO2 atmosphere in DMEM containing 10% fetal bovine serum (FBS) (Life Technologies). For NF-κB activation and invasion experiments, cells were seeded in 24-well plates and 48-well plates (Corning) at densities of 1×105 cells/well and 2×105 cells/well, respectively, and incubated for 24 h prior to subsequent experiments.
S. Typhi and S. Typhimurium strains were pre-cultured in TYE broth as described above. HeLa 57A cells or HEK-293 cells transfected with a NF-κB -luciferase reporter construct were infected with the indicated strains at a final concentration of approximately 106 colony forming units (CFU)/ml. To synchronize the infection, plates were centrifuged for 5 min at 500 g at room temperature. After 3 h, cells were washed with DPBS and incubated at 37°C for an additional 2 h in the presence of DMEM containing 10% FBS. Cells were washed in DPBS, lysed in 0.1 ml of reporter lysis buffer (Promega), and firefly luciferase activity was measured using the luciferase assay system (Promega) in a FilterMax3 microplate reader (Molecular Devices). Results are expressed as percentage of maximum signal elicited in each individual assay. In some experiments, cells were treated 30 min prior to infection until the end of the experiment with either DMSO (vehicle control) or the RIP2-inhibitor SB203580 at a final concentration of 10 µM dissolved DMSO. The NOD1 agonist C12-iE-DAP (Invivogen) was added a final concentration of 100 ng/ml.
For transfection assays [34], HeLa 57A cells were grown to a confluency of about 60% and transiently transfected with a total of 250 ng of plasmid DNA, consisting of 50 ng of the β-galactosidase-encoding vector pTK-LacZ, and either 200 ng of pCMV-myc (control vector) or 100 ng pRAC1-DN and 100 ng of control vector. For co-transfection with pGFP-SopE constructs, 50 ng of pTK-LacZ, 10 ng of the pGFP-SopE plasmid, 90 ng of pEGFP (empty vector), and 100 ng of either pCMV-myc or pRAC1-DN was added. HEK-293 cells were transfected with 25 ng of pTK-LacZ and 25 ng of pNFkB-luc. 48 h after transfection, cells were infected with the indicated Salmonella strains or mock-treated (LB broth) as described above. Efficiency of transfection was normalized by adjusting luciferase values to β-galactosidase values.
Invasiveness of the indicated Salmonella strains was determined using a Gentamicin protection assay as described previously [71]. Briefly, HeLa 57A cells were infected at a MOI of 5 with Salmonella strains pre-cultured in TYE broth. After 1 h, cells were washed and media containing 0.1 mg/ml Gentamicin was added for 90 min. Diluted cell lysates (0.5% Triton-X-100) were spread on LB agar plates to determine the number of CFU per well. Invasiveness was calculated as percentage of recovered bacteria compared to the inoculum.
Overnight cultures of the indicated S. Typhi and S. Typhimurium strains were diluted 1∶50 in TYE broth and incubated at 37°C for 3 h. Total RNA was extracted from approximately 2×109 CFU using the Aurum Total RNA Mini Kit (Biorad). 1 µg of total RNA was subjected to an additional DNase treatment (DNA-free kit, Life Technologies) and converted to cDNA using MuLV reverse transcriptase (Life Technologies) in a 25 µl volume as described previously [71]. 4 µl of this cDNA was used as the template for real time PCR analysis with the primers listed in table 3. Data was acquired on a ViiA 7 real-time PCR instrument (Life Technologies). Relative target gene expression was normalized to mRNA levels of the house keeping gene gmk, encoding guanylate kinase (ΔΔCt method). DNA contamination was less than 1% for all amplicons as determined by a separate RT-PCR mock reaction lacking reverse transcriptase.
Salmonella Typhimurium was cultured in LB broth at 37°C under agitation, followed by subculture in fresh LB (without antibiotics) for 3 hours, at 37°C under agitation. Four 3–4 week-old male healthy Salmonella-free Holstein calves were used in this study. Ligated ileal loops were surgically prepared as previously described [23]. Ligated loops were mock treated with intraluminal injection of sterile LB broth or inoculated with 3 ml of suspensions containing 1×108 CFU of the S. Typhimurium ATCC14028 phoN mutant (AJB715), a phoN::viaB mutant (TH170), a phoN::tviA mutant (SW474), or a phoN invA mutant (SW737). Ligated loops were surgically removed at 5 h after infection for tissue sampling and measurement of intraluminal fluid accumulation. Samples containing the intestinal mucosa and the associated lymphoid tissue were collected with a 6 mm biopsy punch. Each intestinal biopsy was kept in sterile PBS with 50 µg/ml of gentamicin for 1 h, homogenized in 2 ml of PBS, serially diluted, and plated on LB agar plates containing nalidixic acid. Additional biopsies were fixed by immersion in 10% buffered formalin, processed for paraffin embedding, cut and stained with hematoxylin and eosin. Histopathologic changes including hemorrhage, neutrophilic infiltration, edema, and necrosis and/or apoptosis were scored from 0 to 3 (0 for absence of lesions, and 1, 2, or 3 for mild, moderate, or severe lesion, respectively) for a combined total score ranging from 0 to 12.
Animals were obtained from The Jackson Laboratory (Bar Harbor), housed under specific-pathogen-free conditions and provided with water and food ad libitum. Groups of female, 9–12 week old C57BL/6 mice were orally treated with 20 mg Streptomycin. After 24 h, these mice were inoculated as described previously [61] with either 0.1 ml LB broth (mock treatment) or 1×109 CFU of the S. Typhimurium SL1344 fliC fljB phoN mutant (SW793), the fliC fljB phoN::tviA mutant (SW764), the fliC fljB phoN invA mutant (SW794), or the fliC fljB phoN::tviA invA mutant (SW766) suspended in 0.1 ml LB broth. 12 h after infection, animals were euthanized and tissues were collected. The bacterial load was determined by spreading serial 10-fold dilutions of homogenates on LB agar plates containing the appropriate antibiotics. Flash-frozen cecal tissue was homogenized in a Mini-beadbeater (Biospec Products) and RNA was extracted by the TRI reagent method (Molecular Research Center). cDNA was generated using MuLV reverse transcriptase and reverse transcription reagents (Life Technologies). SYBR Green (Life Technologies)-based real-time PCR was performed as described previously [72] using the primers listed in table 3. Data was acquired by a ViiA 7 real-time PCR system (Life Technologies) and analyzed using the comparative Ct method (ΔΔCt method). Murine target gene transcription within each sample was normalized to the respective levels of Gapdh mRNA.
Data obtained from tissue culture experiments, bacterial gene transcription experiments, and the bovine ligated ileal loop model was log-transformed prior to analysis with a paired Student's t-test. To determine statistical significance for relative mucosal mRNA transcription and tissue bacterial load between treatment groups, an unpaired Student's t-test was employed.
|
10.1371/journal.pntd.0004474 | Rabies: Knowledge and Practices Regarding Rabies in Rural Communities of the Brazilian Amazon Basin | The occurrence of outbreaks of human rabies transmitted by Desmodus rotundus in Brazil in 2004 and 2005 reinforced the need for further research into this zoonosis. Studies of knowledge and practices related to the disease will help to define strategies for the avoidance of new cases, through the identification of gaps that may affect the preventive practices.
A semi-structured questionnaire was applied to 681 residents of twelve communities of northeastern Pará state involved in the 2004 and 2005 outbreaks mentioned above. The objective was to evaluate the local knowledge and practices related to the disease. We found a highly significant difference (p<0.0001) in the knowledge of rabies among education levels, indicating that education is a primary determinant of knowledge on this disease. More than half of the respondents (63%) recognized the seriousness of the zoonosis, and 50% were aware of the importance of bats for its transmission, although few individuals (11%) were familiar with the symptoms, and only 40% knew methods of prevention. Even so, 70% of pet owners maintained their animals vaccinated, and 52% of the respondents bitten by bats had received post-exposure vaccination. Most of the respondents (57%) reported being familiarized with rabies through informal discussions, and only a few (23%) mentioned public health agents as the source of their information.
We identified many gaps in the knowledge and practices of the respondents regarding rabies. This may be the result of the reduced participation of public health agents in the transfer of details about the disease. The lack of knowledge may be a direct determinant in the occurrence of new outbreaks. Given these findings, there is a clear need for specific educational initiatives involving the local population and the public health entities, with the primary aim of contributing to the prevention of rabies.
| In 2004 and 2005, the occurrence of outbreaks of human rabies caused by the hematophagous bat, Desmodus rotundus in the Brazilian Amazon, highlighted the role of this bat in the transmission of the disease and the importance of further research on this zoonosis in this region. In the present study, we investigated the local knowledge and practices related to rabies, in some areas affected by the outbreaks, with the aim of identifying gaps, which may affect the preventive practices. Our results show that education influences the level of knowledge, and many residents are aware of the seriousness of the disease and the role of bats in its transmission, although less than half of the respondents knew how to prevent transmission. We also discovered that public health agents were not effective in the transfer of information on rabies, which may be an important determinant of the low levels of knowledge about it. These findings indicated a clear need to increase public consciousness with regard to the potential risk of rabies and the means of avoiding the disease, through educational initiatives directed at the local population, which should involve the public health authorities responsible for the control and prevention of the disease.
| Rabies is an acute form of viral encephalomyelitis, which is almost invariably fatal, and affects mammals on all continents except Antarctica [1]. Transmission occurs through the inoculation of the virus, typically through bites, scratches or contact between skin lesions and the saliva of an infected animal [2–4]. The etiological agent is a member of the order Mononegavirales, family Rhabdoviridae, and the genus Lyssavirus [5]. This genus has a number of different variants that may be hosted by one or more species, acting as regional reservoirs. The classic rabies virus (RABV) is considered to be the most important form of the genus, and it is responsible for more than 55 thousand cases of human rabies worldwide every year, mostly in Asia and Africa [6,7].
In Latin America, dogs have always been considered the principal reservoirs of RABV, although vaccination campaigns for domestic animals have resulted in a 90% reduction in the number of cases of rabies transmission by these animals since the 1980s [8]. In 2004, however, the participation of the hematophagous bat, Desmodus rotundus (E. Geoffroy 1810) in the transmission of rabies on this continent began to attract increasing attention [9]. This shift in the epidemiological profile of the disease was especially relevant in the Brazilian Amazon basin, due to outbreaks of human rabies caused by this bat species in 2004 and 2005 [10–12], which represented a major public health crisis in the rural zone. In fact, during this period, 38 cases were recorded in the northern Brazilian state of Pará, and 24 in the neighboring state of Maranhão [12]. Together, these two states cover 1,579,891.27 km2 of the Brazilian Legal Amazon region [13], an area smaller only than that of Argentina (2,780,092 km2) in comparison with the other 12 countries that make up South America. The establishment of a wild rabies cycle is probably due to the gradual disequilibrium of the natural dynamic of the relationship between the pathogenic agent and its wild host [14], which is likely to have been a response to the increasingly negative environmental impacts affecting this region of Brazil.
The outbreaks occurred primarily in northeastern Pará [12], although after 2005, there were no new cases in humans, and the number of cases in animals declined considerably. Even so, the recent serological study of Costa et al. [15] reported the presence of rabies-neutralizing antibodies in 24 of the 28 bat species currently known to occur on the coast of Pará, indicating that the virus may still be circulating in the region. These authors also found that the most abundant species, Uroderma bilobatum Peters, 1866; Dermanura cinerea (Gervais, 1856); Carollia perspicillata (Linnaeus, 1758) and Artibeus planirostris Spix, 1823, had a seroprevalence of over 40%. While D. rotundus was relatively rare, 43% (3/7) of the specimens collected were seropositive. The two other hematophagous bat species, Diaemus youngi (Jentik, 1893) and Diphylla ecaudata Spix, 1823, have not yet been recorded in the region [15], although they are not directly involved in the transmission of rabies, given their preference for the blood of birds [16]. Currently, D. rotundus has been reported attacking domestic stock in this region, which means that it can still be considered to be a risk zone.
In this context, we evaluated the knowledge and practices related to rabies among the residents of this risk zone, comparing the levels of knowledge in communities where cases of human rabies transmitted by D. rotundus had been recorded on the coast of Pará with those of communities where no cases of human rabies had been recorded. In addition to the community, this investigation considered the age, sex and education level of the residents interviewed, recording their perception with regard to the principal means of transmission of the disease and the practices that help prevent it.
The Institutional Review Board (IRB) at federal Chico Mendes Institute for the Conservation of Biodiversity (ICMBio) authorized this study through license number 39818–1, obtained on June 20 2013. Before administering questionnaires, all the respondents were informed verbally of its aims and objectives, and that their responses would be treated in absolute anonymity. We interviewed only participants who verbally agreed. Oral consent was obtained to ensure anonymity and accommodate illiterate participants, and was documented by the interviewer via voice recording. The ICMBio IRB includes the use of oral consent for the collection of interview data without collecting biological samples from humans in the case of responses that will be kept anonymous, so written consent was not necessary.
The present study focused on protected areas in three municipalities in northeastern Pará, Brazil–(i) the Araí-Peroba Marine Extractivist Reserve in Augusto Corrêa (46°38’06” W, 01°01’18” S), (ii) the Caeté-Taperuçu Marine Extractivist Reserve in Bragança (46°45’56” W, 01°03’13” S), and (iii) the Gurupi-Piriá Marine Extractivist Reserve in Viseu (46°08’15” W, 01°12’15” S). These three municipalities together cover an area of approximately 8098.544 km2, and have a total population of around 210,345 inhabitants [13]. The region is relatively flat, with altitudes of no more than 29 m, and is characterized by a mixture of habitats, with a predominance of Amazon forest, mangroves, and marshlands. The local economy is based on cattle ranching, farming, crabbing, and fisheries.
We applied questionnaires in twelve communities of these municipalities (Fig 1). In six of these communities—Araí, Piçarrera, Cachoeira and Porto do Campo in Augusto Corrêa, and Firmiana and Curupaiti in Viseu—cases of human rabies transmitted by D. rotundus had been recorded during the outbreaks. In the other six communities (Vila Soares and Bacanga in Augusto Corrêa, Benjamin Constant and Treme in Bragança, and Açaiteua and Serra do Piriá in Viseu), no cases of human rabies had been recorded.
In all these communities, the settlements are essentially rural, with a variety of living conditions, including houses made of wattle and daub, timber, and brick, located among forest patches (Fig 2). In some cases, the corral in which the livestock is held is located within a short distance of the landowner’s house, which, together with the proximity of forested areas, contributes to an increased risk of contact with the hematophagous bat, D. rotundus (Fig 2E).
During 2013, we applied the questionnaires (S1 Appendix) in households selected randomly, with residents being selected according to their availability at the moment of the visit. Data were obtained using a paper-based survey. We interviewed a total of 681 residents (approximately 10% of the 12 communities surveyed), of which, 445 were from RR communities, that is, communities in which cases of human rabies transmitted by D. rotundus have been recorded, while the other 236 respondents were from NR communities (no human rabies cases recorded).
To begin with, data were collected on the sex, age, and education level of each respondent. The semi-structured questionnaire was then applied in order to document the perception of the respondents with regard to rabies. Questions were asked on animal ownership, the vaccination of these animals against rabies, possible attacks on these animals and humans by D. rotundus and, when positive, if post-exposure prophylactic vaccination was sought, as well as details on the severity of the disease, its symptoms, transmission, and methods of prevention, as well as the source of this information (S1 Appendix).
The knowledge of the respondents was evaluated using an approach adapted from Kaliyaperumal [17], and classified as Insufficient, Basic, Intermediate or Advanced. The response to each question related to the knowledge of the respondents on rabies was scored 0–3, depending on its completeness and accuracy (S1 Appendix). At the end of the questionnaire, the points were summed. The maximum score is 14, with scores of between 10 and 14 being classified as Advanced knowledge, those between 6 and 9 as Intermediate, 3–5 as Basic, and 0–2 as Insufficient.
Following the application of the questionnaires, the respondents received basic information on different aspects of the rabies zoonosis, such as the means of transmission and prevention.
The Chi-square test (χ2) was used to evaluate possible differences in the knowledge of the respondents on rabies according to their (i) sex (ii) age class, and (iii) education level. The same test was used to evaluate differences in the knowledge of the residents of communities with (RR) and without (NR) recorded cases of human rabies. The level of knowledge on rabies of the respondents was also evaluated in relation to (i) the severity or lethal nature of the disease, and (ii) its prevention, through the vaccination of animals, and the application of the vaccine following attacks by D. rotundus.
A logistic regression (Logit) was used to evaluate the probability that the residents of the RR and NR communities (i) are familiar with different methods of prevention, (ii) vaccinate their animals regularly (in the case of owners of pets or domestic stock), and (iii) seek post-exposure vaccination following attacks by D. rotundus. The odds ratios were calculated with a confidence interval (CI) of 95%. A p = 0.05 significance level was considered for all the statistical analyses, which were run in BioEstat 5.0 [18].
The majority of the respondents were female (n = 436, 64%), adult (n = 555, 81%), and had no more than a primary school education (n = 435, 63%). No significant difference (p = 0.21) was found between communities (RR vs. NR) in relation to the knowledge of the residents with regard to rabies (Table 1).
Perceptions were also investigated in more detail relation to the sex, age class, and education level of these respondents. No significant difference (p = 0.32) was found between male and female respondents, given that approximately 80% of both sexes in both types of community had either basic or insufficient knowledge (Table 2).
Similarly, no significant difference (p = 0.06) was found among age classes, once again, with more than 80% of respondents having basic or insufficient knowledge, although a higher percentage of elderly residents had insufficient knowledge (Table 3).
By contrast, highly significant differences (p<0.0001) were found among the four levels of education (Table 4). Most (around 80%) of the illiterate respondents and those with a primary school education had only basic or insufficient knowledge on rabies, while approximately 70% of those with a high school education had basic or intermediate knowledge. The majority of the 17 respondents with a college education had basic-level knowledge, although a relatively high percentage had advanced knowledge (Table 4).
Most of the respondents were aware that rabies is a grave and potentially lethal zoonosis, in both the RR (n = 276, 62.02%) and NR (n = 156, 66.10%) communities, with no significant difference between localities (p = 0.72). However, when questioned on the symptoms of the disease, the residents of the NR communities had no specific knowledge. The residents of the RR communities were barely more knowledgeable, although 18% were aware of some symptoms. Aggressiveness and intense salivation were the symptoms most frequently cited, for both humans and non-humans (Table 5).
The residents of the two types of community were familiar with the principal rabies transmission routes and thus, the potential vectors. The bites of a number of different animals were mentioned specifically, and bats were mentioned most often in the two areas, being cited by 51% of the residents in the RR communities and 47% in the NR communities (Table 6). While most residents were unfamiliar with measures for the prevention of this zoonosis (52% in the RR communities and 61% in the NR communities), vaccination was the most cited in both types of community (24% in RR and 23% in NR), while washing the bite with soap and water was mentioned by only one resident from an RR community (Table 6).
A majority of the respondents (n = 475, 70%) have pets or domestic stock, with slightly higher ownership being recorded in the NR communities (n = 171, 72%) in comparison with the RR communities (n = 304, 64%). Of the 475 animal owners, 79% (n = 239) in the RR communities, and 74% (n = 127) in the NR communities have their animals vaccinated regularly against rabies, with no significant difference in vaccination rates between the two types of community (p = 0.74). This represents an important preventive measure, considering that 8% of RR residents and 5% of NR residents confirmed that their animals are being attacked by D. rotundus.
Despite these attacks, none of the respondents confirmed being bitten recently (at the time of the survey) by these hematophagous bats. However, 35% (n = 154) of RR residents and 41% (n = 96) of NR residents reported having being bitten at some time in their lives. When attacked by D. rotundus, approximately 60% (n = 87) of the interviewees from RR communities, and 40% (n = 41) from NR communities confirmed having sought post-exposition prophylactic vaccination (PEP), with no significant difference in this response between the two types of community (p = 0.19).
The logistic regression (Logit) provided an estimate of the probability that residents of the two types of community (RR and NR) were familiar with prevention measures for rabies transmitted by D. rotundus (Table 7). This showed that the chance that an RR resident was familiar with a given prevention measure was only slightly higher (odds ratio = 1.09) than that of an NR resident. In fact, the probability that an RR resident was familiar with prevention measures was only 20%, in comparison with 18% in NR residents, with no significant difference being found between the types of community (p = 0.71). A similar tendency was found with regard to animal vaccination rates, with 79% of RR residents vaccinating their animals, in comparison with 74% of NR residents, with no significant difference between types of community (p = 0.27). However, RR residents were significantly more likely (p = 0.02) to seek prophylactic care following attacks by D. rotundus, with 57% against 42% of NR residents (odds ratio = 1.77).
The residents of both types of community identified informal conversations as the principal source of their knowledge on rabies and preventive measure (RR = 60%; NR = 53%). In both cases, less than a quarter of the respondents confirmed receiving information on rabies from public health agents. Importantly, 38% of RR residents and 34% of NR residents had no specific knowledge on this zoonosis (Table 8).
The present study is the first of its kind to be conducted in the Brazilian Amazon basin, with the aim of evaluating the perceptions and practices of the residents of areas of risk for the transmission of human rabies transmitted by the hematophagous bat Desmodus rotundus. The basic knowledge gaps identified among the residents of the study communities were significant and have far-reaching implications for the prevention of this zoonosis, and may contribute to an increase in the risk of new cases or outbreaks.
No significant difference was found in the perceptions of the residents of the two types of community (RR and NR), that is, in which cases of human rabies transmitted by D. rotundus had or had not been recorded, respectively. This similarity between communities is almost certainly a result of the fact that they are separated by relatively short distances, of no more than 50 km, and in some cases, only 5 km (Fig 1). In addition, residents of NR communities obtained information on rabies through informal conversations with neighbors from RR communities, given the existence of family and economic ties in many cases. In this case, it is important to note that the proximity of the study communities may represent a methodological limitation of the present study, and it is possible that the perceptions of the residents of NR communities located further away from RR communities may be far less similar.
The knowledge of the respondents was analyzed according to sex, age class, and education, and our study showed that males and females had similar levels of knowledge, in contrast with the results of studies in Bhutan [19] and Ethiopia [20–22], where the male residents were more knowledgeable than females. In the region of the present study, there are major cultural differences, with females playing a more active role in daily economic activities, in comparison with the regions studied in Asia (Bhutan) and Africa (Ethiopia), where these activities are dominated by males, with a clear influence on the distribution of knowledge [20,21].
The lack of any clear difference among age classes found in the present study was similar to the situation found in these previous studies [20,21], indicating that the age of the individual does not have a significant influence on their knowledge of rabies. Despite this, the present study did identify a tendency for the elderly informants to have more insufficient knowledge. This may be related to the reduced educational opportunities available to this generation, given the limited educational infrastructure of the study region.
In fact, education appeared to be the principal factor determining levels of knowledge on rabies, as shown in the previous studies in Asia and Africa [19–22]. These previous studies also found that knowledge of rabies was directly related to education levels. One possible explanation for this is that individuals with a better education have more access to information, resulting in a better understanding of the features of this zoonosis [21].
It is important to note that one of the limitations of the present study is the differences in the numbers of respondents in the different categories, i.e., sex, age, and education levels (see Table 1). This is related to the fact that the participants of the study were selected according to their availability at the moment of the visit. Given this, it was not always possible to obtain an optimal number of respondents from each category. However, it seems likely that the data set was consistent with reality of the study area.
With regard to the perception of the residents of the two types of community with regard to the seriousness and potential lethality of rabies, more than 60% of the respondents were aware of this aspect of the disease, as in the study of Moran et al. [23]. Despite this, in our study, few of the respondents were able to describe specific symptoms. In fact, only RR resident were familiar with specific symptoms, primarily because they would have had the opportunity to observe the symptoms in relatives o neighbors. Symptoms such as aggressiveness and intense salivation were reported most frequently, in both humans and animals (pets or domestic stock), and in fact, these symptoms are typical of the neurological phase of the disease, while the other symptoms mentioned are observed during either this phase or the prodromal phase [4].
With regard to the transmission of this zoonosis, most respondents referred to bat bites. This is almost certainly due to the fact that these individuals live close to locations at which cases of human rabies transmitted by D. rotundus had been recorded. It is important to note that this species of bat is considered to be the principal reservoir of RABV in Latin America [24–26]. Given this, understanding the potential risks of direct contact with these animals can be considered to be an essential preventive measure for this zoonosis, especially in high risk areas or where this bat is known to attack humans and animals.
While most of the respondents identified bat bites as an important source of the transmission of rabies, around 36% are unaware of the causes of the disease, and perhaps more importantly, more than half were not familiar with any specific prevention measures. Only one of the respondents referred to washing the bite with soap and water as a preventive measure. This measure was also mentioned by few of the residents (8% of the respondents) of high-risk areas in a study in Guatemala, Central America [23]. In fact, this is the primary treatment recommended following an attack by a potential rabies vector, which may reduce by one fifth the risk of developing the disease [27], which reinforces the importance of immediate treatment of the site of the bite, as a preventive measure.
Other post-exposure prophylactic (PEP) measures include (i) disinfection of the wound with alcohol or iodine, in order to inactivate the viral envelope; (ii) application of rabies vaccine on days 0, 3, 7, 14 and 28; and (iii) infiltration of anti-rabies serum with the aim of blocking the proliferation and progression of the virus at the site where it was inoculated [4]. Individuals exposed to a potential risk of rabies should obtain pre-exposure (PrPEP) prevention, which involves the application of three doses of the vaccine at days 0, 7 and 8 [4]. Since the 1980s, the World Health Organization (WHO) has recommended that countries substitute the vaccines produced in animal nerve tissue by those produced in cell cultures. In practice, there are currently only two options of rabies vaccine produced from cell culture—the PCECV (Purified Chick-Embryo Cell Vaccine) and the PVRV or PVCV (Purified Vero Cell Rabies Vaccine), a lineage established from the kidney cells of the green monkey, Cercopithecus aethiops (Linnaeus, 1758) [4].
There have been no recent reports of attacks by hematophagous bats on humans in the study communities. However, approximately 37% of the respondents reported having been attacked at some time during their lives. Some individuals reported that the number of attacks decreased after their community was connected to the national grid of electric power, and that leaving the lights on in the house is an effective measure to keep the bats away, as is keeping the house shut during the night, which stops the animals entering the household. Moran et al. [23] also refer to the sealing of doors and windows as a way of reducing the risk of exposure to the bat, as well as the use of mosquito nets.
In fact, these measures may be effective in reducing the number of attacks and as a consequence, the number of cases of human rabies, given that the outbreaks of rabies transmitted by D. rotundus were recorded in areas with no electricity supply, which were dominated by substandard housing at the time. Schneider et al. [12] and Gilbert et al. [28] concluded that housing conditions are among the principal risk factors for the infection of humans by RABV. They also argued that poor quality housing is typical of many rural areas in Latin America, which have suffered outbreaks of human rabies transmitted by D. rotundus. Finally, the authors concluded that substandard housing may facilitate the access of hematophagous bats to human prey.
Just over half the individuals attacked by bats reports receiving post-exposure vaccination, although an additional limitation of the present study was the lack of proof of vaccination (e.g., vaccination cards) to confirm adequate preventive treatment. Even so, the results of the present study indicate that contact with cases of rabies in humans and animals was important to increase consciousness of the need for post-exposure vaccination, as well as the vaccination of animals. This has also been supported by the intensive animal vaccination campaigns sponsored by the Pará state government, which included the free vaccination of dogs and cats, in some cases, conducted door-to-door by community public health agents. These campaigns have resulted in the elimination of rabies cases in domestic animals in the study area. These vaccination campaigns have contributed to a major reduction—approximately 90%–in the number of cases of rabies in domestic animals and humans in Latin America since the 1980s [8].
Vaccination campaigns have also been established for farm animals, and while not distributed freely, they represent an important government incentive aimed at guaranteeing the vaccination of domestic stocks. Fernandes et al. [29] showed that an increase in the production of beef resulted in an increase in the number of rabies cases in the Brazilian Amazon basin. This emphasizes the importance of maintaining cattle stocks vaccinated, given that beef production tends to be directly proportional to the number of rabies cases. Over the past few decades, the growing number of cases of bovine rabies in many Latin American countries has caused major impacts on both public health and local farming practices [30,31,14,11].
The results of the present study indicate that the majority of the residents of the study area (both types of community) have either Basic or Insufficient knowledge, as in the study Moran et al. [23]. In our study, the respondents were poorly informed with regard to measures that can prevent rabies, and that informal conversations were the primary source of their knowledge. While these conversations may have been based on formal sources of information, the informal transfer of this information among residents may have been subject to distortions and alterations. While community health agents have a primary role in the transfer of information, only 26% of respondents reported receiving information from this source, and even them, only during outbreaks. This emphasizes the need for complementary training with regard to the importance of the transfer of reliable information to local populations.
Overall, then, the implementation of these and other measures designed to guarantee and refine the knowledge of local residents with regard to the potential risks of contracting rabies and means of prevention, may be fundamental to the avoidance of new outbreaks in humans and animals. These objectives may be achieved through the development of educational initiatives, primarily through the relevant public health authorities, and should be directed at both men and women of all ages and education levels. These recommendations are directly relevant to the reality of the Brazilian Amazon basin, although they may provide a practical model for other regions of the world where there is a high risk of lethal outbreaks of human rabies.
|
10.1371/journal.pntd.0000396 | What Will Happen If We Do Nothing To Control Trachoma: Health Expectancies for Blinding Trachoma in Southern Sudan | Uncontrolled trachoma is a leading cause of blindness. Current global trachoma burden summary measures are presented as disability adjusted life years but have limitations due to inconsistent methods and inadequate population-based data on trachomatous low vision and blindness. We aimed to describe more completely the burden of blinding trachoma in Southern Sudan using health expectancies.
Age and gender specific trachomatous trichiasis (TT) prevalence was estimated from 11 districts in Southern Sudan. The distribution of visual acuity (VA) in persons with TT was recorded in one district. Sudan life tables, TT prevalence, and VA were used to calculate Trichiasis Free Life Expectancy (TTFLE) and Trichiasis Life Expectancy (TTLE) using the Sullivan method. TTLE was broken down by VA to derive TTLE with normal vision, TTLE with low vision, and TTLE with blindness. Total life expectancy at birth in 2001 was 54.2 years for males and 58.1 for females. From our Sullivan models, trichiasis life expectancy at the age of 5 years was estimated to be 7.0 (95% confidence interval [CI] = 6.2–7.8) years (12% [95% CI = 11–14] of remaining life) for males and 10.9 (95% CI = 9.9–11.9 ) years (18% [95% CI = 16–20] of remaining life) for females. Trichiasis life expectancy with low vision or blindness was 5.1 (95% CI = 3.9–6.4) years (9% [95% CI = 7–11] of remaining life) and 7.6 (95% CI = 6.0–9.1) years (12% [95% CI = 10–15] of remaining life) for males and females, respectively. Women were predicted to live longer and spend a greater proportion of their lives with disabling trichiasis, low vision, and blindness compared to men.
The study shows the future burden associated with doing nothing to control trachoma in Southern Sudan, that is, a substantial proportion of remaining life expectancy spent with trichiasis and low vision or blindness for both men and women, with a disproportionate burden falling on women.
| Summary measures of population health attempt to express disease burden in terms of a common “currency” and are useful in establishing public health priorities. Disability adjusted life years (DALYs), a health gap measure, have previously been used to estimate burden due to trachoma; however, their methods and results have limitations. This study demonstrates the application of the health expectancies to estimate burden due to trachoma. The study illustrates the future burden associated with doing nothing to control trachoma in Southern Sudan: a substantial proportion of remaining life expectancy spent with trichiasis and low vision or blindness for both men and women, with a disproportionate burden falling on women. The results presented are intuitively meaningful for policy makers and a non-technical audience and compare favourably with other indicators such as mortality and incidence rates or DALYs, which are not generally easily understood. Unless action is taken by further delivery of trachoma control interventions, then populations in Southern Sudan can expect to spend a substantial proportion of their life with low vision or blindness due to trachoma.
| Trachoma is one of the oldest infectious diseases known to mankind and is the leading infectious cause of blindness, estimated to be responsible for 2.9% of blindness worldwide [1]. Recurrent infection with ocular Chlamydia trachomatis results in chronic inflammation, scarring, trichiasis, corneal opacification, and blindness [2]–[4]. Blindness due to trachoma is preventable through the World Health Organization (WHO) SAFE strategy which comprises: Surgery, eyelid surgery to correct in-turned eyelashes which stops pain and minimizes risk of corneal damage; Antibiotic treatment for active trachoma using single-dose oral azithromycin or topical tetracycline; Facial cleanliness, promotion of clean faces especially in children through sustained behaviour change; and Environmental improvements to increase access to water and sanitation [5].
Summary measures of population health, including disability adjusted life years (DALYs) [6] and handicap adjusted life years (HALYs) [7], have been used to estimate the global burden attributable to trachoma. DALYs and HALYs are population health measures permitting morbidity and mortality to be simultaneously described within a single number and estimate the gap between a population's health and some defined goal. The methodology and data sources describing trachoma DALYs and HALYs have differed such that direct comparisons are not possible. For instance, Evans and Ranson estimated global burden of trachoma for the year 1990 to be 80.0 million HALYs [7]; while for the same years the Global Burden of Disease (GBD 1990) project reported trachoma burden to be 1.0 million DALYs [8]. Additionally, studies describing the global burden of trachoma for the year 2000 yielded different estimates of 2.2 million DALYs [6] and 3.6 million DALYs [9]. These previous estimates also have limitations arising from paucity of population-based data on trachomatous low vision and blindness [7],[9],[10].
We aimed to demonstrate the application of the health expectancies approach for trachomatous trichiasis health states (any trichiasis, trichiasis with normal vision, trichiasis with low vision, and trichiasis with blindness) as a summary measure of trachoma burden using population-based survey data from Southern Sudan. Health expectancy is a measure that combines information on both mortality and morbidity to derive lengths of time spent in different states of health. The methods presented can be applied to other trachoma endemic areas and presents estimates of the potential burden of blinding trachoma if control measures are not implemented.
The Institutional Review Board of Emory University approved the survey protocols and clearance to conduct surveys was obtained from the Sudan Peoples Liberation Movement Secretariat of Health (SPLM/Health). Verbal informed consent to participate was sought from the heads of the household, from each individual and the parents of children aged less than 10 years in accordance with the declaration of Helsinki. Consent for household interviews and eye examination was documented by interviewers and examiners on the data collection forms. Personal identifiers were removed from the data set before analyses were undertaken.
Surveys for trachoma were conducted in eleven districts in Southern Sudan between 2001 and 2006 [11]–[13]. For each district, the sample size was calculated to allow for estimation of at least 50% prevalence of active trachoma signs in children aged 1–9 years within a precision of 10% given a 95% confidence limit and a design effect of 5. We also aimed to estimate at least 2.5% prevalence of trachoma trichiasis (TT) in persons aged 15 years and above within a precision of 1.5% at 95% confidence limit and a design effect of 2. The districts were selected on the basis of pragmatic program implementation criteria of: 1) anecdotal reports of blinding trachoma; 2) security and accessibility; and 3) feasibility of initiating trachoma control interventions after the survey.
A two-stage cluster random sampling with probability proportional to size was used to select the sample population in each district. A cluster was defined as the population within a single village. Using a line listing of all the villages in each survey district, villages were grouped into sub-districts. Villages that were inaccessible and/or insecure were excluded from the sampling frame. In the first stage, villages were randomly selected with probability proportional to the estimated population of the sub-district. In the second stage, households were selected from the villages selected in the previous stage using the random-walk method [14], except in Ayod district where the compact segment method [15] was used for sampling households. All residents of selected households were enumerated and those present were eligible for eye examination. It was not possible to return later to the households to pick up any absentees and households where residents were not available were skipped.
Trainee examiners comprising of auxiliary nurses and community health workers were trained using the WHO simplified grading system [16] by a senior examiner experienced in trachoma grading (ophthalmologist or ophthalmic nurse). The minimum accepted inter-observer agreement was set at 80% and reliability assessed in two stages. In the first stage, trainee examiners identified trachoma grades using the WHO sets of trachoma slides [17],[18]. Those examiners who achieved at least 80% agreement then proceeded to the second stage of field evaluation. During field evaluation a reliability study comprising 50 persons of varying age and gender were selected by the ophthalmic nurse to represent all trachoma grades. Each trainee examiner evaluated all 50 subjects independently and recorded their findings on a pre-printed form. Inter-observer agreement was then calculated for each trainee using the senior examiners' observation as the ‘gold standard’. Only trainees achieving at least 80% inter-observer agreement after the field evaluation were included as trachoma graders.
All persons living within each selected household who gave verbal consent were examined using a torch and a ×2.5 magnifying binocular loupe in accordance to the simplified grading system. Alcohol-soaked cotton-swabs were used to clean the examiner's fingers between examinations. All examined participants were assigned a dichotomous outcome for each trachoma sign based on the worst affected eye. TT was defined by the presence of at least one eye lash touching the eyeball or evidence of epilation of the eyelashes. Individuals with signs of active trachoma were offered treatment with 1% tetracycline eye ointment. Patients TT were referred to the health centre where free eyelid surgery was available.
In one district (Mankien), visual acuity (VA) testing was conducted in all eligible participants [19]. Experienced Integrated eye care workers (IECW) were re-trained in VA testing, basic eye examination and trachoma grading and their reliability assessed. Only trainees achieving an inter-observer agreement of 80% and above were eligible to participate as examiners. Prior to the survey, the minimum age for visual acuity (VA) testing was predetermined to be 5 years. VA testing was conducted outdoors in adequate sunlight using the Snellen E chart at 6 meters. In persons with VA<6/60, VA was evaluated with the Snellen chart at 3 meters. Further VA assessment was done in persons with VA<3/60 by counting fingers, hand movement and light perception as appropriate. All participants then underwent basic eye examination. Using a torch and a ×2.5 magnifying binocular loupe, each eye was examined first for in-turned lashes (TT), and the cornea was then inspected for corneal opacities (CO), and the lens examined for cataract. Persons with visual impairment were referred to attend an eye surgery-camp conducted after the survey.
Data were recorded on a customized form and the cause of visual impairment determined for all subjects with a presenting VA of <6/18 for each eye separately. The principal disorder responsible for low vision or blindness was determined for the participant by taking into account the main cause for each individual eye. Vision loss was attributed to trachoma in persons presenting with trichiasis and corneal opacity. In the instance where different causes of vision loss had been identified for each eye separately in a given individual, the principal disorder was chosen to be the one that was most readily curable or, if not curable, most easily preventable (i.e. cataract, trachoma, non-trachomatous CO, and other causes in that order). To define the vision status we adopted the WHO categories of visual impairment based on presenting visual acuity (Box 1).
Our model of the distribution of vision status has been described previously [19]. In brief; using VA data for persons presenting with TT from Mankien survey, age specific distributions of vision status were calculated for 5-year age intervals. We then fitted an ordinal logistic regression model to the observed data to explore the age and gender distribution of the three categories of vision status: normal vision; low vision; and blindness. Persons with visual impairment not directly attributable to trichiasis were excluded from the final model. Children aged 0–4 years were assumed to have normal vision. Predicted probabilities were derived to smooth age-specific curves for the three categories of vision status.
Life tables are frequently used in demography, actuarial science and health services. They trace the life expectancy in pre-determined intervals for a hypothetical population size (frequently 100,000 births) based on parameters usually derived from vital registration data. Abridged life tables for Sudan for the year 2001 were obtained from the World Health Organisation Statistical Information System (WHOSIS) for males and females separately [20]. Demographic estimates for Sudan are based on model life tables because vital registration data are poor or not available. The life tables were derived using the Modified Logit model life table system which is extensively used for countries with poor vital registration. The Modified Logit system has been modelled using data from other populations judged to be similar and is indexed on the number of survivors at age five years and the number of survivors at age 60 years [21].
The health states used to describe the burden due to trachoma were defined as follows: Total Life Expectancy, the total lifespan at birth (years); Trichiasis Free Life Expectancy (TTFLE), the expectation of life without any trichiasis; and Trichiasis Life Expectancy (TTLE), the expectation of life with any trichiasis. TTLE was then broken down into three health states: 1) TTLE with normal vision, the expectation of life with any trichiasis and normal vision (presenting VA≥6/18 in better eye); 2) TTLE with low vision, the expectation of life with any trichiasis and low vision (presenting VA<6/18 but ≥3/60 in better eye); and 3) TTLE with blindness, the expectation of life with any trichiasis and blindness (presenting VA<3/60 in the better eye).
The data analysis framework is summarised in Figure 1. Microsoft Excel spreadsheets developed by the European Health Expectancy Monitoring Unit were adapted for the calculation of health expectancies [22]. Age and gender specific prevalence of trichiasis was estimated from cross-sectional surveys and modelled using logistic regression to smooth the prevalence estimates. The distribution of vision status was derived from Mankien survey whereby visual acuity was categorized into normal vision, low vision and blindness; and modelled by ordinal logistic regression to provide proportions for each category of VA by age and gender [19]. The prevalence of vision status in the sample population was then calculated by multiplying the age and gender specific proportions of vision status with the smoothed prevalence of trichiasis. Life tables were collapsed to represent 5-year age-groups from age zero (0) to 75 years and above. Trichiasis Free Life Expectancy (TTFLE) and Trichiasis Life Expectancy (TTLE) were then calculated using the Sullivan method [23]. The Sullivan method combines use of life tables and age specific prevalence of morbidity to partition life expectancy into years with and without morbidity. The Sullivan health expectancy reflects the current health of the population adjusted for mortality levels and independent of age structure. Health expectancy calculated by the Sullivan method is the number of remaining years, at a particular age, that an individual can expect to live in a specified health state. Trichiasis Life Expectancy was further broken down by vision status to derive TTLE with normal vision, TTLE with low vision and TTLE with blindness.
Table 1 summarises the study population. A total of 23,139 (87.2% of those enumerated) people, in 11 districts, were examined for trachoma of whom males comprised 43%. The overall prevalence of trachomatous trichiasis (all ages) was 6.0% (95% confidence interval [CI] = 5.2–7.0) and varied by district ranging from 0.7% in Katigiri to 10.0% in Kimotong. Of 341 people with TT in Mankien district, 319 were included in modelling of the distribution of vision status (i.e. TT with normal vision, TT with low vision and TT with blindness). The distribution of proportions of vision status by age and gender in persons with trichiasis is shown in Table 2 [19].
Table 3 summarises the age and gender specific prevalence of trichiasis and breakdown of prevalence of trichiasis vision status. The prevalence of TT increased with age and females were more likely to have TT compared to males, age adjusted Odds Ratio (OR) = 1.5 (95% CI = 1.3–1.7). Consistent with prevalence of trichiasis, prevalence of visual impairment (low vision and blindness) increased with age (Table 3).
In the 2001 Sudan life table, women had a higher life expectancy at birth than men. The life expectancy at birth for Sudan was 58.1 years for females and 54.1 years for males. Life expectancy increased in age 5–9 years compared to life expectancy at birth (age 0–4 years) to 56.8 years in males and 60.7 for females; indicating the high under-five mortality rate.
Table 4, Figure 2 and Figure 3 show the life expectancy (LE) and proportions of total life expectancy, trichiasis free life expectancy (TTFLE), trichiasis life expectancy (TTLE), and TTLE with normal vision, TTLE with low vision, and TTLE with blindness. Females had a greater life expectancy at all ages (Figure 2) than males and a larger proportion of remaining life spent with trichiasis low vision or blindness (Figure 3). At age five, TTFLE was 49.8 years (88% of remaining life) and 49.8 years (82% of remaining life) in males and females, respectively. Males expected to live 7.0 (95%CI = 6.2–7.8) years (12% [95% CI = 11–14] of remaining life) with trichiasis at age five; of which 1.9 years (3% of remaining life), 3.5 years (6% of remaining life) and 1.6 years (3% of remaining life) would be lived with normal vision, low vision and blindness, respectively. At age five, females TTLE was 10.9 (95%CI = 9.9–11.9) years (18% [95% CI = 16–20] of remaining life) of which trichiasis with normal vision, low vision and blindness comprised 3.3 years (6%), 4.9 years (8%) and 2.7 years (4%), respectively. For both genders, the proportion of life spent with trichiasis increased with age (Figure 3), by age 50 years, TTLE was 30% (6.4 years) for males and 40% (9.5 years) for females. The proportion of remaining life with trachoma visual impairment (low vision or blindness) at age 50 was 26% (5.5 years) for males and 33% (7.7 years) for females.
This study presents the application of health expectancies in describing the burden due to trachoma by dividing life expectancy into life spent without trichiasis and with trichiasis (trichiasis with normal vision, trichiasis with low vision and trichiasis with blindness). The methods can be applied to other trachoma endemic setting, and presents a technique of estimating the burden associated with uncontrolled trachoma. In Southern Sudan, life expectancy at birth for the year 2001 was 54.2 for men and 58.1 years for women. At age five years, men expected to live an eighth of remaining life with trichiasis and nearly a tenth of remaining life with visual impairment (low vision or blindness) due to trachoma. For women this rose to nearly a fifth of remaining life at age five with trichiasis and an eighth of remaining life with trachomatous visual impairment. Not only were women estimated to live longer, they were also expected to spend a greater amount of time with trichiasis, trichiasis with low vision, and trichiasis with blindness compared to men.
Health expectancy is a measure that combines information on both mortality and morbidity to derive lengths of time spent in different states of health. The Sullivan method for calculating health expectancies has advantages over previous summary measures of trachoma burden because it is presented in units of expected years of life with or without the disease condition. This is intuitively meaningful for policy makers and a non-technical audience, and compares favourably with other indicators such as mortality and incidence rates or disability adjusted life years (DALYs), which are not generally easily understood [24]. Other advantages of the method include simplicity and relative accuracy in addition to using data which are commonly available: prevalence data from surveys and life tables.
In common with other countries for which vital registration is not well established, the WHO life tables are the most authoritative for Southern Sudan. Until recently (April 2008), no population census has been undertaken for many years due to the civil war, and without any vital registration we acknowledge that the true picture for life expectancy in Southern Sudan is difficult to ascertain and may differ from the WHO estimates. Indeed, data collected during the war suggest a total life expectancy for both males and females of just 42 years [25]. However, the Modified Logit model life tables used by WHO were developed in order to address systematic deviations in mortality patterns observed as levels of child and adult mortality deviate from the standard, and this method has been used extensively by WHO to produce life tables for countries with poor vital registration [21].
Our study has a number of potential limitations. The random walk method, whilst acceptable for other purposes is not ideal where the outcome being assessed is one that is obvious to those involved in guiding the survey teams. Bias could have been introduced since the village guides may have been more likely to direct the survey teams to households where they knew there were persons with TT or visual impairment [26]. The Sullivan health expectancies are not appropriate for modelling dynamic changes associated with disease control interventions since it takes a long time for the age specific prevalence of a disability to reach the equilibrium values corresponding to the changes in age specific incidence rates [27]. In addition, cross-sectional data incorporate past recovery, incidence and death rates in the prevalence at particular ages; hence the effects of these rates on health expectancy are more difficult to disentangle [28]. Two other methods are used for calculating health expectancies: multiple-decrement life tables [29] and increment-decrement or multi-state life tables [30]. These methods employ longitudinal data and provide more robust basis for predicting service needs and may be useful in estimating the effects of trachoma control interventions, for instance, eyelid surgery for TT. However, unlike the Sullivan method, these later methods are less used due to lack of appropriate longitudinal data.
Data on distribution of vision status among persons with TT was only available for one district and these were applied to the population prevalence of TT calculated from a large sample from 11 districts, rather than modelling data for each district separately. However, a potential limitation with our VA models is that the effects of co-morbidity of other conditions leading to visual impairment such as refractive error were not controlled for in our model for distribution of VA [19]. Overall, there will be some people for whom this approach is likely to have resulted in an overestimate of trichiasis life expectancy. For others, this approach could have underestimated trichiasis life expectancy, since trachoma may have been causing a proportion of their vision loss, even if not the main cause of vision loss. The effects of co-morbidity thus operate in both directions and the overall bias in estimated potential gain in health expectancy is likely to be very small [31].
Consistent with other studies, our study showed longer life expectancy and trichiasis free life expectancy in females compared to males. In addition, females experienced greater proportions of years lived with trichiasis. Generally, the greater proportion of years lived with disability in females has been suggested to be as a result of longer overall survival or longer survival after the development of disability or disease [32]. Survey data from most trachoma endemic countries have consistently found the prevalence of scarring, trichiasis, and trachoma related blindness to be higher in females compared to males [33]. Therefore, the female excess in low vision or blindness associated with trichiasis is consistent with both greater survival and greater risk of trachomatous blindness among females.
We have presented the burden of trachomatous vision loss by age and gender using health expectancies. These data are of value in advocacy for trachoma control in engagement with politicians and donors. Unless action is taken by further delivery of trachoma control interventions, then populations in Southern Sudan can expect to spend a substantial proportion of their life with low vision or blindness due to trachoma.
|
10.1371/journal.pgen.1004103 | Suicidal Autointegration of Sleeping Beauty and piggyBac Transposons in Eukaryotic Cells | Transposons are discrete segments of DNA that have the distinctive ability to move and replicate within genomes across the tree of life. ‘Cut and paste’ DNA transposition involves excision from a donor locus and reintegration into a new locus in the genome. We studied molecular events following the excision steps of two eukaryotic DNA transposons, Sleeping Beauty (SB) and piggyBac (PB) that are widely used for genome manipulation in vertebrate species. SB originates from fish and PB from insects; thus, by introducing these transposons to human cells we aimed to monitor the process of establishing a transposon-host relationship in a naïve cellular environment. Similarly to retroviruses, neither SB nor PB is capable of self-avoidance because a significant portion of the excised transposons integrated back into its own genome in a suicidal process called autointegration. Barrier-to-autointegration factor (BANF1), a cellular co-factor of certain retroviruses, inhibited transposon autointegration, and was detected in higher-order protein complexes containing the SB transposase. Increasing size sensitized transposition for autointegration, consistent with elevated vulnerability of larger transposons. Both SB and PB were affected similarly by the size of the transposon in three different assays: excision, autointegration and productive transposition. Prior to reintegration, SB is completely separated from the donor molecule and followed an unbiased autointegration pattern, not associated with local hopping. Self-disruptive autointegration occurred at similar frequency for both transposons, while aberrant, pseudo-transposition events were more frequently observed for PB.
| Transposons (“jumping genes”) are ubiquitous, mobile genetic elements that make up significant fraction of genomes, and are best described as molecular parasites. During ‘cut and paste’ transposition, the excised transposon relocates from one genomic location to another. Here we focus on the molecular events following excision of two eukaryotic DNA transposons, Sleeping Beauty and piggyBac. Both transposons are primarily used in a cellular environment that is different from their original hosts, thereby offering a new model to study host-parasite interaction in higher organisms. In the last decade, they have been developed into a technology platform for vertebrate genetics, including gene discovery, transgenesis, gene therapy and stem cell manipulation. Despite the wide range of their application, relatively little is known about their molecular mechanism in vertebrates. We show that these elements are not capable of self-avoidance, as a significant portion of the excised transposons integrates into its own genome in a suicidal process. Despite mechanistic differences, both transposons are affected similarly, and larger transposons are particularly vulnerable. We propose that transposons might recruit phylogenetically conserved cellular factors in a new host that protects against self-disruption. Suboptimal conditions in a new environment could generate abnormal, genotoxic transposition reactions, and should be monitored.
| Mobilization of transposable elements (TEs) is a DNA recombination reaction that can occur either via RNA (retroelement/retrovirus) or DNA intermediates (DNA transposon). In non-replicative, ‘cut and paste’ DNA transposition, the excised transposon relocates from one genomic location to another. In contrast, the ‘copy and paste’ mobilization of a retroelement/retrovirus does not include the excision step, but the downstream events of retroviral integration are highly similar to DNA transposition [1] Many DNA transposons are bracketed by terminal inverted repeats (IRs) that contain binding sites for the recombinase, the transposase. The transposition process is catalysed by the transposase, and can be divided into four steps: (i) the transposase recognizes and binds to the ends of the transposon; (ii) the transposase and two transposon ends form a complex called synaptic or paired end complex; (iii) the transposon is excised from the donor site; and (iv) the excised transposon is transferred to a new location by the transposase reviewed in [2].
TEs are ubiquitous components of both prokaryotic and eukaryotic genomes [3] Even though TEs are best viewed as molecular parasites that propagate themselves using resources of the host cells, their long-term coexistence with their host has provided ample examples of mutual adaptation. The mobility of TEs is regulated by diverse molecular mechanisms, and can be achieved by self-limiting regulatory features intrinsic to the TE itself [4] or mechanisms provided by the host cell. For example, the RNA interference (RNAi) machinery in eukaryotes is probably the best-known cellular mechanism that evolved to control transposition [5], [6]. Notably, generally little is known about the regulation of DNA transposons in eukaryotes. Indeed, our understanding of the mechanisms and the regulation of transposition in eukaryotes are mostly based on assuming analogies to bacterial transposons [2], [7], [8].
In the last decade, the DNA transposition of Sleeping Beauty (SB), a resurrected fish transposon [9] was intensively studied [10]–[13]. Using SB as a model to study host-transposon interaction in eukaryotic cells, a series of evolutionarily conserved (from fish to human) cellular determinants has been identified. HMGB1, a non-histone chromatin factor, is required for synaptic complex formation during SB transposition [11]. Factors of the non-homologous-end-joining (NHEJ) pathway of double strand DNA break (DSB) repair, including Ku70 and the DNA-dependent protein kinase (DNA-PKcs) are required for SB transposition by acting at repairing the transposon excision sites [10]. Through its association with Myc-interacting zinc finger protein 1 (ZBTB17 or Miz1), the SB transposase down-regulates cyclin D1 expression in human cells, resulting in a cell cycle slowdown [12]. A temporary G1 arrest enhances transposition, suggesting that SB transposition is favoured in the G1 phase of the cell cycle, where NHEJ is preferentially active [10]. The HMG-box transcription factor HMGXB4 (HMG2L1), a component of the Wnt-signaling pathway is involved in a feedback regulation of SB transposase expression [13]. These studies indicate that eukaryotic transposons can participate in a complex interactive regulatory platform involving evolutionary conserved cellular mechanisms.
Although, SB is a relatively well-characterised eukaryotic transposon, one part of the transposition reaction, the step following excision but prior reintegration, is yet unexplored. In the process of productive transposition, the excised molecule integrates into a new genomic location. However, in principle, the excised transposon molecule could reinsert, in a self-disruptive process, into its own genome. This suicidal transposition event is called autointegration, self-integration or intramolecular transposition, and is well characterized in prokaryotes [14]–[16]. The best-understood example in bacteria is Tn10 transposition, in which regulation of transposition is a delicate interplay between the transposon and host-encoded factors [17]–[19]. These host factors, namely IHF (integration host factor), HU (heat unstable nucleoid protein) and H-NS (nucleoid structuring protein) are among the most important regulatory factors in E. coli. IHF and HU stimulate the early steps of transposition prior to excision of Tn10 [19]. However, if they remain associated with the transpososome (a DNA-protein complex minimally containing the excised transposon and the transposase), they promote autointegration [7]. By opposing the effects of IHF [20] and HU, H-NS inhibits autointegration and promotes productive transposition [18], [19]. In eukaryotes, autointegration was reported in mariner transposition [21], [22] Curiously, one third of the autointegration events mediated by Mos1 (mariner) were recovered from non-canonical target sites [22]. Self-disruptive autointegration has also been observed during retroviral integration [23]–[25]. A host-encoded protein, barrier-to-autointegration factor (BANF1 or BAF) has been identified by its ability to protect retroviruses from autointegration [23].
Two observations suggest that, similarly to bacterial transposons and retroviruses, autointegration could be a significant factor affecting productive DNA transposition in eukaryotes as well. First, similarly to certain bacterial DNA transposons [26], [27], transposition of SB from a genomic locus frequently occurs into sites that are close to the donor locus [28]; this phenomenon is termed “local hopping”. Obviously, the transposon itself is the closest target to integrate. In Tn10 transposition, the host factor IHF promotes ‘target site channelling’ close to the IR of the transposon [19]. Second, larger transposons are expected to be particularly attractive targets for autointegration. Indeed, it has been observed that, similarly to certain bacterial TEs, longer elements of SB tend to transpose less efficiently [29], [30]. Thus, both ‘local hoping’ and size-sensitivity might be associated with vulnerability of SB transposition to self-integration.
In the present study, we investigated the post-excision fate of two DNA transposons, SB [9] and piggyBac (PB) [31] in vertebrate cells. Although, both SB and PB belong to the superfamily of DDE/D transposases, characterized by a highly conserved catalytic domain [1], they exhibit significant differences in their mechanisms of transposition [32], [33]. For example, the activity of SB is essentially restricted to vertebrates [29], [34], with the exception of a chordate, Ciona intestinalis [35]. By contrast, PB seems to have an extremely wide host range as it can transpose in insects as well as in human cells [36]–[38]. In comparison to PB, SB was reported to exhibit a much stronger ‘local hopping’ phenotype [39], [40]. Furthermore, SB, but not PB was reported to be sensitive to the size of the mobilized element. Specifically, the transposition of PB was reported to be independent on the size of the element below 14 kb [41].
Importantly, both SB and PB are valuable genomic tools for genome manipulation [42], and mostly used in heterologous cellular environments, thereby offering a unique opportunity to investigate various survival strategies of DNA elements in eukaryotes. Indeed, we can model how these elements behave in naïve genomes, and adapt to their new environment. We have used a simple experimental setup, i. e., transfection into cultured cells to monitor the process of establishing a host-parasite relationship in a heterologous environment. This strategy identified BANF1 as a host-encoded factor influencing this process. We propose that deciphering the mechanism and regulation of transposon reactions and translating this knowledge can be effectively used to derive transposon-based genetic tools for genome manipulation or for gene therapy.
To detect and characterise potential autointegration products, the following assay system was established. The test construct, SBrescue, is a plasmid comprising a replication origin (Ori) and an antibiotic resistance cassette for zeocin (Zeo) located between the IRs of the transposon (Figure 1A). Outside of the transposon SBrescue contains the rpsL gene rendering bacteria sensitive to streptomycin [43]. SBrescue and the helper plasmid encoding for the transposase are co-transfected into cells. Plasmid DNA is recovered from the cells two days post-transfection and transformed into E. coli. Bacteria are subjected to double antibiotic selection of zeocin and streptomycin (Figure 1B). Following transposon excision and circularization of the excised transposon, the rpsL is lost, thereby rendering bacteria StrepR (Figure 1B). Autointegrative transposition events can be rescued in the form of either two deletion circles or a single inversion circle, depending on the topology of the strand attack (Figure 1C). The assay can detect autointegration events occurring into regions designated A, B, C and IR (Figure 1A). In addition, integration events into the rpsL gene would render bacteria resistant to streptomycin and recovered by the assay. In contrast, autointegration events into Zeo or Ori would not be detectable with the assay system, because these regions are required for plasmid propagation and maintenance.
To identify conditions affecting autointegration of SB, the following factors were considered: (a) cell type specificity; (b) transposase activity; (c) target site distribution; (d) the size of the transposon; (e) host-transposon interaction. First, SBrescue was introduced into human HeLa cells with or without a helper plasmid expressing the hyperactive SB100X transposase [44] (Figure 1B). Compared to the control (0.03%, 1.19×103/3.88×106), significantly elevated numbers (0.45%, 4×103/9.09×105) of ZeoR/StrepR bacterial colonies were observed when SB100X transposase was present in the experiments (Figure 1D). To characterize potential autointegration events and map the transposon insertion sites, the recovered products were subjected to DNA sequencing. Sequencing data confirmed that similarly to productive transposition, the autointegration events of SB transposition were targeted into TA dinucleotides within the mappable A, B, C and IR regions of the transposon (Figure 1E). To investigate if cellular factors in various vertebrate species might differentially promote or protect against autointegration of SB, the assay was performed in cultured cells of different origin, including AA8 (Chinese hamster, ovarian; 0.85% vs 0.03%, 1.51×103/1.77×105 vs 476/5.52×106), MEF (mouse, embryonic fibroblast; 0.05% vs 0.01%, 8.06×103/1.71×107 vs 332/9.54×105) and PAC2 (zebrafish, fibroblast; 0.13% vs 0.08%, 1.03×103/8.42×105 vs 770/9.69×105) cells (Figure 1D). Our results revealed that the SB-mediated autointegration events were detectable in all tested cell lines, including fish, the natural cellular environment of SB (Figure 1D). Similarly to productive transposition, the frequency of autointegration varied in the different cell types [29]. The highest frequencies of autointegration were detected in HeLa and AA8 cells that generally support efficient transposition [29], suggesting that the frequency of autointegration was primarily dependent on the activity of the transposase, rather than the cell type (Figure 1D). Indeed, compared to the original SB10 transposase [9], autointegration by the hyperactive SB100X transposase [44] was higher by one order of magnitude in human HeLa cells.
Remobilization of the SB transposon from a genomic donor site exhibits a significant bias toward the donor locus (local hopping) [32]. Similarly, the reintegration of Tn10 transposons is not unbiased and targeted to the IRs of the transposon during autointegration, referred as ‘target site channelling’ [19]. In contrast, when launched from an extrachromosomal donor molecule, the genomic distribution of SB insertion sites is fairly random [45]–[47]. Target site selection during transposition of SB from an extrachromosomal plasmid is primarily determined on the level of DNA structure, as insertion sites tend to have a palindromic pattern and a bendable structure [45]. Accordingly, the insertion profile of the SB transposon can be modelled by determining the DNA-deformability scores, called Vstep for each potential TA target site, using the software ProTIS [48]. To determine the autointegration profile of SB, Vstep values were generated for the mappable regions of SBrescue and the observed insertion frequencies were compared to the calculated Vstep values (Figure 2B). Altogether, 53 autointegration products were identified and mapped to the regions of IR, A, B and C. Most of the autointegration events occurred into region B that is farther away from the IRs, and relatively few into regions A and C that are closer to the transposon ends (Figure 2). In regions B and C, there was a correlation between insertion frequencies and Vstep scores (Figure 2B). These results suggest that similarly to transposition from an extrachromosomal donor, insertion site selection during autointegration of SB is largely independent from the donor site and did not exhibit ‘target site channelling’ close to the IRs of the transposon. On the contrary, despite of the predicted high Vstep score, only a single insertion event was recovered from the IRs (Figure 2), suggesting that the transposon ends of SB, embedded in a paired end complex are limited in their abilities to target the IRs or sites close to the IRs during autointegration. Due to the linkage, the autointegration of SB was primarily intramolecular, and no insertions were detected from the rpsL region. Thus, the transposon was fully excised from the flanking donor DNA prior its integration into a new site.
Next, we tested whether self-destructive autointegration could also occur during PB transposition. We have used a transposon donor construct that is identical to SBrescue, except that the SB IRs were replaced by PB IRs [49] (PB2K in Fig. 3B), together with a mouse codon-optimized PB transposase (mPB) [50]. As shown in Figure 3A, autointegration of the PB transposon occurred at frequencies comparable to SB100X (0.49%, 3.2×104/6.4×106) in HeLa cells. As predicted and confirmed by DNA sequencing, autointegration of PB occurred into TTAA motifs, the canonical target site of PB [31] (Supporting Figure S1). Altogether, 23 integration sites were mapped and twelve were recovered from regions B and C (Figures 3B,C). However, unlike with SB, a significant number of integration events (48%, 11/23) mapped outside of the transposon, in the rpsL gene (Figures 3B,C). These non-canonical transposition events also targeted TTAA target sites, but involved only a single end of the transposon. The other IR was not separated from the donor molecule during the reaction. We refer to these non-canonical transposition events as single-ended transposition.
To investigate the phenomenon of single-ended transposition of PB further, a reciprocal construct, PBsingle was generated, where the PB transposon carried an rpsL gene (Figure 3D). In addition to single-ended transposition events detected by PB2K, the PBsingle assay system was suitable to capture various deletion products (Supporting Figure S2). Bacteria that gained StrepR could report on (i) double-ended excision products, (ii) single-ended integration events into either rpsL or (iii) the vector sequence flanking the transposon. The autointegration assay was performed as shown in Figure 1B, except bacteria were exposed to double selection of kanamycin and streptomycin. To capture single-ended events, 336 transposition products were pre-filtered by colony PCR, using primers flanking the PB excision site. Canonical excision products would appear as uniformly sized PCR products, while size difference would report on either single-ended transposition or non-transposase-mediated small deletions/insertion events generated by DNA repair. 31/336 pre-filtered PCR products were analysed further by DNA sequencing, and six out of 31 (19%) products were clearly generated by PB transposase-mediated, single-ended transposition that occurred into TTAA either inside or outside of the transposon (Figures 3D and 3E).
In the ‘single-ended’ transposition of PB, only one of the IRs was mobilized. Still, true single ended events, when the second IR is not involved in any of the steps of transposition, cannot be convincingly demonstrated. In fact, alternative mechanisms can generate similar, hard-to distinguish products. For example, the canonical transposition reaction might fail at the final step, and only one end of the transposon is transferred (lariat model), (Supporting Figure S2). Aberrant transposition might also occur by a mechanism that involves pseudo or cryptic sites mistakenly recognized as IRs. In addition, the ends of the transposon can also be derived from two separate molecules [51] (bimolecular transposition).
To explore the scenario of bimolecular transposition, truncated ‘solo’ transposons were generated. ‘Solo’ substrates, lacking either the left (PBΔleft; SBΔleft) or the right IRs (PBΔright; SBΔright) were tested in a cell culture-based transposition assay [9]. Molecular analysis of the resistant colonies revealed that neither PBΔright nor SBΔleft supported transposition (Table 1). In contrast, the analysis confirmed transposase-mediated transposition of the ‘solo’ substrates, PBΔleft (4.6%) and SBΔright (0.56%) [52] (Table 1), indicating that both transposases are capable of utilizing ‘solo’ substrates. In either cases, the IRs of the ‘solo’ transposons were properly integrated into respective target sites (Supporting Test S1). Notably, in clone PBΔleft#8, we have identified a second right IR integrated into a same genomic locus, confirming that the transposase used the two IRs from separate molecules (Supporting Text SF1). As ‘solo’ transposition occurred ∼8-fold more frequently for PB, we monitored the PB system further in the ‘solo-mixing’ experiments. In this strategy, the PBΔleft and PBΔright constructs were transfected either alone or mixed in equimolar ratios, and tested in the colony forming, transposition assay. If transposition utilizes the IRs from separate molecules, one would expect elevated colony numbers when either PBΔleft or both ‘solo’ substrates are present in the assay, compared to PBΔright that does not support transposition alone (Table 1). The higher number of resistant colonies in the respective experiments indicated that the transposase was able to utilize the IRs from different copies of the transposon, supporting the bimolecular model (Figure 4).
The efficacy of transposition was reported to depend on the size of the transposon [29], [30], [53]–[55]. One potential mechanism responsible for such size-dependence is that following transposon excision, self-disruptive autointegration competes with productive transposition. Since larger transposons have more target sites, they could be particularly attractive targets for autointegration. This hypothesis predicts that the size of the transposon does not affect the frequency of excision, but it shifts the ratio between autointegration and productive transposition. To test this assumption, a series of transposons of different size, ranging from 2679 bp to 7256 bp (SB2K, SB3K, SB4K, SB7K) and 2795 bp to 7319 bp (PB2K, PB3K, PB4K, PB7K) were generated for SB and PB, respectively. Frequencies of transposon excision, autointegration and productive transposition events were determined for the various transposons. Excision frequencies were estimated by quantitative PCR, autointegration was monitored as above. Productive transposition was determined in a cell culture-based assay [9]. Figure 5A shows that excision frequencies declined with increasing size, while autointegration frequencies elevated over 4 kb either moderately or sharply for SB and PB transposons, respectively (Figure 5A). Accordingly, productive transposition frequencies dropped with increasing size of both SB and PB transposons. These results indicated that the size of the transposon affected transposition already at the excision step, thereby arguing against the hypothesis of autointegration being the sole factor that compromises productive transposition with increasing transposon size. Nevertheless, autointegration contributes as an additive element to the less efficient transposition of long transposons. Surprisingly, the two transposons behaved similarly in all three assays (Figures 5A an 5B). Thus, in contrast to general assumptions, and similarly to SB, size affects PB transposition as well.
A cellular protein, BANF1 (BAF) barrier-to-autointegration factor was identified by its ability to protect retroviruses from autointegration [23]. BANF1 binds to double-stranded DNA, including freshly transfected, extrachromosomal plasmid DNA [56], in a non-specific manner [57], [58]. Thus, in principle, BANF1 could affect DNA transposition as well, between the molecular steps of excision and reintegration, when the transposon exists as an extrachromosomal molecule in the cell. To test this assumption, we asked if BANF1 could protect DNA transposons from autointegration. We addressed this question by monitoring autointegration events in HeLa cells, where BANF1 was either knocked-down or transiently overexpressed (Figure 6A). When BANF1 expression was knocked-down by RNA interference (Supporting Figure S3), the frequency of autointegration of SB was increased by two-fold compared to the control (Figure 6A, left panel). In contrast, BANF1 overexpression decreased the frequency of autointegration to one third (Figure 6A, left panel). Similar results were obtained by using the PB transposon (Figure 6A, right panel). No significant effect of BANF1 was observed at the excision step of SB transposition (not shown), suggesting the BANF1 acted specifically following excision.
In addition to BANF1, the effect of another host-encoded factor, the high-mobility group protein (HMGB1) was tested on autointegration. Similarly to BANF1, HMGB1 binds DNA in a non-specific manner [59]. In SB transposition, the transposase physically associates with HMGB1 and recruits it to the transposon DNA [11]. Autointegration was monitored in cells where HMGB1 was either transiently overexpressed or knocked-out [60]. Although, HMGB1 overexpression or deficiency was significantly affecting productive transposition [11], it had no detectable influence on autointegration (Supporting Figure S4). These results indicate that despite their similar non-specific DNA-binding activity, BANF1 and HMGB1 have a clearly distinct effect on DNA transposition.
Alternatively to a non-specific engagement, and similarly to retroviruses, BANF1 might be actively recruited to a preintegration complex of a transposon. In order to distinguish between these two scenarios, a high throughput immunoprecipitation experiment was designed to analyse a protein interactome forming around the SB transposase in mammalian cells. Affinity purification combined with mass spectrometry is a powerful strategy to detect protein-protein interactions among proteins in their native cellular environment [61]. This method is suitable to reveal the composition of entire protein complexes. If we use the analogy to retroviruses [62], one should keep in mind that even if BANF1 is recruited actively to the preintegration complex, it might not be recruited directly by the transposase. To distinguish true interaction partners from non-specific contaminants, we needed an easy-to-detect, confirmed interacting partner of the SB transposase as bait. We can readily monitor interactions of HMGXB4 (HMG2l1) with either the transposon or the transposase in vivo [13]. Thus, HMGXB4 was chosen as bait to analyse higher order complexes formed around SB. The experiments were run in parallel, in the presence and in the absence of the SB transposase. In the control experiment, it is not expected to detect interaction partners of the SB transposase. HEK293T cells were transiently transfected with HA-tagged HMGXB4 protein in the presence/absence of the SB10 transposase [9]. A SILAC pull-down experiment was performed. This experimental strategy identified BANF1 as an interaction partner of HMGXB4− in the presence, but not in the absence of the SB transposase (Figure 6B). The presence of BANF1 was also detectable when the bait, HMGXB4 was used in a co-immunoprecipitation assay (Figure 6C). This observation predicts that BANF1 can be actively recruited into a higher order protein complexes forming around the SB transposase in mammalian cells.
This study focuses on molecular events following the excision steps of two eukaryotic DNA transposons, SB and PB, derived from fish and insect genomes, respectively. The transposition reactions were performed in a heterologous host environment, phylogenetically distant from their natural hosts. The experimental setup mimics the scenario of introducing DNA transposons into a naïve eukaryotic host. We have shown that a significant portion of SB and PB transposon excision events is accompanied by suicidal integration into the transposon's own DNA. Although, different transposons may have different frequency of autointegration depending on the structure of the transpososome and the number of the integration target sites on the transposon, autointegration would influence the success of a transposon in a new environment. Neither SB nor PB was immune to the suicidal process of autointegration. Thus, in general, transposases/integrases in eukaryotes might not be able to distinguish between their own genome form foreign DNA. This would define autointegration as the lack of ability of self-avoidance upon integration. In contrast, certain prokaryotic transposons, including Tn7 and Mu exhibit ‘target immunity’ that prevents the transposon from transposing into its own genome [63], [64]. Both Tn7 and Mu avoid integration into DNA molecules that already have a copy of the transposon. As an alternative to self-encoded ‘target immunity’, some bacterial transposons and eukaryotic retroviruses recruit cellular host factors to protect against autointegration [19], [23]–[25]. In Tn10 transposition a host protein, histone-like nucleoid structuring (H-NS) plays a role in promoting intermolecular and supressing self-destructive intramolecular integration events [19]. Similarly, DNA transposons in eukaryotes might also capture cellular factors to protect their genome against autointegration. This strategy could defend the invading molecule and contribute establishing a stable host-transposon relationship.
BANF1 is involved in several critical processes, including host defence [65], [66]. The usual mode of BANF1 is repressive, due to its propensity to coat DNA. For example, BANF1 acts as a potent inhibitor of virus replication, defending against poxvirus invasion [67]. Intriguingly, and in contrast to its original function in host defence, BANF1 is piggybacked by various retroviruses to protect their viral genome against autointegration. BANF1 inhibits autointegration of the Moloney Murine Leukemia retrovirus, MoMLV [23], [68], [69] or HIV-1 [62]. By physically protecting the retrovirus, BANF1 promotes productive viral integration into the host genome [62]. In our experimental setup, BANF1 was influencing the fate of the excised molecules of two DNA transposons of different origin, SB and PB. Thus, in addition to its reported activity to bind freshly transfected DNA [56] or retroviral cDNA [23], BANF1 might influence the fate of DNA transposons as well. An important ramification of utilizing phylogenetically conserved cellular proteins by transposons might be the ability to survive and establish stable host-parasite relationship in a heterologous host environment. Accordingly, in addition to its role in Tn10 transposition, H-NS was reported to selectively bind the transpososomes of Tn5, and is likely to modulate many other transposition processes in Gram-negative bacteria [70].
SB and PB are members of the superfamily of DDE/D transposases and retroviral integrases, utilizing the same strategy for target joining. Still, how reasonable it is to assume an interaction of BANF1 with both DNA transposons and retroviruses? In fact, BANF1 might be an ideal cellular factor for integrating elements in higher eukaryotes. Due to its non-specific DNA-binding activity to double-stranded DNA [58], a capacity to compact DNA and assemble higher-order nucleoprotein complexes, BANF1 could influence the fate of any extrachromosomal DNA molecule. As in retroviral integration [23], [69], BANF1 may compact the transposon genome to be a less accessible target for autointegration, and promote the integration step. Furthermore, similarly to retroviruses, BANF1 could be even actively recruited to preintegration complexes. The exact manner of recruitment might vary, providing specificity. BANF1 is recruited via physical interaction by the viral matrix protein gag to the retroviral preintegration complex of HIV-1 [62]. In SB transposition, BANF1 was enriched in a higher order complex containing the SB transposase and its interactor HMGXB4. Thus, the enrichment was mediated via protein-protein interaction. Since the experimental setup did not include the transposon DNA, we could not faithfully simulate preintegration complex formation. Nevertheless, HMGXB4 is a specific interaction partner of both the transposon and the transposase of SB [71]. Therefore, it might be reasonable to assume that BANF1 associates with the preintegration complex.
In sum, our strategy to model the process of establishing a host-transposon relationship in a naïve environment identified BANF1 as a host encoded factor influencing this process. Future work will have to clarify if a common role of BANF1 to protect integrating mobile elements in general exists.
Traditional models predict that efficient integration must follow the excision of DNA elements. Strikingly, autointegration was estimated to be over 90% in mariner transposition in vitro, suggesting that under the standard reaction conditions, the vast majority of the excised transposon inserts into itself, rather than into another DNA molecule [21]. This high frequency would establish autointegration as a major factor affecting productive integration. Furthermore, as longer transposons present more potential target sites, autointegration would be a reasonable explanation for size-dependence of transposition, observed for both SB [29], [30] and PB (this work) transposition.
Still, the role of autointegration in counteracting productive transposition might be overestimated. We found that transposon excision, a step prior to integration, is already affected by the size of the transposons (Figure 5A), indicating that a larger transposon might have difficulty to form a synaptic complex. Our data argue that competition between self-integration and productive transposition is unlikely to be the only factor responsible for sensitivity to size. If we assume that unproductive transposition equals suicidal autointegration, the gap between transposon excision and productive transposition could be a good estimate for the effect, and was reported to be around 25% in SB transposition in vivo [32].
In contrast to an earlier report [41], we found that SB and PB transposons were affected similarly by the size of the transposon in three different assays (Figure 5). When the size of the transposon increased from 2683 to 7260 and 2795 to 7319 bps, the frequency of productive transposition dropped by 83% and 89.6% for SB and PB, respectively (Figure 4B). In addition, SB and PB behaved similarly in assays monitoring either excision or autointegration (Figure 5A). Therefore, our data argue against the general assumption that the PB transposon is not sensitive to size below 14 kb [41]. The different observation might be related to the fact that (i) the DNA fragment that Ding et al. used to increase the size of the transposon contained a higher density of TTAA target sites than the existing transposon. Actually, it is impossible to separate the true effects of length and numbers of target sites for a transposon that is highly specific in terms of integrating into a given sequence; (ii) Ding et al. estimated transposition frequencies in transgenic mouse experiments by counting transgenic embryos, regardless of the copy number of the integrated elements per embryo. Therefore, to compare productive transposition of SB and PB transposons, we have adjusted transgenic frequencies by the copy number of the integrated transgenes [72]. Importantly, small size does not seem to be an absolute requirement for mobilization in either case. Decreasing the distance outside the transposon ends of SB was reported to increase transpositional rates under experimental conditions [29]. Moreover, both PB and SB100X were reported to capable of mobilizing giant molecules of DNA, such as BACs (bacterial artificial chromosomes) [37], [73]. These reports indicate that in contrast to viruses, DNA transposons have no strict (if any) upper limit regarding their cargo capacity.
Autointegration of SB, likely due to physical constraints, avoided the IRs, suggesting that the captured events were rather intramolecular than intermolecular. Nevertheless, SB integration is not channelled to the terminal repeats of the transposon as it was observed for Tn10 [19]. Furthermore, the lack of linkage of autointegration sites to nearby regions at the donor DNA molecule would argue against an association between the ‘local hoping’ phenotype and autointegration.
Our experimental approach gave us the opportunity to have a closer insight into the mechanism of both PB and SB transpositions. We have captured autointegration products at comparable frequencies for both SB and PB. We assume that the excision and reintegration steps of autointegration and canonical transposition are mechanistically not significantly different [32], [41], [74] (Figures 1 and S1).
In addition to the autointegration products, our assays detected aberrant, pseudo-transposition events. In the ‘single-ended’ transposition products of PB, one IR of the transposon was clearly separated from the donor site, without obvious involvement of the other IR in the reaction (Figure 3D). The liberated end of PB targeted either the transposon or the backbone DNA (Figures 3C and 3E). SB did not display this feature in a similar assay system. By contrast, both transposons were capable of mobilizing substrates, lacking one of the IRs from separate molecules ([52] and this work). These bimolecular transposition events were eight-fold more frequently detected for PB.
How could aberrant transposition events be generated? In fact, ‘true single ended’ transposition, when a transposase interacts with a single transposon end, performs the cleavage and integration steps without the involvement of another end has not been undoubtedly reported from any system. In fact, alternative mechanisms can generate hard-to distinguish, similar products. For example, the canonical transposition reaction could fail at the final step, and only one end of the transposon is transferred (lariat model). In addition, our ‘solo’ experimental data support the ‘bimolecular model’, when the ends of the transposon derive from separate molecules [51]. In addition to single ended events, small deletions at the donor sites of PB transposition are assumed to be associated with imprecise transposon excision, and involve non-homologous end joining [40]. These structures were reported following PB excision in Drosophila (4.3%), mouse (5%) and in human cells [38], [40].
Aberrant pseudo-transposition can be considered as a fidelity problem of the transposition reaction, and has been observed with P-element in Drosophila, Ds element in Arabidopsis, Ac/Ds elements in maize [51], [75] or Tam3 in Antirrhinum majus [76]–[80]. Small sequence variations generated by NHEJ at the excision sites are unlikely to cause genome rearrangements. By contrast, pseudo-transposition events can generate difficult-to-repair lesions and be genotoxic. Aberrant transposition events were reported to induce deletions, insertions, chromosome translocations and could initiate McClintock's chromosomal breakage-fusion-bridge cycles [51], [81]. Occasional mis-pairing between extrachromosomal molecules would not compromise the safety feature of a transposon-based transfer vector in a heterologous environment. However, fidelity problems could be problematic when the transposon is mobilized from the genome. Thus, cells subjected to PB-based genome manipulation techniques, e.g., transgene-free iPS cells generated by PB excision [82], should be carefully monitored for genome rearrangements.
There seems to be a basic difference in the ways transposons in pro- and eukaryotes control their activity to minimize the potential genotoxicity generated by improper synapsis of the transposon ends. For all classical bacterial transposons characterized to date, including Tn5 transposition, the catalytic steps of the reaction are tightly coupled to the synapsis of the transposon ends [83]. In addition, the coupling of transcription and translation in bacteria also increases the probability of a proper synapsis as the transposase binds tightly to the first IR before searching for nearby ends. In contrast, eukaryotic transposases must search at random for transposon ends when they enter the nucleus. Therefore, regulatory mechanisms promoting accurate double-ended reactions from the same transposon molecule are crucial.
Tc1/mariner transpositions, including SB, might have invented novel “built in regulatory checkpoints” to enforce synapsis prior catalysis [21]. A simple topological filter could also suppress promiscuous synapses of distant ends of the transposon [84]. Furthermore, certain transposition-like reactions, including V(D)J recombination, are also capable of filtering out unpaired reaction products. This regulatory mechanism, assisted by a cellular factor, HMGB1, regulates a highly controlled, ordered assembly process [85], [86]. Similarly to V(D)J recombination, HMGB1 was reported to assist paired end complex formation of SB [11]. In addition to HMGB1, SB transposition requires various vertebrate-specific host factors [10], [11], [13], [29] that render SB transposition restricted to vertebrates. In contrast, PB has an incredibly wide host range (from yeast to human) that could be associated with loose or no host factors requirement.
In comparison to SB, PB transposition results in more frequent, aberrant transposition products in a heterologous environment. Why is it so? If PB does not use host factors to enforce fidelity of the end pairing before excision, the reaction might be less precise by its nature. Alternatively, PB might utilize a host factor in its endogenous host (insect) that guarantees precise regulation. However, this factor is diverged or not available in mammalian cells. Finally, both PB and SB transposons have “built in regulatory checkpoints” that are most effectively filter out aberrant products under optimal conditions and in appropriate hosts. Notably, aberrant transposition events, including single-ended transposition of the Mos1, mariner element were observed under suboptimal conditions [22]. In sum, when a transposon is transferred too far from its original host, the conditions in a new environment could be suboptimal, and the fidelity of the reaction could be compromised. The wide host range of PB can be explained by relative independence from host-encoded factors, perhaps a price to be paid for fidelity.
The IRs of the transposons were identical to the versions published earlier [49], [87] and were not modified for the assays. All the primers used for construct cloning were listed in Supporting Table S1. SBrescue: XmnI/BsaI fragment (Klenow-filled) containing ampicillin gene on pUC19 was replaced by PstI and SalI fragment containing zeocin gene from vector pZEO (isolate SV1, Invitrogen) resulting in pUC19-zeo. Klenow-filled SapI/SspI fragment containing zeocin gene and replication origin was inserted into EcoRI site of PT2/HB to get PT2/SBzeo. The transposon was PCR-amplified with primer AATASB-IR from PT2/SBzeo and ligated to rpsL gene fragment, which was PCR-amplified with primers rps1F/rpslR from nNG639 [43]. SB2K: BspHI/EcoRI fragment containing zeocin gene on SBrescue was replaced by BsaI/BglII fragment containing zeocin and promoter sequences from pFP-Zeo [88]. SB3K, SB4K and SB7K: DNA fragments were PCR-amplified from bacteriophage lamda DNA, using primers lam1kF/lam1kR, lam1kF/lam2kR and lam1kF/lam6kR, respectively, and were inserted into XbaI site (Klenow filled) of SB2K. PB2K: Klenow-filled NotI/HindIII fragment containing zeocin gene from SBrescue was inserted into SpeI site of pUC19PBneo [72] resulting in PUC19XLzeo. PvuII fragment containing PB transposon was ligated to rpsL gene PCR-amplified with primers rps1F/rpslR from nNG639. PB3K, PB4K, PB7K: The AatII/BglII fragments containing lamda DNA from SB3K, SB4K and SB7K were inserted into AatII/BglII sites of PBPr respectively. pcDNA3.1BANF1 (BANF1 gene expressing vector): BANF1 coding sequence was PCR-amplified from pcDNA3.1/HiscBANF1 (a gift from Katherine Wilson, Johns Hopkins University) with primers BAFF/BAFR and cloned into EcoRV site of pcDNA3.1/Zeo (+) (Invitrogen). BAF-RNAi: Oligos of BAF96F/BAF96R were annealed together and cloned into BglII/HindIII site of pFP-Neo-H1 [88]. To generate ‘solo’ substrates PB pUC19XLneo [69] was digested with BamHI to delete the right IR (PBΔright) or with KpnI to remove the left IR (PBΔleft). For “solo” SB, pTneo was digested EcoRI to generate SBΔleft, while the digestion with BamHI yielded SBΔright.
HeLa, AA8 and mouse MEF cells were cultured at 37°C with 5% CO2 in Dulbecco's modified Eagle's medium (DMEM, Gibco/Invitrogen) supplemented with 10% fetal calf serum (FCS, PAA). The zebrafish PAC2 cells were grown at room temperature and atmospheric CO2 concentrations in Leibovitz L15 medium (Gibco/Invitrogen) supplemented with 15% FCS. Cells were transfected at 50–80% confluence with QIAGEN-purified plasmid DNA using jetPEI (Polyplus transfection, for mammalian cells) or FuGene6 (Roche, for fish cells) according to instructions of manufacture. Transfection efficacy of a ∼3 kb and a ∼7 kb plasmid containing GFP cassette was monitored and compared by FACS analysis, but no significant difference was found (not shown).
Cell culture and transfection was done as described [9]. Typically, 1.5×105 cell were subjected to transfection with plasmids containing the transposon (500–1000 ng) and the transposase (60–100 ng). Two days post transfection plasmid DNA was recovered and transformed into bacteria (Invitrogen, ElectroMAX DH10B Cells, Cat. No. 18290-015, Genotype: F– mcrA Δ(mrr-hsdRMS-mcrBC) Φ80lacZΔM15 ΔlacX74 recA1 endA1 araD139 Δ(ara leu) 7697 galU galK rpsL nupG λ–). Bacteria were subjected to either zeocin (to determine total number of plasmids) or zeocin/streptomycin double selection (to determine autointegration events). The number of autointegration events was normalized by total number of plasmids. To confirm autointegration events, individual bacterial colonies were cultured and recovered plasmid DNA was subjected to DNA sequencing using primers of psbLacR3 and PB-F or PB-R for SB- and PB transposon, respectively. For BANF1 overexpression or knockdown experiments, 300 ng of pcDNA3.1BANF1 or BAF-RNAi plasmid was cotransfected with the transposon and helper constructs.
Cell culture and transfection was done as described [9]. Two days post transfection 105 cells were plated on 10 cm dishes and exposed to antibiotic selection (100 ng/ml zeocin, for two weeks). Resistant colonies were visualized by methylene blue staining [9]. Transgene copy number was normalized by using qPCR specific to zeocin.
The plasmid DNA was prepared as described in autointegration assay and dissolved in 50 µl water. Excision frequencies of eight transposon plasmid constructs of various sizes (four SB and four PB) were estimated by using a quantitative, real-time PCR (7700 sequence detection system from ABI, Applied Biosystems, Foster City, CA). To determine the total number of parental plasmid DNA molecules, a ‘parental’ titration curve was established. PCR primers of rpsL-F/rpsL-probe/rpsL-R were used to amplify the rpsL gene on the construct of SBrescue. For the curve, dilutions of 10−2, 10−3, 10−4, 10−5, 10−6 ng of SBrescue plasmid DNA were subjected to a PCR reaction to amplify the rpsL gene (rpsL-F/rpsL and probe/rpsL). To quantify the total number of parental plasmid molecules, total DNA extract was used (3 µl, diluted by 2000-fold, rpsL-F/rpsL-probe/rpsL-R). The excision products were PCR-amplified from the total extract DNA using nested PCR (1st round, primers of rpslexciF1/rpslexciR1, 94°C for 30 s and 30 cycles of 94°C for 30 s, 58°C for 30 s, and 72°C for 30 s; 2nd round, rpslexciF2/rpslexciR2, 1 µl, diluted by 100-fold, 94°C for 30 s and 35 cycles of 94°C for 30 s, 58°C for 30 s, and 72°C for 30 s). The amplified products (10−2, 10−3, 10−4, 10−5, 10−6 ng) were used to establish a second titration curve, specific for the excision products. To quantify excision products, primers of SB-F/SB-probe/SB-R and PB-F/PB-probe/PB-R were used on a total DNA extract (5 µl), for SB and for PB, respectively. The excision frequency was calculated as the ratio of excision products normalized by the total number of parental plasmid molecules. qPCR was performed for each experimental sample in triplicates. Ct values were determined following recommendations by the manufacturer.
Briefly, bacteria were picked by a pipette tip and directly subjected to a PCR assay using primers of PB-F and PB-R (5 pmol of each, Supporting Table S1) and Taq polymerase (Takara) in a total volume of 20 µl. PCR program: 94°C for 1 min; 30 cycles of 94°C for 30 s, 58°C for 30 s, and 2°C for 30 s; and 72°C for 2 min.
A triple SILAC pull-down experiment was performed using anti-HA resin. HEK293T cells were transiently transfected with HA-tagged wild type or mutant HMGXB4 (HMG2l1) [13] and SUMO1 in the presence/absence of Sleeping Beauty, SB10 [9] using Polyplus-transfection jetPEI transfection reagent with 3 µg of plasmids each. We compared proteins co-purifying with HA in cells expressing the empty vector (“light”), HA-tagged HMGXB4− with mutated sumoylation site (“medium”) and HA-tagged wild-type HMGXB4 (“heavy”). The cells were plated on a 15-cm dish and harvested 48 h post-transfection. Two dishes were used for each condition. Detection of interaction partners is performed by mass spectrometry and the results obtained were analyzed by MaxQuant computational platform [89]. Results presented show protein abundance ratios between cells transfected with HMGXB4− and the empty vector control.
Whole-cell extracts were prepared using extraction buffer (Tris-HCl 50 mM at pH 8.0, NaCl 150 mM, 0.1% SDS (Na-dodecylsulphate) Triton X-100 1% and Na-deoxycholate 0.5%) supplemented with protease inhibitor cocktail (Roche, Mannheim, Germany). For immunoprecipitations, equal amounts of lysate (containing 5 mg of total cellular protein from HEK293 cells) were pre cleared with protein G-agarose beads (Sigma, St Louis, MO). Pre-cleared extracts were incubated with EZview Red Anti-HA Affinity Gel (Sigma-Aldrich, USA) for 1 h at 4°C. Precipitates were washed extensively in extraction buffer. Bound complexes were eluted with 2× SDS–PAGE sample buffer and resolved by 7.5–15% SDS–PAGE. Immunoblotting was performed according to standard procedures and proteins detected with the indicated antibodies. Antibodies were detected by chemiluminescence using ECL Advance Western Blotting Detection Kit (Amersham Bioscience).
|
10.1371/journal.pbio.1002481 | Regulation of Smoothened Phosphorylation and High-Level Hedgehog Signaling Activity by a Plasma Membrane Associated Kinase | Hedgehog (Hh) signaling controls embryonic development and adult tissue homeostasis through the G protein coupled receptor (GPCR)-family protein Smoothened (Smo). Upon stimulation, Smo accumulates on the cell surface in Drosophila or primary cilia in vertebrates, which is thought to be essential for its activation and function, but the underlying mechanisms remain poorly understood. Here we show that Hh stimulates the binding of Smo to a plasma membrane-associated kinase Gilgamesh (Gish)/CK1γ and that Gish fine-tunes Hh pathway activity by phosphorylating a Ser/Thr cluster (CL-II) in the juxtamembrane region of Smo carboxyl-terminal intracellular tail (C-tail). We find that CL-II phosphorylation is promoted by protein kinase A (PKA)-mediated phosphorylation of Smo C-tail and depends on cell surface localization of both Gish and Smo. Consistent with CL-II being critical for high-threshold Hh target gene expression, its phosphorylation appears to require higher levels of Hh or longer exposure to the same level of Hh than PKA-site phosphorylation on Smo. Furthermore, we find that vertebrate CK1γ is localized at the primary cilium to promote Smo phosphorylation and Sonic hedgehog (Shh) pathway activation. Our study reveals a conserved mechanism whereby Hh induces a change in Smo subcellular localization to promote its association with and activation by a plasma membrane localized kinase, and provides new insight into how Hh morphogen progressively activates Smo.
| The secreted glycoprotein Hedgehog (Hh) plays a conserved role in embryonic development and adult tissue homeostasis in species ranging from Drosophila to humans. Deregulation of Hh signal transduction contributes to a wide range of human disorders, including birth defects and cancer. The seven-transmembrane protein Smoothened (Smo) is an obligatory and conserved Hh signal transducer, but how Hh stimulates its activity remains unclear. Here we identify a plasma membrane associated kinase, Gilgamesh (Gish)/CK1γ, as a positive regulator of the Hh signaling activity. We find that Gish activates Hh signaling by phosphorylating a specific site in the Smo C-terminal intracellular tail. Phosphorylation of Smo by Gish is required for maximal activation of Smo and depends on membrane association of Gish and prior phosphorylation of Smo by protein kinase A (PKA). We also find that Hh stimulates the association of Smo with Gish after it travels to the plasma membrane, thus facilitating its phosphorylation by Gish. Finally, we provide evidence that CK1γ is found at the primary cilium in mammals and phosphorylates Smo to activate the Hh pathway. Our results uncover a conserved role of Gish/CK1γ in the regulation of Smo phosphorylation and provide new insight into the molecular underpinning of how Hh signal is transduced across the plasma membrane.
| Hedgehog (Hh) signaling plays an essential role in embryonic development and adult tissue homeostasis, and its deregulation has been implicated in congenital diseases and cancers [1–6]. Hh exerts its biological influence through an intracellular signal transduction cascade that emanates from a G protein coupled receptor (GPCR)-family protein Smoothened (Smo) and culminates in the activation of the latent transcription factor Cubitus interruptus (Ci)/Glioma-associated oncogene homologue (Gli) [1,3,7,8]. In the signaling off state, Smo is inhibited by a twelve-transmembrane protein Patched (Ptc). Binding of Hh to Ptc alleviates such inhibition, allowing Smo to be phosphorylated and accumulate on the cell surface in Drosophila or primary cilia in vertebrates, where Smo adopts an open and active conformation to relay the Hh signal to the intracellular signaling components [9–15].
Phosphorylation plays a critical role in the regulation of Smo conformation and subcellular localization [16]. In Drosophila, unphosphorylated or hypophosphorylated Smo adopts a closed, inactive conformation [12] and is ubiquitinated and removed from the cell surface by both proteasome- and lysosome-mediated degradation [17,18]. Upon Hh stimulation, Smo is phosphorylated by protein kinase A (PKA) and casein kinase 1 (CK1), mainly the CK1α/ε isoforms, at three clusters of Ser/Thr residues in its carboxyl-terminal intracellular tail (C-tail) [19–22], which drives a conformational switch of Smo C-tail from the closed inactive to an open active conformation, leading to dimerization/oligomerization of the C-tail [12]. In addition, PKA/CK1α/ε-mediated phosphorylation inhibits ubiquitination of Smo, leading to its cell surface accumulation and activation [17,18]. In addition to PKA and CK1α/ε, Smo activity is also modulated by casein kinase 2 (CK2), atypical protein kinase C (aPKC), and G protein coupled receptor kinase 2 (Gprk2) [23–26]. Gprk2 promotes Smo activity by directly binding and phosphorylating the Smo C-tail to stabilize its active conformation [24]. In mammals, Hh-stimulated phosphorylation of Smo by CK1α and Gprk2 promotes its ciliary localization and active conformation [15]. Cell surface/ciliary accumulation of Smo is thought to be essential for its activation and function [14,27]; however, the underlying mechanism is still poorly understood.
Hh functions as a morphogen to specify different developmental outcomes in a concentration-dependent manner [1,3]. As such, Hh signal transduction needs to be tightly controlled to achieve pathway activities appropriate with the ligand inputs. Although the characterized phosphorylation events contribute to Smo activation, they do not represent all the activation mechanisms, as phospho-mimetic mutations failed to fully activate Smo in both Drosophila and mammals [15,19,24], suggesting that additional mechanisms, either phosphorylation-dependent or independent, may exist. Indeed, Drosophila Smo is phosphorylated at more than 26 Serine (Ser)/Threonine (Thr) residues, many of which have not been well characterized [20].
CK1γ, which is encoded by gilgamesh (gish) in Drosophila [28], is a membrane-associated Ser/Thr kinase of the CK1 family [29]. Gish/CK1γ has been implicated in the regulation of Wingless (Wg)/Wnt signaling by phosphorylating the co-receptor Arrow (Arr)/LRP5/6 [30,31]. Gish is also involved in glial cell migration in Drosophila eye [28], olfactory learning [32], and planar cell polarity (PCP)-mediated morphogenesis [33]. Mammals have three CK1γ isoforms encoded by different genes [29], making it difficult to study the role of CK1γ in development. In this study, we identified Gish as a positive regulator of Hh signaling through a genetic modifier screen. We demonstrated that Hh stimulates the association between Gish and Smo in a manner depending on cell surface localization of both Gish and Smo as well as PKA-mediated phosphorylation of Smo C-tail. We provided evidence that Gish phosphorylates a membrane proximal region of Smo C-tail to promote high levels of Hh pathway activity. However, loss of Gish only caused a minor defect in Hh pathway activity, likely due to a redundancy with another kinase(s). We also found that vertebrate CK1γ is localized at primary cilia depending on its membrane association, and that CK1γ promotes Smo phosphorylation and Sonic hedgehog (Shh) pathway activation depending on the primary cilia. Our results suggest that plasma membrane/ciliary-localized CK1γ plays a conserved role in Hh signaling by promoting the maximal levels of Smo activity.
To identify additional Hh pathway regulators, we have conducted an RNAi-based genetic modifier screen to identify enhancers or suppressors of a "fused wing" phenotype caused by expression of a dominant negative Smo (Smo-PKA12/SmoDN) with a wing-specific Gal4 drivers MS1096 (MS>SmoDN) [19,24,25]. We found that expressing a UAS-RNAi line (V106826) targeting Gish, the Drosophila homologue of mammalian CK1γ, enhanced the “fused wing” phenotype caused by MS>SmoDN (Fig 1A–1C). Using immunostaining with a Gish antibody, we found that Gish RNAi effectively knocked down Gish protein expression in wing imaginal discs (S1A–S1B' Fig). Moreover, two additional UAS-GishRNAi lines, V26003 and BL28066, enhanced the MS>SmoDN-induced phenotype in a similar fashion (S1C–S1F Fig). On the other hand, overexpression of a wild-type Gish, but not a kinase dead form (GishKD), partially suppressed the “fused wing” phenotype caused by MS>SmoDN (Fig 1D and 1E and S1K and S1L Fig).
Gish/CK1γ is localized to the plasma membrane due to its C-terminal palmitoylation [30,33]. Interestingly, overexpression of a soluble form of Gish with its palmitoylation site deleted, GishΔC [33], failed to rescue MS>SmoDN-induced wing phenotype, even though it was expressed at levels similar to the wild-type Gish (Fig 1F and S1M Fig compared with S1K Fig), suggesting that plasma membrane association of Gish is critical for its function in this context. Consistent with this notion, overexpression of other soluble CK1 family members, including CK1α and CK1ε, also failed to rescue the MS>SmoDN phenotype (Fig 1G and 1H).
To determine whether loss- or gain-of-Gish function modified the "fused wing" phenotype through the Hh pathway, we examined the expression of an Hh target gene ptc. In control late third instar wing imaginal discs, Hh induced ptc expression in A-compartment cells near the A/P boundary (Fig 1A'). In MS>SmoDN wing discs, ptc expression near the A/P boundary was greatly reduced (Fig 1B'). Gish RNAi in MS>SmoDN wing discs nearly abolished ptc expression near the A/P boundary in the wing pouch region where MS1096 was expressed (Fig 1C'). On the other hand, overexpression of Gish restored ptc expression close to wild-type levels (Fig 1D'). Hence, gain- or loss-of-Gish activity can modulate Hh pathway activity.
To determine where Gish acts in the Hh pathway, we examined its genetic interaction with Fused (Fu), a Ser/Thr kinase acting downstream of Smo [11,34,35]. Inactivation of Fu by expressing a USA-RNAi transgene with MS1096 (MS>FuRNAi) caused a similar "fused wing" phenotype, albeit more severe than that caused by MS>SmoDN (S1H Fig). However, neither Gish overexpression nor RNAi modified the wing phenotype caused by MS>FuRNAi (S1I and S1J Fig), suggesting that Gish may act upstream of Fu in the Hh pathway.
To confirm the Gish RNAi phenotype, we turned to gish mutants. gishKG03891 is a P-element insertion mutation and a strong allele of gish [32], which is referred to as gishP hereafter. Although gishP heterozygosity did not modify the fused wing phenotype caused by MS>SmoDN (Fig 2B compared with Fig 2A), MS>SmoDN wings carrying gishP homozygous clones exhibited a greatly enhanced phenotype similar to that caused by Gish RNAi in MS>SmoDN wings (Fig 2C compared with Fig 1C and S1E and S1F Fig). Interestingly, we found that heterozygosity for gish deficiency, Df(3R)ED10639 (BL#9481), also enhanced the fused wing phenotype caused by MS>SmoDN (Fig 2D), suggesting that gishP is not a null allele. To obtain a gish null allele, we generated imprecise excision lines from gishP and found several lines including gishΔ4 that could enhance the MS>SmoDN phenotype similarly to gishDf (Fig 2E). Consistent with gishΔ4 being a null allele, gishΔ4 mutant clones in wing discs exhibited diminished Gish immunostaining (S2A and S2B Fig). Furthermore, the enhancement of the MS>SmoDN phenotype by gishΔ4/+ was reversed by coexpression of the wild-type Gish but not by coexpression of either GishKD or GishΔC (Fig 2F–2H).
Our previous study identified Gprk2 as a positive regulator of Hh signaling in a similar genetic modifier screen [24]. Heterozygosity of a gprk2 null allele, gprk2Δ15, also enhanced the fused wing phenotype caused by MS>SmoDN (Fig 2I). Interestingly, taking away one copy of gish in this background (MS>SmoDN gish Δ4/+ gprk2 Δ15/+) further enhanced the wing phenotype (Fig 2J). This dosage-sensitive genetic interaction between Gish and Gprk2 suggests that they may act in close proximity in the Hh signaling pathway.
We then induced gishΔ4 clones using the MARCM system [36] at 24–48 h or 48–72 h after egg laying (AEL). gishΔ4 clones (marked by green fluorescent protein [GFP] expression) induced at 48–72 h AEL survived to late third instar larval stages but did not affect the expression of Hh target genes ptc and engrailed (en) when localized in A-compartment cells near the A/P boundary (S2D–S2D" Fig and Fig 2F–2F" compared with S2C–S2C" and S2E–S2E" Fig). gishΔ4 clones induced at 24–48 h AEL were barely recovered, suggesting that gish mutant cells had a growth disadvantage and were competed out by wild-type cells over time. To recover early-induced clones, we generated gishΔ4 clones in a Minute background, which gave gish mutant cells a growth advantage [37]. As shown in Fig 2 and S2 Fig, gishΔ4 clones (marked by the lack of Gish staining) induced at 24–48 h AEL in the Minute background occupied large areas in late third instar wing discs (Fig 2L and 2N, S2H Fig). ptc expression, which is induced by intermediate levels of Hh, was not affected in gishΔ4 clones (S2H" and S2H‴ Fig); however, en expression in A-compartment cells near the A/P boundary, which is induced by peak levels of Hh, was diminished in gishΔ4 clones (arrow in Fig 2L–2L‴). In addition, Hh-induced Smo accumulation appeared to be attenuated in gishΔ4 clones (arrow in Fig 2N–2N‴). These results suggest that Gish is required for high levels of Hh signaling activity and may regulate the Hh pathway at the level of Smo.
A previous study revealed that Smo derived from Hh-stimulated S2 cells was phosphorylated at 26 Ser/Thr residues in its C-tail, including the three PKA/CK1 phosphorylation clusters and Gprk2 sites (Fig 3A) [20]. However, the kinases responsible for phosphorylating other sites, most notably, a membrane proximal cluster (CL-II) of Ser/Thr residues, 623DlNSSETNDISS634 (underlined S/T residues were phosphorylated sites detected by Mass Spec) [20], have not been identified (Fig 3A). A close inspection of the CL-II site indicates that it contains Ser/Thr residues falling into the consensus sites for the CK1 family kinases: D/E/(p)S/T(X)1-3S/T, in which the underlined S/T is the CK1 site, whereas X represents any amino acid [29]. Indeed, a Glutathione S-transferase (GST) fusion protein containing the intact (GST-Smo601-700) but not the mutated CL-II site (GST-Smo601-700 CL-II SA) was phosphorylated by a recombinant CK1 in an in vitro kinase assay (Fig 3B). To determine whether CL-II is phosphorylated by Gish, we transfected S2 cells with a Smo construct, Myc-Smo△C650, which contains the CL-II site but lacks distal phosphorylation sites such as the PKA/CK1 phosphorylation clusters and Gprk2 sites. We found that Hh stimulation induced a mobility shift of Myc-Smo△C650, which is indicative of Smo phosphorylation [19], and that Hh-induced mobility shift of Myc-Smo△C650 was diminished by Gish RNAi (Fig 3C). Coexpression of Myc-Smo△C650 with Gish also induced a mobility shift of Myc-Smo△C650 (Fig 3C and 3D). Furthermore, mutating the CL-II site in Myc-Smo△C650 (Myc-Smo△C650CL-IISA) abolished its mobility shift induced by either Hh stimulation or Gish overexpression (Fig 3D). Taken together, these results suggest that Hh induces phosphorylation of CL-II through Gish.
To characterize CL-II phosphorylation in the context of full-length Smo, we generated a phospho-specific antibody named Smo4P using the phospho-peptide NDLN(PS)(PS)E(PT)NDI(PS)STW as an antigen (see Materials and Methods). Western blot analysis indicated that purified Smo4P antibody recognized GST-Smo601-700 but not GST-Smo601-700CL-llSA after in vitro phosphorylation by CK1. We found that Smo4P recognized Myc-Smo but not Myc-SmoCL-llSA derived from S2 cells stimulated with Hh or coexpressing Flag-Gish (Fig 3E and 3F). Overexpression of Gish further increased the Hh-stimulated Smo4P signal (Fig 3F). To further demonstrate that Hh stimulates CL-II phosphorylation through Gish, we examined the phosphorylation state of Myc-Smo derived from transiently transfected S2 cells or a stably expressing cell line treated with control or Gish dsRNA and stimulated with Hh. We found that Gish RNAi abolished Hh-stimulated Smo4P signal associated with Myc-Smo (Fig 3G and 3H). A recent study argued that Gprk2 is responsible for CL-II phosphorylation [38]; however, we found that Gprk2 RNAi or overexpression did not affect Smo4P signal intensity associated with Myc-Smo (S3A and S3B Fig). Furthermore, treatment of Myc-Smo-expressing cells with a pharmacological CK1 inhibitor, D4476, also abolished Hh-stimulated Smo4P signal (Fig 3G and 3H). These results suggest that Hh stimulates CL-II phosphorylation through Gish rather than Gprk2.
Gish/CK1γ is attached to the inner leaf of the plasma membrane through its C-terminal palmitoylation (Fig 4A) [30]. Indeed, HA-Gish was mainly associated with cell membrane when expressed in S2 cells (Fig 4B). By contrast, HA-GishCS and HA-GishΔC, which have their C-terminal palmitoylation signal (SRCCCFFKR) substituted (SRSSSFFKR) or deleted, respectively (Fig 4A), exhibited cytoplasmic distribution (Fig 4B) [33]. We then coexpressed HA-Gish, HA-GishCS, or HA-GishΔC with Myc-Smo in S2 cells with endogenous Gish knocked down by dsRNA targeting the 5' UTR of gish. Western blot analysis with Smo4P indicated that only the membrane-associated form of Gish (HA-Gish) but not the cytosolic variants (HA-GishCS and HA-GishΔC) could support CL-II phosphorylation in response to Hh (Fig 4C), suggesting that plasma membrane association of Gish is critical for CL-II phosphorylation. In addition, we found that coexpression of the cytosolic CK1 family members CK1α and CK1ε with Myc-Smo did not significantly increase the Smo4P signal in either the presence or absence of Hh stimulation (Fig 4D and S3C Fig). These results may explain why overexpression of wild-type Gish but not GishΔC, CK1α, or CK1ε could partially rescue the wing phenotype caused by MS-SmoDN (Fig 1D and 1F–1H).
To determine whether cell surface localization of Smo is important for CL-II phosphorylation, we employed a Myc-tagged Smo variant that has a ubiquitin (Ub) moiety fused to its C-terminus (Myc-Smo-Ub) [18]. We confirmed the previous finding that Myc-Smo-Ub failed to accumulate on the cell surface in response to Hh (Fig 4E) [18,39]. However, Myc-Smo-Ub could still be phosphorylated by PKA and accumulate on the cell surface in response to PKA phosphorylation (S4 Fig), suggesting that adding the Ub moiety to the Smo C-terminus did not cause an overall structure change of Smo C-tail to preclude its phosphorylation by any kinase. We then determined whether failure to accumulate on the cell surface affected Hh-induced phosphorylation of Smo at the CL-II site. As shown in Fig 4F, Hh-induced Smo4P signal associated with Myc-Smo-Ub was dramatically reduced compared to that associated with Myc-Smo. This result suggests that cell surface localization of Smo is critical for its phosphorylation by Gish.
Previous studies suggest that Hh stimulates Smo phosphorylation by PKA and CK1α/ε at three clusters of Ser/Thr residues in the middle region of Smo C-tail (Fig 3A), and that these phosphorylation events promote Smo cell surface accumulation and conformational change [12,19]. To determine whether PKA-mediated phosphorylation of the distal sites regulates Gish-mediated phosphorylation of the membrane proximal sites, we first treated Myc-Smo expressing cells with a pharmacological PKA inhibitor H89 and found that inhibition of PKA activity diminished the Hh-stimulated Smo4P signal associated with Myc-Smo (Fig 4G). We then compared CL-II phosphorylation of a PKA-phosphorylation deficient (Myc-SmoSA123) or a phospho-mimetic (Myc-SmoSD123) form of Smo with that of wild-type Myc-Smo [19]. We found that Hh failed to stimulate the Smo4P signal associated with Myc-SmoSA123 (Fig 4H). On the other hand, Myc-SmoSD123 exhibited enhanced basal Smo4P signal, which was further enhanced upon Hh stimulation (Fig 4H). Taken together, these results suggest that PKA/CK1-mediated phosphorylation of Smo in the distal region facilitates Gish-mediated phosphorylation of the juxtamembrane region of Smo C-tail.
We next determined whether PKA site phosphorylation and CL-II phosphorylation were induced by different levels of Hh. S2 cells stably expressing Myc-Smo were treated with Hh-conditioned medium containing different levels of Hh-N (20%, 40%, 60%, 80%, or 100%) for 4 h. Cell lysates were immunoprecipitated with anti-Myc antibody, followed by western blot analysis with either anti-SmoP687, which recognized the phosphorylated PKA site (S687) as well as two downstream CK1 sites [40], or anti-Smo4P. As shown in Fig 4I, SmoP687 signal began to be detected at 40% Hh, whereas Smo4P signal was detected only when the Hh levels exceeded 60% Hh. Myc-Smo expressing S2 cells were also treated with 50% Hh for different periods of time (4, 8, 12, 16, 20, and 24 h). SmoP687 signal began to be detected 8 h after Hh stimulation, whereas Smo4P was not detected until 16 h after Hh stimulation (Fig 4J). Hence, the CL-II phosphorylation appears to require higher levels of Hh or longer exposure to the same level of Hh than the PKA-site phosphorylation.
To determine the functional importance of CL-II phosphorylation, we mutated CL-II in the context of Smo-cyan fluorescent protein (CFP) or SmoSD123-CFP to generate SmoCL-IISA-CFP and SmoSDCL-IISA-CFP, respectively (see Materials and Methods). Consistent with previous findings [11,19], overexpression of Smo-CFP using the MS1096 Gal4 driver (MS>Smo-CFP) induced ectopic expression of dpp-lacZ, which is a low-threshold Hh target gene (Fig 5A); however, SmoCL-IISA-CFP failed to induce ectopic dpp-lacZ expression (Fig 5B). While MS>SmoSD123-CFP induced ectopic expression of not only dpp-lacZ but also ptc-lacZ and en at high levels (Fig 5C–5C"), SmoSDCL-IISA-CFP failed to induce ectopic expression of en and only induced ectopic expression of ptc-lacZ at low levels, although the ectopic expression of dpp-lacZ was not affected by the CL-IISA mutation (Fig 5D–5D"). Hence, mutating the CL-II site in Smo compromised its ability to activate the Hh pathway. Coexpression of Flag-Gish with Smo-CFP increased the ectopic dpp-lacZ expression, leading to more dramatic overgrowth of the wing discs (S5A and S5B' Fig). By contrast, coexpression of Flag-Gish with SmoCL-IISA-CFP did not alter its activity (S5C and S5D' Fig), consistent with the notion that Gish promotes Hh signaling activity through phosphorylating the CL-II site.
To discern the in vivo function of CL-II phosphorylation more precisely, we expressed Smo-CFP and SmoCL-IISA-CFP at low levels using the weak Gal4 driver C765 (C765>Smo) in wing discs carrying smo3 mutant clones. Our previous study revealed that the levels of Smo derived from C765>Smo were only slightly higher than that of endogenous Smo [22]. We found that C765>Smo-CFP completely rescued the expression of both ptc and en in smo3 mutant clones located near the A/P boundary (Fig 5G–5H" compared with Fig 5E–5F"). By contrast, C765>SmoCL-IISA-CFP failed to rescue en expression and only partially rescued ptc expression in smo3 clones (Fig 5I–5J"). In addition, A/P boundary-located smo3 cells expressing C765>SmoCL-IISA-CFP accumulated high levels of full-length Ci (Fig 5I'), suggesting SmoCL-IISA-CFP could still inhibit Ci processing but fail to induce the maturation of full-length Ci into the active but labile form [41]. Furthermore, we found that SmoCL-IISA-CFP, like Smo-CFP, promoted Ci nuclear localization in smo mutant clones (S5E–S5F′′′ Fig). Previous studies revealed that Ci nuclear translocation occurs in A-compartment cells more than ten cells away from the A/P boundary, where there are low levels of Hh [42,43]. Collectively, these results demonstrate that CL-II phosphorylation is essential for Smo to transduce high levels of Hh signaling activity but is dispensable for low levels of Hh pathway activity.
Our previous study demonstrated that Gprk2 promotes high levels of Hh pathway activity by regulating the active state of Smo through both kinase activity dependent and independent mechanisms [24]. Gprk2 phosphorylated Smo at Ser741/Thr742 and S1013/S1015, and mutating these sites (GPSA12) in the context of SmoSD123 (SmoSDGPSA) compromised Hh pathway activity [24]. However, expression of a Smo variant with the Gprk2 sites mutated (SmoGPSA) using the strong Gal4 driver MS1096 fully rescued the expression of Hh target genes including ptc and en in A/P boundary-located smo3 clones. We reasoned that overexpression of SmoGPSA to levels much higher than the physiological level could mask its signaling defect. Indeed, expression of SmoGPSA with C765 only rescued ptc expression but failed to restore en expression in A/P boundary-located smo3 clones (S6A–S6B′′ Fig). Hence, Grpk2-mediated phosphorylation of Smo is also required for its optimal activity.
To further determine the contribution of Gish- and Gprk2-mediated phosphorylation of Smo to Hh pathway activation and how differential Smo phosphorylation generates degraded Hh pathway activity, we introduced phospho-mimetic mutations to the CL-II site (CL-IISD) in the context of SmoSDGPSD, which contains phospho-mimetic mutations in the three PKA/CK1 phosphorylation clusters (SD123) and two Gprk2 sites (GPSD) [24], to generate SmoSDall. Hence, SmoSDall represents Smo with the highest level of phosphorylation, followed by SmoSDGPSD, SmoSD123, and SmoSDCL-IISA. When expressed in wing discs using C765, SmoSD123 induced ectopic ptc-lacZ expression in A-compartment cells both distant from and close to the A/P boundary, albeit at levels lower than that of endogenous ptc-lacZ at the A/P boundary (Fig 6A and 6A'). In addition, SmoSD123 induced weak ectopic en expression in A-compartment cells near the A/P boundary (Fig 6A"). SmoSDCL-IISA induced ectopic ptc-lacZ expression at lower levels than SmoSD123 and failed to induce any ectopic en expression (Fig 6B–6B"), consistent with its activity being weaker than SmoSD123. Similar to SmoSDCL-IISA, SmoSDGPSA also exhibited weaker activity than SmoSD123 because it only induced low levels of ectopic ptc expression but failed to induce ectopic en expression (S6C–S6D′′ Fig). On the other hand, SmoSDGPSD and SmoSDall induced ectopic expression of ptc-lacZ and en at higher levels than SmoSD123 (Fig 6C–6D"). Although both SmoSDall and SmoSDGPSD induced ectopic ptc-lacZ expression at similar levels (Fig 6C' and 6D'), SmoSDall induced ectopic en expression in more anterior cells than SmoSDGPSD (Fig 6C" and 6D"). We also examined the activity of several Smo variants with different phospho-mimetic mutations in Clone 8 (Cl8) cells through a ptc-luciferase reporter assay. As shown in Fig 6E, the levels of ectopic Smo activity correlated with the levels of Smo phosphorylation. Hence, increasing levels of Smo phosphorylation progressively increased its signaling activity.
Our previous studies revealed that phosphorylation of Smo C-tail induced a conformational change resulting in its dimerization/oligomerization, as indicated by an increased fluorescence resonance energy transfer (FRET) between C-terminally tagged CFP and yellow fluorescent protein (YFP) [12,24]. We found that mutating the CL-II site in SmoSD123 reduced its C-terminal FRET (Fig 6F), suggesting that CL-II phosphorylation may promote the active Smo conformation.
We next sought to determine the mechanism by which Hh stimulates Smo phosphorylation by Gish. Because Gish is associated with the plasma membrane, we speculated that Hh might stimulate the formation of a Smo-Gish complex at cell surface. Indeed, when expressed in S2 cells, Myc-Smo coimmunoprecipitated with endogenous Gish as well as coexpressed HA-Gish, and the amount of Gish in the Smo immunoprecipitates dramatically increased after Hh stimulation (Fig 7A and 7B). The formation of the Smo-Gish complex depends on the plasma membrane association of Gish, as Myc-Smo failed to pull down the cytosolic forms of Gish, HA-GishCS, and HA-GishΔC (Fig 7C). In addition, Myc-Smo-Ub pulled down much less Gish compared with Myc-Smo after Hh stimulation (Fig 7D), suggesting Hh-induced Smo cell surface localization of Smo is critical for its interaction with Gish.
When cotransfected into S2 cells, HA-Gish formed a complex with Myc-SmoCT (Smo C-tail) but failed to bind Myc-SmoΔC570, which lacks the C-tail (Fig 7E), suggesting that Smo interacts with Gish through its C-tail. HA-Gish was also associated with Myc-SmoΔC650 in S2 cells (Fig 7E). Furthermore, this association was enhanced upon Hh stimulation (Fig 7F), suggesting that Hh signaling facilitates the binding of Gish to the membrane proximal region of the Smo C-tail.
Finally, we determined whether Smo-Gish association is regulated by PKA-mediated phosphorylation of Smo. We found that mutating the PKA phosphorylation sites to Ala (SA123) attenuated Hh-stimulated Smo-Gish complex formation, whereas the phospho-mimetic mutations of PKA sites and adjacent CK1 sites (SD123) increased both the basal and Hh-stimulated Smo-Gish complex (Fig 7G). The influence of PKA phosphorylation of Smo on Smo-Gish interaction could explain why CL-II phosphorylation is affected by PKA (Fig 4G and 4H).
We next determine whether CK1γ regulates Shh signaling in mammalian cells. When expressed in NIH3T3 cells, an enhanced yellow fluorescent protein (EYFP)-tagged Xenopus CK1γ (EYFP-CK1γ) was accumulated on the plasma membrane as well as on the primary cilium, whereas a cytosolic form of CK1γ, EYFP-CK1γ-ΔC, which lacks the C-terminal palmitoylation site [30], failed to localize on the primary cilium (Fig 8A), suggesting that ciliary localization of CK1γ depends on its plasma-membrane association. In a Gli-luciferase (Gli-luc) reporter assay, EYFP-CK1γ but not EYFP-CK1γ-ΔC stimulated Shh pathway activity (Fig 8B). By contrast, two dominant-negative forms of CK1γ CK1γK73R and CK1γD164N, which specifically inhibited CK1γ activity [30], suppressed Shh-stimulated Gli-luc reporter activity (Fig 8C). Our previous study revealed that Shh stimulated the phosphorylation of mammalian Smo C-tail at multiple clusters of Ser/Thr residues, including a membrane-proximal cluster (S1) important for Smo activation [15]. Western blot analysis using an antibody (PS1), which recognizes phosphorylated S1 site [15], indicated that overexpression of EYFP-CK1γ but not EYFP-CK1γ-ΔC could stimulate Smo phosphorylation at the S1 site (Fig 8E). On the other hand, overexpression of either CK1γK73R or CK1γD164N attenuated Shh-stimulated S1 phosphorylation (Fig 8F). We found that EYFP-CK1γ and Myc-Smo formed a complex when coexpressed in NIH3T3 cells and that CK1γ/Smo association was enhanced upon Shh stimulation (Fig 8H). By contrast, EYFP-CK1γ-ΔC failed to interact with Smo even in the presence of Shh (Fig 8H). Because EYFP-CK1γ-ΔC failed to accumulate on the primary cilium, we speculated that the function of CK1γ in the regulation of Smo phosphorylation and Shh pathway activity might depend on the primary cilium. To disrupt the primary cilium, we overexpressed a dominant negative form of Kif3b (DN-Kif3b), a subunit of the Kinesin-II motor required for the cilium formation [15,44]. Indeed, DN-Kif3b overexpression blocked CK1γYFP/Myc-Smo association (Fig 8I), CK1γ-stimulated S1 phosphorylation (Fig 8G), and Gli-luc reporter activity (Fig 8D), suggesting that the primary cilium is required for CK1γ to bind and phosphorylate Smo.
A recent study revealed that the cell-membrane-permeable Smo agonist SAG could activate Smo without its ciliary accumulation [45]. Indeed, we found that SAG stimulated Smo phosphorylation at the S1 site in NIH3T3 cells transfected with DN-Kif3b, although Shh failed to stimulate Smo phosphorylation in these cells (Fig 8J), suggesting that SAG can induce Smo phosphorylation in the absence of the primary cilia, albeit at lower efficiency than in the presence of the primary cilia. This cilium-independent Smo phosphorylation was abolished when cells were transfected with a dominant negative form of CK1α (CK1αDN [46]); however, the dominant negative forms of CK1γ (CK1γK73R and CK1γD164N) failed to block cilium-independent Smo phosphorylation induced by SAG (Fig 8J). These results further support the notion that CK1γ regulates Smo phosphorylation in the primary cilium.
In this study, we identified a plasma membrane-associated kinase Gish/CK1γ as a conserved positive regulator of the Hh pathway. We found that Hh stimulated the binding of Gish/CK1γ to Smo to phosphorylate a Ser/Thr cluster located in the membrane-proximal region of the Smo C-tail (CL-II site in Drosophila Smo and S1 site in mammalian Smo). We demonstrated that CL-II phosphorylation is required for the optimal Smo activation and Hh pathway activity. Interestingly, we found that Gish/CK1γ regulates Smo phosphorylation and Hh pathway activity depending on its plasma membrane/ciliary localization. We propose that cell surface/ciliary accumulation of Smo, which is facilitated by Hh stimulation, promotes its association, phosphorylation, and activation by Gish/CK1γ (Fig 9).
As an obligatory transducer of the Hh signal, Smo relays the positional information imposed by Hh morphogen gradient to different levels of pathway activity that elicit distinct developmental outcomes. How Smo is differentially activated by Hh morphogen gradient is incompletely understood; however, it has been suggested that different levels of Hh induce different levels of Smo phosphorylation, which are in turn translated into different levels of Hh pathway activity [16]. Indeed, it has been observed that increasing levels of Hh resulted in a progressive increase in the overall levels of Smo phosphorylation, and that increasing the number of phospho-mimetic mutation in the Smo C-tail led to a gradual increase in Hh pathway activity [11,12,15,40]. A previous study revealed that, in cultured Drosophila cells exposed to Hh, Smo was phosphorylated on at least 26 Ser/Thr residues in its C-tail [20]. While the distal phosphorylation sites including the three PKA/CK1 phosphorylation clusters have been well characterized, the biological function of and the kinases involved in the phosphorylation of the membrane-proximal sites, most notably the CL-II site, have remained largely unexplored. In addition, it remains unclear whether phosphorylation of CL-II is constitutive or stimulated by Hh because its phosphorylation state was not determined in the absence of Hh stimulation.
In this study, we generated an antibody (Smo4P) that recognized phosphorylated CL-II and demonstrated that CL-II phosphorylation was stimulated by Hh. Moreover, we found that CL-II phosphorylation was mediated by Gish in a manner depending on its membrane association. Both gish mutation and phosphorylation deficient CL-II mutation affected the expression of only high-threshold Hh target genes such as en, suggesting that Gish-mediated CL-II phosphorylation is not absolutely required for Hh signal transduction but rather plays a modulatory function in fine-tuning Hh signaling strength. We noticed that the reduction of Hh pathway activity in gish mutant clones was less severe compared with that associated with CL-II mutation (Fig 2, S2 Fig and Fig 5). One possibility is that we did not completely overcome the perdurance issue even when gish clones were generated at 24–48 h AEL. The early-expressed gish gene products could persist even after many rounds of cell division, and residual Gish kinase activity might still exist in gish mutant cells to partially phosphorylate CL-II. Indeed, we didn't observe any change in Hh pathway target gene expression in gish mutant clones generated at 48–72 h AEL, many of which are >30 cells in size, suggesting that gish gene products generated at early time points survived many rounds of cell divisions. Another possibility is that in the absence of Gish or when Gish activity is compromised, another kinase(s)—for example, other members of CK1 family—may partially compensate for the loss of Gish, although these kinases do not phosphorylate CL-II as effectively as Gish.
A recent study suggested that CL-II phosphorylation was mediated by Gprk2 without providing direct evidence [38]. However, we found that CL-II phosphorylation was not affected by Gprk2 knockdown or overexpression (S3 Fig). In addition, a recombinant GRK failed to phosphorylate a Smo fragment containing the CL-II site in an in vitro kinase assay [24]. Instead, our previous study indicated that Gprk2 phosphorylated Smo at Ser/741/Thr742 (GPS1) and Ser1013/Ser1015 (GPS2) and that mutating these sites to nonphosphorable Ala in the SmoSD123 background attenuated the activity of this constitutively active form of Smo [24]. Here, we further demonstrated that Smo with Gprk2 site mutated (SmoGPSA) failed to activate en in smo mutant clones located near the A/P boundary (S6 Fig), suggesting that Gprk2-mediated Smo phosphorylation is also required for high-levels of Hh signaling activity. Interestingly, both Gish- and Gprk2-mediated phosphorylation was regulated by PKA (Fig 4) [24], which phosphorylated Smo at S667, S687, and S740 [19–21]. Indeed, mutating the three PKA sites to Ala (SmoSA123) diminished while converting the PKA sites and adjacent CK1 sites to phospho-mimetic residues (SmoSD123) promoted CL-II phosphorylation. Although SmoSD123 could activate both low- and high-threshold Hh target genes when expressed at high levels, expression of SmoSD123 at close to physiological levels resulted in low to intermediate levels of Hh pathway activity (Fig 6A–6A") [19]. Phospho-mimetic mutations at Gprk2 or both Gprk2 and Gish sites in the SmoSD123 background resulted in a progressive increase in its activity (Fig 6D-6D"). Hence, Hh-induced phosphorylation of Smo at PKA sites may confer low to medium levels of pathway activation, while high levels of pathway activity require further phosphorylation by Gish and Gprk2, which is "primed" by PKA-mediated phosphorylation. Consistent with CL-II being critical for high-threshold Hh target gene expression, its phosphorylation requires higher levels of Hh or longer exposure to the same levels of Hh than PKA site phosphorylation (Fig 4I and 4J), suggesting graded Hh signals may differentially regulate PKA- and Gish-mediated phosphorylation of Smo to progressively increase Smo activity (Fig 9).
How does PKA regulate CL-II phosphorylation? Our previous study revealed that Smo adopts a closed conformation in which its C-terminal region folds back and lies in close proximity to its third intracellular loop, whereas Hh-induced phosphorylation at PKA/CK1 sites promotes an open conformation [12]. It is possible that the Gish-binding pocket is masked when Smo C-tails adopt the closed conformation, whereas phosphorylation-mediated conformation switch exposes the Gish-binding pocket. Indeed, we found SmoSD123 exhibited high basal binding to Gish, whereas SmoSA123 lost Hh-stimulated binding to Gish (Fig 7G). However, Hh could further stimulate the binding of Gish to SmoSD123 and increase CL-II phosphorylation in SmoSD123 (Figs 4H and 7G), implying that Hh could regulate Smo/Gish association and CL-II phosphorylation through additional mechanism(s). We noticed that Hh increased the binding of Gish to SmoΔ650 (Fig 7F), which lacks the C-terminal region, including the three PKA sites, suggesting that Hh signaling may induce a conformation change in the transmembrane helixes of Smo, similar to those observed for GPCRs in response to agonist stimulation [47], to expose the juxtamembrane binding site for Gish.
Hh-induced phosphorylation of Smo by PKA promotes its cell surface accumulation [19], which may also contribute the elevated CL-II phosphorylation upon Hh stimulation. In this regard, it is interesting to note that Gish is associated with the plasma membrane through its C-terminal palmitoylation. Cell surface accumulation of Smo will render its close proximity with the plasma membrane-associated Gish and facilitate its binding to Gish. Indeed, Smo failed to bind cytosolic forms of Gish, GishCS and GishΔC, both of which were unable to phosphorylate the CL-II site in response to Hh stimulation. In addition, artificial conjugation of a ubiquitin moiety to Smo (Myc-Smo-Ub) prevented Hh-stimulated Smo cell surface accumulation, Smo-Gish association, and CL-II phosphorylation (Figs 4E, 4F and 7D).
Interestingly, we found that vertebrate CK1γ is localized to the primary cilium depending on its C-terminal palmitoylation. Deleting the palmitoylation site (CK1γ-ΔC) prevented its ciliary localization, its association with mammalian Smo (mSmo), and its ability to phosphorylate mSmo and promote Shh pathway activity. Furthermore, depleting primary cilia by expressing a dominant negative for of Kif3b (DN-Kif3b) also affected CK1γ Smo interaction and phosphorylation by CK1γ. Taken together, these results suggest that only ciliary localized but not the cytosolic CK1γ is capable of phosphorylating Smo to promote Shh pathway activation. As Shh induces mSmo ciliary accumulation, we propose that ciliary accumulation of mSmo may facilitate its binding to and phosphorylation by CK1γ, which contributes to optimal Smo activation (Fig 9). Because abnormal Smo activation contributes to many types of human cancer, the finding that CK1γ is a conserved positive regulator of Hh signaling raises an interesting possibility that interfering with the interaction between CK1γ and Smo may serve as a strategy for cancer treatment.
All flies were raised on standard yeast and molasses-based food at 25°C. Gal4 drivers used in this study are MS1096 and C765 [22,24,48]. UAS transgenes are: UAS-Smo-PKA12/SmoDN, UAS-Smo-PKA123/SmoSA123, and UAS-SmoSD123 [19]; UAS-Smo-CFP, UAS-SmoSD123-CFP, UAS-SmoSDGPSD-CFP, and UAS-Smo-GPS1A2A-Fg/SmoGPSA [24]; UAS-Myc-Gish (BL#41764), UAS-Myc-GishΔC (BL#41769), and UAS-Myc-GishKD (BL#41766); UAS-Flag-CK1α and UAS-Flag-CK1ε [31]; and UAS-GishRNAi (VDRC#106826, VDRC#26003, and BL#28066). Xenopus CK1γ DNA constructs are: pCS-EYFP-CK1γ, pCS-EYFP-CK1γ -ΔC, pCS-CK1γ D164N, and pCS-CK1γ K73R [30]. Mutant flies are: smo3 [10], gishKG03891/gishP (BL#13263), Gprk2Δ15 [24], Df(3R)ED10639 (BL#9481). gishΔ4 was generated by imprecise excision of the P-element from gishKG03891. GishCS, SmoCL-IISA (S626A, S627A, S633A, S634A), SmoCL-IISD (S626D, S627D, S633D, S634D), SmoSDCL-IISA, SmoSDCL-IISD, and SmoSDall were generated by PCR-based site-directed mutagenesis and confirmed by DNA sequencing. Mutant clones were generated by standard FRT/FLP mediated mitotic recombination or the MARCM system as previously described [36,49]. Mutant clones were generated using the following genotype: MARCM clones for gishΔ4 (yw hs-FLP; tub-Gal4; FRT82B tubGal80/FRT82B gishΔ4), gishΔ4 clones in Minute background (yw hs-FLP; FRT82B M(3) hs-CD2/FRT82B gishΔ4), gishP/P wings expressing SmoDN (yw MS1096 UAS-Flp; UAS-Smo-PKA12; FRT82B M(3) hs-CD2/FRT82B gishKG03891), smo clones with or without expressing smo transgenes (yw hs-FLP UAS-GFP; tubGal80 FRT40/smo3 FRT40; tub-Gal4, yw hs-FLP; tubGal80 FRT40/smo3 FRT40; C765/UAS-Smo-CFP or SmoCL-IISA-CFP, or yw hs-FLP UAS-GFP; tubGal80 FRT40/smo3 FRT40; C765/ UAS-Smo-GPSA).
Drosophila S2 cells were cultured in Drosophila SFM (Invitrogen) with 10% fetal bovine serum, 100 U/ml of penicillin, and 100 mg/ml of streptomycin at 24°C. Transfections were carried out using the Calcium Phosphate Transfection Kit (Specialty Media). Hh-conditioned medium treatment was carried out as described [50]. Unless mentioned otherwise, Hh-conditioned medium was used at a 6:4 dilution ratio by fresh medium (referred to as 100% Hh). NIH3T3 cells were cultured in DMEM (Sigma-Aldrich) containing 10% bovine calf serum (ATCC) and were transfected using the GenJet Plus In Vitro DNA Transfection Kit (SignaGen). Immunostaining and western blot analyses were carried out using standard protocols as previously described [49,51]. For immunoprecipitation assay, S2 cells were harvested and washed twice with PBS after transfection for 48 h and then lysed on ice for 30 min with lysis buffer containing 1M Tris pH8.0, 5M NaCl, 1M NaF, 0.1M Na3VO4, 1% NP-40, 10% Glycerol, and 0.5M EDTA (pH8.0). Cell lysates were incubated with protein A–Sepharose beads (Thermo scientific) for 1 h at 4 C° to eliminate non-specific binding proteins. After removal of the protein-A beads by centrifugation, the cleared lysates were incubated with Myc (HA or Flag) antibody for 2 h or overnight. The complexes were collected by incubation with protein A–Sepharose beads for 1 h at 4 C°, followed by centrifugation. The immunoprecipitates were then washed three times for 5 min each with lysis buffer and were fractionated by SDS–PAGE. FRET analysis was carried out as previously described [12]. Antibodies used for this study are rat anti-Ci, 2A1 [52], mouse anti-Ptc (DSHB), mouse anti-En (DSHB), mouse anti-SmoN (DSHB), mouse anti-HA (Santa Cruz), rabbit anti-LacZ (Affinity Bioreagents), rabbit anti-GFP (Invitrogen), and mouse anti-Flag (Sigma). Rabbit anti-Smo4P antibody was generated by Abmart (http://www.ab-mart.com) using the synthetic phosphopeptide NDLN(PS)(PS)E(PT)NDI(PS)STW-C as antigen. The reactive antibody was purified by absorption on a phosphopeptide affinity column and was further purified by subtraction on a column containing a non-phosphopeptide DLNSSETNDISSTW-C. The resulting antibody was initially characterized by western blot analysis using GST-Smo601-700 and GST-Smo601-700CL-IISA.
In vitro kinase assay was performed as previously described [15]. Briefly, GST-fusion proteins were mixed with 0.1 mM ATP, 10 mCi γ-32p-ATP, and CK1δ (New England Biolabs) and incubated at 30°C for 1.5 h in reaction buffer. Phosphorylation of GST-fusion proteins was analyzed by autoradiography after SDS-PAGE. For ptc-luc reporter assays, Cl8 cells were transfected with 1 μg ptc-luc reporter construct and 50 ng RL-PolII renilla construct in 12 well plates together with 1 μg Smo constructs. After 48 h incubation, the reporter assays were performed using the Dual-Luciferase reporter assay system (Promega). Dual-Luciferase measurements were performed in triplicate using FLUOstar OPTIMA (BMG LABTECH). Gli-luc reporter assays were carried out as previously described [15]. Double-stranded (ds) RNA was generated by MEGAscript High Yield Transcription Kit (Ambion). Gish dsRNA targeting the coding region (amino acids 196–474) was generated by PCR using the primers 5′-GAATTAATACGACTCACTATA GGGAG AGGCAGAAC GTCAACAAAACGT-3 and 5′-GAATTAATACGACTCACTATAGGGA GATTTTTGGCGCGTCGATTTCTT-3′. Gish dsRNA targeting the 5' UTR was generated by PCR using the primers 5'-GAATTAATACGACTCACTATAGGGAGAAAAGTGTGTTTGTCAAATTGT-3' and 5'-GAATTAATACGACTCACTATAGGGAG ACTCACCGCCCACACTCACACG-3 '. Gprk2 dsRNA was generated by PCR using DNA template targeting Gprk2 amino acids 124–290 as described [24]. For the RNAi knockdown experiments, S2 cells were cultured in serum-free medium containing the indicated dsRNA for 8 h at 24°C. After adding FBS to a final concentration of 10%, dsRNA-treated cells were cultured for 24 h before transfection. Forty-eight hours after transfection, the cells were collected for analyses.
|
10.1371/journal.pgen.1002337 | Successive Increases in the Resistance of Drosophila to Viral Infection through a Transposon Insertion Followed by a Duplication | To understand the molecular basis of how hosts evolve resistance to their parasites, we have investigated the genes that cause variation in the susceptibility of Drosophila melanogaster to viral infection. Using a host-specific pathogen of D. melanogaster called the sigma virus (Rhabdoviridae), we mapped a major-effect polymorphism to a region containing two paralogous genes called CHKov1 and CHKov2. In a panel of inbred fly lines, we found that a transposable element insertion in the protein coding sequence of CHKov1 is associated with increased resistance to infection. Previous research has shown that this insertion results in a truncated messenger RNA that encodes a far shorter protein than the susceptible allele. This resistant allele has rapidly increased in frequency under directional selection and is now the commonest form of the gene in natural populations. Using genetic mapping and site-specific recombination, we identified a third genotype with considerably greater resistance that is currently rare in the wild. In these flies there have been two duplications, resulting in three copies of both the truncated allele of CHKov1 and CHKov2 (one of which is also truncated). Remarkably, the truncated allele of CHKov1 has previously been found to confer resistance to organophosphate insecticides. As estimates of the age of this allele predate the use of insecticides, it is likely that this allele initially functioned as a defence against viruses and fortuitously “pre-adapted” flies to insecticides. These results demonstrate that strong selection by parasites for increased host resistance can result in major genetic changes and rapid shifts in allele frequencies; and, contrary to the prevailing view that resistance to pathogens can be a costly trait to evolve, the pleiotropic effects of these changes can have unexpected benefits.
| Though much is known about host–parasite coevolution in plants, relatively little is understood in animals. Most studies using animal systems have focused on either generalist parasites or those that do not naturally occur in the host. The sigma virus is specific to Drosophila melanogaster, which provides the unique opportunity to study natural coevolution in a well-established model organism. In order to gain a better understanding of host–parasite coevolution, we have set out to identify novel viral resistance genes using the sigma-Drosophila system. Here we identify two successive mutations that provide increasing resistance to the sigma virus. The first of these, a transposable element insertion within a gene called CHKov1, is already known to provide resistance to insecticides. There is evidence that the novel gene product resulting from this insertion has been under positive selection pressure long before the use of pesticides. Two duplications of this gene region have resulted in further resistance to sigma virus. We believe that selection for resistance to the sigma virus led to the added benefit of resistance to insecticides.
| The presence of a parasite elicits strong selection pressures for the host to evolve increased resistance and the parasite to overcome host defences. This can drive rapid changes in allele frequencies in both organisms and result in “Red Queen” evolution, where both species must constantly evolve just to maintain a fitness status quo [1]. Generation times, population sizes, mutation rates and migration rates all affect the evolutionary potential of hosts and parasites, and these factors mean that in many cases the parasite will be evolving faster than the host [2]. Therefore the host is under constant selection to evolve new forms of resistance to the parasite, and this makes host resistance an excellent model to study the evolution of adaptation.
Identifying the genes underlying the evolution of resistance can provide insights into this process, revealing the types of mutation involved, the nature of selection acting on resistance, and the molecular mechanisms involved in evolving resistance to infection. A substantial amount of work has been done to study the genetics of host-parasite co-evolution in plants, and we have a broad knowledge of plant resistance (R) gene genetics [3]. Unfortunately this is not true for the animal kingdom, especially invertebrates. Aside from a handful of studies on disease vectors, much of the work on invertebrates tends to be purely phenotypic or has not been done with naturally co-evolving systems. Identifying the genes causing variation in the resistance of invertebrates to viruses will allow us to get at many of the mechanisms underlying the evolution of resistance and provide insights to the nature of co-evolution.
The antiviral immune defences of Drosophila have been the target of much research in recent years, with RNAi, autophagy and other pathways proving to be important [4]–[7]. However, on an evolutionary timescale, changes to the immune system are not the only way in which hosts can defend themselves against viruses. Several insects, including Drosophila melanogaster, have developed a symbiosis with the bacterium Wolbachia that provides resistance to a range of RNA viruses [8]–[11]. Viruses also rely on the host cellular machinery for all stages of their replication cycle, and changes to these host factors may also lead to the evolution of resistance, for example by blocking entry in host cells [12].
The discovery of genes causing variation in resistance can also allow us to infer the selection pressures acting on host alleles during co-evolution [3], [13]. Co-evolution can result in two main forms of selection: new resistance alleles may continually arise by mutation and be fixed by directional selection, or negative frequency-dependent selection can maintain polymorphisms of resistant and susceptible alleles [1], [13]–[14]. To complicate matters, selection pressures on host alleles can be very dynamic, not only depending on allele frequencies in the parasite [15], but also on changing environmental conditions. It is also of interest to understand the genetic architecture of resistance and the nature of the mutations involved. For example, is the resistance level primarily controlled by alleles of small or large effect, and is it the result of regulatory or coding changes or both? By addressing all of these questions, the identification of host genes experiencing strong selection will therefore help to develop better models of co-evolution.
We have investigated the genetics of resistance to the sigma virus, the only naturally occurring host-specific parasite known in D. melanogaster [16]–[17]. Host specificity is important, as when a parasite infects a single host species there is particularly strong selection for reciprocal adaptation, and such “tight” co-evolution simplifies the arduous task of understanding how co-evolution operates. The sigma virus is a member of the rhabdovirus family, and has a negative-sense RNA genome [18]. It is only transmitted vertically from parent to offspring [18]. In this study we have investigated a resistance gene called ref(3)D, which had previously been mapped between two visible markers on the right arm of the 3rd chromosome [19].
The two fly lines that we began our experiments with differed dramatically in their resistance to the sigma virus —11 days after injection less than 5% of the flies from the resistant OOP line showed the symptom of being paralysed by CO2, compared to over 95% of flies from the susceptible 22a line (Figure 1). Previous work has mapped a gene called ref(3)D, which affects sigma virus replication, to the third chromosome of this fly stock [19]. However, it also contains a gene with an allelic variant that reduces transmission of the sigma virus through sperm, so we first removed this allele to avoid complications in identifying ref(3)D. This was accomplished by crossing OOP and 22a to generate a line with a recombination event between the suspected locations of each gene (92–94 cM region). The resulting line was homozygous for both the resistant allele of ref(3)D and the allele of the other gene that results in high rates of transmission through sperm, and this was used in subsequent experiments.
To map ref(3)D, we produced lines that carried a homozygous third chromosome that was a recombinant between the resistant and susceptible stocks. We used molecular markers to screen 191 recombinant flies to identify those that had recombined in a 12 cM interval believed to contain the gene, and created 21 homozygous recombinant lines in this anticipated region. These lines were injected with the sigma virus and genotyped with molecular markers across the region (Figure 2A). There was a clearly bimodal distribution of infection rates, with some lines being highly resistant and others highly susceptible. Furthermore, there was a perfect association between infection rates and genotype across a 182 kb region (Figure 2A; Wilcoxon Rank Sum Test: W = 110, P = 1.2×10−4).
This process was repeated to generate recombinants in the 2 cM interval that contains the resistance gene. This time we screened 1920 flies for informative recombinants and 32 new homozygous recombinant lines were generated in this new region. Again, after injecting the virus these could be clearly categorized into resistant and susceptible lines. After genotyping the lines, this experiment reduced the region where there is a perfect association between genotype and phenotype to 60 kb (Figure 2B; Wilcoxon Rank Sum Test: W = 256, P = 1.5×10−6).
To select for recombinants in this smaller region we used phenotypic markers rather than molecular markers. We combined two P-elements carrying eye-color markers to produce a susceptible mapping stock (2GT1), crossed this to a resistant fly line, and selected recombinants that carried just one of the two markers. Using this approach we generated 10 lines that were homozygous for the recombinant chromosome. As before, these lines were assayed for resistance to the sigma virus and genotyped for several markers across the 60 kb candidate region. This reduced the region that could contain the gene to 36 kb (Figure 2C; Wilcoxon Rank Sum Test: W = 16, P = 0.04).
To map the gene within in this region, we induced site-specific recombination in males using P-elements. In this experiment we crossed transposable element lines that were susceptible to the sigma virus (data not shown) to a resistant line, and induced recombination at the location of the P-element. We successfully produced four recombinants that were viable as homozygotes. To control for the effects of genetic background, lines that lacked a recombination event were also generated using the same crossing scheme, so they either had the susceptible chromosome containing the transposable element or the resistant chromosome. To check that recombination had occurred, we scored molecular markers flanking the transposable element positions in each line. We injected the recombinant lines and respective controls with the sigma virus (Figure 3), and found that there was a striking difference between the resistance of recombinants between two sites located just 3089 bases apart in the published genome (3R:21155073..21158162, D. melanogaster genome version 5.31.). This region contains all of CHKov2 plus the 3′ end of CG10669 in the published genome sequence (part of the fifth exon, all of the sixth exon and the 3′UTR).
To identify the polymorphisms that could be causing resistance, we sequenced the region around CHKov2 and found that there had been a complex rearrangement in the resistant line (Figure 4C, highly resistant line). The susceptible line had a gene order that is the same as the published Drosophila genome (Figure 4B, note that this is described as ‘resistant’ in the figure as a more susceptible allele is described below). As is the case in the published genome sequence, in both our resistant and susceptible lines a naturally occurring Doc transposable element has inserted into the protein coding region of CHKov1, which is a paralog and neighbour of CHKov2. Previous research has shown that this insertion results in two short transcripts being produced, which are predicted to encode truncated proteins [20]. However, in the resistant line there are two duplications, both of which involve partial sequences of both CHKov1 and CHKov2. The first duplication includes a large portion of the 5′ end of CHKov1 (including some upstream intergenic sequence) and approximately two-thirds of the 3′ end of CHKov2. The second duplication is in the reverse orientation, and includes all of CHKov2 and the 5′ end of CHKov1 (compared to the first duplication, this includes less of the Doc element insertion, exactly the same protein coding region and an identical region of the upstream intergenic sequence). It is highly likely that this rearrangement is causing the difference in resistance, as in the region mapped by male recombination there is only one single nucleotide polymorphism (SNP) outside of the rearrangement that differs between the resistant and susceptible lines.
This rearrangement could confer resistance to viruses either by altering the expression of the genes involved, or due to coding changes (the only coding sequence which is altered is the truncation of one of the duplicates of CHKov2). We therefore used quantitative rtPCR to examine whether the expression of CHKov1 or CHKov2 is different in the resistant and susceptible flies. It has previously been shown that neither of the CHKov genes change expression after injection with sigma [21], so any novel changes in expression could be attributed to the rearrangement. Six days after injection with the sigma virus CHKov2 expression was 5.6-fold greater in the resistant lines than the susceptible lines (Wilcoxon Rank Sum Test: W = 86, P = 0.0001) and 12 days after injection it was 9.6-fold greater (Wilcoxon Rank Sum Test: W = 88, P = 2.6×10−5). In contrast there was no evidence for a change in the expression of CHKov1, despite this gene being amplified to three copies in the resistant line (1.9 fold greater expression in susceptible lines on day 6, Wilcoxon Rank Sum Test: W = 24, P = 0.11; 1.4 fold greater expression in susceptible lines on day 12, Wilcoxon Rank Sum Test: W = 29, P = 0.24).
In the experiments above we have used a symptom of infection — paralysis on exposure to CO2 — to test if flies are infected. To check whether the resistance gene is reducing viral titres rather than simply altering CO2 sensitivity itself, we used quantitative PCR to estimate the relative copy number of the viral genome in resistant and susceptible flies. Using the same samples that we used to examine gene expression, we found that there was an approximately 79–fold decrease in sigma virus load in resistant lines 6 days after the virus was injected (Wilcoxon Rank Sum Test: W = 0, P = 3×10−5) and a 138–fold decrease after 12 days (Wilcoxon Rank Sum Test: W = 0, P = 3×10−5).
As the rearrangement of the CHKov1 and CHKov2 genes that confers resistance to the sigma virus was originally found in a natural population in Europe, we examined its frequency in nature. To do this we used Freeze 1 of the Drosophila Genetic Reference Panel (DGRP) (http://www.hgsc.bcm.tmc.edu/project-species-i-Drosophila_genRefPanel.hgsc), which is a set of highly inbred North American fly lines whose genomes have been sequenced. As the genome sequences were produced from short-read data, rearrangements and transposable element insertions are not reliably assembled. We therefore used PCR to genotype all the lines for both the Doc element in CHKov1 and the complex rearrangement. The Doc insertion was present in most of the lines (155 were homozygous for the insertion, 29 were homozygous without it, and 8 were heterozygous, likely due to insufficient inbreeding.), but the rearrangement was not found in any of the 192 lines tested. Therefore this rearrangement is not an important cause of virus resistance in this population.
As the truncated version of CHKov1 has been duplicated in the most resistant allele (Figure 4C), we tested whether the Doc element insertion in CHKov1 was itself associated with resistance. We injected 11870 flies from 186 of the DGRP lines with the sigma virus and tested them for infection with the CO2 assay 13 days later. We found that the insertion is associated with a highly significant drop in infection rates (Bayesian generalised linear mixed model: P<0.001). Using this statistical model, we estimate that the Doc insertion is associated with a 52% drop in infection rates from 82% to 30% (95% C.I. on drop: 42%–64%). It should be noted that the susceptible line used in the mapping experiment above contains the Doc insertion. Therefore the three alleles in this region shown in Figure 4 have a hierarchy of resistance, with the ‘rearranged’ allele being most resistant (Figure 4C) and the Doc insertion having intermediate resistance (Figure 4B). The sequence in Drosophila simulans has neither the Doc insertion nor the rearrangement, indicating that the most susceptible allele is the ancestral state (Figure 4A), the allele of intermediate resistance arose next following the Doc insertion, and then a rearrangement occurred that lead to a further increase in resistance.
To examine whether any other polymorphisms in this region are associated with resistance we used the data from the genome sequences of the DGRP lines. Using 150 lines whose genomes have been sequenced we examined the 60 kb region which we mapped in our first set of experiments (Figure 2B). In the regions flanking CHKov1 we found that 32 of 468 SNPs in the region were significantly associated with resistance to the sigma virus after Bonferroni correction (Figure 5A). However, there is extensive linkage disequilibrium between the Doc insertion and surrounding sites (see below; [20]), so all of these associations could all be caused the same polymorphism. We therefore repeated the analysis, but this time included the presence or absence of the Doc insertion in the model. We found that none of the associations were significant (Figure 5B), so the most parsimonious interpretation is that a single polymorphism in this region is causing resistance. As the Doc insertion has such a dramatic effect on the protein encoded by CHKov1, this is most likely to be the cause of resistance.
The mapping data together with these association studies therefore provide strong evidence that there are two different polymorphisms in this region that make flies resistant involving the Doc insertion and its subsequent duplication. However, we still wished to confirm that none of the other SNPs associated with resistance in the DGRP lines could contribute to the difference between the resistant and susceptible lines used in the mapping experiments. We therefore sequenced this entire 60 kb region from both the resistant and susceptible lines (OOP and 22a), and identified 191 SNPs and 11 indels that differed between these lines and were present in the DGRP genomes. None of these polymorphisms were significantly associated with resistance to sigma (after corrections for multiple testing; Figure 5B), and only two of them fell within a 30 kB region around the duplication implicated in resistance. This confirms that different genetic changes are affecting resistance in the DGRP lines and causing the difference between the two lines we used in the mapping experiments.
Previous studies have examined the pattern of genetic variation around the Doc insertion in CHKov1 [20], but the sequences of all 192 DGRP lines provides us with a more complete dataset. We found that there is extensive linkage disequilibrium between the Doc insertion and surrounding sites that extends at least 25 kB to the 3′ end of the gene and a much shorter distance in the 5′ direction (Figure 6). In the region where sites are in linkage disequilibrium with the Doc insertion, there is greater genetic variation among the susceptible chromosomes than the resistant chromosomes (Figure 6), despite the resistant allele being most common. These observations are consistent with the conclusion of Aminetzach et al [20] that the Doc insertion has recently increased in frequency under directional selection.
We have found that two events have led to successive increases in resistance to the sigma virus (Figure 4). The first of these is a Doc transposable element insertion into the coding sequence of CHKov1. The second is a complex rearrangement that results in two duplications of CHKov1 and the Doc element, further increasing resistance to sigma. As infection with the sigma virus reduces the fitness of infected flies [22], it is likely that selection for resistance to this common pathogen has led to the major structural changes in this gene and large shifts in resistance to the sigma virus. The first of these events, involving the insertion of the Doc element, caused the infection rate in our experiments to drop from 82% in flies with the susceptible allele to 30% in flies with the insertion. Transposable element insertions are known to be important in causing a number of major-effect mutations that are important in adaptations such as insecticide resistance [23]–[24]. In contrast to most of these changes, which tend to affect the regulatory regions upstream of genes [23], this Doc element has inserted into an exon and is expected to cause major changes to the structure of the protein. In its ancestral state, CHKov1 is comprised of four exons that produce a single transcript. Previous research has shown that, by interrupting the original transcript, this Doc insertion results in two derived transcripts being produced, each of which contains both Doc element sequence and CHKov1 sequence [20]. Assuming these transcripts are translated, this is likely to result in the protein losing its original enzymatic function, as neither of the new transcripts include the two protein domains encoded by the original transcript (a choline kinase domain and the PFAM domain DUF227) [20].
The second event to occur was a complex rearrangement of this region, which resulted in an even greater increase in resistance to the sigma virus than the original Doc insertion. The rearrangement leaves the fly with two full copies of CHKov2, a partial copy of CHKov2, and three full copies of the first derived transcript of CHKov1 caused by insertion of the Doc element (Figure 4). The simplest explanation of how this rearrangement increases resistance is that the amplification of the region coding for the first derived transcript of CHKov1 increases the expression of this new gene, and this in turn increases resistance. However, we were unable to find any evidence for the expression of CHKov1 changing, suggesting that this is not the case. Furthermore, the coding region of CHKov1 is unaffected by the rearrangement. However, the rearrangement is associated with a 6- to 9-fold increase in expression of CHKov2, suggesting that this may be the cause of resistance. CHKov2 is a paralog of CHKov1 which also has a predicted choline kinase activity [20], so it is possible that the two genes could both have antiviral effects through a similar mechanism.
These complex, sequential modifications to the CHKov1 region are similar to a series of alleles of the gene Cyp6g1 which increase resistance to the pesticide DDT [25]. In the case of Cyp6g1, successive increases in resistance to DDT were caused by the insertion of an Accord transposable element into the promoter followed by a gene duplication event and the insertion of an HMS-Beagle transposable element and a partial P-element [25]. Together with our results, this suggests that both transposable element insertions and gene-duplications can be important sources of major-effect mutations that contribute to phenotypic evolution.
It is well known that genes that increase resistance to pathogens often have pleiotropic effects on other components of fitness. For example, in Drosophila, selection for increased resistance to parasitoid wasps results in a decrease in competitive ability [26] and flies that are resistant to bacteria have reduced fecundity [27]. As these pleiotropic effects tend to be harmful, it is commonly thought that resistance to pathogens is a costly trait to evolve, and these costs are assumed in many theoretical models of coevolution [1]. However, previous research has found that the Doc element insertion in CHKov1 increases resistance to organophosphate insecticides [20]. Therefore, contrary to received wisdom, this pleiotropic effect of this antiviral resistance allele would appear to be beneficial to the fly.
Although CHKov1 is involved in pesticide and viral resistance, the molecular basis of these effects are not clear. Neither CHKov1 nor CHKov2 appear to be part of an induced response to the sigma virus, as they are not upregulated in infected flies [21]. It has been suggested that CHKov1, which contains a choline kinase domain, might make flies resistant to organophosphates by affecting choline metabolism in general or the target of organophosphate insecticides, acetylcholine esterase [20]. If this is the case, it is possible that it could be linked to the mechanism of virus resistance as Rhabdoviruses use acetylcholine receptors to enter cells [28].
Did the Doc insertion initially function as a defence against viruses or insecticides? Previous work has shown that there is extensive linkage disequilibrium between the Doc insertion and surrounding sites [20]. These observations, which we confirmed using a much larger dataset, provide compelling evidence for a partial selective sweep in which the Doc insertion has very recently increased in frequency. However, the number of sequence changes that have accumulated in the Doc element suggest that the insertion occurred approximately 90,000 years ago, which long predates the use of insecticides [20]. The most recent common ancestor of present-day sigma virus isolates existed roughly 2,000 years ago [17], and the infection may have been present in fly populations for much longer than this. Therefore, the Doc element would initially have only played a role in defending flies against viral infection, but these flies found themselves with an unexpected advantage once organophosphate insecticides were introduced.
The duplication of this region that resulted in the allele with the highest level of virus resistance has occurred very recently. There are only 2 sequence differences between our mapping lines in the 30 kB region surrounding the duplication, compared with over 550 polymorphisms among the DGRP lines. For this reason it is unsurprising that this highly resistant allele is still rare in the wild (although we have not tested flies from the population where this allele was first found). It is possible that given sufficient time this allele may replace the partially resistant allele that dominates today's populations.
Taken together, our results show that successive changes to the same genomic region have caused large shifts in the resistance of flies to the sigma virus. These mutations have all resulted in substantial structural changes to the genes involved, and the first of them has swept through populations under directional selection. This has not only increased the resistance of flies to viral infection, but it may also have pre-adapted flies to the introduction of insecticides in the middle of the last century.
A susceptible (22a) and resistant (OOP) fly line was provided by Didier Contamine. The third chromosome of OOP is derived from the Paris line [19] and carries both the resistant allele of the ref(3)D gene and an allele of a gene called ref(3)V which reduces the transmission of the virus through sperm [19]. The remaining chromosomes of OOP are from the susceptible Oregon R lab stock. Before attempting to map ref(3)D we first separated it from ref(3)V by crossing OOP and 22a. The F1 progeny were then crossed to TM6B, Tb/Sb, and the resulting TM6B,Tb/+ male progeny back-crossed to the balancer stock. These flies were then genotyped using molecular markers located at 92 cM and 94 cM on the standard genetic map. As these markers lie between ref(3)D and ref(3)V, this allowed us to identify a recombinant that carried the resistant allele of ref(3)D but not ref(3)V.
To map ref(3)D we created stocks that carried homozygous chromosomes that were recombinants between the resistant and susceptible chromosomes. We crossed the resistant stock to 22a, and crossed the F1 progeny to TM6B,Tb/Sb. Single male TM6B,Tb/+ progeny were then crossed back to the balancer. A few days after setting up this cross the males were removed from the tube and genotyped using molecular markers at 80 cM and 92 cM, which flank the region thought to contain ref(3)D [19]. This allowed us to retain just the 21 genotypes that had recombined in this region. In the next generation we crossed sibling TM6B,Tb/+ flies, and then selected for homozygous recombinants in the subsequent generation. Once we had mapped the gene to a smaller region (see below), we then repeated the experiment using different molecular markers to produce another 33 recombinants between 86 cM and 88 cM.
To select recombinants in even smaller regions we used phenotypic markers flanking the region of interest rather than molecular markers. First, we selected two lines, w1118; P {GT1}BG02256 and w1118;P {GT1}jigr1BG00794, which carry P-elements flanking the region of interest. These elements both carried the mini-white gene, and flies that carry a single heterozygous element have lighter colored eyes than flies carrying two heterozygous elements [29]. This allowed us to cross them and select a 3rd chromosome mapping line that carries both elements (2GT1). This was then crossed to a resistant 3rd chromosome recombinant line (D2-6) generated in the experiment described above. Recombinants between 2GT1 and D2-6 were then generated as in the previous experiment, except that this time the 3rd chromosome recombinant lines were balanced with w−;TM3,Sb/H and recombinants were detected from their eye color. Ten homozygous 3rd chromosome recombinant lines were generated along with controls with either no recombination event or a recombination event outside the region of interest.
To generate recombinants at defined sites in the vicinity of the resistance gene we used P-element-induced male recombination [30]. Four different lines with transposable element insertions (P-elements) (Text S1) were used with a resistant line to generate recombinants via male induced recombination. The crossing scheme was kindly provided by Kevin Cook (Text S1) and w−;TM3,Sb/H was used to balance the lines. Non-recombinant lines with either the 3rd chromosome derived from the susceptible P-element line or the resistant parental stock used in this cross (see Text S1) were generated as controls. All four transposable insertion lines contained the second allelic variant (Figure 4B) of CHKov1 (data not shown).
DNA was extracted using either a protocol using Chelex resin (Sigma-Aldrich, St Louis) [31] or a Tissue Genomic DNA Kit (Metabion, Munich). Genotyping was done using microsatellites, indels, SNP specific primers or via sequencing (Table S1). To score length differences in indels and microsatellites, short PCR products were run on 2% agarose gels, while larger products were run on 1% agarose gels. PCR products for sequencing were cleaned up by incubating with the enzyme Exonuclease I and Shrimp Alkaline Phosphotase at 37°C for 1 hr, followed by a 15 min incubation at 72°C to deactivate the enzymes. The sequencing reaction consisted of 25 cycles of 95°C (30 sec), 50°C (20 sec) and 60°C (4 min) using BigDye reagents (ABI). Sequencing was carried out at either Source BioScience LifeSciences (Cambridge) or The GenePool (Edinburgh).
The Hap23 strain of the sigma virus [32] was extracted from an infected line of D. melanogaster (Om), and this extract was used in all assays except one. One hundred flies were ground in 1 ml of Ringer's solution, centrifuged at 13000 rpm for 30 seconds, and the supernatants from several replicate tubes mixed together. The viral extract was then separated into small aliquots and stored at −80°C. When this ran low, the same procedure was followed, this time using susceptible flies two weeks after they were injected with the previous stock of sigma virus. This new stock was tested on susceptible and resistant lines and then used in the 3rd chromosome 2GT1 experiment.
Female D. melanogaster were injected in the abdomen with sigma virus until slight extension of the proboscis was observed. They were then maintained on either Lewis media or apple juice-agar media. Flies were tipped onto new media two days after injection and then two more times before they were tested for infection. The flies were then exposed to 100% carbon dioxide for 15 minutes at 12°C on day 10 after injection (the first two recombinant assays) or day 14 (all subsequent assays). Flies were given 2 hours to recover from the carbon dioxide and then the number of dead or paralyzed individuals was counted as well as the total number of individuals in each vial. Four replicate vials each containing approximately 15 flies on average were used in each experiment except for the first recombinant assay with the third chromosome line (three replicates).
DNA for sequencing was extracted using the kit described above. The majority of the 59.6 kb region on chromosome 3 that we had identified by mapping using recombinant lines (3R:21126075..21185688; release 5.31 of the Drosophila genome) was sequenced from both OOP and 22a (GenBank accession numbers JN247668–JN247669). Primer pairs were designed to amplify these regions in overlapping fragments (Table S1), and the sequencing was performed as described above. The sequencing of a small region involving the genes CHKov1 and CHKov2 was made more difficult by a complex rearrangement in which certain sequences had been duplicated. This region was therefore sequenced by designing PCR primers that amplified just single copies of the duplicated region.
Diagnostic PCR primers were designed to genotype flies for a Doc element insertion in CHKov1 and a complex rearrangement involving CHKov2. The forward primer CHK2-8F (5′ GCAGCACGATCGTCAAATAG 3′) and the reverse primer CHK2-8R (5′ AATGCTTCAAAGGTTTTGTTGA 3′) were used to detect the absence of the insert near CHKov2. The forward primer CHK2-7F (5′ TCTTCTCATCTTCCGGGACT 3′) and the reverse primer FlipR (5′ GTAGTTACTGGACCACAAGTTGAAG 3′) were used to identify the presence of the 5′ end of the insertion near CHKov2. The forward primer CHK_F (5′ CTCTTGGCTCCAAACGTGAC 3′) and reverse primer CHK_R (5′ AAGGCAAACGACGCTCTT 3′) were used to detect the absence of the Doc1420 element in CHKov1. The forward primer Doc1420_F (5′ CTTGTTCACATTGTCGCTGAG 3′) was used with the reverse primer CHK_R to detect the presence of the Doc1420 element in CHKov1. The genotype of another resistance gene, ref(2)P, was scored using the PCR test described in [33].
To examine the expression of candidate resistance genes and estimate viral titers we used quantitative rtPCR. Four biological replicates of 8 resistant and 11 susceptible recombinant lines were injected with sigma virus, and RNA was extracted from two of the replicates after 6 days and the other two replicates after 12 days (1 resistant line missing second 12-day replicate). From each biological replicate we extracted RNA from 10 individuals using Trizol (Invitrogen) following the manufacturer's instructions. RNA was reverse transcribed into cDNA using MMLV (Invitrogen) and random hexamer primers. Viral load was determined using quantitative PCR using SYBR Green and the forward primer DmelSV_F1 (5′ TTCAATTTTGTACGCGGAATC 3′) and reverse primer DmelSV_R1 (5′ TGATCAAACCGCTAGCTTCA 3′), which amplify a region of the viral genome spanning the L gene and 5′ trailer (and therefore amplify genomic RNA but not mRNA). Expression of CHKov1 was measured using the forward primer CHKoV1-qPCR-F1 (5′ GAACTCCGTGGGATCGACTA 3′) and reverse primer CHKoV1-qPCR-R2 (5′ CATGGGACAGGTGTTTGTCA 3′). These primers span the first intron of the gene, and amplify a region of the gene that is present in the truncated form of the gene (described below). Expression of CHKov2 was measured using the forward primer CHK2_3F (5′ CACCAAAAATCTCCGTGGTT 3′) and reverse primer qPCR_Chkov2_3_R (5′ TCGTTCTCATAAGCGACTATACATC 3′). Expression of Actin 5C was used as a control in all assays using the primers qActin5c_for2 (5′ GAGCGCGGTTACTCTTTCAC 3′) and qActin5c_rev2 (5′ AAGCCTCCATTCCCAAGAAC 3′). We performed three technical replicates of each PCR and used the mean of these in subsequent analyses.
To test which naturally-occurring polymorphisms are associated with resistance we used the Drosophila Genetic Reference Panel, which is a panel of highly inbred fly lines from North America whose genomes have been sequenced (http://www.hgsc.bcm.tmc.edu/project-species-i-Drosophila_genRefPanel.hgsc). To measure the resistance of these lines, we injected 186 of the lines with the virus and tested them for infection 13 days later. In total we tested 11870 flies for infection, and on average 4 different replicate vials of each line containing an average of 16 flies were tested. As far as was possible, each replicate vial of each line was injected on a different day and on each day we used different combinations of lines.
R version 2.11.1 was used for statistical analyses. Our data from the infection experiments consists of numbers of infected and uninfected flies, which we treat as a binomial response in a generalized linear mixed model. The parameters of the model were estimated using the R library MCMCglmm [34], which uses Bayesian Markov chain Monte Carlo (MCMC) techniques. To test for an association between Doc1420 status and resistance to sigma virus we used the model:Where νi,j is the probability of flies in vial i from line j being infected. β is a vector of the fixed effects of ref(2)P genotype and Doc1420 genotype, and XiT is a row vector relating the fixed effects to vial i. αj is a random effect of line j. The residual, εi,j, includes over-dispersion due to unaccounted for heterogeneity between vials in the probability of infection. The estimated effect of Doc1420 on infection rates was back-transformed from logits into a proportion, and the number quoted in the text is based on estimates for lines that have the susceptible allele of ref(2)P. The 95% highest posterior density of the MCMC sample was used as an estimate of the credible intervals (C.I.) of parameters.
This Bayesian approach is computationally intensive and slow to implement, so when testing larger numbers of SNPs from the DGRP dataset for effects on resistance we used a maximum likelihood method. The model was essentially the same as that described above except the SNP in question was included as a fixed effect (and Doc1420 status was not always included). The model was fitted using the R function lmer, and the significance of the fixed effects was assessed using the Wald statistic. When sample sizes are small this can give anti-conservative results [35], but this should not be important in our analysis as common SNPs were found to be highly significant (see below).
For each fly line in which we measured viral titres or gene expression by quantitative RT-PCR, we first calculated ΔCt as the difference between the cycle thresholds of the gene of interest and the endogenous control (actin 5C). The viral titre or gene expression in resistant flies relative to susceptible flies was calculated as 2−ΔΔCt, where ΔΔCt = ΔCtresistant−ΔCtsusceptible, where ΔCtresistant and ΔCtsusceptible are the means of the ΔCt values of the resistant and susceptible lines. To assess whether these differences were statistically significant, we used a Wilcoxon Rank Sum Test to compare ΔCt in the resistant lines and the susceptible lines. This calculation assumes that the PCR reactions are 100% efficient. To check whether this assumption is realistic we used a dilution series to calculate the PCR efficiency. Using this approach we found that the actin PCR is 103% efficient, the virus PCR is 101.5% efficient, the CHKov1 PCR is 100.0% efficient and the CHKov2 PCR is 102.5% efficient.
|
10.1371/journal.pcbi.1003819 | Phase Transitions in the Multi-cellular Regulatory Behavior of Pancreatic Islet Excitability | The pancreatic islets of Langerhans are multicellular micro-organs integral to maintaining glucose homeostasis through secretion of the hormone insulin. β-cells within the islet exist as a highly coupled electrical network which coordinates electrical activity and insulin release at high glucose, but leads to global suppression at basal glucose. Despite its importance, how network dynamics generate this emergent binary on/off behavior remains to be elucidated. Previous work has suggested that a small threshold of quiescent cells is able to suppress the entire network. By modeling the islet as a Boolean network, we predicted a phase-transition between globally active and inactive states would emerge near this threshold number of cells, indicative of critical behavior. This was tested using islets with an inducible-expression mutation which renders defined numbers of cells electrically inactive, together with pharmacological modulation of electrical activity. This was combined with real-time imaging of intracellular free-calcium activity [Ca2+]i and measurement of physiological parameters in mice. As the number of inexcitable cells was increased beyond ∼15%, a phase-transition in islet activity occurred, switching from globally active wild-type behavior to global quiescence. This phase-transition was also seen in insulin secretion and blood glucose, indicating physiological impact. This behavior was reproduced in a multicellular dynamical model suggesting critical behavior in the islet may obey general properties of coupled heterogeneous networks. This study represents the first detailed explanation for how the islet facilitates inhibitory activity in spite of a heterogeneous cell population, as well as the role this plays in diabetes and its reversal. We further explain how islets utilize this critical behavior to leverage cellular heterogeneity and coordinate a robust insulin response with high dynamic range. These findings also give new insight into emergent multicellular dynamics in general which are applicable to many coupled physiological systems, specifically where inhibitory dynamics result from coupled networks.
| As science has successfully broken down the elements of many biological systems, the network dynamics of large-scale cellular interactions has emerged as a new frontier. One way to understand how dynamical elements within large networks behave collectively is via mathematical modeling. Diabetes, which is of increasing international concern, is commonly caused by a deterioration of these complex dynamics in a highly coupled micro-organ called the islet of Langerhans. Therefore, if we are to understand diabetes and how to treat it, we must understand how coupling affects ensemble dynamics. While the role of network connectivity in islet excitation under stimulatory conditions has been well studied, how connectivity also suppresses activity under fasting conditions remains to be elucidated. Here we use two network models of islet connectivity to investigate this process. Using genetically altered islets and pharmacological treatments, we show how suppression of islet activity is solely dependent on a threshold number of inactive cells. We found that the islet exhibits critical behavior in the threshold region, rapidly transitioning from global activity to inactivity. We therefore propose how the islet and multicellular systems in general can generate a robust stimulated response from a heterogeneous cell population.
| Most biological systems exist as dynamic multicellular structures where distinct functionalities are generated through cellular interactions. While important for proper function, the complexity in network architecture, cellular dynamics, as well as the presence of heterogeneity, noise and biological variability make the overall function of multicellular structures difficult to understand. Approaches to understanding coupled dynamical systems have handled this complexity by explaining system structure and function individually [1], [2]. These two aspects are both of central importance when it comes to understanding the way living systems are organized and how their anatomy supports their function. Therefore, by employing network theory to inform or predict the architectural aspects of dynamical system models, we can better understand how structural properties can impact functional behaviors. One living system exhibiting complex multicellular dynamics, yet with a scale tractable for study with these approaches, is the islet of Langerhans where dysfunction generally leads to diabetes. As such the islet provides a physiologically relevant system in which we can examine properties of multicellular dynamical systems and discover behavior that is broadly applicable.
The islets of Langerhans are multicellular micro-organs located in the pancreas which maintain glucose homeostasis through the secretion of hormones such as insulin. Glucose-stimulated insulin secretion (GSIS) from β-cells within the islet is driven by glucose-dependent electrical activity. The metabolism of glucose and increased ATP/ADP ratio inhibits ATP-sensitive K+ (KATP) channels, causing membrane depolarization. Activation of voltage-dependent Ca2+ channels elevates intracellular free-calcium activity ([Ca2+]i) to trigger insulin granule exocytosis [3], [4]. Defects at several points in this signaling pathway, including the KATP channel, can cause or enhance the risk of developing diabetes [5]–[8]. Despite the importance of this pathway, it is important to recognize β-cells do not act autonomously. Rather, like many tissues, there are extensive cell-cell interactions within the islet that govern overall function. For example, isolated β-cells exhibit heterogeneous sensitivities to glucose with a low overall dynamic range of GSIS [9]–[11], yet β-cells within the islet robustly release insulin. Connexin36 (Cx36) gap junctions mediate the electrical coupling between β-cells [12]–[14] which coordinates oscillations in electrical activity and insulin release across the islet, enhancing the pulsatile release of insulin and glucose homeostasis [13]–[15]. In the absence of coupling many cells in the islet also show spontaneous elevations in [Ca2+]i; likely as a result of heterogeneities in glucose sensitivity [10], [16]. Therefore, another equally important role gap junctions play is to coordinate a suppression of spontaneous electrical activity at lower glucose levels [17]. Given that basal regulation is integral to glucose homeostasis, electrical coupling and the coordinated electrical dynamics are a critical factor in the regulation of islet function and in diabetes.
Multicellular electrical dynamics in the islet have been described as functional networks where synchronized changes in [Ca2+]i indicate functional connectivity between cells [14], [18], [19]. Such network analysis has been applied to examine the dependence of [Ca2+]i dynamics on the level of coupling and its regulation, and has indicated that β-cell connectivity is non-homogeneous with a small subset of connections dominating synchronized behavior. As part of this analysis, the network of functional connectivity can be approximated by a Boolean network which quantitatively describes changes in multicellular behavior, including changes in coupling strength, network size or network shape [14], [20]–[22]. These studies have generally focused on the synchronization of [Ca2+]i oscillations, and such synchronized oscillatory/pulsatile behavior has been similarly examined in other physiological multicellular systems [23]–[25]. However, few studies have theoretically examined the suppressive effect of electrical coupling in the islet and its ability to shape the glucose-regulation of electrical activity. This is particularly warranted given a recent study that showed how severe diabetes caused by expression of mutant KATP channels could be prevented through a modulation in gap junction coupling [26]. Therefore, details for how the network structure and composition facilitate a highly sensitive and robust response from a heterogeneous cell population remain to be determined.
In this study we examine how electrical coupling within β-cell networks in the islet provide resilience against heterogeneous cell populations to generate robust network responses. We first develop quantitative predictions derived from a Boolean approximation of the β-cell network, where the dependence of [Ca2+]i on variations in the constituent cellular excitability and coupling is described. We then test these predictions using two experimental systems involving transgenic mice that express mutant KATP channels with increased or decreased ATP-sensitivity [27], [28]. This creates defined populations of cells within the islet which are ‘excitable’ or ‘inexcitable’, and can be further used to examine how our theoretical predictions and experimental data extend to physiological regulation of glucose homeostasis. We next link the static Boolean network model predictions and experimental findings with a dynamic multicellular model of the islet which incorporates recent understanding of β-cell electrophysiology [29], [30]. We finally extend these experimental and theoretical measurements to a general case with a continuum of heterogeneous cellular behavior.
A consistent feature in this study is the emergence of critical behavior as a result of β-cell electrical coupling, where the islet exhibits a phase transition between globally active and inactive states as cellular excitability approaches a critical threshold value. We discuss how the robust functionality that emerges at the multicellular level is not only relevant to the islet of Langerhans and its dysfunction in diabetes, but also to the function of other multicellular biological systems.
Based on prior approximations of heterogeneity in cellular excitability and coupling, Boolean networks of connectivity were simulated to predict how general multicellular electrical activity depends on the relative excitability of the constituent cell population and the coupling between individual cells [14], [20]. Nodes within a cubic lattice had a probability Pexc of being active, and adjacent nodes were functionally coupled with a ‘coupling probability’ p (figure 1A), and resultant clusters of coupled nodes were identified. ‘Inexcitable’ β-cells can suppress activity in excitable β-cells via electrical coupling [17], with <30% inexcitable cells necessary for this suppression. To simulate this, a logic rule was used for each cluster of coupled nodes within a given lattice, where greater than a threshold percentage of inexcitable cells (Sp) can suppress activity in all other cells in its coupled network. Simulations of the resultant average network activity were run with varying values of Pexc, Sp, and p to represent differing cellular excitabilities and electrical coupling (figure 1B, SI figure S1).
An increase in electrical activity is predicted as Pexc is increased; however the functional form is highly dependent on p (figure 1C). In the absence of coupling (p = 0), a trivial linear response is obtained where Pexc represents the level of electrical activity. With increasing p, the activity becomes increasingly non-linear as a function of Pexc. For higher values of p (0.3 to 1) a sharp transition between active and quiescent behavior is observed, representing a phase transition with emerging critical behavior. These higher values of p lead to network-spanning coupling (figure S1, S2), and as such the ‘rule’ governing suppression acts over the whole network. For low values of p (0 to 0.2), the network is composed of coupled ‘clusters’ (figure S2), and the simulation is close to linear without a strong transition. This level of p corresponds to insufficient coupling to span the network, which is similar to the critical coupling probability (∼0.25) in percolation theory [31]. As such for p>0.25 there are 3 specific regions of emergent network behavior: a small (∼10%) decrease for Pexc>0.85 (‘pre-critical state’); a rapid ∼75% drop at Pexc = 0.85 (‘critical state’), then a small linear decrease for Pexc<0.85 (‘post-critical state’). The critical state emerges when Pexc approaches 1-Sp; with the sharpness of the transition as well as behavior in the pre- and post-critical states being dependent on p.
Overall, this transition can be understood by considering the well-defined threshold for activity (Sp) and the network-spanning connectivity that occurs above the critical coupling probability (p>0.25). For values of Pexc<(1-Sp) there is a gradual decrease in network activity with increasing p (figure 1D), representing the suppressive effect of coupling. For Pexc>(1-Sp), the network activity remains high, although a slight drop occurs for low levels of coupling. Therefore in a general Boolean network, electrical coupling is predicted to lead to critical behavior, where a phase transition in the network activity occurs as a function of constituent cellular activity.
To test the Boolean network model predictions, we measured intracellular free-calcium activity ([Ca2+]i) in islets which had defined levels of cellular excitability. Islets were isolated from mice with inducible, β-cell specific expression of a mutant ATP-insensitive KATP channel subunit (Kir6.2[ΔN30,K185Q]) under CreER-recombinase control [27]. Expression of these over-active KATP channels render β-cells functionally inexcitable, causing an absence of insulin release, marked hyperglycemia and diabetes [27]. Tamoxifen induction of CreER controls Kir6.2[ΔN30,K185Q] expression levels which can be monitored via GFP co-expression, leading to both controllable and quantifiable cellular excitabilities (SI figure S3).
At 20 mM glucose islets show [Ca2+]i which decreased with increasing expression of GFP and therefore Kir6.2[ΔN30,K185Q], similar to model predictions. This showed critical behavior with 3 specific regions (figure 2A): For low GFP expression <15% (few Kir6.2[ΔN30,K185Q] expressing cells), [Ca2+]i was active over the entire islet with similar behavior to wild-type islets lacking GFP (GFP = 0) and Kir6.2[ΔN30,K185Q] (figure 2A, BI–II, ‘pre-critical’ behavior). Oscillations were almost fully synchronous in each case (not shown). For GFP expression at 10–20% there was a sharp drop-off in islet [Ca2+]i, where small changes in GFP resulted in highly disproportionate changes in [Ca2+]i (figure 2A, BIII, ‘critical’ behavior). Activity was focused to clustered areas of synchronization (not shown). For high GFP expression >25% (high number of Kir6.2[ΔN30,K185Q] expressing cells), islets showed sporadic [Ca2+]i restricted to increasingly smaller clusters (Figure 2DIV, ‘post-critical behavior’). Although islets with <20% GFP has similar overall activity compared to wild-type islets (0% GFP), there was a marked reduction in the plateau fraction in the <20% GFP group (15–40%) compared to the wild-type group (50–70%); indicating that even small numbers of inactive cells impacts global behavior. In comparison, islets with high levels of GFP (>25%) had a low plateau fraction (15–20%) in those cells that were active. GFP+ cells showed similar activity to GFP− cells albeit with a small increase in activity, likely due to a few inactive non-β-cells included in the GFP− analysis (SI figure S4).
Comparison of experimental data to the Boolean model can be seen for a number of values of p (figure 2C) and Sp (figure 2D). Varying p (gap junction coupling) matches the sharpness of the transition, whereas varying Sp (number of inexcitable cells required to suppress activity) matches the position of the transition. A p = 0.3 (95% CI: 0.280–0.311) and Sp = 0.135–0.15 best fits the experimental [Ca2+]i data (figure 2A,E). The distribution for fitted p was relatively broad but for Sp was well defined (figure 2E). These values of p are similar to those found in prior studies examining the synchronization of [Ca2+]i oscillations, which indicated a limited level of functional coupling in the islet (p = 0.31–0.36) [20]. These values of Sp are also consistent with experimental studies that suggest between 1 and 30% of inactive cells can suppress activity in other cells [17], [20].
Therefore, introducing inexcitable cells into the islet experimentally generates critical behavior which quantitatively agrees with a Boolean network model and predicts the importance of electrical coupling in regulating multicellular excitability.
β-cell [Ca2+]i drives insulin release to regulate glucose homeostasis. Given that the behavior in [Ca2+]i following varied expression of over-active KATP channels, we next tested whether this also occurred in downstream physiological parameters. Averaged over each mouse, similar [Ca2+]i was observed in wild-type islets lacking GFP and islets with low-level GFP (<20%, ‘pre-critical’), while both were significantly greater than [Ca2+]i in islets with high GFP expression (>20%, ‘post-critical’) (figure 3A). Plasma insulin also showed a similar transition, with pre-critical (GFP<20%) plasma insulin being significantly greater than post-critical (GFP>20%) plasma insulin (figure 3B). However mice lacking GFP did show significantly greater insulin than mice with low-level GFP, correlating with the reduced plateau fraction observed. Insulin reduces glucose levels, and as expected pre-critical mice (GFP<20%) had normal glucose levels (figure 3C), while post-critical mice (GFP>20%) demonstrated elevated glucose levels. Glucose-stimulated insulin secretion from isolated islets showed similar behavior to that of plasma insulin (figures 3D–F), where again islets lacking GFP showed significantly greater GSIS than islets with low level GFP.
Therefore insulin dynamics and blood glucose levels follow similar behavior as the driving [Ca2+]i following varied expression of Kir6.2[ΔN30,K185Q], demonstrating a physiological link in the critical behavior in [Ca2+]i activity as a function of Pexc,.
The Boolean model accurately predicts the impact of variable cellular excitabilities (Pexc) on [Ca2+]i suppression at elevated glucose through expression of over-active KATP channels (Kir6.2[ΔN30,K185Q]). However, the Boolean model also predicts how [Ca2+]i suppression varies as a function of gap junction coupling (p) (Figure 1D). To test this, we measured [Ca2+]i in islets from mice with β-cell specific mosaic expression of an inactive KATP channel subunit (Kir6.2[AAA]). This was combined with a knockout of Cx36, yielding 100% (Cx36+/+), 50% (Cx36+/−) or 0% (Cx36−/−) gap junction coupling, as well application of the gap junction inhibitor 18-α-glycyrrhetinic acid [10]. Expression of inactive KATP channels render β-cells constitutively (glucose-independent) active, yet islets which have a majority (but not all) of their cells expressing inactive KATP channels show glucose-dependent electrical activity similar to wild-type islets [17]. GFP co-expression indicates ∼70% of β-cells express inactive KATP channels (SI figure S3) such that Pexc = 0.7. With increasing gap junction coupling [Ca2+]i progressively decreased until residual activity was observed at full coupling, similar to that in the post-critical state upon Kir6.2[ΔN30,K185Q] expression. There was strong agreement between experimental measurements and the Boolean Network model, with a p at normal (Cx36+/+, 100%) gap junction coupling of 0.38 (95% CI: 0.372–0.394) and a suppression threshold Sp = 0.15 giving the best fit (figure 4A). This is similar to p, Sp derived in the first experimental system (figure 2). Varying Sp affects the gap junction dependence in [Ca2+]i, with little effect between 0.05–0.2, but strong divergence above 0.2 (figure 4B). Therefore the Boolean network model can accurately predict behavior in a different experimental model with defined levels of cellular excitability (Pexc) and gap junction coupling (p).
The Boolean network model accurately describes how [Ca2+]i critically depends on cellular excitability and coupling. Nevertheless it is a static framework of a dynamical system and does not take into account limit-cycle behavior. To investigate whether similar behavior exists in a coupled dynamical oscillator model of the islet, we generated a multi-cellular version of a recent β-cell model which includes a comprehensive description of β-cell electrophysiology [30]. Our model also included a more realistic quasi-spherical architecture, heterogeneity in gap junction coupling [14], [32], and heterogeneity in endogenous cellular activity [11], [14]. To model KATP-overactivity resulting from Kir6.2[ΔN30,K185Q] expression, a defined fraction of cells with reduced ATP-inhibition of KATP activity was introduced to render them inexcitable.
As with the Boolean network model and experimentally measured [Ca2+]i, a clear phase transition was observed at 20 mM glucose in the coupled oscillator model with ∼15% KATP-overactivity (figure 5A). Again critical behavior manifested in three regimes. Simulated islets without KATP over-activity showed [Ca2+]i dynamics closely matching previously published models (figure 5BI) [30]. Simulated islets with low KATP-overactivity (<15%) showed a linear decrease in activity with a reduced plateau fraction as experimentally observed (figure 5A, BII, ‘pre-critical’ behavior), while maintaining near-full synchronization. Simulated islets with KATP-overactivity at 10–30% again showed a sharp drop-off in [Ca2+]i, with small changes in KATP-overactivity leading to highly disproportionate changes in [Ca2+]i (figure 5A, BIII). Simulated islets with high KATP-overactivity (>30%) showed only sporadic low level [Ca2+]i (Figure 5A, BIV, ‘post-critical’ behavior). A physiological mean gap junction conductance of 120 pS [14], [32] was found to best describe experimental data (figure 5A,C). The sharpness and position of the phase transition was highly dependent on the mean coupling conductance, with increasing conductance leading to a sharper transition occurring at lower KATP-overactivity (figure 5C).
The islet is commonly modeled as a cubic lattice or other regular geometry [14], [21], [33]. A spherical islet-like structure which has a heterogeneous number of cell-cell connections (mean,SD = 5.3,1.7) generated a less-sharp transition compared to a regular cubic geometry, and this better matched experimental data (SI figure S5). Similarly, a heterogeneous level of coupling conductance generated a less-sharp transition (SI figure S6A). This indicates the importance of coupling heterogeneity, in terms of connection geometry, connection number and connection strength. The endogenous heterogeneity in cellular activity did not significantly impact the phase-transition indicating the dominating effect of Kir6.2[ΔN30,K185Q] expression (SI figure S6B). A similar phase-transition was also observed for simulations run at 11 mM glucose (not shown).
Therefore critical behavior also emerges in a dynamic coupled β-cell oscillator model with quantitative agreement with experimental measurements and a static Boolean network model.
We have examined how the coupling between heterogeneous cells leads to critical behavior by introducing defined mutant populations of inexcitable cells (Kir6.2[ΔN30,K185Q]) or excitable cells (Kir6.2[AAA]). However, endogenous β-cells are themselves highly heterogeneous under physiological ranges of glucose, showing a continuum of excitabilities rather than being constitutively excitable/inexcitable [9]–[11].
To examine how gap junction coupling leads to critical behavior in the presence of endogenous heterogeneity, we applied a ‘ramp’ of increasing diazoxide concentrations to uniformly promote KATP channel opening. At 11 mM glucose, [Ca2+]i in wild-type islets at 0 µM and 50 µM diazoxide was similar, but at 100 µM there was a rapid ∼60% drop (figure 6A), where only a few remaining cells were active (figure 6B). Similar low-level [Ca2+]i was observed at 250 µM. In islets from mice lacking Cx36 gap junction coupling, similar [Ca2+]i was observed to wild-type islets at 0 µM diazoxide, albeit with no synchronization. Upon increasing diazoxide, a more gradual decrease in [Ca2+]i was observed, with less [Ca2+]i observed at 50 µM diaozixde but more [Ca2+]i remained at 100 µM diazoxide (figure 6A,B).
These data were also well described using the coupled dynamic oscillator model. In the presence of endogenous heterogeneity at 11 mM glucose, a uniform reduction in ATP-sensitive KATP inhibition led to a clear phase transition in islet [Ca2+]i in the presence of normal coupling (120 pS) (figure 6C,D). However, in the absence of coupling a more gradual change occurred in good agreement with experimental measurements (figure 6C,D); where [Ca2+]i was elevated in the absence of coupling over a certain range of uniform KATP inhibition. As such, <50 µM diazoxide lies in the ‘pre-critical’ regime, >100 µM diazoxide lies in the ‘post-critical’ regime, and the transition lies at 50–100 µM.
In experiments with mutant KATP subunit expression, cells were considered ‘inexcitable’ if they showed GFP and Kir6.2[ΔN30,K185Q] expression. In this case of endogenous heterogeneity, for a given concentration of diazoxide, we can consider a cell is ‘inexcitable’ if it is quiescent in the absence of electrical coupling. By plotting activity in the presence of coupling (representing the resultant activity) against activity in the absence of coupling (representing intrinsic cellular excitability) similar phase-transitions are apparent; with quantitative agreement between experimental data, dynamic coupled oscillator model and static network model (figure 7A–C). The phase transition in the dynamic coupled oscillator model was dependent on how heterogeneity was generated, where heterogeneity in multiple factors rather than any one factor was required for agreement with experimental data (SI figure S7). Therefore critical behavior can occur more generally from the coupling between heterogeneous cellular populations within the islet, as exemplified here experimentally and theoretically.
The islet of Langerhans shows unique functional properties that result from the underlying network interactions between constituent cells. One important property is that β-cells within the islet show globally quiescent behavior at lower levels of glucose despite showing a heterogeneous range of glucose sensitivities when in isolation. A proposed mechanism underlying this behavior is that at a given glucose stimulation, inactive cells suppress cells that otherwise would be active, via gap junction coupling. We applied predictive mathematical models to quantify this behavior and determined the relative role of KATP channel activity (controlling cellular excitability) and gap junction activity (controlling cellular coupling) in shaping this islet response. We then experimentally verify predicted behavior using two independent experimental models.
In line with previous work describing coupled electrical dynamics, we showed that the structure and function of the islet cellular network can be described through principles of network theory [20], [21]. Both the Boolean network and dynamic oscillator models predict the emergent behavior upon coupling between a heterogeneous cell population. The islet rapidly transitions between globally coordinated active and inactive states upon disproportionally small changes in the excitability of the constituent cells as they approach a critical ‘threshold’ excitability. This occurs under both conditions of β-cell heterogeneity we examined: the imposed bimodal β-cell populations achieved through expression of Kir6.2[ΔN30,K185Q] or Kir6.2[AAA] mutations; and endogenous β-cell heterogeneity with diazoxide activation of KATP.
The Boolean model reveals that there is an imbalance in the ability of excitable and inexcitable cells to respectively propagate stimulation or suppression. A low Sp in the model indicates a preference for excitable cells to be suppressed by inexcitable cells. This describes how gain-of-function Kir6.2[ΔN30,K185Q] expressing cells (which are glucose-unresponsive) suppress activity in coupled normal cells at high glucose, and how loss-of-function Kir6.2[AAA] expressing cells are suppressed by normal cells at low glucose (figure 2, 4). The role of p (gap junction coupling) determines the spatial extent over which suppression occurs. As shown in figure S1 and S2, a low p results in coupled behavior restricted to a few cells and therefore inactive cells are unlikely to couple to many active cells and mediate suppression. When p exceeds the critical coupling probability (∼0.25) then coupling spans the whole network and inactive cells can couple to and suppress most active cells in the network. The sharp transition that emerges upon p>0.25 can be understood by considering that the threshold for activity (Sp) is well defined with a sharp cutoff for the Pexc which determines whether the cluster is active or inactive. The agreement with experimental data indicates that there is little variability between cells in this threshold for suppression, as also supported by the distributions of fitted Sp (figure 2E).
While the coupled dynamic oscillator model also predicts and describes the phase transitions present, the Boolean model describes the essential features that govern multicellular regulation of islet excitability. Results suggest that the islet may fundamentally behave in a binary fashion in terms of gap junction coupling and KATP-regulated excitability. Given the proportion of cells that intrinsically (i.e. in the absence of coupling) show activity at a given glucose stimulation and the level of coupling, the overall response of the islet can be approximated through this reductionist model. Of course dynamical features are missing from the Boolean model which is only described by the coupled dynamic oscillator model: including the altered oscillatory characteristics in the pre-critical state. The low p (0.30–0.38) required for the Boolean network model to quantitatively describe experimental data points to incomplete coupling present; and this can explain the residual activity in the post-critical state (figure 1C). Recent studies of coordinated [Ca2+]i oscillations and waves in the islet have indicated a ‘backbone’ of a few strong connections dominate coupled [Ca2+]i dynamics, which is equivalent to a similarly low p [19], [20]. The ability of the coupled dynamic oscillator model to also describe the transition between globally active and inactive states, suggests that the dynamics of the islet may behave according to general principles of coupled dynamical systems. Further work is needed to examine this critical behavior in more detail, including power law scaling and its dependence on network parameters and cellular properties.
The phase-transition behavior can also be explained through a mean-field theory analogy (SI figure S8). Cells expressing the mutant Kir6.2[ΔN30,K185Q] are intrinsically inexcitable (figure S8A). In the ‘pre-critical’ regime the number of these cells is below a critical threshold and insufficient to suppress glucose-stimulated activity via coupling; therefore all cells are recruited to elevate [Ca2+]i. When the number of these inactive cells approaches the critical threshold (Sp = 0.15 for Kir6.2[ΔN30,K185Q] expression) critical behavior emerges and coupling mediates suppression of other active cells. In normal islets treated with diazoxide, endogenous β-cell heterogeneity leads to variable intrinsic excitabilities and we expect diazoxide renders cells less glucose sensitive to be inexcitable (figure S8B). In the absence of coupling these are observed to be inactive (figure 6). Low concentrations of diazoxide (<50 µM) render only a few cells inexcitable, which is below the critical threshold (Sp∼0.5) and insufficient to suppress [Ca2+]i. At higher concentrations of diazoxide (>100 µM) more cells are rendered inexcitable, and when this number exceeds the critical threshold, coupling mediates suppression of other normally excitable cells. We predict that observed glucose-dependent activity and the coupling dependence can also be explained in this way (see below).
The Sp for endogenous heterogeneity is higher than that for an imposed biomodal distributions (i.e. diazoxide treatment versus Kir6.2[ΔN30,K185Q] or Kir6.2[AAA] expression) suggesting a more even balance between the ability of excitable and inexcitable cells to respectively propagate stimulation or suppression in wild-type islets (figure 6, 7). This balance may arise from the different distribution of heterogeneity present, but a phase-transition still emerges in the presence of coupling indicating a more general regulation of multicellular excitability.
Therefore through limited coupling of heterogeneous populations of cells, critical behavior emerges in the islet dynamical system where large changes in activity result from small changes in the constituent cellular excitabilities.
Gap junctions impact islet behavior in two main ways. At high glucose (KATP channel-closure), gap junctions coordinate oscillatory dynamics of membrane depolarization and [Ca2+]i to generate a robust pulsatile insulin secretion [13], [14], [32]. A number of recent studies have examined this aspect, including multicellular modelling and quantitative analyses [14], [18], [19], [21]. Equally important however is that at lower glucose (KATP channel-opening), gap junctions mediate a suppression of membrane depolarization, [Ca2+]i, and insulin secretion [10], [13], [17]. The mechanisms involved in mediating suppression are not well characterized, and several experimental perturbations have yielded unexpected results or have not been well described theoretically [17], [34], [35]. Here, we were able to quantitatively describe suppressive behavior resulting from coupling, which yields a more complete understanding for how the islet functions under conditions of KATP channel opening.
At 6–7 mM glucose, the islet sharply transition between global quiescence and globally synchronized [Ca2+]i oscillations. In the absence of coupling, the progressive elevation in the number of cells showing [Ca2+]i elevations is gradual [10]. This follows the same behavior as variable Kir6.2[ΔN30,K185Q] expression and diazoxide concentration (figures 2,6). At <6 mM glucose, global suppression is equivalent to >15% Kir6.2[ΔN30,K185Q] expression or >100 µM diazoxide; whereas at >7 mM glucose global activity is equivalent to <15% Kir6.2[ΔN30,K185Q] expression or <50 µM diazoxide. The 6–7 mM glucose transition is therefore equivalent to behavior at ∼15% Kir6.2[ΔN30,K185Q] expression or 50–100 µM diazoxide. As such, we propose results from the Boolean model, as illustrated by the mean-field theory analogy, have greater implications by describing physiological glucose-dependent islet electrical activity (figure S8C). Coupling heterogeneity and islet architecture lead to variability in the number and strength of connections, impacting the phase transition. These factors may therefore play a role in shaping the physiological regulation of glucose-stimulated [Ca2+]i and insulin secretion (Figures S5,S6).
At 11 mM glucose, heterogeneity leads to a small population of β-cells (<10%) remaining inactive in the absence of coupling [10]. In the presence of coupling there is global activity with a lower plateau fraction compared to higher glucose levels (e.g. 20 mM) [36]. This matches the behavior at 5–10% Kir6.2[ΔN30,K185Q] expression or 50 µM diazoxide in the respective absence and presence of coupling. Therefore an alternative view for how oscillatory dynamics are shaped at an islet-wide level is that less-active cells within the β-cell network have a modulatory effect on overall oscillation waveform, rather than oscillations being shaped by purely intrinsic properties of the β-cells. Importantly, the reduced plateau fraction of [Ca2+]i bursts at 5–10% Kir6.2[ΔN30,K185Q] expression correlates with a significant decrease in insulin secretion (figure 3). A decrease in burst duration has previously been suggested to reduce insulin release [37], as supported by these results. Thus subtle alterations in the balance of constituent cell excitabilities have a strong physiological effect on islet function.
Our results also have implications for neonatal diabetes mellitus (NDM), where the majority of cases result from mutations to Kir6.2 or SUR1 KATP channel subunits [8], [38]. Kir6.2[ΔN30,K185Q] expression models this disease [27]. Our results show that NDM mutations gives rise to a disproportionate suppression in [Ca2+]i and insulin release, thereby causing diabetes due to the critical behavior that emerges from coupling and network dynamics. This also explains how an absence of coupling elevates [Ca2+]i and insulin release (figures 1, 6) to prevent the progression of diabetes. This was experimentally demonstrated in a recent study [26], and the rescue of diabetes can only be understood mechanistically at the multicellular level. Clearly in human diabetes, mutations are not expressed mosaically. However, the diazoxide results which depend on a continuum of heterogeneity (figure 6) demonstrate that critical behavior exacerbates NDM upon uniform KATP channels overactivity. Other monogenic diabetes causing mutations that affect β-cell excitability, such as Glucokinase [6], may also have similar effects on islet excitability and lend themselves to analysis by the Boolean model and coupled oscillator model.
Mutations causing NDM are functionally equivalent to >15% Kir6.2[ΔN30,K185Q] or >100 µM diazoxide, effectively residing in a post-critical state suppressing global [Ca2+]i. There exists a spectrum of KATP channel mutations linked to diabetes, where weaker mutations to Kir6.2 and SUR1 elevate the risk of type2 diabetes [39]–[41]. These mutations likely have a more subtle effect on islet excitability and as a consequence we predict that islets residing in the pre-critical state (<15% Kir6.2[ΔN30,K185Q] or <50 µM diazoxide) would still be susceptible to diabetes following metabolic stress. Further, while gap junction reduction recovers insulin release and glucose control in the post-critical regime (i.e. NDM), we predict a gap junction increase would be beneficial in the pre-threshold regime (i.e. type2 diabetes).
Converse to this, results from Kir6.2[AAA] islets show how critical behavior provides the islet with a resilience to over-excitable β-cells. Given the ∼85% threshold of excitable cells required to elevate [Ca2+]i, with ∼70% over-excitable Kir6.2[AAA]-expressing cells, many of the ∼30% normal β-cells would also need to be active (e.g. >50% at ∼5.5 mM glucose [10]). This explains only the minor shift in glucose-stimulated [Ca2+]i that occurs following KATP inactivity and highlights the role electrical coupling plays in protecting islets against hyper-excitability [17].
Therefore we describe the emergence of critical behavior linking multiple levels including molecular and cellular behavior, multicellular behavior, in-vivo physiology, disease and treatment.
This study also has implications for general understanding of physiological systems composed of coupled dynamic units. Previous theoretical studies have shown how the introduction of non-oscillatory elements above a critical level in a generalized coupled oscillator system can lead to cessation of global oscillations with phase transitions [42], [43]. Our study experimentally and theoretically demonstrates this in a disease relevant system. Further, prior studies theoretically demonstrated that the fraction of excitable elements (i.e. Pexc) and coupling strength (i.e. p) exist in a phase plane where increased coupling decreases the number of inactive elements required for suppression [43], [44]. We demonstrated this experimentally and theoretically, with Kir6.2[ΔN30,K185Q] and diazoxide-induced suppression. While strong coupling promotes robust synchronization, it will increase suppression from non-oscillatory inexcitable units. Given a small population of inactive β-cells exists as a result of cellular heterogeneity, these generalized theoretical studies imply inappropriate elevations in coupling would be deleterious, by reducing glucose-stimulated [Ca2+]i. As such the level of coupling is likely at an optimal level to balance global synchronization and suppression.
The strong link between dynamical β-cell networks and generalized coupled oscillators implies similar behavior can be expected in other physiological systems. In the heart, electrical activity is initiated by pacemaker cells and propagates to excite contractile myocytes. In culture, non-excitable fibroblasts proportionally reduced cardiomyocyte wave propagation bursts frequency with Cx43 dependence [45]. No activity was reported for >30% fibroblast penetrance and modulation of action potential frequency occurred at <30% fibroblast penetrance (implying Sp = 0.3). While the mechanisms of coupling-dependent suppression are very different compared to our study, global responses are similar implying similar governing principles. Similarly, pacemaker cells exhibit dominance over myocytes at an optimal gap junction conductance [46]; where high coupling leads to arrhythmias and low coupling leads to poor synchronization [47]. Neurons also display intrinsic oscillatory behavior, and the effects of coupling and presence of inhibitory and excitatory neurons on synchronization and phase modulation is an active area of research [48]–[53]. Critical dynamics have been described theoretically to emerge from excitatory and inhibitory units in neuronal networks [54], and a computational study which introduced ‘contrarian elements’ into neural coupled oscillator networks found that a similar threshold of 15% suppressed global dynamics [49].
We also anticipate that critical behavior resulting from coupling of heterogeneous units may be considered a general regulatory mechanism. Many systems respond to a stimulus by transitioning between inactive and active states (e.g. contractile, hormone-secretory). Our study implies that constituent cellular units need not themselves have a uniform or robust response to generate a robust multicellular response. Rather, a robust response can emerge from coupling a heterogeneous collection of cells; where coupling and architecture need only have sufficient strength and connection number on average. This makes the overall system robust against noise and variability, and loosens the requirement for tight regulatory mechanisms within the constituent cells. Similarly, given a constant stimulus, a robust transition between globally active and inactive states could be achieved by remodeling connectivity with a small number of inhibitory units. For example, down-regulating connections from ∼20% to ∼5% inhibitory cells would transition the system from inactive to active requiring minimal system remodeling. This also suggests how inappropriate changes in coupling or constituent cells may lead to global non-responsiveness and disease. We speculate these principles may apply to other neuroendocrine cell systems such as GH-cells or the adrenal medulla, where functional remodeling elevates hormone secretion upon physiological stimuli [18], [55]. Indeed many of these principles have been linked with GnRH neuron function during development [56], [57]. Therefore we suggest that new and robust functionalities can be generated at the multicellular level from the coupling of non-robust constituent cell function, requiring minimal system resources compared to the requirements were cells to act autonomously.
Any living system cannot avoid deterioration through mutation or other pathological insult. This study experimentally and theoretically demonstrates that if the fraction of inactive elements exceeds a coupling-dependent threshold, the global activity of the system can be abolished. In the case of the islet this explains how inactive cells can suppress the activity of other cells, thereby preventing the secretion of insulin. In the case of KATP mutations, this quantifies the threshold of inexcitable cells required for pathogenic symptoms and explains how coupling can eliminate the emergence of diabetes or exacerbate it. Overall, this gives a new understanding for how emergent properties of the islet as a β-cell network are generated; as well as for understanding islet dysfunction in diabetes and novel ways to overcome dysfunction. More broadly, this generates insight into emergent behavior of multicellular systems in general.
All experiments were performed in compliance with the relevant laws and institutional guidelines, and were approved by the University of Colorado Institutional Biosafety Committee (IBC) and Institutional Animal Care and Use Committee (IACUC).
The generation of Rosa26-Kir6.2[ΔN30,K185Q] (‘gain-of-function’ KATP subunit with GFP co-expression), Pdx-CreER (β-cell specific inducible Cre), Kir6.2[AAA] (‘loss-of-function’ KATP subunit with GFP tag), and Cx36−/− (Connexin36 global knockout) have been described previously [27], [28], [58], [59]. Expression of variable Kir6.2[ΔN30,K185Q] was achieved in β-cells by crossing Rosa26-Kir6.2[ΔN30,K185Q] and Pdx-CreER mice, and inducing Kir6.2[ΔN30,K185Q] expression in 8–16 week old mice by 1–5 daily doses of tamoxifen (50 mg/g body-weight). Littermates lacking Rosa26-Kir6.2[ΔN30,K185Q] and/or Pdx-CreER were used as controls.
Blood glucose was measured daily and averaged over day 27–29 post tamoxifen induction using a glucometer (Ascensia Contour, Bayer). Plasma insulin was measured at day 29 from blood samples centrifuged for 15 minutes at 13,900RCF, and assayed using mouse ultrasensitive insulin ELISA (Alpco).
Islets were isolated by collagenase injection into the pancreas through the pancreatic duct; the pancreas was harvested and digested, and islets were handpicked [23]. Islets were maintained in RPMI medium (Invitrogen) supplemented with 10% FBS, 11 mM glucose, 100 U/ml penicillin, 100 µg/ml streptomycin, at 37°C under humidified 5% CO2 for 24–48 hours prior to study. For insulin secretion measurements, islets (5/column, duplicates) were pre-incubated in Krebs-Ringer buffer (128.8 mM NaCl, 5 mM NaHCO3, 5.8 mM KCl, 1.2 mM KH2PO4, 2.5mM CaCl2, 1.2 mM MgSO4, 10 mM HEPES, 0.1% BSA, pH 7.4) plus 2 mM glucose; then incubated for 60 minutes in Krebs-Ringer buffer plus 20 mM glucose. After incubation, the medium was sampled and insulin concentration assayed using mouse ultrasensitive insulin ELISA. To estimate insulin content, islets were lysed in 1% TritonX-100 and frozen at −20°C overnight.
To measure [Ca2+]i dynamics, isolated islets were loaded with 4 µM FuraRed-AM (Invitrogen) in imaging medium (125 mM NaCl, 5.7 mM KCl, 2.5 mM CaCl2, 1.2 mM MgCl2, 10 mM Hepes, 2 mM glucose, and 0.1% BSA, pH 7.4) at room temperature for 90–120 minutes and held in polymdimethylsiloxane (PDMS) microfluidic devices [17]. FuraRed fluorescence was imaged on a spinning-disk confocal microscope (Marianas, 3I) with a 40× 1.3NA Plan-NEOFluar oil-immersion objective (Zeiss) maintained at 37°C. Images were acquired at 1 frame/sec using a 488 nm diode laser for excitation and a 580–655 nm long-pass filter for emission. Time-courses were acquired 10 minutes after change in glucose concentration, diazoixide or 18-α-glycyrrhetinic acid application. Time-courses are displayed as normalized to the average fluorescence.
All images were analyzed using custom MATLAB (Mathworks) routines or using Slidebook (3I). To calculate islet activity, images were smoothed using a 5×5 average filter. The variance of pixel time-courses was first calculated for a quiescent reference cell; manually selected from an area which displayed no fluctuations in intensity over time compared with image noise. A pixel was considered ‘active’ if its time-course showed a variance >2 standard deviations above the variance of the quiescent reference cell [10], [20]. Photobleaching was accounted for through a linear fit, and time-courses were rejected if excessive motion artifacts occurred. The area of active cells in terms of pixels, was determined for each condition and expressed as a fraction of total islet pixel area as defined by mean FuraRed fluorescence. GFP+ regions were defined as having a mean pixel fluorescence intensity above that measured in GFP− wild-type cells. The area of active cells in GFP+ regions was expressed as a fraction of the total GFP+ area. Information describing activity is represented as a false-color HSV image where Hue is set to 1 (red), Saturation is set to 1 for active cells and 0 (no color) for inactive cells, and Value (intensity) is set to the average FuraRed fluorescence. Data are presented as mean±SEM. For comparison of two means, Student's t-test was utilized. For comparison of multiple means, ANOVA was utilized along with Tukey's HSD test.
Bond percolation is a sub-model of percolation theory [31], [60] which can be used to simulate the islet [14], [20], [21]. For a lattice of nodes (cells) in a given geometry, adjacent nodes are connected with a ‘coupling probability’ p, or not connected with a probability (1-p). Connected nodes are considered ‘functionally coupled’, where activity is synchronized at high glucose and suppression mediated at low glucose. We implemented simulations of bond percolation lattices (figure S1) as previously described [20]. Briefly, cubic lattices with alternating node and bond sites (length L = 11) were generated. Probabilities were assigned to each bond site with a uniform distribution (0 to 1). Neighboring nodes were coupled if the bond site probability was less than or equal to the coupling probability p. Clusters of coupled nodes were identified and potential bond sites removed to establish a matrix of identified coupled nodes (figure S1A).
Coupling-mediated suppression is based on the principle that a threshold fraction of non-responsive (‘inexcitable’) cells can suppress all other cells to which they are coupled [17], [20]. Within a cluster of coupled nodes, if the fraction of inexcitable cells is greater than a threshold fraction (Sp), then all cells within the cluster are inactive. Experimental studies indicate this threshold is <30% [17] and has been modelled to be ∼15% for MIN6 aggregates [20]. The probability Pexc defines the fraction of cells within the islet that are intrinsically excitable; where in the absence of coupling they would be active. Within each cluster of coupled nodes a binomial distribution was used to estimate the probability of there being a threshold number of inactive cells to lead to suppression. Given a threshold number of inactive cells required for suppression (k), a total number of cells in a cluster (n), and the fraction of inactive cells (q); the probability that a coupled cluster is active (Pr) is:(1)where(2)
The cumulative distribution for k to n, P(X≤k), represents the probability of suppression (sufficient inexcitable cells) in a coupled cluster. To obtain the resultant % activity, P(X≤k) was averaged over all clusters within the islet, weighted by the number of cells n in each cluster. ‘k’ normalized to ‘n’ gives the fraction of inexcitable cells required for suppression (Sp). ‘1-q’ gives the fraction of excitable cells (Pexc). 500 simulations were run for each Pexc = [0 1] and p = [0 1], at 0.01 increments for given values of Sp. The p and Sp parameters that generated the best fit for the simulation mean to experimental data were determined by a chi-squared minimization. To determine the probability distribution for p and Sp, 4000 simulations were run and each simulation was separately fitted for p and Sp by chi-squared minimization.
The islet model was based on the Cha-Noma β-cell model [29], [30], itself based on the Fridlyand β-cell model [61], [62], and adapted to include cell-cell coupling. We also included further aspects of cell-cell coupling and altered KATP channel function. A list of parameters used in the model is included in SI Table S1. The membrane potential (Vi) of each β-cell i is related to the total transmembrane current as:(3)
Where the kinetics of each current is described in [29], [30].
To simulate gap junction coupling and a multicellular islet, multiple ‘cells’ were simulated with a coupling current between each neighboring cell. The membrane potential for each cell i was modified to account for coupling to j neighboring cells:(4)
Heterogeneity in coupling was introduced by randomly assigning the gap junction conductance between cells i and j, according to an experimentally measured distribution [unpublished data], with SD/mean = 70%. To more accurately model β-cell coupling architecture, random cell lattices were created using a position- and availability-based sphere-packing algorithm (mean,SD number of cell-cell connections = 5.3,1.7) [63] (figure S3).
IK(ATP) was described in [29], [30] as:(5)
Where is the open channel conductance and represents the mean open probability which is given by:(6)
Endogenous heterogeneity was modelled by randomizing all parameters indicated in Table S1 between cells about a mean value according to a Gaussian distribution with SD/mean = 10%. To generate heterogeneity in electrical responses equivalent to experimental measurements, the open channel conductance was randomized between cells about a mean value according to a Gaussian distribution with SD/mean = 25%. This heterogeneity achieves variability in activity that matches experimental measurements in islets lacking Cx36 [10], [14].
To model Kir6.2[ΔN30,K185Q] expression, the open probability was modified in a proportion (Pexc) of simulated cells:(7)where γ is a constant representing the fraction of ATP-insensitive current (), and was set to 0.5.
To model diazoxide treatment, the fraction of ATP-insensitive current was increased in all cells uniformly according to:(8)such that α = 1 represents an untreated islet, and α = 0.5 is equivalent to 100% expression of Kir6.2[ΔN30,K185Q].
All simulations were initially written and verified in MATLAB, then rewritten in C++ and simulated on the University of Colorado JANUS supercomputer. The model was solved using a constant time-step Euler integration scheme (Boost C++ Libraries) with 100 µs step-time and 100 ms sampling-time. Rendering simulations was performed with Mathematica 9.0 (Wolfram Research).
|
10.1371/journal.pgen.1006380 | Coordinately Co-opted Multiple Transposable Elements Constitute an Enhancer for wnt5a Expression in the Mammalian Secondary Palate | Acquisition of cis-regulatory elements is a major driving force of evolution, and there are several examples of developmental enhancers derived from transposable elements (TEs). However, it remains unclear whether one enhancer element could have been produced via cooperation among multiple, yet distinct, TEs during evolution. Here we show that an evolutionarily conserved genomic region named AS3_9 comprises three TEs (AmnSINE1, X6b_DNA and MER117), inserted side-by-side, and functions as a distal enhancer for wnt5a expression during morphogenesis of the mammalian secondary palate. Functional analysis of each TE revealed step-by-step retroposition/transposition and co-option together with acquisition of a binding site for Msx1 for its full enhancer function during mammalian evolution. The present study provides a new perspective suggesting that a huge variety of TEs, in combination, could have accelerated the diversity of cis-regulatory elements involved in morphological evolution.
| Acquisition of cis-regulatory elements is a major driving force of evolution, but its whole evolutionary significance is poorly understood. Here, we found an unprecedented case in that three TEs cooperatively and synergistically act as a distal enhancer for wnt5a expression in mammalian frontonasal region, being involved in the evolution of the secondary palate formation. We elucidated the stepwise evolutionary history of getting the enhancer activity via (retro-)transposition of TEs during mammalian evolution. Taking various types of TEs constituting nearly half of mammalian genomes into consideration, this study provides a new perspective for the potentiality that almost infinite combinations of proximally-located different TEs could have generated a huge diversity of developmental enhancers and had a great impact on mammalian morphological evolution.
| Morphogenesis is generally controlled by spatiotemporal expression of a number of specific gene sets [1]. The acquisition of novel phenotypic traits during mammalian evolution has been posited to result from changes in gene expression patterns, which are mediated by gain of new cis-regulatory elements such as enhancers [2]. Mammalian genomes contain hundreds of thousands of conserved non-coding elements (CNEs), which, in humans, occupy 3–8% of the genome [3]. Because CNEs are expected to include a number of transcriptional enhancers [4], they are recognized as highly important clues to understanding the key gene regulatory mechanisms involved in mammalian evolution [5–8].
Mammals have acquired a variety of morphological features during evolution. One of the striking evolutionary events is the development of the bony secondary palate [9]. Complete closure of the mammalian secondary palate during development (palatogenesis) separates the oral cavity from the nasal cavity, which allows breathing while eating and efficient suckling. This closure begins with the formation of bilateral palatal shelves (PS) in the embryonic maxillary prominences, and then the PS grow horizontally to fuse with each other at the midline [10] (S1 Fig). Dozens of genes such as the msx1 and wnt family are known to be involved in palatogenesis [11, 12]. Agenesis of the secondary palate, known as cleft lip/palate, is one of the most common congenital defects in humans, occurring once in every 700 newborns [11]. In this regard, wnt5a is one of the possible responsible genes identified by genetic association studies of human cleft lip/palate [13]. Correspondingly, mice lacking wnt5a or its non-canonical receptor Ror2 are born with cleft palate [14, 15]. Therefore, revealing the molecular regulatory mechanisms of such genes, which remain largely unknown, is essential to our understanding of the molecular basis of mammalian-specific morphological evolution as well as that of the cleft lip/palate defect in humans.
Transposable elements (TEs), i.e., retroposons and DNA transposons, occupy nearly half of mammalian genomes. Retroposons such as SINEs propagate their copies via reverse-transcription of their RNA intermediates, with reintegration of the copied DNA, whereas DNA transposons simply directly relocate within the genome [16–18]. Although TEs are, in general, regarded as genomic parasites or sometimes as harmful dynamic mutagens, we for the first time proposed, together with the Bejerano’s group, that some TEs are involved in macro-evolution by showing that they overlap with CNEs [7, 19]. This fact implies that TEs under purifying selection acquired functions during evolution [20, 21], which is called exaptation [22] or co-option, and that many types of TEs such as SINEs might have contributed to various morphological innovations during mammalian evolution [8]. Indeed, we previously demonstrated that hundreds of AmnSINE1 sequences are evolutionarily conserved among mammals [7, 23, 24]. One AmnSINE1 is an enhancer of fgf8 in the diencephalon, and another acts as an enhancer of satb2 expression in the deep layer of the neocortex, especially in callosal projection neurons [23, 25, 26]. Further, the LF-SINE locus, which is shared among tetrapods, serves as a distal enhancer of the neurodevelopmental gene Isl1 [19], and Pomc has two neuronal enhancers derived from CORE-SINE and MaLR [27, 28]. Thus, it has been clearly established that TEs are one of the main sources of cis-regulatory elements [29].
Mammalian genomes harbor a variety of TEs as exemplified by the human genome, which has >1,100 types (subfamilies) that occupy >45% of the genome. These facts prompted us to consider whether multiple TEs of different origin and sequence, being located proximal to one another, could be co-opted/exapted as a single enhancer element. If so, a huge diversity of developmental enhancers could have been generated by combining different TE types during evolution. This possibility has never been examined, however, and in all the known cases of co-option/exaptation, a developmental enhancer was found to consist of only a single TE [19, 23, 27, 28], although one interesting case in which the promoter region of decidual prolactin was reported to be derived from two TEs [30].
Here, we report that a CNE containing three TEs, including AmnSINE1, acts as a distal enhancer of wnt5a during palatogenesis. This is an unprecedented example, to our knowledge, in which three different TEs inserted side-by-side play a cooperative role in the distal enhancer function within a CNE. TEs located proximal to one another may have potential as genetic sources of diversity of regulatory elements and that such cooperative enhancers might have contributed to mammalian morphological evolution through controlling spatiotemporally diverse expression of various genes.
The 1.2-kb AS3_9 locus is located at chr3:54916774–54917973 of the human genome (GRCh38/hg38) (Fig 1A and 1B). This locus is one of the hundreds of AmnSINE1-derived CNEs in mammals identified by our group [24]. In this locus, the AmnSINE1-related region is 71.4% identical to nucleotide positions 391–501 of its original consensus sequence (Fig 1B, S2 Fig) [7]. To test whether the evolutionarily conserved AS3_9 locus possesses an enhancer function—as is the case for other AmnSINE1-derived CNEs [23, 25, 26]—we performed a transgenic mouse enhancer analysis using a construct containing AS3_9 and a lacZ reporter gene (Fig 1B). The transgenic mice (AS3_9-lacZ) consistently displayed strong lacZ expression in the frontonasal region at embryonic day 13.5 (E13.5) (Fig 1C, S3A Fig). Especially, lacZ was expressed in the frontonasal prominence including the medial and lateral nasal processes, the maxillary processes that give rise to the upper lip and PS, and mandibular processes that form the lower lip. The lacZ expression patterns in the frontonasal prominence during embryogenesis were consistent among the three AS3_9-lacZ mouse lines we established in this study (S3B Fig).
We expected that this TE-derived CNE serves as a distal enhancer of a gene responsible for the development of the frontonasal region in mammals. The ~2-Mb region surrounding AS3_9 contains eight genes (Fig 1A). We carried out in situ hybridization (ISH) for each of the eight candidate genes by using the respective mRNA as a probe; this revealed that only wnt5a is expressed in the frontonasal region (S4 Fig). This is consistent with a previous report that wnt5a is expressed in the frontonasal prominence and anterior side of PS [14] and is responsible for secondary palate development [12, 15]. We also found that lacZ expression in the AS3_9-lacZ embryos coincided exactly with wnt5a expression in the frontonasal prominence at E10.5 (S5A Fig) as well as in the anterior side of PS at E13.5–14.5 (Fig 1C–1E; S5B Fig). These results suggested that AS3_9 is a distal enhancer for wnt5a expression in the frontonasal region including PS.
We next assessed whether AS3_9 serves as an enhancer of wnt5a expression during secondary palate formation. Using AS3_9-ko mouse embryos in which the TE-derived 800-bp region of AS3_9 was targeted (see S6 Fig), we carried out both ISH (wnt5a mRNA probe) and a histological analysis. AS3_9 homozygous mutant mice established from two lines were viable and fertile. Expression of wnt5a in the frontonasal region of E14.5 homozygous AS3_9-ko embryos was weak and/or irregular compared with wild-type littermates (Fig 2). For example, one of the AS3_9-ko mice (#1 in Fig 2) showed no wnt5a expression on the anterior side of PS (Fig 2B) and very weak expression in the PS and mandibular processes in the intermediate region (Fig 2F). Another mouse (#2) displayed moderate wnt5a expression in mandibular processes but little in the PS (Fig 2C and 2G). Therefore, none of the AS3_9-ko embryos showed strong wnt5a expression, i.e., equivalent to that of wild type. This result demonstrated that AS3_9 is indeed an enhancer of wnt5a expression. To investigate whether reduced wnt5a expression could affect PS development, we performed a histological analysis of E14.5 wild-type and AS3_9-ko littermates; notably, the knockout embryos did not exhibit any distinguishable agenesis or delayed palatogenesis (S7A Fig). Moreover, at E15.5, the PS of AS3_9-ko embryos were completely closed as was observed in the wild-type embryos (S7B Fig). Therefore, even when wnt5a expression was unstable or weak in the AS3_9 mutants, palatogenesis progressed essentially normally, presumably owing to a compensation mechanism involving other cis-elements (see Discussion).
The conservation pattern of AS3_9 implies that this CNE can be divided into four sub-elements (conservation graph in Fig 1B), prompting us to investigate the origins of the conserved sub-elements. Interestingly, we found that, in addition to the AmnSINE1 region, two other conserved sub-elements were derived from other TEs, namely X6b_DNA and MER117, which are 74.4% and 72.8% identical to their full-length consensus sequences, respectively (S8A and S8B Fig). X6b_DNA is a non-autonomous DNA transposon distributed in Theria (placental mammals and marsupials), whereas MER117 is a hAT-type non-autonomous DNA transposon distributed only in placental mammals. Consistent with these distributions, we found that the orthologs of the X6b_DNA and MER117 elements in AS3_9 are only found in therian and placental mammals, respectively (Fig 1B; S8 Fig). The other conserved region is not derived from a known TE or repetitive sequences (gray bar in Fig 3), as we confirmed that a RepeatMasker analysis and a blast search against the human genome returned no significant hit.
Because each of the three TEs in AS3_9 was conserved as an independent sub-element of the locus (S2 and S8 Figs), we expected that they make distinct contributions to overall enhancer function. To clarify each role, deletion constructs lacking various combinations of the TE regions were used for enhancer assays with transgenic mice (Fig 3A–3I, S9A–S9I Fig). At E13.5, enhancer activities were evaluated based on lacZ expression in the ventral region including the maxillary and mandibular processes, in the rostral region including the medial and lateral nasal processes, and in PS. Embryos harboring a construct with only the three TE regions (Fig 3B and 3B') showed strong lacZ expression equivalent to that of AS3_9-lacZ mice (Fig 1C, Fig 3A and 3A'). Conversely, the construct lacking the three TE-derived regions lacked enhancer activity in the frontonasal region (Fig 3E and 3E'). These results indicated that the three TE regions were sufficient to recapitulate the full AS3_9 enhancer activity and that other regions, such as the non-TE conserved region (gray bar in Fig 3), probably do not contribute to enhancer function. Constructs lacking the MER117 region showed no or very weak enhancer activity in the medial and lateral nasal processes (Fig 3C and 3C'), whereas only the constructs carrying MER117 yielded a lacZ signal in the apex of the nose (compare Fig 3D' with 3E', 3F' with 3I' and 3G' with 3H'). The MER117 region is, therefore, responsible for enhancer function in nasal processes, especially in the nose apex. The presence of X6b_DNA in the constructs always yielded lacZ expression in the ventral region, namely, the maxillary and mandibular processes (Fig 3A', 3B', 3C', 3F' and 3I'). Consistently, only the lack of X6b_DNA resulted in no lacZ expression in this region (compare Fig 3A' with 3G', 3D' with 3F' and 3E' with 3I'). Accordingly, the X6b_DNA region is responsible for the enhancer activity in the ventral region. The AmnSINE1 region alone did not yield lacZ expression (Fig 3H and 3H'); however, this region increased the range and intensity of the enhancer activity of X6b_DNA and MER117. For example, X6b_DNA alone supported enhancer activity mainly in the maxillary process and weak activity in the mandibular process (Fig 3I and 3I'); addition of AmnSINE1 yielded strong lacZ expression in the mandibular process as well as limited parts of the nasal prominence (Fig 3C and 3C'). Likewise, when AmnSINE1 was present, the enhancer signal of MER117 at the nose apex (Fig 3D') extended somewhat further toward the upper region of the medial nasal processes and part of the lateral nasal processes (Fig 3G and 3G'). Notably, enhancer activity in PS was observed only when all three TE regions were included in the same construct (Fig 3A'' and 3B''). These results suggested that each TE plays a distinct role in wnt5a enhancer function.
To elucidate the molecular mechanism of the AS3_9 enhancer, we utilized the yeast one-hybrid system to search for transcription factors that bind the AS3_9 sequence. Twelve candidate genes were identified (S1 Table, S10A–S10C Fig), of which three (Msx1, Msx2, Gtf2ird1) are known to be involved in mammalian craniofacial development [31, 32]. The most noteworthy finding was msx1 because it has been demonstrated as one of the genes responsible for cleft palate in humans [33] and mice [31, 34, 35]. We found an Msx1-binding motif (TAATTG) [36] within the X6b_DNA-derived sequence of AS3_9. Mutation of this site abrogated Msx1 binding (S10D–S10F Fig). Furthermore, we conducted enhancer analysis using the AS3_9 sequence in which an identical mutation was introduced in the Msx1-binding site. Intriguingly, the transgenic embryos showed limited enhancer activity in the medial nasal process and maxillary process as well as loss of activity in PS (Fig 3J', S9J Fig), similar to the X6b_DNA-deleted constructs (Fig 3D' and 3G'). These results indicated that Msx1-binding is essential for the full enhancer function of AS3_9 and suggested that the msx1 and wnt5a signaling pathways may interact closely during the secondary palate development.
Our analysis of the orthologs within the AS3_9 locus revealed that the AmnSINE1, X6b_DNA, and MER117 elements are present only among Mammalia, Theria, and Eutheria, respectively, suggesting that they were integrated in this order during evolution (Fig 1A, S2 and S8 Figs). Fig 4 shows the evolutionary scenario for the establishment of the AS3_9 enhancer. Although the AmnSINE1-derived region alone lacks enhancer function, this region is evolutionarily conserved even between humans and platypus (S2 Fig). Therefore, the AmnSINE1-derived region might have had another/unknown function in a common ancestor of mammals before 186 million years ago (Mya) [37], and it might have acquired a new additional role as the AS3_9 enhancer during evolution. After divergence of monotremes, integration of X6b_DNA and subsequent acquisition of the Msx1-binding site resulted in co-option/exaptation 170–186 Mya [37]. The AmnSINE1 and X6b_DNA region might serve as the wnt5a enhancer in the developing maxillary and mandibular processes. After divergence of marsupials (170 Mya) [37], MER117 was integrated, and finally AS3_9 was established as the current complete enhancer with extended activity to the medial and lateral nasal processes as well as PS. This is the first demonstration, to our knowledge, of stepwise evolution via co-option/exaptation of a developmental enhancer.
The full activity of the AS3_9 enhancer in the whole frontonasal region and PS was not observed with any one of AmnSINE1, X6b_DNA, and MER117 alone and was only attained with the combination of all three sub-elements (Fig 3A'' and 3B''). Therefore, the three TEs act cooperatively and synergistically as one complex distal enhancer element during palatogenesis. The division of roles among the TEs (Fig 3) implies that they undergo different epigenetic modifications in the different tissues. To address this possibility, we investigated the ChIP-seq data of the ENCODE project available in the UCSC database (S11A Fig) and the Roadmap Epigenomics project data (S11B Fig; http://www.roadmapepigenomics.org/). The UCSC genome browser shows that the chromatin states of AmnSINE1 and X6b_DNA regions of AS3_9 are open in various fibroblast cell lines (black bar in S11A Fig). This strong signal for open chromatin in the X6b_DNA region is also observed in the Roadmap Epigenomics data (see DNase column in S11B Fig). For histone modifications, the AmnSINE1 and MER117 regions show weak inactive/heterochromatin states (e.g., H3K9me3/H3K27me3/H3K36me3) in many other cells (S11B Fig). The ChIP-seq data for transcription factors show the binding of CTCF to the AmnSINE1 + X6b_DNA region and the bindings of NR2C2 and SRF to the MER117 (S11A Fig). Although involvement of NR2C2 or SRF in palatogenesis has not been reported, it is possible that these proteins are involved in the secondary palate formation in mammals. Unfortunately, these epigenetic states has not been tested in the frontonasal region or PS during the corresponding developmental stages of mice. Future examination for these epigenetic signals of AS3_9 may lead us to further understanding of the molecular mechanism for the formation of the secondary palate in mammals.
It is generally considered that robust gene expression is ensured by the presence of a backup cis-regulatory system such as primary and secondary enhancers [38, 39]. Actually, many studies have demonstrated that deletion of one enhancer can perhaps be compensated by another enhancer with little effect on phenotype [40–42]. Therefore, complete palatogenesis in the AS3_9-ko mice was probably due to the remaining of weak wnt5a expression (Fig 2, S7 Fig). It is likely that AS3_9 serves as one of multiple cis-regulatory elements responsible for wnt5a expression during secondary palate development. This hypothesis can be rationalized from paleontological evidence that suggests that acquisition of the bony secondary palate dates back 200 Mya [43]. Therefore, before the divergence of monotremes (184 Mya), other enhancer(s), the presence of which was suggested above, might have been responsible for the formation of the bony secondary palate of early mammals.
Because wnt5a-deficient mice reduced expression of msx1, bmp4, and shh in the anterior palate, wnt5a is considered to act upstream of these genes [12, 15]. Little is known, however, about the molecular mechanisms by which wnt5a expression is regulated. Msx1 is also associated with human non-syndromic cleft palate [33], and msx1-deficient mice have cleft secondary palate [34] as well as reduced expression of bmp4 and shh [35]. In the anterior PS, msx1 up-regulates bmp4 and vice versa, and bmp4 controls the downstream shh signaling that triggers PS growth [12, 35]. As we showed in the present study, it is remarkable that the Msx1-binding site in AS3_9 is necessary for its enhancer function (Fig 3J and 3J', S9J Fig), suggesting that wnt5a expression in the anterior palate is controlled by msx1. Therefore, taking the previous study suggesting that msx1 is one of the downstream genes of wnt5a into consideration [15], wnt5a and msx1 may have a synergistic effect on palatogenesis, as is the case with msx1 and bmp4 [35]. Therefore, palatogenesis is presumably controlled not by simple hierarchical signaling but rather by various interdependent cis-regulatory elements.
The present study showed that in the distal enhancer of wnt5a three TEs take their part cooperatively in palatogenesis (Fig 3). Notably, each of the X6b_DNA and MER117 regions of AS3_9 possesses a distinct tissue-specific enhancer property by itself (Fig 3D' and 3I'), indicating that different function can be evolved independently by multiple TEs located close to one another. In general, hundreds or more of TE types (subfamilies) constitute 20–50% of vertebrate genomes [44]. For example, it has been reported that several TEs contain motifs of functional sequences such as the CTCF-binding motif in rodent B2 SINE [45, 46] or in the MER20 DNA transposon [47], the Nfi-binding motifs in MER130 [48], and the OCT4-binding site in LTR7 or MER74A [49, 50], clearly demonstrating that certain TEs have the potential to acquire a function during evolution. Therefore, by multiple TEs being inserted close to one another, it is possible that they subsequently acquired a new regulatory function during evolution. We expect that many such coordinated TE-derived enhancers are hidden in mammalian genomes. To find clues that support this hypothesis, we searched the human genome for AmnSINE1 copies located proximal to other TEs, all of which overlap CNEs (S3 Table). Among the 626 conserved AmnSINE1 loci, 54 elements (8.6%) accompany other TEs that have been evolutionarily conserved, including all the major TE classes such as SINEs, LINEs, LTR-retrotransposons, and DNA transposons. Thus, the possibility arises that, at least at some loci among these 54 CNEs, several TEs located proximal to one another cooperate to modulate cis-regulatory networks that have been involved in the evolution of morphological innovations. This perspective extends the potential of TEs as genetic sources of a broader diversity of cis-regulatory elements. Further functional analysis of these TE-derived cis-regulatory elements will enhance our understanding of their involvement in morphological innovation during evolution.
The mouse strains B6C3F1, C57BL/6, and ICR were purchased from Sankyo Laboratory Service Corporation (Tokyo, Japan). Animals were kept in ventilated cages under a 12-h light/dark cycle at 24°C. This study was approved by the Ethics Committee of Tokyo Institute of Technology and Institutional Animal Care and Use Committee of RIKEN Kobe Branch.
A 2.1-kb DNA fragment of the mouse AS3_9 locus was amplified by PCR using primers AS3_9-F and AS3_9-R (S2 Table) containing Hind III recognition sites. The product was cloned into the Hind III site of plasmid HSF51 harboring the mouse heat-shock protein 68 promoter followed by the bacterial lacZ reporter gene and the SV40 poly-A signal, yielding the AS3_9-HSF51 construct.
The AS3_9 deletion constructs containing various combinations of TEs (Fig 3A–3I) were generated by overlap extension PCR as described [26]. Briefly, internal primers overlapping complementary sequences (S1 Table) were designed to carry out the deletion of each TE region. The first PCR was performed with one of the internal primers and either one of the vector primers (HSF51-F or HSF51-R; S2 Table) using AS3_9-HSF51 as template. The resulting PCR fragments were used as templates for the second PCR with the two vector primers, and the PCR products were cloned into HSF51 upstream of the heat-shock protein 68 promoter via Sal I and Hind III sites.
Transgenic mice were produced as described [23, 26]. Briefly, the constructs were linearized with Sal I and Xho I. After purification using the Gel Extraction kit from Qiagen, the DNA fragments were dialyzed against microinjection buffer (5 mM Tris-HCl, 0.1 mM EDTA) at 4°C overnight. Pronuclear microinjection was performed using 6–10 ng/μl of the DNA solution into a B6C3F1 zygote, and microinjected zygotes were transferred to the oviduct of pseudopregnant ICR females.
Transgenic mouse embryos were identified by PCR genotyping using primers LZgt-F02 and LZgt-R01 from yolk samples. The transgenic embryos were fixed for 1 h in phosphate-buffered saline (PBS) containing 1% formaldehyde, 0.1% glutaraldehyde, and 0.05% (v/v) NP-40 and then stained with PBS containing 500 μg/ml X-gal, 5 mM K3Fe(CN)6, 5 mM K4Fe(CN)6, 2 mM MgCl2, 0.02% NP-40, and 0.01% sodium deoxycholate for >3 h at 37°C. Consistency among lacZ expression patterns was confirmed by multiple microinjection experiments. For staining sections, the fixed embryos were permeated with 30% sucrose in PBS overnight at 4°C and embedded with O.C.T. compound (Tissue-Tek, Sakura, Torrance, CA) for sectioning. Coronal sections produced with a cryostat (Leica CM 1850) were counterstained with kernechtrot (Nuclear Fast Red). In addition to the analysis of transient transgenic embryos, we generated three stable lines derived from AS3_9-lacZ transgenic mice (S3B Fig). The E9.5–15.5 embryos of each heterozygote transgenic mouse were harvested and stained with X-gal solution as described above.
Mouse embryos for ISH for wnt5a and lacZ were prepared from ICR and stable transgenic lines, respectively. E11.5 embryos for whole-mount ISH were fixed overnight in 4% paraformaldehyde dissolved in diethylpyrocarbonate-treated PBS at 4°C. For section ISH (E13.5–15.5), embryos were permeated overnight with 30% sucrose dissolved in diethylpyrocarbonate-treated PBS after fixation. Then, the embryos were embedded with O.C.T. compound and frozen. Sections were prepared on a Leica sledge microtome at 14 μm and individually mounted on slides. Digoxigenin-labelled antisense RNA probes were synthesized from linearized plasmids with T3 and T7 polymerase (Roche, Basel, Switzerland). Respective plasmids carrying subcloned coding regions of wnt5a and lacZ were prepared. ISH was performed as described [51].
ISH was also performed on E13.5 coronal cryosections as described [25] to examine the expression of the AS3_9-proximal genes. Briefly, eight genes (wnt5a, erc2, lrtm1, caca2d3, selk, actr8, il17rb, and chdh) surrounding AS3_9 were identified with the UCSC Genome Browser (http://genome.ucsc.edu/). The cDNAs were amplified by PCR from an E14 or E17 cDNA pool, cloned into pGEM T-easy vectors (Promega, Madison, WI, USA), and used for probe syntheses. Probes prepared by in vitro transcription using the DIG RNA Labeling kit (Roche) were purified by lithium chloride precipitation and used for ISH as described [25].
For the ISH of the AS3_9-ko mice for the wnt5a probe, the plasmid containing the mouse wnt5a cDNA (1.4 kb) was generated via PCR with the primers in S2 Table and sub-cloned into pBluescript (KS-). The embryos were fixed, embedded in Paraplast, and serially sectioned (10 μm thickness). Sections were subjected to ISH as described [51, 52].
The AS3_9 sequence (chr14:29028538–29029337 of GRCm38/mm10), including the three TEs, was targeted (S6A Fig). Two arm fragments (7.1 kb of the long arm and 3.6 kb of the short arm) were amplified using an LA PCR kit (Takara, Japan) with the following primers: AS3_9-LongArm-F2 and AS3_9-LongArm-R2 for the long arm, and AS3_9-ShortArm-F and AS3_9-ShortArm-R for the short arm (S2 Table). The long-arm fragment was sequenced and cloned into the 5’ (Not I and Sal I) cloning site of the PGK-Neo-pA cassette in the Targeting vector (DT-A-pA/loxP/PGK-Neo-pA/loxP; see http://www2.clst.riken.jp/arg/cassette.html for details); the short arm was also sequenced and cloned in the 3’ (Xho I) site in the same vector. Homologous recombination was conducted using TT2 embryonic stem cells [53], and two recombinants were used to produce chimeric AS3_9-deficient mice (#49 and #154)
For Southern hybridization, DNA probes for the 5’ and 3’ regions of the targeted sequence as well as the Neo sequence were amplified with specific primers (S2 Table) and labelled with [α-32P]dCTP. Genomic DNA (10 μg each) from F1 individuals was digested with Bln I and Bsp1407 I (for the 5’ probe), EcoR V (for the 3’ probe), or Sac I (for the Neo probe). Signals of 12.0, 6.2, and 6.3 kb for the mutant alleles were detected by Southern blotting using the 5’, 3’, and Neo probes, respectively (S6C–S6E Fig). PCR genotyping was performed with AS3_9-gtF1(KO) and AS3_9-gtR for detection of the knockout allele and AS3_9-gtF1(WT) and AS3_9-gtR for the wild-type allele (S6B Fig; S2 Table).
The Matchmaker Gold Yeast One-Hybrid Library Screening System kit (Clontech, Palo Alto, CA, USA) was used to identify proteins that could bind to the AS3_9 sequence. A SMART cDNA library was constructed using the kit from the frontonasal tissues of 21 mouse embryos at E14.5. The bait DNA (chr14:29028539–29029109 of GRCm38/mm10) consisted of AmnSINE1 and X6b_DNA regions because the sequence is responsible for the enhancer activity in maxillary processes, the origin of PS outgrowth. Yeast one-hybrid screening was conducted with two consecutive selection steps with 100 ng/ml (first step) and 500 ng/ml (second step) of the antibiotic aureobasidin A. Among 11.4 million clones of the library screened, 14 positive clones (S1 Table) were isolated consisting of 12 genes in total, which included craniofacial developmental genes (Msx1, Msx2, and Gtf2ird1). Mutation in the Msx1-binding site of the bait sequence was introduced (TAATTGG -> gccgTGt) using appropriate primers (S2 Table), and the yeast one-hybrid assay was performed according to manufacturer's protocol with aureobasidin A (250 ng/ml).
We performed an in silico screen of the human genome (hg38) to identify AmnSINE1-derived CNEs proximal to other TE-derived CNEs. By comparing a list of conserved elements (phastConsElements100way) from the UCSC genome database and the latest TE annotation list by RepeatMasker (with the repeat library 20140131; http://www.repeatmasker.org/species/hg.html), all TEs overlapping >30 bp with the conserved elements (LOD score >100) were extracted. Among the 626 evolutionarily conserved AmnSINE1 sequences found, those proximal to (<600 bp) another TE-derived CNE were collected and listed in S3 Table.
Details of the AS3_9-ko lines are available at http://www2.clst.riken.jp/arg/mutant%20mice%20list.html (Accession No. CDB0941K).
|
10.1371/journal.pntd.0007569 | Mortality by cryptococcosis in Brazil from 2000 to 2012: A descriptive epidemiological study | Cryptococcosis is a neglected and predominantly opportunistic mycosis that, in Brazil, poses an important public health problem, due to its late diagnosis and high lethality.
The present study analysed cryptococcosis mortality in Brazil from January 2000 to December 2012, based on secondary data (Mortality Information System/SIM-DATASUS and IBGE).
Out of 5,755 recorded deaths in which cryptococcosis was mentioned as one of the morbid states that contributed to death, two distinct groups emerged: 1,121 (19.5%) registered cryptococcosis as the basic cause of death, and 4,634 (80.5%) registered cryptococcosis associated with risk factors, mainly AIDS (75%), followed by other host risks (5.5%). The mortality rate by cryptococcosis as the basic cause was 6.19/million inhabitants, whereas the mortality rate by cryptococcosis as an associated cause was 25.19/million inhabitants. Meningitis was the predominant clinical form (80%), males were the more affected (69%), and 39.5 years old was the mean age. The highest mortality rate due to cryptococcosis as basic cause occurred in the state of Mato Grosso (10.96/million inhabitants). Mortality rates due to cryptococcosis as associated cause were highest in the states of Santa Catarina (70.41/million inhabitants) and Rio Grande do Sul (64.40/million inhabitants), both in the South Region. Southeast, Northeast and South showed significant time trends in mortality rates.
This study is relevant because it shows the magnitude of cryptococcosis mortality linked to AIDS and removes the invisibility of a particular non-AIDS-related disease, accounting for almost 20% of all cryptococcosis deaths. It can also contribute to control and surveillance programs, beyond highlighting the urgent prioritization of early diagnosis and proper treatment to reduce the unacceptable mortality rate of this neglected mycosis in Brazil.
| Cryptococcosis is an invasive, global, and neglected mycosis. Species of the Cryptococcus neoformans complex cause opportunistic infections in immunosuppressed hosts, particularly AIDS patients, while infections by species of the C. gattii complex predominate as a primary endemic mycosis in tropical and subtropical areas. In Brazil, it is an important and hidden public health problem, mainly in its meningitis form, whose lethality ranges from 45 to 65%, but remains as a not-notifiable disease. Brazilian studies placed it as the first cause of mortality among all AIDS-associated systemic mycoses and the second cause of mortality among systemic mycoses in general. This national study used data from the Brazilian Mortality Information System (SIM) in which all mentioned causes were considered, allowing the analysis of the associated causes of deaths and showing two different patterns of infections: a primary and an opportunistic cryptococcosis. Primary cryptococcosis presented a peculiar epidemiological regional profile and the opportunistic cryptococcosis was hidden by several immunosuppressive conditions. The authors expect that this study can support a better understanding of this infection and encourage more research and public health initiatives to prevent and control the cryptococcosis, both primary and opportunistic.
| Encapsulated yeasts of the Cryptococcus neoformans/Cryptococcus gattii species complexes are the causative agents of cryptococcosis, a systemic mycosis of humans and animals, acquired by inhalation of their spores—desiccated yeast cells or basidiospores—from the environment [1,2]. Although usually regressive, some cases develop cryptococcal lung injury, which can spread to other sites or organs. On reaching the central nervous system, it may cause meningoencephalitis, the most severe form of cryptococcosis that, without early diagnosis and proper treatment, is highly lethal or disabling [3,4].
Cryptococcus neoformans infection predominates in immunocompromised hosts, being globally a threat to people living with HIV/AIDS, causing approximately 15% of AIDS-related annual mortality [5,6]. Cryptococcosis by C. gattii occurs mainly in otherwise immunocompetent hosts, but some immune deficiency not detected by routine tests may predispose individuals to this infection [7,8].
It is estimated that more than 300 million people worldwide, of which about 3.8 million in Brazil, suffer from a serious fungal infection every year, resulting in more than 1,350,000 deaths [9,10]. Among these diseases is cryptococcosis, with an overall incidence varying from 0.04 to 12% per year among people with HIV [5].
The global incidence of cryptococcosis in people living with HIV/AIDS in 2008 was estimated in approximately 1 million meningitis cases annually (range 371,700–1,544,000) causing around 625,000 deaths [5]. The highest number of yearly cases was estimated to occur in sub-Saharan Africa (720,000), followed by South–East Asia, and Latin America as the second and third regions with the most cases of cryptococcal meningitis (54,400) [5].
Since then, due to extensive antiretroviral therapy (ARVT) expansion, AIDS-related deaths have been reduced by 45% [6]. In 2014, the global incidence cases of cryptococcal meningitis was estimated at 223,100 (95% CI 150,600–282,400) and the annual global deaths were estimated at 181,100, with 135,900 (75%; [95% CI 93,900–163,900]) deaths in sub-Saharan Africa [6]. Latin America’s annual burden of cryptococcal meningitis estimate was 5300 (2600–8900 interval) and deaths from cryptococcal meningitis were 2400 (1100–4400) [6]. But, even so, cryptococcosis is not on the WHO neglected tropical diseases list [5,11].
Besides the well-known outbreak in North America [12,13], cryptococcosis by C. gattii presents a peculiar epidemiological profile in South America, especially in Brazil, where it is endemic in large areas of the Amazon region and the semi-arid Northeast region [14–19].
However, data available on cryptococcosis in Brazil is fragmented and circumscribed, mostly based on indirect data on AIDS programs and some based on analyses of series of cases, diagnosed in regional centres. According to studies regarding mortality related to systemic mycoses in the nationally, cryptococcosis is the second cause of mortality among them [20]. Moreover, cryptococcosis is highlighted as the most frequent among the systemic mycoses associated with AIDS [21], assuming its essentially opportunistic character.
The cryptococcosis lethality rate in Brazil is substantial, reported in the range of 45% to 65% [22], independent of the presence of risk factors, dominated by association with AIDS, as well as the primary form of the disease. A different scenario is seen in developed countries, as for example in Canada, in non-HIV hosts, where the diagnosis of pulmonary forms is more frequent than meningitis, the overall lethality is about 8% and there is a control program and surveillance for primary cryptococcosis [23].
Cryptococcosis is a major public health problem in Brazil, most cases are diagnosed as central nervous clinical forms, mainly meningitis. Only a few cases are diagnosed in a pulmonary form, which usually disseminates to meningoencephalitis, increasing hospitalizations and lethality. Late diagnosis of cryptococcosis slows crucial therapeutic measures to reduce sequelae and avoid lethal outcomes. Nevertheless, cryptococcosis is not a reportable disease in Brazil, and the real magnitude of its mortality is unknown [15].
In order to improve epidemiological surveillance, regional strategies and priorities for early diagnosis and treatment of cryptococosis in HIV as well as in non-HIV groups, this study aims to characterize the mortality by cryptococcosis as a health problem with a diverse geographical pattern in Brazil. This paper shows the magnitude of cryptococcosis mortality and points to cryptococcosis as a severe and often fatal neglected mycosis in Brazil. The vast majority of deaths are hidden by several immunosuppressive conditions.
This is a descriptive epidemiological Brazialian study, based on secondary data for the period 2000 to 2012, covering a historical series of 13 years.
The study was approved by the Ethics Research Committee of the Sérgio Arouca Brazilian National School of Public Health, number 37353614.5.0000.5240.
The research used secondary data from the DATASUS/Ministry of Health (MS) Mortality Information System (SIM) and the Brazilian Institute of Geography and Statistics (IBGE). Therefore, the individuals whose information was extracted were not identified individually. Furthermore, there was no direct intervention with the patient and / or relatives, ensuring anonymity.
The DATASUS/Ministry of Health (MS) SIM is the official source of death data for infectious and parasitic diseases (IPDs). SIM compounds the National Epidemiological Surveillance System (NESS), providing data about deaths in Brazil through information registered on death declaration (OD), including basic and associated cause, based on the 10th International Classification of Disease (ICD). This data is collected by Municipal Health Secretaries (MHS) and registered in a national database and available for consultation. SIM data collection methodology did not change during the study period. Demographic data of the population and cartographic bases of the Brazilian federal units and regions were obtained from IBGE.
The following variables were considered: cryptococcosis as basic or associated cause of death, gender, age, and place of residence. Data was distributed and analyzed according to country, regions and states. Deaths were studied according to their frequency by place of residence and their temporal and spatial distribution, estimating mortality and trend coefficients and analyzing their geographical distribution.
Basic cause of death was defined as a disease or condition that initiated the chain of pathological events that led directly to death. Associated cause of death was defined as a pathological condition that had an unfavourable effect and contributed to death, mentioned in the death certificate. The classification between basic or associated cause was attributed by the physician who completed the death certificate. Only recently, data on deaths according to multiple causes is available in the mortality database.
The mean mortality rate was estimated taking as numerator the number of basic cause of death by cryptococcosis at specific locations during the study period (2000–2012). The utilized denominator was the mean size of the Brazilian population, in the same period, multiplied by 1,000,000 inhabitants. The same methodology was used for cryptococcosis as an associated cause. The mean mortality rate for all mentions was also estimated in the death certificates, that is, by the sum of both conditions above.
To highlight the particularity of cryptococcosis, the total number of times cryptococcosis was mentioned, either as the basic or associated cause, that is, the total number of mentions among the diseases that contributed to death, was used. The ratio was then estimated by dividing the frequency of cryptococcosis as a mentioned cause by frequency as the basic cause (ratio: total mentions/basic cause) [24].
In order to analyze association between gender and associated or basic cause, we used a chi-square test, with significance level of 5%. We used a Poisson model with offset term to model the mortality rate by cause (associated or basic), age groups and gender. The incidence density ratios (IDR) and 95% confidence intervals (95% CI) were obtained from this model.
The information on mortality by cryptococcosis with reference to each region or federated unit was analyzed according to its geographic distribution and presented through tables and thematic maps.
We analyzed the time trends of mortality rate by Joinpoint analysis for basic and total cause of death. For this, we modelled the rates by Poisson model with quasilikelihood estimation, in order to solve the overdispersion problem. After, we used a segmented regression to determine the breakpoints in which we observed a significant change in trend of mortality rate. The Annual Percentage Change (APC) in each trend was obtained, with a 95% confidence interval (CI). Graphs of mortality rates observed (squared points) and of mortality rates predicted by the Poisson segmented regression (lines) were provided.
Tabwin, Microsoft Excel 2010, R 3.5.1 and package segmented and QGIS were used to obtain the database, tabulation, trends and graphing.
From 2000 to 2012, a total of 5755 deaths were recorded in Brazil in which cryptococcosis was mentioned. Of these, cryptococcosis was recorded as the basic cause of death in 1121 deaths (19.5%), representing a mean mortality rate of 6.09/ million inhabitants. The remaining 4634 (80.5%) deaths from cryptococcosis were recorded as an associated cause with a mortality rate of 25.19/million inhabitants. Male deaths were more common in both the basic and associated causes (Table 1).
The frequency rate of basic cause (mentions/basic cause) was 5.13 (5755/1121).
Of the 4314 cases associated with AIDS, 71.5% of deaths occurred in males, prevailing in the age range of 20 to 59 years old, accounting for 95.8% (n = 4133) of the deaths. In the group of other risk factors (n = 320), males represented 66.9% of the deaths (Table 1). The IDR found corroborates the increased risk of death in males, the age group of 20 to 59 years and associated cause (Table 2).
Among deaths of those younger than 20 years of age, (2.9% of the total), cryptococcosis mentioned as basic cause accounted for 6.7% (n = 76), and 5.3% (n = 17) of deaths by cryptococcosis due to other risk factors, excluding HIV+, as compared to 1.5% (n = 68) of cryptococcosis AIDS-related deaths. In the basic cause group, cryptococcosis deaths among those older than 60 represent 51% of total mentions in this age group (225/445) and among those younger than 20 years old, represent 47% of total mentions in this group (76/161) (Table 1).
Several known immunosuppressive conditions were recorded as basic cause in 80% (n = 4634) of the deaths where cryptococcosis was mentioned as associated cause. AIDS was the major immunosuppressive disorder with 75% (n = 4314 deaths), followed by other immunodeficiency conditions or risk factors with 5.5% (n = 320) of deaths: non-Hodgkin lymphoma (27), unspecified immunodeficiency (17), lymphoid leukemia (13), chronic renal failure (12) and other causes (251), reflecting the opportunistic face of this mycosis.
All clinical presentations of registered cryptococcosis have pointed to a severe disease, especially cryptococcal meningitis. Cerebral cryptococcosis–ICD (International Classification of Diaseases) B45.1 - (cryptococcal meningitis) predominated as by far the most frequent form, with 4743 deaths (82.4%) of the total mentions. In the AIDS group, this form occurred in 83.6% (3609) of deaths, whereas where cryptococcosis was the underlying cause of death, it was 79.9% (895). It is worth noting that the pulmonary form was diagnosed with cryptococcosis as a basic cause of death in 65 cases (5.8%), when associated with other risk factors 18 (5.7%) and when associated with AIDS it was recorded only in 31 deaths (0.7%) (Table 3).
The distribution of deaths and the mean mortality rate by other infectious meningitides, according to the basic cause, were also analyzed in order to assess the relevance of the central nervous system in cryptococcosis among the other meningitides.
During the study period, 21,333 meningitis deaths occurred, with a mortality rate of 115,97/million inhabitants. Among meningitis with specified cause, the meningococcal etiology was responsible for 8.6% (1,830), with a mortality rate of 9.95/million inhabitants, being the most frequent, followed by cryptococcal meningitis, with 895 deaths (4.2% of the total) and mortality rate of 4.87/million inhabitants. Also relevant were toxoplasma meningitis with 806 deaths (3.8%) and a mortality rate of 4.38/million inhabitants; viral meningitis with 753 deaths (3.5%) and a mortality rate of 4.09/million inhabitants; and tuberculous meningitis with 624 deaths (2.9%) and a mortality rate of 0.26/million inhabitants. Meningites of unknown cause were included as “other meningites” (Table 4).
In the same period, there were 608,314 deaths from other infectious diseases listed in Chapter 1 from ICD 10. Cryptococcosis was the thirteenth cause of death between chronic and recurrent infectious disease, 1121 by basic cause. The proportion of cryptococcal deaths in the study period compared to the other infectious diseases was 0.18% (S1 Table).
Deaths from cryptococcosis were recorded in all Brazilian states, but their distribution was not homogeneous. Thematic maps show the geographic profile of cryptococcosis mortality rates in the period, as basic cause as well as an associated cause of death (Table 5) (Figs 1 and 2).
Total mentioned cause of death by cryptococcosis shows that the South Brazilian region has the highest rates, followed by the Midwest and Southeast. The North and Northeast had the lowest rates (Fig 3). The Southeast, Northeast and South showed significant time trends in mortality rates (S2 Table). The Southeast region showed a decreasing trend of mortality rate (-4.82%) in years 2000–2006. The Northeast region showed 105.20% of increasing trend in 2001 and 10.35% of increasing between 2005 and 2012. Meanwhile, the South region showed decreasing trends of mortality rates in 2000–2005 (-2.89%) and 2009–2012 (-7.27%).
The basic cause of death by cryptococcosis show that the Northeast region had the lowest rates (Fig 4). The North and Northeast showed significant time trends in mortality rates (S3 Table). The North region showed a decreasing trend of mortality rate (-64.28%) between years 2009–2010. The Northeast region showed 12.42% of increasing trend between 2000 and 2008 and 43.92% of increasing between 2009 and 2012.
This study points to cryptococcosis as a neglected, severe and often fatal opportunistic condition, since the vast majority of deaths (80%) is hidden by a serious immunosuppressive disease, especially AIDS. In fact, Two patterns of infection were revealed: 1) primary cryptococcosis and 2) opportunistic cryptococcosis, both expressed mainly in the form of meningoencephalitis, indicating late diagnosis, ineffective treatment and difficult access to the national care network.
The study of cryptococcal mortality considering only the basic cause presented limitations, since the presence of underlying immunodeficiencies predominates in the scenario. Thus, when the total causes mentioned in the death certificates was considered, a broad picture of the mycosis in Brazil was revealed, leading to an important reflection on the neglected diseases associated with a host with immunodeficiency. Eighty percent of cryptococcosis deaths were revealed through this approach [25]. In addition, the poorer regions of the north and northeast of Brazil still have high proportions of deaths due to ill-defined causes, which may hide both cryptococcosis and other AIDS-related infectious causes of death [25–27].
The distribution of cryptococcosis deaths according to gender and by all mentioned causes showed a preponderance among males. When associated with AIDS (71.2%), crypto mortality was greater than the mortality due to cryptococcosis as basic cause (62.3%), which corresponds with data from the literature [16, 21].
We observed in primary cryptococcosis an age-matched progressive curve from childhood to adulthood, consistent with progressive environmental exposure to the agent. However, the age-related pattern of cryptococcosis associated with AIDS reflects the predominance of this risk factor, specially in the age group of 20 to 59 (Tables 1 and 2).
In the over 60 age group within the total number of deaths an important differential was also revealed: AIDS-related deaths accounted for 2.7% of the total, while deaths due to cryptococcosis as basic cause represented about 20%, i.e. about seven times higher and, deaths due to cryptococcosis associated with other risk factors represented 33.6%, that is, twelve times higher. This set of evidence seems to corroborate the double profile of cryptococcosis (Table 1).
This age-related profile is consistent with recent reports showing individuals affected by cryptococcal meningitis caused by C. gattii, with high lethality rates (30 up to 50%) and frequent relapses in the North and Northeast regions of Brazil, along with increased frequency of cryptococcosis in AIDS in young male adults, the involvement of immunocompetent children, adolescents and young adults and the involvement of elderly individuals [15,16,28,29].
The mortality rate of infectious diseases which are difficult to diagnose and that require specialized care, usually expresses the tip of an iceberg. This study hypothesized cryptococcosis as an underestimated causa mortis, because the laboratory resources for timely agent identification and with the needed accuracy are often unavailable. The lack of extensive diagnostic laboratory coverage is evident, given that 75% of all meningitis-related deaths had no defined etiology [27].
In this study, the vast majority of cryptococcosis deaths, according to the total number of mentions, was due to cryptococcal meningitis (82.4%), and when associated with AIDS, caused 83.6% of deaths. In Brazil, studies have shown the relevance of cryptococcosis as the main mycosis associated with AIDS death [21], and as the second cause of mortality among systemic mycoses [20]. Cryptococcal meningitis was the second most frequent opportunistic neurological infection in HIV/AIDS [28–30], only surpassed by neurotoxoplasmosis [31–33].
Further, between 1980 and 2002, about 13,000 individuals had cryptococcosis at the time of diagnosis of HIV infection, six percent of the 215,810 registered cases of AIDS in Brazil [22]. Brazilian autopsy studies of the Central Nervous System of AIDS patients showed high cryptococcal involvement: 12% (17/138), 13.5% (34/252) and 15.8% (45/284), respectively [31–33].
The high lethality of cryptococcal meningitis in Brazil results from the convergence of factors such as late suspicion and diagnosis, difficult access to care network, unavailability of rapid laboratory tests, together with inadequate or unavailable antifungals. The screening of Cryptococcal Antigen (CrAg) in HIV infected persons with CD4 count below 100 cells/mm3 is highly recommended by the WHO [34–36]. According to international advised protocol to reduce mortality by cryptococcosis, another important issue regarding treatment is to associate 5-flucytosine with amphotericin as a combined initial therapy [37,38], but is as yet unavailable in the national therapeutic arsenal, despite institutional efforts to import this drug [39].
It is worth noting that the mortality rate related to cryptococcal meningitis was higher than that of Toxoplama CNS infection (neurotoxoplasmosis), as well as higher than meningitis caused by all viral infections and by tuberculosis. In Africa, cryptococcal meningitis is the most common cause of meningitis in adults [6]. In the US, cryptococcal meningitis hospitalizations were more frequent than all bacterial meningitides combined, with an incidence of 1.1 per 100,000 inhabitants versus 0.728 per 100,000 respectively [38].
This study detected two patterns for cryptococcosis in Brazil: the first, a primary cryptococcosis drawn by deaths recorded as the basic cause, an emerging disease and the second, a cryptococcosis registered as an associated cause of death, an opportunistic infection affecting individuals who present some immunodepression, mainly AIDS-related patients [8,40].
C. gattii species complex occurs in tropical, subtropical and temperate areas, affecting mainly apparently healthy hosts in contact with environmental sources of infection [8]. C. neoformans complex is cosmopolitan and affects mainly individuals who present some immunodepression [8,40]. The geographic distribution of C. gattii in Brazil shows a higher prevalence in the North and Northeast regions compared with the other regions [18,19], while C. neoformans is more prevalent in the South and Southeast regions [14].
In our study, we did not find significant differences between cryptococcal deaths in the North and Northeast regions, but we found a great difference in the South, Southeast and Central-West regions, where crypto deaths as associated cause were more frequent than basic cause. Furthermore, the majority of individuals infected by HIV in Brazil were concentrated in the South and Southeast regions [41], reinforcing the two profiles of cryptococcosis: the South, Southeast and Central-West with predominant opportunistic infection by C. neoformans and the North and Northeast with, side by side, the opportunistic infection by C. neoformans and the primary infection by C. gattii [14,18,19].
The geographic distribution and joinpoint analysis show that the highest mortality rates due to cryptococcosis reported as basic cause was observed in the North, folowed by the Central West and the South. The state of Mato Grosso, Pará, Mato Grosso do Sul, Amazonas and Santa Catarina showed the highest rates by state. These regions are economically heavily based on agricultural activity. The North and Central West are the new Brazilian agricultural frontiers with intensive population mobility [42,43]. As previously pointed out, recent studies documented the presence of an endemic primary cryptococcosis in the Amazon region, the north, and northeast of Brazil [15–17,19].
The geographic distribution of mortality rates due to cryptococcosis as associated cause evidenced that the highest mortality rates occurred in the most economically dynamic regions of the country. These rates occurred in the South, Central West and Southeast. The highest rates were reported in the states of Santa Catarina, Rio Grande do Sul and Mato Grosso do Sul. This distribution is analogous to the distribution of AIDS deaths in the period, which it is also consistent with the interiorization spreading of the AIDS epidemic in Brazil [41].
The major limitation of this study is in relation to the use of secondary data, which underestimates the true number of deaths related to neglected diseases. The lack of specialized laboratories and medical resources in the poorest regions of the country result in a large number of deaths of indeterminate cause. Furthermore, the SIM does not have access to medical records, only to diagnoses reported on death certificates. Therefore, it is impossible to know how the diagnosis of cryptococcosis was made.
This study is the first one to apply a holistic approach to cryptococcosis mortality in Brazil. It provides needed visibility to cryptococcosis, revealing two distinct profiles, one primary and the other opportunistic associated mainly with AIDS.
The high frequency of deaths by cryptococcosis meningitis and other severe clinical presentations indicates late diagnosis, unavailability of rapid diagnostic methods, lack of effective antifungal treatments, and difficult access to the care network in the country.
This study can support surveillance and improvement actions aimed at preventing many avoidable deaths by this neglected systemic mycosis.
|
10.1371/journal.pgen.1000757 | Homeobox Transcription Factors Are Required for Conidiation and Appressorium Development in the Rice Blast Fungus Magnaporthe oryzae | The appropriate development of conidia and appressoria is critical in the disease cycle of many fungal pathogens, including Magnaporthe oryzae. A total of eight genes (MoHOX1 to MoHOX8) encoding putative homeobox transcription factors (TFs) were identified from the M. oryzae genome. Knockout mutants for each MoHOX gene were obtained via homology-dependent gene replacement. Two mutants, ΔMohox3 and ΔMohox5, exhibited no difference to wild-type in growth, conidiation, conidium size, conidial germination, appressorium formation, and pathogenicity. However, the ΔMohox1 showed a dramatic reduction in hyphal growth and increase in melanin pigmentation, compared to those in wild-type. ΔMohox4 and ΔMohox6 showed significantly reduced conidium size and hyphal growth, respectively. ΔMohox8 formed normal appressoria, but failed in pathogenicity, probably due to defects in the development of penetration peg and invasive growth. It is most notable that asexual reproduction was completely abolished in ΔMohox2, in which no conidia formed. ΔMohox2 was still pathogenic through hypha-driven appressoria in a manner similar to that of the wild-type. However, ΔMohox7 was unable to form appressoria either on conidial germ tubes, or at hyphal tips, being non-pathogenic. These factors indicate that M. oryzae is able to cause foliar disease via hyphal appressorium-mediated penetration, and MoHOX7 is mutually required to drive appressorium formation from hyphae and germ tubes. Transcriptional analyses suggest that the functioning of M. oryzae homeobox TFs is mediated through the regulation of gene expression and is affected by cAMP and Ca2+ signaling and/or MAPK pathways. The divergent roles of this gene set may help reveal how the genome and regulatory pathways evolved within the rice blast pathogen and close relatives.
| Pathogens have evolved diverse strategies to cause disease. Magnaporthe oryzae is the fungal phytopathogen that causes rice blast and is considered an important model for understanding mechanisms in fungal development and pathogenicity. Asexual reproduction and infection-related development play key roles in M. oryzae disease development. The conidium of M. oryzae differentiates a specialized structure, an appressorium. The appressorium generates turgor pressure that allows penetration through the mechanical rupture of host cuticle layers. After colonizing host cells, the fungus produces massive conidia via conidiogenesis, serving as secondary propagules for the polycyclic disease. To elucidate molecular mechanisms in asexual reproduction and appressorium-mediated disease development, we identified eight homeobox transcription factors through a genome-wide in silico analysis. Characterization using deletion mutants revealed that each homeobox TF functions as a stage-specific regulator for conidial shape, hyphal growth, conidiation, appressorium development, and invasive growth during M. oryzae development. Notably, conidiation and appressorium development were entirely abolished in ΔMohox2 and ΔMohox7, respectively. This study also provides evidence that M. oryzae is able to cause rice blast by means of hypha-driven appressoria upon responses to host signaling factors. This study will aid in the understanding of regulatory networks associated with fungal development and pathogenicity.
| Magnaporthe oryzae is an ascomycete fungus and the causal agent of rice blast, the most destructive disease of rice worldwide. The annual yield loss from rice blast is equivalent to rice that could feed about 60 million people [1]. Rice blast has served as an important model for studying molecular plant-pathogen interactions because of its economic significance and genetic tractability of the host and pathogen. More recently, the availability of the genome sequences of both rice and the fungal pathogen has provided a new platform to understand molecular pathogenesis at the genome level [2]–[4].
Like most fungal pathogens, conidia (asexual spores) of M. oryzae play a central role in the disease cycle. The conidia attach to the surface of host plants upon hydration and produce germ tubes. This fungus develops a specialized infection structure, an appressorium, at the tip of the germ tube. The appressorium generates enormous turgor pressure (>8 MPa) by accumulating osmolytes including glycerol for penetration through the mechanical rupture of host cell barriers [5]. After penetration, the fungus develops invasive hyphae, colonizes host cells, and produces massive conidia via conidiogenesis, serving as secondary inocula for new infections. This infection cycle may occur many times during the growing season, resulting in explosive disease development. Therefore, understanding the molecular mechanisms involved in conidiation and appressorium development is a prerequisite to provide novel strategies for disease management.
During the last couple of decades, considerable progress has been made in understanding signaling pathways that regulate the infection-related development of this fungus. These include the mitogen-activated protein kinase (MAPK) signaling cascade [6],[7], signaling pathways dependent on secondary messengers such as Ca2+ and cAMP [8]–[13], and G-protein-mediated signaling pathways [14],[15]. For example, deletion of genes involved in cAMP and Ca2+ signaling pathways has revealed that both are required not only for infection-related fungal development but also for pathogenicity [16],[17]. Disruption of Gα subunits and the MAP kinase gene also indicated the involvement of G-protein and MAP kinase signaling in vegetative growth, sexual reproduction, and pathogenicity in M. oryzae [15]. Most studies have focused on well-known upstream signaling pathways, but relatively little information is available on the downstream regulators of appressorium development.
Conidiogenesis is a complex process that involves a cascade of morphological events. M. oryzae produces three-celled conidia on a conidiophore, a specialized structure elongated through apical extension of an aerial hypha. Unlike vegetative hyphae, conidiophores rarely branch and their growth is modestly determinate. Several conidia, mostly three to five are arrayed at the tip of a conidiophore in a sympodial pattern after the occurrence of rounds of mitosis. The molecular biology of conidiation has been characterized in detail for Aspergillus nidulans and Neurospora crassa [18],[19]. The transcription factor (TF) genes brlA, abaA, and wetA are key regulators in the central regulatory pathway of A. nidulans conidiation, which coordinately regulates the order of gene activation during conidiophore development and spore maturation. Several other genes, FlbB, FlbC, FlbD, and FlbE, act as early regulators that activate the central regulatory pathway in response to the product of FluG activity [19]. In N. crassa, conidium-specific con genes have been described [20]. The fl gene, which encodes a TF, and genes (frq, wc-1, and wc-2) that act in the Neurospora circadian clock regulate the morphological transition from filamentous growth to conidiation [18],[21].
Relatively little information exists on conidiation in M. oryzae despite its central role in the disease cycle. Deciphering the conidiation pathways may reveal key determinants of initiation and the progress of conidiation, which may provide potential targets for disease control. A few genes are involved in conidial morphology in M. oryzae. Mutations at the SMO locus cause abnormally shaped conidia [22]. A mutation of the Acropetal gene causes head-to-tail arrays of conidia [23]. Deletion of the CON7 gene encoding a zinc finger TF also causes abnormal conidia with less septa and/or protuberances at the basal scar [24]–[26]. In addition to genes involved in conidial ontology, a recent study showed that a zinc finger TF-coding gene, named COS1 for conidiophore stalk-less1, is indispensible for an early stage of conidiophore development [27].
Transcriptional regulation is a major mechanism by which alterations in the expression of specific subsets of genes determine development and differentiation in cells. DNA-binding TFs play a pivotal role in the transcriptional regulation of specific target genes necessary for such processes in response to physiological or environmental stimuli. A comparative genome-wide analysis revealed that a variety of TFs are abundantly present in metazoans, including fungi [28]. M. oryzae appears to possess over 400 TF genes, but only a few have been characterized [10],[24],[27],[29],[30]. Homeobox TF genes contain highly conserved sequences coding for the DNA-binding motif called the homeodomain. This group of homeobox TFs was first discovered in Drosophila melanogaster in which they specify the body plan along the antero-posterior axis [31],[32]. Numerous genes for homeobox TFs have since been identified in eukaryotes, including fungi. Several studies have established the involvement of homeodomain proteins in mating and sexual differentiation in fungi [33]–[37]. A gene encoding a homeodomain protein also controls hyphal morphology and conidiogenesis in Podospora anserina [38]. In Ustilago maydis, homeobox TFs regulate the hyphal growth, pathogenicity and sexual cycle [39]. It is therefore evident that homeobox TFs play important regulatory roles in morphogenesis and pathogenesis in plant-pathogenic fungi.
As a first step in deciphering the biological functions of TF genes in M. oryzae, we comprehensively searched currently available sequences at the genome-wide level for the existence of homeobox TFs in the fungal kingdom. This analysis revealed that a total of eight genes encoding putative homeobox TFs exist in M. oryzae, which were here named MoHOX1 to MoHOX8. Similarly, other fungal species appear to possess multiple copies of these genes. To further characterize the roles of MoHOXs in M. oryzae biology, deletion mutants of each MoHOX gene were generated through a homology-dependant gene replacement strategy. Analyses of the various transformants including ΔMohox mutants demonstrated that homeobox TFs function as stage-specific regulators in fungal development and pathogenicity in M. oryzae. In particular, MoHOX2 and MoHOX7 are indispensable for conidiation and appressorium development, respectively. Our findings would provide new insight into the transcriptional regulation of infection-related morphogenesis at the genome level.
Members of the homeobox TF family possess a conserved DNA-binding motif called the homeodomain [40]. Using InterPro terms (IPR001356 and IPR003120) for homeodomains, 216 homeobox TFs were retrieved from 22 eukaryotic microbe genomes, including eight (MoHOX1 to MoHOX8) in M. oryzae (Table S1). The MoHOX genes separated into eight distinct clades in the phylogenetic tree (Figure 1). This suggests that expansion of the DNA-binding TF family may be linked to functional divergence, as hypothesized previously [41]. The MoHOX-clades, except the MoHOX8-clade, which contains a divergent form of the homeodomain, embrace homeobox TFs belonging only to the subphylum Pezizomycotina, but not to the subphylum Saccharomycotina or phylum Basidiomycota, suggesting that homeobox TFs have evolved in a lineage-specific manner [41],[42].
In order to characterize the roles of homeobox genes in M. oryzae development and pathogenicity, constructs for the targeted disruption of the MoHOX genes were generated using a split-marker deletion method or double-joint PCR method (Figure S1). Eight deletion mutants were generated in which all of the MoHOX genes were replaced with a hygromycin resistance cassette, as confirmed by DNA blot and RT-PCR analyses using a gene-specific probe and sets of PCR primers (Figure S1, Table S2). The strains of the wild-type and transformants generated in this study are presented in Table 1.
The effects of the deletion of MoHOX genes on M. oryzae development and pathogenicity are summarized in Table 2. In brief, deletion mutants of each MoHOX gene exhibited unique phenotypes in fungal development and pathogenicity, such as mycelial growth, conidial morphology, conidiation, and appressorium formation. The ΔMohox1 mutant showed a significant reduction in vegetative growth, but increase in melanin pigmentation (Figure S2). The ΔMohox6 mutant also exhibited a significant reduction in vegetative growth, whereas other phenotypes in the ΔMohox6 mutant were indistinguishable from those of the wild-type. Conidiation (asexual reproduction) was completely abrogated in the ΔMohox2 mutant, but this defect was recovered when the wild-type copy of the MoHOX2 gene was transformed into the mutant. The ΔMohox4 mutant produced shorter and smaller conidia in both length and width compared to those of the wild-type (Figure S3). The ΔMohox7 mutant was unable to form appressoria on hydrophobic surfaces, while its other phenotypes were similar to those of the wild-type. However, the ΔMohox8 mutant formed appressoria, but was nonpathogenic due to a defect in penetration. The mst12 mutant, carrying a partial deletion in the MST12 gene, shows the same phenotypes as the ΔMohox8 mutant, confirming that they are the same gene [30]. No distinguishable phenotypes were observed in ΔMohox3 and ΔMohox5, as compared to the wild-type. Based on our phenotypic observations of MoHOX deletion mutants, the mutants ΔMohox2 and ΔMohox7 were chosen for detailed studies as their phenotypes are associated with important developmental stages in the M. oryzae disease cycle.
Quantitative measurement of conidia reconfirmed that conidial production was completely abolished in the ΔMohox2 mutant on V8 juice or oatmeal agar media. However, this defect in conidiation was fully recovered in the complemented transformant Mohox2c, to a similar extent as in the wild-type (Table 2). The other phenotypes in the ΔMohox2 mutant were quite similar to those in the wild-type and complemented transformant Mohox2c (Table 2, Figure S4), indicating that MoHOX2 is specifically involved in conidiation. Microscopic observation was performed to carefully define the effect of the MoHOX2 deletion on conidial formation (Figure 2). The wild-type and Mohox2c developed pear-shaped conidia on a conidiophore in a sympodial pattern 18 h after incubation (Figure 2A and 2C). However, no conidia formed in ΔMohox2 after prolonged incubation under conidial induction conditions, even though conidiophore development appeared to be normal in the mutant. In order to determine if MoHOX2 is also involved in conidiophore development, we stained patches of aerial mycelia with lactophenol aniline blue to distinguish conidiophores from other aerial hyphae [27]. Microscopic examination revealed that conidiophores developed in the ΔMohox2 mutant, as in the wild-type and Mohox2c (Figure 2B). These results indicate that MoHOX2 is a specific regulator that is essential for conidial development.
To evaluate the role of MoHOX2 during M. oryzae disease development, a pathogenicity assay was performed by inoculating susceptible rice leaves with mycelial agar plugs, rather than a conidial suspension, because ΔMohox2 is unable to produce conidia. The inoculation with ΔMohox2 mycelial blocks caused blast lesions similar to the wild-type (Figure 3A). Given that successful lesion development requires the development of appressoria on germ tubes of conidia for penetration into plant cells, this result led us to examine diseased tissue using microscopy. As expected, many appressorium-like structures were observed on the surfaces of rice leaves inoculated with either ΔMohox2 or wild-type mycelial agar plugs (data not shown). To investigate whether these appressorium-like structures contribute to disease development, we carried out a penetration assay, in which a mycelial agar plug (6 mm in diameter) was placed on onion epidermal cells. Both the wild-type and ΔMohox2 developed appressoria specifically at tips of hyphae on onion cells and invasive hyphal growth was subsequently observed underneath the appressorium inside the cell (Figure 3B). Appressoria that formed on hyphal tips of the wild-type and ΔMohox2 were indistinguishable in shape, size, and melanization. In order to evaluate whether a hydrophobic surface is conductive to appressorial formation by hyphae, an agar plug containing actively growing hyphae was placed on either the hydrophobic or hydrophilic surface of Gelbond film. Unexpectedly, both surfaces induced appressorial formation at hyphal tips of the wild-type and ΔMohox2 (Figure 3C). As seen in pathogenicity assays with ΔMohox2 mycelia, conidia from Mohox2c caused typical necrotic lesions on foliar parts of the host plant, similar to those caused by the wild-type (Figure 3A).
During appressorium-mediated penetration, lipid droplets abundant in conidium move into the incipient appressorium and degrade at the onset of turgor generation [43]. To understand the functional role of hypha-driven appressoria, we examined the temporal and spatial occurrence of lipid droplets using Nile red staining and epifluorescence microscopy. Unlike in conidia, lipid droplets were not initially detected in hyphae of the wild-type or ΔMohox2 on appressorium-inductive surfaces until 24 h (Figure 3D). However, lipid droplets became abundant in hyphae, and translocated into nascent appressoria 48 h after inoculation. The process was completed by 72 h. Considering that translocation of lipid droplets into nascent appressoria on conidial germ tubes occurs within 4 h [43], such a delayed process seems to be associated with the de novo synthesis of lipid droplets and hypha-driven appressorium development. Taken together, these results indicate that MoHOX2 is essential for conidiogenesis, but dispensable for appressorium formation and pathogenicity. Also, M. oryzae can form hypha-driven appressoria that can cause foliar disease.
Quantitative real-time RT-PCR (qRT-PCR) was used to examine the expression pattern of MoHOX2 under various conditions. The MoHOX2 gene was found to be constitutively expressed during development (Figure 4A). MoHOX2 transcript levels were dramatically higher during conidiation but lower during invasive growth, as compared to other stages. This pattern of MoHOX2 expression appears to be correlated with a functional role for MoHOX2 in conidiation.
To evaluate the impact of upstream signaling pathways on MoHOX2 expression, we measured the expression levels of the MoHOX2 gene in mutant backgrounds related to signal transduction (Figure 4B, Table 1). Significantly reduced MoHOX2 expression (greater than two-fold) was found in the adenylate cyclase mutant Δmac1 and the phospholipase C mutant ΔMoplc1. The expression of the MoHOX2 gene changed slightly, but not significantly, less than two-fold in other mutants, including the mutants Δmocrz1 for a calcineurin-responsive TF, ΔcpkA for a cAMP-dependent protein kinase catalytic subunit, Δmck1 for a MAPKKK, and Δpmk1 for a MAPK. These results suggest that the expression of MoHOX2 is co-regulated by cAMP and Ca2+-dependent pathways.
Since MoHOX2 is a putative homeobox TF, specifically required for an earlier stage of conidiation, it is reasonable to speculate that MoHOX2 acts as a TF that modulates the expression of other conidiation-related genes. To determine the impact of MoHOX2 deletion on the expression of conidiation-related M. oryzae genes and other M. oryzae genes that are orthologs to conidiation-related genes in other fungi (Table 3), their expression levels were measured in the ΔMohox2 mutant background. The expression of the genes MoAPS1, COS1, and Con7 was significantly upregulated (Figure 4C), as were the M. oryzae genes MoCon6 and MoCon8, orthologs to N. crassa Con-6 and Con-8 (Figure 4D). These results indicate that genes tested are not direct targets of MoHOX2. However, the altered levels of gene expression in ΔMohox2 suggest that MoHOX2 functions as a key TF of downstream gene expression leading to conidiogenesis.
The deletion of MoHOX7 completely abolished the ability to form appressoria while other phenotypes were not affected (Table 2). Microscopic examination revealed that wild-type conidia formed appressoria on the tip ends of germ tubes 6 h after incubation on a hydrophobic surface (Figure 5A). In contrast, the ΔMohox7 mutant failed to develop appressoria; instead, the germ tubes in ΔMohox7 elongated abnormally and appeared to have several rounds of swellings and hooking until 12 h. During prolonged incubation (24 h) of the ΔMohox7 mutant, its germ tubes grew like vegetative hyphae, with branches rather than repeating recognition steps (Figure 5A). Nile red staining of germ tubes of the ΔMohox7 mutant revealed a large accumulation of lipid droplets until swelling and hooking occurred, before switching to vegetative hyphal growth (Figure 5A). These defects in the ΔMohox7 mutant were repaired in the complemented transformant Mohox7c. This suggests that MoHOX7 plays a critical role in appressorium development rather than the recognition of cues that induce appressorium formation.
Next, we tested whether MoHOX7 is also required to form appressoria at hyphal tips. Not surprisingly, the ΔMohox7 mutant did not form an appressorium on hydrophobic and hydrophilic surfaces, although non-melanized swellings and hooking were found on tips of hyphae (Figure 5B), as observed with conidial germ tubes in Figure 5A. In contrast, the wild-type, Mohox7c, and MoHOX7e formed appressoria at the tips of mycelia (Figure 5A). This supports the idea that MoHOX7 is essential for the development of appressoria, both from hyphae and germ tubes.
Infection assays on rice were performed to test whether the ΔMohox7 mutant can cause disease on host tissues. Conidial suspensions were sprayed onto 3-week-old rice seedlings. The wild-type caused blast lesions on the plant, but the ΔMohox7 mutant produced no lesions in plant cells (Figure 6A). These defects in the ΔMohox7 mutant were fully restored to the wild-type levels in Mohox7c. To test the role of MoHOX7 in penetration, onion epidermis and rice leaf sheath cells were inoculated with a conidial suspension. The wild-type penetrated into epidermal cells and grew invasively (Figure 6B). In contrast, the ΔMohox7 mutant was unable to penetrate into plant surfaces. Germ tubes appeared to have rounds of swellings and hooking before vegetative growth began on plant surfaces (Figure 6B). To determine whether the ΔMohox7 mutant can perform invasive growth in plant cells, a conidial suspension was infiltrated into rice leaves by injection using a syringe. Both the wild-type and ΔMohox7 developed blast lesions, indicating that the mutant still had the ability to grow inside plant cells (Figure 6C). These results suggest that the functioning of MoHOX7 is not required for further growth inside the host, and is strictly limited to the stage of appressorial development.
qRT-PCR was performed to determine the transcription level of MoHOX7 during developmental stages in M. oryzae. The transcription of MoHOX7 was dramatically upregulated during appressorium formation (>29-fold), conidiation (>12-fold), and in invasive growth in planta (>four-fold) at 78 h after inoculation, compared to its expressions during mycelial growth (Figure 7A). The highest expression of MoHOX7 during appressorium development was consistent with its role in appressorium-mediated disease development, as shown in Figure 5.
To investigate whether signaling pathways associated with appressorium-mediated penetration affect the expression of the MoHOX7 gene, the levels of MoHOX7 transcripts were measured in several mutants related to signal transduction (Table 1, Figure 7B). Much lower levels of MoHOX7 transcripts were observed in ΔMoplc1 and Δmac1 mutants, whereas the expression of the MoHOX7 gene was not significantly affected in the mutants Δcpka, Δmck1, or Δpmk1. This suggests that the expression of MoHOX7 is regulated by Ca2+ and cAMP-dependent signaling pathways (Figure 7B). To test if such signaling molecules can restore appressorium formation in the ΔMohox7 mutant, a conidial suspension of ΔMohox7 on hydrophilic and hydrophobic surfaces was treated with the chemicals 1,16-hexadecanediol (HDD), cAMP, CaCl2, and 1,2-dioctanoyl-sn-glycerol (DOG), a diacylglycerol analogue. None of these molecules complemented the defects in appressorium formation in the ΔMohox7 mutant on both the hydrophilic and hydrophobic surfaces, while these molecules increased appressorium formation in the wild-type on the hydrophilic surface (Table S3). This result suggests that MoHOX7 is a key downstream regulator of appressorium development.
Conidiogenesis and appressorium development are key steps in the colonization of host plants by many fungal pathogens. These processes are controlled by a precise developmental program in response to stimuli from the host and environment. Organisms have evolved regulatory networks to ensure the correct timing and spatial pattern of the developmental events. Transcription factors (TFs) play important roles in fungal development and pathogenicity as regulators in biological networks. The functional analysis of TFs provides new insights into a controlling network that governs fungal development and pathogenicity.
In an effort to understand developmental biology in M. oryzae, we identified eight homeobox TFs (MoHOX1 to MoHOX8) as candidates for development-controlling genes since they are well-known regulators of development and differentiation in other organisms [29],[35],[36]. MoHOX2 proved essential for conidiogenesis. The disruption of MoHOX2 completely abolished the ability of M. oryzae to produce conidia, even though conidiophore development was normal. The deletion of this gene did not affect any other developmental stages, such as hyphal growth, appressorium formation, and penetration, except for a subtle difference. Our study suggests that MoHOX2 is a stage-specific regulator of an earlier step of conidiogenesis. This phenotypic feature in the mutant is interesting and unique because the deletion of any of the other genes related to conidiogenesis caused pleiotropic defects in M. oryzae. The Δcon7 mutant, affected in a zinc finger TF, develops morphologically abnormal conidia and never forms an appressorium [24]. A very recent study reported the interesting finding that COS1, which encodes another zinc finger TF, is a determinant of conidiophore formation and melanin pigmentation [27]. The authors also observed that the Δcos1 mutant, unlike the wild-type, developed appressorium-like structures on the host surface and disease symptoms in a mycelial inoculation test, and speculated that COS1 may have a role in an unknown mechanism involved in mycelia-mediated infection. However, we believe that the wild-type can form appressoria at hyphal tips as a strategy for host infection, although there may be strain-dependent differences. Such hypha-driven appressoria in the ΔMohox2 and wild-type mediated penetration into host cells caused typical symptoms of rice blast, indicating a functional similarity to conidial germ tube-driven appressoria in M. oryzae disease development. Hyphal appressorium-mediated penetration by M. oryzae may not predominantly occur in nature due to limitations in the spatial distribution of hyphae and on hyphal longevity. However, members of non-sporulating fungi develop sclerotia as survival and inoculum structures, in which hyphae become interwoven, aggregated, melanized, and dehydrated [44],[45]. These fungal pathogens penetrate host epidermal cells by means of infection cushions, aggregated forms of branched hyphae, after their perception of host factors [46]. The ΔMohox7 mutant, defective in appressorium formation on conidial germ tubes, was also unable to form appressoria at hyphal tips, and the defect was not recovered with the addition of exogenous cAMP. This indicates that MoHOX7 regulates appressorium development in both hyphae and conidial germ tubes. The role of MoHOX7 in appressorium development has been predicted by the study of a large deletion mutant, Δpth12 (NCBI accession number DQ060925), which was generated by restriction enzyme-mediated integration (REMI), but no detailed characterization of Δpth12 has been performed. Although entry into the host is possible through natural surface openings such as stomata, this is not believed to be the predominant mode for spread of infection; the ΔMohox7 mutant consistently failed to colonize unwounded host leaves. Independent of appressorium-mediated infection, M. oryzae has evolved other strategies to cause disease on hosts. During root infection, hyphae swellings resembling the hyphopodia of root-infecting fungi are associated with root invasion, leading to systemic disease development similar to typical foliar disease [47]. This evidence supports the idea that M. oryzae has evolved hypha-mediated infection structures to gain entry into their hosts.
A variety of signals induce appressorium formation, including surface hardness, hydrophobicity, adhesion quality, and host molecules [48]–[51]. The hyphae of the ΔMohox2 mutant and wild-type developed appressoria upon sensing hydrophobic surfaces, consistent with a previous observation [52]. In addition, the hyphae were also able to form appressoria on hydrophilic surfaces, unlike conidial germ tubes. This suggests that there is an unknown pathway that regulates the development of the hypha-driven appressoria after sensing environmental cues such as the surface hardness. The strong adhesion of conidial germ tubes to substrates is needed for appressorial development, and we also observed that tight adhesion of hyphae is a prerequisite to formation of appressoria at the tips. Adhesion, therefore, appears to be a critical step prior to the initiation of infection, not only for the prevention of dislodgement, but also for subsequent correct development. After perceiving inductive signals, germ tubes of conidia exhibit hooking and swelling related to appressorium formation. Thus, the continued hooking and swelling in the ΔMohox7 mutant indicates that MoHOX7 is not involved in signal recognition. In contrast, the deletion of MagB, which encodes a Gα subunit involved in sensing a surface cue, causes the formation of long and straight germ tubes [15].
The biological processes that mediate a functional appressorium from a conidial germ tube are very complicated, based on molecular and cytological evidence such as autophagy [53], changes in metabolism [43], cell cycle control [53], and the generation of turgor pressure [5]. A body of evidence has demonstrated that conserved signaling pathways are associated with the coordination of biological changes related to appressorial development and maturation. However, the downstream molecular pathways that are activated in developing appressoria remain mostly uncharacterized. Filling this gap would require a specific aim at characterizing tissue-specific regulators. Signaling pathways often converge on transcriptional regulation during cell development. Interestingly, the MoHOX2 and MoHOX7 genes were significantly downregulated in the two signaling-defective mutants Δmac1 and ΔMoplc1 [12],[17]. Given that both adenylate cyclase and phospholipase C are associated with the generation of signaling molecules, such as cAMP and Ca2+, it is possible that cAMP and/or Ca2+-dependent signaling pathways are involved in modulating the functions these two MoHOX genes.
As most homeobox TFs in other species are constitutively expressed [54], it is not surprising that MoHOX2 and MoHOX7 were constitutively expressed during development, but most highly expressed during conidiation and appressorium formation, respectively. Many TFs are also post-translationally regulated, especially by (de)phosphorylation [55],[56]. This event typically occurs in response to external stimuli, which lead to the modulation of the DNA-binding activity of homeobox TFs. This suggests that the two MoHOX proteins are phosphorylated or dephosphorylated by a kinase or phosphatase, respectively. The identification of such regulators responsible for (de)phosphorylating MoHOX proteins is in progress using a yeast two-hybrid system. Previously, a yeast two-hybrid study showed that PMK1 interacts with MST12 (MoHOX8), suggesting that PMK1 regulates MST12 to control invasive growth [30],[57]. PMK1 is a well-known MAP kinase essential for appressorium formation and invasive growth in M. oryzae [6]. Since MST12 is not involved in appressorium formation, there may be other TFs regulated by PMK1 that are involved in appressorium formation. Considering that MoHOX7 is another homeobox TF that is crucial for appressorium formation, PMK1 might interact with MoHOX7 for the regulation of appressorium development.
In summary, we have demonstrated that members of the homeobox TF family function as stage-specific regulators during M. oryzae development and pathogenicity. MoHOX1, -2, -4, -6, -7, and -8 are specifically associated with hyphal growth and pigmentation, asexual reproduction, conidial morphology, mycelial growth, appressorium development, and invasive growth, respectively. Detailed molecular and cytological analyses revealed that deletion of the MoHOX2 gene entirely abolished asexual reproduction, while other stages, including conidiophore development appeared normal. Also, the data showed that MoHOX7 is a key regulator, essential for appressorium development on both hyphal tips and conidial germ tubes. These results provide evidence that M. oryzae is able to cause foliar disease via hypha-driven appressoria, after sensing environmental cues. Our studies will help to unveil the regulatory mechanisms involved in conidiation and appressorium formation and contribute to development of novel strategies for rice blast control.
Magnaporthe oryzae strain KJ201 was obtained from the Center for Fungal Genetic Resources (CFGR) and was used as the wild-type stain in this study. The strain and its transformants were routinely grown at 25°C under continuous fluorescent light on oatmeal (50 g oatmeal per liter) agar medium or V8 (4% V8 juice) agar medium. DNA and RNA were isolated from mycelia, which were grown in liquid complete medium (0.6% yeast extract, 0.6% tryptone, 1% sucrose) for 4 days. Conidia were obtained from 10-day cultures on oatmeal agar media by rubbing the mycelia with water followed by filtration through Miracloth (Calbiochem, San Diego, USA). Germinated conidia were obtained by placing drops of conidial suspension on hydrophobic surfaces for 3 h. Appressoria were obtained by holding germinated conidia for an additional 3 h. For the phenotype assay, complete medium was used to measure the vegetative growth and colony characteristics [58]. Oatmeal agar medium and V8 juice agar medium were used to measure conidiation and conidial morphology.
Genomic DNA was isolated using two different methods, depending on the experimental purpose. Genomic DNA for general experiments was isolated according to a standard method [59]. Genomic DNA for PCR screening of transformants was prepared using the quick and safe method [60]. Restriction enzyme digestion, agarose gel separation, and DNA gel blotting were performed following standard procedures [59]. DNA hybridization probes were labeled with 32P using the Rediprime II Random Prime Labeling System kit (Amersham Pharmacia Biotech, Piscataway, NJ, USA) according to the manufacturer's instructions. The hybridization membrane was exposed to a Phosphorimager (BAS-2040, Fuji Photo Film, Tokyo, Japan) and visualized with the Phosphorimager software. Total RNA was isolated from frozen fungal mycelia using the Easy-Spin total RNA extraction kit (Intron Biotechnology, Seongnam, Korea) following the manufacturer's instructions. To measure the relative abundance of MoHOX2 and MoHOX7 transcripts in mutant backgrounds listed in Table 1, RNAs of the mutants were extracted from mycelia grown in CM liquid medium for 4 days at 25°C in a 120-rpm orbital shaker. The primer sets used to detect transcripts of conidiogenesis-related genes from M. oryzae, A. nidulans, and N. crassa, sets of primers are listed in Table 3 and Table S2.
For RT-PCR and quantitative real time RT-PCR (qRT-PCR), 5 µg of total RNA were reverse transcribed into first-strand cDNA using the oligo (dT) primer with the ImProm-II Reverse Transcription System kit (Promega, Madison, WI, USA) according to manufacturer's instructions. For detecting transcripts of the two complements, RT-PCR was conducted with primer pairs MoHOX2_ORF_F/MoHOX2_ORF_R and MoHOX7_ORF_F2/MoHOX2_ORF_R (Table S2). RT-PCR was performed in 20-µl reaction mixtures containing 100 ng cDNA, 2.5mM of dNTP mix, 2 µl 10×PCR buffer, 1 µl (10 pmol) of each primer, and 1 unit of Taq polymerase. In all, 30 cycles of RT-PCR were run on a Perkin-Elmer 9720 DNA thermal cycler. The β-tubulin gene was included as a control.
Real-time quantitative reverse transcription PCR (qRT-PCR) reactions were performed following previously established procedures [61]. The AB7500 Real-Time PCR system (Applied Biosystems, Foster city, CA, USA) was used for PCRs that consisted of 3 min at 95°C (1 cycle) followed by 15 s at 95°C, 30 s at 60°C, and 30 s at 72°C (40 cycles). Each qRT-PCR mixture (final volume 10 µl) contained 5 µl of Power SYBR® Green PCR Master Mix (Applied Biosystems), 3 µl of forward and reverse primers (100 nM concentrations for each) and 2 µl of cDNA template (12.5 ng/µl). The oligonucleotide sequences used for each gene are listed in Table S2. To compare the relative abundance of target gene transcripts, the average threshold cycle (Ct) was normalized to that of β-tubulin (MGG00604) for each of the treated samples as 2−ΔCt, where −ΔCt = (Ct, target gene−Ct, β-tubulin). Fold changes during fungal development and infectious growth in liquid CM were calculated as 2−ΔΔCt, where −ΔΔCt = (Ct, target gene−Ct, β-tubulin)test condition−(Ct, WT−Ct, β-tubulin)CM [10]. qRT-PCR was performed with three independent pools of tissues in two sets of experimental replicates.
Vegetative growth was measured on complete agar medium on 10 days after inoculation, with three replicates. The ability to produce conidia was measured by counting the number of conidia from 6-day-old V8 juice agar plates as described previously [7]. Conidia were collected by flooding the plate with 5 ml of sterilized distilled water. The number of conidia was counted using a hemacytometer under a microscope. Conidiophore development was monitored as previously described [23]. Conidial size was measured as width by length under a microscope. Conidial germination and appressorium formation were measured on a hydrophobic coverslip. Conidia were harvested from 10-day-old oatmeal agar culture plates using sterilized distilled water. A conidial suspension of 40 µl was dropped onto a coverslip following adjustment of its concentration to approximately 5×104 spores/ml. Drops were placed in a moistened box and incubated at 25°C. After 9 h of incubation, the percentage of conidia germinating and germinated conidia-forming appressoria was determined by microscopic examination of at least 100 conidia per replicate in at least three independent experiments, with three replicates per experiment.
Plant penetration assays were performed using onion epidermis and rice sheaths, as previously described [62]. These experiments were replicated three times. For the pathogenicity assay, conidia were harvested from 8 to 10-day-old cultures on oatmeal agar plates and 10 ml of conidial suspension (105 conidia/ml) containing 250 ppm Tween 20 were sprayed onto susceptible rice seedlings (Oryza sativa cv. Nakdongbyeo) at the three- to four-leaf stage. The inoculated plants were kept in a dew chamber at 25°C for 24 h in darkness and moved to a growth chamber with a photoperiod of 16 h with fluorescent lights. Disease severity was measured at 7 days after inoculation, as previously described [63]. For the infiltration infection assay, 100 µl of conidial suspension were injected into three points per leaf of 4-week-old rice plants. These experiments were replicated three times.
Homeobox TFs were identified using InterPro terms (IPR001356 and IPR003120) for homeodomains via the pipeline of Fungal Transcription Factor Database (http://ftfd.snu.ac.kr/) [64].
Amino acid sequences of homeobox TFs were aligned using CLUSTAL W 1.83 [65]. Phylogenetic trees were constructed using the neighbor-joining method [66] with the aid of FTFD [64]. All sequence alignments were tested with a bootstrap method using 10,000 repetitions. The domain architecture of homeobox TFs was determined by InterProScan [67] and presented using the CFGP (http://cfgp.snu.ac.kr/) [68].
|
10.1371/journal.ppat.1002234 | Maturation-Induced Cloaking of Neutralization Epitopes on HIV-1 Particles | To become infectious, HIV-1 particles undergo a maturation process involving proteolytic cleavage of the Gag and Gag-Pol polyproteins. Immature particles contain a highly stable spherical Gag lattice and are impaired for fusion with target cells. The fusion impairment is relieved by truncation of the gp41 cytoplasmic tail (CT), indicating that an interaction between the immature viral core and gp41 within the particle represses HIV-1 fusion by an unknown mechanism. We hypothesized that the conformation of Env on the viral surface is regulated allosterically by interactions with the HIV-1 core during particle maturation. To test this, we quantified the binding of a panel of monoclonal antibodies to mature and immature HIV-1 particles by immunofluorescence imaging. Surprisingly, immature particles exhibited markedly enhanced binding of several gp41-specific antibodies, including two that recognize the membrane proximal external region (MPER) and neutralize diverse HIV-1 strains. Several of the differences in epitope exposure on mature and immature particles were abolished by truncation of the gp41 CT, thus linking the immature HIV-1 fusion defect with altered Env conformation. Our results suggest that perturbation of fusion-dependent Env conformational changes contributes to the impaired fusion of immature particles. Masking of neutralization-sensitive epitopes during particle maturation may contribute to HIV-1 immune evasion and has practical implications for vaccine strategies targeting the gp41 MPER.
| The conformation of HIV-1 Env is of tremendous importance from an immunological standpoint. While several human monoclonal antibodies that exhibit broadly neutralizing activity have been identified, efforts to elicit such antibodies have met with minimal success. Here, we show that the conformation of Env is altered on the surface of immature vs. mature HIV-1 particles in such a way that certain epitopes recognized by some broadly neutralizing antibodies are more exposed on immature virions. This maturation-dependent conformational masking may represent an important mechanism of HIV-1 immune evasion.
| HIV-1 fusion is mediated by the Env glycoprotein, a trimeric complex of heterodimers composed of the surface glycoprotein (SU) gp120 and the transmembrane glycoprotein (TM) gp41. Fusion of virions with target cells takes place through a series of events initiated by binding of gp120 to CD4 on the surface of the target cell (reviewed in [1]). CD4 binding induces conformational changes in gp120 that permit exposure of the coreceptor-binding site, composed of the bridging sheet (consisting of four discontinuous anti-parallel beta strands) and the third hypervariable (V3) loop. Subsequent engagement of CD4-bound gp120 by a chemokine coreceptor—either CCR5 or CXCR4—triggers dramatic conformational changes in gp41 that result in fusion of viral and cellular membranes.
A common feature of lentiviruses is that their TM proteins have long cytoplasmic tails. HIV-1 gp41 encodes a 152 amino acid cytoplasmic tail (CT), while TM proteins of simple retroviruses have tails of 20–50 amino acids in length [2]. Several activities have been attributed to the gp41 CT, including polarized budding of HIV-1 particles from epithelial cell monolayers [3], rapid internalization of Env from the cell surface [4], [5], incorporation of Env into virions during particle assembly [6], [7], and interaction with Pr55Gag during virion assembly [5], [6], [7], [8], [9].
To become infectious, newly formed HIV-1 particles must undergo a process of maturation involving specific cleavage of the major structural polyprotein Pr55Gag by the viral protease. Immature HIV-1 particles contain stable cores and are non-infectious due to defects in early post-entry steps of the life cycle [10]. However, recent studies have demonstrated that immature virions are also impaired for fusion with target cells and that the gp41 CT plays a key role in repressing immature HIV-1 particle fusion [11], [12], [13].
The detailed mechanism by which HIV-1 fusion is regulated by structural changes within the core has not been determined, but one recent study attributed the repression to a change in physico-mechanical properties (i.e. “stiffness”) that accompanies HIV-1 maturation [14]. An alternative hypothesis is that maturation triggers a conformational change in the ectodomain of the Env glycoprotein complex, releasing it into a fusion-competent state. Such a mechanism might also limit the exposure of neutralization-sensitive epitopes in gp120 and gp41, thus promoting immune evasion. Previous work has revealed that the gp41 CT modulates Env conformation on HIV-1, HIV-2, and SIV, thus lending support to the latter hypothesis [15], [16], [17].
To test whether HIV-1 particle maturation alters the conformation of the Env proteins, we used a sensitive and quantitative imaging-based antibody-binding assay to probe the conformations of full-length and CT-truncated Env proteins on mature and immature HIV-1 particles. The results revealed specific epitopes in gp120 and gp41 that exhibit greater exposure on immature vs. mature virions, including two in the membrane-proximal external region (MPER). Therefore, Env trimers on immature virions are present in an exposed conformation, and neutralization-sensitive epitopes undergo conformational masking during particle maturation.
To analyze Env conformation on HIV-1 particles, we developed a sensitive and quantitative imaging-based assay for binding of antibodies to virions. The assay was designed to permit the use of a panel of available conformation-specific monoclonal antibodies (mAbs) specific for gp120 and gp41. HIV-1 particles, containing a GFP-Vpr fusion protein, were immobilized on glass cover slips and stained using an indirect immunofluorescence protocol. To avoid potential artifacts resulting from fixation, Env-specific primary antibodies were bound under native conditions. After washing to remove unbound antibodies, antibody-bound virions were fixed with paraformaldehyde and detected by addition of a fluorescent, Cy5-conjugated secondary antibody. By this approach, the gp120-specific mAb 2G12 readily detected wild-type HIV-1 particles, as seen by the colocalization between the GFP-labeled virions and the Cy5 antibody fluorescence (Fig. 1, top row). As a control, Env-defective HIV-1 particles were not bound by 2G12, thus establishing the specificity of the assay (Fig. 1, bottom row).
The imaging-based assay detects antibody binding to individual HIV-1 particles, allowing for analysis of the distribution in staining intensities among a population of virions. The binding signal generally exhibited a broad distribution, with the bulk of the particles exhibiting low-level binding with a wide tail toward higher binding. Wild-type HIV-1 particles have been reported to contain approximately 10 Env trimers, with considerable variability among virions [18]. In the example shown for mAb 4E10 binding, a larger percentage of mature, wild type virions exhibited a lower average intensity than immature virions while more immature virions had high average intensities (Fig. S1A). The change in median average intensity per particle observed over the distribution was also apparent in the fluorescence images (Fig. S2). These are examples of data used to quantitatively compare the binding of antibodies to mature and immature HIV-1 particles. Representative histograms for each antibody are included in the supporting information.
As an additional control, the fluorescence distribution for Env-defective immature particles was determined for mAb 4E10 and mAb b12 binding. As expected, there was no difference between the binding to mature and immature particles lacking HIV-1 Env (Fig. S3), and the binding to each was minimal.
To facilitate interpretation of the particle imaging data, immunoblotting of viral lysates was performed to compare Env levels on virions (Fig. S4A). Quantification of band intensities revealed a difference of no more than 20% between mature and immature particles (Fig. S4B). Consistent with our previous work and that of others [8], [13], we observed higher TM/SU ratios on both mature and immature particles lacking the gp41 CT, owing to the elevated TM protein association with HIV-1 particles. Since our analyses focused mainly on pairwise comparisons between mature and immature particles, the elevated gp41 levels on CT-truncated particles does not represent a confounding factor. We conclude that the quantitative differences in immunodetection of Env on the surface of mature and immature HIV-1 particles result from differences in Env conformation.
As an initial test to compare conformations of Env proteins on mature and immature virions, we tested the binding of HIV-Ig, a polyclonal IgG pool isolated from HIV-1 infected individuals. Under the optimized experimental conditions, Env-defective HIV-1 particles were not bound by HIV-Ig, indicating that staining was specific for Env (Fig. 2). We observed that the median staining intensity was 28% higher for immature versus mature particles. Immature HIV-1 particles lacking the gp41 cytoplasmic tail also exhibited elevated staining with HIV-Ig relative to the corresponding CT-truncated mature particles; however, this increase was not statistically significant. Because the imaging-based assay permits quantification of antibody binding to each particle, we also examined the distribution of HIV-Ig antibody binding (Fig. S5). Analysis of the distribution revealed that the bulk of the particles bound lower quantities of antibodies, with a tail toward higher binding levels. These results indicate that while some of the HIV-1 Env epitopes recognized by HIV-Ig exhibit enhanced exposure on immature particles, the gp41 CT is not required for the enhanced antibody binding. However, some antigenic elements on the surface of HIV-1 virions could be masked by cleavage of the Gag polyprotein inside the virions. Therefore, we asked whether specific Env epitopes differ between immature and mature virions and whether such differences depend on the gp41 CT.
MAb 2G12 specifically recognizes N-linked glycans in the C2, C3, C4, and V4 domains of gp120 [19]; binding of this antibody is thus independent of gp120 conformation. Accordingly, we observed no significant difference in mAb 2G12 binding to mature and immature virions (Fig. 3A, Fig. S6A). However, mAb 2G12 binding was increased by 35% on immature virions containing a truncated gp41 CT, consistent with a previous report that truncation of the CT enhances mAb 2G12 binding to Env expressed on the cell surface [15].
MAb B4e8 recognizes the V3 loop in gp120. We observed that mAb B4e8 binding to immature particles was increased by 15% relative to mature HIV-1 virions (Fig. 3B, Fig. S6B). On CT-truncated particles, the 15% increase in binding to immature particles was retained. However, neither of these differences was statistically significant, suggesting that exposure of the mAb B4e8 epitope is not markedly altered during HIV-1 maturation. The data from these two conformation-independent antibodies also corroborates the conclusion from the immunoblot analysis that the levels of gp120 are not significantly different on mature vs. immature virions.
A previous study reported that mature and immature HIV-1 particles are equally competent for binding to CD4+ T cells [13]. To test whether HIV-1 maturation alters the conformation of the CD4 binding site on gp120, we quantified the binding of mAb b12, which recognizes an epitope overlapping the CD4 binding site. MAb b12 bound mature and immature virions to an equivalent extent, suggesting that the region of gp120 recognized by this antibody is not structurally altered on the surface of immature HIV-1 particles (Fig. 4A, Fig. S7A). Curiously, the CT-deleted Env bound significantly more mAb b12 when present on immature vs. mature particles. Overall, these data suggest that the epitope recognized by mAb b12 is exposed to a similar extent on mature and immature HIV-1 particles and that the gp41 CT appears to modulate exposure of this epitope differentially on mature vs. immature particles.
To quantify CD4 binding to HIV-1 particles, we utilized CD4-IgG2, a fusion protein in which four copies of the V1 and V2 domains of human CD4 replace the heavy and light chain Fv portions of human IgG2. Because sCD4 has been shown to induce gp120 shedding to a different extent on mature vs. immature virions [20], we asked whether differential shedding contributed to the results obtained for CD4-IgG binding. For this purpose, we incubated HIV-1 particles with soluble CD4 (sCD4) and stained with mAb 2G12, whose binding should be unaffected by the conformational state of gp120. We found that while the mAb 2G12 signal decreased by 22% for mature HIV-1, the signal for immature HIV-1 decreased by only 11% (Fig. S8). Likewise, the mature CT-truncated virus signal decreased by 20%, and the immature tail-truncated virus signal decreased by only 3%. When the binding results were normalized by the observed levels of gp120 shedding, we observed no difference in binding of sCD4-IgG2 to mature and immature virions (Fig. 4B, Fig. S7B). Collectively, these results suggest that the CD4 binding site on gp120 is not structurally altered on the surface of immature virions. The data further demonstrate that the fusion impairment associated with immature viruses is not owing to a quantitative defect in CD4 binding.
CD4 binding induces structural rearrangements in gp120, exposing epitopes recognized by the mAbs E51, A1g8, and 17b, which overlap the coreceptor-binding site in gp120 [21], [22]. We employed these antibodies in combination with soluble CD4 (sCD4), and taking into account the levels of sCD4-induced gp120 shedding, to quantify CD4-induced conformational changes on the surface of mature or immature HIV-1 particles. In the absence of sCD4, mAbs E51 and A1g8 binding to gp120 was approximately 20% greater on immature vs. mature HIV-1 particles (Fig. 5A and B, black bars, Fig. S9E and H). By contrast, binding of mAb 17b to immature particles was approximately 40% less than binding to mature particles (Fig. 5C, Fig. S9B). We also tested the effects of sCD4 on the binding of these antibodies (Fig. 5A–C, white bars, Fig. S9C,F,I). Binding of each antibody to immature virions was stimulated by sCD4 to a greater or equal extent vs. mature particles, with mAb 17b exhibiting the greatest increase (Fig. 5D). Truncation of the CT abolished the enhanced sCD4-induced binding of mAbs A1g8 and 17b to immature particles. Collectively, these results demonstrate that CD4 binding triggers exposure of some epitopes to an equal extent on immature and mature virions and other epitopes to a greater extent on immature virions. Furthermore, the involvement of the gp41 CT links the enhanced epitope exposure to the fusion impairment associated with immature HIV-1 particles.
The fusion protein gp41 contains several epitopes that are exposed preferentially when Env is in a fusion-active conformation. MAb 50–69 binds at the C-terminal end of the N-terminal heptad repeat region of gp41 [23]. Binding of mAb 50–69 to immature particles was approximately 75% greater than to mature particles, a difference that was highly significant (Fig. 6A, Fig. S10A and B). Increased binding also was observed on particles lacking the gp41 CT. Thus, the Env protein on the surface of immature virions exhibits greater exposure of the mAb 50–69 epitope, but this conformational difference does not depend on the CT.
MAb 5F3 is a gp41-specific antibody that recognizes an epitope adjacent to the fusion peptide. We observed that mAb 5F3 binding to immature virions was approximately twice that of mature virions (Fig. 6B, Fig. S10C and D). When the CT was truncated, mAb 5F3 binding to mature particles was increased, but binding to mature and immature particles was equivalent. Thus, virion maturation masks the epitope recognized by mAb 5F3, and truncation of the gp41 CT abolishes this effect.
The MPER of gp41 is the target of several broadly neutralizing antibodies (bnAbs). The antigenic epitopes in this region of gp41 may be hidden from immune recognition, since the presence of such bnAbs is rare in infected patients. We tested the MPER-specific mAbs 2F5, Z13e1, and 4E10 for binding to mature and immature HIV-1 particles. Binding of mAbs 4E10 and Z13e1 to immature particles was greater than to mature virions (70% and 100% increase, respectively), and in both cases the increase was abolished by truncation of the gp41 CT (Fig. 7A and B, Fig. S1A–D). For mAb 2F5, slightly greater binding to immature particles was apparent, but the observed difference was not statistically significant (Fig. 7C, Fig. S1E and F).
On average, the MPER-specific mAbs 4E10 and Z13e1 bound immature particles approximately 1.5 to 2 times as well as mature particles when the median binding signals were compared. Examination of the distributions revealed that the differences resulted from increases in the percentage of immature particles at higher antibody binding levels (Fig. S1A and C). Analysis of mAb Z13e1-bound particles with an average intensity above 15 a.u. revealed a greater than 2-fold increase in binding to immature HIV-1 virions (Fig. S1D); analysis of mAb 4E10-bound particles with an average intensity above 50 a.u. revealed a 6-fold increase in binding to immature HIV-1 virions (Fig. S1B). We conclude that MPER-specific mAbs exhibit enhanced binding to immature HIV-1 particles, and the gp41 CT plays a role in controlling accessibility of MPER epitopes.
To test whether the increased binding was an effect of avidity caused by bivalent interaction with full-length IgG forms of the mAbs rather than increased affinity of the antibody combining site with the epitopes, we tested the binding to particles of Fab fragment forms of mAbs that were directly labeled with fluorophore. Consistent with the results using whole antibodies, Fab 4E10 exhibited significantly enhanced binding to immature vs. mature HIV-1 particles (Fig. S11), and the effect was abolished by truncation of the gp41 CT. By contrast, binding of Fab b12 did not exhibit a statistically significant difference. We conclude that MPER-specific antibodies exhibit preferential binding to immature HIV-1 particles and the gp41 CT contributes to the enhanced binding. These results reinforce the conclusion that the epitope recognized by mAb 4E10 is more accessible on immature HIV-1 particles.
In this study, we observed quantitative differences in epitope exposure on the surface of mature and immature HIV-1 particles using a novel single particle imaging-based binding assay. Several important epitopes exhibited increased exposure on the surface of immature HIV-1 particles. Specifically, exposure of the gp41 epitopes recognized by mAbs 50–69 and 5F3, and the MPER-specific mAbs Z13e1 and 4E10, were markedly increased on the surface of immature particles. By contrast, binding of mAb 17b, which recognizes a CD4-induced epitope, was lower for immature vs. mature particles. However, the accessibility of the CD4-induced epitope recognized by mAb 17b was generally increased on immature vs. mature particles. Taken together, the results indicate that the conformation of Env is altered on the surface of immature vs. mature HIV-1 particles, suggesting that the conformational of Env is altered during HIV-1 particle maturation.
Understanding the conformations of Env is of tremendous importance from an immunological perspective. HIV-1 rapidly evolves to evade the host antibody response, yet a current view is that an effective HIV-1 vaccine will require both humoral and cellular virus-specific responses. Despite considerable effort, attempts to elicit antibody responses that neutralize a wide range of HIV-1 isolates have been unsuccessful. Nonetheless, several human mAbs have been identified that exhibit broadly neutralizing activity, indicating that such antibodies can be produced in vivo, although they are rare. An area of intense study is the MPER region of gp41, which is the target of several bnAbs. In the current study, we show that the epitopes recognized by mAbs Z13e1 and 4E10 are bound more readily when present on immature particles, indicating that these epitopes become masked during HIV-1 particle maturation. Such conformational masking may represent an important mechanism of HIV-1 immune evasion, and immunization strategies targeting the MPER may benefit from focused approaches utilizing structurally engineered antigens informed by studies of immature HIV-1 particles. Therefore, a detailed understanding of the conformation of Env on immature particles could aid in the design of recombinant HIV-1 immunogens.
In the present study, we primarily analyzed the binding of full-length IgG antibodies rather than monovalent Fabs. Because IgG molecules are bivalent, their binding could be stabilized by avidity effects, which in turn could depend on the proximity of Env complexes on the virion surface. Our previous work showed that the Env proteins form a stable complex with the core within immature HIV-1 particles that depends on the gp41 CT. In mature virions, or particles lacking the gp41 CT, the Env trimers may be free to move laterally on the virion surface, thus allowing for patching of Env trimers. Thus, differences in antibody binding might be due, at least in part, to avidity effects resulting from altered trimer mobility in the viral membrane. To address this, we tested the binding of b12 and 4E10 Fab proteins, and the results agreed with the binding of the corresponding full-length IgG antibodies. Therefore, avidity effects seem unlikely to account for the enhanced antibody binding observed for immature particles, on which the patching of Env would be limited. Thus, our results are more consistent with the interpretation that immature particles exhibit increased exposure of specific epitopes vs. differential antibody-induced patching of Env trimers.
Previous work by our group and another demonstrated that immature HIV-1 particles are repressed for fusion with target cells by an activity that requires the gp41 CT [11], [13]. In the present study, we sought to test whether the repressed fusion of immature particles might be due to restricted conformational changes on the surface of immature particles. The tight association of both Env subunits with immature particles, observed even following detergent treatment, suggested that the Env might be locked in a “cloaked” fusion-inactive conformation owing to association of the gp41 CT with the highly stable Gag polyprotein lattice [8]. This cloaking mechanism would have the benefit of protecting the HIV-1 Env complex from neutralizing antibodies during the maturation process. Our results were surprising: exposure of several neutralization-sensitive Env epitopes was greater on immature particles. Several of the maturation-dependent conformational differences we observed were abolished by truncation of the gp41 CT, indicating that the CT couples Env conformation to particle maturation. CD4-induced binding of the gp120-specific antibody 17b also was enhanced on immature particles, and truncation of gp41 suppressed the enhanced binding. Therefore, our results establish that the gp41 CT alters the conformation of the Env ectodomains on immature particles, potentially interfering with the receptor-dependent conformational changes required for fusion. It is possible, however, that the enhanced sCD4-induced binding of 17b is unrelated to the fusion impairment associated with immature particles. Murakami et al. previously reported that immature HIV-1 particles are not impaired for binding of a mAb (NC-1) that recognizes the 6-helix bundle (6HB) conformation, which is thought to form at a late stage of the fusion process [13]. Specifically, the authors observed that addition of sCD4 induced equivalent binding of mAb NC-1 to mature and immature particles. While this result could suggest that the Env trimers on immature particles are competent for fusion-dependent conformational changes, it is not clear that mAb NC-1 binding to cell-free virions necessarily reflects a conformational change specific to fusion. Indeed, a recent study has shown that NC-1 recognizes other forms of gp41 besides the 6HB [24]. Therefore, it remains plausible that the differential binding of selected mAbs to mature and immature HIV-1 particles reflects conformational differences that contribute to the fusion impairment exhibited by immature HIV-1 virions.
Based on our results, we propose a model in which a strong Gag-Env association constrains the Env subunits, particularly gp41, into an exposed conformation, and that receptor engagement is insufficient to drive the additional conformational changes necessary for fusion. The model also implies that the MPER-specific bnAbs act by trapping gp41 in a conformational intermediate formed during particle maturation. Our results do not exclude other potential effects contributing to impaired fusion, such as altered physico-mechanical properties associated with immature particles bearing full-length Env proteins [14]. It should be noted that the present work employed a laboratory-adapted HIV-1 clone, which is likely hypersensitive to CD4. Fusion of viruses containing Env proteins from primary HIV-1 isolates is also regulated by maturation [25], and it will be important to extend the present studies to such Env proteins.
293T cells were cultured at 37°C and 5% CO2 in Dulbecco's Modified Eagle medium (DMEM; Cellgro) supplemented with fetal bovine serum (10%), penicillin (50 IU/mL), and streptomycin (50 µg/mL). The proviral DNA constructs used for the production of HIV-1 have been described previously [11] and are as follows: R9, wild-type HIV-1; R9.PR-, protease-defective HIV-1 containing a triple alanine substitution in the protease active site; R9Tr712, HIV-1 containing a truncation of the gp41 C-terminal 144 amino acids; R9Tr712.PR-, protease-defective HIV-1 containing a truncation of the gp41 C-terminal 144 amino acids. Viruses were produced by transient transfection of 293T cells in 10 cm dishes with 20 µg of proviral DNA and 7 µg of a GFP-Vpr fusion protein expression vector [26] using a calcium phosphate-based method [27]. Virus stocks were harvested 48 h after transfection and clarified through 0.45 µm syringe filters. Aliquots were buffered with 10 mM HEPES pH 7.3 prior to storage at −80°C. HIV-1 stocks were assayed for p24 by enzyme-linked immunosorbent assays (ELISAs) as previously described [8], [28], after boiling in SDS-PAGE loading buffer to solubilize the hyperstable immature particles.
The following reagents were obtained through the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: CD4-IgG2 from Progenics Pharmaceuticals; HIV-1 gp120 mAb 2G12 and HIV-1 gp41 mAbs 5F3, 2F5, and 4E10 from Dr. Hermann Katinger; HIV-1 gp120 mAbs F425 B4e8 and F425 A1g8 from Dr. Marshall Posner and Dr. Lisa Cavacini; HIV-1 gp120 mAb IgG1 b12 from Dr. Dennis Burton and Carlos Barbas; HIV-Ig from NABI and NHLBI; HIV-1 gp120 mAbs 17b and E51 from Dr. James E. Robinson; HIV-1 gp41 mAb 50–69 from Dr. Susan Zolla-Pazner; HIV-1 gp41 mAb IgG1 Z13e1 from Dr. Michael Zwick; sCD4-183 from Pharmacia, Inc.
Amino acid sequences for Fabs b12 and 4E10 were obtained from the Protein Data Bank (PDB IDs 2NY7 and 2FX7). DNAs encoding the Fab protein sequences were designed and optimized for mammalian cell expression systems and synthesized commercially (GeneArt, Regensburg, Germany). The Fab DNAs were cloned into the pEE6.4 heavy chain expression vector (Lonza Group Ltd, Basel, Switzerland) with a recombinant stop codon placed before the constant region, to specify Fabs. The light chain cDNAs also were expressed in recombinant form using a mammalian cell optimized sequence that had been cloned into the pEE12.4 light chain vector (Lonza). The plasmids were transformed into DH5 strain E. coli cells for EndoFree Plasmid Maxi DNA preparation (Qiagen, Hilden, Germany). Transient transfection of each heavy and light chain combination for expression of Fab proteins in the serum-free HEK 293F cell expression system (Invitrogen, Carlsbad, CA) was accomplished using PolyFect reagent (Qiagen) according to the manufacturer's instructions. Supernatants were collected after 120 hours of expression and Fabs were purified using FPLC with a KappaSelect prepacked column (GE HealthCare Life Sciences, Piscataway, NJ) in D-PBS and concentrated in 15 mL centrifugal filter units with 30 kDa molecular weight cut-off (Millipore, Billerica, MA). A high level of purity of the Fabs was confirmed using a non-reducing SDS-PAGE and a Coomassie Blue stain (Invitrogen) that did not reveal contaminating proteins. The Fabs were shown to bind in ELISA as expected to HIV-1 virus-like particles or to recombinant gp120 molecules [29] to confirm their functionality before further study. Purified 4E10 and b12 Fabs were directly labeled using the Alexa Fluor 647 Microscale Protein Labeling Kit (Invitrogen) according to the manufacturer's instructions.
HIV-1 virions were plated on poly-D-lysine coated dishes (MatTek). Virions were incubated with an Env-specific mAb (except in the case of polyclonal HIV-Ig) for two hours at room temperature, fixed in 2% paraformaldehyde for 15 min at room temperature, washed five times with PBS, incubated with a Cy5-conjugated anti-human IgG (Jackson ImmunoResearch Laboratories, Inc.) at a concentration of 14 µg/mL for one hour at room temperature, washed five times with PBS, mounted, and imaged using a Zeiss LSM 510 META inverted confocal microscope. Image acquisition was performed using a 63× objective lens with 2× optical zoom with line averaging for a 1024×1024 pixel image. GFP imaging was performed with an excitation wavelength of 488 nm and band pass 505–550 emission filter. Cy5 imaging was performed with an excitation wavelength of 633 nm and a long pass 650 emission filter. For preincubation with sCD4, viruses were incubated in suspension with 0.25 µg/mL 2-domain sCD4-183 for 30 min at room temperature prior to plating and addition of primary antibody. To optimize the assay for detecting differences in epitope exposure, each antibody was first titrated on wild-type and Env-deficient virions. Antibody concentrations giving minimal staining of Env-deficient virions and a non-saturated level of staining with wild-type virions were employed (Table S1). This allowed each antibody to be used within its dynamic range, so that any differences in binding to the viruses could be detected. Final antibody concentrations are provided in Table S1. CD4-IgG2 was used at 0.25 µg/mL. These concentrations were selected as the optimal concentrations that exhibited minimal background and were not saturating thus allowing detection of quantitative changes in epitope exposure.
Labeled 4E10 and b12 Fabs (1 µg/mL) were used in virus binding assays as described in the text. The only alteration to the protocol was eliminating the secondary antibody incubation and mounting immediately after the post-fixation washes.
MetaMorph software (Molecular Devices) was used to quantify the average intensity of Cy5 staining for each GFP-positive particle. Particles were defined as adjacent groups of green pixels, or regions, above a background intensity of 20 graylevels in an area with a width and height between 0.2 and 0.8 microns. These regions were overlayed on the Cy5 image, and both the area and average signal intensity were calculated for each region. As can be seen in Figs. 1 and S2, not all Cy5 stained spots colocalize with a GFP region. We hypothesize that these spots are either HIV-1 particles that did not incorporate sufficient GFP-Vpr to be detected, or are microvesicles containing HIV-1 Env. Since Cy5 intensity was measured within regions defined by the presence of GFP, these GFP-negative particles were excluded from analysis. To verify that analyzing the average intensity for each particle would not skew the data, size distributions were generated for each virus type using the data from two independent mAb 4E10 and mAb b12 staining experiments (Fig. S12). These distributions, together with the absence of a correlation between the average GFP intensity per particle and particle area (Fig. S13B), showed that the apparent particle areas were not different for each type of HIV-1 virus used. In addition, there is no correlation between the average GFP intensity per particle and the average Cy5 (4E10) intensity per particle (Fig. S13A) and also no correlation between average Cy5 (4E10) intensity per particle and particle area (Fig. S13C). Of note, the observed higher levels of 4E10 binding occurred over the full range of particle sizes (Fig. S13C).
Data points are the median average intensity per particle of six independent fields (each contained approximately 200–600 particles) obtained from at least three independent experiments. These results were used in Wilcoxon rank-sum tests to evaluate statistical significance of observed differences. As a reference point, the particle size (Fig. S12) and the antibody binding histograms (Fig. S1,5,6,8,9,10) indicate the number of particles from one independent experiment (the combination of six images) used to generate the plots.
Viruses particles were pelleted through a 20% sucrose cushion. Pellets were resuspended in SDS-PAGE sample buffer. Virion quantities were normalized based on p24 ELISA results prior to electrophoresis on 4%–20% SDS-PAGE gels (Bio-Rad) and transfer to nitrocellulose. Nitrocellulose was blocked in 5% nonfat dry milk in PBS containing 0.1% Tween-20. Primary antibodies were used as follows: gp120 mAb 2G12 (1 µg/mL), gp41 mAb 2F5 (1.25 µg/mL), CA 183-H12-5C (0.75 µg/mL). Secondary antibodies were donkey anti-human IgG IRDye800 (Rockland) and goat anti-mouse IgG DyLight680 (Thermo Scientific). Blots were scanned using the LI-COR Odyssey Imaging System, and bands were quantified using the instrument software.
To generate example histograms for each antibody, the intensity values for each particle from six independent images acquired from one experiment were binned into 1 arbitrary unit (a.u.) bins. The percentage of particles falling within each bin was calculated based on the total number of particles, as indicated in the legend for each histogram, in the population. The y-axis scale was set to maximally visualize the WT, PR-, ΔCT, and PR- ΔCT virus populations. Therefore, the maximal value for the Env- population is indicated in parentheses in the top left corner of each plot.
To determine the percentage of particles stained above a cutoff intensity, as indicated by the white arrow and dotted line on the histograms, the average intensity values for each particle from six independent images acquired from one experiment (with the corresponding histogram representing one such experiment) were combined and analyzed as a single unit. The percentage of particles above the cutoff for each virus population in each experiment was calculated. These values were averaged for each virus and presented with the standard error of the mean. Each data point includes values from between three and six independent experiments.
To assess the level of gp120 shedding from mature and immature HIV-1, we quantified the level of mAb 2G12 binding following incubation with sCD4 under the conditions used for quantifying CD4-induced binding of mAb 17b. The HIV-1 viruses were plated on poly-D-lysine coated MatTek dishes, incubated with 0.25 µg/mL sCD4 for 30 min at room temperature, incubated with 1 µg/mL mAb 2G12 for 2 hours at room temperature, fixed in 2% paraformaldehyde for 15 min at room temperature, washed five times with PBS, incubated with a Cy5-conjugated anti-human IgG for one hour at room temperature, washed five times with PBS, mounted, and imaged. Image acquisition and data analysis were performed exactly as described in the text.
|
10.1371/journal.pgen.1000393 | Mutation of the Mouse Syce1 Gene Disrupts Synapsis and Suggests a Link between Synaptonemal Complex Structural Components and DNA Repair | In mammals, the synaptonemal complex is a structure required to complete crossover recombination. Although suggested by cytological work, in vivo links between the structural proteins of the synaptonemal complex and the proteins of the recombination process have not previously been made. The central element of the synaptonemal complex is traversed by DNA at sites of recombination and presents a logical place to look for interactions between these components. There are four known central element proteins, three of which have previously been mutated. Here, we complete the set by creating a null mutation in the Syce1 gene in mouse. The resulting disruption of synapsis in these animals has allowed us to demonstrate a biochemical interaction between the structural protein SYCE2 and the repair protein RAD51. In normal meiosis, this interaction may be responsible for promoting homologous synapsis from sites of recombination.
| Production of sperm and eggs, also known as gametes, requires a reduction in the number of copies of the genome, from the two found in most cells of the body to the single copy found in gametes. This is a complex process, made even more complex because it is coupled with recombination, a process that is an important contributor to genetic diversity. Mammals and many other organisms achieve reduction and recombination through a process called meiosis, which is recognisable by the presence of a distinctive structure—the synaptonemal complex—that links the chromosomes together and is essential for meiosis to complete. We have made mice that lack SYCE1, a protein component of the synaptonemal complex. In these animals, meiosis is blocked at a particular stage, and this has allowed us to detect co-localisation and interactions—likely indirect—between enzymes involved in recombination and structural proteins involved in meiosis. This provides a starting point to understand in biochemical detail the protein links between structure and function in meiosis. Mutations or variants in the genes encoding such proteins are likely contributors to variations in fertility and to abnormalities in chromosome number.
| Meiosis is a specialised process in which the replicated diploid genome undergoes two rounds of cell division without an intervening DNA replication. Production of haploid gametes from the diploid germ line is a complex process requiring the accurate separation of the two parental genomes to avoid the aneuploidy which would result from errors. Meiotic recombination imposes the additional requirement that the two genomes be precisely aligned for exchange of genetic information. In organisms from budding yeast to humans a key component of the meiotic cellular machinery used to enforce this is the synaptonemal complex (SC). This is a widely occurring, proteinaceous structure which physically links the pairs of sister chromatids (for review see [1]) and is visualised in the electron microscope as a zipper like structure with two lateral elements (LE) and the central element (CE) in between. Lateral elements are derived from axial elements (AE) that connect sister chromatids after premeiotic DNA replication. To date, numerous protein components of the SC have been defined in a variety of organisms (reviewed in [1]). They can be classified as components either of the LE/AE or of the CE. In mammals AE proteins include cohesins and coiled coil domain proteins such as SYCP3 and SYCP2 [2]–[4]. The CE contains the recently described proteins SYCE1, SYCE2 and TEX12 [5],[6]. SYCP1 is a key protein, which links AEs to the CE through its central coiled coil domain and by having C and N terminal globular domains anchored in AE and CE respectively [7]–[9]. In many organisms the formation of the SC is dependent on double strand breaks (DSBs) which can be processed to crossover or, more frequently, non crossover pathways. The SC may play a role in regulating the non random distribution of crossovers known as interference. However the requirement for and intact SC is sexually dimorphic in mice and it is not required for interference in female meiosis [10].
In male mice the fully assembled SC is required to complete crossover recombination and genetic exchange. Mutations in axial element components Sycp2 and Sycp3 result in failure of SC formation and infertility in the male. Milder meiotic defects in female meiosis result in increased aneuploidy and reduced litter sizes [11]–[13]. To date mutagenesis of known components of the CE in mouse suggest that an intact CE is required in both sexes. In Sycp1 null mice synapsis is completely abolished and although the MSH4 foci indicative of intermediate stages of recombination are present neither sex forms the MLH1 foci, which are the cytological markers of crossover, and both sexes are infertile [14]. Syce2 null mice, in which the axial elements align but do not synapse, also do not form MLH1 foci in either sex although again proteins indicative of earlier stages of the recombination process such as RAD51 and MSH4 are present [15]. TEX 12, a central element protein which interacts with SYCE2, has recently been shown to have a similar null phenotype with the absence of crossover recombination in both sexes [16]. Since these proteins are mutually dependent for localisation to and formation of the CE this similarity is not surprising.
Based on known interactions between SYCP1, SYCE1, SYCE2 and TEX12 (Figure S1) we have suggested that the assembly of the SC is a multi-step process which is blocked at different stages by the absence of SYCE1 and 2 and probably TEX12 [15]. In the presence of SYCE2 and the absence of SYCE1 the prediction is that points of synapsis, as observed in the Syce2−/− animals, do not occur. Here we report the phenotype of such mutant animals. Importantly this phenotype has suggested interactions between these structural components of the SC and the recombination machinery.
We disrupted the mouse Syce1 gene by gene targeting in AB2.2 ES cells. The targeting vector was designed to replace exons 2–11 of the Syce1 gene with the LacZ- Neor selection cassette (Figure S2A). Correct targeting was confirmed by Southern Blot analysis (Figure S2B). Correctly targeted ES cells were injected into C56BL/6 blastocysts and produced two germline transmitting chimeras. Offspring produced by mating these chimeras to C56BL/6 females were genotyped by PCR (Figure S2C) and Syce1+/tm1HGU animals intercrossed. Animals were produced from these matings with all genotypes in Mendelian ratios. To confirm the absence of the SYCE1 protein in the Syce1tm1HGU /tm1HGU (Syce1−/−) animals we used Western blotting. A polyclonal antibody raised against C-term of SYCE1 detects a protein band of the expected size (45 KDa) in wild-type testis extracts but not in the Syce1−/−, confirming the specificity of antibodies as well as indicating that the Syce1 disruption described here results in a null mutation (Figure S2D). The lack of detectable proteins demonstrates the absence of splicing between the Neor gene and remaining Syce1 exons which might produce truncated proteins.
Syce1−/− mice are infertile. Mating of both sexes with wild-type animals failed to yield any offspring although Syce1−/− males produced copulatory plugs suggesting normal sexual behaviour. Syce1 mutant ovaries were minute and testes size was only 20–30% of wild-type littermates, which is similar to other meiotic mutants [12], [14]–[16]. We observed no phenotypes in other tissues of these animals.
Histological analysis of adult Syce1−/− gonads revealed an almost complete lack of follicles in ovaries (Figure 1A), suggesting a disruption during meiosis followed by apoptosis, and lack of postmeiotic cells in the testis (Figure 1B). Primary spermatocytes were the most common germ cell type indicating a spermatogenesis arrest at prophase I. Elevated levels of apoptosis were detectable in some tubules by TUNEL staining (Figure 1B, insets) suggesting that arrested cells are eliminated by this mechanism. The high number of positive cells in a fraction of tubules indicates that most of the cells undergo apoptosis at the same epithelial stage, which was determined to be stage IV (data not shown). Syce1−/− females show a meiotic prophase phenotype similar to males indicating that SYCE1 plays the same role in both male and female meiosis. The lack of mature gametes is consistent with the expected role of SYCE1 protein in meiosis and demonstrates that Syce1 is an essential gene for both male and female fertility.
To investigate the cause of the meiotic defect in more detail we prepared surface spread chromosomes from Syce1−/− spermatocytes. Normally during meiotic prophase I homologous chromosomes are closely juxtaposed and are then physically connected by the SC along the entire length of chromosome axes. Immunostaining for SYCP3, SYCP2 and STAG3 proteins revealed that AEs are formed normally in the absence of SYCE1 (Figure 2 and S3) and that homologous chromosomes align in close juxtaposition. The sex chromosomes are an exception to this; as in Sycp1, Tex12 and Syce2 null mutants the pseudoautosomal regions do not pair and a sex body is not formed (Figure 2D, arrows). Wild-type spermatocytes at pachynema are characterised by the presence of ribbon-like structures seen by staining for SYCP1. These represent fully formed SCs linking homologous chromosomes (Figure 2A). In Syce1−/− cells, although AEs are formed and aligned SCs do not assemble between them as indicated by the absence of continuous SYCP1 staining (Figure 2B,D). Interestingly a weak discontinuous SYCP1 signal was observed associated with AE whether they are closely aligned or not (Figure 2B, D). We used immunostaining for SYCE2 and TEX12, two other markers of synapsis that in the wild-type co-localise with SYCP1 (Figure 2E) to further investigate synaptic failure. Although SYCE2 and TEX12 foci co-localise as expected, immunostaining for SYCE2 or TEX12 does not resemble that of the wild-type animals. Instead they were found in intermittent foci between closely aligned AEs (Figure 2F). This is consistent with the observations that their localisation to the SC is co-dependent and their known interactions (Figure S1) [6],[15],[16]. Unlike in wild-type spermatocytes, in Syce1−/− spermatocytes SYCE2 does not always follow SYCP1 signal either locally within a pair of homologs or globally in one nucleus (Figure 2D, B respectively). A subset of cells shows accumulation of SYCP1 on both AEs without accompanying SYCE2, suggesting that the SYCP1 C-terminal region can bind to AEs in the absence of SYCE1. Additionally in Syce1/Syce2 double knockout SYCP1 still binds to aligned AEs suggesting that it is the presence of SYCE1 that restricts SYCP1 binding to synapsed axes when all components are present (not shown). Syce1−/− oocytes display very similar defects in chromosome synapsis to males (Figure 2G–H). AE are fully formed and homologous chromosomes align, however tripartite synaptonemal complex is not formed along the length of chromosomes.
In some cases AEs are in very close apposition along their length with spacing similar to that of the normal SC with SYCE2 and SYCP1 co-localised between them. In order to determine whether these sites of co-localisation of CE proteins represent SC formation we have performed electron microscopy on testis sections from Syce1−/− animals. Extensive analysis of the mutant material revealed presence of parallel AEs but failed to find any signs of the CE (Figure 3). This is in contrast to the Syce2 or Tex12 nulls, where CE-like structures were observed [15],[16]. Based on the observations from all three mutants we propose that the SYCE1 protein is required not only to stabilise SYCP1 dimers within central element but also to stack the transverse filaments into layers to form CE and determine the thickness of the SC.
Meiotic recombination is initiated by SPO11-mediated double strand breaks (DSB) [17]. The generation and the repair of these breaks are required for chromosomal synapsis in most organisms including mammals [18]–[21]. The appearance of these breaks is accompanied by the phosphorylation of histone H2AX on large domains of chromatin around the break. As meiosis proceeds to the pachytene stage γH2AX is removed from synapsed chromosomes and is restricted to the largely asynapsed sex chromosomes in the XY body [22]–[24] (Figure 4A). Syce1−/− spermatocytes showed extensive γH2AX staining in early cells that persisted to the most advanced spermatocyte stages (Figure 4B)(in these animals the sex body does not form). Oocytes show the same pattern of staining (Figure 4J). This suggests that DSB are generated in the Syce1−/− mutants but are not efficiently repaired.
To assess the state of DSB repair in mutant spermatocytes and oocytes we analysed the distribution of proteins involved in different steps of meiotic repair and recombination [25],[26]. First the strand exchange proteins RAD51 and DMC1 are recruited to the sites of DSB and form early recombination nodules (EN). RAD51/DMC1 mediate the homology search and the single end invasion of the homologous chromosome [27]. Cytologically, RAD51 and DMC1 manifest as numerous foci along chromosome cores, typically several hundred occur in a mouse meiotic nucleus [28]. During normal meiosis numbers of RAD51/DMC1 foci peak at leptonema and disappear by mid-pachynema except along asynapsed cores of sex chromosomes in males (Figure 4C and K). RAD51 foci are highly abundant in both Syce1−/− spermatocytes and oocytes and are localised to both aligned and unaligned chromosome cores (Figure 4D and L). Fifteen percent of cells lack RAD51 foci entirely.
The MutS homologs MSH4 and MSH5 have been proposed to function in stabilization or resolution of recombination intermediates and possibly also during synapsis in earlier stages of prophase I [29]–[31]. In normal meiosis MSH4 foci appear concurrently with synapsis at early zygotene, peaking at late zygotene and starting to decrease at early pachytene (Figure 4E and M). In Syce1−/− spermatocytes and oocytes MSH4 foci appear without synapsis and are found only between aligned chromosome cores (Figure 4F and N). This indicates that MSH4/MSH5 mediated DNA-DNA interactions between homologous chromosomes can occur in the absence of SYCE1. Spermatocytes of mice lacking other proteins such as SYCP1 and SYCE2 which are required for synapsis also have MSH4 foci.
After MutS homologs MSH4/MSH5 associate with DNA a complex of MutL homologs MLH1/MLH3 is recruited to sites now termed late recombination nodules (RN). Together they are implicated in the processing of DSB through the double Holliday junction (dHJ) recombination intermediates that result in crossover. Mlh1 was shown to be essential for crossover formation in mammals and yeast [32]–[34]. In wild-type meiosis MLH1 appears at late prophase in pachytene and is present in a few sites that correspond in number and distribution to the number of crossover events estimated genetically [35](Figure 4G and O). We stained Syce1−/− spermatocytes and oocytes with an anti-MLH1 antibody and failed to observe any MLH1 foci (Figure 4H and P). This indicates that despite MSH4 associated recombination intermediates MLH1 can not be recruited to resolve them into crossover in the absence of SYCE1 and synapsis or that cell death occurs before that stage.
Taken together, analysis of the progress of meiotic recombination suggests that SYCE1 is dispensable for the initiation of recombination but is essential for stable homologue interactions mediated by the SC and crossover formation.
Recombination and synapsis are co-dependent and physically linked in yeast where synapsis is initiated at sites of recombination destined to be crossovers [36],[37]. To our knowledge no such link has been described in the mouse.
In Syce1−/− spermatocytes we noticed that the pattern of SYCE2/TEX12 foci between closely juxtaposed AEs resembles that of RAD51. To confirm our observations we immunostained Syce1−/− spermatocytes with anti-SYCE2 and -RAD51 antibodies. A subset of cells (42%, n = 435) with high number of RAD51 foci (approximately two hundred per nucleus) did not have any SYCE2 staining (Figure S4) However, cells with approximately half the number of RAD51 foci, located between aligned AE, showed co-localised staining for SYCE2 (43% n = 435) (Figure 5B). SYCE2 was almost always accompanied by a RAD51 signal in these cells (Figure 5B, lower panel in offset). To test if this co-localisation reflects a biochemical interaction between SYCE2 and RAD51 we used immunoprecipitation (IP) from wild-type and Syce1−/− testicular extracts. We have immunoprecipitated proteins using both anti-SYCE2 antibody and preimmune serum as a control, and checked for interacting proteins by probing western blot with anti-RAD51 antibodies. We were able to detect RAD51 as a band of approximately 37 KDa in the input as well as weakly in the wild-type and Syce1−/− IP samples but not in the control (Figure 6A). As a further control we have used Syce2−/− testis extract for IP with anti-SYCE2 antibodies and failed to detect RAD51(Figure 6B). To check if this interaction is specific and not due to the precipitation of the whole SC we tested SYCE2 IP samples with antiSYCP3 antibodies and did not detect SYCP3 in the immunoprecipitated sample (Figure 6C). Although we detect SYCE2 and RAD51 in the same complex we can not and do not conclude that this interaction is direct. Our attempts to demonstrate that using an in vitro assay have been inconclusive due to insolubility of proteins when co-overexpressed or to RAD51-GST interactions in pull down reactions. We proceeded to check if SYCE2 also co-localises with MSH4 which appears when chromosomes synapse and which succeeds RAD51 in the recombination nodules. Co-immunostaining of Syce1−/− spermatocytes for SYCE2 and MSH4 revealed that these two proteins only partially co-localise. (Figure 5D, and inset). There are different classes of cells: one which has only SYCE2 signals and no MSH4 (7.5%, n = 189, not shown), another which stains for both (36%, n = 189) (Figure 5D) and the remaining largest group shows only MSH4 foci (50%, n = 189) (Figure S4). This would suggest that as RAD51 is displaced by MSH4, SYCE2 is no longer associated with chromosomes in the Syce1−/− animals. Altogether, this data suggests that central clement protein SYCE2 interacts, directly or indirectly, with the recombination protein RAD51.
Is synapsis dependent on the RAD51/SYCE2 interaction? Spo11 null mice are unable to generate meiotic DSB and as a result RAD51 is absent from the nucleus. Despite this, various degrees of synapsis, mostly nonhomologous, were observed in the Spo11 null, on the basis of SYCP1 staining [20],[21]. We have stained Spo11−/− spermatocytes for SYCE1 and SYCE2 to check if these proteins are components of this DSB independent synapsis. Our results show that both SYCE1 and SYCE2 co-localise with SYCP1 on the SC in the Spo11 mutants indicating that apparently normal synapsis can form in the absence of RAD51 and DSB (Figure S5), but in this case between random chromosomes.
Successful completion of meiosis in mouse depends on the assembly of the SC. Recent work using targeted mutagenesis to make null mutations in three (Sycp1, Syce2 and Tex12) of the four known protein components of the CE has shown that the CE is a critical component of this structure [14]–[16]. Here we complete the set by mutating the remaining known component SYCE1. As predicted from the known multiple interactions of the proteins (Figure S1) Syce1−/− animals have a phenotype which is very similar to that of the other three null mutations. DNA repair is incomplete, the SC and the sex body are absent, homologous alignments at variable distances of the AEs occur, early (RAD51) and intermediate (MSH4) markers of recombination are present but there is a complete absence of MLH1 marking crossovers. In the testis cells are eliminated by apoptosis and both sexes are infertile. Complete assembly of the SC is co-dependent on the presence of all four proteins (SYCP1, SYCE1, SYCE2 and TEX12) and perhaps on others as yet undiscovered. However the mice null for different CE components are likely blocked in different states of SC assembly and provide tools to dissect this essential process.
There are distinct features of the Syce1−/− phenotype. In the absence of SYCE1 transverse filament protein SYCP1 binds to AEs when they are closely aligned and presumably forms N-termini associations [9]. This may reflect the protein's ability to form polycomplexes with dimensions corresponding to SCs [38]. However SYCP1 is also associated with AEs that are further apart confirming the proposal in our model that SYCP1 N-terminal associations alone are insufficient to promote SC assembly and require SYCE1 for stability in physiological conditions. The extensive association of SYCP1 with AEs in the Syce1−/− animals suggests that SYCE1 could play a role in restricting SYCP1 binding in wild-type synapsis. These associations with unpaired AEs are absent in the Syce2−/− and Tex12−/− males where SYCE1 is present [15],[16].
The Syce1−/− phenotype further supports the idea that SYCE2 and TEX12 act in concert. From published data we know that their localisation to the SC is co-dependent [15],[16] and in the absence of SYCE1 (this paper) both SYCE2 and TEX12 co-localise as foci between aligned AEs, therefore their recruitment to chromosome axes is SYCE1 independent. Previously, in our model for synaptonemal complex assembly we suggested that SYCE1 stabilises N-terminal interactions of SYCP1 in the CE and that SYCE2/TEX12 is required for the elongation of the SC. The Syce1−/− phenotype is consistent with this model.
Given the presence of three out of four CE components and interactions between SYCP1 and SYCE2 we expected some form of CE to be present in Syce1−/− spermatocytes as found in Syce2−/− and Tex12−/− spermatocytes. Our extensive analysis of testis sections at the EM level failed to detect a CE. Our model for CE assembly was two dimensional, reflecting observations in the light microscope and in EM sections but the SC has a thickness which we had not taken into account and of which SYCE1 may be a component [39]. In a revised model although the three CE proteins (SYCP1, SYCE2 and TEX12) co-localise they do not produce a visible CE in the microscope due to the absence of multiple layers of proteins dependent on SYCE1. We propose that SYCE1 stabilises the N-termini associations of SYCP1 (width) and regulates formation of transverse filament stacking (thickness) in addition to being required for SC extension through its interactions with SYCE2 and SYCP1.
Studies of the SC functions in various organisms revealed that the SC is essential for normal progression of meiotic recombination and formation of crossovers in yeast, plants and mammals [14],[40],[41]. It has been also shown that proper assembly of the SC between homologous chromosomes depends on recombination. In the absence of the SPO11 induced DSBs that initiate recombination, levels of SC formation are highly reduced or form between nonhomologous chromosomes [20],[21]. Additionally, the correct processing of DSBs at the early stages of recombination is essential for synapsis to occur [29],[31],[42],[43]. Impaired recombination in mouse mutants lacking the CE points to the possibility that interactions between the structural components of the CE and the recombination machinery occur and are essential for crossover. Prior to synapsis the recombinase RAD51 is recruited to the DSBs and disappears as chromosomes synapse. In mutants that lack the SC RAD51 persists longer and is associated with the AEs. It is not possible to study the function of RAD51 in meiosis due to embryonic lethality of the Rad51 mutation [44]. However, the phenotypes of recently reported mutations in the Tex15 and Tex11 (Zip4H) genes show that both recruitment as well as timely disappearance of RAD51 are crucial for synapsis and meiotic recombination. In the Tex15 mutant RAD51 foci are highly reduced in number whereas in the Tex11 (Zip4H) mutant the number of these foci increases, probably as a result of delayed processing of DSB. Both mutants show synapsis defects. In Tex11 null some chromosomes do not synapse at all and in Tex15−/− spermatocytes synapsis is completely abolished. As a result the number of MLH1 foci present in spermatocytes is reduced or eliminated, respectively [45]–[47]. In wild-type meiosis several different types of structures containing recombination proteins have been described based on immuno-histochemsitry. In leptotene RAD51/DMC1 foci have been termed early nodules (EN), later they begin to contain RPA in addition to RAD51/DMC1 and when synapsis is complete RAD51 is absent in RPA containing transition nodules (TN). The MLH1 containing recombination nodules (RN) appear last [26].
Based on our observation that SYCE2 and RAD51 co-localise in a subset of the Syce1−/− spermatocytes and that interactions between these proteins can be detected in testis extracts we propose that this interaction promotes synaptonemal complex assembly/extension. From a yeast two hybrid assay and in vitro pull down experiments it was previously suggested that SYCP1 interacts with RAD51 but not with DMC1 [48]. SYCP1 was also shown to recruit SYCE1 and SYCE2 to the SC as these proteins are not chromosomally localised in Sycp1−/− spermatocytes [5],[6] and hence must be involved in the RAD51/SYCE2 interaction. Although all four CE proteins are needed for complete synapsis, structures suggestive of sites of initiation of synapsis can be seen at both light and electron microscope resolution in the absence of SYCE2 or TEX12 but not in the absence of SYCE1. In the SYCE1 null animals we observe co-localisation of SYCE2 and RAD51 which we suggest occurs in normal mouse meiosis but is obscured by the subsequent rapid assembly of the SC. This concentration of SYCE2 may function to promote SC extension. We can not exclude that TEX12, a SYCE2 binding partner, plays a specific role in its interaction with RAD51. Interestingly, it was shown that in DSB deficient mutants, when breaks are introduced artificially, the number of RAD51 foci representing induced DSB correlate with the extent of synapsis [49]. This also points out the link between RAD51 and synapsis. However, it seems that RAD51 is not required in Spo11 mutants for initiation and partial assembly of the SC [20],[21] but in these animals the SC is not formed between homologous chromosomes. Perhaps the presence of RAD51 at the sites of DSB favours the extension of homologous SC assembly over that of non homologous SC in a competitive and (in terms of aneuploidy) potentially disastrous situation.
Feedback from SC assembly must be required for the maturation of a small set of TN into the RN marking sites of recombination. The combination of cytology and enzymology has pointed to the ability of cellular structures to recruit and perhaps modify the function of repair enzymes for use in meiosis [50]. Our results here suggest that this process may also operate in the reverse direction with repair proteins playing a role in the assembly of structures essential for meiosis and fertility.
To inactivate the Syce1 gene, we designed a targeting vector to replace exons 2–11 by selection cassette. This construct was based on a modified pBluescript vector containing DTA cassette, En2SA-IRES-LacZ-pA and floxed tk-NEO gene. A 5.2 kb ApaI fragment containing part of intron 1 of the mouse Syce1 gene was cloned between DTA and LacZ-Neo cassettes and a 2.2 kb SacI fragment containing exons 12–13 of the Syce1 gene was cloned downstream of Neo cassette. The linearised Syce1 targeting construct was electroporated to AB2.2 ES cells. After selection with G418 ES cell clones were screened by PCR (FP: CAACCTCCCTCACCACCTTA, RP: TTGCTGAAGTTGTGCCAGAC). Potential positive clones were expanded and DNA was extracted for Southern blot analysis. DNA was digested with EcoRI and hybridised with external probe (See Figure S2). Cells from one of the correctly targeted ES clones were injected into C57/B6 blastocysts to obtain chimeras. Chimeric males were mated to C57/B6 females and progeny was genotyped using primers (FP:CCAGAAGCCTGAACATCTGACA, RP:TACCATCCTCCATGAGCTGTCT, Neo:AGGACATAGCGTTGGCTACCC). To produce Syce1ko mice we intercrossed heterozygous offspring. Tissues for histological examinations were dissected and fixed in Bouin's fixative. Subsequently, tissues were embedded in paraffin and 6 µm sections were cut. Mounted sections were deparaffinised, rehydrated, and stained with hematoxylin and eosin. Apoptosis was assayed using DeadEnd Fluorometric TUNEL System (Promega) according to the manufacturer's protocol
Spread chromosomes from males and females were prepared and stained as previously described [5], Images were captured using a system comprising a charge-coupled device camera (Orca-AG; Hamamatsu), a fluorescence microscope (Axioplan II; Carl Zeiss MicroImaging, Inc.) with Plan-neofluar objectives (100× NA 1.3), a 100-W Hg source (Carl Zeiss MicroImaging, Inc.), and quadruple band-pass filter set (model 86000; Chroma Technology Corp.), with the single excitation and emission filters installed in motorised filter wheels (Prior Scientific Instruments). Image capture was performed using in-house scripts written for IPLab Spectrum (Scanalytics). Images were processed using Adobe Photoshop.
Electron microscopy was performed using ultra thin sections of testis tissue fixed in 2.5% glutaraldehyde and 1% OsO4 as described previously [51].
The primary antibodies used were: rabbit anti-SYCE1; rabbit anti-SYCE2 [5]; guinea pig anti-SYCE1; guinea pig anti-SYCE2; guinea pig anti-TEX12 [6]; rabbit anti-SYCP1 (Abcam); mouse anti-SYCP3 [52]; rabbit anti-SYCP3 (Abcam); rabbit anti-STAG3 [53]; rabbit anti-SYCP2 [54]; rabbit anti-γH2AX (Upstate Biotechnology); mouse anti-Rad51 (Upstate Biotechnology); mouse anti-MLH1 (BD Biosciences); rabbit anti-Msh4 (Abcam). Secondary antibodies used were Alexa Dyes (AlexaFluor-488, 594 and 647) conjugates (Molecular Probes).
Protein extraction, immunoprecipitation and detection were carried out as previously described [5]
|
10.1371/journal.ppat.1000071 | Role for A-Type Lamins in Herpesviral DNA Targeting and Heterochromatin Modulation | Posttranslational modification of histones is known to regulate chromatin structure and transcriptional activity, and the nuclear lamina is thought to serve as a site for heterochromatin maintenance and transcriptional silencing. In this report, we show that the nuclear lamina can also play a role in the downregulation of heterochromatin and in gene activation. Herpes simplex virus DNA initiates replication in replication compartments near the inner edge of the nucleus, and histones are excluded from these structures. To define the role of nuclear lamins in HSV replication, we examined HSV infection in wild-type and A-type lamin–deficient (Lmna−/−) murine embryonic fibroblasts (MEFs). In Lmna−/− cells, viral replication compartments are reduced in size and fail to target to the nuclear periphery, as observed in WT cells. Chromatin immunoprecipitation and immunofluorescence studies demonstrate that HSV DNA is associated with increased heterochromatin in Lmna−/− MEFs. These results argue for a functional role for A-type lamins as viral gene expression, DNA replication, and growth are reduced in Lmna−/− MEFs, with the greatest effect on viral replication at low multiplicity of infection. Thus, lamin A/C is required for targeting of the viral genome and the reduction of heterochromatin on viral promoters during lytic infection. The nuclear lamina can serve as a molecular scaffold for DNA genomes and the protein complexes that regulate both euchromatin and heterochromatin histone modifications.
| Transcription of eukaryotic genes is regulated by sequence-specific DNA-binding proteins that bind to the DNA as well as the structure of the chromatin associated with the specific gene. Posttranslational modification of histones plays a major role in the higher order structure of the chromatin and whether it serves as heterochromatin or inactive chromatin versus euchromatin or active chromatin. The nuclear lamina has been shown to promote the maintenance of heterochromatin in mammalian cells, but little is known about where heterochromatin is modulated. In this study, we find that the A-type lamins are required for the targeting of herpes simplex virus genomic DNA to the periphery of the infected cell nucleus and for the prevention or reduction of heterochromatin on the viral genome and transcriptional silencing of the viral genome. This raises the broader function of the nuclear lamina in the regulation of both euchromatin and heterochromatin. We propose that the nuclear lamina is a platform for the organization of chromatin remodeling and histone modification enzymes that regulate both euchromatin and heterochromatin.
| Herpes simplex virus (HSV) undergoes productive infection through transcription and replication of its viral genome within the nucleus [1]. HSV gene expression involves the temporal expression of immediate-early (IE), early (E), and (L) genes [2] and the sequential remodeling of the infected cell nucleus by viral proteins [3]. One of the earliest demonstrations of the compartmentalization of nuclear processes, such as DNA replication, was the observation of replication compartment formation in the nuclei of HSV-infected cells. HSV replication compartments are the site of viral DNA replication, late gene transcription, and viral DNA encapsidation [4]. Replication compartments and parental genome complexes form at the nuclear periphery during early times of infection [4]–[6]. Furthermore, lamin A/C and the nuclear envelope emerin protein co-precipitated with the HSV DNA replication protein ICP8 in a proteomics study [7], suggesting an association of the replication compartment with the nuclear lamina and/or nuclear envelope. The nuclear lamina is disrupted at late times postinfection [8],[9], at least in part to allow access of the nucleocapsids to the inner nuclear envelope for budding and primary envelopment. However, nothing is known about the role of the nuclear lamina at early times postinfection.
A-type and B-type lamins are major components of the nuclear lamina that underlies the inner nuclear membrane and provides structural integrity to the nucleus [10]. The A-type lamins (lamins A, C, AD10, and C2) are expressed in differentiated cells and are encoded by the LMNA gene, whose products are encoded by transcripts generated by alternative splicing [11]. A-type lamins are found along the inner side of the nuclear envelope and within the nucleoplasm where they form a veil-like structure [12],[13]. Lamins are believed to function in higher order chromatin organization by acting as part of a molecular scaffold with integral membrane proteins to tether peripheral heterochromatin and chromatin remodeling complexes to the nuclear envelope [14],[15]. Evidence of lamin A/C function in chromatin organization has been provided by studies showing that mutations in the human LMNA gene lead to premature aging and progressive loss of heterochromatin [16],[17], indicating a role for the nuclear lamina in heterochromatin maintenance. Furthermore, immortalized mouse embryonic fibroblasts from Lmna−/− knockout mice exhibit alterations in nuclear envelope integrity, mislocalization of lamin binding proteins, and reduced peripheral heterochromatin [18],[19]. Targeting of genes to the nuclear periphery has been associated with gene silencing in several cases [20],[21]; however, in other cases the movement of active genes to the periphery is believed to be due to association of actively transcribing genes with nuclear pores [22],[23] and not the nuclear lamina. Hence, the role of nuclear targeting in regulation of gene expression remains to be fully defined. The available data suggest a role for the nuclear lamina in maintenance of heterochromatin and gene silencing.
Viruses have served as sensitive probes for the study of mechanisms of cellular processes, and chromatin plays an important role in regulation of HSV gene expression [24] . HSV DNA in the virion is not associated with histones [25]. As DNA enters the nucleus, cellular mechanisms attempt to silence the incoming genome through assembly of heterochromatin onto the DNA molecules, as first observed in transfected cells [26]. The HSV VP16 tegument protein plays a role in reducing histone association with viral DNA and in increasing the euchromatin modifications on the histones associated with viral genes [27],[28]. The HSV immediate-early (IE) ICP0 protein acts as an inhibitor of histone deacetylases [28]–[30]. In terms of the cellular mechanisms regulating chromatin structure, little is known about the nuclear location where chromatin modification takes place or is regulated beyond the role of the nuclear lamina in heterochromatin maintenance. This study provides evidence that lamin-dependent targeting of the HSV genome to the nuclear periphery is associated with a reduction of heterochromatin on viral lytic genes.
Based on the localization of early viral replication compartments at the nuclear periphery and the co-precipitation of lamin A with the HSV ICP8 DNA replication protein, we hypothesized that the nuclear lamina plays a role in HSV transcription and DNA replication through recruitment of viral DNA and assembly of replication compartments at the inner nuclear membrane at early times postinfection. To define the role of lamin A/C in the formation of replication compartments in the nuclei of HSV-infected cells, we examined HSV infection in WT (Lmna+/+) and lamin A/C knockout (Lmna−/−) immortalized mouse embryonic fibroblasts (MEFs) [18]. We first used immunofluorescence to define the role of lamin A/C in the assembly of viral replication compartments. Lmna+/+ and Lmna−/− MEFs were either mock-infected or infected with HSV at a multiplicity of infection (MOI) of 10 PFU/cell, fixed at 8 hours post-infection (hpi), and stained with antibodies specific for the HSV ICP8 DNA replication protein and for histone H1. Mock-infected MEFs showed diffuse intranuclear histone H1 staining in both Lmna+/+ and Lmna−/− cells, but the Lmna−/− cells showed reduced H1 staining near the nuclear envelope, consistent with reduced chromatin attachment to the nuclear envelope (Figure 1A, panels a and c). HSV-infected Lmna+/+ MEFs contained intranuclear replication compartments, as evidenced by ICP8 staining at 8 hpi, which filled much of the interior of the nucleus and excluded histone H1 to the periphery and certain internal regions of the nucleus (Figure 1A, panels b, f and j). Surprisingly, Lmna−/− MEFs infected with the same amount of virus showed fewer cells containing replication compartments as detected by immunofluorescence, and the compartments observed were much smaller (Figure 1A, panels d, h and l). Second, the punctate ICP8 foci were more densely packed in the Lmna−/− cells than in the Lmna+/+ cells (Figure 1A). Third, histone H1 was not segregated from the small replication compartments (same panels). Finally, the nuclei did not enlarge in Lmna−/− infected MEFs, in contrast to what was observed previously in primate cells [31] and in Lmna+/+ MEFs (Figure 1A). Similar experiments looking at the IE ICP4 transactivator protein at 4 hpi also showed smaller replication compartments and a diffuse distribution of histone H1 in Lmna−/− MEFs as compared with Lmna+/+ MEFs (Figure 1B). The smaller replication compartments observed in Lmna−/− cells were also observed at later times postinfection, e.g., 12 hpi (results not shown).
To examine the role of lamin A/C in the intranuclear location of replication compartments, we infected Lmna+/+ and Lmna−/− MEFs with HSV at a low MOI for 36 hours to allow for the development of discrete plaques. Plaques were smaller on the Lmna−/− cells and formed at an 8-fold lower efficiency on Lmna−/− cells as compared with Lmna+/+ cells (L. Silva and D. Knipe, unpublished results). Previous studies had shown that in cells at the periphery of a developing plaque, replication compartments and genome complexes form along the inner nuclear envelope nearest the plaque [5],[6]. This was likely due to tethering of the viral genome and/or replication compartments at the nuclear periphery near the nuclear pore where the viral genome enters the nucleus. Immunofluorescence detection of the HSV immediate-early ICP4 protein was used to define early complexes as ICP4 is reported to associate with the parental viral genome [6], and detection of ICP8 was used to define early replication compartments [3]. In Lmna+/+ MEFs at the edge of a plaque, developing replication compartments, as detected by ICP4 and ICP8 immunofluorescence, were assembled within the nucleus in an asymmetric distribution along one edge of the nucleus nearest the plaque (Figure 2A). However, this asymmetric ICP4 and ICP8 distribution was lost in the absence of lamin A/C (Figure 2B). To quantify this difference, we scored Lmna+/+ and Lmna−/− MEFs according to the distribution of ICP4 foci. Lmna−/− MEFs displayed a 5-fold decrease in asymmetric ICP4 foci distribution as compared with Lmna+/+ MEFs (Figure 2C). These results argued that a loss of lamin A/C may lead to an inability of the viral genomes to target to the nuclear periphery due to the absence of lamins or lamin-associated proteins that are required for recruitment of the incoming parental genomes, which ultimately develop into replication compartments.
The reduced levels of ICP8 immunofluorescence in HSV-infected Lmna−/− MEFs suggested that viral early gene expression was reduced. We therefore measured viral RNA and protein levels in Lmna+/+ and Lmna−/− MEFs by northern and western blotting, respectively. Viral ICP27 (IE) and ICP8 (E) mRNA levels were reduced in Lmna−/− MEFs at 4 hpi (Figure 3A, lane 4). In addition, we observed that levels of ICP8 were reduced in Lmna−/− MEFs as early as 4 hpi and showed at least a 3-fold reduction at 8 hpi, as compared to the Lmna+/+ MEFs (Figure 3B, lanes 4 and 6). These results argued that lamin A/C was required for early viral gene expression. Similar reductions in the immediate early proteins ICP0, ICP4, and ICP27 were observed in infected Lmna−/− MEFs as compared with Lmna+/+ MEFs (Figure 3B, lanes 4 and 6). Thus, the earliest defect in viral gene expression in the Lmna−/− MEFs was reduced expression of IE genes.
The increased levels of histone H1 co-localizing with replication compartments suggested that the reduced level of viral gene expression might be due to repressive effects of chromatin on viral genes. During productive infection with wild type virus, limited amounts of nucleosomes are associated with the viral genome [32],[33]. Histones that are associated with HSV DNA during productive infection have modifications that allow for active transcription [27]. Mature replication compartments exclude histone H1 [31] and cause the marginalization of the host chromatin [8]. To determine if heterochromatin was associated with viral replication compartments, we first examined mock- or HSV-infected cells using immunofluorescence to detect the trimethylated form of histone H3 lysine 9 (H3K9Me3) and histone H4 lysine 20 (H4K20Me3), both markers of heterochromatin. In Lmna+/+ infected cells, heterochromatin was excluded from replication compartments (Figure 4A and 4B, panels b, f and j). In contrast, heterochromatin appeared coincident with the small replication compartments observed in HSV-infected Lmna−/− MEFs (Figure 4A and 4B, panels d, h and l), suggesting that heterochromatin was associated with replication compartments in these cells. Similar results were seen in immunofluorescence experiments detecting heterochromatin protein 1α (HP-1α), which recognizes and binds to trimethyl H3K9 (Figure 4C).
To further test the hypothesis that association of heterochromatin with viral promoters in HSV-infected Lmna−/− MEFs inhibited gene expression, we conducted chromatin immunoprecipitation (ChIP) experiments. The amount of HSV DNA associated with histones was measured by ChIP using antibodies specific for histone H3 or the heterochromatin markers trimethyl H3K9, H3K27, and H4K20. The immunoprecipitated DNA was quantified by real-time PCR for the ICP4 gene transcription start site and mouse GAPDH gene promoter sequences [34]. The relative amounts of viral promoters associated histones bearing the different modifications were expressed as the fraction of viral promoter sequence immunoprecipitated with the specific antibody normalized to the fraction of GAPDH DNA immunoprecipitated in the same reaction. The levels of viral DNA associated with histone H3, and thus with total chromatin were less than that for GAPDH but similar for Lmna+/+ and Lmna−/− MEFs (Figure 5). In contrast, there was a 65-fold increase in trimethyl H3K9, a 6-fold increase of trimethyl H3K27 and a 23-fold increase in H4K20 associated with the ICP4 promoter sequences in Lmna−/− MEFs as compared with the Lmna+/+ MEFs (Figure 5).
To confirm that association of heterochromatin with HSV DNA and replication compartments was truly the result of the lamin deficiency in Lmna−/− MEFs, we examined the distribution of heterochromatin in Lmna−/− cells transfected with plasmids encoding GFP or GFP-lamin A. In Lmna−/− cells expressing GFP, we observed co-localization of heterochromatin (HeK9Me3) with replication compartments (ICP8) (Fig. 6, panels a–e), as described above. In contrast, in Lmna−/− cells expressing GFP-lamin A, there was increased expression of ICP8, and replication compartments, as evidenced by ICP8 staining, were larger and showed an exclusion of heterochromatin (Fig. 6, panels f–j). Therefore, the observed changes in ICP8 expression, replication compartment formation, and heterochromatin distribution reverted to wild-type by expression of lamin A, arguing that the mutant cell phenotype was due to the absence of lamin A. These results in total support an important role for lamin A in reduction of heterochromatin on HSV DNA during lytic infection.
The reduced levels of IE and E viral gene expression in the Lmna−/− MEFs predicted that the replication cycle was not being completed efficiently in these cells. Viral DNA replication was examined by real-time PCR measurement of viral DNA levels. Consistent with the reduced viral gene expression and small replication compartments, viral DNA replication was reduced by at least 3-fold in Lmna−/− MEFs compared with Lmna+/+ MEFs at 8 and 16 hpi (Figure 7A). Similar viral DNA levels were seen in Lmna+/+ and Lmna−/− MEFs infected in the presence of the viral DNA inhibitor phosphonacetate (PAA) for 2 hpi, arguing that the amounts of viral DNA entering the two cell types were equivalent. Viral growth was assayed by measurement of viral yields in infections of Lmna+/+ and Lmna−/− MEFs at different multiplicities of infection (MOI). Viral yields were reduced by approximately 5-fold in Lmna−/− MEFs infected at an MOI of 10 (plaque-forming units per cell) as compared to Lmna+/+ MEFs at 8–24 hpi (Figure 7B). Therefore, viral replication was reduced modestly in infections performed at high MOI. However, at low MOI (0.01 PFU/cell), there was an approximately 100-fold reduction in viral yield in Lmna−/− MEFs compared to Lmna+/+ MEFs, arguing that the magnitude of the requirement for lamin A/C in HSV replication was multiplicity-dependent.
To ensure that the chromatin phenotype was also observed at low MOI, we examined replication compartment formation and heterochromatin distribution in cells infected at low MOI (0.1). At 8 and 12 hours postinfection in Lmna+/+ cells, replication compartments were observed that nearly filled the infected cell nucleus and heterochromatin was marginalized along the inner nuclear membrane (Fig. 8, panels a–f). In contrast, at 8 and 12 hours postinfection in Lmna−/− cells, replication compartments were small and co-localized with heterochromatin (Fig. 8, panels g–l). Therefore, lamin A/C is needed for replication compartment formation and heterochromatin exclusion at both low and high MOI's of infection but plays a more essential role at low MOI.
Heterochromatin is associated with the nuclear lamina, and A-type lamins have been shown to promote the maintenance of heterochromatin in mammalian cells. Thus, it is believed that the nuclear lamina is the site of heterochromatin maintenance. However, little is known about the sites or structures involved in modulation of heterochromatin. In this study we find that the A-type lamins are required for targeting of herpes simplex virus genomic complexes to the periphery of the infected cell nucleus and for preventing or reducing heterochromatin on the viral immediate-early lytic gene promoters. This raises the potential of a broader role for the nuclear lamina in the regulation of both euchromatin and heterochromatin. We propose that the nuclear lamina is a platform for the organization of chromatin remodeling and histone modification enzymes that regulate both euchromatin and heterochromatin. In HSV-infected cells, viral regulatory proteins shift the activity of these chromatin regulatory complexes to prevent assembly of or reduce heterochromatin on the viral genome so that optimal viral gene transcription can occur.
During lytic infection, only limited amounts of nucleosomes are associated with viral DNA [32],[33]. Furthermore, the histones that are associated with viral DNA bear euchromatic modifications [27],[35]. Viral gene products are believed to play a role in regulating histone association and chromatin modification on HSV DNA [1],[24]. In this study we have shown that the host nuclear lamin A/C gene products are required for histone modifications that occur on the ICP4 gene promoter. We speculate that viral proteins, such as VP16 and ICP0, function on the nuclear lamina or in the nucleoplasmic lamin to organize enzymatic complexes that carry out euchromatic modifications of histones on the HSV genome.
We have demonstrated that the type A lamins are required for targeting of the HSV genome to the nuclear periphery for assembly of the early replication compartments, as shown previously at early times of infection [4] and in cells along the edge of a plaque [5],[6]. Localization to the nuclear periphery is correlated with reduced levels of heterochromatin on viral IE gene promoters, arguing that viral DNA located at the nuclear periphery is protected from chromatin silencing by the host cell machinery. The HSV VP16 virion protein and the ICP0 IE protein have been shown to play roles in promoting the acetylation of histone H3 on HSV DNA. ICP0 is not required for localization of viral genomes to the nuclear periphery [6], but there is no information about VP16 as yet. Thus, the viral and cellular proteins involved in tethering HSV DNA on the nuclear periphery remain to be defined. Also, the stage in viral replication at which the HSV genome associates with the nuclear lamina or nuclear periphery is not known. Association of the viral genome with the nuclear lamina could occur at the time of IE gene transcription, E gene transcription or initiation of viral DNA replication, although our data and those of others [6] argue that this may occur at or before IE gene transcription.
Replication of the genomes of RNA viruses in the cytoplasm has also been proposed to occur on a surface but in that case on membranes (reviewed in [36]). Thus, nuclear DNA viruses may use the nuclear lamina and inner nuclear envelope as a platform for replication while cytoplasmic RNA viruses use cytoplasmic membranes as a platform for replication. It has been proposed that these surfaces provide a two-dimensional lattice or platform for assembly of replication complexes [37]. The nuclear lamina may play an additional role in providing a platform for recruitment of viral DNA as well as chromatin-modifying enzymes that keep the viral genome in an active chromatin conformation.
We found that the requirement for lamin A by HSV replication was multiplicity-dependent in that the reduction of replication in Lmna−/− cells was about 5-fold at high MOI (10 PFU/cell) while at low MOI (0.01 PFU/cell), the reduction in Lmna−/− cells was approximately 100-fold. Defects in replication compartment formation and heterochromatin association with replication compartments were observed under both conditions; thus, we believe that A-type lamins exert similar effects on viral replication at low and high MOI. At high MOI, however, the virus can circumvent the heterochromatin block. It is conceivable that at high MOI the large number of input viral genomes titrates out the finite amount of histones in the infected cell and the genomes are transcribed. Alternatively, at high MOI, the increased numbers of viral genomes eventually encounter histone-modification enzymes by less efficient means than in the assemblies located on the nuclear lamina. It is worth noting that the replication requirement for ICP0, which inhibits histone deacetylases [38], is also multiplicity-dependent [39],[40]. Therefore, the viral and cellular functions that HSV uses to prevent chromatin silencing appear to be more important at lower MOI's.
Previous studies have largely documented a role for the A type lamins in maintenance of heterochromatin. Mutations in the human LMNA gene lead to premature aging and progressive loss of heterochromatin [16],[17], while immortalized mouse embryonic fibroblasts from Lmna−/− knockout mice exhibit alterations in nuclear envelope integrity, mislocalization of lamin-binding proteins, and reduced peripheral heterochromatin [18],[19]. In contrast, our results argue that type A lamins are necessary for preventing assembly or for removal of heterochromatin on HSV IE genes during lytic infection. Although these results may seem to be inconsistent, we propose that lamin A can serve as a platform for the organization of enzyme complexes that, under the appropriate conditions, can lead to heterochromatin or euchromatin formation on DNA sequences associated with the lamina. We further hypothesize that during HSV lytic infection viral gene products act to shift the balance towards euchromatin through the assembly of chromatin and enzyme complexes on viral lytic genes associated with the nuclear lamina that lead to euchromatic modifications of histones. In contrast, during latent infection the HSV latency-associated transcript promotes the assembly of heterochromatin on viral DNA during latent infection [34]. Thus, by regulating the type of chromatin on the viral chromosome, HSV determines whether it will undergo a productive or latent infection in different cell types. Further studies should determine the precise mechanism by which the nuclear lamins are exploited by HSV to keep its genome transcriptionally active during productive infection. These studies should provide the basis for mechanisms operative on cellular chromatin as well.
Immortalized Lmna−/− murine embryonic fibroblasts (MEFs) and litter-matched Lmna+/+ control MEFs were provided by Brian Kennedy, University of Washington [18]. Cells were grown in Dulbecco's modified Eagle medium (DMEM; Gibco) supplemented with 5% fetal bovine serum (FBS)+5% bovine calf serum (BCS), 2 mM L-glutamine, 100 U/ml penicillin and 100 µg/ml streptomycin at 37°C in 5% CO2. Wild type HSV-1 KOS strain virus was grown and titrated on Vero cells as described previously [8] and used for infections at the multiplicity of infection (MOI) as described. Cells were seeded one day before infection. Virus was diluted in phosphate-buffered solution (PBS) containing 0.1% glucose and applied to cells for 1 h at 37°C. The cells were washed three times for 30 seconds with an acid wash buffer (135 mM NaCl, 10 mM KCl, 40 mM citric acid buffer, pH 3), and washed with DMEM before incubation in DMEM-1% FBS medium at 37°C for the indicated periods of time. For viral growth curve experiments, HSV-1 KOS was used to infect Lmna+/+ and Lmna−/− MEFs at a multiplicity of infection (MOI) of 10 or 0.01 PFU/cell. At 1 hpi, cells were washed three times for 30 seconds with acid wash buffer before incubation in DMEM plus 1% FBS for the indicated time period.
Lmna+/+ and Lmna−/− MEFs were seeded at 1×105 cells/well on glass coverslips in 24-well plates overnight at 37°C prior to infection at the indicated MOI for immunofluorescence experiments as described previously [34]. Cells were incubated with antibodies specific for histone H1 (Upstate), histone H3K9 (Abcam), histone H4K20 (Abcam), heterochromatin protein 1α (Cell Signal Technology), HSV-1 ICP4 4040II rabbit polyclonal (Kent Wilcox), HSV-1 ICP4 mouse monoclonal (Abcam), HSV-1 ICP8 mouse monoclonal 39S [41], or HSV-1 ICP8 rabbit polyclonal 3-83 [42]. Secondary antibodies conjugated to Alexa 594, Alexa 488, and Alexa 350 dyes and prolong antifade reagent were obtained from Molecular Probes Inc. Cells were imaged on an Axioplan 2 microscope (Zeiss) with a 63× objective and Hamamatsu CCD camera (model C4742-95). Images were deconvolved using the inverse filter algorithm in the Axiovision (Rel.4.5) software.
Plasmid pGFP-LA, kindly provided by D.M. Gilbert, contains lamin A cDNA cloned into the pEFGP-C1 expression vector [43]. pEGFP-C2 (Clontech) described as pGFP for simplicity, was used for the expression of GFP. Three days prior to infection, Lmna−/− MEFs were seeded at 5×105 cells per well in a 24-well plate with glass coverslips. Transfections were performed on day two, using Genejuice (Novagen) and 0.5 µg of plasmid pGFP-C2 or pGFP-LA DNA per well diluted in Optimem (Invitrogen) and 1% DMEM media without antibiotics. At day 4, or 48 hrs post-transfection, cells were infected with HSV-1 at an MOI of 20. Cells were fixed for immunostaining at 8 and 12 hpi and processed as described above.
Plasmid vectors pCI-ICP27 [44] and pBS-ICP8 (Kevin Bryant, unpublished results) were used to generate hybridization probes for the ICP27 and ICP8 mRNAs, respectively. The plasmid inserts were labeled with 32P dCTP (Perkin-Elmer) using Ready-To-Go DNA labeling beads (Amersham). Unincorporated nucleotides were removed from the probe using a Microspin G-50 column (GE Healthcare).
Lmna+/+ and Lmna−/− MEFs were either mock-infected or infected with HSV at an MOI of 20. RNA was extracted using 1mL of Trizol LS reagent (Invitrogen) per 100 mm dish. For Northern blotting, 10 µg of RNA was denatured in a solution of 50% formamide, 1.1 M formaldehyde, and 1mg/ml ethidium bromide and subjected to electrophoresis in a 10% agarose gel containing 1% formaldehyde in 1× MOPS buffer as described previously [45]. The RNA was transferred to a nitrocellulose membrane overnight in 20× SSC. The blot was incubated in QuickHyb solution (Stratagene) for 15 minutes at 68°C and then in a solution containing 32P labeled probes and 20 mg/ml denatured salmon sperm DNA at 68°C for 1 h. The blot was washed twice with 2× SSC-0.1% SDS for 15 minutes at 60°C and once with 0.1×SSC-0.1% SDS for 30 minutes at 60°C. The blots were exposed to a phosphorimager screen (Amersham) overnight.
Lmna+/+ and Lmna−/− MEFs were either mock-infected or infected with HSV-1 at an MOI of 20. At 4 and 8 hpi, cells were harvested in Laemmli sample buffer containing one protease inhibitor cocktail tablet (Roche) per 10 ml and boiled for 5 minutes. Aliquots of whole cell lysates were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), and proteins were electrically transferred to a nitrocellulose membrane. Membranes were blocked with a solution of 5% milk in PBS for an hour at room temperature and washed three times for 5 minutes in Tris-buffered saline with Tween-20 (TBST). Membranes were incubated for 2 hours at room temperature with antibodies specific for ICP0 (1∶1000), ICP4 (1∶1000), ICP8 (3-83; 1∶5000), ICP27 (1∶10,000), lamin A/C (1∶5000) or actin (1∶1000) diluted in TBST. Membranes were washed three times for 5 minutes in TBST prior to a two-hour incubation at room temperature with secondary antibodies conjugated to horseradish peroxidase diluted 1∶1000. Horseradish peroxidase signal was detected using chemiluminescence reagents (ECL; Amersham) and luminescence was detected using X-ray film (Kodak).
Lmna+/+ and Lmna−/− MEFs were seeded at 3×106 cells per 100 mm dish overnight at 37°C and were either mock infected or infected with HSV-1 at an MOI of 20. At 4 hpi cells were fixed with formaldehyde (final concentration 1% v/v) and fixation stopped with glycine (125 mM) [35]. The cells were collected by centrifugation, resuspended in lysis buffer (1% SDS; 10 mM EDTA; 50 mM Tris-HCl, pH 8.1), and incubated on ice for 10 minutes. The cell lysates were sonicated to shear DNA into lengths of ∼400 bp. The sheared chromatin was diluted 10-fold in radioimmunoprecipitation assay (RIPA) lysis buffer (0.1% SDS/1% sodium deoxycholate/150 mM NaCl/10 mM Na2PO4/2 mM EDTA/1% Nonidet P-40) with protease inhibitors. The diluted chromatin was pre-cleared with protein A agarose beads (Upstate) for 2 hours at 4°C with rotation followed by centrifugation. An aliquot (1%) of each chromatin supernatant was reserved as the input sample. The chromatin supernatant was incubated with 2.5 µg of antibody specific for histone H3 (Abcam) or 5 µg of antibodies specific for histone H3 lysine 9 trimethyl (Abcam), histone H3 lysine 27 trimethyl (Upstate), or histone H4 lysine 20 trimethyl (Abcam) at 4°C with rotation. An aliquot was incubated without antibody as a control to determine background binding. Immunocomplexes were collected by incubation with protein A agarose beads for 30 minutes at 4°C with rotation. The beads were washed three times for 5 minutes at room temperature with a low salt wash buffer (150 mM NaCl; 20 mM Tris-HCl, pH 8.1; 2 mM EDTA; 1% Triton X-100; 0.1% SDS) with protease inhibitors, followed by one wash for 5 minutes with a high salt buffer (500 mM NaCl; 20 mM Tris-HCl, pH 8.1; 2 mM EDTA; 1% Triton X-100; 0.1% SDS) with protease inhibitors. Immunocomplexes were eluted by incubation at 65°C for 30 minutes and room temperature for 15 minutes with fresh elution buffer (1%SDS; 0.1 M NaHCO3). Crosslinks were reversed by incubation for 4 hours at 65°C with a final concentration of 0.2M NaCl. The eluates were incubated with proteinase K, and DNA was purified by phenol: chloroform extraction, ethanol precipitation, and used as a template for real-time PCR.
Real-time PCR was performed by using SYBR Green and an ABI Prism 7700 sequence detection system (Applied Biosystems) as described previously [34]. PCR reactions were conducted for 40 cycles (30 s at 95°C, 60 s at 60°C) in duplicate. Mouse GAPDH Primers were previously described [34]: (GeneBank accession no. NML008084 nucleotides 781–900: 5′-CAATGT-GTCCGTCGTGGATCT-3′ and 5′-T TGAAGTCGCAGGAG-ACAACC-3′) and ICP4 gene transcript (nucleotides: 131105-131160: 5′-GCCGGGGCGCTGCTTGTTCTCC-3′ and 5′-CGTCCGCCGTCGCAGCCGTATC-3′). The amount of DNA precipitated in the ChIP assays was quantified by comparison with a standard curve, which was obtained by running a 10-fold dilution series of ICP4 or mGAPDH DNA. The amount of DNA in the no antibody control was subtracted from the amount immunoprecipitated by the appropriate antibody. The fraction of ICP4 DNA immunoprecipitated compared to the input sample was normalized to the fraction of GAPDH immunoprecipitated in the same reaction, and this value is defined as fold enrichment/GAPDH.
For quantification of viral DNA replication, cells were infected as described above and the DNA extracted using the DNeasy kit (Qiagen). Aliquots of DNA (100 ng) were used for real-time PCR and the samples run in duplicate. Viral DNA levels at each time point were quantified relative to the 2 hour postinfection sample by the ΔCt method as described [34]. To determine the relative DNA content at various times, average Ct values for the ICP4 promoter primer were subtracted by the average Ct values for GAPDH. The calibrator value (HSV sample 2-hpi) was subtracted by the GADH Ct value. To obtain the ΔΔCt value, the Ct value was subtracted by the Ct value of the input time point. ΔΔCt = (Cttest-Ctreference)−(Ct2H ICP4-Ct2H GAPDH). The fold enrichment value is 2−ΔΔCt.
ICP0: NC_001806; NP_044660; (GeneID: 2703390)
ICP4: NC_001806; NP_044662; (GeneID: 2703392)
ICP8 (UL29): NC_001806; NP_044631; (GeneID: 2703458)
ICP27 (UL54): NC_001806; NP_044657; (GeneID: 2703426)
Mouse GAPDH: NC_000072; NP_032110 (GeneID: 14433)
Human Lamin A/C: NC_000001; NP_733821; (GeneID: 4000)
Mouse Lamin A/C: NC_000069; NP_001002011; (GeneID: 16905) |
10.1371/journal.ppat.1007053 | The adaptor molecule CD2AP in CD4 T cells modulates differentiation of follicular helper T cells during chronic LCMV infection | CD4 T cell-mediated help to CD8 T cells and B cells is a critical arm of the adaptive immune system required for control of pathogen infection. CD4 T cells express cytokines and co-stimulatory molecules that support a sustained CD8 T cell response and also enhance generation of protective antibody by germinal center B cells. However, the molecular components that modulate CD4 T cell functions in response to viral infection or vaccine are incompletely understood. Here we demonstrate that inactivation of the signaling adaptor CD2-associated protein (CD2AP) promotes CD4 T cell differentiation towards the follicular helper lineage, leading to enhanced control of viral infection by augmented germinal center response in chronic lymphocytic choriomeningitis virus (LCMV) infection. The enhanced follicular helper differentiation is associated with extended duration of TCR signaling and enhanced cytokine production of CD2AP-deficient CD4 T cells specifically under TH1 conditions, while neither prolonged TCR signaling nor enhanced follicular helper differentiation was observed under conditions that induce other helper effector subsets. Despite the structural similarity between CD2AP and the closely related adaptor protein CIN85, we observed defective antibody-mediated control of chronic LCMV infection in mice lacking CIN85 in T cells, suggesting non-overlapping and potentially antagonistic roles for CD2AP and CIN85. These results suggest that tuning of TCR signaling by targeting CD2AP improves protective antibody responses in viral infection.
| Enhancing the production of protective antibodies in response to infection or vaccine is critically important for host protection. However, we have only limited knowledge about molecular targets to enhance functions of CD4 helper T cells that are essential for antibody affinity maturation and class switching. In this work, we found that inhibiting the function of the adaptor molecule CD2AP results in enhanced antibody responses and improved protection of mice from chronic infection by LCMV. Mice lacking CD2AP specifically in T cells showed enhanced CD4 T cell differentiation towards the follicular helper subset, which is a critical regulator of antibody responses, and generated more germinal center B cells leading to production of mutated, protective antibodies. This effect was specific to CD4 T cells in type-I immune responses, associated with viral infection, while deletion of CD2AP had little impact on CD4 T cells in type-II immune responses or CD8 T cells. Our results thus suggest that CD2AP can be a specific target to enhance antiviral protective immunity during viral infection or vaccination.
| CD4 T lymphocytes are critical mediators of the adaptive immune response to infection [1]. Upon encountering cognate antigen as peptide-MHC (pMHC) complexes on antigen presenting cells in the lymphoid organs, they initiate clonal expansion and undergo differentiation into effector cells, depending on the cytokine milieu initially established by innate immune cells [2]. In response to viral infection, differentiation of activated CD4 T cells is directed towards an IFN-γ-producing subset, referred to as TH1, that also produces IL-2 and TNF-α and enhances cellular immune responses by CD8 T cells and macrophages. Activated CD4 T cells also differentiate into the follicular helper (TFH) cells that migrate to germinal centers (GCs) [3]. During viral infection, TFH cells provide help signals to B cells for antibody affinity maturation through their expression of co-stimulatory ligands and cytokines, such as CD40L and IL-21, and direct immunoglobulin class switching to the antiviral IgG2a/c subclass by secreting IFN-γ [4]. Thus, understanding the molecular mechanisms governing CD4 T cell responses through either TCR-dependent or cytokine-dependent signals can yield insight into host protection against viral infection or protective antibodies induced by vaccines.
Following ligation of the TCR by cognate pMHC complexes, re-organization of membrane proteins and post-translational modification of signaling components ultimately result in activation of genes necessary for their proliferation and effector differentiation [5]. At the TCR-juxta-membrane regions, the TCR itself and its proximal kinases are densely packed and surrounded by adhesion molecules, such as LFA1 integrin, which together comprise the immunological synapse (IS) and facilitate quality control of TCR signals [6]. A number of previous studies have demonstrated that repeated cycles of this brief activation of TCR fine-tune the quality and magnitude of the activation program [7]. However, it remains unknown whether such fine-tuning mechanisms are common across different effector subsets, or whether the altered regulation may have impact on CD4 T cell immune responses in vivo.
The closely related proteins, CD2AP and CIN85, are scaffolding molecules that possess three tandem SH3 domains, a proline rich domain, and a C-terminal coiled-coil domain that mediates their hetero/homodimerization [8]. CD2AP was originally identified via a yeast two-hybrid screen as an interacting partner of the adhesion molecule CD2 and a dominant negative form expressed in Jurkat T cells prevented formation of the immunological synapse [9]. However, in naive T cells from AND TCR transgenic mice lacking CD2AP, formation of the immunological synapse was intact, suggesting that CD2AP functions independently of CD2 in primary T cells [10]. Interestingly, following stimulation with pMHC, activation of proximal kinases and degradation of TCR are delayed, leading to prolonged TCR activation and enhanced cytokine production in vitro [10]. However, the role of CD2AP in tuning of TCR signaling in T cell immune responses or the impact of altered TCR signaling quality in vivo caused by Cd2ap deficiency has not been defined due to a fatal kidney disease that develops around 3 weeks of age in Cd2ap−/− mice [11].
Here, we generated T cell-specific Cd2ap−/− deficient mice and show that CD2AP deficiency enhances control of chronic viral infection by augmenting antiviral TFH and GC responses. In contrast, TFH and GC responses were minimally altered when the mice were immunized with sheep red blood cells (SRBCs), which are prone to elicit type 2 responses [12]. TCR signaling in Cd2ap−/− CD4 T cells was prolonged specifically under TH1 conditions in vitro, resulting in increased production of IFN-γ, while CD2AP was dispensable for temporal tuning of TCR signals in TH2 or TH17 cells. These results demonstrate that CD2AP-dependent tuning of TCR signaling in CD4 T cells is TH subset-specific and that CD2AP may be an effective target to accelerate the development of protective antibodies in antiviral immune responses.
To conduct analyses of function of CD2AP in T cells in vivo, we generated the Cd2ap-flox allele, in which expression of cre recombinase results in deletion of exon 2 (S1A Fig). This allele was bred to Cd4-cre to inactivate Cd2ap in T cells. Numbers of mature CD4 and CD8 single positive thymocytes and CD4 and CD8 T cells in the spleen were comparable between Cd4-cre+ Cd2apF/F and control Cd2apF/F mice (S1B and S1C Fig).
To define the role of CD2AP in T cells in vivo, we infected Cd4-cre+ Cd2apF/F and control Cd2apF/F littermate control mice with the Armstrong strain of lymphocytic choriomeningitis virus (LCMV), which elicits acute antiviral responses and is cleared by CD8 T cells. At the peak of the T cell response on day 8 after infection, no difference was seen in numbers of total CD8 T cells, KLRG1+ short-lived effector CD8 T cells or that of LCMV glycoprotein (gp)-specific CD8 T cells between Cd4-cre+ Cd2apF/F and Cd2apF/F mice (Fig 1A, 1B and 1D), suggesting that CD2AP is not required for normal CD8 T cell responses to acute LCMV infection. Comparable numbers of LCMV-gp-specific CD4 T cells also were detected in both genotypes eight days after infection (Fig 1C and 1D), suggesting that CD2AP is dispensable for priming and initial expansion of LCMV-specific CD4 T cells. Both Cd2ap−/− and control LCMV-specific CD4 T cells differentiated into Ly6C+ TH1 effector cells or TFH cells as defined by CXCR5 and PD-1 expression on day 8 after infection (Fig 1E and 1F). Accordingly, we did not observe a difference in either frequency or absolute number of Fas+ GL7+ GC B cells at this time point (Fig 1E and 1F). At a later time-point (day 22), however, the frequency and absolute number of total and LCMV-specific CXCR5+ PD-1+ TFH cells as well as GL7+ Fas+ GC B cells was significantly increased in Cd4-cre+ Cd2apF/F mice (Fig 1G and 1H). These results indicate that Cd2ap-deficiency enhances CD4 T cell differentiation into the TFH subset following initial bifurcation between TH1 effector and TFH fates early in infection, and augments GC B cell responses.
To determine whether TFH differentiation of Cd2ap−/− CD4 T cells is also enhanced by immunization that induces other types of CD4 T cell responses, we analyzed the GC response after immunization of Cd4-cre+ Cd2apF/F and Cd2apF/F mice with SRBCs, which predominantly induces development of IL-4 producing CD4 T cells [13, 14]. While in the absence of CD2AP, a statistically insignificant trend in increased TFH and GC responses was seen at day 12 following immunization, a relatively late time point when antigen levels are declining [15] (Fig 2A–2H), numbers of TFH and GC B cells were comparable between Cd4-cre+ Cd2apF/F and Cd2apF/F mice both at a relatively early time point, day 6, and a later time point day 22 (S2A–S2D Fig). In addition, we found no significant differences in TFH and GC B cell numbers between Cd4-cre+ Cd2apF/F and Cd2apF/F mice following immunization with NP-CGG precipitated in aluminum salts (S2E and S2F Fig), which also induces a preferential TH2 biased CD4 T cell response [16, 17]. These results suggested that CD2AP suppresses CD4 T cell differentiation towards the TFH subset dependent on the context in which the response is induced.
While we observed enhanced TFH differentiation in Cd2ap-deficient mice to LCMV-Armstrong, CD8 CTLs rather than LCMV-specific antibodies are required for control of acute LCMV infection [18]. To determine whether enhanced TFH differentiation seen in CD2AP deficiency had an impact on antiviral antibody-mediated immunity, we utilized the LCMV-clone13 (LCMV-c13) model of chronic viral infection since TFH-dependent high affinity antibody responses in the GCs and class-switch recombination to IgG2a/2c subclasses are required for control of the infection [19–21]. After infection, LCMV replicated and established viremia to a similar extent in Cd4-cre+ Cd2apF/F and control cre−Cd2apF/F mice on day 8 (Fig 3A). In control cre−Cd2apF/F mice, viral abundance, as determined by the quantities of the LCMV gp transcript in the plasma of mice, started to gradually decline around day 30 (Fig 3A), coinciding with expansion of TFH and GC B cells in response to a surge of IL-6 production by follicular dendritic cells [19]. The decline in LCMV abundance was significantly accelerated in Cd4-cre+ Cd2apF/F mice, with the viral titers below the limit of detection in some Cd4-cre+ Cd2apF/F mice by day 45 (Fig 3A and S3A Fig). Consistent with the accelerated clearance of LCMV, the frequencies and numbers of total and LCMV gp-specific TFH cells in the spleen 22 days after infection were significantly increased (Fig 3B and 3D; S3B and S3C Fig), although total numbers of LCMV-specific CD4 T cells were not changed in Cd4-cre+ Cd2apF/F mice compared to control mice (Fig 3D). Accordingly, frequencies and numbers of GC B cells were significantly increased in Cd4-cre+ Cd2apF/F mice compared to control mice (Fig 3C and 3E; S3D Fig). It has been demonstrated that T follicular regulatory (TFR) cells can suppress the GC response [22]; thus we analyzed the frequency of Foxp3+ cells within the TFH compartment, however, we did not find differences between genotypes indicating TFR differentiation was unimpaired (S3E and S3F Fig). Moreover, plasma from Cd4-cre+ Cd2apF/F mice exhibited elevated neutralizing activity against LCMV in vitro, while titers of total anti-LCMV antibodies with dominant IgG2c and sub-dominant IgG1 subclasses were comparable between Cd4-cre+ Cd2apF/F and control Cd2apF/F mice (Fig 3F; S3G Fig). As a measure of GC output we also analyzed the number of memory B cells in the spleen and plasma cells in the bone marrow at 60 days following LCMV-c13 infection and observed no difference in overall numbers between Cd4-cre+ Cd2apF/F and control Cd2apF/F, suggesting enhanced acquisition of mutations rather than total output contributed to the phenotype (S4 Fig). Given the importance of CD4 T cell help in promoting long-lived polyfunctional CD8 T cells we also analyzed the CD8 T cell response at day 22 [23, 24]. Importantly, we did not detect differences in overall CD8 T cell numbers or LCMV-specific H-2Db(gp33-41) tetramer binding T cells in Cd4-cre+ Cd2apF/F mice compared to control mice (S5A and S5B Fig). In addition, we did not observe differences in the polyfunctionality of LCMV-specific CD8 T cells as determined by IFN-γ and TNF-α production (S5C Fig). These results indicate that while inactivation of CD2AP only minimally affects the CD8 T cell response, it enhances TFH differentiation and generation of neutralizing antiviral antibodies by augmenting B cell help.
To determine whether the increased TFH differentiation in Cd4-cre+ Cd2apF/F mice was cell-intrinsic, we generated BM chimeras by transferring a mixture of BM cells from Cd4-cre+ Cd2apF/F or cre−Cd2apF/F mice (CD45.2) and CD45.1 WT mice into lethally irradiated CD45.1 WT mice. Following reconstitution of donor-derived hematopoiesis, the recipient mice were infected with LCMV-c13 and contribution of Cd2ap-deficient and -sufficient cells to the TFH compartment was examined 22 days after infection (Fig 4A). While the percentages of CD45.2 cells in the B cell and CD44lo PD-1− naive CD4 T cell compartments were equivalent between chimeras reconstituted with Cd4-cre+ Cd2apF/F or cre−Cd2apF/F BM cells, the contribution of CD45.2 cells to the CXCR5+ PD-1+ TFH pool was significantly greater in chimeras receiving Cd4-cre+ Cd2apF/F compared to control chimeras receiving cre−Cd2apF/F BM cells (Fig 4B). In addition to the quantitative increase in TFH numbers in Cd4-cre+ Cd2apF/F mice, Cd2ap-deficient TFH cells expressed elevated levels of genes that are associated with TFH functions, including Il21, Il21r and Il17ra, and also those associating with active cell cycling (Fig 4C) [25]. We confirmed these results by analysis of frequencies of Ki67+ cells within TFH cells and found them to be increased in KO cells compared to WT (S3H Fig). In contrast to these changes, we did not observe changes in expression of TFH associated molecules such as ICOS and OX-40, suggesting these costimulatory pathways might not contribute to the observed phenotype (S3I Fig). These results suggest that Cd2ap-deficient CD4 T cells are maintaining highly activated states, potentially through enhanced TCR-dependent signaling, and establish a stable TFH program potentially by the IL-21-IL-21R feed-forward loop and thus better support GC B cell responses against viral infection.
Our data demonstrated that TFH differentiation was enhanced by Cd2ap-deficiency specifically in viral infection, but not in immunization with SRBCs or NP-CGG in alum. These results suggest that CD2AP regulates TCR signaling in a TH-subset-specific manner. To dissect the subset-specific role of CD2AP in CD4 T cells, we activated naive CD4 T cells from Cd4-cre+ Cd2apF/F and control Cd2apF/F cells under polarizing conditions that promote generation of TH1 or TH2 cells and compared their responses to antigen receptor stimulation (Fig 5A). While production of IL-4 under TH2 conditions was comparable between Cd2ap-deficient and -sufficient CD4 T cells, the production of IFN-γ was increased by 3-fold in Cd2ap-deficient TH1 cells compared to control cells (Fig 5B). Consistently, activation of the MAPK pathway was prolonged specifically in Cd2ap-deficient TH1 compared to control TH1 cells, whereas we did not observe the difference in TH2 cells (Fig 5C). Previous studies of CD4 T cells using germline Cd2ap-deficient mice showed that Cd2ap−/− TCR-transgenic naive T cells are defective in TCR downregulation [10]. In polarized TH1 cells, downmodulation of surface TCRs was significantly impaired in the absence of CD2AP, while the difference was not observed under TH0, TH2 or TH17 conditions between Cd2ap-deficient and -sufficient CD4 T cells (Fig 5D). Furthermore, when Cd2ap-deficient and control TH1 cells were re-stimulated with plate-bound anti-CD3 and anti-CD28 antibodies or with PMA and Ionomycin for 4 hours, similar frequencies of cells expressed IFN-γ and TNF-α (Fig 5E and 5F). However, when they were stimulated with plate-bound anti-CD3 and anti-CD28 antibodies for 24 hours, the frequencies of cells that expressed IFN-γ or TNF-α during the last 2 hours of the stimulation were significantly increased in Cd2ap-deficient TH1 cells relative to WT TH1 cells presumably due to prolonged retention of TCR on the cell surface, which was crosslinked by the plate-bound antibodies (Fig 5E and 5F). These results indicate that CD2AP deficiency prolongs TCR signaling specifically in CD4 T cells under TH1 conditions.
CIN85 and CD2AP are closely related and may have redundant functions [8,12]. To test this, we also generated mice in which T cells are deficient for CIN85 by breeding a Cin85-flox allele to both Cd4-cre [26]. Cd4-cre+ Cin85F/F mice exhibited no differences in CD4 T and CD8 T cell response to acute LCMV infection (S6 Fig). However, in contrast to Cd4-cre−Cd2apF/F mice, Cd4-cre+ Cin85F/F mice were unable to control chronic LCMV infection with significantly higher viral burden in the peripheral blood at days 45 and 60 after infection compared to cre− littermate controls (Fig 6A). In addition, while all WT mice had cleared virus from the plasma by day 80, approximately half of Cd4-cre+ Cin85F/F still exhibited high viral load (Fig 6B). The clearance phenotype was more pronounced when we infected Vav1-icre+ Cin85F/F mice, in which all hematopoietic cells, including B cells, were deficient for Cin85 (S7A Fig), suggesting CIN85 plays additional roles in other hematopoietic cells in the context of LCMV-c13 infection, potentially through B cells as previously reported [26]. However, when we analyzed mice at day 30 following infection we did not find any significant differences in either frequency or absolute number of CD8 T cell, TFH, or GC response (S7B–S7D Fig). Consistently, when we analyzed Cd4-cre+ Cin85F/F mice, the frequency of Fas+ GL7+ GC B cells was only marginally, but insignificantly, decreased in Cd4-cre+ Cin85F/F mice compared to control Cin85F/F mice (Fig 6C) with anti-LCMV IgG titers unchanged, suggesting no severe defects in the GC response (Fig 6D). However, plasma from infected Cd4-cre+ Cin85F/F mice exhibited a significantly reduced neutralizing activity compared to plasma from cre− littermate controls (Fig 6E). These results indicate that CIN85 is required for protective antibody responses in a T cell-dependent manner presumably through controlling TFH function and potentially the quality of help provided to GC B cells with respect to generation of high-affinity neutralizing antibodies required for clearance of chronic LCMV [27]. These results suggest that although CD2AP and CIN85 share significant homology, their functions in antiviral CD4 T cell responses are non-redundant, and potentially antagonistic.
This study demonstrated that deletion of CD2AP in T cells results in skewing of CD4 T cell differentiation towards TFH cells in response to viral infection, leading to enhanced control of LCMV that requires GC-derived high affinity antibody responses [19, 21, 28]. TFH differentiation was correlated with sustained TCR signaling under TH1 conditions, while TCR signaling in vitro under non-TH1 conditions was not altered. Thus, our work revealed a specific role of CD2AP in subset-specific CD4 T cell responses.
Sustained TCR stimulation during chronic LCMV infection or in the cancer microenvironment causes deregulation of CD8 T cells, a phenomenon known as exhaustion [1], [29]. Frequent interactions with cognate pMHC-I result in the persistent upregulation of several inhibitory receptors which act to dampen T cell proliferation and effector functions, a hallmark of the “exhausted state” [2, 29]. However, the impact of sustained TCR stimulation on the function of CD4 T cells has been less clearly understood. In chronic LCMV infection, CD4 T cells exhibit less IL-2 production and increased IL-10 production, a phenomenon that is similar in nature to CD8 T cell “exhaustion” [3,30–32]. However, these CD4 T cells with the altered activation state acquire the capability of producing IL-21, a key cytokine that enhances the GC response and also supports the CD8 T cell response; both are required for control of the viral infection [4, 31–33]. Thus, although sustained TCR signaling compromises CD8 T cell functions, CD4 T cells are able to tolerate sustained signaling through TCR to mediate pathogen control.
Several recent studies indicate that during chronic LCMV infection, CD4 T cells exhibit a relatively unique propensity to acquire TFH features, a process that is dependent on continuous antigen stimulation [5, 34]. The acquisition of TFH phenotype in chronic infection appears to be different compared to acute LCMV infection [35]. Interestingly, in late phases >day 20 of LCMV-c13 infection B cells do not appear to be absolutely required for the development of CXCR5+ cells, suggesting other types of antigen presenting cells could contribute to the sustained TFH response as this does not occur in MHCII KO [34]. Preferential TFH accumulation has also been shown to be dependent on type-I Interferon [6, 36] which has not been explicitly observed in acute contexts, via a cell-extrinsic mechanism, suggesting other soluble or cell-associated factors could have a more direct influence [36]. These results illustrate the complexity in direction of the CD4 T cell response, and illustrate the variety of mechanisms that influence the humoral response in response to the nature of the insulting pathogen. However, in several contexts it appears that modulation of TCR affinity and or signal duration impacts the differentiation of TFH cells, supporting a more broadly applicable role for the TCR.
Our study with CD2AP-deficient T cells confirm and extend previous studies that suggest that sustained TCR signaling promotes TFH differentiation [37, 38]. Specifically, it was shown following infection by Listeria monocytogenes that TCR:pMHCII dwell time correlated well with GC-TFH differentiation [37]. In addition, in antigen-specific cells elicited by immunization with pigeon cytochrome c, TFH phenotype cells tended to have higher affinity for tetramer compared to non-TFH cells, and transfer of TCR transgenic cells with higher affinity for pMHCII revealed a preference for TFH differentiation [38]. One potential explanation is that TFH and TH1 cells may share a transitional stage, at which sustained TCR signaling may restrict T-bet upregulation to direct the activated CD4 T cells into the TFH subset, which favors low-T-bet [39], while TFH differentiation during a type-2 immune response may require a distinct mechanism. This may explain why TFH differentiation is enhanced when CD4 T cells receive sustained TCR stimulation as antigen levels decline. Interestingly, enhanced TFH differentiation and increased IL-21 expression by Cd2ap-deficient CD4 T cells did not noticeably alter the CD8 T cell response, which is also dependent on CD4 T cell help. These results potentially indicate that the nature of CD4 T cell help to B cells and CD8 T cells may be independent.
While downregulation of the TCR and subsequent termination of TCR downstream signaling is broadly shared by T cells, our results indicate that CD2AP-dependent regulation of TCR signaling duration is TH1-specific in vitro, which correlates with enhanced TFH differentiation in LCMV infection. However, it still remains unclear whether the in vivo functions of CD2AP are completely explained by TCR signaling via the immunological synapse, interactions with CD2, or potentially via signaling via other cytokines such as IL-21 or Type I interferon and is an area of further investigation. Notably, a recent study reported TH2-specific regulation of the duration of TCR signaling by the Rab-GTPase Dennd1b [40]. Dennd1b-deficiency in CD4 T cells results in sustained TCR signaling and increased cytokine production specifically under TH2 conditions. Furthermore, a DENND1B SNP in human, causing reduced expression of its protein, is associated with asthma. These two studies, contrasting lineage-specific requirement for CD2AP and Dennd1b in controlling TCR signaling in TH1 and TH2 cells, respectively, imply that the control of TCR signaling may be uniquely regulated by different CD4 effector subsets. Despite strong correlations between sustained duration of TCR signaling in vitro and enhanced T cell polarization in vivo in Cd2ap-deficient mice and Debbd1b-deficient mice, however, there may be additional mechanisms, such as cell-cell contact or signaling through CD2, by which T cell polarization are enhanced in these genetically modified mice.
Our study highlights CD2AP as a potential target for immunotherapy to enhance B cell immune responses. Given the importance of high-affinity neutralizing antibodies in the clearance of both acute and chronic viral infections or vaccines, boosting the GC response is an attractive therapeutic target. The development of small molecules that inhibit the function of CD2AP in the context of viral infection or in the context of vaccines, could enhance the antibody response. TFH responses, however, have also been associated with the development of autoantibodies and autoimmunity.
Collectively, our results demonstrate that the quality of TCR signaling is regulated by distinct mechanisms in CD4 effector T cells, and regulates differentiation of antigen-specific CD4 T cells towards distinct subsets.
Overall care and use of the animals was consistent with The guide for the Care and Use of Laboratory Animals from the National Research Council and the USDA Animal Care Resource Guide and were performed according to a protocol approved by Washington University’s Animal Studies Committee under a protocol number of 20150187, which was approved on 10/26/2015 and expires 10/9/2018. Euthanasia procedures are consistent with the “AVMA guidelines for the Euthanasia of Animals 2013 edition.”
C57BL/6N and B6-CD45.1 mice were purchased from Charles River Laboratory. Cd4-cre mice in the C57BL/6 background were purchased from the Jackson Laboratory. A Cin85 (Sh3kbp1)-floxed allele was described previously [26]. To generate a Cd2ap-flox allele, a genomic fragment containing Cd2ap exon2 was flanked with two loxP sites and subcloned into a vector containing a loxP-flanked neomycin resistant gene, a diphtheria toxin A (DTA) gene, and a 6.0-kb 5’ and a 2.1-kb 3’ homology arm. The targeting vector was electroporated into C57BL/6-derived Bruce4 embryonic stem (ES) cells. After the selection with G418 (Thermofisher), correctly targeted clones were identified by PCR. The neomycin resistant gene was removed by a transient transfection of a cre expression vector. The targeted ES clones were injected into blastocysts of BALB/c mice. The chimeric mice were crossed with C57BL/6 mice to obtain germline transmission. To generate BM chimeras, B6-CD45.1 mice were lethally irradiated (10.5 Gy) and reconstituted with donor BM cells for at least 8 weeks before experiments. All mice were housed in a specific pathogen-free facility at Washington University in St. Louis, and were analyzed at 8 to 10 weeks of age, unless stated otherwise. Both sexes were included without randomization or blinding.
Mice were infected with 2 x 105 plaque-forming units (PFU) of LCMV-Armstrong strain via the intraperitoneal route or 2 x 106 (PFU) of LCMV-c13 by intravenous injection. For the quantification of plasma viral load, RNA was extracted from 10 μL of plasma using Trizol (Life Technologies). Before RNA extraction a spike-in of RNA extracted from 293T cells (American Type Culture Collection, ATCC) expressing gfp mRNA was added to the plasma samples. The amounts of the LCMV GP transcript relative to that of ‘‘spiked- in” gfp RNA were determined by real-time qRT-PCR as previously described [21, 41]. For plasma viral titers, plasma was frozen at -80°C before performing plaque assay on Vero cells as previously described. Briefly, 7.5 x 105 Vero cells (ATCC) were plated per well of a 6-well plate 24 hour prior to incubation with serial 10-fold dilutions of plasma in 200 μL of MEM/1%FBS (Thermofisher) for 1 hour at 37°C. Following the incubation Vero cells were overlaid with 0.5% Agarose (Thermofisher) solution in complete MEM and incubated for 5–6 days. Plaques were visualized following fixation in 1% PFA (Electron Microscopy Sciences) and staining with 0.1% crystal violet (Sigma).
In vitro neutralization assay was performed according to a published protocol with some modifications [27]. Heat-inactivated (1 hr at 55°C) plasma were serially diluted with media and incubated with an equal volume of viral supernatant containing ~30–40 focus forming units (FFU) of LCMV clone 13 at 37°C for 90 min. 4 x 104 MC57G cells (ATCC) were added to the virus mixture and once the cells had adhered to the plate ~4 hours, were overlaid with a 1% methylcellulose solution in completed DMEM. 40 hours following infection, overlay was removed, and cells were fixed with 4% PFA for 1 hour at RT. Cells were permeabilized with 0.5% Triton X-100 in PBS and blocked with 10% FBS in PBS. Cells were stained with 1ug/mL of anti-LCMV NP (VL-4; BioXCell) for 1 hour at RT followed by anti-rat IgG HRP. Foci were developed using KPL TrueBlue Substrate.
Mice were immunized with 8–10 x 108 SRBC (Lampire) in 200 μL or with 100μg of NP-CGG (Biosearch Technologies) precipitated Potassium Aluminum Sulfate (Sigma), via the intraperitoneal route.
Single-cell suspensions were prepared by manual disruption of spleens with frosted glass slides. Absolute cell counts were determined using Vi-Cell (Becton-Dickson). The following monoclonal antibodies were used: anti-CD4 (GK1.5; Biolegend), anti-CD8a (53–6.7; Biolegend), anti-CD25 (PC61; Biolegend), anti-CD44 (IM7; Biolegend), anti-CD45.2 (104; Biolegend), anti-CD45R (B220) (RA3-6B2; Biolegend), anti-CD62L (MEL-14; Biolegend), anti-CD95 (Fas) (Jo2; BD Biosciences), anti-KLRG1 (2F1/KLRG1; Biolegend), anti-IFN-γ (XMG1.2; eBioscience), anti-TCRb (H57-597; eBioscience), anti-PD-1 (29F.1A12; Biolgend), anti-CXCR5 (2G8; BD Biosciences) or anti-CXCR5 (L138D7; Biolegend), anti-Ly6C (HK1.4; Biolgend), and anti-GL7 antigen (GL7; Biolegend).
For sorting of naive CD8 T cells, splenocyte samples were initially depleted of B220+ cells through the use of magnetic beads (Thermofisher). CD4 T cells were negatively selected using the Dynabeads FlowComp Mouse CD4 Kit (Thermofisher) and then were stained with monoclonal antibodies to purify CD62L+ CD44− CD25− cells on a FACSAria II or FACSAriaIII (BD Biosciences). For phenotypic analysis, cells were stained with monoclonal antibodies, phycoerythrin (PE)-conjugated H-2Db-gp(33–41) (MBL International) for 1 hour at RT or PE-I-Ab(gp66-77) (obtained from NIH Tetramer Core) for 90 minutes at 37°C. Dead cells were excluded by staining with DAPI (4,6-diamidino-2-phenylindole; Sigma). Data were analyzed with FlowJo software (TreeStar).
Sort purified naive (CD62L+ CD44lo CD25−) CD4 or CD8 T cells were cultured in RPMI 1640 medium supplemented with 10% FBS (Thermofisher) and Gluta-MAX in the presence of plate-bound anti-CD3 (145-2C11; Biolegend) and soluble anti-CD28 (37.51; Bio X Cell) at concentrations of 0.1 μg/ml and 0.5 μg/ml, respectively, unless specified otherwise, in multiwell tissue culture plates coated with rabbit antibody to hamster IgG (0855395; MP Biomedicals). For TH1 polarization 10 ng/mL of IL-12 (R&D Systems) and 10 μg/mL of anti-IL-4 (11B11) were added to culture media. For TH2 polarization 10 ng/mL IL-4 (eBioscience), 10 μg/mL anti-IL-12 (BioXCell), and 10 μg/mL anti-IFN-γ (Biolegend) were added to culture media. Following 3 days of anti-CD3 and anti-CD28 stimulation cells were removed from stimulation and rested for 3 days in conditioned medium at concentration of 0.5–1 x 106 cells/mL. For CD8 T cell cultures, cells were cultured in 40 U/mL IL-2 during the resting period.
Day 6 CD4 T cells were harvested, washed in fresh media, and restimulated using 50 ng/mL PMA (Sigma) and 1 μM Ionomycin (Sigma) or plate-bound anti-CD3 (Biolegend) and soluble anti-CD28 (BioXcell) as described above for the indicated times. For intracellular cytokine analysis, Brefeldin A (Biolegend) was added to the cultures 2 hours before harvest. Supernatant was collected 24 hours post-stimulation.
ELISA for cytokines was performed on Nunc Maxisorp plates using ELISA MAX IFN-γ and IL-4 kits (Biolegend) and developed using TMB substrate (Dako). OD450 values were read on a spectrophotometer. Standard curves for these ELISAs were generated with purified cytokines.
Anti-LCMV ELISA was performed as previously described [42]. Nunc Polysorp plates were coated with 10ug/mL sonicated cell lysate from LCMV-infected BHK-21 cells (gift from Marco Colonna, Washington University) as capture antigen or uninfected BHK-21 cell lysate overnight followed by UV irradiation (300 mJ in Stratalinker 1800; Stratagene). Plasma antibody was detected with biotinylated anti-mouse IgG1 antibody (BD) and anti-mouse IgG2c antibody (1077–08; Southern Biotech). Endpoint titers were calculated by a sigmoidal-dose response curve using a Graphpad Prism 7 software.
Cells were subject to LIVE/DEAD Aqua staining (Thermofisher) for 30 minutes at 4°C before being fixed with 4% PFA for 10 minutes at RT. Cells were washed twice with 0.03% saponin (Sigma) in 2% FBS/PBS before being stained with the indicated antibodies in 0.3% Saponin in 2% FBS/PBS for 20 min at 4°C.
In vitro polarized T cells were rested for 1 hour in cytokine-free RPMI medium before being stained with 5 μg/mL anti-CD3 and 5 μg/mL anti-CD28 on ice for 20 minutes. TCR was then crosslinked by addition of 20 μg/mL of anti-Armenian Hamster IgG to the cells on ice. Cells were stimulated in 5 ml polystyrene tubes for the indicated times in a water bath at 37°C. Cells were lysed in NP-40 Lysis Buffer (1% NP-40, 50 mM Tris-HCl (pH7.5), 150 mM NaCl, 2 mM EDTA, phosphatase inhibitors) and cleared via centrifugation at 14,000 rpm for 10 minutes, then denatured in SDS sample buffer.
Lysates from equal numbers of cells were separated by 8% or 10% SDS PAGE and transferred to nitrocellulose membranes (Bio-Rad), which were incubated with primary antibodies (identified below), followed by detection with horseradish peroxidase–conjugated species-specific antibody to immunoglobulin light chain (115-035-174 or 211-032-171; Jackson ImmunoResearch) and a Luminata HRP substrate (Millipore). The following antibodies were used: anti-CD2AP [10], anti-ERK2 (Santa Cruz, C-14), anti-p-MEK1/2 (Cell Signaling Technology, 9121).
Total RNA was extracted from ~300,000 sorted cells using manufacturer’s instructions using the RNA XS Kit (Macherey Nagel). cDNA synthesis and amplification were performed with a Pico Input RNA Kit, according to the manufacturer’s instructions (Clontech). Libraries were sequenced on a HiSeq3000 (Illumina) in single-read mode, with a read length of 50 nucleotides producing ~25 million reads per sample. Sequence tags were mapped onto the NCBI37 mm9 with TopHat [43], followed by transcript assembly and Reads Per Kilo base of exon per Million reads (RPKM) estimation using Cufflinks [44–46] on the Galaxy platform (https://usegalaxy.org/). The raw data has been deposited at NCBI GEO with an accession number GSE112778.
Statistical analyses were performed with a two-tailed unpaired Student’s t-test or a Mann-Whitney test using GraphPad Prism software.
|
10.1371/journal.pgen.1002147 | Extracellular Matrix Dynamics in Hepatocarcinogenesis: a Comparative Proteomics Study of PDGFC Transgenic and Pten Null Mouse Models | We are reporting qualitative and quantitative changes of the extracellular matrix (ECM) and associated receptor proteomes, occurring during the transition from liver fibrosis and steatohepatitis to hepatocellular carcinoma (HCC). We compared two mouse models relevant to human HCC: PDGFC transgenic (Tg) and Pten null mice, models of disease progression from fibrosis and steatohepatitis to HCC. Using mass spectrometry, we identified in the liver of both models proteins for 26 collagen-encoding genes, providing the first evidence of expression at the protein level for 16 collagens. We also identified post-transcriptional protein variants for six collagens and lysine hydroxylation modifications for 14 collagens. Tumor-associated collagen proteomes were similar in both models with increased expression of collagens type IV, VI, VII, X, XIV, XV, XVI, and XVIII. Splice variants for Col4a2, Col6a2, Col6a3 were co-upregulated while only the short form of Col18a1 increased in the tumors. We also identified tumor specific increases of nidogen 1, decorin, perlecan, and of six laminin subunits. The changes in these non-collagenous ECM proteins were similar in both models with the exception of laminin β3, detected specifically in the Pten null tumors. Pdgfa and Pdgfc mRNA expression was increased in the Pten null liver, a possible mechanism for the similarity in ECM composition observed in the tumors of both models. In contrast and besides the strong up-regulation of integrin α5 protein observed in the liver tumors of both models, the expression of the six other integrins identified was specific to each model, with integrins α2b, α3, α6, and β1 up-regulated in Pten null tumors and integrins α8 and β5 up-regulated in the PDGFC Tg tumors. In conclusion, HCC–associated ECM proteins and ECM–integrin networks, common or specific to HCC subtypes, were identified, providing a unique foundation to using ECM composition for HCC classification, diagnosis, prevention, or treatment.
| The microenvironment can have a profound influence on cellular behavior and survival and on growth of developing tumor cells. We present the first comprehensive analysis of the extracellular matrix (ECM) and associated receptor proteomes, applied here to the study of hepatocellular carcinoma (HCC). This study demonstrates the utility of mass spectrometry-based approaches to characterize, at the protein level, gene families with extensive sequence homology, post-transcriptional regulations, and post-translational regulations. This is also the first study to analyze and compare liver proteome changes occurring during the transition from fibrosis and steatohepatitis, common preneoplastic conditions in humans, to HCC, using two mouse models. This approach identifies ECM and integrin components, which could play an important role in the early steps of hepatocarcinogenesis, and provides a path to identifying ECM–tumor cell networks that may contribute to the heterogeneous features of HCC.
| Cirrhosis, the result of end-stage fibrosis, and steatohepatitis are common pre-neoplastic conditions associated with hepatocarcinogenesis [1]. It is therefore important to understand the mechanisms leading to the transition from fibrosis and steatosis to HCC. Mice with liver-specific transgenic (Tg) expression of platelet-derived growth factor-C (PDGFC) represent a relevant model for such a study as members of the PDGF family play major roles in angiogenesis and fibrosis [2], [3]. Moreover, these mice develop liver fibrosis resembling human alcoholic and nonalcoholic fatty liver disease, which precedes development of HCC [4], [5]. Another relevant model is mice with liver specific deletion of the phosphoinositide 3-kinase (PI3K)/phosphatase and tensin homolog (Pten). PTEN loss of function in hepatocytes leads to steatohepatitis, fibrosis and HCC later in life [6], [7]. While the liver tumors in PDGFC Tg mice show characteristics of HCC, the tumors in the Pten null model present a mixed phenotype of HCC and cholangiocarcinoma [8], [9]. Up to 40% of human HCCs potentially arise from progenitor-like tumor initiating cells and tend to have a more aggressive phenotype [10]. In addition, the presence of intermediate cells co-expressing both hepatocyte and biliary markers is associated with HCC occurrence [11] and acquisition of cholangiocarcinoma-like expression traits plays a critical role in the heterogeneous progression of HCC [12]. It is therefore of particular relevance to compare liver proteome changes in both the PDGFC Tg and the Pten null models.
Through mass-spectrometry-based profiling of the liver tissues collected at different disease stages in these two mouse models, we have characterized changes in the liver proteome occurring in fibrotic and steatotic tissue, as well as in tumors. We previously reported that the extensive mass-spectrometry-based approach we used in this study reaches depth and allows for quantitative estimates of protein abundance [13]. Changes in specific protein families or networks can be characterized as shown here for proteins of the extracellular matrix (ECM) and their receptors. The ECM is a key component of the microenvironment that is in immediate contact with the tumor cells and is a critical source for growth, survival, motility and angiogenic factors that significantly affect tumor biology and progression. In addition, cell adhesion to the ECM through integrins and other cell surface receptors triggers intracellular signaling pathways that can regulate cell cycle progression, migration and differentiation. While hepatic ECM has been extensively studied in the context of liver fibrosis, little attention has been given to the role of the ECM in the early steps of hepatocarcinogenesis. Therefore, delineating and comparing the liver proteome changes of ECM components in two mouse models of liver cancer represents a unique and important contribution to our understanding of the molecular mechanisms of early hepatocarcinogenesis, and to ongoing efforts to identify novel diagnostic and therapeutic targets.
The liver specific PDGFC Tg and Pten null mouse models reproduce the steps of HCC development observed in humans progressing from steatohepatitis and fibrosis to hepatocyte dysplasia and tumorigenesis. These events are associated with significant modification of the stroma and associated extracellular matrix. Preceding tumor development, there is accumulation of collagen fibers in the liver of these mice (Figure 1A). Steatosis is also particularly pronounced in the Pten null liver as shown by the accumulation of lipid droplets (Figure 1A). Interestingly, the phenotypes of the tumors are different in these two models with characteristics of HCC in the PDGFC Tg model and with mixed cell characteristics of HCC and cholangiocarcinoma in the Pten null mice (Figure 1B). Extensive mass spectrometry analysis following a multi-dimensional protein separation strategy composed of two-dimensional HPLC followed by SDS-PAGE was applied as described [13] to livers collected from these two models at the fibrosis and steatosis stage as well as on small HCCs. For each sample group, livers from three or four mice were pooled. A total of 10,707 protein isoforms, products of 8,278 individual genes were identified with high confidence. For each identified protein, protein abundance was calculated using the frequency of tandem mass spectra assigned to that protein. We previously reported that this label-free approach provides a good estimate of protein abundance in liver [13].
Out of the 44 alpha chains of the murine collagen family, a total of 26, corresponding to 16 collagen types, were identified: COL1A1, COL1A2, COL2A1, COL3A1, COL4A1, COL4A2, COL4A3, COL4A4, COL4A5, COL4A6, COL5A1, COL5A2, COL5A3, COL6A1, COL6A2, COL6A3, COL7A1, COL8A1, COL10A1, COL14A1, COL15A1, COL16A1, COL18A1, COL22A1, COL27A1, COL28A1 (Table 1). All 26 collagens were identified with ProteinProphet scores of 0.97 or higher corresponding to a false discovery rate of 0.003 and all 26 collagens were identified in both mouse models. For 16 of these 26 collagens, this study represents to date the first evidence at the protein level (www.uniprot.org). The information on protein annotation, peptide numbers and sequences is summarized in Table S1. As expected, collagens types I and III were predominant in abundance. Among the remaining collagens, collagens types IV, V and VI were the most abundant. Significant changes in abundance during disease progression were observed for all 26 collagen proteins. In human liver, type I and III collagen levels increase up to eight times during fibrogenesis, with a significantly higher increase of type I collagen than of type III collagen. Similarly, type I and III collagen levels strongly increased in the fibrotic liver of 2-month-old PDGFC Tg mice with a significantly higher increase of COL1A1 and COL1A2 than of COL3A1 (Figure 2A). Eight collagens of lower abundance were also up-regulated in the PDGFC Tg fibrotic liver. These included: COL2A1, COL5A1, COL5A2, COL5A3, COL8A1, COL22A1, COL27A1 and COL28A1 (Figure 2B).
The protein abundance of the remaining 15 identified collagens increased in the tumors of 8-month old PDGFC Tg mice as shown in Figure 3A for the proteins of moderate to high abundance (COL4A1, COL4A2, COL6A1, COL6A2, COL6A3, COL14A1 and COL18A1) and in Figure 3B for the proteins of lower abundance (COL4A3, COL4A4, COL4A5, COL4A6, COL7A1, COL10A1, COL15A1 and COL16A1). The same abundance changes were observed for these 15 collagens in the tumors of 9-month old Pten null mice (Figure 3C, 3D). In both models, the tumor-associated abundance increase was particularly significant for collagens type IV, VI, XIV, XV and XVI. Validation at the transcript level was performed for Col4a2 and Col15a1. Col4a2 mRNA was strongly up-regulated in tumors of both models with 10.8-fold increase (p = 0.008) in PDGFC Tg mice (Figure 4A) and 4.8-fold increase (p = 0.002) in Pten null mice (Figure 4B). Col4a2 mRNA expression was also significantly higher in tumors compared to adjacent tissue of both models (p = 0.01 in PDGFC Tg mice and p = 0.001 in Pten null mice). Col15a1 mRNA was not detected in control liver tissues and was only weakly expressed in fibrotic and steatotic liver in both models. Its expression was significantly increased in tumors in both models with 6.8-fold increase (p = 0.009) in PDGFC Tg mice (Figure 4C) and 6.3-fold increase (p = 0.003) in Pten null mice (Figure 4D). Col15a1 mRNA expression was also significantly higher in tumors compared to adjacent tissue in both models (p = 0.04 in PDGFC Tg mice and p = 0.02 in Pten null mice).
Peptides specific to post-transcriptional variants were identified for Col1a1, Col6a2, Col6a3 and Col18a1 (Figure 5). These variants result from alternative splicing for Col1a1, Col6a2 and Col6a3 and from alternative promoter usage for Col18a1. Validation at the transcript level was performed for Col6a2 and Col18a1 using primers specific to the post-transcriptional variants. Col6a2 canonical mRNA was strongly up-regulated in tissue adjacent to tumors in both models with 23.4-fold increase (p = 0.01) in PDGFC Tg mice (Figure 6A) and 6.0-fold increase (p = 0.02) in Pten null mice (Figure 6B). A correlated up-regulation was observed for Col6a2 splice variant with 8.5-fold increase (p = 0.01) in PDGFC Tg mice (Figure 6C) and 3.4-fold increase (p = 0.02) in Pten null mice (Figure 6D). Col18a1 canonical mRNA also called NC1-764, was unchanged in liver tissue and tumors of PDGFC Tg mice (Figure 7A) and decreased in tumors of Pten null mice (5.5-fold, p = 0.002) (Figure 7B). In contrast, Col18a1 variant NC1-301 strongly increased in tumors in both models with 32.8-fold increase (p = 0.01) in PDGFC Tg mice (Figure 7C) and 118.7-fold increase (p = 0.0006) in Pten null mice (Figure 7D). Col18a1 variant NC1-301 also strongly increased in adjacent tissue in both models with 16.9-fold increase (p = 0.01) in PDGFC Tg mice (Figure 7C) and 50.7-fold increase (p = 0.001) in Pten null mice (Figure 7D).
Lysine hydroxylation is a well-known post-translational modification of type I, III and V collagens and contributes to matrix remodeling and stiffening. We investigated whether lysine hydroxylation occurs on other collagens and changes in abundance during tumor development, by researching the mass spectrometry raw data using criteria allowing for the identification of lysine hydroxylation modifications. Extensive lysine hydroxylation modification was observed as expected for COL1A1, COL1A2, COL3A1 and COL5A1 with 9, 12, 7 and 5 modified lysine residues identified, respectively. Other collagens with modified lysine residues included all six type IV collagens, COL6A2, COL16A1 and COL27A1 (Table 2). The lysine hydroxylation status was particularly high for COL3A1 with 94% of identified peptides presenting with lysine modifications in both PDGFC Tg fibrotic liver and Pten null steatotic liver; and with 100% of identified peptides presenting with lysine modifications in the tumors collected from both models (Figure 8). The lysyl hydroxylation status of COL1A1 and COL1A2 also slightly increased in the tumors compared to the fibrotic and steatotic livers in both models increasing from 33% to 43% for COL1A1 and from 21% to 25% for COL1A2 in PDGFC Tg mice and increasing from 27% to 37% for COL1A1 and from 15% to 17% for COL1A2 in Pten null mice (Figure 8). Inversely, the lysyl hydroxylation status of COL6A2 slightly decreased in the tumors compared to the fibrotic and steatotic livers in both models decreasing from 29% to 17% in PDGFC Tg mice and from 15% to 13% in Pten null mice (Figure 8). For the type IV collagens, the hydroxylation status was below 5%. Low hydroxylation was also found for COL5A1, COL5A2, COL16A1 and COL27A1.
The ECM is also composed of non-collagenous proteins such as laminins. Laminins are large macromolecules constituted by the association of one α, one β and one γ chain. Laminin α5, laminin β2 and laminin γ1 were up-regulated in the tumors of both mouse models (Figure 9), suggesting that laminin 521 (previously called laminin 11) is the most abundant laminin in HCC. Laminin β3 and laminin γ2 were specifically up-regulated in Pten null tumors while laminin β1 was specifically up-regulated in PDGFC Tg tumors (Figure 9). Perlecan, also known as HSPG2, decorin and nidogen 1 were up-regulated in tumors of both models. Validation at the transcript level was performed for laminin α5 and nidogen 1. Laminin α5 mRNA was only weakly expressed in fibrotic and steatotic liver in both models but was significantly up-regulated in tumors in both models (6.4-fold (p = 0.01) in PDGFC Tg mice and 10.5-fold (p = 0.0002) in Pten null mice) (Figure 10). Laminin α5 mRNA expression was also significantly higher in tumors compared to adjacent tissues, in both models (p = 0.02 in PDGFC Tg mice and p = 0.03 in Pten null mice) (Figure 10). Nidogen 1 mRNA was increased by 7.2-fold (p = 0.008) and by 8.9-fold (p = 0.0003) in tumors from PDGFC Tg and Pten null mice, respectively (Figure 11A, 11B). Similarly, nidogen 1 protein was increased by 6.1-fold (p = 0.01) and by 15.3-fold (p = 0.001) in tumors from PDGFC Tg and Pten null mice, respectively (Figure 11C, 11D).
Because of the similarity in ECM composition in both models, we investigated whether PDGF ligands were up-regulated in Pten null liver. While Pdgfb mRNA was undetected, both Pdgfa and Pdgfc mRNAs were up-regulated in Pten null tumors by 3.0-fold (p = 0.0007 and p = 0.002, respectively) and in adjacent tissue by 2-fold (p = 0.003 and p = 0.02, respectively) (Figure 12). The up-regulation of both PDGFA and PDGFC may therefore explain the common ECM changes observed in the PDGFC Tg and Pten null tumors and adjacent tissue.
Cell-ECM interactions are largely mediated through receptors called integrins made up of α and β chains. While integrin α5 was the most abundant and commonly up-regulated integrin chain in the tumors of both mouse models, the pattern of the other identified integrins was significantly different between the two models (Figure 13). Integrins α2b, α3 and β1 were specific to Pten null tumors while integrins α8 and β5 were specific to PDGFC Tg tumors. The up-regulation of integrin α6 and of CD44 was also much stronger in Pten null tumors compared to PDGFC Tg tumors. The differential expression of integrins α6 and α8 was further validated. Integrin α6 mRNA was increased by 39.0-fold (p = 0.001) in Pten null tumors and by 11.6-fold (p = 0.002) in PDGFC Tg tumors (Figure 14A, 14B). Integrin α8 mRNA was increased in PDGFC Tg liver tissue at all disease stages by 16.2- to 24.0-fold (Figure 14C) but remained unchanged in Pten null liver (Figure 14D).
In summary (Figure 15), this study identified collagens type IV, VI, VII, X, XIV, XV, XVI and the short variant of COL18A1, NC1-301, as tumor-associated collagens in HCC. Laminin 521 was the most abundant laminin in HCC and integrin α5 the most abundant integrin subunit. High ratios of COL18A1 variant NC1-301 over COL18A1 variant NC1-764, high ratios of integrin α6 over integrin α8 and high levels of integrin α3 were specifically observed in the Pten null tumors.
The microenvironment can have profound influences on cellular behavior, survival and growth of developing tumor cells [14]. Increased rigidity of the ECM is commonly associated with HCC [15] and ECM deposition and matrix remodeling has been shown to affect liver progenitor cell expansion [16]. We characterized and compared ECM protein changes occurring during tumor development in the PDGFC Tg mouse model, a model of HCC associated with fibrosis and angiogenesis [4] and in the Pten null mouse model, a model of liver tumors of mixed cholangio- and hepatocytic features [6]–[8], [17]. This study represents the most comprehensive analysis of the ECM and associated receptor proteome reported to date, and demonstrates the utility of mass spectrometry-based approaches to study gene families with extensive sequence homology, post-transcriptional and post-translational regulations. It is also the first study to analyze and compare proteome changes occurring during the transition from fibrosis and steatosis to HCC in two mouse models.
Collagens, the most abundant structural components of the ECM are homo- and heterotrimeric molecules whose subunits, the alpha chains, are distinct gene products. Forty-four different alpha chains have been sequenced, several of them being differentially spliced, which adds to the diversity of the collagen family. To date, 28 different combinations of the alpha chains (collagen types I–XXVIII) have been identified or predicted to exist (www.uniprot.org). While only ten collagen types have been described in the liver [15], this extensive proteomic study resulted in the identification of 16 types. Fibril-forming types I and III collagens are predominantly synthesized by hepatic stellate cells and are used as markers for liver fibrogenesis. In the fibrotic liver of PDGFC Tg mice, type I and III collagen levels strongly increased with a significantly higher increase of type I collagen than of type III collagen changing the I/III ratio from 1∶1 in the healthy liver to over 2∶1, as observed in human fibrotic liver. Collagens type V (COL5A1, COL5A2 and COL5A3) and type II (COL2A1), the other fibril-forming collagens, were also up-regulated in the fibrotic liver of the PDGFC Tg mice. This ECM composition in the fibrotic liver of PDGFC Tg mice is consistent with a signature of activated hepatic stellate cells, a hallmark of PDGFC Tg mice [4]. Out of the 26 alpha chains identified, 15 were up-regulated in tumors of both models. These include the six alpha chains of collagen IV, the three alpha chains of collagen VI, COL7A1, COL10A1, COL14A1, COL15A1, COL16A1 and COL18A1. Collagen VI, a component of microfibrillar structures in many tissues, is a heterotrimer with the chain composition (6a1)(6a2)(6a3). Type VI collagen binds cells and may be involved in cell migration and differentiation and embryonic development. All collagen VI subunits, including splice variants for COL6A2 and COL6A3, were up-regulated in the tumors and adjacent tissue of both models. The most abundant structural component of basement membranes is collagen IV. The six different alpha chains 4a1–4a6 were up-regulated in the tumors of both mouse models. Besides the heterotrimeric molecule (4a1)2(4a2) composed of the two most abundant collagen IV subunits, the other combinations between alpha chains, particularly those including the subunits of minor abundance, are not yet established. Whereas COL4A1 and COL4A2 are found in all basement membranes studied, COL4A3, COL4A4 and COL4A5 are found only in a subset of basement membranes and are always found together [18]. Strong deposition of collagen type IV was described in dysplastic areas and small HCCs in human cirrhotic livers indicative of early events in hepatocarcinogenesis [19]. The multiplexin collagens XV and XVIII are also localized to basement membranes. COL15A1 was the collagen alpha chain that showed the stronger up-regulation in both the Pten null and PDGFC Tg tumors.
COL18A1 was also up-regulated in the mice tumors. Interestingly, increases in this protein correlated with the increase of a specific isoform of Col18a1 mRNA, isoform NC1-301, resulting from alternative promoter usage. NC1-301 mRNA was increased by over 30-fold in PDGFC Tg tumors and over 100-fold in Pten null tumors. In contrast, the canonical Col18a1 mRNA, NC1-764, was unchanged or slightly decreased in the PDGFC Tg tumors and strongly decreased in Pten null tumors. It was previously reported that NC1-764 mRNA expression decreases in advanced HCCs [20] and that cholangiocarcinoma cells expressed NC1-301 which was deposited in tumor basement membrane [21]. This is in good agreement with the changes we observed in both COL18A1 isoforms in the mice tumors, with a greater ratio NC1-301/NC1-764 in Pten null tumors compared to PDGFC Tg tumors. These results suggest that the ratio of COL18A1 isoforms could directly correlate with the expansion of intermediate cells co-expressing both hepatocytes and biliary markers. Finally, collagen VII, the main constituent of anchoring fibrils, was also up-regulated in tumors of both models. It has been reported that human epidermal cells devoid of collagen VII did not form tumors in mice, whereas those retaining the specific N-terminal NC1 domain were tumorigenic [22].
Other glycoproteins in basement membranes such as laminins and nidogen 1 increased in the mice tumors. Nidogens are believed to connect laminin and collagen IV networks, hence stabilizing the basement membrane structure and appear critical for anchoring other components such as perlecan. At present, five laminin α (α1–α5), three β (β1–β3) and three γ (γ1–γ3) chains and 16 trimers have been characterized in mouse and human [23]. Based on the chain identification, laminins 511 or 521 (previously called laminin 10/11) and laminin 522 are likely up-regulated in PDGFC Tg and Pten null tumors. It was reported that laminins containing the α5 chain serve as functional regulators of HCC progression [24]. Up-regulation of laminin β3, a major component of laminin 332 (previously called laminin 5) was observed specifically in Pten null tumors. Interestingly, laminin β3 was reported up-regulated in cholangiocarcinoma cell lines compared to HCC cell lines [25] and laminin 332 is present in almost all intrahepatic cholangiocarcinoma cases [26].
The global expression of the ECM in liver during tumor development results from the combined expression profiles of tumor cells, stromal cells, and non-tumor hepatocytes. Activated hepatic stellate cells and myofibroblasts express a wide spectrum of ECM molecules but an important fraction of ECM is also synthesized by other liver cells, notably sinusoidal and portal endothelia, bile duct epithelia and hepatocytes [27], [28]. While this study increases our knowledge of HCC-specific matrix composition, future studies should focus on the cellular distribution of the described proteins. Overall, beside a subset of laminins, the ECM changes were remarkably similar in the tumors and adjacent tissues of both mouse models, suggesting a common molecular and cellular mechanism. We therefore investigated the possibility that PDGF factors were up-regulated in Pten null mice. While Pdgfb mRNA was undetected in Pten null liver, both Pdgfa and Pdgfc transcripts were increased in Pten null tumors and adjacent tissues. We also observed an up-regulation of both PDGF receptor alpha and PDGF receptor beta in Pten null tumors and adjacent tissues (LB, personal communication).
Most cell types have the ability to bind to the surrounding ECM and certain ECM components can transmit signals to cells via transmembrane receptors [29]. Such matrix sensors are mainly integrins. Integrins comprise a large family of cell surface glycoproteins which consist of alpha and beta subunits and that regulate cell adhesion, migration, proliferation and apoptosis [30]. There are 18 α and 8 β subunits, each of which can bind to several partners giving rise to at least 24 distinct integrin heterodimers with different functions and ligand binding activities. Laminins are ligands for both α6β1 and α3β1 integrins. These integrins were specifically up-regulated in Pten null tumors. Interestingly, laminin 511 modulates human embryonic stem cell aggregation and adherence through binding of the α6β1 integrin receptor highly expressed in the membranes of undifferentiated stem cells [31], [32]. It was also reported that oval cells express integrin α6 [33]. Similarly, laminin 332 and integrin α3 were co-up-regulated in Pten null tumors. It was reported that cells lacking integrin α3 do not proliferate in response to laminin 332 treatment [34]. Altogether, these results suggest that laminin 332/integrin α3–induced HCC growth and that laminin 511-integrin α6β1 interaction is specific to Pten null tumors.
The identified HCC–associated ECM and integrin components could play an important role in the promotion of the early steps of hepatocarcinogenesis, providing a foundation for novel strategies to prevent, diagnose and treat HCC. Inhibiting the expression of ECM components or blocking their interactions with signaling integrins could serve as a means for establishing a non-permissive microenvironment that may prevent tumor development. Integrin inhibitors such as humanized antibodies or blocking peptides against integrin α5β1 are currently under clinical investigation. Our results suggest that these novel drugs should be evaluated for the treatment of HCC. In addition, integrin α3β1-laminin 332 and integrin α6β1-laminin 511 networks may be promising targets to prevent laminin-tumor cell interaction in HCC with dysregulated PTEN function.
The PDGFC Tg mice used for this study were previously described [4]. Liver tissue samples were collected by necropsy from 1.5-month old PDGFC Tg mice with hepatic fibrosis, 8-month old PDGFC Tg mice with small HCCs and from 1.5-month and 8-month old wild-type controls. Mice carrying Pten conditional knockout alleles were crossed with an Albumin (Alb)-Cre-transgenic mouse. The Alb-Cre-transgenic mice were genotyped using Cre specific primers. For this model, control animals are PtenloxP/loxP; Alb-Cre−. Liver tissue samples were collected by necropsy from 6-month old Pten null mice with steatosis, 9-month old Pten null mice with small HCCs and from 6-month and 9-month old control mice. All tissues were immediately snap-frozen in liquid nitrogen or fixed in 10% neutral buffered formalin overnight, processed to paraffin blocks, sectioned, and stained with hematoxylin/eosin or Masson's trichrome by using standard techniques. This study was carried out in strict accordance with the regulations of the U.S. National Institutes of Health. All of the work with animals was performed in adherence to the “Guide for the Care and Use of Laboratory Animals” published by the U.S. National Research Council, including the use of appropriate anesthesia whenever recommended by these guidelines. The protocol was approved and reviewed annually by the Institutional Animal Care and Use Committee of the Fred Hutchinson Cancer Research Center (File #1662). Every effort was made to minimize the number of animals required for the study and to minimize the pain and discomfort experienced.
Liver tissues from three or four mice in each study group were separately ground on dry ice and subsequently pooled. Proteins were extracted twice from 40 mg of each pooled group in 1 ml lysis buffer (5 M urea, 2 M thiourea, 2% w/v n-Octyl-β-D-Glucopyranoside, 40 mM Tris and 1 mM phenylmethylsulfonyl fluoride). Following centrifugation at 16,100×g at 4°C for 1 hr, the pellet fraction was solubilized in Laemmli buffer and the proteins from the supernatant were separated using the Alliance 2-D Bioseparations System (Waters Corporation, Milford, MA) as previously described [13]. Briefly, an anion exchange column, BioSuite Q 10 µm, (Waters Corporation, Milford, MA) was used for the first dimension. Eight stepwise gradients were performed consisting of 0, 100 mM, 200 mM, 350 mM, 500 mM, 650 mM, 800 mM and 1000 mM NaCl. The reversed phase columns, Symmetry300 C4 3.5 µm, (Waters Corporation, Milford, MA) were used for separation of the fractions obtained from the first dimension steps. Two reversed phase columns were switched through the column selector. A total of ∼260 fractions was collected for each pooled group. Some adjacent fractions were combined leading to a final number of 34 samples for each pooled group. All fractions were lyophilized and resuspended in Laemmli buffer. Proteins obtained from the 2-D HPLC separations were further separated by 12% SDS PAGE. Gel pieces were combined into 37 individual samples for each pooled group according to protein size and abundance, dehydrated with 100% acetonitrile and dried using a speed vacuum. Gel pieces were incubated with 10 µl of 6.7 ng/µl trypsin in digestion buffer overnight at 37°C. The reaction was stopped with 15 µl of extraction buffer (2% formic acid/3% acetonitrile) and the supernatants were collected.
The generated peptide samples were desalted using Symmetry C18 de-salting columns (Waters Corporation, Milford, MA) and subjected in duplicate to nanoflow LC-MS/MS analysis with a nano-UPLC system (Waters Corporation, Milford, MA) coupled to a hybrid 7-Tesla linear ion-trap Fourier-transform ion cyclotron resonance mass spectrometer (LTQ-FT, Thermo Scientific, Waltham, MA). Peptides were separated on a reversed phase column (75 µm×250 mm) packed with Magic C18AQ (5-µm 100 Å resin; Michrom Bioresources, Auburn, CA) and directly mounted on the electrospray ion source. We used a 60 min gradient from 10% to 40% acetonitrile in 0.1% formic acid at a flow rate of 300 nl/min. A spray voltage of 1600 V was applied. The LTQ-FT instrument was operated in the data-dependent mode, switching automatically between MS survey scans in the FTICR (target value 1,600,000, resolution 100,000, and injection time 1.5 s) with MS/MS spectra acquisition in the linear ion trap. The five most intense ions from the FT full scan were selected for fragmentation in the linear ion trap by collision-induced dissociation with a normalized collision energy of 30% at a target value of 10,000 (injection time 400 ms). Selected ions were dynamically excluded for 60 s. The absolute average mass accuracy for the parent ion was <5 ppm.
Acquired data were processed using the X!Tandem search algorithm [35] and PeptideProphet and ProteinProphet statistical tools [36], [37]. The tandem mass spectra were searched against the mouse International Protein Index protein sequence database (IPI, version 3.34, http://www.ebi.ac.uk/IPI/). The following search criteria were used in all cases: trypsin specificity, 2.5 Da of mass accuracy for the parent ion and methionine oxidation as a variable modification and when specified, lysine oxidation was added as a variable modification. Relative abundance scores were calculated for individual proteins based on total peptide counts normalized to account for the total amount of protein in the mixture.
Total RNA was extracted from individual liver tissue samples and purified using the Trizol reagent (Invitrogen, Carlsbad, CA). RNA samples were then submitted to DNAse digestion, reverse transcription using random hexamers and real-time PCR using specific primers listed in Table S2. For each sample, the cDNA equivalent to 1 ug total RNA per 20 µl reaction was amplified with the iCycler MyiQ using SYBR Green Supermix and analyzed by MyiQ software (Bio-Rad Laboratories, Hercules, CA) and relative quantification of RNA expression was calculated with the 2−ΔΔCt method. Actin quantification was used for normalization. The specificity of qPCR products was confirmed by melting curve analysis and gel-based analysis of the PCR products.
Proteins from individual liver tissues were extracted from 100 mg of liver, in 1 ml lysis buffer consisting of 5 M urea, 2 M thiourea, 2% w/v n-Octyl-β-D-Glucopyranoside (Sigma-Aldrich, St Louis, MO), 50 mM Tris (Fisher Scientific) and 1 mM phenylmethylsulfonyl fluoride (GE Healthcare, Little Chalfont, UK). The homogenate was centrifuged at 19,000× g at 4°C for 1 hour and supernatant was collected. Proteins (40 ug) were loaded onto 12% SDS PAGE gels and immunoblotting was performed using rat monoclonal anti-Nidogen 1 antibody (Santa Cruz Biotechnology, Santa Cruz, CA) at 1/100 dilution. Immunoreactivity was revealed by enhanced chemiluminescence using ECL kit (GE Healthcare, Little Chalfont, UK) and quantification was performed using ImageJ (http://rsbweb.nih.gov/ij/).
|
10.1371/journal.ppat.1000779 | Highly Differentiated, Resting Gn-Specific Memory CD8+ T Cells Persist Years after Infection by Andes Hantavirus | In man, infection with South American Andes virus (ANDV) causes hantavirus cardiopulmonary syndrome (HCPS). HCPS due to ANDV is endemic in Southern Chile and much of Argentina and increasing numbers of cases are reported all over South America. A case-fatality rate of about 36% together with the absence of successful antiviral therapies urge the development of a vaccine. Although T-cell responses were shown to be critically involved in immunity to hantaviruses in mouse models, no data are available on the magnitude, specificity and longevity of ANDV-specific memory T-cell responses in patients. Using sets of overlapping peptides in IFN-γ ELISPOT assays, we herein show in 78 Chilean convalescent patients that Gn-derived epitopes were immunodominant as compared to those from the N- and Gc-proteins. Furthermore, while the relative contribution of the N-specific response significantly declined over time, Gn-specific responses remained readily detectable ex vivo up to 13 years after the acute infection. Tetramer analysis further showed that up to 16.8% of all circulating CD3+CD8+ T cells were specific for the single HLA-B*3501-restricted epitope Gn465–473 years after the acute infection. Remarkably, Gn465–473–specific cells readily secreted IFN-γ, granzyme B and TNF-α but not IL-2 upon stimulation and showed a ‘revertant’ CD45RA+CD27−CD28−CCR7−CD127− effector memory phenotype, thereby resembling a phenotype seen in other latent virus infections. Most intriguingly, titers of neutralizing antibodies increased over time in 10/17 individuals months to years after the acute infection and independently of whether they were residents of endemic areas or not. Thus, our data suggest intrinsic, latent antigenic stimulation of Gn-specific T-cells. However, it remains a major task for future studies to proof this hypothesis by determination of viral antigen in convalescent patients. Furthermore, it remains to be seen whether Gn-specific T cells are critical for viral control and protective immunity. If so, Gn-derived immunodominant epitopes could be of high value for future ANDV vaccines.
| In man, hantavirus cardiopulmonary syndrome (HCPS) caused by Andes Virus (ANDV) is endemic in the Southern cone of Chile and Argentina but cases of HCPS are being increasingly reported all over South America since 1995. HCPS is characterized by fulminant pulmonary edema which progresses to shock and death in about 36% of patients with HCPS. Nevertheless, to date, neither antiviral treatments nor vaccines inducing neutralizing antibodies (NAb) have proven effective against HCPS-causing hantaviruses. We set out for the first study on human cellular immunity towards ANDV in 78 convalescent survivors of ANDV infection. We found that Gn-specific responses were predominant as compared to N- and Gc-specific responses, even up to 13 years after the infection. Surprisingly, most of the Gn-specific responses were restricted to two neighboring epitopes within the Gn carboxyterminus. Interestingly, among HLA-B*3501+ patients, Gn465−473-specific CD8+ T-cells showed highly differentiated but resting phenotype and functions. It remains to be seen in future studies whether the immunodominace of Gn-specific T-cells is crucial for protective immunity. Most intriguingly, titers of neutralizing antibodies increased in 10/17 individuals months to years after the acute infection and independently of whether they were residents of endemic areas or not. Thus, our data suggest viral persistence or latency in part of ANDV-convalescent patients. However, it remains a major task for future studies to proof the concept of latent/persistent human ANDV infection by the determination of viral antigen in convalescent patients.
| The family Bunyaviridae is comprised of five genera of tri-segmented negative-stranded RNA viruses, which are responsible for a considerable burden of zoonotic disease in man. While most are tick- or mosquito-borne, members of the genus Hantavirus are transmitted from chronically- and asymptomatically-infected rodents to humans via aerosols, which may derive from urine, feces or saliva. Globally hantaviruses may cause as many as 200,000 cases of human disease per year. In man, two clinical conditions may arise: hemorrhagic fever with renal syndrome, caused by the Asian and European strains (e.g. Hantaan, HTNV and Puumala, PUUV) or hantavirus cardiopulmonary syndrome (HCPS), which is caused by Sin Nombre virus (SNV) and Andes virus (ANDV), among others in the Americas. HCPS is an emerging infectious disease in North- and South America [1]-[5] and, currently, Chile represents among the most endemic regions for HCPS with more than 580 cases since 1995 [6].
As for ANDV, transmission to man is followed by infection of lung endothelial cells and, after an incubation period of 7 to 39 days [7], the development of a vascular leakage syndrome, eventually leading to massive pulmonary edema, shock and, in many cases, death. The high case-fatality ratio (mean 36%), the absence of a proven antiviral treatment or a vaccine, their mode of transmission and their potential use as weapons for bioterrorism, have rendered HCPS-causing hantaviruses Category A pathogens within NIAID's biodefense program [8]. Importantly, ANDV is the only hantavirus for which person-to-person transmission has been repeatedly documented [9]–[11].
The hantavirus virion contains a lipid-bilayer envelope into which both constituents, the Gn and Gc antigens of the heteromeric glycoprotein, are inserted via transmembrane domains. In the viral core, there are three nucleocapids each consisting of the RNA-binding N or nucleocapsid protein in complex with one of the genomic RNAs. These mRNAs encode the RNA-dependent RNA polymerase or L protein on the large or L segment (2153aa), the Gn (650aa) and Gc (490aa) glycoproteins on the middle or M segment, and the N protein (430aa) on the S segment [12].
Currently, there is a big discrepancy regarding the role of T cells in either pathogenesis or immunity of hantavirus infections. On one hand some studies in SNV-infected patients describe a correlation between the severity of HCPS and either the frequency of SNV-specific CD8+ T-cells [13] or the HLA-B35 haplotype [14], suggesting a T-cell driven pathogenesis of HCPS. On the other hand, several early reports highlight the importance of lymphocytes for immunity of mice towards hantaviruses, such as HTNV [15]–[17]. Likewise, clearance of HTNV in newborn mice was dependent on TNF-α production and cytotoxic activity of specific CD8+ T cells [18]. In addition, (HTNV-) N-protein-specific memory T-cells conferred partial protection and cross-protection towards N-expressing vaccinia virus [19] or hantaviruses [20] in mice and Syrian hamster [21],[22], respectively. In line with these findings we have recently reported that clearance of ANDV-RNA from peripheral blood cells of a patient was closely related to the appearance of cytotoxic CD8+ T cells about two months after the acute infection [23]. This observation together with the finding that we were unable to detect memory T-cells in many of the survivors of ANDV-induced HCPS (see below) led us to the hypothesis that T-cells may be crucial for protection and immunity towards ANDV rather than the pathogenesis in ANDV-infected patients. Concisely, knowledge of targeted epitopes and functional properties of ANDV-specific T-cells in ANDV-survivors may be important for both future studies in acutely ill patients and possibly for vaccine development.
Despite its importance, the knowledge of the human cellular immune response to hantaviruses is limited. To date, only few studies have assessed SNV-, HTNV- or PUUV- specific T cells [13], [24]–[28] in rather small patient cohorts and based on individual in-silico-predicted peptides and/or T-cell lines that had been expanded in vitro. Thus, the overall magnitude of human T-cell responses in vivo and the epitope-hierarchy within the memory T-cell pool in convalescent patients remains uncertain. Also, phenotype, effector functions and longevity of specific memory T-cells in humans remain to be elucidated. Greater knowledge of these matters would likely to be of high value to potential vaccine developers.
In an effort to gain insight into human cellular immunity to ANDV and to establish an immuno-hierarchy among ANDV-antigens, we carried out a study of the viral protein-specific T cell responses in 78 Chilean patients with past ANDV-infection. Our findings on the immuno-hierarchy among major structural hantavirus proteins and the frequencies as well as the functional features of CD8+ memory T cells may be of special interest for vaccine development since all attempts to induce long-lasting neutralizing humoral immunity have been unsuccessful so far [29].
In order to quantify circulating ANDV-specific T-cells ex vivo and to determine the immunodominant epitopes of ANDV, we first challenged PBMC of 78 Chilean convalescent patients (between 4 months and 13.2 years after hospitalization due to infection) with 310 overlapping peptides (distributed in 13 pools) spanning the entire N- and Gn/Gc precursor proteins [30] in IFN-γ ELISPOT assays.
Based on the criteria we used to score a sample as “positive” (see Material and Methods), 51 (66%) of the 78 patients showed significant responses against epitopes of at least one of the three viral antigens (Fig. 1A). Among these patients, 33/51 (65%) showed significant responses against epitopes of the N-protein, while 13/51 (25%) showed Gc-specific T-cells (Fig. 1B). However, 80% (41/51) of the positive patients launched significant responses towards Gn-derived epitopes. Moreover, while mean responses among the 51 individuals reached the sum of 1809 Spot Forming Units (SFU)/106PBMC when considering all viral antigens, Gn-specific responses accounted for more than half of the total response, at 973 SFU/106PBMC (Fig. 1C) as compared to 697 and 139 SFU/106 PBMC for N- and Gc-epitopes, respectively. Since all patients were BCG-vaccinated twice during childhood, we determined BCG-specific T cells (n = 10), resulting in a mean of 162 SFU/106 PBMC. Thus, Gn is the immunodominant antigen in ANDV-convalescent individuals.
We next asked whether differences in the longevity of each of the specific T-cell categories could account for the relative immunodominance of Gn among ANDV antigens, e.g., whether Gn-specific cells might persist longer than did cells responsive to the other antigens. We therefore considered the numbers of circulating N-, Gn- and Gc-specific T cells of each patient in relation to the time between the patient's hospital admission due to HCPS and the timepoint at which T cells were applied to ELISPOT assays (Fig. 2A–C). Although this approach does not allow to draw conclusions on the slope of antigen-specific responses at the level of a given individual, it is possible to directly compare the different antigen-specific responses within the overall cohort over time. Similar approaches have been previously performed on a cohort of smallpox vaccinees in order to estimate longevity and half-life of cellular and humoral memory responses [31].
Interestingly, only the N-specific response (Fig. 2A) exhibited a negative and significantly descending linear regression slope over time (r = −0.11, p<0.05), as when compared to Gn- and Gc-specific responses (Fig. 2B, C) (r = 0.07 and r = 0.02, respectively). We next segregated the 51 patients with positive T-cell responses into six groups according to the time past since hospitalization due to ANDV-infection (<1, 1–2, 2–4, 4–6, 6–9 and >13 years, respectively). As can be seen in Figure 2D, up to four years after infection N-specific responses were predominant, whereas afterwards Gn-specific relative contributions to the overall T-cell response increased from approximately 40% to more than 80% (Fig. 2D). Taken together, these data suggest that Gn- and Gc-specific T-cell responses are more stably maintained as compared to N-specific responses. However, in light of the limited value of the cross-sectional data available to us, future prospective studies assessing individual T-cell responses from the acute to the convalescent phase in individual patients would be needed to determine the absolute half-life of the N-, Gn- and Gc-specific T-cell responses.
Subsequently, results from IFN-γ ELISPOT assays revealed that regions Gn1-230 and Gn221–450 elicited a mean of 102 and 209 SFU/106 PBMC in 34% and 38% of all responsive patients, respectively (Fig. 3A). However, Gn441–650 elicited a mean response of 623 SFU/106PBMC (range 0–5506 SFU/106PBMC) in 24/51 (46%) patients. These data clearly indicate that epitopes within the carboxyl-terminus of Gn of ANDV are responsible for the immunodominance of Gn among ANDV-convalescent individuals. In contrast to the strong Gn-specific responses, Gc641–815 and Gc806–980 elicited a response in 24% (77 SFU/106PBMC) and 11% (25 SFU/106PBMC) of all patients, whereas Gc971–1140 was targeted by only 4% (30 SFU/106PBMC) of patients (data not shown).
We next stimulated cryopreserved PBMC from patients with Gn441–650-specific response using peptide pools representing Gn441–505, Gn496–560, Gn551–615 and Gn606–650, respectively. This approach revealed that region Gn441–505 comprised the major epitopes of Gn441–650 (data not shown). We then challenged cryopreserved PBMC of 11 patients with individual peptides spanning region Gn441–505. As shown in Figure 3B, only three patients (p30, p40, p59) exclusively recognized peptides Gn451–465 and Gn456–470, whereas p17 and p57 additionally recognized Gn461–475 and Gn466–480. By contrast, six patients (p6, p10, p15, p28, p32, p53) showed exclusive recognition of Gn461–475 and Gn466–480. Thus, a total of five patients recognized Gn451–465/Gn456–470, whereas eight patients recognized Gn461–475/Gn466–480. Subsequently, four additional individuals with exclusive and significant responses towards Gn461–475/Gn466–480 were identified (data not shown).
Together with a previous report from our lab [23], these epitopes are the first described within the Gn-region of hantaviruses. In addition, we determined immunodominant regions of ANDV N-protein, which included N1–70 (24.3% of responsive patients elicited a significant response) and N121–190 and N181–250 (21.2% and 19.5%, respectively) eliciting mean responses of 113–147 SFU/106PBMC (data not shown). Downmapped individual epitopes within the N-protein are summarized in Table 1.
We next wondered, which state of differentiation was expressed by the IFN-γ producing memory CD8+ T cells. Intracellular cytokine staining showed that IFN-γ+CD8+CD45RO+ T-cells expressed varying levels of CD45RA but consistently expressed a CCR7−CD28−CD27− effector memory phenotype (Fig. 3C). In line with this terminally differentiated phenotype, 25% of all IFN-γ+ T cells secreted granzyme B upon stimulation with their cognate peptide (Fig. 3D) whereas no IL-2 could be detected (data not shown). In addition, up to 45% of these IFN-γ+CD8+ memory T cells co-expressed TNF-α as determined by ICS (data not shown). These findings suggest that even years after acute ANDV infection (e.g. p32 and p40 were investigated 5.4 and 13.2 years after hospitalization, respectively), high frequencies of cytolytic memory CD8+ T-cells are maintained in the periphery.
As both the Gn451–465/Gn456–470 and Gn461–475/Gn466–480 epitopes share 10 amino acids in sequence within the pairs, we reasoned that one CD8+ T-cell epitope may be located within each overlapping sequence, respectively (e.g. Gn456–465 and Gn466–475). In support of this hypothesis, the analysis of HLA-A, -B, -DR and -DQ alleles revealed that 5/5 patients with response towards Gn451–465/Gn456–470 exclusively shared the HLA-A*24 allele (data not shown), whereas the HLA-B*35 allele was the only allele shared by all 12 patients recognizing Gn461–475/Gn466–480 (Fig. 4A). These data suggest the existence of two separate but neighboring CD8 T-cell epitopes in the carboxyterminal region of Gn that are restricted by HLA-A*24 and HLA-B*35, respectively. Indeed, as shown in Figure 4B, a significant response in a Gn461–475–specific T-cell line could only be detected when the HLA-B*35 allele was present on heterologous APCs (B-LCL).
It was previously suggested that severe HCPS due to SNV is associated with the HLA-B*35 allele [14] and with CD8 T-cell responses restricted to it [13]. We therefore were interested in the relation of memory T-cell responses and outcome of their ANDV infection in HLA-B*35-positive and negative patients (Fig. S1). Among all 78 patients, that is patients with (n = 51) and without (n = 27) significant memory T-cell responses, no differences in overall T-cell responses could be observed when comparing HLA-B*35-negative patients with mild or severe HCPS (Fig. S1A). By contrast, we found an about 3-fold higher overall T-cell response in HLA-B*35-positive patients with mild HCPS as compared to both HLA-B*35-positive patients with a history of severe disease and either group of HLA-B*35-negative patients (Fig. S1B). Likewise, 10/12 (83%) HLA-B*35-positive patients with significant responses to Gn461-475 had a history of mild HCPS (Fig. S1C). Thus, these data suggest that HLA-B*35-restricted memory T-cell responses are related to mild rather than to severe disease outcome.
We next sought to determine the optimal epitope of Gn461–475/Gn466–480 (Fig. 4C). Because we consistently observed stronger immune responses towards Gn461–475 (SLFSLMPDVAHSLAV) than towards Gn466–480 (MPDVAHSLAVELCVP), we reasoned that Leucine at position 465 may increase either the binding affinity or the TCR-recognition of the overlapping sequence Gn466–475 (MPDVAHSLAV). We therefore decided to generate Gn461–475–specific T-cell lines from HLA-B*3501 individuals and then challenged these cells with cleaved peptides of Gn465–475 (LMPDVAHSLAV). As shown in Figure 4C, cleavage of the carboxyterminal Leucine at position 473 led to a complete loss of epitope recognition. Similarly, elimination of the aminoterminal Methionine at position 466 was critical for epitope recognition. Most interestingly, virtually identical results were obtained when cells from HLA*B3501, HLA-B*3502 and HLA-B*3505 individuals were challenged with cleaved peptides, indicating that the Gn466–473 epitope is equally immunogenic among different HLA-B*35 subtypes.
We next were interested in comparing the phenotype of Gn465–473 restricted T cells and with other HLA-B*3501-restricted virus-specific T cells in seven HLA-B*3501 positive ANDV-convalescent patients (Fig. 5A, B). The mean abundance of Gn465–473 specific T-cells was higher than those specific for N131–139 and Gc664–673 –epitopes, described by Kilpatrick et al. [13], the latent EBV-epitope EBNA3A458–466 [32], the Influenza A NP418–426 epitope [33], or the Rv2903c201–209 epitope of Mycobacterium tuberculosis, known to be recognized by BCG-vaccinated individuals [34].
In the seven HLA-B*3501-positive individuals we had detected between 0 and 4394 SFU/106 PBMC (0%–0.0044%) following stimulation with peptide Gn461–475 in IFN-γ ELISPOT assays. When normalizing the results by the percentages of CD3+CD8+ cells in these individuals (range 9.9–33.9% of PBMC, mean 19.7%), one would have expected between 0% (p45) and 0.036% (p17) of CD3+CD8+ cells being tetramer positive. However, we found 0.3% and 5.9% of all CD3+CD8+ T cells being positive for Gn465–473:HLA-B*3501 tetramer complexes, respectively. In addition, the highest frequencies for tetramer-positive cells were found in patient 10 (16.8% of CD3+CD8+ cells), whereas only 0.0111% of his CD3+CD8+ cells produced IFN-γ in ELISPOT assays. This discrepancy between both detection methods is in line with previous reports [35].
We next determined the state of differentiation of Gn465–473-specific T-cells, where a clear dichotomy was observed (Fig. 6A, B). In patients with positive responses towards Gn461–475 in IFN-γ ELISPOT assays (IFN-γ ++), Gn465–473 specific T cells were mostly CD45RA+CCR7− and significantly more of a differentiated CD28−CD27− phenotype as compared to IFN-γ− samples. In addition, we found significant differences with regard to the IL-7Rα (CD127), which is crucially involved in maintenance of memory T-cells in the periphery in the absence of cognate antigen [36]. Patients with IFN-γ+ ELISPOT results showed mainly CD127− Gn465−473 T cells, whereas T cells of IFN-γ− patients were mostly CD127+ (Fig. 6A, C). Thus, Gn465−473-specific CD8+ T cells showed a phenotype that is clearly distinct relative to that described for other self-limited diseases such as those caused by influenza A and respiratory syncytial virus but more resembled the pattern associated with latent infections, such as past exposure to CMV [37]. Because a CD28−CD27−CD127− phenotype was previously described to be a result of ongoing antigen-stimulation, as found in latent CMV infection, we next determined the expression of activation markers, such as of KLRG-1, CD69, CD38 and CD25 on Gn465–473 and Influenza A-specific T cells within the seven HLA-B*3501+ ANDV-convalescent (Fig. 6D). No significant differences could be observed between IFN-γ+ and IFN-γ− Gn465–473-specific populations or between Gn465–473- and Influenza A NP418–426-specific T-cells.
In a next step we assessed whether re-exposure to viral antigens could have led to a boost in the donor's immune response. We therefore compared memory T-cell responses in patients who got infected during recreation (R-patients) with those of residents in endemic areas (E-patients) (Fig. 7A). No significant differences were observed between the two groups in those responses, although endemic patients revealed about double as many Gn-specific memory T-cells than recreational patients (mean 765 vs 361 SFU/106 PBMC) for unclear reasons. However, these results are not in line with the hypothesis of repeated viral exposure in patients who reside in endemic areas, since N- and Gc-specific responses were virtually identical in both groups.
In addition, we identified seven individuals, which had been infected during recreation (R-patients, Fig. 7B, C) and ten individuals residing in endemic areas (E-patients, Fig. 7D, E), for all of which two prospective serum samples were available. The time period between sample 1 and sample 2 was 0.3–6.9 years and 1.2–4.1 years in R- and E-patients, respectively. Surprisingly, in R- and E-patients anti-N titers raised four- to 64-fold between samples 1 and 2 in 4/7 and 5/10 patients, respectively. Most intriguingly, however, also neutralizing antibody (NAb) titers rose two- to eight-fold in 4/7 and 6/10 of R- and E-patients, respectively. Importantly, NAb titers, measured by a blinded worker, increased two- to four-fold in patients R1, R3, R5, E1 and E4 between sample timepoint 1 and 2, although in all cases sample 1 was taken months to years after the acute phase. Taken together, these results suggest that re-exposure to extrinsic, environmental virus is not responsible for the observed rise in NAb titers or high frequencies of memory T cells.
Finally, we sought to prospectively study Gn-specific T-cells in three patients. When Gn465–473-specific T cells were phenotyped over a time period of two years (Fig. 8A–C), no dynamic changes of the CD27− population could be observed, indicating that differentiated Gn465–473-specific T cells are able to stably persist at high frequencies without the need for B7:CD28- or CD70:CD27-mediated survival signals. However, in patients E9 and E2 (Fig. 8A, C, respectively), Gn465–473-specific T cells actually increased over time, paralleling the two- to eight-fold increase in NAb titers observed in these two individuals (Fig. 7E).
Infection with ANDV is the predominant cause for HCPS in South America. Case-fatality rates of currently 36%, person-to-person transmission and the absence of a proven effective antiviral treatment urge the development of a vaccine. Although the protective potential of neutralizing antibodies against the hantavirus surface glycoproteins Gn and Gc, but not the N-protein, was established in vitro [38]–[40] and in animal models [41]–[43], efforts to induce long-lasting neutralizing antibodies in human volunteers have been unsuccessful so far [29],[44] or remain to be proven effective and long lasting [45].
On the other hand, several early reports highlight the importance of lymphocytes for immunity of mice towards hantaviruses, such as HTNV [15]–[17]. Likewise, appearance of virus-specific CD8+ T cells with cytotoxic activity and the ability to produce IFN-γ and TNF-α was associated with clearance of HTNV in newborn mice. In contrast, HTNV infection was not cleared when TNF-α production and cytotoxic activity of specific CD8+ T cells were impaired [18]. In another report, (HTNV-) N-protein-specific CD8+ memory T-cells, induced by a DNA vaccine, conferred partial protection against re-challenge with a vaccinia virus expressing the N-protein [19]. One study in mice showed that N-protein specific T cells rather than antibodies mediated protection and cross-protection upon re-challenge with homologous and heterologous hantaviruses [20]. Finally, ANDV infection of Syrian hamsters - the sole animal model for human HCPS–could be prevented for at least 10 months by previous vaccination with ANDV N-protein [21], again indicating that protection can be achieved independently of neutralizing, Gn/Gc-specific antibodies. Most recently Safronetz et al. confirmed these findings in Syrian hamsters vaccinated with Gn-protein expressing Adenovirus vectors. Interestingly, these animals were protected from lethal ANDV infection independently of neutralizing antibodies and showed no or very low levels of ANDV-RNA up to 9 days after ANDV infection [22]. As for ANDV-infection in man, we recently showed that clearance of ANDV-RNA from peripheral blood cells was closely related to the appearance of cytotoxic CD8+ T cells, but not NAb, in a patient about two months after the acute infection [23].
Taken together, these reports suggest that cytotoxic T cells are crucially involved in clearance and protection from hantaviruses. Conversely, establishment of hantavirus-specific cytotoxic memory CD8+ T cells prior to infection, e.g. by a vaccine, may provide protective, albeit not sterilizing, immunity to the host. However, limited information on human cellular immunity to hantaviruses is available and, to date, only one study addresses ANDV-specific T-cell responses [23].
Using a panel of 310 overlapping peptides spanning the entire N-, Gn- and Gc-protein of ANDV allowed us to study most, existing T-cell epitopes in 78 convalescent survivors of ANDV infection in a non-HLA-restricted manner. In contrast to other reports with a similar approach, the majority of responses were specific for Gn- but not N-protein-derived epitopes. Thus, our results in ANDV-infected patients seem to contradict the current dogma of N-protein being the principal T-cell immunogenic hantavirus antigen [19], [20], [22], [24], [25], [27], [46]-[48]. However, in previous studies, epitope-specific T cells were detected either by in vitro expansion prior to testing or using individual peptides or tetramer complexes for ex vivo detection in a small and HLA-selected patient populations [13], [24]–[28]. Thus, differences in the experimental design, rather than its elevated immunogenicity, may explain, why we found Gn being the immunodominant antigen of ANDV, whereas no single Gn-epitope had been described for other hantaviruses. In fact, 92–96% of the amnio acid sequence of the two Gn-epitopes described herein, are conserved within the PUUV and SNV sequence, respectively. An alternative explanation may derive from the differences between our study and other studies in the timing of T-cell testing after the acute phase or differences in infection kinetics between the different hantaviruses. Specifically, as can be seen in Figure 2D, N-derived epitopes seemed to be relatively predominant up to four years after the acute infection, whereas Gn-derived epitopes were predominant in patients with a longer convalescence phase. In addition, the kinetics of NAb titers in our patients suggest that viral antigen may be present for months or years after the acute infection, a phenomenon which has not been described for other hantaviruses. Thus, differences in epitope avidity and/or precursor expansion over time may have contributed to the relative predominance of Gn-specific T-cells in our study.
Of note, 80% of all patients with detectable T-cell responses recognized Gn-derived epitopes (Fig. 1B). To our surprise, however, responses against Gn were not broad but rather focused to the carboxyterminus of Gn, namely the region of aa 451–480. Interestingly, the cytoplasmic tail of Gn has been shown to contain important virulence factors as it suppresses type 1-interferon responses in infected cells [49],[50]. On the other hand, the carboxyterminal 142 residues of pathogenic (namely of ANDV and HTNV), but not non-pathogenic hantaviruses, prone the C-terminal tail of Gn towards degradation by the proteasome, which then leads to the presentation of epitopes by MHC I molecules to CD8+ T-cells [51]. This mechanism could explain the relative immunodominance of Gn-derived epitopes seen in our study and also may represent a virulence factor of ANDV suggesting that T cells are causative for HCPS. Nonetheless, an early and vigorous cytotoxic T-cell response towards epitopes of the C-terminal Gn may also be able to restrict the virulence of ANDV infection. Indeed, among HLA-B*35-positive patients mild disease outcome seemed to be associated with stronger responses towards the Gn-carboxyterminus than in patients with severe HCPS (Fig. S1B, C). In line with this finding, a recent study in 87 Chilean ANDV-infected patients found that the HLA-B*35 allele was the most frequent allele among patients with mild disease and almost twice as frequent as in patients with severe disease [52]. Although these data seem to contradict previous reports describing both the expression of the HLA-B*35-allele [14] as well as HLA-B*35-restricted T-cell responses [13] as risk factors for severe HCPS by SNV, it remains speculative whether our results indicate an pivotal role for T cells in disease outcome. The size of the memory T-cell pool is only indirectly linked to the effector T-cell response by the original burst size [53] and therefore may not reflect the size and composition of the effector T-cell pool during the acute phase. Moreover, 33% of HLA-B*35-negative patients and 48% of patients without any detectable memory T-cell responses had a history of mild HCPS (data not shown). Likewise, 52% of patients with severe HCPS did not show memory T-cell responses (data not shown). Both seem to argue against an exclusive role of T cells for disease outcome. In addition, other hand, we also showed a discrepancy between ELISPOT and tetramer-derived T-cell frequencies (see above), which indicates the existence of IFN-γ-negative ANDV-specific T-cells. In fact, three of the seven studied HLA-B*3501 positive donors did not show significant responses in initial IFN-γ ELISPOT assays (Fig. 1) but showed substantial numbers of tetramer-positive CD3+CD8+ T cells. This is in line with previous reports comparing determination of T-cell frequencies by ELISPOT and tetramer analysis [35]. In this report tetramer analysis revealed on average ten-fold higher frequencies than IFN-γ ELISPOT assays of T cells specific for a HLA-A2-restricted HIV Gag-derived peptide. This report as well as our data suggests that the vast majority of virus-specific T cells may not readily secrete IFN-γ when stimulated by peptides in ELISPOT assays. It is also possible that these T cells were functional during the acute phase but not during convalescence. Alternatively, differences in the infectious dose (e.g. low versus high infectious dose) or the kinetics and doses of the evolving neutralizing antibodies may interfere with the functional quality of the memory T-cell pool. Taken together, additional studies with HLA-tetramers in acutely ANDV-infected patients will be necessary to better understand the role of HLA-B*35 and T-cell kinetics for ANDV disease outcome.
Another interesting aspect of our study concerns the longevity of the memory T cells in association with their highly differentiated phenotype. After more than 13 years after infection we still detected 564–2152 SFU/106 PBMC by ELISPOT, most of which were directed towards Gn-derived epitopes.
With regards to the longevity of CD8 memory T cells, our results nicely confirm a previous study by Van Epps et al. in PUUV-convalescent patients, in which up to 100–300 SFU/106 PBMC of N-epitope-specific CD8 T cells were found in three patients up to 15 years after acute infection [24],[25],[27]. However, by tetramer analyses we detected still-higher frequencies with up to 16.8% of all CD3+CD8+ T cells proving positive for the single epitope Gn465–473 while displaying a late effector memory phenotype (CD127−CD28−CD27−CCR7−). This phenotype was in line with the cells' ability to readily secrete granzyme B and TNF-α without IL-2. Surprisingly, we also found that up to 5% of tetramer+ cells, which, again in accordance with their CD28+CD27+CCR7+CD127+ phenotype, did not exert any immediate effector functions, such as IFN-γ secretion upon stimulation with their cognate epitope. While these data again show that screening by ELISPOT underestimates the real proportion of Gn-specific T cells [35], it is also tempting to speculate, albeit virtually impossible to prove, that patients with either of the observed phenotypes differ at their stage of immunity towards a possible ANDV re-challenge.
In absence of antigenic stimulation as well as of autocrine or paracrine IL-2 and co-stimulation via CD28:B7 and CD27:CD70 interaction, late effector memory T-cells are heavily prone to apoptosis. However, we herein were able to show in three HLA-B*3501+ patients that peripheral Gn465–473-specifc T cells were maintained in the periphery for at least two years, despite a consistent CD27− phenotype, thereby lacking the receptor for crucial anti-apoptotic signals provided by CD70. While the former perception was that senescent end-differentiated CD28−CD27− T cells were unable to divide, recent evidence suggests, that highly differentiated granzymeB+CD8+ memory T cells are actually dividing upon stimulation equally well as naïve CD8+ T cells [54]. In man, the CD45RA+CD28−CD27−CCR7− late effector memory phenotype has been mainly described in patients with latent CMV infection [37],[55]. In this model repetitive or latent antigen stimulation is supposed to drive CD8+ memory differentiation and/or the recruitment of new memory T-cells. However human hantavirus infections are not known to cause latent or chronic infections and we were not able to detect viral RNA in plasma or peripheral blood cells of patients with sustained and high T-cell responses (data not shown). Also, we failed to detect a fingerprint of recent antigenic stimulation of tetramer-positive cells through assessment of the expression of additional activation markers, including CD69, CD38 and CD25. In addition, expression of IL-7Rα (CD127) was described to be a critical factor for long-term survival of CD8+ memory T cells in absence of their cognate antigen. Since CD127 is usually downregulated upon antigen exposure and rapidly re-expressed after antigen clearance, it is consistent that mainly virus-specific CD127+CD8+ memory T-cells are found in studies on Influenza-, respiratory syncytial virus- and HBV- specific T-cells. By contrast, in persistently HIV-, CMV- or EBV-infected individuals T cells are maintained despite their lack of CD127 expression [56],[57]. Finally, KLRG1 is mainly expressed on antigen-experienced T-cells with immediate effector functions [55]. Thus, considering ANDV a self-limiting transient infection in man, a CD127+KLRG1− phenotype would have been expected years after the infection. However, IFN-γ +, but not IFN-γ −, Gn465–473-specific T cells expressed substantially less CD127 than their Influenza A virus (NP418–426)-specific counterparts, whereas no clear pattern could be observed regarding the KLRG1 expression. However, although no difference could be observed between ANDV (Gn465–473) and Influenza (NP418–426)–specific T-cells with regards to CD25, CD38 and CD69 expression, the lack of CD127 suggests persistent antigenic stimulation in individuals with IFN-γ+ T cells. When employing BLAST, we were not able to identify other organisms that share the amino acid sequence of Gn465–473, thereby making cross-reactivity an unlikely explanation. Consistently, no further serology testing was performed in convalescent patients.
We next hypothesized that residents of endemic areas might have received intermittent antigen-boosters due to viral re-exposure and therefore should show somewhat higher T-cell responses to all viral antigens. However, we did not find significant differences in ANDV-specific T cell numbers when comparing patients who reside in endemic areas and those, which got during recreation got infected in an endemic region (Fig. 7A), while residing in non-endemic areas.
Moreover, in the majority of prospective serum samples of ten patients from endemic regions and of seven patients from non-endemic regions, we surprisingly found an increase in both anti-N and NAb titers despite the fact that the second sample was taken years after the first samples in most cases. The fact that this was also observed in patients who never had returned to endemic regions since their primary ANDV infection, suggests that re-exposure to extrinsic (environmental) virus does not account for high antibody titers and, conversely, not for high ANDV-specific T-cell frequencies. Regarding the increase in NAb titers between sample 1 and 2, we cannot exclude that NAb titers continued to rise after sample 1 was drawn during or shortly after the acute phase. Therefore, it is possible that in these patients (Fig. 7) titers of sample 2 in fact were identical or even lower than the maximum titer achieved during the acute phase. However, NAb titers of 13/17 individuals were still relatively high (≥1∶400) at the timepoint of sample 2, that is 1.2–11.3 years after the acute infection. This argues for continuous antigen-exposure in both E- and R-patients, since NAb titers, in contrast to non-neutralizing antibodies, strictly depend on the presence of their cognate antigen. Specifically, in absence of antigen, murine NAb titers fall below the detection limit after 100–200 days [58]. Most importantly, however, we also found that NAb titers increased two- to four-fold in five patients (R1, R3, R5, E1 and E4, Fig. 7C, E) in which sample 1 was taken months to years after the acute phase. Since not only maintenance [58] but also kinetics of NAb titers heavily depend on the presence of viral surface antigens [59], these results support the hypothesis that re-exposure to viral surface (Gn/Gc) antigen is responsible for high and rising NAb titers in both R- and E-patients. Due to the fundamental differences between R- and E-patients in their risk for re-exposure to extrinsic virus, this is turn suggests that intrinsic viral antigen is responsible for the relative “immune-inflation” and the terminal differentiation of Gn-specific CD8+ T-cells. Thus, it may be that intermittent release of low doses of viral antigen from intrinsic virus (e.g. that never completely cleared from tissue reservoir(s)) is sufficient to maintain and boost of NAb titers and T-cell frequencies, whereas changes in activation marker expression on ANDV-specific T-cells are too short-lived (with the exception of low CD127 expression) to be consistently different (e.g. KLRG1) from that of Influenza A virus-specific T cells. However, as long as viral antigen or genome cannot be detected in convalescent patients as those described herein, the concept of latent or persistent ANDV infection in convalescent patients remains speculative. Future studies should therefore focus on antigen detection in tissues (e.g. surgical or post-mortem specimen) from solid (e.g. lung, kidney) and immuno-privileged (e.g. brain) organs.
Our data suggest that long-lived effector memory T cells can be maintained at high numbers in the periphery over years independently of IL-7. Notably, our findings resemble those found in murine Sendai and Influenza A virus infections, where epitope-specific T-cell clonal expansions occurred in absence of antigen throughout the CD8 memory pool [60]. As in our study, but in contrast to models of persistent infection, clonally expanded effector memory T-cells in these studies retained potent functionality despite their highly differentiated phenotype. Although we could not formally show clonal expansion for all our patients (with the exception of two individuals, see Fig. 8A, C) due to our non-prospective study design, similar underlying, but yet undefined, mechanisms may explain our findings in human ANDV infection.
The induction of a highly differentiated, resting e.g. Gn465–473–specific, memory T-cell subset might be of major interest in the context of vaccine development for several reasons. First, years after infection high numbers of these CCR7− Gn465–473–specific cells remain available for immuno-vigilance in the periphery. Second, as shown, this subset possesses the ability to readily secrete antiviral (e.g. IFN-γ, TNF-α) as well as lytic (granzyme B) effector molecules. Third, although most human studies seem to focus on epitopes restricted to HLA-A*02 because of its wide distribution among the Caucasian population (about 25%, [61]), it should be noted that within the Amerindian population, the frequency of HLA-B*35 positive individuals is about 70% higher than in Caucasians [61]. In fact, 25% (range 22–30%) of the inhabitants in ANDV endemic regions in Southern Chile express the HLA-B*35 and/or the HLA-A*02 allele, respectively [52]. Thus, for this population, HLA-B*35-restricted epitopes, like Gn465–473, might be of similar impact as HLA-A*02-restricted epitopes. However, it first has to be established in future studies (e.g. Syrian hamster models) whether and to which extend Gn-derived T-cell epitopes may contribute to protective immunity. Furthermore, although hantaviruses are not known to mutate, additional epitopes have to be identified in future studies in order to prevent failure of a T-cell based vaccine due to mutations within the Gn465–473 epitope.
Taken together, our results suggest that infection with ANDV may lead to a strong highly differentiated effector memory response. The findings concerning the predominant immunogenicity of ANDV-Gn protein may have implications for the understanding of immunity not only to ANDV, but also to other hantaviruses.
A total of 78 patients were enrolled between 4 months and 13.2 years after hospitalization due to either mild or moderate/severe HCPS. All patients had a previous confirmed hantavirus diagnosis done in Chilean reference laboratories by IgG serology to SNV and ANDV antigens by enzyme-linked immunosorbent assay (ELISA), as previously described [62]–[64].
Mild HCPS was defined by the sole support of the patient by symptomatic therapy, including respiratory support by an oxygen mask. On the other hand, ANDV-infected patients who required intensive care by mechanic ventilation and/or anti-shock treatment with vasoactive drugs were defined as moderate/severe HCPS.
All patients included were Chilean citizens and volunteered to participate without receiving monetary incentive. Prior to enrollment all patients enrolled signed informed consent, which was previously informed by IRB committees of Clínica Alemana de Santiago, the Chilean Ministry of Health and regional IRB committees. Before enrollment, patients were extensively informed about the intention of the study by the local study nurse. Upon enrollment patients did not suffer from any signs of active disease and were only enrolled if considered healthy donors. Samples consisted in 45 cc of peripheral venous blood, using tubes containing Sodium Heparin (BD vacutainer). Samples were shipped within 24 hours to our laboratory and were processed immediately upon receipt. PBMC were isolated by Ficoll-Hypague gradient and fresh PBMC were applied to ELISPOT assays. PBMC, which were not used immediately were cryopreserved in liquid nitrogen.
96-well filterplates (Millipore) were coated with 5 µg/ml of anti-hIFN-γ (Endogen, clone M700A) or 15 µg/ml anti-granzyme B (mabtech, clone GB10) at 4°C overnight one day prior to the assay. For granzyme B assays, prior to coating membranes were activated by incubation of the wells with 15 µl/well of 35% Ethanol for 1 minute. After washing and blocking of the plate, fresh or cryopreserved PBMC or T-cell lines were applied and incubated for 20 hours in an incubator (Nuraire) at 37°C and 5% CO2 in the presence of 310 overlapping 15mer peptides (Mimotopes, Australia) organized in 13 pools of 12 to 44 peptides (final concentration 1 µg/ml, each) of continuous sequence spanning the entire genome of the N and GPC protein of the Chilean ANDV [30]. For mapping experiments, cells were incubated with 10 µg/ml of each individual peptide. As negative controls, corresponding dilutions of DMSO (Sigma) were used, whereas a 1∶100 dilution of PHA (M form, Invitrogen) was used as positive control. After the incubation period, plates were washed and incubated with biotinylated secondary antibodies (IFN-γ: clone M701B, granzyme B: GB11) according to the manufacturer's manual. After incubation with Streptavidine-Alkaline phospahtase (Vector, at 1∶1000 for 2 hours), plates were incubated with NCIP/BPT substrate (BioRad), and analyzed using the ELI.Scan (A.EL.VIS GmbH) analyzing unit. Results were expressed as Spot Forming Units (SFU), representing the numeric difference between specific spots and the spots in the negative control (DMSO).
All FRNT studies were carried out in an approved (C20041018-0267) biosafety level 3 laboratory. Plasma samples from the patient were serially diluted in fourfold increments, mixed with equal volumes of approximately 60 focus forming units (f.f.u.) of a human Chilean virus isolate [65] before incubation on Vero E6 cells, processed and analyzed as described before [66]. The neutralization activity of an antibody was expressed as the highest plasma dilution capable of reducing the number of foci by at least 80%.
Cryopreserved PBMC of four different patients were challenged in vitro for 1.5 hours by a previously determined individual immunogenic peptide (10 µg/ml), the corresponding DMSO dilution or PMA/ionomycin (500/50 ng/ml) in the presence of 1 µg/ml anti-CD49d (clone 9F10) and anti-CD28 (clone CD28.2) (both BD Pharmingen), respectively, and cultured for 4.5 additional hours in the presence of GolgiStop/monensin. Finally, intracellular cytokine staining was performed by fixation, permeabilization of cells and subsequent staining for surface markers and intracellular IFN-γ and TNF-α according to the manufacturer's protocol (BD, California).
2×105 PBMC/well were stimulated with 10 µg/ml of the Gn461–475 in the presence of 10 ng/ml IL-7 and 300 pg/ml IL-12 (R&D systems). On day 2 after setup and every 3–4 days 10 U/ml and 150 µg/ml IL-15 were added to the culture. On day 7 and 14, T-cells were re-stimulated using irradiated (30Gy) PBMC or irradiated autologous (100Gy) B-LCL. On day 21 T-cells were assayed in IFN-γ ELISPOT assays using truncated peptides as indicated.
Was performed on a 4-colour FACSCalibur (Becton Dickinson) or CyAn (Dako) and using either CellQuest (Becton Dickinson®) or Summit 4.0 (Dako) analyzing software. We used the following antibodies (all BD Pharmingen): CD3-FITC (UCHT1), CD4-FITC (SK3), CD45RA-FITC (HI100), CD27-FITC (M-T271), TNF-α-FITC (Mab11), CD45RO-PE (UCHL1), CD28-PE (L293), CD8-PercP (SK1), IFN-γ-APC (B27), CD3-APC (UCHT1). FITC-, PE- and APC- (MOPC-21) as well as PercP- (X40) conjugated mouse IgG1κ were used as isotype controls. Anti-KLRG-1-Alexa488 was kindly provided by Prof H.P. Pircher (University of Freiburg, Germany). Tetramer complexes were custom-synthesized by the NIAID tetramer facility (Gaithersburg, MD) according to the published protocol (http://research.yerkes.emory.edu/tetramer_core/protocol.html), and Gn465–473:tetramers were either APC- or PE-labeled. All other tetramer complexes were APC-labeled. Table 2 shows a summary of HLA-B*3501 tetramer complexes used in the present study.
The patients were genotyped for the HLA loci A, B, DRB1 and DQB1, using the SSP PCR (Sequence Specific Primer–Polymerase Chain Reaction) technique. Low and high resolution SSP kits from Dynal (Oslo, Norway) and Invitrogen Corporation (USA) were used.
For analysis of ELISPOT results for each patient an unpaired Student's t-test was applied in order to calculate significant results as compared to the internal negative (DMSO) control. To be evaluated as positive a sample (that is response to a individual or a pool of up to 40 ANDV-derived 15mer peptides) had to fulfill three criteria: (i) a significant difference (p<0.05) between sample and negative control, (ii) specific SFU had to be superior of 50/106 PBMC, (iii) value had to be above a cut-off, defined as a mean + 2xSD, which was previously established in 20 healthy controls for each peptide pool.
For differences in frequencies of tetramer cell populations an unpaired Student's t-test was applied. We further studied the association of time since ANDV infection until blood sampling with the T-cell responses against N, Gn and Gc as determined by IFN-γ ELISPOT at the time point of blood sampling. A decreasing response with increasing time since infection corresponds to a loss of T-cell memory in time and is reflected by a negative correlation. We reject the null-hypothesis of no association between time and the IFN-γ-ELISPOT response at the two-sided alpha level of 0.05. Due to skewness of the response data, we log-transformed ELISPOT responses and fitted our linear regression models on the log-transformed responses and assessed the model assumptions by inspecting residual values against time.
|
10.1371/journal.pcbi.1003567 | Evolutionary Game Dynamics in Populations with Heterogenous Structures | Evolutionary graph theory is a well established framework for modelling the evolution of social behaviours in structured populations. An emerging consensus in this field is that graphs that exhibit heterogeneity in the number of connections between individuals are more conducive to the spread of cooperative behaviours. In this article we show that such a conclusion largely depends on the individual-level interactions that take place. In particular, averaging payoffs garnered through game interactions rather than accumulating the payoffs can altogether remove the cooperative advantage of heterogeneous graphs while such a difference does not affect the outcome on homogeneous structures. In addition, the rate at which game interactions occur can alter the evolutionary outcome. Less interactions allow heterogeneous graphs to support more cooperation than homogeneous graphs, while higher rates of interactions make homogeneous and heterogeneous graphs virtually indistinguishable in their ability to support cooperation. Most importantly, we show that common measures of evolutionary advantage used in homogeneous populations, such as a comparison of the fixation probability of a rare mutant to that of the resident type, are no longer valid in heterogeneous populations. Heterogeneity causes a bias in where mutations occur in the population which affects the mutant's fixation probability. We derive the appropriate measures for heterogeneous populations that account for this bias.
| Understanding the evolution of cooperation is a persistent challenge to evolutionary theorists. A contemporary take on this subject is to model populations with interactions structured as close as possible to actual social networks. These networks are heterogeneous in the number and type of contact each member has. Our paper demonstrates that the fate of cooperation in such heterogeneous populations critically depends on the rate at which interactions occur and how interactions translate into the fitnesses of the strategies. We also develop theory that allows for an evolutionary analysis in heterogeneous populations. This includes deriving appropriate criteria for evolutionary advantage.
| Population structure has long been known to affect the outcome of an evolutionary process [1]–[4]. Evolutionary graph theory has emerged as a convenient framework for modelling structured populations [4], [5]. Individuals reside on vertices of the graph and the edges define the interaction neighbourhoods.
A variety of processes have been investigated on a number of graph classes. However, few analytical results exist in general, since an arbitrary graph may not exhibit sufficient symmetry to aid calculations. The most general class of graphs for which analytical results are known is the class of homogeneous (vertex-transitive) graphs. Such a graph has the property that for any two vertices and there exists a structure-preserving transformation of such that . It is worth noting that not all regular graphs are homogeneous; an extreme example is the Frucht graph [6], which is regular of degree and has only the trivial symmetry. Intuitively, this class consists of graphs that “look” the same from any vertex. The amount of symmetry in such graphs has allowed for a complete set of analytical results for restricted types of evolutionary processes and weak selection [7]–[9]. Despite the tractability of calculations on homogeneous graphs, natural population structures are seldom homogeneous. Therefore it is important to understand the effects of heterogeneous population structures on evolutionary processes [4], [8], [10] and, in particular, on the evolution of cooperation.
In the simplest case there are two strategic types: cooperators that provide a benefit to their interaction partner at some cost to themselves (), whereas defectors provide neither benefits nor incur costs. This basic setup is known as an instance of the prisoner's dilemma and reflects a conflict of interest because mutual cooperation yields payoff and hence both parties prefer this outcome over mutual defection, which yields a payoff of zero. However, at the same time each party is tempted to defect in order to avoid the costs of cooperation. The temptation of increased benefits for unilateral defection thwarts cooperation – to the detriment of all. This conflict of interest characterizes social dilemmas [11], [12].
More general kinds of interactions between two individuals and two strategic types, and , can be represented in the form of a payoff matrix as in Table 1. The payoffs garnered from these game interactions affect an individual's expected number of offspring by altering their propensity to have offspring (their fitness) or their survival. The expected number of offspring is determined by the fitness of the individuals and some population updating process, which will be made precise in the next section. The offspring produced during the population update have the potential to change the strategy composition of the population. An increase in the abundance of one strategy over a sufficiently large time scale indicates that strategy is favoured by evolution.
It can be shown, for replicator dynamics, for example [13], [14], that any payoff matrix can be reduced to the matrix in Table 1 without loss of generality because adding a constant term to the payoff matrix does not affect the dynamics and multiplying the payoffs by a positive factor merely rescales the time. Therefore we can always shift the payoffs such that --encounters return a payoff of zero and scale all other payoffs such that --encounters yield a payoff of . In the Averaged versus Accumulated Payoffs section, we show that the generality of the matrix in Table 1 extends to other forms of stochastic dynamics in finite populations based on the frequency dependent Moran process [15].
The (additive) prisoner's dilemma introduced before corresponds to the special case with and . Rescaling the payoff matrix in Table 1 by yields the traditional form, Table 2. More generally, the prisoner's dilemma requires and to result in the characteristic conflict of interest outline above. The special case of the additive prisoner's dilemma, Table 2, effectively reduces the game to a single parameter with (and ). Moreover it has the special property that when an individual changes its strategy, the payoff gain (or loss) is the same, regardless of the opponents' strategy – the so-called equal-gains-from-switching property [16].
In the absence of structure, cooperators dwindle and disappear in the prisoner's dilemma. In contrast, structured populations enable cooperators to form clusters, which ensures that cooperators more frequently interact with other cooperators than they would with random interactions [17], [18]. Such assortment between cooperators is essential for the survival of cooperation [19].
In heterogeneous graphs not all vertices have the same number of connections and hence the fitnesses of individuals may be based on different numbers of interactions. Because of this, some vertices are more advantageous to occupy than others. However, which sites are favourable depends on the type of population dynamics. In particular, for the Moran process in structured populations it is important to distinguish between birth-death and death-birth updating [10], [20], [21], i.e. whether first an individual is randomly selected for reproduction with a probability proportional to its fitness and then the clonal offspring replaces a (uniformly) randomly selected neighbour – or, if first an individual is selected at random to die and then the vacant site is repopulated with the offspring of a neighbouring individual with a probability proportional to its fitness. Even in homogenous populations the sequence of events is of crucial importance but becomes even more pronounced in heterogenous structures [10], [20].
In order to illustrate that the population dynamics may bestow an advantage to individuals occupying certain sites in a heterogeneous population, consider neutral evolution, where game payoffs do not affect the evolutionary process and all individuals have the same fitness. For birth-death updating every individual is chosen to reproduce with the same probability but neighbours of individuals with few connections are replaced more frequently. Hence vertices with fewer neighbours are more favourable than those with many connections. Conversely, for death-birth updating every individual has the same expected life time but highly-connected individuals, or, hubs, get more frequently a chance to produce offspring, since one of their many neighbours dies, and are thus more favourable than vertices with few neighbours [21]–[23]. A simple example of this is a -line graph, one central vertex connected to two end vertices. In the birth-death process, the central vertex is replaced with probability , while either end vertex is replaced with probability , while in the death-birth process, the central vertex replaces either end vertex with probability and either end replaces the centre with probability [21]. The upshot is, even though the fitness of all individuals is the same, the effective number of offspring produced depends on the dynamics as well as an individual's location in the population.
The intrinsic advantage of some vertices over others can be further enhanced through game interactions leading to differences in fitness that depend on an individual's strategy as well as its position on the graph. For example, a cooperator occupying a favourable vertex can more easily establish a cluster of cooperators, which creates a positive feedback through mutual increases in fitness. Conversely, a favourable vertex also supports the formation of a cluster of defectors but this results in a negative feedback and lowers the fitness of the defector in the favourable vertex. The fact that heterogeneity can promote cooperation was first observed for the prisoner's dilemma and snowdrift games [24], [25] and has subsequently been confirmed for public goods games [26], [27]. However, the detailed effects not only crucially depend on the dynamics but also on how fitnesses are determined. For example, heterogenous population structures favour cooperation if payoffs from game interactions are accumulated but that advantage disappears if payoffs are averaged [28]–[30].
The effects of population structure on the outcome of evolutionary games is sensitive to a number of factors: population dynamics [10], [20], [31], translation of payoffs into fitness [28], [30], [32]–[35], the diversity of players [27], [34], [36], and the type of game played – for example, spatial structure tends to support cooperation in the prisoner's dilemma but conversely, in the snowdrift game, spatial structure may be detrimental [37]. Macroscopic features of the evolutionary process on the level of the population, such as frequency and distribution of cooperators, are determined by microscopic processes on the level of individuals. In the current article, we discuss some of these microscopic processes, such as averaging and accumulating payoffs, and the rate at which interactions take place, and illustrate how they affect an evolutionary outcome. Crucially, we also illustrate that the conditions for evolutionary advantage commonly found in the literature are not applicable to evolution in finite, heterogeneous populations. We modify these conditions and develop a general framework to determine evolutionary advantage in finite, heterogeneous populations.
The manuscript is organized as follows. Sections “Accumulated and Averaged Payoffs” and “Criteria for Evolutionary Success” create a critical synthesis of the existing literature concerning evolution on heterogeneous graphs. In these sections we extend existing results to general games and focus on an imitation process. We also discuss the inapplicability of approaches used in homogeneous populations and present our novel conditions for evolutionary success in heterogeneous populations. Interspersed in these sections are new observations and results that aid in establishing a consistent framework on which we base further novel results presented in the section “Stochastic Interactions and Updates”.
In heterogenous population structures individuals naturally engage in different numbers of interactions. This renders comparisons of the performances of individuals more challenging. One natural approach is to simply accumulate the game payoffs. This clearly puts hubs with many neighbours in a strong position as scoring many times even a small payoff may still exceed few large payoffs. To avoid this bias in favour of hubs, game payoffs can be averaged. Interestingly, these two approaches not only play a decisive role for the evolutionary outcome but also entail important biological implications. In this section we extend previous work on payoff accounting [29] to general games and provide a thorough discussion of why different payoff accounting schemes can result in markedly different evolutionary outcomes.
Consider two different ways to translate the total, accumulated payoffs of an individual into its fitness :(1a)(1b)where denotes the strength of selection and is the number of interactions experienced by . The limit recovers the neutral process, where selection does not act. Note that the payoff matrix in Table 1 can still be used without loss of generality because adding a constant merely changes the (arbitrary) baseline fitness from to and multiplying the payoffs by is identical to simply changing the selection strength to .
The exponential form of fitness in the above equations is mathematically convenient since it guarantees that the fitness is always positive, irrespective of the strength of selection and payoff values. It is worth noting that if the strength of selection is weak, that is, , then(2a)(2b)which represents another common form for fitness found in the literature [8].
In order to determine the evolutionary success of a strategic type in a finite population we consider three fixation probabilities: and . The first, , indicates the probability that a single type in an otherwise population goes on to supplant all s, while the second, , refers to the probability of the converse process where a single type takes over a population of types. These fixation probabilities are important whenever mutations can arise in the population during reproduction or through errors in imitating the strategies of others. The last probability, , denotes the fixation probability of the neutral process, which is defined as the dynamic in a population with vanishing selection, . In such a case the game payoffs do not matter and everyone has the same fitness. Based on these fixation probabilities two distinct and complementary criteria are traditionally used to measure evolutionary success [15], [20]:
The above conditions (11) and (14) are based on the implicit assumption of homogenous populations or averaged payoffs and randomly placed mutants. In the present context of heterogenous populations and with mutants explicitly arising through errors in reproduction or imitation, both conditions require further scrutiny and appropriate adjustments.
The first condition implicitly assumes that an mutant appears in a monomorphic population with the same probability as a mutant in a monomorphic population. However, in heterogenous populations with accumulated payoffs this is not necessarily the case. Even in monomorphic states hubs may have a higher fitness and hence are more readily imitated, or reproduce more frequently, than low degree vertices. This can result in a bias of the rates at which and mutants arise. Thus, the condition for evolutionary advantage, Eq. (11), must read(15)In general, and depend on the population structure as well as the payoffs and their accounting. The star structure serves as an illustrative example in the next section.
Similarly, the second condition also needs to be made more explicit. In general, to determine whether a mutation is beneficial its fixation probability should exceed the probability that in the corresponding monomorphic population one particular individual eventually establishes as the common ancestor of the entire population. We denote these monomorphic fixation probabilities by , and , respectively. Thus, the second condition, Eq. (14), should be interpreted as(16a)(16b)i.e. that the fixation probability of a single (or ) mutant in a () population exceeds that of one () individual turning into the common ancestor of the entire population.
If mutations occur during an updating event, then in heterogeneous populations mutants occur more frequently in some vertices than in others. For our imitation process, high degree vertices serve more often as models than low degree vertices and hence the mutation is likely to occur in neighbours of high degree vertices. Note that this is different from placing a mutant on a vertex chosen uniformly at random from all vertices [47]. A randomly placed neutral mutant fixates, on average, with a probability corresponding to the inverse of the population size. This is not necessarily the case if neutral mutants arise in reproductive events or errors in imitating or adopting other strategies. In fact, the distinction between and is only required on heterogenous graphs with accumulated payoffs and non-random locations of mutants. In all other situations the (average) monomorphic fixation probabilities are the same and equal to , where is the population size.
In summary, due to the fitness differences in a monomorphic population with accumulated payoffs the turnover is accelerated and more strategy updates take place and hence more errors occur than in the corresponding monomorphic population. This means that, on average, mutant s more frequently attempt to invade an population than vice versa. Overall, this leads to new conditions for evolutionary success in heterogeneous populations, summarized as follows. Type (i) has an evolutionary advantage or is favoured if where is the probability a mutant arises in an all- population (and vice-versa), and, is beneficial if , where is the probability a single individual goes on to become the common ancestor in an all- population. Analogous conditions hold for a mutant type. We apply these novel conditions to an example found in the literature [47], the star graph.
As we have seen, when payoffs are averaged, members of a heterogeneous population are possibly playing different games, while if they are accumulated, all individuals play the same game. Therefore, only accumulating payoffs allows for meaningful comparisons of different heterogeneous population structures. A common simplifying assumption is that each individual interacts once with all its neighbours, see Figure 3. For heterogeneous populations this assumption means that those individuals residing on higher-degree vertices are interacting with their neighbours at a higher rate than those on lower-degree vertices. This leads to a separation of time scales, where interactions occur on a much faster time scale than strategy updates.
Realistically, all social interactions require a finite amount of time and hence the number of interactions per unit time is limited. This constraint already affects the evolutionary process in unstructured populations [48] but becomes particularly important in heterogenous networks where, for example, in scale-free networks some vertices entertain neighbourhood sizes that are orders of magnitude larger than that of other vertices. For those hubs it may not be possible to engage in interactions with all neighbours between subsequent updates of their strategy or the strategies of one of their neighbours. In order to investigate this we need to abandon the separation of the timescales for interactions and strategy updates.
A unified time scale on which interactions and strategy updates occur can be introduced as a stochastic process where a randomly chosen individual initiates an interaction with probability with a random neighbour and reassesses its strategy with probability by comparing its payoff to that of a random neighbour according to Eq. (9). Interactions alter the payoffs of both individuals (and hence their fitnesses, , see Eq. (1a)) according to the game matrix in Table 1. If individual adopts the strategy of its neighbour, then its payoff (and interaction count) is reset to zero, , regardless of whether the imitation had resulted in an actual change of strategy. Simulation results for various are shown in Figure 6.
For small few interactions occur between strategy updates and in the limit neutral evolution is recovered because no interactions occur. Conversely, in the limit many interactions occur between strategy updates, which allows individuals to garner large payoffs as well as build up large payoff differences. The average number of interactions initiated by any individual between subsequent reassessments of the strategy is , the relative ratio of the time scales of game interactions versus strategy updates. However, the distribution of the number of interactions is biased: individuals with a large number of interactions tend to score high payoffs and hence are less likely to imitate a neighbours' strategy, which in turn results in a further increase of interactions. On heterogenous graphs and scale-free networks, in particular, this bias is built in by the underlying structure because highly connected hubs engage, on average, in a much larger number of interactions than vertices with few neighbours. Moreover, hubs are more likely to serve as models when neighbours are reassessing their strategy – simply because hubs have many neighbours. Thus, hubs are not only more resilient to change but also have a stronger influence on their neighbourhood. When this ratio begins to get large, interactions dominate strategy updates and the resulting game dynamics on heterogeneous and homogeneous graphs becomes indistinguishable.
Interestingly, a similar bias in interaction numbers spontaneously emerges on homogenous graphs, lattices in particular. Since all vertices have the same number of neighbours, no vertices are predisposed to achieve more interactions than others but some inequalities in interaction numbers occur simply based on stochastic fluctuations. As above, those vertices that happen to engage in more interactions tend to have higher payoffs and hence are less likely to imitate their neighbours and keep aggregating payoffs. This positive feedback between interaction count and resilience to change spontaneously introduces another form of heterogeneity, which becomes increasingly pronounced for larger . In fact, for large it rivals the structurally imposed heterogeneity of scale-free networks, see Figure 7.
Regardless of the structure, the positive feedback between payoff aggregation and the diminishing chances to change strategy (and hence reset payoffs) means that a small set of nodes forms an almost static backdrop of the dynamics and hence has a considerable effect on the evolutionary process. This set is a random selection on homogenous structures and consists of the hubs on heterogenous structures. As a consequence, the initial configuration of the population has long lasting effects on the abundance of strategies.
A more detailed view on the effects of on the evolutionary process is provided by restricting attention to the prisoner's dilemma and additive payoffs, c.f. Table 2. This can be accomplished by setting with . The equilibrium levels of cooperation in the -plane are shown in Figure 8 for lattices and scale-free networks.
Altering the relative rates of interactions versus strategy updates has interesting effects on the evolutionary outcome. For lower rates of interaction (), scale-free networks outperform lattices in their ability to promote cooperation. As interaction rates increase and strategy updates become more rare (), scale-free networks and lattices become virtually indistinguishable in their ability to support cooperation. For both lattices and scale-free networks an optimal ratio between strategy updates and interactions exist: for lattices this is roughly , suggesting that lattices support the greatest amount of cooperators when interactions occur at the same rate as strategy updates, whereas for scale-free networks the optimum lies around , which suggests that scale-free networks provide the strongest support for cooperation if there are roughly three updates per interaction.
Evolutionary dynamics in heterogenous populations, scale-free networks in particular, have attracted considerable attention over recent years. Somewhat surprisingly, the underlying microscopic processes and their implications for the macroscopic dynamics and the corresponding biological interpretations have received little attention.
Here we have shown that established criteria to measure success in evolutionary processes make different kinds of implicit assumptions that do not hold in general for heterogenous structures. Instead, for such structures it becomes imperative to reconsider, revise and generalize these criteria, which was done in the Criteria for Evolutionary Success section. If errors arise in imitating the strategic type of other individuals, or mutations occur during reproduction, then mutations are more likely to arise in some locations than in others. For example, on the star graph mutants likely occur in the leaf nodes for birth-death updating and imitation processes but in the hub for death-birth processes. Moreover, in heterogenous populations the fixation probabilities generally depend on the initial location of the mutant and hence even the fixation probability of a neutral mutant may no longer simply be the reciprocal of the population size but rather intricately depend on the population structure.
Another crucial determinant of the evolutionary dynamics in heterogenous populations is the aggregation of payoffs from interactions between individuals. Individuals on vertices with a higher (lower) degree expect to have more (fewer) interactions than on average. Even though the choice between averaging or accumulating payoffs may seem innocuous, it has far reaching consequences. Previous authors [29] have found that averaging payoffs in a prisoner's dilemma game on a scale-free network eliminates such a network's ability to promote cooperation as observed in earlier studies [24], [25], [27], [49]. We have extended this result to general games and provide a detailed rationale for this phenomenon which is summarized as follows. If payoffs are accumulated, some individuals are capable of accruing more payoffs than others strictly by virtue of them having more potential partners. Averaging payoffs removes the ability of hubs to accrue greater payoffs, but simultaneously makes it difficult to compare results for different population structures (e.g. lattices versus scale-free networks) even if their average degrees are the same because the type of game played depends on the location in the graph. Hence, accumulating payoffs seems a more natural choice to compare evolutionary outcomes based on different population structures because it ensures that everyone engages in the same game. However, if we assume all interactions are realised then those individuals with more neighbours interact at a much greater rate than those with less.
In order to investigate the disparity in the number of interactions on the success of strategies on heterogenous graphs we introduced the time-scale parameter , which determines the probability that an interaction or a strategy update occurs. When increasing the rate of strategy updates (small ), heterogeneous graphs are able to support higher levels of cooperation than lattices. Conversely, increasing the rate of interactions (large ) results in small differences between lattices and scale-free networks; both support roughly the same levels of cooperation. For imitation processes, individuals with high payoffs are unlikely to change their strategies and hence are likely to keep accumulating more payoffs. On scale-free networks, hubs are predestined to become such high performing individuals but on lattices they spontaneously emerge, triggered by stochastic fluctuation in the interaction count and driven by the positive feedback between increasing payoffs and increasing resilience to changing strategies (and hence to resetting payoffs).
For intermediate an optimum increase in the level of cooperation is found: lattices support cooperation most efficiently if a balance is struck between interactions and strategy updates (), whereas scale-free networks work most efficiently if slightly more updates occur (). For lattices a related observation was reported for noise in the updating process [50]. If the noise is large, updating is random but if it is small the game payoffs become essential. Interestingly, cooperation is most abundant for intermediate levels of noise – which is similar to having some but not too many interactions between strategy updates.
Previous work has found that heterogeneous graphs support coordination of strategies, where all individuals are inclined to adopt the same strategy, while homogeneous graphs support co-existence [51], [52]. The time scale parameter introduced in the Stochastic Interactions and Updates section seems to aid in promoting coexistence in both types of graphs, based on the large green region in Figures 3, 6, and 8. Exactly how the time scale parameter promotes coexistence is a topic worthy of further investigation.
Naturally there is no correct way of modelling the updating of the population or the aggregation of payoffs but, as so often, the devil is in the detail and implicit assumptions originating in traditional, homogenous models may be misleading or have unexpected consequences in more general, heterogenous populations.
In [47], the authors calculate expressions for the probability that a single mutant fixes on a star graph. These expressions are in terms of state transition probabilities. Denote by the transition probability from a state with individuals on the leaves and an individual on the hub to a state with individuals on the leaves and a on the hub. With this notation, the fixation probability of a single on a leaf vertex is(27)and for a single on the hub,(28)where, in both cases,(29)For the imitation process defined by Eq. 9 and accumulated payoffs we have(30a)(30b)(30c)(30d)and for averaged payoffs,(31a)(31b)(31c)(31d)These are incorporated into the Eqs. (27) and (28) to yield the fixation probabilities and . The fixation probabilities and are obtained in a similar way. The averages and are then calculated using Eqs. (17), (18), and (19). Finally, a first-order approximation in is found for the above. For example,(32a)The other fixation probabilities are found in a similar way:(32b)(32c)(32d)Assuming , and employing the appropriate condition for evolutionary advantage, yields Eqs. (25a–25d) in the main text.
|
10.1371/journal.pntd.0006419 | Geospatial-temporal distribution of Tegumentary Leishmaniasis in Colombia (2007–2016) | Tegumentary Leishmaniasis (TL) is a neglected disease with worldwide distribution and considered a public health problem, especially in Latin America. In Colombia, the governmental epidemiological surveillance system (SIVIGILA) is responsible for collecting information on the presentation of cases of TL from each of the municipalities and departments. In absence of a study compiling and analyzing currently available metadata of TL in Colombia, this study describes the geospatial-temporal distribution of TL and identifies the regions of the country on which prevention measures should be established in order to control the disease.
This is an exploratory descriptive analysis of the distribution of TL in Colombia. Information was collected on new cases of the disease during the years 2007–2016 from the Colombian reporting system (SIVIGILA). Incidence calculations were made based on population estimates by departments and biogeographical regions. Time evolution is shown in biennial maps. A 10-year series was analyzed, showing that the Amazon region is the most affected in terms of incidence, while the Andean region has the highest number of cases with a high variability among the departments that make it up. In those departments where there is a greater reported diversity of vector species, a large number of cases was observed.
Transmission dynamics of TL in Colombia in the past 10 years have been variable, with a greater concentration of cases in the central and southern departments. The present study contributes to improve the understanding of the patterns of distribution of TL in Colombia and can be a basis for future studies of impact evaluation of Health policies in the country and the region.
| Colombia is among the countries with the highest number of Tegumentary Leishmaniasis cases worldwide. Despite public health efforts, and the existence of a national epidemiological surveillance system, articulated with the regional SisLeish system, the trends followed by the disease’s prevalence and incidence have not been explored. This work presents a retrospective analysis of Tegumentary Leishmaniasis in Colombia between the years 2007 and 2016, depicting the spatiotemporal distribution of the disease throughout the Colombian territory. The results show a sustained transmission rate, especially in the Amazon and Orinoco regions, and high intraregional variability that could be attributed to each region’s characteristics (vector species diversity and low adherence to preventive measures). The Amazon region is of particular importance due to its direct connection to other endemic countries such as Peru and Brazil. This connection highlights the need to prioritize the implementation of control and prevention strategies in this region. The surveillance system displays multiple flaws, which contribute to under-reporting and have a direct impact on the regional system. Strengthening the system should become a key public health objective in the country. The results of this research can contribute to the understanding of the transmission dynamics of Leishmaniasis in Colombia, as well as to the bolstering of public policy efforts in towards prevention and control of the disease.
| Leishmaniasis, considered one of the most neglected diseases worldwide, is a grouping of parasitic pathologies caused by protozoans of the Leishmania genus, and transmitted to humans primarily by insects of the Psychodidae family [1]. Leishmaniasis displays a wide range of manifestations and clinical forms of which the most common, both worldwide and in the American continent, is Cutaneous Leishmaniasis (CL) characterized by ulcerative skin lesions [1–3]. CL is considered a widespread public health issue with an elevated prevalence, considered endemic in 98 countries, adding up to a population at risk of over 350 million people. Meanwhile, Mucocutaneous Leishmaniasis (ML) affects a smaller percentage of the population, producing disfiguring lesions which compromise both the oral and nasal mucosae [1, 2, 4]. In the clinical practice, CL and ML are known as Tegumentary Leishmaniasis (TL). The American continent constitutes a special scenario for TL, given that the biological and clinical complexity of the disease is compounded by the presence of sociodemographic variability, geographical diversity, and the presence of internal armed conflicts in different countries. These features facilitate the parasite’s spread in the Americas, now present in 20 countries, of which 18 are endemic with a yearly average of 56,262 cases between 2001 and 2015. Seventy percent of the cases reported in 2015 were found in just three countries: Brazil, Colombia and Peru [3, 5, 6].
Colombia is one of the 6 countries that account for over two thirds of TL cases worldwide [7]. This clustering of the disease has been associated with various circumstances, including the presence of armed conflict, which leads to internal displacement and increases the probability of contact between vectors, reservoirs and hosts. This occurs in both the civilian and military populations. Additionally, social inequality, the absence of preventive measures in rural areas, deforestation, and geographical diversity allow the circulation of a wide array of species of both vector and parasite in up to 30 of the 32 departments in the country [3, 6, 8]. It is worth noting that Colombia is the country with the greatest number of circulating species in the world, 10 in total (L. panamensis, L. braziliensis, L. guyanensis, L. infantum chagasi, L. mexicana, L. lainsoni, L. amazonensis, L. colombiensis, L. equatoriensis, L. naiffi), and that the most prevalent species exhibit high genetic diversity [9–11]. This set of conditions conspires to produce an incidence of 33.6 per 100,000 inhabitants in 2015, leading to the categorization of the country as one of Intense Transmission by the Panamerican Health Organization (PAHO) according to its Leishmaniasis Compound Index [3]. Considering the particular characteristics found in the country, it is important that robust epidemiological surveillance models be implemented, which can detect and categorize cases not only by geographical location, but also by species. These models would allow for the execution of prevention and control strategies in order to gradually reduce the burden of disease in the country.
Currently, public health authorities in Colombia collect data on various diseases by means of a system called SIVIGILA, which gathers data on a variety of illnesses that are of public health concern [12]. In spite of the existence of this system, it is presumed that inadequate reporting is the norm in rural areas with no direct access to information systems. Likewise, and in response to the challenges around surveillance and control of leishmaniasis throughout the continent, the Regional Leishmaniasis program of the PAHO, along with representatives from six endemic countries, created a regional information system called SisLeish. The program gathers data from the epidemiological surveillance systems in each country, continually evaluating the disease’s distribution and producing regional indicators [13].
Despite the existence of information systems which compile data on TL, the geospatial and temporal distribution of the disease throughout the Colombian territory has not yet been studied. The trends and changes in TL’s distribution for recent years remains unknown. Therefore, this study aimed to analyze the spatiotemporal distribution of TL in Colombia between the years 2007 and 2016, providing a quantitative characterization of the disease’s behavior and spread in the country. Using the current available information, we also constructed a descriptive model which estimates the standardized incidence ratio in each of the departments and regions in each time period. This analysis allows us to formulate hypotheses regarding the areas most in need of attention from public health authorities.
We report a geospatial analysis of TL data in Colombia. The data were readily obtained from existing public access databases (SIVIGILA). Hence, there are no specific ethical considerations.
The governmental surveillance system in Colombia is carried out by the National Public Health Surveillance System (SIVIGILA), regulated in 2006 by the office of the President of the Republic of Colombia, which is tasked with the information collection from Primary Data Generating Units (UPGD). These units correspond to Institutions Providing medical Services (IPS) in which cases of the various diseases that require mandatory reporting (Common source and transmissible events, Non transmissible prevalent diseases and Avoidable mortality events) are detected [12, 14]. Mandatory reporting events are logged in SIVIGILA’s website, and the data for TL was used to construct summary tables by municipality and department for every year from 2007 to 2016. These tables were then complimented with demographic data from the national statistical service (DANE) presented in the “Estimación y proyección de población nacional, departamental y municipal total por área 1985–2020” report [15]. The data was organized with the Microsoft Excel software in its 2016 version. The data for each geographical unit was collated both annually and biennially. Additionally, departments were grouped by geographical regions (Andean, Amazon, Caribbean, Pacific, Orinoco, and Insular) (See S1 Text for more information about geographical regions). Then, inferences were conducted regarding the geospatial distribution. The data collection and analysis process are depicted in Fig 1.
Descriptive statistics were used to summarize the data. Preliminarily, the crude incidence (Io) was calculated by dividing the number of cases by the projected population for each of the scales and each one of the years considered. The data were then grouped, and Io was calculated for biennial periods. Io by department was split in quartiles and graphed, using a different color for each quartile. Maps depicting incidence quartile for each department were generated from an aggregation of data from biennial periods from 2007 to 2016. Administrative polygons for Colombia were obtained from the Global Administrative Areas database, managed by the University of California Berkeley [16]. Regional polygons were constructed using R version 1.0.136 [17], and packages sp and rgdal [18, 19]. The maps were generated using QGIS version 2.18.7 Las Palmas [20]. Standardized Incidence Ratios (SIR) were calculated from the total number of cases in Colombia for the period considered using Empirical Bayes smoothing method described by Clayton and Kaldor [21].
In order to visualize the detail of the annual tendencies for each of the studied regions, a time series graphic was generated and smoothed by means of Lowess (Locally-Weighted Scatterplot Smoother, 0.3 degrees, 2 steps) using Minitab v. 18. Additionally to visualize cases and vector distributions, total case data was overlaid with vector species data reported by Ferro et al. [8] on which accumulated cases in all study periods were added. The map was constructed displaying the number of cases detected in each department for all the periods, with circles centered on each department’s centroid and radii proportional to the number of cases.
Generally, it was observed that TL has had a stable behavior in the country in the past few years, with a stable incidence between 9.99 and 34.17 cases per 100,000 inhabitants. Closer examination of TL behavior by geographical region shows that the Amazon region exhibits the highest incidence rates for all observed periods. Incidence in this region varies from 76.75 to 240.93 cases per 100,000 inhabitants. The region with the lowest incidence is the Insular region, in which only one case occurred during the periods examined. Likewise, the Caribbean region displayed incidences between 5.11 and 13.64 cases per 100,000 inhabitants. The Andean region had the highest number of cases for all periods, reaching a total of 12,847 cases between 2009 and 2010. The Orinoco region displayed the lowest case number, with 1031 cases for the 2007–2008 period (S1 Table).
The behavior of TL by departments was highly variable. The Antioquia department presented a highest number of cases in all the time periods, registering 20,951 in total across all years studied, with an average between 1000 and 3000 cases each year. The Meta department displayed the second highest number of TL cases in the country, reaching a peak of 4019 cases in 2009, this considered as the highest number of cases in a single department in any of the studied periods. The departments with the fewest cases were Bogotá with 37 cases and San Andrés with 1 case in all the periods (S2 Table).
Fig 2 shows intraregional and interdepartmental variations in annual incidence. Across the study periods, high Io’s (>51.6 cases per 100,000 inhabitants) occurred in departments of the Amazon and Orinoco regions. Despite presenting high Io’s at the starting and ending periods, the departments in the Andean region presented mostly high to intermediate incidences (15.4–51.6 cases per 100,000 inhabitants). Departments in the Pacific and Caribbean regions displayed intermediate Io’s in 8 out of the 10 years considered. The Caribbean region displayed a greater frequency of low Io’s (<3.92 cases per 100,000 inhabitants) in later time points. Departmentally, Putumayo, Guaviare, Guainía, Caquetá, Vaupés, Meta, Chocó, and Vichada showed high Io’s for most periods considered (Fig 2).
Maps in Fig 3 display biennial incidence divided in quartiles across departments in a choropleth format. Departments in the Amazon and Orinoco regions overall display the highest incidences, with values above the third quartile. The segmentation of the data by quartiles shows that the largest values are extreme, with a much larger range in the last group of incidences. Some departments in the Andean region, such as Antioquia, Santander, and Tolima, showed intermediate to high incidences for all time periods, while departments such as Valle and Quindío showed intermediate to low incidences.
SIR showed high incidences in departments belonging to the Orinoco and Amazon regions in all the time periods. The Caribbean region displays the lowest SIRs across all time periods. Centrally-located departments show stable incidences across time periods, with the notable exceptions of Norte de Santander and Tolima, which trend towards higher values (Fig 4).
Superposition of vector species and leishmaniasis cases shows that departments with greater reported vector diversity tend to display an elevated number of cases, particularly in the departments belonging to the Andean and Orinoco regions (Fig 5). Lutzomyia gomezi and Psychodopygus panamensis are present in 21 departments each, considered the most widely spread species in the country. They are followed by Psathyromyia shannoni and Nyssomyia trapidoi, which are present in 16 and 11 departments, respectively. Antioquia showed the highest number of vector species, 11 overall, followed by Meta, Caldas, Boyacá, and Amazonas, with 9 species each.
Yearly tendency of TL varied according to the region under study, showing a stable pattern in the Andean, Caribbean and Insular regions. The Amazon, Orinoco, and Pacific regions displayed a progressive increase in incidence up to the year 2010, followed by a stable behavior, mainly in the Amazon and Orinoco regions, and finally decreasing from the year 2014 to 2016. Case behavior in relation to Io backs the likely influence of the population base in the annual tendency of the event (Fig 6).
TL remains a public health issue in Colombia and worldwide. This study, shows that the disease trends have been stable in the past decade, particularly in the Caribbean and Insular regions (Figs 2–4 and 6). Both regions are adjacent to the Caribbean sea, with high average temperatures, and frequent droughts, which inhibit vector spread [8]. The Andean region showed an increasing tendency, which might be due to a variety of factors, including outbreaks, and changes which facilitate the infection’s permanence, urbanization, and deforestation [22–26]. The Orinoco, Amazon and Pacific regions displayed an increasing tendency during the first years considered, with a decrease in incidence in the later years (Fig 6). The initial tendency coincides with the inception of the epidemiological vigilance system, which could have altered the behavior artificially. Additionally, the geographical characteristics in these regions allow for the optimal development of vectors, thus facilitating the disease’s transmission and permanence [8].
On an intraregional level, we observed highly fluctuating incidences, mainly in the inner Andean Region and the Orinoco region, where low incidences in departments such as Huila, Quindío, Arauca and Casanare can be contrasted with the high incidences found in Santander, Tolima, Meta, and Vichada (Figs 2–4). These large differences among departments belonging to the same regions, some of which are contiguous, shows that individual departmental characteristics modify disease burden. The Andean region, which gathers over 50% of the country’s population, shows great socioeconomic diversity and highly variable geography [27]. These features have been studied extensively in previous investigations, and have demonstrated a strong influence on disease presentation, particularly in those region in which the biogeographical and social conditions allow not only for vector development, but also for its propagation due to deforestation and urbanization [28, 29]. Another important aspect is the degree of sophistication of the individual departmental reporting systems, which has a direct impact not only on diagnosis, but also on the disease’s treatment. This can ultimately result in an increase in risk to all the population, since some of the population inhabits distant areas where it is difficult to diagnose and treat them, thus becoming reservoirs and perpetuating the disease’s cycle [28].
On the other hand, the Amazon region demands special attention, not only from Colombia, but also from other countries in the region, due to its connections with countries as Brazil and Peru, which allow for the continued transit of potentially infected humans, and the continuity across borders of vector habitats [3]. This was one of the most affected areas by the disease, with high incidences across all study periods. This can be attributed to the wealth of vector species and of mammals that serve as disease reservoirs, which helps maintain the epidemiological circuit and, in turn, transmission rates [8]. The population in this region is relatively low, when compared with other regions with similar transmission rates (S1 Table), due in part to the large percentage of the territory occupied by tropical rainforest. The opposite situation is seen in the Andean region, which is made up of the 11 departments that make up the majority of the country’s population, in which the large number of cases does not result in elevated incidences (Figs 1 and 5; S1 Table). Some of the factors that may have produced these population concentrations are the armed conflict that was active during the studied periods, and a lack of public policies that favored the agricultural economic sector, which combined led to a reduction in rural population growth, mass displacement from rural to urban areas and an increase in internal migration [30, 31]. It is worth noting that some departments in the Orinoco and Amazon regions, have shown high transmission rates and prevalence in independent studies [9, 32]. These same departments report fewer cases to the central surveillance system than departments in other regions with more robust healthcare infrastructure. These same areas are often the most accessible to research groups, which strengthens the surveillance network, particularly in the Andean region.
Likewise, the great diversity of circulating Leishmania species in Colombia plays a fundamental role in the permanence and propagation of the disease, given that neither epidemiological surveillance, nor treatment, take this variable into account. This assumes that all species are equally susceptible to treatment, which has been refuted in multiple studies [4, 11, 33–35]. The presence of some vectors in over half of the departments in Colombia (Fig 5) shows the importance of this factor in the perpetuation of the epidemiological cycle. This underlines the need to address this issue in order to control Leishmaniasis in the country [8, 36]. Likewise, the superposition of the vector species with the TL cases is observed most often in zones with high incidences, reinforcing the importance of the vector’s role (Fig 6). Studies carried out on vectors in Mexico and Spain have shown similar behaviors to those herein presented and have highlighted the need to analyze not only the distributions of the species, but also the bio-climatic and sociodemographic factors as transcendental elements in the transmission of the disease [37, 38]. Likewise, studies conducted in Brazil have associated factors such as urbanization and poor sanitary conditions with the adaptation and maintenance of the vector life cycle [39, 40]. This highlights the importance of complementing the epidemiological surveillance with the phlebotominae analysis, more when Ferro et al., shows through predictive analysis the possible dispersion of each of the species of the vectors involved in the transmission, where it is evident that almost throughout of all the Colombian territory and especially in bordering zones, there can be presence of diverse species of sandfly vectors [8]. The above, denotes the priority that the strengthening of the Epidemiological Surveillance System should have, not only at a national level, but also at a regional level, since the joint analysis of the different elements that make up the epidemiological circuit of the disease will allow not only a better understanding of the system, but also establish more efficient prevention and control measures.
Despite the multiple efforts undertaken by governmental agencies to control the disease, many characteristics of the surveillance system must be expanded on to be able to formulate public policies that lead to a reduction in case presentation. Retrospective studies play a significant role in this process, since trends can be identified, and then prevention and control measures can be formulated. Of note, studies in Brazil and Iran have evaluated the spatial distribution and epidemiological characteristics of leishmaniasis, becoming the starting point for a more effective allocation of resources, guiding national and regional public health policy [41, 42]. Despite the existence of a reporting system, underreporting is common in the most remote rural areas. These areas coincide with areas highly affected by armed conflict, which makes it challenging to estimate the real number of cases with certainty. It is expected that the signing of the peace accords in Colombia will lead to a more sustained and widespread governmental presence in these areas, allowing for more reliable information gathering in both the civilian population and in the population belonging to illegal armed groups. The weaknesses of the national epidemiological surveillance system have direct repercussions over the Regional SisLeish system [13], which collates data from all countries in the Americas in order to guide the efforts of the PAHO. Non-adherence to the surveillance system and delays in reporting have been detected in the SisLeish system, and can be traced back directly to local systems [13]. The data presented here differ significantly from those reported by Maia-Elkhoury et al. [43] with regards to incidence. This is due to differences in the definition of the population at risk. This study considered the entire population of the department as being at risk, instead of just accounting for the population of municipalities in which cases were reported. Given that underreporting is a major hurdle to analysis, we considered that the totality of the population should be considered at risk, instead of the population of the few municipalities which report cases to the national system. The case of the Amazonas department is illustrative, where some of the municipalities report significant numbers of cases for some years, and none for others (Fig 2). The population cannot be considered as not having been at risk during the years when no cases were reported. Differences may also have arisen due to base populations being different. The base populations considered in our analysis were extrapolated from census projections of 2005, since current population dynamics are unknown.
The SIR model allows us to highlight departments that are at particularly high risk, and which merit increased attention from the authorities. Although more sophisticated modeling techniques were attempted, high interdepartmental variability in the occurrence of cases resulted in inadequate fits. Geo-political changes introduced further variability, which could correspond to a high rate of internal migration. It should be noted that the highest rates of internal migration during the period studied were found in the departments of Boyacá, Cundinamarca, Meta, and the Orinoco and Amazon regions, which would reflect the displacement of potentially infected populations to urban areas [29, 44–46]. It is important to note that variables that could aid in modelling, like deforestation and socioeconomic characteristics, are not reported directly by SIVIGILA, and can have dramatic effects on disease occurrence [25, 38].
Disease dynamics may change significantly in the future due to historical transitions that are underway. The resolution of the armed conflict with the FARC guerrilla, and the establishment of rural transitional zones, will likely lead to many former combatants to be detected by the epidemiological surveillance system. Public health agencies should prepare not only for a significant increase in the number of resources required for diagnosis and treatment, but also for the redistribution of TL cases in the country, which will require changes in resource allocation. This abrupt change may lead to case number spikes in some areas previously considered as low transmission rate areas, which do not correspond to outbreaks, but which will be detected as such by the current system. In spite of SIVIGILA’s shortcomings, it is worthwhile noting that the system must be strengthened and extended, since the data collected by the system are the basis for the analysis of the country’s health situation, giving birth to all of the national public health initiatives.
In conclusion, to our knowledge, this is the first study analyzing the TL epidemiological situation for a 10-year period in Colombia, accounting for all geographical subdivisions, and vector species distribution in the country. The Andean, Amazon, and Orinoco regions account for over 75% of the cases of TL in Colombia, so that public health efforts should be extended there, particularly away from urban areas. Enhanced control can be achieved through the implementation of disease vigilance and control systems, and the strengthening of public health networks that could reduce the disease’s impact on the population. Adequate geospatial modelling of the disease could be achieved by means of a more extensive case description which includes socioeconomical and biogeographical data. This information would be required in order to develop predictive models that could more accurately guide public health efforts towards this neglected disease.
|
10.1371/journal.pcbi.1006952 | Prediction of VRC01 neutralization sensitivity by HIV-1 gp160 sequence features | The broadly neutralizing antibody (bnAb) VRC01 is being evaluated for its efficacy to prevent HIV-1 infection in the Antibody Mediated Prevention (AMP) trials. A secondary objective of AMP utilizes sieve analysis to investigate how VRC01 prevention efficacy (PE) varies with HIV-1 envelope (Env) amino acid (AA) sequence features. An exhaustive analysis that tests how PE depends on every AA feature with sufficient variation would have low statistical power. To design an adequately powered primary sieve analysis for AMP, we modeled VRC01 neutralization as a function of Env AA sequence features of 611 HIV-1 gp160 pseudoviruses from the CATNAP database, with objectives: (1) to develop models that best predict the neutralization readouts; and (2) to rank AA features by their predictive importance with classification and regression methods. The dataset was split in half, and machine learning algorithms were applied to each half, each analyzed separately using cross-validation and hold-out validation. We selected Super Learner, a nonparametric ensemble-based cross-validated learning method, for advancement to the primary sieve analysis. This method predicted the dichotomous resistance outcome of whether the IC50 neutralization titer of VRC01 for a given Env pseudovirus is right-censored (indicating resistance) with an average validated AUC of 0.868 across the two hold-out datasets. Quantitative log IC50 was predicted with an average validated R2 of 0.355. Features predicting neutralization sensitivity or resistance included 26 surface-accessible residues in the VRC01 and CD4 binding footprints, the length of gp120, the length of Env, the number of cysteines in gp120, the number of cysteines in Env, and 4 potential N-linked glycosylation sites; the top features will be advanced to the primary sieve analysis. This modeling framework may also inform the study of VRC01 in the treatment of HIV-infected persons.
| The two Antibody Mediated Prevention (AMP) clinical trials are testing whether intravenous infusion of VRC01 (an antibody that can neutralize most HIV-1 viruses) can prevent HIV-1 infection. Since the outer envelope (Env) protein of HIV-1 is highly genetically diverse, the AMP trials will evaluate in an “amino acid sequence sieve analysis” whether VRC01 prevents infection differentially depending on Env amino acid features of exposing viruses. To maximize power of sieve analysis, the number of amino acid features tested should be limited to those most likely associated with whether the virus is sensitive to neutralization by VRC01. We used machine learning to analyze a database of 611 HIV-1 Envelope pseudoviruses, with data on how well VRC01 neutralizes each pseudovirus, to identify models that best predict neutralization sensitivity from the amino acid features and to identify the most predictive features. We identified models that could predict HIV-1 sensitivity (as opposed to resistance) to VRC01 very well, and found that several amino acid residues in Env locations where both VRC01 and the CD4 receptor bind were important for making correct predictions. Our modeling approach will enable a focused AMP sieve analysis and may be useful for studying the use of VRC01 in the treatment of HIV-infected persons.
| The immense genetic and antigenic diversity of the trimeric HIV-1 envelope (Env) glycoprotein spike [precursor form = (gp160)3, proteolytically cleaved to (gp120/gp41)3], the major target of neutralizing antibodies, poses a significant problem in the development of an effective prophylactic vaccine. Broadly neutralizing monoclonal antibodies (bnAbs) isolated from individuals with chronic HIV-1 infection have demonstrated significant promise by targeting a wide spectrum of this diversity [1–5]. These bnAbs generally target conserved elements in one of five regions of gp160: the V2 variable region, the N332 region in the V3 region, the CD4 binding site (CD4bs), the gp120–gp41 interface, and the membrane proximal external region [6]. Knowledge of Env amino acid (AA) signature patterns that are associated with a neutralization phenotype of interest [7] informs our understanding of the relevant immunogenic characteristics of HIV-1 and has important implications for bnAb regimen and HIV-1 vaccine design.
The IgG1 monoclonal antibody (mAb) VRC01 neutralizes more than 80% of 600 viral strains tested in vitro [1, 8], and targets a region in the relatively conserved CD4bs [1, 2, 9]. Evidence from several experimental animal infection models [10–13] highlights the potential of bnAbs such as VRC01 to prevent HIV-1 infection when administered via passive immunization in pre-exposure or post-exposure prophylactic strategies [14, 15]. VRC01 has moved through four phase 1 clinical trials (VRC601 [16], VRC602 [17], A5340 [18], and HVTN 104 [19]) and is now being evaluated in the phase 2b Antibody Mediated Prevention (AMP) trials [HVTN 704/HPTN 085 (ClinicalTrials.gov identifier NCT02716675) and HVTN 703/HPTN 081 (ClinicalTrials.gov identifier NCT02568215)] [20], the first proof-of-concept efficacy trials in adults to determine whether passive administration of a bnAb can prevent HIV-1 acquisition in men who have sex with men and transgender persons, and women who are at risk of HIV infection. A detailed description of the AMP trials and the statistical considerations of their designs can be found in Gilbert et al. [20].
Following the conclusion of the AMP trials, we will conduct a series of “sieve analyses,” which investigate the extent to which intervention-mediated protection from infection varies with phenotypic (phenotypic sieve analysis) and AA sequence (genotypic sieve analysis) characteristics of the exposing viruses [21]. The phenotypic sieve analysis in the AMP trials will compare functional properties (such as sensitivity to VRC01-mediated neutralization) of the breakthrough founding viruses from infected VRC01 recipients versus infected placebo recipients; this kind of analysis was previously conducted in the VAX004 efficacy trial of a candidate preventive HIV-1 gp120 vaccine [22]. The other type of sieve analysis to be conducted in the AMP trials–AA sequence or genotypic sieve analysis–compares AA features of breakthrough founding Env sequences from infected VRC01 recipients versus infected placebo recipients, similar to what we and others have done in preventive vaccine efficacy trials for HIV-1 [23–29], malaria [30], and dengue [31]. An AA sequence sieve effect for a particular HIV-1 AA sequence feature is defined as significant variation in prevention efficacy against viruses with different levels of this feature.
A major challenge posed to AA sequence sieve analysis is the large number of Env AA sequence features that could be considered for analysis, as an exhaustive search for sieve effects would have low statistical power after multiple-testing adjustment. Therefore, “down-selection” of a set of top-ranked AA sequence features to the primary sieve analysis is important for conserving statistical power. To address this challenge in vaccine efficacy trials, our approach first conducts pre-specified primary analyses that focus on a limited subset of AA features based primarily on knowledge of the specificity of the vaccine-elicited immune responses, or aggregates AA information into distances to vaccine-insert sequences [24, 31, 32]. Following these primary analyses, we conduct pre-specified exploratory sieve analyses in which we search for sieve effects across a much more exhaustive set of features, considering the full proteomic sequence and other genomic features, with the goal of generating additional hypotheses about how prevention efficacy depends on pathogen proteomics/genomics.
For the AMP bnAb efficacy trials, our guiding criterion for including an AA sequence feature is evidence that it helps predict the sensitivity of the virus to VRC01-mediated neutralization, as measured in vitro by the TZM-bl assay that will be used for the phenotypic (neutralization) sieve analysis. Two specific objectives of our work are: (1, “model selection”) to develop a best model or best few models for predicting TZM-bl neutralization sensitivity to VRC01 and advance this model or these models for use in the primary AA sequence sieve analysis, where we refer to predicted outcomes from these models based on given virus AA sequences as “proteomic antibody resistance” (PAR) scores; and (2, “feature selection”) to rank AA sequence features by their importance for predicting TZM-bl neutralization sensitivity to VRC01, and select the most important features to advance to the primary AA sequence sieve analysis. In particular, for (1), the VRC01 resistance level of different viral Env sequences can be compared using PAR scores. As such, the AA sequence sieve analysis will estimate how prevention efficacy (PE) against HIV-1 acquisition varies with the defined PAR score of the virus, similar to the neutralization sieve analysis that assesses how PE varies with the measured IC50, IC80, or slope of the virus. This analysis will also allow a comparison of how well AA sequence information vs. measured neutralization information discriminates PE. For (2), for each advanced top-ranked feature, the AA sequence sieve analysis will consist of point and confidence interval estimates of PE against each HIV virus proteomic type defined by a level of the feature, as well as a statistical test of whether PE varies across the different virus types, with multiplicity-adjustment across the top-ranked features.
We addressed both objectives using data from the Compile, Analyze and Tally NAb Panels (CATNAP) database [8], which collates IC50 and IC80 neutralization values for specific mAbs versus HIV-1 Env pseudoviruses used in the assay, as well as the corresponding Env viral sequences. We used two machine learning approaches, both of which used a set of pre-defined AA sequence features to predict each of five TZM-bl neutralization outcomes: two dichotomous outcomes indicating a virus’s resistance vs. sensitivity status based on IC50, and three quantitative outcomes (log IC50, log IC80, and the estimated neutralization slope of the dose-response curve). A strength of our approach compared to previous approaches for predicting neutralization resistance from AA sequence features is that we provide formal statistical inferences (i.e., confidence intervals) for cross-validated parameters that quantify prediction accuracy. We also apply a recently proposed variable importance measure that is interpretable without requiring a particular parametric model to be correctly specified and describes importance relative to the population rather than relative to a fitted machine learning algorithm. The entire analysis was done on two independent splits of the available VRC01 CATNAP data to enable a simple way to evaluate replicability of the findings and to cross-check prediction accuracy. Objective (1) was achieved with the identification of models that provide PAR scores highly predictive of a resistant vs. sensitive virus. Objective (2) was achieved in that we identified 42 AA sequence features with high variable importance measure (VIM) scores, indicating that they were highly predictive of VRC01 neutralization sensitivity, and that were also significantly associated with neutralization.
The distributions of neutralization sensitivity outcomes of Env pseudoviruses in dataset 1 (comprising half of the Env sequences retrieved from the CATNAP database, see Methods) and in dataset 2 (comprising the other half of the Env sequences retrieved from the CATNAP database, see Methods) are shown in Fig 1. Sixteen percent of all viruses included in this analysis have right-censored IC50 values, and hence are considered resistant for both dichotomous outcomes. For the analysis of the sensitive/resistant only outcome, 22% of all viruses were excluded from the analysis because they qualified as neither sensitive (IC50 < 1 μg/mL) nor resistant (right-censored IC50 value). In this reduced population, 79% of viruses are sensitive and 21% are resistant. The quantitative IC50 readouts have broader variability than the quantitative IC80 readouts. In all analyses, we included each virus’s geographic region of origin, to control for possible geographic confounding. All confidence intervals provided in parentheses represent 95% confidence intervals.
To address objective 1, we applied nonparametric ensemble-based cross-validated learning (a form of stacking [33]) as the primary learning method, which is often referred to as super learning or the Super Learner [34] (see Methods, Statistical Learning Approaches). We compared the performance of Super Learner with each of its component learners as comparative benchmarks. The Super Learner enjoys an oracle property ensuring in large samples that its error (e.g., mean-squared error or AUC for predicting neutralization resistance from AA sequence features) is essentially the same or better than any of its individual component learners [34]. The Super Learner has also been shown to perform well in many simulation studies and in real data applications (e.g., [34–36]). We quantified model performance by rigorous inference on data-adaptive target parameters [37], including cross-validated area under the receiver operating characteristic curve (CV-AUC) for binary outcomes [38] and cross-validated nonparametric R-squared (CV-R2) for quantitative outcomes, the latter of which is a version of cross-validated MSE that is scaled by the variance of the outcome for the sake of interpretability [39]. Cross validation is used for an initial comparison to confirm that Super Learner performs equivalently or better than its component learners, while our primary criterion for evaluating a model’s performance is with the holdout data, using the area under the receiver operating characteristic curve (AUC) for binary outcomes and nonparametric R-squared for continuous outcomes (R2).
While the most auspicious models for the AMP sieve analysis will have maximally high CV-R2 or CV-AUC, it is difficult to define specific thresholds for these metrics for qualifying a model for use in the AMP sieve analysis. However, we propose that a bare minimal requirement is that the 95% confidence interval (CI) about the chosen cross-validated prediction accuracy metric indicates significantly greater prediction accuracy than a pure-noise model—this benchmark is a 95% CI for CV-R2 above 0 and a 95% CI for CV-AUC above 0.5.
To define our terminology, we use the term “prediction” to define the general case of predicting a neutralization endpoint, either dichotomous or quantitative, and we also use it in the specific case of regressing quantitative endpoints. We use the term “classification” to specify the prediction of dichotomous neutralization endpoints.
We structured the objective 2 variable importance analysis by pre-specifying thirteen distinct input variable groups of Env AA sequence features that could potentially be relevant for VRC01 neutralization based on statistical and biological (structural, immunological, and virological) grounds. The input variable groups are, with the first six based on individual Env AA positions and residues at those positions: 1) VRC01 binding footprint sites, 2) CD4 binding sites, 3) sites with sufficient exposed surface area, 4) sites identified as important for glycosylation, 5) sites with residues that covary with the VRC01 binding footprint sites, 6) sites associated with VRC01-specific potential N-linked glycosylation (PNGS) effects, 7) sites in gp41 associated with VRC01 neutralization sensitivity or resistance, 8) indication of potential N-linked glycosylation sites (PNGS), 9) majority virus subtypes, 10) region-specific counts of PNG sites, 11) viral geometry, 12) cysteine counts, and 13) steric bulk at critical locations. The Methods section provides details about the features (“Envelope amino acid feature input variable groups”).
For each of the five outcomes, we calculated VIMs for all features included in any of the 13 input variable groups with two VIM analysis approaches. The first applied a Monte Carlo Cross-Validation (MCCV)-based algorithm-specific approach that did not take into account the input variable grouping, whereas the second applied an ensemble-based approach using the Super Learner (see Methods, Statistical Learning Approaches) to assess variable importance of each of the 13 variable groups and individual features. We report as top-ranked features those with a Holm-Bonferroni adjusted 2-sided p-value < 0.05 from a univariate regression adjusted for geographic region, and that rank among the top 50 features by either of the two VIM approaches. The primary results from these VIM analyses pertain to the IC50 censored outcome and the log IC50 outcome, which are best-predicted dichotomous and quantitative outcomes with the largest sample size.
For the IC50 censored dichotomous outcome analysis, Table 1 reports all features that ranked among the top 50 features by either VIM method, and that had a Holm-Bonferroni 2-sided p-value less than 0.05 for an association with the outcome in a logistic regression model using both datasets (with adjustment for geographic region as in all analyses). This p-value criterion was more stringent than the pre-specified criteria, and was added to ensure that any individual features selected by our VIM method were individually predictive after strict multiplicity adjustment using a well-understood standard method, logistic regression. Table 2 shows the results for the quantitative log IC50 outcome, under the same rule for reporting except replacing logistic regression with linear regression. There is much overlap between the features found in Tables 1 and 2, and if we take the union of all features found for both endpoints, the result is a set of 42 unique features.
The majority of the top-ranked features for the two outcomes pertain to presence/absence of specific Env residues, with the most important residues located at CD4 contact sites shown previously to be associated with VRC01 neutralization sensitivity or resistance. The most important of the features predictive of non-censored IC50 (which we refer to here as neutralization sensitivity) were an arginine at position 456 (R456) and a glycine at position 459 (G459), which were identified as top-ranked features for both outcomes and by both VIM methods; when referring to an AA present at a given position, we give the one-letter code of the AA followed by its position in HBX2 coordinates. Other highly ranked features predictive of sensitivity were N280 (top-ranked by both VIM methods for IC50 censored), G458 (top-ranked by the algorithm-specific method for the IC50 censored outcome), and D279 (top-ranked by the algorithm-specific method for the log IC50 outcome). The most highly ranked residue predictive of neutralization resistance was I471 (highly ranked by both VIMs for both outcomes). In total, 16 residues predictive of neutralization sensitivity and 10 residues predictive of neutralization resistance made the top-ranked list in Table 1. Of these, 6 (37.5%) residues predictive of neutralization sensitivity and 3 (30%) residues predictive of neutralization resistance also made the top-ranked list in Table 2. Visualizations of the locations, magnitudes, and distributions of the most predictive residues in Tables 1 and 2 are provided in Fig 6. Clusters of predictive sites are found just prior to and within the V5 variable loop, and within Loop D. The logo plots in Fig 6C and 6D show the distributions of amino acids within neutralization-sensitive and -resistant viruses, respectively. These figures demonstrate that, with the majority of the predictive sites, minority mutations are entirely or strongly associated with resistance (e.g., with anything but arginine (“R”) at site 456 conferring VRC01 resistance), and that no strong discriminating signal exists at the individual site level, implying that a multivariate predictor would be necessary for effective performance.
Besides site-specific residue information, other AA sequence features were also highly ranked. A longer length (in AAs) of Env and of gp120 was found to be important for predicting resistance for both outcomes, and more cysteines in Env were found to be important for predicting resistance in the IC50 censored outcome. For the dichotomous outcome, the presence of a PNGS at either site 156 or site 616 was important for predicting sensitivity, and the presence of a PNGS at either site 229 or site 824 was important for predicting resistance. The number of PNG sites in the V5 region was important for predicting neutralization sensitivity, for the dichotomous outcome only, and subtype A1 was also important for predicting neutralization sensitivity, for the log IC50 outcome only. VIM results for the other three outcomes are given in S7 Table.
In addition to evaluating the predictive importance of individual AA sequence features, the ensemble-based Super Learner approach also estimated VIMs for the groups of pre-specified AA sequence features (S11 Fig) by feature group. This analysis showed that the group of AAs in the VRC01 binding footprint and the CD4 binding sites were the most important predictive groups, a result found for both the dichotomous and the log IC50 outcome.
Excellent research has been done to understand HIV-1 Env genotypic/AA features that affect VRC01 resistance [2, 9, 41–44]. Our approach complements this work by applying state-of-the-art machine learning and methodology for unbiased, nonparametric statistical inference (our area of expertise) − a contribution not yet provided in previous computational approaches to define AA sequence signatures for various bnAbs [7, 45–49] − and then interpreting the results in comparison with those from literature. Other advantages of our modeling approach are that it (i) combines the strengths of several machine learning algorithms to improve prediction accuracy in an automated, unbiased way; and (ii) bases all decisions on hold-out validation, where only data not used in model building is used for measuring predictive performance and for selecting features.
Within each of two separately analyzed halves of the CATNAP data, we used the metrics of cross-validated AUC and cross-validated nonparametric R2 for summarizing prediction accuracy to compare the performance of different learners and feature sets, with 95% confidence intervals that are valid in the context of “inference after model selection” [38]. We then applied models fit with one dataset to the other completely separate dataset, validating their performance and effectively providing a simple conclusion of replicability of the results. In addition, our novel Super Learner-based variable importance measure (VIM) has a useful incremental predictive value interpretation for a given feature, as the additional proportion of variance in the outcome explained by including the feature in addition to including all other features [50]. An advantage of this nonparametric VIM is that its interpretation does not require any particular parametric model for the data to hold, in contrast to methods such as logistic regression.
In our first objective (model selection), we identified several models via machine learning that provided strong and similarly accurate performance to classify viruses (based on specific AA sequence features in Env) as resistant vs. sensitive, measured by whether neutralization IC50 is right-censored or by restricting the “sensitivity” category to IC50 < 1 μg/mL. Due to its advantageous theoretical properties, strong performance in simulation studies and data analyses, and consistently high performance in the present work, the Super Learner is the learning approach that will be advanced to the primary sieve analysis. We have already created a preliminary Super Learner-based model to predict neutralization sensitivity of virus sequences obtained from HIV-1 infections occurring during the AMP trials. When predicting the quantitative endpoints, our models had weaker performance than those built to classify the dichotomous outcomes. It is possible that the dichotomous IC50 outcome is easier to classify because the value “censored” has clear meaning as neutralization activity not being detected in the experiment [8], whereas log IC50 and log IC80 may have more noise due to natural variability in the TZM-bl assay and the assignment of specific numeric values to censored values. In addition, of the three quantitative outcomes studied, prediction performance was best for log IC50, intermediate for log IC80, and worst for neutralization slope. This may be explained in part because 29.6% percent of the 611 viruses in CATNAP were missing data on IC80 (and by extension, neutralization slope), and perhaps the missing data is contributing to a diminished predictive signal. In addition, the slope readout is a ratio-readout with the additive difference term log IC80 –log IC50 in the denominator, such that if noise makes IC80 close to IC50, then the denominator is small, and the impact of noise is amplified. This can occur because each of IC80 and IC50 are estimated imperfectly based on a percent neutralization by concentration curve (with IC50 estimated with somewhat more precision given the fiftieth percentile is in the center of the percent neutralization distribution). Thus, it remains an open question whether slope is a meaningful neutralization outcome.
Our novel findings in objective 2 pertain to the variable importance measures of individual AA sequence features, which identified 14 specific residues at certain AA sites (M181, E279, S280, T280, C397, R425, M428, Q455, H456, S456, W456, D459, I471, and N655) and 6 other AA sequence features (longer gp120, longer Env, more cysteines in gp120, more cysteines in Env, and PNG sites at 229 and 824) that had high variable importance for predicting neutralization resistance. We also identified 18 specific residues (P124, N156, L179, D279, N280, S365, H374, N425, Q428, R456, D457, G458, G459, E466, R469, G471, T569, and D589) and 4 other AA sequence features (PNGS at sites 156 and 616, more PNG sites in V5, and subtype A1) that had high variable importance for predicting neutralization sensitivity.
Most of the important features identified in this study were based on the documented VRC01 footprint and the CD4 binding sites, which largely overlap with each other. Seven of the predictive residues that we identified, however, fall outside of these regions. Three of these seven sites were included for analysis because of their covariation with sites within the VRC01 footprint: L179 (lending to VRC01 sensitivity), and M181 and C397 (lending to VRC01 resistance). The latter finding is interesting, in that it is rare to find cysteines in the surface-exposed region of gp160 outside the context of the 10 canonical disulfide bond-forming pairs (described in [51]). Neither site 179 nor site 397 has been described in the literature to be associated with VRC01 activity, although glycans at 397 have been found to be important for the binding activity of other bnAbs [52].
Three of the four remaining sites identified as predictive by this study were included as part of the sites in gp41 that have been found to be associated with VRC01 binding: T569 and D589 were associated with increased sensitivity, and N655 was associated with resistance. These results agree with previous reports of similar mutations altering neutralization sensitivity globally [53–55], and highlight how variation in the HR-1 and HR-2 domain of gp41 can modulate sensitivity to neutralization by VRC01.
The final site identified by this study is N156 (lending to VRC01 sensitivity), which was included as part of Group 6, based on the probability that the presence or absence of a PNGS at this site would result in improved or reduced VRC01 binding [56]. Indeed, site 156 was observed to be a PNGS in 94% (576 of 611) of the CATNAP sequences included in this study, where disruption of the PNGS motif was more likely to confer VRC01 resistance: 51% of sites without a PNGS at site 156 were found to be VRC01 resistant, while only 14% of sites with a PNGS at site 156 were VRC01 resistant. A glycan at this position has been hypothesized as being important for recognition by other bnAbs [57], but to the best of our knowledge this is the first report to associate N156 with sensitivity to neutralization by VRC01.
Many of the residues we identified as highly predictive of at least one of the outcomes are supported by experimental evidence as being important for VRC01 binding. Four of the top-ranked AAs found in this study (D279, N280, R456, and G459) have been shown to be sites of common interactions with potent VRC01-like Abs [58], and D279 and E459 have been identified as making critical interactions with VRC01 [9, 52]. Moreover, mutation of residue D279 to E279 (D279E) was shown to be part of the VRC01 escape pathway within the donor from whom VRC01 was isolated [41]. For objective 2, we also found that AA sequence features in the VRC01 binding footprint sites and in the CD4bs have greatest variable importance, a result that is not surprising given previous work. This finding supports conducting AMP sieve analyses that focus on groups of AA sites that define these two regions.
Diverse approaches have been taken to identify Env sequence patterns associated with bnAb neutralization sensitivity (bnAb signatures) [7, 45–49]. We next discuss our work in the context of sequence-based approaches that have been taken to predict sensitivity to VRC01-mediated neutralization. Using non-linear support vector machines to predict neutralization sensitivity of pseudoviruses with different Env AA sequences, Hake and Pfeifer identified N186, N276, N279, N280, G459, and K232 as strong predictors of susceptibility to VRC01-mediated neutralization [59]. Of these 6 AAs, we found that G459 ranked extremely highly for contribution to prediction to 4 out of the 5 outcomes (Table 2, S6 Table). N280 also ranked highly for contribution to prediction of the quantitative log IC50 outcome (Table 2). Moreover, considering that for each of the outcomes, only a small number of sites (between 2 to 4 for each outcome) met our criteria for being highly predictive of a given outcome (high VIMs across both methods, a low FWER p-value), and that our sensitivity analysis was able to achieve equivalent classification performance of the IC50 censored outcome with only five features (see Results), our results support the overall conclusion of Hake and Pfeifer that, in general, only a few key residues are needed to well-predict neutralization resistance.
While we have reported our specific findings based on all available VRC01 CATNAP data as of the initiation of this work in March 2017, the CATNAP database is being continuously updated to include new results in the scientific literature; at the time of this writing, the CATNAP database includes 54 neutralization results that were not available when the datasets for this analysis were finalized. When finalizing the plan for the AMP sieve analysis (expected in 2020), we plan to re-run this predictive analysis with an up-to-date version of the CATNAP database, to (a) ensure that our selected PAR scores are based on the maximum number of pseudoviruses/sequences; (b) verify that the most predictive AA features remain the same, and (c) update our analysis plans accordingly if new AA features are found that outperform the top-performing AA features identified here. We will also consider including AA sequence features identified by others in complementary analysis approaches.
We now discuss the limitations of this study. To maximize our sample size, we used all available sequences with TZM-bl neutralization data in the CATNAP database, regardless of their subtype. The AMP trials, however, are conducted in regions where circulating HIV-1 viruses are largely subtypes B (Americas and Switzerland) and C (southern Africa). As such, the results of this analysis may be influenced by characteristics of viral subtypes that will not play a role in the AMP trials; however, we note that our analysis did assess whether subtype helped predict neutralization resistance, and only subtype non-A1 vs. subtype A1 was found to be an important feature (while subtype C was associated with greater neutralization resistance, its variable importance measures did not rank it among the most predictive features).
While the TZM-bl assay is validated and the multiple labs contributing data to CATNAP undergo certification through proficiency testing, nevertheless it is common for different labs to produce IC50 or IC80 readouts with two-to-three-fold variation on the same samples [60]. This variability creates noise in the outcome variable that dampens prediction accuracy. In addition, the outcome predicted best by our models–whether the IC50 outcome was reported as right-censored (i.e., resistant) in the original publication cited by CATNAP [8]–has noise stemming from unknown differences among labs in factors that were considered in defining the outcome.
Another limitation of our approach is that we considered prediction of neutralization sensitivity of a single Env pseudovirus based on its gp160 AA sequence, but some AMP efficacy trial participants are expected to be infected with multiple founder viruses. How to properly account for the number of founders and the accompanying multiple gp160 sequences in predicting neutralization sensitivity of an exposing viral quasispecies is an open question for future research. An additional caveat is that we only analyzed VRC01 neutralization readouts obtained by one particular assay, which uses TZM-bl target cells and is performed in vitro, only approximating a real-life exposure event of the genital mucosa to HIV-1 in the presence of VRC01; however, given that this assay is the standardized and validated platform for HIV-1 vaccine evaluation, developing predictors of this assay’s readouts is an important goal, with a test of in vivo validation forthcoming from the AMP sieve analysis. Prior evaluation of the ability of multiple bnAbs to prevent HIV-1 infection using a mucosal tissue explant model has shown that neutralizing activity as assessed by the TZM-bl assay is moderately correlated with inhibitory activity in penile and cervical tissue, but not correlated with inhibitory activity in colorectal tissue [61]. In addition, VRC01 may protect against HIV-1 acquisition in additional ways not captured by a neutralization assay, such as via non-neutralizing Fc effector functions. In support of this idea, VRC01 (or serum from participants infused with VRC01) has been shown to mediate low levels of antibody-dependent cell-mediated cytotoxicity [62] and higher levels of antibody-dependent cellular phagocytosis (ADCP) of gp140-coated microspheres and of virion capture [63], which may also be important for preventing HIV-1 acquisition. These findings make it of interest in future work to build models predicting ADCP and other non-neutralizing Fc effector functions based on AA sequence features.
With this study, we have created and applied modeling tools to help design the primary AA sequence sieve analysis in AMP, such that the analysis will be based on the hypothesis that Env AA-based predictors of in vitro resistance measured by the TZM-bl assay will also discriminate prevention efficacy. For each of the top-ranked features we identified, the AMP sieve analysis could test whether the level of prevention efficacy differs across HIV-1 variants of the feature. Beyond preparation for sieve analysis in bnAb prevention efficacy trials, another application of our predictive modeling framework includes scoring AA signature types for bnAb resistance, prior to using a particular bnAb as a therapeutic in an HIV-1 infected individual.
A total of 624 Env viral AA sequences, their associated pseudovirus IC50 and IC80 values for neutralization by VRC01 as assessed by the TZM-bl assay, and other associated annotations were retrieved from the CATNAP database [8]. [The 50% and 80% inhibitory concentrations (IC50 and IC80, respectively) are defined as the concentration of VRC01 required to cause either a 50% or 80% reduction in Env pseudovirus replication (as measured in relative luminescence units) relative to the level of replication in the absence of VRC01. This reduction in replication in the presence of VRC01 is inferred to be a consequence of VRC01-mediated neutralization. Hence, a low IC50 or IC80 value for VRC01 indicates that the given Env pseudovirus is relatively sensitive to VRC01-mediated neutralization, whereas a higher or right-censored IC50 or IC80 value indicates that the given Env pseudovirus is relatively resistant to VRC01-mediated neutralization.] Some of the provided annotation was unstructured or unsuitable for analysis, so these data were refactored appropriately. (Additional details about the data processing step can be found in the S1 Text.) Thirteen sequences/pseudoviruses were excluded from the analysis, because their IC50 measurement was recorded as right-censored at 1 μg/ml. According to the study that produced these results [64], this unusually low limit of censorship was due to a lack of reagent. As such, we excluded these pseudoviruses from the study, as their neutralization results were regarded as unreliable.
This resulted in a total of 611 sequences/pseudoviruses to include in the analysis, of which 48.0% (293) were subtype C (the predominant subtype in the HVTN 703/HPTN 081 trial) and 13.3% (81) were subtype B (the predominant subtype in the HVTN 704/HPTN 085 trial) (S8 Table). Of these 611 pseudoviruses, all 611 had quantitative log IC50 neutralization readouts, which means that the IC50 censored outcome and the quantitative log IC50 outcome (defined below in “TZM-bl resistance outcomes used in this analysis”) were available for all 611 of them. We were able to derive a sensitive/resistant only outcome for 474 (77.6%) of these pseudoviruses, and 430 (70.4%) of the pseudoviruses had an IC80 neutralization readout, which means that we were only able to derive quantitative log IC80 and neutralization slope outcomes for these 430 pseudoviruses. All of the data used in this analysis, including the identifiers of the sequences and their outcomes used, are posted at https://github.com/benkeser/vrc01/tree/1.0.
The restructured dataset was randomly partitioned into two datasets (“dataset 1” and “dataset 2”) for the statistical learning analyses. The two datasets were mutually exclusive, each with half of the data [n = 306 (dataset 1) and n = 305 (dataset 2)]. The random partitioning process was stratified by the viruses’ country of origin. We chose to stratify the data by country instead of HIV-1 subtype because subtype was to be included as an input feature in the analysis, as it was considered to be a potential sequence-based predictor of resistance, whereas country was not used as an analytical feature but was controlled for as a potential confounder. Of the 611 HIV-1 sequences analyzed, 51.9% (317) originated from a country in which one or more study sites for the HVTN 703/HPTN 081 trial are located and 9.5% (58) originated from a country in which one or more study sites for the HVTN 704/HPTN 085 trial are located (S8 Table). The geometric means of the imputed log10 IC50 values of the pseudoviruses whose Env sequences were included in this analysis are shown by region/subtype in S12 Fig.
AA sites of potential relevance to VRC01-mediated neutralization of HIV-1 were included as input features for the statistical learning analyses. For features defined by residue content at a given AA site, only AA sites passing a minimum variability filter were included. Specifically, the residue had to differ from the consensus residue in at least 3 sequences in the entire analysis dataset (i.e., before splitting into the two analysis sets). Furthermore, indicators for the presence of a residue at a specific site were only included if that residue was present in at least three viral sequences at that site across the entire dataset.
Those sites that passed the minimum variability requirement were included in the analysis if they fell into any of the following pre-defined groups (many sites are found in more than one pre-defined set):
In addition to AA sites, the following features within feature groups were also included as input features:
All of these groups, and the sites contained within them, are outlined in Table 3A. Fig 7 provides a schematic visualization of the AA sites in Feature Groups 1–7 before application of the minimum variability filter (specific sites in each group are listed in S1 Text).
For this analysis, new univariate IC50 and IC80 outcome variables for each pseudovirus were created that incorporate imputations from the right-censored IC50 and IC80 values retrieved from CATNAP and that account for occasions that values were reported from multiple studies. This process is described in detail in S1 Text.
Using these IC50 and IC80 univariate outcomes, the statistical learning processes aimed to predict from the AA sequence features five TZM-bl neutralization resistance outcomes: (1) dichotomous resistance outcome of whether the IC50 is right-censored (the “IC50 censored” outcome) as recommended in [8]; (2) dichotomous resistance outcome (the “sensitive/resistant only” outcome), with resistance defined by the IC50 being right-censored as in (1) and sensitive defined as IC50 < 1 μg/ml; (3) log10 IC50 resistance outcome as a quantitative measure (the “quantitative log IC50” outcome); (4) log10 IC80 resistance outcome, also as a quantitative measure (the “quantitative log IC80” outcome); and (5) estimated dose-response curve slope of neutralization, which is a function of IC50 and IC80, calculated as in equation (6) of [83], equal to log(4) divided by (log IC80 –log IC50 (the “neutralization slope” outcome). A caveat of the “IC50 censored” outcome analysis is that 50 pseudoviruses that were included were right-censored at value IC50 > 10 μg/ml (no pseudoviruses right-censored at a lower value were included), yet some pseudoviruses had observed IC50s only incrementally larger than 10 μg/ml (e.g., 14 pseudoviruses had IC50 greater than 10 μg/ml and less than 50 μg/ml); this issue could add noise to the analysis.
|
10.1371/journal.pntd.0007133 | A chronic bioluminescent model of experimental visceral leishmaniasis for accelerating drug discovery | Visceral leishmaniasis is a neglected parasitic disease with no vaccine available and its pharmacological treatment is reduced to a limited number of unsafe drugs. The scarce readiness of new antileishmanial drugs is even more alarming when relapses appear or the occurrence of hard-to-treat resistant strains is detected. In addition, there is a gap between the initial and late stages of drug development, which greatly delays the selection of leads for subsequent studies.
In order to address these issues, we have generated a red-shifted luminescent Leishmania infantum strain that enables long-term monitoring of parasite burden in individual animals with an in vivo limit of detection of 106 intracellular amastigotes 48 h postinfection. For this purpose, we have injected intravenously different infective doses (104—5x108) of metacyclic parasites in susceptible mouse models and the disease was monitored from initial times to 21 weeks postinfection. The emission of light from the target organs demonstrated the sequential parasite colonization of liver, spleen and bone marrow. When miltefosine was used as proof-of-concept, spleen weight parasite burden and bioluminescence values decreased significantly.
In vivo bioimaging using a red-shifted modified Leishmania infantum strain allows the appraisal of acute and chronic stage of infection, being a powerful tool for accelerating drug development against visceral leishmaniasis during both stages and helping to bridge the gap between early discovery process and subsequent drug development.
| Visceral leishmaniasis is a neglected disease that poses a significant threat to impoverished human populations of low-income countries. Due to the unavailability of vaccines, pharmacological treatment is the only approach to control the disease that otherwise can be lethal. To date, drug management in endemic regions is based on combinations of a handful of mostly unsafe drugs, where the emergence of resistant strains is an additional problem. To accelerate the discovery of new drug entities, several gaps from the early discovery of a compound to its public use, should be filled. One of these gaps is the need of a rapid go/no-go testing system for compounds based on robust preclinical models. Here, we propose a new long-term model of murine visceral leishmaniasis using in vivo bioluminescent imaging. For this purpose, a red-shifted bioluminescent Leishmania infantum strain was engineered. This strain has allowed the appraisal of the disease in individual animals and the monitoring of parasite colonization in liver, spleen and bone marrow. As proof of concept of this platform, mice were infected with the transgenic L. infantum strain treated with a standard schedule of miltefosine, the only oral drug available against Leishmania parasites. Bioluminescence and parasite load in the target organs were compared showing a good correlation. Our findings provide a robust and reproducible tool for drug discovery in a chronic model of murine visceral leishmaniasis.
| Leishmaniasis is a complex of neglected parasitic diseases affecting the poorest people in 98 countries, particularly those with weak or non-existent health systems. [1]. There are at least three different forms of clinical presentations; cutaneous, mucocutaneous and visceral leishmaniasis, the latter being fatal if left untreated [2]. Visceral leishmaniasis (VL) is estimated to produce 300.000 new cases and between 20.000–40.000 deaths every year. Most of the cases are localized in three geographical regions; South Asia and East Africa where the disease is caused by Leishmania donovani and the transmission is mostly anthroponotic. By its part, in Brazil, where the disease is produced by L. infantum chagasi, the transmission is zoonotic and occurs mainly from infected dogs [3].
Nowadays, therapeutic or prophylactic human vaccines are still lacking, and the cure of patients is based on chemotherapy [4, 5]. Treatment of VL was mainly based on painful intramuscular injections of pentavalent antimonials, such as sodium stibogluconate (SSG). SSG has been the first-line antileishmanial drug in India, although its clinical efficacy in some areas of North Bihar State has gradually declined, due to the emergence of fully resistant L. donovani strains. SSG is being substituted by liposomal amphotericin B (AmBisome) as first-line treatment, despite slow intravenous administration of the drug is needed [6–8]. In East Africa, SSG was the first-line regimen for decades, but due to its toxicity and following WHO recommendations in 2010, SSG + paromomycin combination therapy became the treatment of choice [9]. However, the administration of this drug combination is painful and requires patient hospitalization, and therefore, more friendly alternatives were implemented. These include single dose of AmBisome plus 10 consecutive days of SSG, single dose of AmBisome plus 10 days of miltefosine or miltefosine alone for 28 days. However, none of these combinations improved the results of the treatment of choice in Phase II clinical trials [10].
Miltefosine is the last drug successfully introduced against VL. It is also the only drug that has a good oral bioavailability. However, an increase in relapse rates has been reported in India and Nepal, probably associated with low drug exposure [11, 12]. In addition, miltefosine is potentially embryotoxic and fetotoxic in experimental animals and thereby, its administration is not recommended in women during pregnancy [13].
For all these reasons, there is an unmet need to fill the antileishmanial drug discovery pipeline with safer drugs that display new mechanisms of action, likely allowing combination therapy in order to prevent the emergence of resistant strains [14]. During this process, and once compounds have shown high in vitro potency, selectivity, specificity, low toxicity and good predictable pharmacokinetic/pharmacodynamic properties, a proof of concept that undoubtedly shows the in vivo efficacy of lead compounds, is required. Both mice and hamsters are used as models of acute and chronic VL, respectively, during the evaluation of the proof of concept [15, 16]. The most frequent technique to evaluate the infection course after drug treatment has been microscopic counting of amastigotes in liver, spleen and bone marrow smears stained with Giemsa dye. However, these are labour-intensive techniques that require specific skill training and present low sensitivity when parasite burdens are low after treatment [17], therefore, they are limited by tissue sampling biases, which require large animal cohorts.
In vivo real-time imaging combined with modified parasites expressing bioluminescent or fluorescent reporters may accelerate the initial stage of drug discovery at the preclinical level. Fluorescent reporters in the near infrared wavelength avoid interference with haemoglobin and do not require the addition of substrate. However, and despite the number of near infrared proteins currently available [18, 19] further reporters with longer emission wavelengths are still required in order to increase sensitivity. In this regard, red-shifted bioluminescent reporters [20] are currently allowing the appraisal of different infections produced by Trypanosomatids in vivo in real-time without the need to kill animals. This tool allows to run longitudinal studies with a reduced number of animals since they are not sacrificed, and in addition each animal is its own control, therefore the variability of experimental outcomes is limited [21–25]. In summary, in vivo real-time imaging allows to develop the proof of concept in a record time, accelerating the drug discovery process.
Nowadays, the mouse is used as acute preclinical model of VL, being liver the main affected organ when experimental treatments are initiated at early times postinfection. On the contrary, hamster is a more stringent and relevant model to recreate human VL [26, 27]. Generally speaking, during chronic infections the persistence of pathogens yields a state of T cell dysfunction known as exhaustion that is characterized by the loss of effector functions, low recall response and suboptimal T cell proliferation [28]. This is a hallmark feature shared by mice, dogs and humans and it is associated with disease progression [29–31].
Here, we describe a chronic murine model of VL that combines in vivo real-time image with stably modified strain expressing red-shifted luciferase (luc) aiming to track the presence of parasites in target organs during a long-time course of infection that can be used for preclinical drug-discovery. Miltefosine was used as proof of concept to assess the suitability of this technique during drug discovery.
The animal research described in this manuscript complies with Spanish Act (RD 53/2013) and European Union Legislation (2010/63/UE). The protocols were approved by the Animal Care Committee of the Centro de Biología Molecular Severo Ochoa (CBMSO, Madrid, Spain), project licence number JMJ/bb. Animals were maintained under specific pathogen-free conditions in individually ventilated cages. They experienced a 12 h light/dark cycle and had access to food and water ad libitum.
Seven to eight weeks-old female Balb/c mice were obtained from Janvier Labs (St Berthevin Cedex, France) and housed in specific pathogen-free facilities in the P2-facility of CBMSO for this study. L. infantum (strain MCAN/ES/96/BCN 150) promastigotes (previously obtained from infected dogs) were a gift from J.M. Requena (CBMSO, Madrid, Spain). Parasites were routinely cultured at 26 °C in M199 medium supplemented with 25 mM HEPES pH 6.9, 10 mM glutamine, 7.6 mM hemin, 0.1 mM adenosine, 0.01 mM folic acid, 1x RPMI 1640 vitamin mix (Sigma-Aldrich), 10% (v/v) heat-inactivated fetal calf serum (FCS) and antibiotic cocktail (50 U/ml penicillin, 50 μg/ml streptomycin).
The 1647-bp PpyRE9h coding region was amplified by PCR from pGEX-6P-2 HCO RE9h vector, a kind gift from Dr. Bruce Branchini, (Department of Chemistry, Connecticut College, CT, USA). The oligonucleotides used as primers (RBF919 and RBF920 in Table 1) introduced NcoI-NotI as restriction sites for cloning into pLEXSY-PAC vector (Jena Bioscience) and XhoI restriction site in the forward primer for cloning into pSK II vector. The 1647-bp PCR amplified fragment containing the PpyRE9h coding region was digested with XhoI and NotI and cloned first into pSK II vector previously cut with the same restriction enzymes to yield pSK-PpyRE9h plasmid. Then, this plasmid was cut with NcoI-NotI and the PpyRE9h ORF was cloned into pLEXSY-PAC vector to yield the pLEXSY-PAC-PpyRE9h construct. Parasites expressing red-shifted luc were obtained after electroporation of L. infantum BCN150 promastigotes with the linear SwaI-targeting fragment obtained from pLEXSY-PAC-PpyRE9h vector. Transfections were performed by electroporation (Gene Pulser X cell System, Biorad) using 10 μg of DNA fragments under the following conditions: 25 μF, 1500 v, 9 ms in 4 mm gap cuvettes. Subsequent plating on semisolid media containing 200 μg/mL puromycin as selection antibiotic, allowed the isolation of individual colonies that were subcultured in liquid media under antibiotic pressure. The correct integration of each fragment into the 18S rRNA locus of the resulting clones (PpyRE9h+L. infantum) was confirmed by PCR amplification analysis, using appropriate primers (Table 1).
The Luciferase Assay System (Promega) was used to assess luc expression in PAC-isolated clones from semisolid in vitro cultures. Briefly, 100×106 parasites were washed off with PBS and then lysed with 1 ml of Cell Culture Lysis Reagent provided by the manufacturer. Cell lysate was serially diluted in 96-well plate and then, ten microlitres of each cell lysate dilution were mixed with 90 μL of luciferin substrate. Luminescence was measured immediately using a Synergy HT microplate reader (BioTek).
In order to recover the infectivity of PpyRE9h+L. infantum strain after cloning, 108 metacyclic promastigotes were inoculated intravenously (IV) in the tail vein. Mice were sacrificed 4 weeks later and spleens were used to recover infective amastigotes. Amastigotes were isolated from the spleen by passing the tissue through a wire mesh. Then, splenocytes were disrupted by passing sequentially through 27G1/2 and 30G1/2 needles. Finally, cell debris was retained by passing successively through polycarbonate membrane filters with pore sizes of 8 μm, 5 μm and 3 μm (Isopore, Millipore). Released amastigotes (free of host cells), were washed twice with PBS (4000 x g for 20 min at 4°C) and counted by direct microscopy. To assess luciferase expression in intracellular infections, phorbol 12-myristate 13-acetate (PMA)-differentiated THP1 human monocytic leukemia cells, were grown on 8-well chambered coverslips (IBIDI) and incubated with infective amastigotes recuperated from mice in a ratio of 1:10 for further 4 hours. Extracellular parasites were removed by extensive washing with warm PBS and processed for immunofluorescence analysis 5 days after infection. Briefly, coverslips were fixed with 2% (v/v) paraformaldehyde in PBS and incubated with 0.1% (v/v) Triton X-100 in PBS for 10 min at room temperature in order to permeabilize the cells. Slides were then probed with 1:1000 red firefly luciferase polyclonal antibody (ThermoFisher); followed by 1:500 DyLight 633-conjugated goat anti-rabbit IgG secondary antibody for 30 min at room temperature. DNA was labelled using 1 μM Hoechst 33342 before mounting with Vectashield mounting medium. Images were acquired on a Zeiss LSM800 microscope with airyscan at 60X magnification.
Different infective doses of stationary phase PpyREh9+L. infantum promastigotes (ranging from 5x104 to 5x108) were IV injected to 6–8 weeks-old female Balb/c mice. Every week animals were placed in a Charge-Coupled Device (CCD) IVIS 100 Xenogen system (Caliper Life Science) for BLI analysis and images were acquired 10–20 min after intraperitoneal D-luciferin injection (150 mg/kg). Briefly, the animals were lightly anesthetized with 2.5% isofluorane (then reduced to 1.5%), before being placed on the camera. To standardize image capture and in order to allow comparison between mice, the images presented in the figures correspond to an acquisition time of 1 min duration, taken once luminescence plateaued.
To estimate the parasite burden in living mice, Regions Of Interest (ROIs) around liver and spleen in ventral and lateral animals’ positions were drawn using Living Image v.4.3 to quantify BLI expressed as radiance (p/s/cm2/sr). The detection threshold for in vivo imaging was estimated using uninfected mice placed in different positions (ventral and lateral), using ROIs of whole animals (n = 16).
At the end of some experiments, animals were euthanized and dissected, to confirm parasite burden by more conventional methods. Briefly, spleens and livers were reimaged ex vivo and used to quantify parasite load by Limit Dilution Assay (LDA). LDA was calculated as the geometric mean of the titer obtained from quadruplicate cultures x reciprocal fraction of the homogenized organ added to the first well. The titer was the reciprocal value of the last dilution in which parasites were observed [32].
Individual animal values were used as the unit of analysis of in vivo and ex vivo experiments. Statistical differences between groups were evaluated using t-student test using SigmaPlot v.14.0. Differences of P < 0.05 were considered significant.
PpyRE9h was stably integrated into the 18S rRNA promoter using pLEXSY vector (Fig 1A). Correct integration of reporter gene into the resulting clones (PpyRE9h+L. infantum) was confirmed by PCR amplification analyses using the primers of Table 1 (Fig 1B). The PCR-confirmed clones were screened for luciferase activity and those having higher activity were selected for in vivo experiments. Promastigote cultures of PpyRE9h+L.infantum grew at the same rate as wild-type parasites (Fig 1C). There was a linear relationship between the luciferase activity in vitro and the number of parasites independently of the parasite stage (logaritmic, metacyclic or freshly isolated amastigotes) and independently of the instrument for measuring (luminometer or IVIS camera). Fig 1D shows the relationship between luciferase activity and the number of logaritmic promastigotes (PpyRE9h+L.infantum and wild-type strains). Cell lysates from promastigotes were serially diluted into 96-well plate, D-luciferin was added and luciferase activity was measured using a luminometer (103–106 cells range, r2 = 0.999), being the detection limit of 103 promastigotes/well. The PCR-clone with the highest luciferase activity was selected for recovering infectivity through mouse (see Materials and Methods). Fig 1E shows the outcomes using metacyclic promastigotes and free amastigotes in the IVIS camera. In this experiment 6x106 parasites (metacyclic and amastigotes) were serially diluted in a 96-well plate, 100 μl D-luciferin were added and luciferase activity was measured in the IVIS camera (4.68x104-6x106 cells range, r2 = 0.987 for metacyclic and r2 = 0.994 for amastigotes). The detection limit was 9.37x104 metacyclic/well and 7.5x105 amastigotes/well. This suggests that more amastigotes than metacyclic promastigotes are required for their detection by the IVIS camera.
Spleen-isolated amastigotes were used to infect PMA differentiated THP-1 macrophages. Four hours later the non-phagocytosed parasites were gently washed off with warm PBS and left for further 96 h. The infection was stained using anti-firefly luciferase antibody and its localization was confirmed to be cytosolic by confocal microscopy (Fig 1F).
The infectivity of the selected clone was enhanced by passing through Balb/c mice that were successively infected with 108 promastigotes by IV route until spleen weight increased up to 0.7–1 g (4–5 passages through mice). To establish the in vivo sensitivity of the PpyRE9h+L. infantum strain, Balb/c mice were IV injected with different doses of infective metacyclic promastigotes (5x104-5x106) and photographed 1 h postinfection. At this time, the bioluminescent signal was detected in the liver but only with the highest doses (5x105 and 5x106). Forty-eight hours later when most promastigotes have transformed into amastigotes; BLI signal was only detected from mice infected with 5x106 parasites (Fig 2A). The appraisal of the infection showed that BLI signal in the liver peaked ~3 weeks post-infection (acute phase), then disappeared slowly from this organ and increased in the spleen (chronic phase) (Fig 2A). To estimate the in vivo limit of detection with PpyRE9h+L. infantum parasites, we used BLI signal expressed as radiance (p/s/cm2/sr). The limit of detection in vivo was estimated to be above 1x106 parasites at 48 h postinfection, when metacyclic parasites have transformed into intracellular amastigotes (Fig 2B).
We were interested in developing a chronic model of infection to use as proof of concept for well-established infections in spleen and bone marrow. In order to evaluate the stability of bioluminescent signal through time in chronic infections, 5x108 metacyclic promastigotes were IV injected and animals were photographed in ventral and lateral positions from 5 to 16 weeks post-infection. The spleen infection was detected independently of the animal positions. BLI signal was increasing from week 5 to reach the maximum radiance 12 weeks after infection (Fig 2C). Moreover, bone marrow radiance was detected only in ventral images from 8 to 16 weeks, and the bioluminescent signal was increasing during this time (Fig 2C).
To establish a correlation between BLI signal detected in vivo and the parasite burden in liver and spleen, mice (n = 12, one animal died before the end of the experiment) were infected with parasite dose ranging from 5x106-5x108. Nine weeks postinfection, animals were imaged and the luminescence was recorded in vivo in the regions of interest (ROI) previously drawn around the spleen and liver. Animals were euthanized and the liver and spleen processed to determine parasite burden. Both organs showed a good correlation with the in vivo recorded BLI signal (Fig 2D and 2E).
PpyRE9h+L. infantum infected mice were treated with miltefosine as a proof-of-concept to validate this model in a long-term follow-up infection.
Mice (n = 30) were IV infected with 5x108 metacyclic promastigotes and imaged for BLI after 3, 7, 12 and 14 wpi confirming that infection was established (Fig 3A). Animals were divided in groups and half of them were treated with miltefosine 40 mg/kg/day for 5 days by oral gavage. Once miltefosine treatment was ended, animals were imaged and sacrificed at different times (48 h, 1 and 6 weeks post-treatment that corresponded to 15, 16 and 21 wpi). Fig 3B (top panel) shows that BLI signal in whole animals was almost undetectable after miltefosine treatment (48 h post-treatment) in a chronic infection, and that the BLI reduction persisted for 6 weeks after the end of the treatment. Quantification of the BLI signal revealed that radiance in untreated animals (3x105 p/s/cm2/sr) decreased significantly to 2.32x104 and 1.61x103; one and six weeks, respectively after the end of the treatment, reaching BLI values similar to non-infected animals (Fig 3B; bottom panel P<0.05).
Animals were sacrificed at different times after the end of treatment (48h, 1 week and 6 weeks posttreatment) and the organs (spleen and liver) were photographed after injecting D-luciferin ex vivo (Fig 3C). There was a significant marked reduction in the weight of the spleen, which reached values similar to those of the uninfected animals at 6 weeks after the end of the treatment (Fig 3D P<0.001). Ex vivo bioluminescent values recorded from treated and untreated animals over the time were plotted (Fig 3E) confirming the BLI reduction seen in vivo. The BLI decrease was significant at all analysed times in both organs (P<0.001) with the exception of the liver at 48h posttreatment that was not significant and liver at 6wpi (P<0.01). Both organs showed logarithmic reduction of BLI from the end of the treatment to 6 weeks later. Ex vivo parasite burden was estimated using limiting dilution assay confirming parasite load reductions of 98%, 99,9%, and 99,9%, at 48h, 1 week and 6 weeks postinfection (Fig 3F).
The introduction of new medicines against VL from the initial concept to public release is a time-consuming and expensive process. Moreover, the clinical recurrences after treatment failure and the emergence of resistances are worsened by the shortage of new clinical entities and the long period needed to release a new medicine [33]. To bridge the gap between early drug identification and in vivo preclinical studies, new bioimaging tools have recently been introduced to accelerate the drug discovery process while drastically reducing the number of animals used.
To develop robust preclinical in vivo platforms, several aspects related to the genetic modifications introduced in the pathogen and the suitability of the animal model should be addressed before their validation with a proof of concept [34, 35]. In such a way, we present here the generation of the strain PpyRE9h+L. infantum and its utility to quantify the parasite load in vivo in infected mice in real time. As the virulence of the modified strain can be lost after genetic manipulation and passage in culture, as soon as the correct integration of the construct was confirmed, the selected clone was passed through mice to recover its infectivity [36, 37, 38].
Once D-luciferin was administered, the light detected by CCD camera and transformed to pseudocolor images, enabled parasite traceability in the whole body and the estimation of parasite burden in a murine model of chronic VL, reducing the number of animals to be analysed in longitudinal studies. During in vivo infections amastigotes enter into a semi-quiescent physiological stage in which major energetic processes are specifically repressed [39], explaining the differences in light emission between metacyclic promastigotes and freshly isolated amastigotes. However, our results show that parasites emitted light enough to provide accurate and rapid radiance that allow the appraisal up to 21 wpi. In this study, light could be detected in Balb/c mice in the liver during the acute phase of infection and later in the spleen and bone marrow during the chronic phase, allowing a continuous and long-term follow-up of the infection. Under these conditions, light detected in vivo–that corresponded to ROIs drawn around liver and spleen—correlated well with parasite burden calculated from LDA, which it would allow to estimate the parasite burden without the sacrifice of animals.
The location of parasites (peripheral or deeper tissues) within the mammalian host has been pointed as a key factor affecting the limit of parasite detection in vivo [40]. The light emitted by freshly isolated amastigotes from splenic lesions in our system showed a detection threshold similar to the previously reported by other authors [22].
In experimental VL, the hamster is considered the best experimental model since it reproduces many clinicopathological features of the human disease and can be fatal in the absence of treatment [41]. High-dose murine models of VL develop hallmarks of progressive human, primate, and canine disease with loss of gp38 stromal cells [42], remodelling of splenic marginal zone region [43], altered migration of DCs [42] and loss of follicular germinal centers [44]. For this reason, Balb/c mice have been proposed as an adequate model of chronic VL [45–46]. In addition, during chronic infections the persistence of pathogens yields a state of T cell dysfunction known as exhaustion that is characterized by the loss of effector functions, low recall response and suboptimal T cell proliferation [28]. In VL, this stage of T cell exhaustion is associated with disease progression in mouse, dogs and human infections [29–31]. For this reason and in order to have a murine BLI model of chronic VL, the inoculum size was increased to 5x108 metacyclic parasites per mouse.
In previous studies we have used the same L. infantum/Balb/c model showing hallmarks of progressive infection [47]. PpyRE9h+L. infantum strain allowed a continuous monitoring of parasite load from the beginning of the infection up to animal’s sacrifice, detecting both acute and chronic infections. In view of these results, PpyRE9h+L. infantum constitutes an ideal tool for the appraisal of drug efficacy in in vivo preclinical models.
The assay was validated by the treatment with miltefosine, starting 14 weeks post-infection and extended for long-term appraisal (6 weeks after drug withdrawal). In rodents miltefosine is known to produce significant parasite burden reduction (90–99% depending on parasite strain) in liver and spleen, along with no-sterile cure (when the treatment is initiated at 7–21 days postinfection [48–50]. These no-sterile curative results were confirmed later in hamster models of chronic VL treated with high-dose miltefosine (20 mg/kg/10 days), started 40 days post infection, although it resulted in 100% survival measured 20 wpi [23]. In our study, both radiance and parasite burden values dropped immediately after the end of treatment and they remained decreasing during the long-term follow up, although sterile cure was never achieved. A possible disease recurrence due to incomplete parasite suppression was expected. However, and despite the long-term follow up, much beyond the half-life of miltefosine [51], no recurrence was seen.
Several studies with AmBisome and stibogluconate in mice have shown that the infection status influenced treatment outcome, so that treatments were less effective in the chronic infection model than in acute infection models [52, 53]. The changes that occur in liver and spleen structure and function during early and late stages on infection might be the cause [46]. The mouse model proposed in this work would provide accurate information about potential drugs and their efficacy on the later stages of infection when it has been described that the efficacy of several drugs might be more compromised.
In conclusion, the gold standard methods used to evaluate the efficacy of antileishmanial drugs based on LDA or microscopic examination are laborious and time-consuming, have intrinsic variability, require intensive use of animals and cannot be monitored in real time. However, novel in vivo bioimaging models based on bioluminescent L. infantum parasites are highly sensitive, easily traceable, and yet provide statistically valuable outcomes with the use of far fewer animals than traditional methods. This technology is reproducible; less expensive because it reduces the number of animals needed, it is barely distressing for animals and can be easily adapted to different experimental models being thereby suitable to accelerate drug development.
|
10.1371/journal.pntd.0005828 | Local environmental and meteorological conditions influencing the invasive mosquito Ae. albopictus and arbovirus transmission risk in New York City | Ae. albopictus, an invasive mosquito vector now endemic to much of the northeastern US, is a significant public health threat both as a nuisance biter and vector of disease (e.g. chikungunya virus). Here, we aim to quantify the relationships between local environmental and meteorological conditions and the abundance of Ae. albopictus mosquitoes in New York City. Using statistical modeling, we create a fine-scale spatially explicit risk map of Ae. albopictus abundance and validate the accuracy of spatiotemporal model predictions using observational data from 2016. We find that the spatial variability of annual Ae. albopictus abundance is greater than its temporal variability in New York City but that both local environmental and meteorological conditions are associated with Ae. albopictus numbers. Specifically, key land use characteristics, including open spaces, residential areas, and vacant lots, and spring and early summer meteorological conditions are associated with annual Ae. albopictus abundance. In addition, we investigate the distribution of imported chikungunya cases during 2014 and use these data to delineate areas with the highest rates of arboviral importation. We show that the spatial distribution of imported arboviral cases has been mostly discordant with mosquito production and thus, to date, has provided a check on local arboviral transmission in New York City. We do, however, find concordant areas where high Ae. albopictus abundance and chikungunya importation co-occur. Public health and vector control officials should prioritize control efforts to these areas and thus more cost effectively reduce the risk of local arboviral transmission. The methods applied here can be used to monitor and identify areas of risk for other imported vector-borne diseases.
| This paper examines the ecological underpinnings of the invasive mosquito Ae. albopictus and the associated risk of arboviral transmission in New York City. We aim to quantify the relationships between local environmental and meteorological conditions and Ae. albopctus abundance. Further, we explicitly determine risk of local arbovirus disease transmission by Ae. albopictus by overlaying imported chikungunya cases from the epidemic year of 2014. Our overarching objective is to determine the extent of Ae. albopictus infestation and the distribution of viremic human hosts to predict risk of localized chikungunya outbreaks in New York City, and use these predictions to focus vector control and community education interventions to localities at greatest risk. We develop a model incorporating both local environmental and meteorological conditions to predict Ae. albopictus populations at fine spatial scale. We find that peak imported chikungunya cases and Ae. albopictus populations are temporally synchronous but primarily spatially asynchronous. The areas that do have high arboviral importation and Ae. albopictus populations should be prioritized for vector control and education interventions.
| Aedes albopictus Skuse 1984, also known as the Asian tiger mosquito, is an invasive mosquito of growing consequence and concern especially for temperate areas [1, 2]. Originating from Southeast Asia, this mosquito has expanded its range globally over the past three decades [3]. Its invasiveness is linked to its ability to exploit a range of container habitats, to lay desiccation resistant eggs that can survive without water for up to a year, and to oviposit eggs that hatch in installments [3]. In North America it was first observed in Texas in 1985 and its spread to the northeastern US was linked to the highway network [4]. To date there are over 500 counties in 34 states as well as the District of Columbia where Ae. albopictus has been reported [5, 6].
In the last two decades, the Americas have witnessed the emergence of a number of epidemic arboviruses of public health significance: Beginning in the 1990s the resurgence and spread of dengue (DENV), in 1999 the arrival of West Nile virus (WNV), and in 2013 the explosive spread of chikungunya (CHIKV). In the past year, the western hemisphere has experienced yet another arbovirus, Zika (ZIKV). These diseases incur significant costs to local economies and health care systems. Acute symptoms are typically not life-threatening; however, chronic conditions associated with these arboviruses are serious and in the case of the link between ZIKV and congenital microcephaly, particularly devastating.
Ae. aegypti readily transmits arboviruses to humans due to its anthropophilic biting tendencies; this vector lives in close proximity to humans and almost exclusively bite people. In contrast, Ae. albopictus is often considered a secondary vector of human arboviruses, because it inhabits a wider range of environments, including suburban and rural, and bites a wider variety of hosts, including birds [7]. These factors mitigate its transmission potential to humans. The principal argument cited for its secondary role is that in areas where it is present and Ae. aegypti is absent outbreaks are limited [8, 9]. However, the role of Ae. albopictus as a vector has not been fully elucidated across much of its range, particularly in places where it has recently been introduced, such as Europe (1979) and North America (1985) [5, 6, 10]. Its role may be secondary to Ae. aegypti; it may still be evolving; it may be the primary vector in more suburban and rural areas; it may be an important vector bridging sylvatic and urban cycles; or it may have an important role maintaining viruses between epidemics [11, 12]. It is also possible that Ae. albopictus behaves differently depending on its environment, whether urban, suburban or rural [13].
In its native range, Ae. albopictus mainly occurs in vegetated and rural habitats, especially where it co-occurs with Ae. aegypti [12]. However in areas where Ae. aegypti is absent, Ae. albopictus pullulates in urban areas [14]. As its range increases, Ae. albopictus appears to be more closely associated with humans [15]. Additionally, there is growing evidence that in human-dominated landscapes, Ae. albopictus favors humans, with 68–100% of blood meals taken from humans across nine studies recently reviewed [16]. Finally, its importance as a nuisance-biter further underscores its predilection for human blood when it is available [17, 18].
In temperate areas, where Ae. aegypti populations are limited by freezing temperatures, Ae. albopictus is the only endemic vector of DENV, YFV, CHIKV, and ZIKV. While temperate outbreaks occur they tend to be mild due to: the seasonality of mosquito populations limiting outbreaks at the onset of cold temperatures; sanitation services and piped water that reduce breeding habitats; infrastructural barriers, including screens and air conditioning that limit vector-host contact; and surveillance systems and other vector control resources that limit transmission if a local outbreak should arise [6, 19]. However temperate outbreaks do occur and may even be increasing in frequency.
Ae. albopictus has been implicated in the local spread of arboviruses in Asia, Europe, and the US. Ae. albopictus was responsible for frequent and widespread DENV epidemics in Japan during WWII, a DENV outbreak in Hawaii during 1943 [11], and DENV transmission in tropical regions of Asia until its displacement by Ae. aegypti in the 1950s [11]. More recently, Ae. albopictus was identified as the vector of the 2005–2007 CHIKV epidemic outbreak on La Reunion and in some of the outbreaks in India during the same time period [20]. In Europe, the first CHIKV outbreak occurred in Ravenna, Italy during 2007 with over 200 cases traced back to a single infected returning traveler and spread by established local populations of Ae. albopictus [21]. Subsequently, in France, local transmission of CHIKV by Ae. albopictus occurred in 2010 [22] and again in 2014 [23]. Ae. albopictus was also responsible for outbreaks of DENV in Asia: during 2001 and 2010 in China [24, 25] and 2014 in Japan [26]. In the US, Ae. albopictus mosquitoes caused a DENV outbreak in Hawaii during 2001 and a single locally acquired case in New York was attributed to Ae. albopictus in 2013 [27]. The recent invasion of Ae. albopictus in Gabon in 2007 was linked to the emergence of DENV, CHIKV, and ZIKV there [28].
In addition to the many arboviral outbreaks linked to Ae. albopictus, there are numerous other arboviruses that Ae. albopictus is known to carry, although its vectorial role remains largely un-described. Regardless, its broad viral susceptibility suggests that it may be implicated as an important, if not primary, vector in the transmission of other arboviruses now and in the future [11].
Even in the absence of disease transmission, infestation with Ae. albopictus may accrue negative health outcomes. In the eastern US, it has become the most common nuisance mosquito, aggressively biting humans during the day—so much so that it is a leading deterrent of outdoor recreation in cities [11, 17, 18, 29].
New York City (NYC) is a hub for international travel, which increases the chance of arbovirus introduction into local Ae. albopictus populations. There have been many arbovirus cases imported into New York: during 2014, 803 imported CHIKV cases representing 29% of all US imported cases, and during 2016, 1001 ZIKV cases representing 21% of all US imported cases [30, 31]. True importation rates are likely higher given the asymptomatic rates of these diseases (25% for CHIKV and 80% for ZIKV [32, 33]). Given this high rate of importation, it is logical to investigate whether the conditions necessary for local arbovirus transmission—the mosquito vector, the virus, and the ecological and epidemiological conditions suitable for transmission—co-occur in NYC. Our aims for this study are to identify the factors affecting Ae. albopictus abundance and the importation of arbovirus cases, and to use these findings to develop spatial-temporal risk maps that can inform vector control strategies.
The New York Department of Health and Mental Hygiene’s (NYC DOHMH) Office of Vector Surveillance and Control has 52 permanent mosquito surveillance sites spanning the five boroughs of NYC (S1 Fig). These 52 sites were established in 1999 after the introduction of WNV to NYC, and remained in operation each season from June 1st to October 31st. The trap locations and trap types deployed (gravid and light traps) are specifically targeted to collect WNV vectors (i.e. Culex mosquitoes). While not as effective as BG Sentinel traps for detecting the presence (especially low numbers) of Ae. albopictus [34, 35], these traps have been used to determine Ae. albopictus distribution and abundance [15, 36]. A recent study found BG and CDC light traps baited with dry ice like those in NYC to have equivalent Ae. albopictus trapping efficiency [36]. Weekly data from the light and gravid traps were combined as has been done previously to reduce bias and increase the power of analysis [15].
Our modeling approach exploits links between meteorological and local environmental factors and Ae. albopictus populations in the northeastern US (see Supporting Information). To measure temporal differences in meteorological factors in NYC we used the North American Land Data Assimilation System (NLDAS) dataset, a combined NASA/NOAA product, which provides gridded estimates of near-surface meteorological conditions at 13 km x 13 km spatial resolution [37]. Hourly estimates of precipitation measured in millimeters per hour, temperature measured in Kelvin 2-m above ground, and specific humidity measured in kilograms per kilograms 2-m above ground were used to calculate monthly averages for the years 2006–2016.
To measure fine-scale spatial differences in the urban environment we used 3 foot spatial resolution land cover data [38]. This land cover dataset defines 7 land cover classes (trees, grass, bare, building, road, other paved, and water). We further calculated the Shannon diversity index (SDI) at the same 3 foot spatial resolution, which provides an estimate of environmental heterogeneity accounting for both the total proportional area of each land cover class (abundance) as well as the number of land cover classes present (evenness):
S D I = ∑ i = 1 R p i l n ( p i ) (1)
where the proportion of land cover class i relative to the total number of classes (pi) is multiplied by the natural logarithm of this proportion (lnpi), summed across classes, and multiplied by −1.
To determine the area covered by one or two family residential buildings, open spaces, and vacant lots we used data from PLUTO, a geographically registered dataset created by the Department of City Planning at the tax lot level for the city of New York [39]. We created raster grids of the PLUTO data at the same spatial resolution as the land cover classes.
We calculated the proportion of each of the 11 environmental variables (7 land cover, SDI, and 3 PLUTO) within 200m of every pixel in the mapped domain representing NYC. Because Ae. albopictus has a flight range under 200m [40], each pixel (which supplies an accounting of each of the 11 environmental variables within the 200m radius) provides a synopsis of the environmental conditions Ae. albopictus would be exposed to if present at that location in NYC. Next, we standardized these values by subtracting the mean and dividing by the standard deviation across the whole domain [41]. We extracted the standardized values at each of the 52 permanent trap locations to estimate local environmental conditions in order to model annual Ae. albopictus abundance.
The surveillance data provide a record of the invasion and establishment of Ae. albopictus in NYC. This mosquito was first trapped in the Bronx during 2000, between 2000 and 2005 was caught in increasing trap numbers across the city, and between 2006 and 2016 was caught in over 96% of traps. We thus restricted our analysis to the period after invasion from 2006 to 2016. Between 2006 and 2016, 61,977 Ae. albopictus mosquitoes were caught in gravid and light traps across the 52 permanent trap locations. In 2016, BG Sentinel traps were added to the 52 permanent trap locations, trap counts from these BG Sentinel traps and the CDC light traps were significantly correlated (r = 0.21; p< .001). The annual numbers of traps collecting Ae. albopictus (traps positive), the total Ae. albopictus mosquitoes caught in gravid and light traps, and the abundance (calculated as the number caught per trap location divided by the 23 weeks of surveillance) for gravid, light, and both trap types together are shown in Table 1 and Fig 1.
The subset of important parameters from the spatial and temporal modeling efforts include February specific humidity, April precipitation, June temperature, and June precipitation, as well as the extent of residential buildings, open spaces, vacant lots, water, and grass. With these nine variables we fit generalized linear negative binomial models using all combinations of these variables. Of those tested, 137 were significant and 10 were included in the ensemble model set (Table 2; Fig 2).
The temporal ensemble model predictions (made using monthly mean estimates of meteorological conditions) shows broad confidence intervals that are similar across all 52 permanent trap locations (Fig 3). This near uniformity is due to the small differences in meteorological conditions within NYC (S3 Fig). We used root mean squared error (RMSE) to compare the accuracy of the temporal, spatial, and spatiotemporal model predictions with the observed values for 2016. RMSE is largest for temporal ensemble predictions (2.58), followed by spatial ensemble predictions (2.25), and lowest for spatiotemporal ensemble predictions (1.75). RMSE for the LOOTCV model spanning all 11 years of analysis (RMSE = 2.38) and the full spatiotemporal model (RMSE = 2.34) predictions were comparable (S4 Fig), indicating that out-of-sample prediction is possible and that no single year overly dominates the model structure. Further we test the sensitivity and specificity of the spatiotemporal ensemble model predictions. We use the mean value of both the observed and predicted values (2.37) as the cut-off point for the analysis. we find that the sensitivity (to truly predict above average observed values) is 69% and the specificity (to truly detect below average observed values) is 77%.
To map predicted Ae. albopictus abundance for 2016 across NYC at fine spatial resolution we used the ensemble coefficient estimates from the spatiotemporal modeling effort and the surface raster grids created for each parameter (Fig 4, Panel I; S5 Fig). Ae. albopictus are predicted to be most abundant in parts of Staten Island, and southern Brooklyn and Queens.
During 2014 both imported CHIKV cases and Ae. albopictus abundance peaked in August suggesting that epidemic risk coincided temporally with mosquito abundance. In Fig 4 (Panel II) the spatial distribution of imported CHIKV cases is presented by zipcode. Zipcodes with higher risk are in northern Manhattan and the Bronx. Overlaid are the results from the spatiotemporal Poisson probability model run in SatScan (Fig 4, Panel II, bottom). Through this analysis we find a significant cluster of imported CHIKV cases between the months of July and October across 28 zipcodes verifying increased risk in upper Manhattan and the Bronx.
Using the mean predicted values of Ae. albopictus annual abundance by zipcode in conjunction with the distribution of imported CHIKV cases from 2014 we are able to ascribe risk for local transmission in NYC. We find that the distribution of imported CHIKV cases and areas of high Ae. albopictus abundance are mainly discordant; however, there are some areas of concordance, including parts of southern Queens in the vicinity of John F. Kennedy airport, as well as the Bronx (Fig 4; Panel III). These delineated areas of higher risk should inform vector control and public health personnel where to target control for Ae. albopictus-borne disease.
Here, we examined the separate temporal and spatial influences, as well as the combined spatiotemporal influences, on annual Ae. albopictus abundance in NYC using ensemble modeling methods. We find that spatial variability is greater than temporal variability, suggesting that local environmental conditions are a stronger determinant of Ae. albopictus abundance than inter-annual differences in meteorological conditions. This may be due to a general availability of hospitable meteorological conditions in NYC (S3 Fig) or may reflect the finer spatial resolution of the local environmental conditions compared to that of the meteorological data used in the analysis. Taken at face value, this finding underscores a greater importance of local environmental predictors over meteorological effects on annual Ae. albopictus abundance. However the improvement of model fit with the inclusion of both meteorological and environmental conditions indicates the importance of both for predicting annual Ae. albopictus abundance.
Meteorological conditions in the spring and early summer (February specific humidity, April and June precipitation, and June temperature) positively influence Ae. albopictus abundance. Higher February specific humidity indicates wetter, warmer conditions in February—conditions that may improve survivorship of overwintering eggs. April and June precipitation may increase container habitat for Ae. albopictus, leading to an increase in overall annual Ae. albopictus abundance. The influence of early season rainfall may be because rainfall early in the season is more directly linked to Ae. albopictus production than rainfall later in the breeding season which is decoupled from mosquito production by human watering activities [51]. Warmer temperatures in early summer, i.e. June, may lead to an acceleration of Ae. albopictus reproduction early in the season which may in turn lead to higher annual numbers. The importance of early season meteorological conditions suggests that annual predictions can be made before Ae. albopictus populations peak in NYC, which may help vector control initiatives target and reduce these pestiferous mosquitoes.
Of the environmental parameters tested we find that open spaces, residential areas, vacant lots, water, and grass influence Ae. albopictus abundance. The land use classifications (residential, open spaces, and vacant lots) were more important than individual land cover categories or the SDI in predicting Ae. albopictus abundance. Land use classifications depict a particular configuration of land cover types. Within open spaces, mainly parks in NYC [39] we find 35% of the area is trees, 39% grass, and only 1% buildings. These areas had a negative influence on the annual abundance of Ae. albopictus. Vacant lots have a similar composition of trees (25%), grass (37%), and buildings (7%) as open spaces albeit with fewer trees and more buildings. In contrast to open spaces, vacant lots had a positive influence on annual Ae. albopictus abundance. This difference may be explained by how humans engage with these different land use classifications. Unlike open spaces, vacant lots tend to be unmanaged areas where weedy vegetation is left and trash accumulates; characteristics noted by others to be associated with higher Ae. albopictus infestation [52–55].
Residential areas in NYC have a more equitable distribution of trees (21%), grass (15%), and buildings (25%) and were the most important environmental parameter predicting high Ae. albopictus populations. Residential areas and vacant lots, likely have more available containers than open spaces; however, the types of containers may differ substantially between areas designated as residential or vacant lots; with more permanent water holding containers more closely linked with human watering in residential areas compared to more discarded water holding containers more closely linked with rainfall in vacant lots [56]. The positive influences of both residential and vacant areas on annual Ae. albopictus populations suggest that habitat requirements are met in these locations. Further, because these mosquitoes do not travel far during their lifetimes, this indicates that the habitat requirements of this vector are met at both immature and adult life stages. While water holding containers suitable for mosquito development are present in both environments, the types of containers likely differ substantially [54, 57]. Thus when it is dry Ae. albopictus populations may only flourish in residential areas and future analysis should investigate the interactive effects of meteorological conditions, in particular precipitation, with land use classifications to further examine the influence of sociological processes on Ae. albopictus populations.
The data used for this analysis are somewhat limited by the trap types and locations of collection. The 52 permanent trap locations were installed in 1999 after the introduction of WNV in NYC and both the trap locations and trap types deployed are specifically targeted to collect WNV vectors (i.e. Culex mosquitoes). The trap types, light and gravid traps are not as well suited to capture Ae. albopictus compared to other traps such as BG Sentinel traps. BG Sentinel traps were deployed for the first time in 2016, and the results of this analysis can be used to further inform placement of BG Sentinel traps to areas predicted to have high Ae. albopictus populations. 70% (n = 37) of the permanent trap locations are within park boundaries in NYC. While other researchers have found that small green islands within urban areas are hot spots for Ae. albopictus and disease transmission [26, 58, 59], the results of this analysis suggest that residential areas are likely to have higher Ae. albopictus populations than park land in NYC. Thus, while the current surveillance provides an important time series of annual Ae. albopictus abundance, an expansion of trap locations to reflect local environmental conditions that favor Ae. albopictus such as in residential areas and vacant lots may provide better population estimates.
While the spatiotemporal ensemble model predictions for 2016 capture the range of observations, they overestimate annual Ae. albopictus abundance when observations are low and underestimate annual Ae. albopictus abundance when observations are high (Fig 3). This limitation may be due to the spatial or temporal scales on which we based our measurements. Indeed, the predictive capability of the spatiotemporal model may be improved by incorporating measures of meteorological and environmental conditions at different scales. However, the sensitivity and specificity tests support the ability of the model to distinguish between above and below average years of Ae. albopictus production which is important for informing vector control initiatives.
In evaluating the risk of local arbovirus transmission, we find that the distribution of Ae. albopictus and imported CHIKV cases is temporally aligned (Fig 4, Panel II, bottom) but primarily spatially discordant, which provides a check on local transmission in NYC. However, we do identify locales at higher risk (Fig 4, Panel III), which should provide guidance for future vector surveillance and control as well as public health educational campaigns. The distribution of imported DENV and ZIKV cases should be compared to the CHIKV cases mapped here to determine any similarities or differences in the distribution of imported arboviruses across NYC and assess if the spatiotemporal distribution of imported CHIKV case is suitable for ascribing overall risk of arboviral introduction into local Ae. albopictus populations.
Local transmission of CHIKV by Ae. albopictus has not been reported in NYC likely due to a combination of the strain currently circulating in the western hemisphere and socioeconomic conditions in the northeastern US that limit vector-host contact rates. The CHIKV strain circulating in the western hemisphere belongs to the Asian lineage, while local CHIKV transmission by Ae. albopictus in temperate Europe is linked to the CHIKV variant (E1—226V) which is more readily transmitted by Ae. albopictus [60–62]. A future introduction of the E1—226V variant might thus lead to local CHIKV transmission in the northeastern US by Ae. albopictus.
In temperate areas, Ae. albopictus is the only endemic vector of CHIKV as well as DENV and ZIKV. Its broad viral susceptibility suggests that it may be implicated as an important, if not primary, vector in the transmission of other arboviruses now and in the future [11]. Blood titers from imported human cases have documented levels sufficient to infect endemic mosquito vectors [63]. Therefore, the introduction of just one case could trigger a local outbreak [21], especially if vector densities are high [64, 65]. In the northeastern US, because human population density and susceptibility are high and the population is unfamiliar with protective behaviors, arboviruses could spread quickly [1].
Socioeconomic factors, in particular, window screens and access to air conditioning (AC) that limit vector-host contact rates, have restricted the temperate spread of mosquito borne disease in the US [66, 67]. While these barriers are typically sufficient against vector borne diseases in the US, their distribution remains inequitable and their permanence is not guaranteed. In NYC access to AC is variable, with up to 40% of senior citizens in areas of Brooklyn and the Bronx reporting no access [68]. Further analysis could incorporate social risks such as these to better focus vector control and public health education efforts. Additionally, climate-related extreme weather events are expected to produce increased damage to infrastructure and power outages, which could significantly alter mosquito-human contact rates.
Ae. albopictus is a pestiferous mosquito that reduces outdoor use and effectively transmits a number of emergent arboviruses [4]. Currently there are no vaccines or treatments available for these arboviruses. Limiting disease transmission still hinges on effective vector control, which depends on removal and/or regular maintenance of containers, efforts that require concerted, coordinated efforts between vector control officers and communities. Entomological surveillance records widespread and abundant Ae. albopictus populations in NYC (Table 1; Fig 1) despite ongoing vector control efforts. Because these mosquitoes are so difficult to control informed, targeted vector control efforts are essential. To this end, we have identified key meteorological and local environmental conditions associated with Ae. albopictus abundance, developed spatiotemporal models of Ae. albopictus, and generated spatially explicit forecasts of this risk in NYC. By overlaying the spatiotemporal ensemble model of Ae. albopictus abundance with potential arbovirus introduction risk as determined by the spatiotemporal distribution of imported CHIKV cases in 2014, we delineate fine scale spatial differences in local arbovirus transmission risk in NYC that may be used to guide vector control and public health educational campaigns.
|
10.1371/journal.pgen.1002350 | Small RNAs Prevent Transcription-Coupled Loss of Histone H3 Lysine 9 Methylation in Arabidopsis thaliana | In eukaryotes, histone H3 lysine 9 methylation (H3K9me) mediates silencing of invasive sequences to prevent deleterious consequences including the expression of aberrant gene products and mobilization of transposons. In Arabidopsis thaliana, H3K9me maintained by SUVH histone methyltransferases (MTases) is associated with cytosine methylation (5meC) maintained by the CMT3 cytosine MTase. The SUVHs contain a 5meC binding domain and CMT3 contains an H3K9me binding domain, suggesting that the SUVH/CMT3 pathway involves an amplification loop between H3K9me and 5meC. However, at loci subject to read-through transcription, the stability of the H3K9me/5meC loop requires a mechanism to counteract transcription-coupled loss of H3K9me. Here we use the duplicated PAI genes, which stably maintain SUVH-dependent H3K9me and CMT3-dependent 5meC despite read-through transcription, to show that when PAI sRNAs are depleted by dicer ribonuclease mutations, PAI H3K9me and 5meC levels are reduced and remaining PAI 5meC is destabilized upon inbreeding. The dicer mutations confer weaker reductions in PAI 5meC levels but similar or stronger reductions in PAI H3K9me levels compared to a cmt3 mutation. This comparison indicates a connection between sRNAs and maintenance of H3K9me independent of CMT3 function. The dicer mutations reduce PAI H3K9me and 5meC levels through a distinct mechanism from the known role of dicer-dependent sRNAs in guiding the DRM2 cytosine MTase because the PAI genes maintain H3K9me and 5meC at levels similar to wild type in a drm2 mutant. Our results support a new role for sRNAs in plants to prevent transcription-coupled loss of H3K9me.
| Methylation of histone H3 at the lysine 9 position (H3K9me) is a fundamental chromatin modification that suppresses expression from invasive and repetitive sequences such as transposons. In plant genomes, regions modified by H3K9me are maintained with precise boundaries. However, at junctions where H3K9me target regions are subject to read-through transcription from outside promoters, the stability of H3K9me patterns is jeopardized by transcription-coupled processes that remove this modification. We show that maintenance of H3K9me patterns at such vulnerable sites requires small RNAs corresponding to the H3K9me target region. We use a sensitive reporter system to show that, in the absence of small RNAs, target regions subject to read-through transcription undergo an immediate reduction in H3K9me levels, followed by further losses in progeny plants upon inbreeding. Our results support a new function for small RNAs in maintaining accurate H3K9me patterns in the plant genome.
| The eukaryotic cell is under constant threat from transposons and other invasive sequences. Transposons can drain cellular resources for RNA and protein synthesis and can damage the cell through expression of aberrant gene products or activation of transposon movement. A major mechanism to protect against these deleterious effects is to target transposons and other repetitive sequences for silencing mediated through chromatin modifications. In most eukaryotes, transposon chromatin is marked by methylation of histone H3 at the lysine 9 position (H3K9me). In some eukaryotes including mammals and plants transposon chromatin is also marked by cytosine methylation (5meC). An important question is how H3K9me and 5meC are accurately maintained on transposons but not on host genes.
A conserved strategy to maintain H3K9me and 5meC is to use the modifications as methyltransferase (MTase) binding recognition motifs. For example, in Arabidopsis thaliana, dimethylation of H3K9 (H3K9me2) maintained by three partially redundant histone MTases—SUVH4 (also known as KYP, At5g13960), SUVH5 (At2g35160), and SUVH6 (At2g22740)—is associated with 5meC maintained by the CMT3 cytosine MTase (At1g69770) [1]–[4]. The SUVH MTases contain a 5meC binding domain and CMT3 contains an H3K9me binding domain, suggesting that the SUVH/CMT3 pathway involves an amplification loop that can perpetuate both H3K9me and 5meC [5], [6]. Consistent with this model, mutations in the CMT3 or MET1 (At5g49160) cytosine MTases, which act to maintain 5meC in non-CG and CG sequence contexts respectively, result in reduced H3K9me2 levels on transposons and repetitive sequences [3], [7]–[10]. In addition, a suvh4 suvh5 suvh6 triple H3 K9 MTase mutant displays similar reduced non-CG methylation patterns to a cmt3 mutant [4]. Although the H3K9me/5meC amplification loop provides a mechanism to stably maintain both modifications in untranscribed regions of the genome, at junctions where modified sequences are transcribed through from nearby unmodified promoters, H3K9me can be removed by transcription-associated histone replacement or histone demethylation [7], [11]. What prevents transcriptional destabilization of H3K9me patterns?
Duplicated Arabidopsis genes encoding the tryptophan synthesis enzyme phosphoribosylanthranilate isomerase (PAI) provide an ideal system to understand the balance between transcription and SUVH-mediated H3K9me2/CMT3-mediated 5meC. In most Arabidopsis strains there are three unlinked PAI gene duplications that lack 5meC [12]. However, in the Wassilewskija (Ws) strain one of the PAI loci is rearranged as a tail-to-tail inverted repeat (IR) of two genes PAI1–PAI4 (At1g07780), which triggers the recognition of PAI sequences as invaders. The PAI1–PAI4 IR as well as two unlinked singlet genes PAI2 (At5g05590) and PAI3 (At1g29410) are modified by H3K9me2 and 5meC, coextensive with their regions of shared sequence identity [3], [13] (see Figure S1 for PAI gene maps). The PAI1–PAI4 IR is fused to a heterologous promoter with a transcription start site approximately 500 base pairs (bp) upstream of the PAI1 5meC boundary, which drives constitutive expression of PAI1 transcripts [14]. The polyadenylated transcripts that accumulate from this locus consist of a majority class that terminates normally in the PAI1 3′ untranslated region at the center of the IR and a minority class that extends through PAI1 into palindromic PAI4 sequences to provide a source of fold-back double-stranded RNA (dsRNA). Therefore the PAI1–PAI4 locus is able to stably maintain H3K9me2 and 5meC on the IR sequences even in the face of substantial read-through transcription. The PAI2 and PAI3 singlet genes also stably maintain H3K9me2 and 5meC even though they are likely to be only partially silenced by limited upstream modifications: at PAI2 5meC extends only 250 bp upstream of the predicted transcription start site, and at PAI3 5meC extends only as far as the predicted transcription start site [12], [13].
Arabidopsis uses three cytosine MTase pathways to control 5meC: the CMT3 pathway maintains 5meC mainly in non-CG contexts in conjunction with the SUVH H3K9 MTases, the MET1 pathway maintains 5meC mainly in CG contexts, and the DRM2 (At5g14620) pathway initiates 5meC on new invasive sequences under the guidance of small RNAs (sRNAs), as well as contributing to maintenance of non-CG methylation at some loci [15]. In a cmt3 or a suvh4 suvh5 suvh6 mutant, the Ws PAI genes are depleted for 5meC in non-CG contexts [4], [16]. In addition, in a cmt3 met1 double mutant the PAI genes are depleted for 5meC in all contexts [3]. Therefore, the DRM2 pathway plays a minimal role in the maintenance of PAI 5meC patterns. However, genetic or epigenetic changes that impair the production of transcripts that read through from PAI1 into palindromic PAI4 sequences at the PAI1–PAI4 IR cause reduced levels of PAI 5meC in non-CG contexts [13], [14], [17], [18]. In light of these results, we hypothesized that sRNAs processed from dsRNAs might underlie a mechanism to prevent the loss of SUVH/CMT3-mediated modifications due to read-through transcription, independently of the role for sRNAs in guiding DRM2.
To test the hypothesis that sRNAs control the SUVH/CMT3 pathway, we used mutations in Arabidopsis dicer-like (DCL) ribonucleases to block processing of sRNAs from dsRNAs, and monitored the effects on Ws PAI gene H3K9me2 and 5meC levels. Arabidopsis encodes four DCLs (reviewed in [19]). DCL1 (At1g01040) is specialized for processing 21 nucleotide (nt) microRNAs (miRNAs) needed for developmental gene regulation, whereas DCL2 (At3g03300), DCL3 (At3g43920), and DCL4 (At5g20320) have partially redundant roles in processing sRNAs used in other silencing pathways. For example, DCL3 processes 24 nt sRNAs used to guide DRM2 to matching target sequences such as transgene insertions and transposons [20], [21]. In a dcl3 mutant DCL2 and DCL4 can partially compensate by processing 22 nt and 21 nt sRNAs respectively corresponding to the same genomic target sequences [22], [23].
Here we show that the dcl2 dcl3 dcl4 mutant has reduced levels of H3K9me2 and non-CG methylation on PAI sequences relative to wild type, corresponding to loss of PAI sRNAs. We also show that a drm2 mutant maintains similar levels of PAI H3K9me2 and 5meC relative to wild type. Therefore the PAI genes illustrate that DCL-dependent sRNAs help maintain SUVH/CMT3-mediated modifications through a distinct mechanism from their role in guiding DRM2. In the dcl mutant there is a weaker reduction in PAI 5meC levels but a similar or stronger reduction in PAI H3K9me2 levels compared to a cmt3 mutant, indicating a connection between sRNAs and maintenance of H3K9me2 patterns independent of CMT3 function. We also show that upon inbreeding in the absence of DCL function, remaining PAI 5meC is destabilized. Our results reveal a new pathway for sRNA control of H3K9me2 and associated 5meC patterns in plants. This pathway provides a homeostatic mechanism to use a product of read-through transcription—sRNAs—as a means to counteract transcription-coupled loss of H3K9me2 on transposons and repeats.
To test whether dicer-dependent sRNAs contribute to maintenance of PAI 5meC patterns, we generated strains where dicer mutations were combined with the three methylated PAI loci from Ws and assayed PAI 5meC patterns using both DNA gel blot and bisulfite sequencing assays. For PAI DNA gel blot analysis we cleaved genomic DNA with each of three 5meC-sensitive restriction enzymes that have cleavage sites within methylated PAI sequences: HincII (sensitive to methylation of the outermost non-CG cytosines in 5′ atGTCAACag 3′, where the recognition sequence is shown in uppercase), MspI (sensitive to methylation of the outer non-CG cytosines in 5′ CCGG 3′, and HpaII (sensitive to methylation of either the inner CG or outer non-CG cytosines in 5′ CCGG 3′). HincII cleaves at the translational start codons of PAI1, PAI2 and PAI4 but not PAI3, and the MspI/HpaII isoschizomers cleave in the second introns of PAI2, PAI3, and PAI4 but not PAI1 (Figure S1).
We found that genomic DNA prepared from dcl2, dcl3, and dcl4 single insertional null mutants and the dcl2 dcl4 and dcl3 dcl4 double mutants had similar PAI cleavage patterns to wild type Ws genomic DNA when assessed by HincII, MspI, or HpaII DNA gel blot assays (Figure 1). In contrast, genomic DNA prepared from the dcl2 dcl3 and dcl2 dcl3 dcl4 mutants displayed increased cleavage with HincII at PAI1–PAI4 and PAI2, and with MspI at PAI1–PAI4, PAI2, and PAI3 relative to wild type Ws, diagnostic of partially reduced non-CG methylation levels at all three PAI loci. Bisulfite sequencing of PAI1 and PAI2 proximal promoter/first exon regions in the dcl2 dcl3 dcl4 mutant compared to wild type Ws and Ws cmt3 showed that there was a partial loss of 5meC in CHG and CHH contexts. Therefore, the bisulfite sequencing data are consistent with the DNA gel blot assays. The results indicate that DCL2 and DCL3 act redundantly to maintain PAI non-CG methylation patterns.
To determine whether the miRNA processing dicer DCL1 contributes to the remaining PAI non-CG methylation in the dcl2 dcl3 dcl4 triple mutant relative to cmt3, we included a dcl1 dcl2 dcl3 dcl4 quadruple mutant strain in the DNA gel blot analysis (Figure 1). Because dcl1 null alleles are embryo-lethal we used a partial-function dcl1-9 allele that is viable but female-sterile [24], [25]. The dcl1 dcl2 dcl3 dcl4 mutant displayed similar cleavage patterns to the dcl2 dcl3 dcl4 mutant, indicating that the dcl1-9 mutation does not enhance the partial loss of 5meC conferred by mutation of the other three DCL genes. In subsequent studies we focused on the dcl2 dcl3 dcl4 mutant, which has global depletion of sRNAs other than miRNAs [23].
For comparison to the dcl mutants we included genomic DNA prepared from cytosine MTase mutants in the Ws background (Figure 1). DRM2 is the major cytosine MTase controlling initiation of 5meC, but the related DRM1 MTase (At5g15380) could also contribute to this pathway [26]. Therefore we used a drm1 drm2 double null insertional mutant. DNA from the Ws drm1 drm2 mutant displayed similar PAI cleavage patterns to wild type Ws in all three DNA gel blot assays, and similar 5meC patterns to wild type Ws in bisulfite sequencing analysis of PAI1 and PAI2 proximal promoter regions. DNA from a Ws met1 mutant displayed increased cleavage at all three PAI loci with HpaII, and partially increased cleavage at PAI1–PAI4 and PAI2 with MspI, but no difference from wild type PAI cleavage patterns with HincII in DNA gel blot assays, diagnostic of a partial loss of 5meC in CG and CCG contexts. DNA from the cmt3 mutant displayed nearly complete cleavage with HincII and MspI, and partially increased cleavage with HpaII in DNA gel blot assays, diagnostic of strong loss of 5meC mainly in CHG and CHH contexts. Compared to the cytosine MTase mutants across the three DNA gel blot assays, the dcl mutant PAI demethylation phenotypes are consistent with a partial defect in the SUVH/CMT3 pathway rather than a defect in the DRM or MET1 pathways.
DCL3 and DRM2 are key factors in establishing new 5meC imprints. We used a previously developed genetic assay combining the Ws PAI IR dsRNA source locus with an unmethylated PAI2 target gene from another strain background to show that dcl3 and drm1 drm2 mutations impair the acquisition of new 5meC on PAI2 (Text S1, Figure S2). Therefore the PAI genes use the same DCL3/DRM pathway for establishing 5meC imprints as other characterized loci. However, once PAI 5meC patterns are established, the DCL3/DRM pathway plays a minimal role in long-term maintenance (Figure 1).
We used chromatin immunoprecipitation (ChIP) analysis with H3K9me2-specific antibodies on chromatin prepared from the dcl2 dcl3 dcl4 mutant compared to chromatin prepared from wild type, suvh4 suvh5 suvh6, cmt3, drm1 drm2, or cmt3 drm1 drm2 strains to determine whether the dcl mutations affect levels of H3K9me2 as well as non-CG methylation on PAI sequences. Chromatin was analyzed by quantitative PCR with primer pairs specific for the PAI1 arm of the PAI1–PAI4 IR locus or the PAI2 singlet gene.
At both PAI1–PAI4 and PAI2 the dcl2 dcl3 dcl4 mutant had reduced levels of H3K9me2 relative to wild type, although not as strongly as in the suvh4 suvh5 suvh6 H3K9 MTase mutant (Figure 2). Comparing the ChIP results to the assays for 5meC (Figure 1), the reduced PAI H3K9me2 levels in the dcl2 dcl3 dcl4 mutant are still sufficient to support substantial CMT3 activity. Therefore CMT3 might be able to use even sparsely distributed H3K9me2 as a localization signal.
At both PAI loci the cmt3 mutant also had partially reduced levels of H3K9me2, presumably because reduced 5meC levels impair SUVH localization to PAI sequences. At PAI2 increased transcription due to proximal promoter demethylation in a cmt3 mutant could also contribute to reduced H3K9me2 levels, perhaps accounting for a stronger relative reduction at PAI2 than at PAI1–PAI4 [3], [16]. The dcl2 dcl3 dcl4 mutant had similar reduction in H3K9me2 levels to the cmt3 mutant at PAI2, but a stronger reduction at PAI1–PAI4. In contrast, the dcl2 dcl3 dcl4 mutant had weaker reductions in PAI2 and PAI1–PAI4 non-CG methylation levels compared to cmt3 (Figure 1). This comparison indicates that the dcl2 dcl3 dcl4 mutations impair maintenance of H3K9me2 independently of effects on CMT3 function. If the dcl mutations acted by impairing CMT3 to cause a partial reduction in PAI 5meC levels as the primary consequence, then the resulting reduction in H3K9me2 levels would be expected to be less than in the cmt3 mutant.
At both PAI1–PAI4 and PAI2, the drm1 drm2 mutant displayed similar levels of H3K9me2 to wild type, and the drm1 drm2 cmt3 mutant displayed similar levels of H3K9me2 to cmt3 (Figure 2). Therefore, the DRM cytosine MTases do not contribute to maintenance of PAI H3K9me2 patterns.
To determine whether reduced PAI non-CG methylation and H3K9me2 levels in dcl2 dcl3 dcl4 correlate with loss of PAI sRNAs, we used RNA gel blot analysis to detect PAI sRNAs (Figure 3). As a negative control we used a mutant derivative of Ws, Δpai1–pai4, where the PAI1–PAI4 IR source of dsRNA DCL substrates has been deleted by homologous recombination between flanking direct repeat sequences [17]. As a positive control we used the Δpai1–pai4 strain transformed with a PAIIR transgene consisting of an IR of approximately 700 bp of PAI cDNA sequences transcribed by the strong constitutive Cauliflower Mosaic Virus 35S promoter [14]. We previously determined that PAI sRNAs could be detected in the Δpai1–pai4(PAIIR) transgenic strain but not in wild type Ws using RNA gel blot analysis with a PAI cDNA riboprobe. Furthermore, high-throughput sRNA sequencing in the C24 strain that has a similar PAI1–PAI4 IR to Ws detected PAI sRNAs at low levels [27]. We therefore optimized detection of rare PAI sRNAs by designing a high-affinity locked nucleic acid (LNA) probe corresponding to the sense strand of a 35 nt sequence in the PAI fifth exon.
In wild type Ws the LNA probe detected low levels of PAI sRNA species of both shorter and longer sizes, between 21 and 24 nt relative to size markers, above the background signal in the Δpai1–pai4 negative control strain (Figure 3). This pattern is consistent with processing of PAI1–PAI4 palindromic transcripts into sRNAs by more than one dicer. Correspondingly, the C24 high-throughput sequencing analysis detected PAI sRNAs between 21 and 24 nt long covering the entire IR region [27]. The wild type Ws levels of endogenous PAI sRNAs were comparable to levels detected in a hundred-fold dilution of RNA prepared from the Δpai1–pai4(PAIIR) positive control transgenic strain (Figure 3). The transgenic strain produced mostly smaller PAI sRNAs, as previously observed for this strain using a PAI cDNA riboprobe [14], presumably due to differences in PAI IR expression patterns and structure.
PAI sRNA species were depleted in the dcl2 dcl3 dcl4 strain to similar background levels as detected in Δpai1–pai4 (Figure 3). This result supports the hypothesis that loss of PAI sRNAs underlies the reduction in SUVH-dependent H3K9me2 and CMT3-dependent 5meC on PAI sequences. The lack of residual PAI sRNAs in dcl2 dcl3 dcl4 indicates a minimal contribution of DCL1 to generating these species, although DCL1 is functional in processing microRNAs such as miR167.
To determine whether loss of sRNAs causes destabilization of remaining PAI silencing modifications upon inbreeding by self-pollination, we introduced the dcl2 dcl3 dcl4 mutations into a Ws pai1 reporter background where silencing of the PAI2 singlet gene can be monitored by visual inspection. In wild type Ws, PAI1 expressed from the heterologous upstream promoter is the major source of PAI enzyme; expression of PAI2 is impaired by H3K9me2/5meC on proximal promoter sequences and PAI3 and PAI4 do not encode functional enzyme due to polymorphisms [12]. In the Ws pai1 missense mutant, the impairment of PAI2 expression is revealed through tryptophan deficiency phenotypes including reduced size and blue fluorescence under ultraviolet (UV) light caused by accumulation of the tryptophan precursor anthranilate [28]. The stable maintenance of PAI2 H3K9me2/5meC in pai1 is reflected in stable maintenance of blue fluorescence across generations of inbreeding. Mutations that decrease PAI2 H3K9me2 and/or 5meC levels in the Ws pai1 background, including cmt3, met1, and suvh4, result in reduced fluorescence [2], [16], [28].
The initial pai1 dcl2 dcl3 dcl4 strain displayed partially reduced PAI 5meC patterns similar to the PAI1 dcl2 dcl3 dcl4 strain (Figure 1, Figure 4). The pai1 dcl2 dcl3 dcl4 plants were larger and less fluorescent than pai1 plants, reflecting the partial reduction of non-CG methylation levels on PAI2 (Figure 4).
Examination of pai1 dcl2 dcl3 dcl4 inbred populations revealed that blue fluorescence diagnostic of PAI2 silencing was not stably maintained. In a population of 191 pai1 dcl2 dcl3 dcl4 plants we found two non-fluorescent segregants (1.0%). In contrast, no non-fluorescent individuals were found in control populations of thousands of pai1 plants, consistent with our previous results. Each of the non-fluorescent pai1 dcl2 dcl3 dcl4 plants yielded approximately 75% non-fluorescent and 25% fluorescent second-generation progeny (54 non-fluorescent out of 70 total progeny plants [77%] for one line and 31 non-fluorescent out of 42 total progeny plants [74%] for another line). Approximately one third of the second-generation non-fluorescent plants lacked remaining PAI2 5meC in a HincII DNA gel blot assay, and these individuals yielded 100% non-fluorescent third-generation progeny, whereas the remaining second-generation non-fluorescent plants had partial levels of PAI2 5meC and yielded approximately 75% non-fluorescent third-generation progeny. For example, 12 out of 28 [43%] non-fluorescent second-generation progeny from one line had fully demethylated PAI2 phenotypes in the HincII assay and each of these individuals yielded 100% non-fluorescent progeny; the pai1 dcl NF line shown in Figure 4 is derived from one of these individuals. The segregation patterns are consistent with reduced PAI2 5meC levels and increased expression occurring on just one of the two chromosomes in the parental non-fluorescent plant and being inherited in a Mendelian fashion.
DNA gel blot analysis of a non-fluorescent pai1 dcl2 dcl3 dcl4 line indicated a nearly complete loss of CCG methylation monitored by MspI cleavage and partially reduced CG methylation monitored by HpaII cleavage at PAI2 relative to the fluorescent parental line, consistent with the reversion of tryptophan deficiency phenotypes (Figure 4). However, PAI1–PAI4 and PAI3 maintained similar 5meC patterns to the fluorescent parental line. Therefore, in pai1 dcl2 dcl3 dcl4 the loss of PAI2 5meC and silencing detected by the blue fluorescence screen is not coupled to destabilization of 5meC at the other PAI loci.
Both the initial loss of PAI non-CG methylation and the stochastic further loss of PAI2 silencing and 5meC in dcl2 dcl3 dcl4 are similar to patterns we previously observed in the Δpai1–pai4 mutant [17]. This comparison indicates that regardless of whether PAI sRNAs are depleted by loss of DCL function or by loss of the source of PAI dsRNA substrates for DCL cleavage (Figure 3), PAI2 silencing is similarly destabilized. The destabilization could be due to a combination of effects at PAI2 including impairment of H3K9me2 maintenance, increased transcription, and impairment of the DRM2/DCL3 pathway in resetting 5meC imprints (Text S1, Figure S2).
To determine whether other SUVH/CMT3 target loci besides the PAI genes have reduced 5meC levels in the dcl2 dcl3 dcl4 mutant, we used a survey approach with MspI DNA gel blot assays (Figure 5). We monitored representative sequences of three types: a degenerate (86% identical) inverted repeat locus IR1074 [29], highly repetitive 5S rDNA and 180 bp centromeric sequences (CEN), or low-copy transposons Ta3 [7] and Mu1 [30]. At all of these sequences, there was no difference in MspI cleavage between wild type and the drm1 drm2 mutant, but greatly increased cleavage in the cmt3 mutant, indicating that CCG methylation at the monitored MspI sites is dependent on CMT3 with a minimal contribution from the DRM MTases. For each sequence, we tested both the pai1 dcl2 dcl3 dcl4 parental fluorescent strain and an isogenic non-fluorescent progeny line to determine whether these two strains had differences in 5meC patterns at loci other than PAI2 (Figure 4).
The IR1074, 5S rDNA, and CEN sequences showed partially increased MspI cleavage in both of the dcl mutant strains relative to wild type and cmt3 (Figure 5). For the highly repetitive sequences the increased cleavage was evident as a slight shift downwards in the peak intensity of the ladder of cleaved bands (Figure 5B). Therefore, similarly to the PAI genes, these sequences require DCL function for maintenance of CMT3-dependent 5meC. The sequences had similar 5meC patterns in the fluorescent versus non-fluorescent pai1 dcl2 dlc3 dcl4 lines, indicating that the loss of PAI2 5meC in the non-florescent line is not coupled to more general destabilization of 5meC.
In contrast, Ta3 and Mu1 showed no differences in MspI cleavage between the dcl mutant lines and wild type controls (Figure 5). The Mu1 transposon has a polymorphic arrangement between Ws and the Columbia (Col) strain in which the dcl alleles were originally isolated, and the dcl2 dcl3 dcl4 strain carries both Mu1 arrangements. Despite this complication, comparisons to wild type versus cmt3 controls in each strain background showed no evidence of demethylation in the dcl mutant lines. Therefore, not all SUVH/CMT3 targets display DCL-dependent maintenance of 5meC. In light of our hypothesis that sRNAs prevent loss of H3K9me2 due to read-through transcription, the different effects of the dcl mutations at different SUVH/CMT3 target loci could reflect the extent to which read-through transcription occurs across the modified sequences. For example, the dcl-sensitive locus IR1074 is likely to be transcribed across to make fold-back dsRNA because this locus produces sRNAs even in an RNA-dependent RNA polymerase rdr2 mutant background [29].
We also monitored H3K9me2 levels at the single-locus targets IR1074 and Ta3 by ChIP in the dcl2 dcl3 dcl4 mutant and the same control strains used for analysis of the PAI genes (Figure 6). At IR1074 H3K9me2 levels were reduced in the dcl2 dcl3 dcl4 mutant relative to wild type, although not as strongly as in the suvh4 suvh5 suvh6 mutant. In contrast, at Ta3 H3K9me2 was maintained at similar levels between the dcl2 dcl3 dcl4 mutant and wild type. The H3K9me2 ChIP results agree with the 5meC results indicating that full modification of IR1074 but not Ta3 depends on DCL function (Figure 5A, Figure 5C).
The cmt3 mutant had reduced levels of H3K9me2 at both IR1074 and Ta3, presumably due to impaired SUVH localization and/or increased transcription caused by loss of non-CG methylation (Figure 5, Figure 6). However, at both loci the drm1 drm2 mutant maintained H3K9me2 at similar levels to wild type, and the drm1 drm2 cmt3 mutant maintained H3K9me2 at similar levels to cmt3 (Figure 6). Therefore, CMT3 but not the DRM cytosine MTases contributes to maintenance of H3K9me2 patterns at IR1074 and Ta3.
At IR1074, the dcl2 dcl3 dcl4 mutant and the cmt3 mutant displayed similar reductions in H3K9me2 levels relative to wild type (Figure 6). However, dcl2 dcl3 dcl4 had a weaker reduction in IR1074 non-CG methylation levels than cmt3 (Figure 5). This relationship is similar to that observed for PAI1–PAI4 and PAI2 (Figure 1, Figure 2), and supports the view that loss of sRNAs in the dcl2 dcl3 dcl4 mutant impairs maintenance of H3K9me2 patterns independently of CMT3 function at loci subject to read-through transcription.
In Arabidopsis, H3K9me2 maintained by the SUVH4, SUVH5, and SUVH6 histone MTases is used to guide 5meC in non-CG contexts maintained by the CMT3 cytosine MTase [1]–[4]. The SUVH MTases contain 5meC-binding domains, and CMT3 contains an H3K9me-binding domain, leading to the model that the SUVH/CMT3 pathway involves an amplification loop between 5meC and H3K9me2 [5], [6]. However, this amplification loop is not sufficient to maintain full levels of 5meC and H3K9me2 on the Ws PAI gene duplications, including a constitutively transcribed IR locus PAI1–PAI4 and partially silenced singlet genes PAI2 and PAI3. In previous work we showed that production of palindromic transcripts from the PAI1–PAI4 IR is also required for maintenance of PAI non-CG methylation [13], [14], [17], [18]. For example, in a Δpai1–pai4 mutant the PAI2 and PAI3 genes have reduced non-CG methylation, and the remaining 5meC on PAI2 is destabilized upon inbreeding [17]. Here we use mutations in the DCL dicer ribonucleases to show that PAI sRNAs processed from PAI dsRNAs are the key species that reinforce the SUVH/CMT3 amplification loop between H3K9me2 and non-CG methylation.
Arabidopsis uses DCL-dependent sRNAs incorporated into argonaute (AGO) effector proteins as nucleic acid sequence-specificity guides in a variety of pathways including miRNA control of development, RNA interference, and guidance of 5meC mediated by the DRM2 cytosine MTase 19,23. The DRM2 pathway contributes together with the SUVH/CMT3 pathway to maintenance of non-CG methylation at many 5meC target loci [15], [31]. This overlap has obscured whether sRNAs have an independent role in the SUVH/CMT3 pathway. However, the Ws PAI genes maintain 5meC in non-CG contexts almost entirely through the SUVH/CMT3 pathway once initial 5meC is established (Figure 1, Text S1, Figure S2). The reduction in PAI non-CG methylation and H3K9me2 levels in dcl mutant backgrounds therefore indicates a direct connection between DCL-dependent sRNAs and the SUVH/CMT3 pathway (Figure 1, Figure 2). DCL2 and DCL3 are the key dicers required for maintaining H3K9me2 and 5meC patterns on the PAI genes, suggesting a preference for longer 22 and 24 nt sRNAs in this pathway.
ChIP analysis shows that the dcl2 dcl3 dcl4 mutant has reduced H3K9me2 levels at PAI loci similar to or stronger than in the cmt3 mutant (Figure 2). However, the dcl2 dcl3 dcl4 mutant has weaker reductions in PAI 5meC levels than the cmt3 mutant (Figure 1). Therefore reduced H3K9me2 levels in the dcl mutant cannot be accounted for as a secondary effect of impaired CMT3 function. Instead, the ChIP results support the view that the dcl mutations directly impair maintenance of H3K9me2 patterns. In this view, the partial loss of PAI H3K9me2 in dcl2 dcl3 dcl4 reduces CMT3 localization, resulting in reduced PAI non-CG methylation as a secondary effect.
Similarly to the PAI genes, a subset of other SUVH/CMT3 target loci including highly repetitive 5S rDNA and CEN sequences have partially reduced non-CG methylation levels in the dcl2 dcl3 dcl4 mutant (Figure 5). However, some loci such as the low copy transposons Ta3 and Mu1 can maintain full 5meC levels relative to wild type despite the loss of DCL function. This variation could reflect the degree to which different SUVH/CMT3 target loci are transcribed across. This variation could also reflect which RNA polymerases are most active at different loci. Arabidopsis encodes five RNA polymerases: the conserved eukaryotic RNA polymerases POLI, POLII, and POLIII, and plant-specific POLIV and POLV implicated in targeting DRM2-dependent 5meC [32]. In particular, POLV is proposed to transcribe across target loci to make “scaffold” transcripts that recruit sRNA/AGO complexes and components of the DRM2 pathway [33], [34]. Because of its specialized role in making silencing-associated transcripts, POLV might be less disruptive of H3K9me2 than other RNA polymerases designed to express host genes. In this case, protein-encoding loci like the PAI genes that are transcribed by RNA POLII, and 5meC targets that depend on RNA POLII for scaffold transcript synthesis [35], might have a stronger dependence on an sRNA-based mechanism to maintain H3K9me2 than POLV-transcribed regions of the genome.
Our previous studies with allelic variants of the PAI1–PAI4 IR locus support the hypothesis that the level of transcription across the locus determines the extent to which sRNAs are needed for maintenance of PAI 5meC levels. In one study we used transgene-expressed sRNAs to direct 5meC and transcriptional silencing to the upstream promoter that drives transcription through PAI1–PAI4, thereby impairing production of PAI dsRNAs and sRNAs [14]. In this transgenic strain the PAI1–PAI4 locus was able to maintain full 5meC levels, whereas the PAI2 and PAI3 singlet genes had partially reduced non-CG methylation levels. These patterns are consistent with a model where the decreased transcription of PAI1–PAI4 specifically reduces its dependence on PAI sRNAs. In a second study we characterized a mutant derivative of Ws where a rearrangement in the center of the PAI1–PAI4 IR introduces a new polyadenylation site and reduces the levels of transcripts that extend into palindromic PAI4 sequences, without altering promoter sequences or the level of transcription across the locus [18]. In the rearrangement mutant the PAI1–PAI4 IR locus as well as the PAI2 and PAI3 singlet genes had partially reduced non-CG methylation levels. These patterns are consistent with a model where read-through transcription at all three loci together with reduced PAI sRNAs results in loss of H3K9me2 and 5meC at all three loci, similarly to the situation in the dcl2 dcl3 dcl4 mutant. Taken together, our results support a homeostatic mechanism where sRNAs produced from heterochromatic regions by read-through transcription feed back to counteract depletion of H3K9me2 and associated 5meC levels caused by read-through transcription.
The mechanistic relationship between sRNAs and maintenance of H3K9me2 patterns remains to be determined. In the fission yeast Schizosaccharomyces pombe, an sRNA-loaded AGO protein in the RITS effector complex interacts with nascent transcripts at centromeric repeats to recruit the Clr4 H3K9 MTase (reviewed in [36]). Plants could use an analogous effector complex interaction mechanism to target SUVH H3K9 MTases to specific regions of the genome. Consistent with this possibility, in a suvh4 suvh5 mutant background, the remaining SUVH6 MTase maintains levels of H3K9me2 and associated 5meC similar to wild type at the PAI1–PAI4 IR but not at the PAI2 singlet gene [4]; this locus-specific activity could reflect preferential interactions between SUVH6 and effector complexes that assemble near a site of dsRNA synthesis. Alternatively, sRNA-AGO complexes could recruit intermediate factors that then promote SUVH activity at specific targets. A third possibility is that sRNAs could guide a pathway that protects heterochromatic sequences from H3K9 demethylation. For example, the IBM1 JumonjiC domain H3K9 demethylase acts to prevent H3K9me2 and non-CG methylation from accumulating in transcribed genes 11,37,38. IBM1 could be excluded from also acting at heterochromatic sequences through a mechanism that involves sRNA-AGO complexes. Furthermore, sRNA-dependent mechanisms that promote addition of H3K9me2 or prevent removal of H3K9me2 could operate in concert.
Pathways where sRNA-AGO complexes guide H3K9me to appropriate regions of the genome have been identified in organisms ranging from fission yeast to the protozoan Tetrahymena thermophila to the insect Drosophila melanogaster, even though these organisms lack 5meC [39]–[41]. Our discovery that Arabidopsis also uses sRNAs to maintain H3K9me could represent a plant-specific variation on this fundamentally conserved strategy. In this case, the sRNA/SUVH/CMT3 pathway and the sRNA/DRM2 pathway could have both evolved from a basal mechanism involving sRNA-AGO guidance of H3K9 MTases. Consistent with this possibility, SUVH variants that lack catalytic activity but maintain methyl-DNA binding are required for DRM2-dependent 5meC [42].
The plant sRNA/H3K9me maintenance mechanism is interwoven with the SUVH/CMT3 chromatin binding amplification loop and partially redundant functions of the MET1 and DRM pathways to create a reinforced silencing network. However, loss of the sRNA/H3K9me maintenance mechanism cannot be completely buffered by the other pathways, and results in both immediate reductions and longer-term destabilization of H3K9me2 and 5meC. The unique properties of the PAI genes make them ideal reporters to further understand how sRNAs are harnessed to control maintenance of H3K9me2 on appropriate target sequences in plant genomes.
T-DNA insertional dcl alleles were obtained from the Arabidopsis Biological Resource Center (ABRC) or from the laboratory of James Carrington at Oregon State University. The dcl2-1, dcl3-1, and dcl4-2 mutations are likely null alleles originally isolated in the Col strain [21], [23]. The dcl1-9 mutation is a partial function allele originally isolated in Ws, but then crossed five times to the Landsberg erecta (Ler) strain 24,25. Each dcl mutant was crossed to Ws. PCR-based genotype markers were used to identify dcl mutant progeny homozygous for the three PAI loci from Ws (Table S1). Each dcl allele was crossed a second time with Ws to increase the proportion of the genome contributed by the Ws parent. The resulting dcl single mutant strains were then crossed with each other to generate double, triple, and quadruple mutant combinations. The dcl mutations were also crossed into the Ws pai1 reporter strain [28]. The Ws drm1 drm2 double T-DNA insertional null strain was obtained from the laboratory of Steven Jacobsen at UCLA [43]. The Col cmt3-11T T-DNA insertional null strain was obtained from the ABRC [44]. The Ws pai1, Ws Δpai1–pai4, Ws Δpai1–pai4(PAIIR), Ws cmt3i11a, Ws x met1-1, and pai1 suvh4R302* suvh5-1 suvh6-1 strains were previously described [4], [14], [16], [17], [28]. Ws cmt3illa and Ws drm1 drm2 mutants were crossed to make the Ws drm1 drm2 cmt3 strain.
Plant genomic DNA preparation and DNA gel blot assays for 5meC were performed as previously described [12]. Bisulfite sequencing of the top strands of PAI1 and PAI2 proximal promoter regions was performed as previously described [14]. PAI bisulfite sequencing primers are listed in Table S1.
Total RNA was extracted with TRIzol reagent (Invitrogen) using the manufacturer's protocol. Low molecular weight (LMW) RNA was enriched by precipitating high molecular weight RNA out of solution with 0.5 M NaCl, 10% polyethylene glycol (MW 8000). The remaining LMW RNA was precipitated with 100% ethanol and resuspended in water treated with diethyl pyrocarbonate. LMW RNA was fractionated on a 17% acrylamide 7 M urea gel and transferred to a Hybond-N membrane (GE Healthcare). sRNA 5′ ends were chemically crosslinked to the membrane as previously described [45]. Membranes were hybridized in OligoHyb buffer (Ambion) overnight at 42°C with 32P 5′ end-labeled oligonucleotide probes. Probes were either an LNA modified PAI1 exon 5 sense 35-mer (Exiqon) or an miR167 antisense 21-mer (Table S1). Probed membranes were washed three times with a 2× SSC, 0.1% SDS solution. sRNA sizes were estimated from an ethidium bromide-stained low molecular weight DNA ladder (USB), and by comparison to the PAI sRNA species observed in the Δpai1–pai4(PAIIR) control strain [14].
Formaldehyde crosslinking and chromatin preparations were performed as previously described [46] starting with two grams of aerial tissue from three-week-old plants grown in soilless potting medium (Fafard mix 2) under continuous illumination. Chromatin was immunoprecipitated with anti-H3K9me2 monoclonal antibody [47] or carried through the protocol with no antibody added as a control (mock precipitation). Immunoprecipitations were performed as previously described [3]. Each ChIP assay was performed in at least three independent biological replicates. Quantitative PCR amplification of immunoprecipitated DNA was performed using the 7300 Real-Time PCR System (ABI), with three replicate reactions for each sample. ChIP primer sequences are listed in Table S1.
|
10.1371/journal.pgen.1005657 | A Simple Model-Based Approach to Inferring and Visualizing Cancer Mutation Signatures | Recent advances in sequencing technologies have enabled the production of massive amounts of data on somatic mutations from cancer genomes. These data have led to the detection of characteristic patterns of somatic mutations or “mutation signatures” at an unprecedented resolution, with the potential for new insights into the causes and mechanisms of tumorigenesis. Here we present new methods for modelling, identifying and visualizing such mutation signatures. Our methods greatly simplify mutation signature models compared with existing approaches, reducing the number of parameters by orders of magnitude even while increasing the contextual factors (e.g. the number of flanking bases) that are accounted for. This improves both sensitivity and robustness of inferred signatures. We also provide a new intuitive way to visualize the signatures, analogous to the use of sequence logos to visualize transcription factor binding sites. We illustrate our new method on somatic mutation data from urothelial carcinoma of the upper urinary tract, and a larger dataset from 30 diverse cancer types. The results illustrate several important features of our methods, including the ability of our new visualization tool to clearly highlight the key features of each signature, the improved robustness of signature inferences from small sample sizes, and more detailed inference of signature characteristics such as strand biases and sequence context effects at the base two positions 5′ to the mutated site. The overall framework of our work is based on probabilistic models that are closely connected with “mixed-membership models” which are widely used in population genetic admixture analysis, and in machine learning for document clustering. We argue that recognizing these relationships should help improve understanding of mutation signature extraction problems, and suggests ways to further improve the statistical methods. Our methods are implemented in an R package pmsignature (https://github.com/friend1ws/pmsignature) and a web application available at https://friend1ws.shinyapps.io/pmsignature_shiny/.
| Somatic (non-inherited) mutations are acquired throughout our lives in cells throughout our body. These mutations can be caused, for example, by DNA replication errors or exposure to environmental mutagens such as tobacco smoke. Some of these mutations can lead to cancer. Different cancers, and even different instances of the same cancer, can show different distinctive patterns of somatic mutations. These distinctive patterns have become known as “mutation signatures”. For example, C > A mutations are frequent in lung caners whereas C > T and CC > TT mutations are frequent in skin cancers. Each mutation signature may be associated with a specific kind of carcinogen, such as tobacco smoke or ultraviolet light. Identifying mutation signatures therefore has the potential to identify new carcinogens, and yield new insights into the mechanisms and causes of cancer, In this paper, we introduce new statistical tools for tackling this important problem. These tools provide more robust and interpretable mutation signatures compared to previous approaches, as we demonstrate by applying them to large-scale cancer genomic data.
| Cancer is a genomic disease. As we lead a life, DNA within our cells acquires random somatic mutations, mainly caused by DNA replication errors and exposures to mutagens such as chemical substances, radioactivities and inflammatory reactions. Although most mutations are harmless (called “passenger mutations”), a small portions of mutations at some specific sites in cancer genes (“driver mutations”) affect cell growth, causing autonomous proliferation, tissue invasion, and contributing to oncogenesis [1]. Cancer genome studies typically focus on identifying driver mutations, to help understand the mechanism of cancer development. However, passenger mutations can also yield important information, because they often show patterns (“mutation signatures”) which can provide insights into the forces that cause somatic mutations. For example, classical studies of mutation patterns revealed that C > A mutations are abundant in lung cancers in patients with smoking history, and these are caused by benzo(a)pyrene included in tobacco smoke [2]. Also, C > T and CC > TT mutations are abundant in ultraviolet-light-associated skin cancers, and these are caused by pyrimidine dimers as a result of ultraviolet radiation [3].
The potential for classical studies to yield insights into somatic mutation processes was limited in several ways. Due to limited sequencing throughput, most classical studies focused on a few cancer genes, such as TP53, where high mutation frequencies could be expected. They then contrasted mutation pattern profiles among different cancer types, aggregating mutations across multiple individuals within the same cancer type to yield sufficient mutations for analysis. However, since many of the mutations in cancer genes are driver mutations causing cell proliferation, the resultant mutation profiles are a biased representation of the underlying mutation process. Furthermore, the paucity of mutation data made it effectively impossible to assess variation in mutation patterns among individuals.
Recent advances in high-throughput sequencing provide new opportunities to investigate sample-by-sample mutation signatures in an unbiased way using genome-wide somatic mutation data. For example, a large-scale study using 21 breast cancer samples identified an association of C > [AGT] mutations at TpC sites, which was later proved to be caused by APOBEC protein family [4–6], and a novel phenomenon called kataegis [7]. Moreover, a landmark study of 7,034 primary cancer samples, representing 30 different cancer classes, has provided the first large-scale overview of mutation signatures across a large number of cancer types [8]. This has lead to great hopes that detection of novel mutation signatures and associated mutagens can lead to identification of novel mutagens and prevention of cancer.
To make the most of these opportunities requires the development of efficient and effective statistical methods for analyzing mutation signatures in vast amounts of somatic mutation data. Current statistical approaches [9, 10] are excellent starting points, and have helped generate the new insights noted above. However, we argue here that these existing methods have two important limitations, caused by the fact that they use an unconstrained model for each “mutation signature”. First, although using an unconstrained model might appear to be a good thing in terms of flexibility, in practice it can actually reduce flexibility, because the price of using an unconstrained model is that one must limit the domain of mutation signatures considered. For example, most recent analyses of mutation signatures consider only the immediate flanking 5′ and 3′ bases of each substitution to be part of the signature, even though it is known that more distal bases—and particularly the next flanking base on each side—can contain important contextual information [11]. These recent analyses take this approach because, in the unconstrained model, incorporating the more distal bases into the signature very substantially increases the number of parameters, making estimated mutation signatures unstable. Secondly, and just as important, the unconstrained model means that each signature is a probability distribution in a high-dimensional parameter space, which can make signatures difficult to interpret.
In this paper, we present a novel probabilistic approach to mutation signature modelling that addresses these limitations. In brief, we first simplify the modelling of mutation signatures by decomposing them into separate “mutation features”. For example, the substitution type is one feature; flanking bases are each another feature. We then exploit this decomposition by using a probabilistic model for signatures that assumes independence across features. This approach substantially reduces the number of parameters associated with each signature, greatly facilitating the incorporation of additional relevant sequence context. For example, our approach can incorporate the two bases 3′ and 5′ of the substitution and transcription strand biases using only 18 parameters per signature, compared with 3,071 parameters per signature with current approaches. We demonstrate the benefits of this simplification in data analyses. These benefits include more stable estimation of signatures from smaller samples, refinement of the detail and resolution of many mutation signatures, and possibly identification of novel signatures.
Assuming independence among features in a signature may initially seem unnatural. However, its use here is analogous to “position weight matrix models” which have been highly successful for modelling transcription factor binding motifs. Indeed, an important second contribution of our paper is to provide intuitive visual representations for mutation signatures, analogous to the “sequence logos” used for visualizing binding motifs. Finally, we also highlight the close connection between mutation signature models and the “mixed-membership models”, also known as “admixture models” [12] or “Latent Dirichlet Allocation” models [13] that are widely used in population genetics and document clustering applications. These connections should be helpful for future elaboration of computational and statistical methods for cancer mutation signature detection.
Software implementing the proposed methods is available in an R package pmsignature (probabilistic mutation signature), at https://github.com/friend1ws/pmsignature. The core part of the estimation process is implemented in C++ by way of the Rcpp package [14], which enables handling millions of somatic mutations from thousands of cancer genomes using a standard desktop computer. In addition, a web-based application of our method is available at https://friend1ws.shinyapps.io/pmsignature_shiny/.
The term “mutation signature” is used to describe a characteristic mutational pattern observed in cancer genomes. Such patterns are often related to carcinogens (e.g., frequent C > A mutations in lung cancers with smoking histories).
What constitutes a mutational pattern varies among papers. The simplest approach is to consider 6 possible mutation patterns, corresponding to 6 possible substitution patterns (C>A, C>G, C>T, T>A, T>C, T>G; the original base is often fixed to C or T to remove redundancy of taking complementary strands). However, in practice we know that DNA context of a substitution is often important, and so it is common to go the next level of complexity, and include the immediate 5′ and 3′ flanking bases in the mutation pattern. This results in 96 (6 × 4 × 4) patterns. Further incorporating the strand (plus or minus) of each substitution extends this to 192 patterns [8, 9].
Mathematically, mutation signatures have previously been characterized using an unconstrained distribution over mutation patterns [9, 10]. Thus, if the number of mutation patterns considered is M then each mutation signature is characterized by a probability vector of length M (which must sum to 1, so M − 1 parameters). A problem with this approach is that it requires a large number of parameters per mutation signature. As noted above, even accounting only for immediately flanking bases gives M = 96. Furthermore, M increases exponentially if we try to account for additional context: to take account of up to n bases 5′ and 3′ to the mutated site (henceforth referred to as the −n position and the +n position, respectively) gives M = 6 × 42n. Having a large number of parameters per signature causes two important problems: i) estimates of signature parameters can become statistically unstable; ii) signatures can become difficult to interpret.
The first contribution of this paper is to suggest a more parsimonious approach to modelling mutation signatures, with the benefit of producing both more stable estimates and more easily interpretable signatures. In brief, we substantially reduce the number of parameters per signature by breaking each mutation pattern into “features”, and assuming independence across mutation features. For example, consider the case where a mutation pattern is defined by the substitution and its two flanking bases. We break this into three features (substitution, 3′ base, 5′ base), and characterize each mutation signature by a probability distribution for each feature (which, by our independence assumption, are multiplied together to define a distribution on mutation patterns). Since the number of possible values for each feature is 6, 4, and 4 respectively this requires 5 + 3 + 3 = 11 parameters instead of 96 − 1 = 95 parameters. Furthermore, extending this model to account for ±n neighboring bases requires only 5 + 6n parameters instead of 6 × 42n − 1. For example, considering ±2 positions requires 17 parameters instead of 1,535. Finally, incorporating transcription strand as an additional feature adds just one parameter, instead of doubling the number of parameters.
Since the aim of a mutation signature is, in some ways, to capture dependencies among features, the independence assumption may seem counter-intuitive. However, the idea is exactly analogous to the use of a “position weight matrix” (PWM) to represent motifs in sequence data. In this analogy, a motif is analogous to a mutation signature, and each location in the motif is analogous to a “feature”. Just as we use a probability vector for each feature, a PWM defines a probability vector for each location in a motif, and these probabilities at each location can be multiplied together to yield a probability distribution on sequences. Even though a PWM cannot capture complex higher-order dependencies, some of which likely do exist in practice, it has been a highly successful tool for motif analysis—likely because it can capture the most important characteristics of transcription factor binding sites (that some locations will show strong preference for a particular base, whereas others will not), and also because it can be represented in an easily interpretable way via sequencing logos [15]. For similar reasons—in addition to the empirical demonstrations we present later—we believe our mutation signature representation will prove useful for mutation signature analysis.
Fig 1 illustrates the way that our new representation of signatures can simply capture a previously identified signature [9, 16] and provides an easily interpretable visualization of the signature that is analogous to sequencing logos [15]. We particularly note how the key elements of this mutation signature are more immediately visually apparent than with visualizations of the full vector of probabilities used by existing approaches.
Suppose each somatic mutation has L mutation features, m = (m1, m2, ⋯, mL), where each ml can take Ml discrete values. Also, let M: = (M1, ⋯, ML). For example, when taking account of 6 substitution patterns and ±2 flanking sites, M = (6, 4, 4, 4, 4). See S1 Table for other examples.
Suppose we have observed mutations in I sampled cancer genomes, and let Ji denote the number of observed mutations in the i-th cancer genome. Further, let xi, j = (xi, j, 1, ⋯, xi, j, L), (i = 1, ⋯, I, j = 1, ⋯, Ji) denote the observed mutation feature vector for the j-th mutation of i-th cancer genome, where xi, j, l ∈ {1, ⋯, Ml}.
Our model assumes that each mutation arose from one of K possible mutation signatures. Each cancer sample has its own characteristic proportion of mutations of each signature type (which might depend on lifestyle, genetic differences, etc.).g We let qi, k denote the proportion of signature k in sample i, so qi = (qi, 1, qi, 2, ⋯, qi, K)∈ΔK, (i = 1, ⋯, I) where Δ S = { ( t 1 , ⋯ , t S ) : t s ≥ 0 ∀ s , ∑ s = 1 S t s = 1 } denotes the S-dimensional simplex of non-negative vectors summing to 1. Further, each mutation signature is characterized by parameter vectors Fk: = (fk, 1, …, fk, L), where fk, l is a probability vector for the l-th feature in the k-th signature. That is, fk, l = (fk, l, 1, …, fk, l, Ml) ∈ ΔMl.
Our generative model for the observed mutations {xi, j} in each cancer sample can now be described as a two-step process.
This generative model is summarized in Fig 2. This model is essentially a “mixed-membership model”, also known as an “admixture model” [12] or “Latent Dirichlet Allocation” [13]. For example, the membership proportions for each sample are analogous to admixture proportions in an admixture model; the mutation signatures are analogous to populations, and the mutation signature-specific parameters are analogous to population-specific allele frequencies.
The key parameters in this model are the membership proportions for each sample, qi, and the mutation signature parameters, Fk. We estimate these parameters by maximizing likelihood using an EM algorithm. A simulation study demonstrates that the estimation method can reproduce the mutation signature very accurately provided enough mutations and samples are available (see S1 Text). See Methods for more detailed models, parameter estimation, further discussion on relationships with mixed membership models, how to select K, etc.
The intrinsic composition of genome sequence, if unaccounted for, can undesirably influence estimated mutation signatures. For example, since the di-nucleotide CpG is underrepresented in most genomic regions (other than promoters), a signature with substitutions from a C base can have weaker signals of G base at the +1 position. In previous work [10], this background problem was dealt by explicitly incorporating “mutation opportunity” coefficients into the model. Here, to reduce the influences of intrinsic sequence composition on our signature estimates, we introduce a special “background signature” { F 0 } ∈ Δ M 1 × ⋯ × M L, which is designed to capture biases in intrinsic genome sequence composition and is calculated from the composition of consecutive nucleotides of the human genome sequence (See Methods for the detail).
Here we compare our new “independent model” for mutation signatures, which assumes independence among mutation features, with the “full model”, which corresponds to existing approaches. We compare mutation signatures obtained by the two approaches and investigate the robustness of each approach by down-sampling experiments.
The data consist of 14,717 somatic substitutions collected from a study of 26 urothelial carcinomas of the upper urinary tract (UCUT) [18]. The original study identified a novel mutation signature in these data: T > A substitutions at CpTpG sites with a strong transcription strand specificity caused by aristolochic acids (AA).
We consider a mutation pattern to consist of the substitution pattern, the ±2 flanking bases, and the transcription strand direction. Thus each signature is characterized by 18 parameters in our independent model, and by 3,071 parameters in the full model. After analyzing the data with various numbers of mutation signatures K, we selected K = 3 signatures for these analyses (see S2 Text).
The inferred APOBEC signature under the independent model shows a clear depletion of G base at the −2 position, which is consistent with the previous study [9] and results in the next subsection (Fig 3A and 3B). In contrast, for the full model, this tendency is rather mild (Figs 3C, 3D, and S1). The inferred AA mutation signature has no clear characteristics at the −2 position. These results suggest that our independent model has the potential to identify signatures in more detail and with less data than existing approaches based on the full model.
To investigate this further we performed down-sampling experiments. Using the mutation signatures obtained using all 14,717 substitutions as a gold standard, we assessed performance of the proposed method on down-sampled data consisting of r% of the original data, where r = (1%, 2.5%, 5%, 10%, 25%, 50%). To measure robustness we used the cosine similarity on the full dimensional vector space, which allows comparison between the full model and the independent model. We repeated each down-sampling experiment 100 times for each model.
The results (Fig 3E and 3F) confirm that the results of the independent model are substantially more robust to reductions in data size than the full model. Indeed, mutation signatures inferred using the independent model with only 10% of the data remain highly similar to the signatures inferred from the full data; by comparison the full model shows a much larger drop-off in similarity, especially in the APOBEC signature where even using 50% of the data gives a substantial drop-off in similarity. Both methods found the AA signature easier to recover than the APOBEC signature. We believe that this is because the number of T > A substitutions at GpTpC sites are far more frequent in this dataset.
To provide a more comprehensive practical illustration of our method, we applied it to somatic mutation data from 30 cancer types [8]. We applied the method to each cancer type separately to assess similarity of estimated signatures across cancer types. For each cancer type we selected the number of signatures K by fitting the model with increasing K and examining the log-likelihood, bootstrap errors, and correlation of membership parameters. The selected values of K are given in S2 Table. Also, we simply removed somatic mutations located in an intergenic region to include transcription strand biases as mutation features. Finally, we merged similar mutation signatures across different cancer types (when their Frobenius distances were < 0.6, where the Frobenius distance between two matrices (F1, F2) is Tr ( ( F 1 - F 2 ) ( F 1 - F 2 ) t ) (Tr means the trace of square matrices).
Figs 4 and 5 summarize the results. In total, we identified 27 mutation signatures. Many of these signatures show reassuring similarities with signatures identified in previous studies. However signatures from our independence model, because of its ability to effectively and parsimoniously deal with both ±2 flanking base context and strand bias, are often more refined, highlighting additional details or features not previously evident. By comparing the composition of nucleotides and cancer types exhibiting the signatures with results of previous studies, we were able to associate many of the detected signatures with known mutational processes. In addition, as we reviewed these signatures and compared them with previous work, we noticed connections that, while not directly related to our new model, appear novel and noteworthy. The remainder of this section provides a comprehensive discussion of these findings.
Signatures 1 and 8 (C > A at TpCpT and C > T at TpCpG, respectively) observed in colorectal and uterine cancers appear likely to be associated with deregulated activity of the error-prone polymerase Pol ϵ. In previous analyses of these data [8], the signature for Pol ϵ dysfunction was represented by a single signature (their “signature 10”, see S2 Fig). In contrast our new approach uses two signatures. Since these signatures are highly correlated, and appear connected by a single biological mechanism, we certainly do not argue that inferring them as a single signature is “wrong”. However, splitting them into two signatures does help highlight certain features. Specifically, signature 1 shows a transcription strand bias whereas signature 8 does not, and this is true for both colorectal and uterus cancers (S3C and S3D Fig). This strand bias may be connected with the enrichment of C >A at TpCpT mutations in leading strands of replication forks observed by [19]. Although replication strand bias is different from transcription strand bias, these two biases may be connected through the fact that replication origins prefer transcription start sites [20].
These signatures also illustrate the ability of our model to help highlight sequence context effects beyond the immediate flanking bases. Specifically, both signatures 1 and 8 show an elevated frequency of the T base at position −2, and signature 1 also shows slightly elevated frequency of the T base at position +2 (Figs 6B, 6C, S3C, and S3D). A previous study of Pol ϵ [19] found that a nonsense mutation R23X of TP53 is enriched in cancers with Pol ϵ defects. In fact, the pattern of this mutation is C > T at TpTpCpGpA, closely matching signature 8. This illustrates that the inclusion of ±2 bases into signatures may be helpful for identifying underlying mechanisms.
Signature 2 (C > A at [CT]pCpT) is observed solely in low grade gliomas, and appears related to, but slightly different from, the signature previously detected in the same cancer types (“signature 14”, [9]). Indeed, the corresponding signature in the previous study shows very complex patterns (C > A at NpCpT or C > T at GpTpN). Further investigation revealed that this signature is driven by a single sample with an extremely high mutation rate (see S4A and S4B Fig), and signature 2 disappeared when we removed this sample (S5 Fig). It may be that the complex low-grade-glioma specific signature detected in the previous study is driven by the same single sample. We suggest that these signatures should be treated with caution until validated in additional samples.
Signature 4 (C > A at CpCpG) observed in kidney clear cell carcinomas, lung adenocarcinomas and melanomas seems to correspond to the “signature R2” detected in the same cancer types (plus lung squamous carcinomas) in [9] (see their Supplementary Figures). Again our analysis highlights additional contextual information, with a strikingly elevated frequency of base C at the −2 position (S6A, S6B and S6C Fig). However, for each cancer type, only a few samples support this signature (see S6D, S6E and S6F Fig), and the corresponding signature could not be validated in the previous study: most somatic mutations corresponding to that signature could not be validated by re-sequencing or visual inspection of BAM files using genomic viewers. Again, further investigation yielded a potential explanation for this finding: this signature largely matches that of a putative artifact caused by oxidation of DNA during acoustic shearing [21], and we conclude that this signature, and the corresponding signature in previous work, are likely artefactual. Although not of direct biological interest, identifying artefactual signatures could be helpful in removing false positive mutations.
Signature 13 (T > [AGT] at TpCpN sites) was observed in 12 cancer types, and is surely related to the activity of the APOBEC family. The 12 inferred signatures were highly consistent among cancer types except for B-cell lymphoma (see S3A Fig), highlighting the robustness of our approach. Almost all of them show enrichment of A and T and depletion of G base at the −2 position (Figs 6A and S3A), consistent with the UCUT data above and previous analyses [9]. The estimated transcribed strand specificities varied among cancer types, suggesting that there is not consistent strand-specificity in APOBEC signatures (and the observed variation may be due to estimation errors). Signatures 15 and 16 may also be related to APOBEC, although the estimated signatures are sufficiently different from 13 that they were not merged into a single cluster by our specified clustering criteria.
Signatures 10, 11, 12, 19 and 21 provide further examples of our method refining previously observed signatures, highlighting strand biases and/or sequence context effects, particularly 2 bases upstream of the substitution. Signature 10 (C >T at [CT]pCpC) was observed in head and neck cancers and melanomas, and probably relates to ultraviolet light. Consistent strand specificities among the two cancer types (S3E Fig) matches previous results [9], but our analysis additionally highlights elevated abundance of T at the −2 position (Figs 6D and S3E). Signature 11 (C > T at GpCp[CG]) appears in small-cell lung cancers and stomach cancers, and seems to be the same as “signature 15” in the previous study, whose function remains unclear. Again our analysis highlights elevated abundance of G at the −2 position (Figs 6E and S3F). Signatures 12 (C > T at [CG]pCp[CT]), 19 (T > C at GpTpN) and 21 (T > [CG] at CpTpT) observed in pilocytic astrocytomas, stomach cancers and oesophagus cancers, respectively, agree well with those detected in the same cancer types in the previous study [9]. However our analysis again refines these signatures, highlighting a strand bias in all three, and sequence context effects at the −2 position in Signatures 12 and 21.
One signature, Signature 20, appears not to match any signatures in the previous analysis [9] and represents a potentially novel signature. This signature (T > C at [AC]pTpN) is observed in thyroid cancers, and shows a very strong strand specificity, which could be due to transcription-coupled nucleotide excision repairs. This signature may have been too weak for previous methods to detect, perhaps because the mutation ratio of thyroid cancer is low, possibly reflecting improved sensitivity of our more parsimonious model.
The remaining signatures largely recapitulate previous results. Signature 3 and 5 (C > A at NpCpN) observed in head-and-neck cancers and three types of lung cancers are probably associated with tobacco smoking. The estimated signature in each cancer type shows higher mutation prevalence on the template strand (S3B Fig), which is consistent with the previous study [2, 9]. Signature 6 (C > A at NpCp[AT]) observed in neuroblastomas matches the pattern detected in the same cancer type in the previous study. Signature 7 (C > T at NpCpG sites) was observed in 25 out of 30 cancer type, and arguably relates to deamination of 5-methyl-cytosine. Signature 9 (C > T at NpCp[CT]) was observed in melanomas and glioblastomas, and is probably associated with a chemotherapy drug, temozolomide. Signature 18 (T > C at ApTp[AG]) observed in liver cancers has been shown to be more common in Asian cases than in other ancestries [16], though the source of this signature is still not clear. In this signature, we observe a very strong strand specificity as shown in [9, 16], suggesting a possible role for transcription-coupled nucleotide excision repairs.
In this paper, we presented new methods for inferring and visualizing mutation signatures from multiple cancer samples. The new methods exploit simpler, more parsimonious, models for mutation signatures than existing methods. This improves stability of statistical estimation, and easily allows a wider range of contextual factors (e.g. more flanking bases) to be incorporated into the analysis. In addition, we provide a new intuitive way to visualize the inferred signatures.
We have also emphasized the connection between mutation signature detection, and the use of mixed-membership models in other fields, particularly admixture analysis and document clustering. This connection naturally raises the possibility of improving the proposed approach by learning from experiences in those other fields. For example, in admixture analysis, [22] found that the use of a correlated prior on allele frequencies improved sensitivity to detect population clusters; this suggests that it might be fruitful to consider a correlated prior distribution on signatures, to allow that some signatures—perhaps in different cancers—may be similar to one another (though not identical). More generally, introducing certain prior distributions or penalty terms, such as sparsity-promoting penalties [23, 24] and determinantal point process priors [25, 26] could improve both accuracy and interpretation. Further, as the scale of cancer genome data becomes large, more sophisticated computational approaches for estimating parameters may become necessary. We can potentially borrow a number of computational techniques such as those using EM-algorithm [27, 28], sequential quadratic programming [29], Gibbs sampling [12, 30] and variational methods [13, 31, 32]. Finally, to address the problem of determining the number of signatures, it may be fruitful to extend the framework to the Hierarchical Dirichlet processes [33].
Although we have focused on point substitution mutations in this paper, many other types of mutations occur in cancer genomes, including insertions, deletions, double nucleotides substitutions, structural variations and copy number alterations [34, 35]. Our framework could incorporate these additional mutation types, by summarizing them using appropriate mutation features. In some cases, choice of appropriate features may need investigation. For example, longer deletions could be represented by the length of deletion and the adjacent bases; for short deletions (a few bases) it may be fruitful to include the actual deleted bases as part of the features.
We have detected a number of mutation signatures having transcription strand biases, which are naturally considered to be associated with transcription activities. Therefore, to further understand the effect of transcription activities on mutational mechanisms, we can include gene expression or RNA polymerase II occupancies to mutation features, so that the relationships of strand biases and transcription activities will be clarified. Also, it may be interesting to devise a probabilistic model for mutation signatures somewhere between complete independence and non-independence assumption, for example, using ideas analogous to those in [36] that uses a Markovian structure for transcription factor binding sites. This may help improve the modelling flexibility of mutation signatures while keeping the number of parameters moderate.
Although we believe our new methods already provide useful gains compared with existing approaches, the methods are perhaps even more important for their future potential to incorporate other contextual data, including epigenetic data, into mutation signature analysis. This is important, because local mutation densities are closely related to a number of genomic and epigenetic factors, such as GC content, repeat sequences, chromatin accessibility and modifications, and replication timing [37–40]. A recent study found that epigenetic information in the cell types of origin of the corresponding tumors is the most predictive [41] for local mutation densities. A growing range of epigenetic data from many cell types are now available, and it will be interesting to integrate these epigenetic factors into mutation signature analysis to help understand how these epigenetic factors influence DNA damage and repair mechanisms. Our work here provides a straightforward way to do this: epigenetic data can be simply added as features to the mutation signature. This has the potential to improve accuracy of signature detection (e.g. S7 Fig), and to produce novel biological insights. We believe that the value and impact of our work, and specifically our proposed approach to modelling mutation signatures via independent features, will grow as more and more features are incorporated into the analysis.
The parameters {fk, l} and {qi} must be estimated from the available mutation data {xi, j}. Here we adopt a simple approach that uses an EM-algorithm to maximize the likelihood.
Let gi, m denote the number of mutations in the i-th sample that have mutation feature vector m. In the E step of the EM algorithm, we calculate values of auxiliary variables θi, k, m defined as
θ i , k , m = q i , k ∏ l = 1 L f k , l , m l ∑ k ′ = 1 K q i , k ′ ∏ l = 1 L f k ′ , l , m l . (2)
Then, in the M-step, we update the parameters {fk, l} and {qi, k} as
f k , l , p = ∑ m : m l = p g i , m θ i , k , m ∑ p ′ ∑ m : m l = p ′ g i , m θ i , k , m , (3) q i , k = ∑ m g i , m θ i , k , m ∑ k ′ ∑ m g i , m θ i , k ′ , m . (4)
We use the R package SQUAREM [42] to accelerate convergence of this EM algorithm (SQUAREM implements a general approach to accelerate the convergence of any fixed-point iterative scheme such as an EM algorithm). To address potential problems with convergence to local minima, we apply the EM algorithm several times (10 times in this paper) using different initial points, and use the estimate with the largest log-likelihood. See S3 Text for the derivation of the above updating procedures.
Here, we describe how the background mutation signature is obtained in the case where mutation features are the substitution patterns, the ±2 flanking bases, and the transcription strand. Since the majority of the data used in this paper is exome sequencing data, and since we consider transcription strand as a mutation feature, we use the exonic regions of the human genome reference sequence to obtain the background mutation signature. First we calculate the frequencies of 5-mers with transcription strands of the corresponding exon, where we take complement sequences and flip the strand for those whose central bases are A or G. Then, assuming alternated bases are equally likely from each central base C and T, the frequency of each mutation feature is derived directly from those of the 5-mers and transcription strands. Finally, the probability of each mutation feature is derived by normalizing each frequency to sum to one.
We use the non-parametric bootstrap [43] to calculate standard errors for parameter estimates. This involves resampling somatic mutations from the empirical distribution of the original data {xi, j} for each cancer genome. For each of 100 such bootstrap samples, we re-fitted the model, using parameters obtained for the original data as initial points. We then used sample standard errors of the inferred mutational signatures as estimates of parameter standard errors.
Determining an appropriate number of mutation signatures K is a challenging task. One approach is to utilize some statistical information criteria such as AIC [44] or BIC [45]. In the population structure problems, for example, the Bayesian deviance [12] and cross-validation [46] have been suggested. One previous study on mutation signature problems [10] utilized BIC. On the other hand, the problem of using these statistical information criteria is that most of them are based on the likelihood, where slight deviations between the specified probabilistic models and the reality sometimes leads to additional (possibly spurious) mutation signatures being selected to compensate for those deviations.
In this paper, instead of utilizing a statistical information criteria, we adopt the following strategy:
These strategies are not claimed as optimal, but appeared to provide satisfactory results in our applications here. The development of automated and practical approaches for choosing K is a possible area for future development.
Previous approaches to mutation signature modelling in [8, 9] are a special case of our framework. Specifically, they correspond to combining all possible combinations of mutation features into a single “meta-feature”, which takes M1 × M2 × … × ML possible values. Thus, instead of having L features with M = (M1, …, ML), existing approaches have one feature with M = (M1 × … × ML) (see S1 Table). The resulting model allows for arbitrary distributions on the M1 × … × ML feature space, and we call the resulting model the “full model”. The full model can represent complicated dependencies in a single signature. For example, a situation where C > A is frequent at ApCpG sites and C > T is frequent at TpCpA sites could be represented with one signature. This may be desirable in some settings and not in others. However, when many mutation contextual factors are taken into account and the number of free parameters becomes huge, estimated results can be unstable and unreliable. Furthermore, there is a risk of over-interpreting the complex features of estimated signatures.
Our model is closely related to mixed-membership models that have been adopted in other applications, such as document classification and population structure inference problems. In this subsection, we outline these relationships, slightly abusing notation to contrast the relationships.
In the topic model [13, 27], which are a form of mixed-membership models frequently used in document classification problems, each document is assumed to have K different “topics” in varying proportions (qi ∈ ΔK), where each topic is characterized by a word frequency (a multinomial distribution on a set of words W (fk ∈ ΔW). And each word is assumed to be generated by one of K multinomial distributions (topics). The detailed generative process of the j-th word in the i-th document xi, j is:
Actually our “full model” (L = 1) is essentially the same as a topic model.
On the other hand, in population structure inference problems [12, 47], each individual is assumed to be an admixture of K ancestries in varying proportions, where each ancestry is characterized by the allele frequency at each SNP locus. Each SNP genotype of an individual is assumed to be generated by the two step model: first, an ancestry (“population”) is chosen according to the admixture proportion for each individual, and then the SNP genotype is generated according to the allele frequency of the selected ancestry at that locus. The relationships among the mutation signature models, topic models and population structure models are summarized in Table 1.
As pointed out by [48], there is a close relationship between mixed-membership models and nonnegative matrix factorization, which has been successfully used in the previous studies for mutational signature problems [7–9]. In fact, the proposed method can be seen as non-negative matrix factorization with additional restrictions. See S4 Text for details of the relationship between the proposed approach and nonnegative matrix factorization.
|
10.1371/journal.pgen.1006932 | Control of RUNX-induced repression of Notch signaling by MLF and its partner DnaJ-1 during Drosophila hematopoiesis | A tight regulation of transcription factor activity is critical for proper development. For instance, modifications of RUNX transcription factors dosage are associated with several diseases, including hematopoietic malignancies. In Drosophila, Myeloid Leukemia Factor (MLF) has been shown to control blood cell development by stabilizing the RUNX transcription factor Lozenge (Lz). However, the mechanism of action of this conserved family of proteins involved in leukemia remains largely unknown. Here we further characterized MLF’s mode of action in Drosophila blood cells using proteomic, transcriptomic and genetic approaches. Our results show that MLF and the Hsp40 co-chaperone family member DnaJ-1 interact through conserved domains and we demonstrate that both proteins bind and stabilize Lz in cell culture, suggesting that MLF and DnaJ-1 form a chaperone complex that directly regulates Lz activity. Importantly, dnaj-1 loss causes an increase in Lz+ blood cell number and size similarly as in mlf mutant larvae. Moreover we find that dnaj-1 genetically interacts with mlf to control Lz level and Lz+ blood cell development in vivo. In addition, we show that mlf and dnaj-1 loss alters Lz+ cell differentiation and that the increase in Lz+ blood cell number and size observed in these mutants is caused by an overactivation of the Notch signaling pathway. Finally, using different conditions to manipulate Lz activity, we show that high levels of Lz are required to repress Notch transcription and signaling. All together, our data indicate that the MLF/DnaJ-1-dependent increase in Lz level allows the repression of Notch expression and signaling to prevent aberrant blood cell development. Thus our findings establish a functional link between MLF and the co-chaperone DnaJ-1 to control RUNX transcription factor activity and Notch signaling during blood cell development in vivo.
| Tight regulation of proteins level is required for proper development. Notably, the aberrant expression of key transcription factors or signaling pathway components controlling blood cell development contributes to hematological diseases such as leukemia. In this report, we use Drosophila as a model to study the function and mode of action of a family of conserved but poorly characterized proteins implicated in leukemia called Myeloid Leukemia Factors (MLF). By combining proteomic, transcriptomic and genetic approaches, we show that Drosophila MLF acts in concert with an Hsp40 co-chaperone to control the level and activity of a RUNX transcription factor and therefore RUNX+ blood cell number and differentiation. Furthermore, we show that RUNX dosage directly impinges on the activity of the Notch signaling pathway, which is critical for RUNX+ cell survival and differentiation, by regulating the transcription of the Notch receptor. These findings shed light on a new mode of regulation of RUNX level and Notch activity to prevent abnormal blood cell accumulation, which could be involved in leukemogenesis.
| Proper blood cell development requires the finely tuned regulation of transcription factors and signaling pathways activity. Consequently mutations affecting key regulators of hematopoiesis such as members of the RUNX transcription factor family or components of the Notch signaling pathway are associated with several blood cell disorders including leukemia [1, 2]. Also, leukemic cells often present recurrent chromosomal rearrangements that participate in malignant transformation by altering the function of these factors [3]. The functional characterization of these genes is thus of importance not only to uncover the molecular basis of leukemogenesis but also to decipher the regulatory mechanisms controlling normal blood cell development. Myeloid Leukemia Factor 1 (MLF1) was identified as a target of the t(3;5)(q25.1;q34) translocation associated with acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) more than 20 years ago [4]. Further findings suggested that MLF1 could act as an oncogene [5–8] or a tumor suppressor [9] depending on the cell context and it was shown that MLF1 overexpression either impairs cell cycle exit and differentiation [10], promotes apoptosis [11, 12], or inhibits proliferation [13, 14] in different cultured cell lines. Yet, its function and mechanism of action remain largely unknown.
MLF1 is the founding member of a small evolutionarily conserved family of nucleo-cytoplasmic proteins present in all metazoans but lacking recognizable domains that could help define their biochemical activity [15]. Whereas vertebrates have two closely related MLF paralogs, Drosophila has a single mlf gene encoding a protein that displays around 50% identity with human MLF in the central conserved domain [16, 17]. In the fly, MLF was identified as a partner of the transcription factor DREF (DNA replication-related element-binding factor) [16], for which it acts a co-activator to stimulate the JNK pathway and cell death in the wing disc [18]. MLF has been shown to bind chromatin [18–20], as does its mouse homolog [21], and it can either activate or repress gene expression by a still unknown mechanism [18, 20]. MLF also interacts with Suppressor of Fused, a negative regulator of the Hedgehog signaling pathway [19], and, like its mammalian counterpart [13], with Csn3, a component of the COP9 signalosome [22], but the functional consequences of these interactions remain elusive. Interestingly the overexpression of Drosophila MLF or that of its mammalian counterparts can suppress polyglutamine-induced cytotoxicity in fly and in cellular models of neurodegenerative diseases [17, 23–25]. Moreover phenotypic defects associated with MLF loss in Drosophila can be rescued by human MLF1 [17, 26]. Thus MLF function seems conserved during evolution and Drosophila appears to be a genuine model organism to characterize MLF proteins [15].
Along this line, we recently analyzed the role of MLF during Drosophila hematopoiesis [26]. Indeed, a number of proteins regulating blood cell development in human, such as RUNX and Notch, also control Drosophila blood cell development [27]. In Drosophila, the RUNX factor Lozenge (Lz) is specifically expressed in crystal cells and it is absolutely required for the development of this blood cell lineage [28]. Crystal cells account for ±4% of the circulating larval blood cells; they are implicated in melanization, a defense response related to clotting, and they release their enzymatic content in the hemolymph by bursting [27]. The Notch pathway also controls the development of this lineage: it is required for the induction of Lz expression and it contributes to Lz+ cell differentiation as well as to their survival by preventing their rupture [28–31]. Interestingly, our previous analysis revealed a functional and conserved link between MLF and RUNX factors [26]. In particular, we showed that MLF controls Lz activity and prevents its degradation in cell culture and that the regulation of Lz level by MLF is critical to control crystal cell number in vivo [26]. Intriguingly, although Lz is required for crystal cell development, mlf mutation causes a decrease in Lz expression but an increase in crystal cell number. In human, the deregulation of RUNX protein level is associated with several pathologies. For instance haploinsufficient mutations in RUNX1 are linked to MDS/AML in the case of somatic mutations, and to familial platelet disorders associated with myeloid malignancy for germline mutations [1]. In the opposite, RUNX1 overexpression can promote lymphoid leukemia [32, 33]. Understanding how the level of RUNX protein is regulated and how this affects specific developmental processes is thus of particular importance.
To better characterize the function and mode of action of MLF in Drosophila blood cells, we used proteomic, transcriptomic and genetic approaches. In line with recent findings [20], we found that MLF binds DnaJ-1, a HSP40 co-chaperone, as well as the HSP70 chaperone Hsc70-4, and that both of these proteins are required to stabilize Lz. We further show here that MLF and DnaJ-1 interact together but also with Lz via conserved domains and that they regulate Lz-induced transactivation in a Hsc70-dependent manner in cell culture. In addition, using a null allele of dnaj-1, we show that it controls Lz+ blood cell number and differentiation as well as Lz activity in vivo in conjunction with mlf. Notably, we found that mlf or dnaj-1 loss leads to an increase in Lz+ cell number and size due to the over-activation of the Notch signaling pathway. Interestingly, our results indicate that high levels of Lz are required to repress Notch expression and signaling. We thus propose a model whereby MLF and DnaJ-1 control Lz+ blood cell growth and number by promoting Lz accumulation, which ultimately turndowns Notch signaling. These findings thus establish a functional link between the MLF/Dna-J1 chaperone complex and the regulation of a RUNX-Notch axis required for blood cell homeostasis in vivo.
To better characterize the molecular mode of action of MLF, we sought to identify its partners. Accordingly, we established a Drosophila Kc167 cell line expressing a V5-tagged version of MLF close to endogenous levels in a copper-inducible manner (Fig 1A). After anti-V5 affinity purification from whole cell extracts of control or MLF-V5-expressing cells, isolated proteins were identified by mass spectrometry. Five proteins reproducibly co-purified with MLF and were either absent or at more than 4 fold lower levels in each control purification (Fig 1B): the Hsp40 co-chaperone DnaJ-1 (also known as DROJ1; [34]), the constitutively expressed Hsp70 chaperones Hsc70-4 and Hsc70-3, the RNA binding protein Squid (Sqd), and the retrotransposon-encoded protein Copia. Of note, as this manuscript was in preparation, Dyer et al. also identified DnaJ-1 and Hsc70-4 as partners of MLF using a similar proteomic approach in the Drosophila S2 cell line [20]. Since DnaJ-1 was the strongest hit in our analysis, we focused on this candidate and we further characterized its interaction with MLF as well as its function both in cell culture and in vivo.
First, we confirmed the interaction between MLF and DnaJ-1 by co-immunoprecipitation assays in Kc167 cells transfected with expression plasmids for tagged versions of these proteins using anti-tag antibodies (Fig 1C and 1D, and S1 Fig) or an anti-MLF antibody (S1C Fig). In addition, consistent with the hypothesis that these proteins interact in the cell, immunostainings showed that DnaJ-1 and MLF co-localize in the nuclei of Kc167 transfected cells (S1D Fig). Finally, we also observed a specific interaction between MLF and DnaJ-1 by in vitro GST pull down assays (S1E Fig).
We then mapped the domains required for the interaction between DnaJ-1 and MLF. Hsp40/DnaJ co-chaperones play a crucial role in the regulation of protein folding and degradation; they chiefly act by delivering substrates to Hsp70/DnaK chaperones and stimulating their ATPase activity [35, 36]. DnaJ-1 belongs to the DnaJB/class II subfamily of Hsp40/DnaJ proteins, which are characterized by an N-terminal J-domain required to stimulate Hsp70 ATPase activity (amino acids 4 to 57 in DnaJ-1), a central glycine/phenylalanine (G/F)-rich region (amino acids 64 to 144), and a conserved C-terminal region (amino acids 157 to 320) that contains the client proteins binding domain followed by a dimerization interface [36]. Immunoprecipitations of GFP-MLF expressed with different HA-tagged DnaJ-1 variants indicated that the DnaJ-1 C-terminal region mediates MLF binding (Fig 1C). In contrast, a point mutation (P32S) in the highly conserved HPD loop crucial for Hsc70 activation [36], deletion of the J-domain or deletion of the J and G/F domains did not affect the interaction between DnaJ-1 and GFP-MLF. MLF does not harbor characteristic domains apart from a central “MLF homology domain” (MHD, amino acids 96 to 202) conserved between MLF family members [15]. Using GFP-DnaJ-1 as bait and MLF deletion mutants as preys, we found that the MHD was sufficient for binding DnaJ-1, while MLF N- and C-terminal regions were dispensable (Fig 1D). Finally, consistent with the above results, the C-terminal region (amino acids 157 to 334) of DnaJ-1 bound to the MHD of MLF (S1F Fig). In sum, these data indicate that MLF and DnaJ-1 specifically bind each other through their conserved central and C-terminal region, respectively.
We have previously shown that MLF is required for Lz stability and transcriptional activity [26]. Interestingly, Dyer et al. reported that the knockdown of DnaJ-1 or of its chaperone partner Hsc70-4 leads to a destabilization of exogenously expressed Lz in S2 cells [20]. However, the relationships between DnaJ-1, MLF and Lz were not further explored. We thus asked whether DnaJ-1 also controls Lz activity. As shown in Fig 2A, transfection of a Lz expression plasmid in Kc167 cells induced a robust activation of the 4xPPO2-Fluc reporter gene [37], which was significantly decreased when either mlf or dnaj-1 expression was knocked down by dsRNA treatment. Consistent with previous results [20, 26], Western blot analyses showed that mlf and dnaj-1 knockdowns caused a drop in Lz protein level (Fig 2B). Moreover, RT-qPCR experiments showed that mlf and dnaj-1 knockdowns did not affect the expression of each other or decrease lz transcript level, while they did cause a significant reduction in the expression of Lz target gene ppo2 (S2A–S2D Fig). Hence, like MLF, DnaJ-1 controls Lz protein stability and activity in Kc167 cells.
Next, we tested the effect of DnaJ-1 overexpression on Lz’s activity and protein level. Reminiscent of MLF [26], we observed that DnaJ-1 over-expression was associated with an increase in Lz-induced transactivation and Lz level (Fig 2C and 2D). The overexpression of C-terminally-truncated DnaJ-1 proteins did not affect Lz-induced transcription or its expression. In contrast, the overexpression of DnaJ-1 carrying the P32S point mutation or a deletion of its J-domain caused a decrease in Lz-induced transcription and a drop in Lz level (Fig 2C and 2D), suggesting that the activation of Hsc70 by DnaJ-1 is required for Lz’s stable expression and activity. In line with this hypothesis, knocking down Hsc70-4, which interacts with DnaJ-1 and MLF [20], caused a strong decrease in Lz-induced transactivation and a concomitant reduction in Lz protein level (S2E and S2F Fig). In sum, our results support the idea that MLF acts with DnaJ-1 in a Hsc70 chaperone complex to promote Lz stability and activity.
Given the impact of MLF and DnaJ-1 on Lz activity, we then asked whether these two proteins bind this RUNX transcription factor. Upon transfection of the corresponding expression plasmids, both HA-DnaJ-1 and HA-MLF were co-immunoprecipitated by GFP-tagged Lz but not by GFP alone (Fig 2E and 2F). Furthermore, in vitro translated Lz bound to E. coli-purified GST-MLF and GST-DnaJ-1 but not to GST alone in pull down assays (S2G Fig). Using different MLF variants in co-immunoprecipitation assays, we found that the N-terminal part of the MLF homology domain (amino acids 96 to 147) was crucial for the interaction with Lz (Fig 2G). Similarly the C-terminal domain of DnaJ-1 was required for binding Lz, while its J domain was dispensable (Fig 2H). Therefore it appears that MLF and DnaJ-1 interact with Lz through conserved domains and our results suggest that the MLF/DnaJ-1 complex regulates Lz stability and activity in Kc167 cells by binding it.
Since DnaJ-1 interacts with MLF and controls Lz level ex vivo, we then sought to analyze DnaJ-1 function in circulating larval crystal cells, whose proper development requires Lz stabilization by MLF [26]. Given that no mutant for dnaj-1 was available, we used a CRISPR/Cas9 strategy to generate dnaj-1 null alleles (S3 Fig) [38]. In the following experiments, we used an allelic combination between two mutant lines obtained from independent founder flies (dnaj-1A and dnaj-1C), which harbor a complete deletion of the dnaj-1 coding sequence (S3 Fig). Around 65% of the dnaj-1A/C mutants reached the larval stage and 15% emerged as adult flies but they did not show obvious morphological defects. Reminiscent of mlf phenotypes [26], bleeding of third instar larvae revealed that dnaj-1 mutants exhibited a ±1.8-fold increase in the number of circulating lz>GFP+ blood cells as compared to wild-type (Fig 3A). In addition, as in the mlf mutant, crystal cells from dnaj-1 mutant larvae still expressed the differentiation marker PPO1 and were capable of melanization upon heat treatment (Fig 3C–3H). A closer examination also revealed the presence of unusually large lz>GFP+ cells in the dnaj-1 mutant and quantitative analyses confirmed that dnaj-1 loss caused a significant increase in lz>GFP+ cell size whereas lz>GFP- cells were unaffected (Fig 3B). Interestingly, a similar phenotype is observed in mlf mutant larvae (Fig 3B), suggesting that both genes not only control crystal cell number but also their differentiation (see below). Importantly, lz>GFP+ cell number and size was restored to wild-type when DnaJ-1 was re-expressed in the crystal cell lineage of dnaj-1A/C mutant larvae using the lz-GAL4 driver (Fig 3A and 3B). This demonstrates not only that these phenotypes are specifically caused by the dnaj-1 mutation, but also that DnaJ-1 acts cell autonomously and after the onset of lz expression in the crystal cell lineage. Of note, we did not observe a rescue when we expressed a DnaJ-1 protein lacking its J-domain, suggesting that the interaction with Hsp70 chaperones is critical for the function of DnaJ-1 in the crystal cell lineage (S3C and S3D Fig). Furthermore, the increase in crystal cell number and size was also observed when we monitored crystal cell presence by immunostaining against PPO1 in larvae carrying a dnaj-1A or dnaj-1C homozygous mutation or over a deficiency uncovering the dnaj-1 locus (S3E and S3F Fig). Overall, these results demonstrate that, like mlf, dnaj-1 controls circulating larval lz>GFP+ cell number and size.
Since MLF and DnaJ-1 bind to each other, we tested whether they genetically interacted to regulate crystal cell development. While heterozygous mutation in either mlf or dnaj-1 did not significantly alter circulating lz>GFP+ cell number or size, mlfΔC1/+,dnaj-1A/+ transheterozygote larvae displayed a significant increase of both parameters (Fig 3I and 3J). We thus conclude that DnaJ-1 and MLF act together to control crystal cell development. In sum, these results reveal a functional interaction between MLF and DnaJ-1 in vivo.
Next we assessed whether DnaJ-1 affects Lz stability in vivo as it does in cell culture. Unexpectedly, immunostaining against Lz did not reveal a decrease in Lz expression in dnaj-1 mutant crystal cells while the level of Lz was clearly lower in the mlf mutant (Fig 4A–4C). Actually quantitative analyses revealed a slight (30%) but significant (p = 0.006) increase in Lz level in dnaj-1 mutant as compared to wild-type, whereas Lz level dropped by more than 2 folds in mlf mutant (Fig 4D). Thus, unlike mlf, dnaj-1 loss is not sufficient to destabilize Lz in vivo. We then tried to understand the reason for this discrepancy. One potentially important difference between Kc167 cells, in which DnaJ-1 is required to stabilize Lz, and crystal cells, in which it is not, is MLF expression. Indeed, in Kc167 cells, MLF is mainly detected in the cytoplasm and is present at low levels in the nucleus (S4A Fig). In contrast, MLF is present at high levels in the nucleus of larval crystal cells (S4B Fig). Moreover, MLF expression in this lineage is not affected by dnaj-1 loss (S4C and S4F Fig). We thus surmised that the presence of high levels of nuclear MLF might prevent Lz degradation in the absence of DnaJ-1.
To test this hypothesis, we designed two complementary experiments. On the one hand, we assessed whether MLF over-expression in Kc167 cells could protect Lz from degradation following dnaj-1 knockdown. Lz level dropped when Kc167 cells were treated with a dsRNA targeting dnaj-1 (Fig 4G) and increased upon over-expression of MLF (Fig 4F). Strikingly though, and in line with the observations in dnaj-1 mutant crystal cells, the level of Lz was not reduced but further increased when dnaj-1 was knocked down in MLF-overexpressing cells (Fig 4H and 4I). On the other hand, we asked whether Lz would still be stable in dnaj-1 mutant crystal cells if MLF level is decreased. Accordingly, we expressed a dsRNA directed against mlf in lz>GFP+ cells, which caused a significant and similar knock-down of MLF in wild-type and dnaj-1 mutant larvae (S4D–S4F Fig). Remarkably, we found that the drop in Lz protein level caused by mlf down-regulation was significantly enhanced in dnaj-1 deficient larvae, while the dnaj-1 mutation alone increased Lz level (Fig 4J–4N). Hence it appears that in the absence of DnaJ-1, high levels of MLF prevent Lz degradation.
Given that chaperones are important for proper protein folding [35, 36], we postulated that Lz proteins accumulating in crystal cells in the absence of DnaJ-1 might be less active. Thus increasing Lz expression might be sufficient to rescue lz>GFP+ cell number and size. In addition, although re-expressing Lz is sufficient to restore lz>GFP+ cell number in mlf mutant larvae [26], it is not known whether this also rescues lz>GFP+ cell size. Interestingly, lz>GFP+ cell count and cell size were restored to wild-type levels when we enforced Lz expression in this lineage either in mlf or dnaj-1 mutant larvae (Fig 4O and 4P). We thus conclude that DnaJ-1 and MLF act together to control crystal cell development by regulating Lz activity in vivo
In parallel, to gain further insights into the function of MLF in the control of crystal cell development, we established the transcriptome of circulating lz>GFP+ blood cells in wild-type and mlf larvae. Heterozygous lz-GAL4,UAS-mCD8-GFP L3 larvae carrying or lacking a mlf null mutation were bled, lz>GFP+ cells were collected by FACS and their gene expression profile was determined by RNA sequencing (RNAseq) from biological triplicates. Using Drosophila reference genome dm3, we detected the expression of 7399 genes (47% of the total fly genes) in each of the 6 samples (Fig 5B and S1 Table). Consistent with the role of the crystal cells as the main source of phenoloxidases [39], the two most strongly expressed genes were PPO1 and PPO2. In addition, lz expression as well as that of several other crystal cell markers was readily detected (see below). It was recently shown that larval circulating Lz+ cells derive from plasmatocytes, which express Hemolectin (Hml) and Nimrod C1 (NimC1), and transdifferentiate into crystal cells [40]. Accordingly, we detected the expression of these genes, as well as other “plasmatocytes” markers such as peroxidasin and croquemort (which were actually shown to be also expressed in crystal cells [41, 42]) in lz>GFP+ cells.
Using DESeq2 to identify differentially expressed genes between wild-type and mlf mutant lz>GFP+ cells, we found 779 genes with significantly altered expression (adjusted p-value <0.01): the transcript level of 469 genes was decreased and that of 310 genes was increased in the absence of MLF (Fig 5A and 5B, and S2 Table). In line with our previous in situ hybridization results [26], RNAseq analysis did not reveal a significant modification of PPO1 or PPO2 expression in the absence of mlf. However, the lz transcript level was reduced by ±2 fold (p-value = 0.0018), which could be due to defective maintenance of the lz auto-activation loop [43]. To assess whether other crystal cell markers were affected by mlf, we established a compilation of genes expressed in (embryonic or larval) crystal cells based on Flybase data mining and re-examination of Berkeley Drosophila Genome Project in situ hybridizations (http://insitu.fruitfly.org/cgi-bin/ex/insitu.pl) (S3 Table). Among these 129 genes (i.e. excluding mlf itself), 44 (34%) were differentially expressed in the absence of mlf (19 repressed and 25 activated) (Fig 5C), indicating a strong over-representation of deregulated gene in the “crystal cell” gene set as compared to all expressed genes (p-value = 2.6x10-13, hypergeometric test) and showing that mlf plays a crucial role for proper crystal cell differentiation.
To substantiate these results, we analyzed by in situ hybridization the expression of 4 genes that were either down-regulated (CG7860 and Oscillin) or up-regulated (CG6733 and Jafrac1) in the mlf mutant. CG7860 and Oscillin were specifically expressed in lz>GFP+ but not in the surrounding lz>GFP- hemocytes in wild-type conditions (Fig 5D and 5G). Consistent with our RNAseq data, the expression of CG7860 and Oscillin was strongly reduced in mlf mutant larvae. Although CG6733 is expressed in embryonic crystal cells [43], we did not detect its expression in circulating hemocytes of wild-type larvae, but it was expressed in the lz>GFP+ lineage in mlf larvae (Fig 5J and 5K). Finally, Jafrac1 expression increased in lz>GFP+ cells of mlf mutant larvae as compared to wild-type, whereas its (lower) expression in lz>GFP- blood cells seemed similar (Fig 5M and 5N). These data thus confirm the RNAseq results and demonstrate that MLF controls the expression of several crystal cell markers. Since the above results indicate that MLF functionally interacts with DnaJ-1 during crystal cell development, we also tested whether these four genes were deregulated in the dnaj-1 mutant. As for mlf, we observed that a dnaj-1 mutation caused a down-regulation of CG7860 and Oscillin and an up-regulation of CG6733 and Jafrac1 expression in lz>GFP+ blood cells (Fig 5F, 5I, 5L and 5O).
In sum it appears that the loss of mlf or dnaj-1 leads to a deregulation of the crystal cell gene expression program characterized both by the overexpression and the downregulation of crystal cell markers. Therefore mlf and dnaj-1 are required for proper differentiation of the Lz+ blood cell lineage.
Interestingly, the levels of Notch receptor transcripts were significantly higher in the mlf mutant (p = 1.3x10-6) (Fig 5C). Notch signaling plays a key role in crystal cell development [27]: Notch is first activated by its ligand Serrate to specify Lz+ cells (crystal cell precursors) and its activation is subsequently maintained in Lz+ cells in a ligand-independent manner to promote crystal cell growth and survival [29–31, 40, 44]. The rise in lz>GFP+ cell number and size observed in mlf and dnaj-1 mutant could thus be due to increased ligand-independent Notch signaling. However, the role of Notch signaling in crystal cell growth and survival has been mainly investigated in the larval lymph gland [30, 31]. In agreement with these investigations, inhibiting the Notch pathway in circulating Lz+ cells, either by down-regulating the expression of Suppressor of Hairless [Su(H)], the core transcription factor in the Notch pathway, or by overexpressing Suppressor of Deltex [Su(dx)], a negative regulator of Notch [45], resulted in a decrease in lz>GFP+ cell number and impaired their growth, whereas the overactivation of Notch signaling consecutive to the expression of a constitutively active Su(H)-VP16 fusion protein [46], caused a strong increase in lz>GFP+ cell number and size (S5 Fig).
Then we further investigated the level of Notch expression and activation in mlf and dnaj-1 mutant blood cells. Immunostaining using an antibody against the Notch extracellular domain (NECD) showed that Notch accumulated at higher levels in lz>GFP+ cells of mlf and dnaj-1 mutant larvae than in wild-type conditions (Fig 6A–6C). Quantitative analyses confirmed that mlf loss caused a significant increase in Notch level in lz>GFP+ cell, whereas the (lower) expression of Notch in lz>GFP- blood cells was not affected (Fig 6D). Similar results were obtained when we measured Notch protein levels using an antibody directed against its intra-cellular domain (NICD) (Fig 6E and S6 Fig). Thus Notch level is specifically increased in lz>GFP+ cells of mlf and dnaj-1 mutants. Next, we tested whether this resulted in increased signaling by monitoring the expression of two Notch signaling pathway reporter genes expressed in larval crystal cells: Klumpfuss-Cherry [31] and NRE-GFP [47]. Both mlf and dnaj-1 loss were associated with a strong increase in the expression of these reporters (Fig 6F–6J). Thus mlf and dnaj-1 are required to tune down Notch signaling in the crystal cell lineage. Finally, we asked whether the rise in lz>GFP+ cell size and/or number observed in mlf and dnaj-1 mutants depends on Notch. Strikingly, when we reduced Notch dosage by introducing one copy of the N55e11 null allele in these mutants, both parameters were restored to control levels, while N55e11 heterozygote mutation had no effect per se (Fig 6K and 6L). Collectively, these data strongly support the hypothesis that the increase in Notch level underlies lz>GFP+ cell expansion in mlf and dnaj-1 mutants.
It was shown that crystal cells tend to increase their size as they mature in response to Notch signaling [31, 40], which is consistent with the results we obtained by manipulating Notch signaling activity in Lz+ cells (S5 Fig). To better characterize the defects associated with mlf or dnaj-1 loss, we analyzed the distribution of lz>GFP+ cells as well as Notch level according to lz>GFP+ cell size categories. Whereas cells more than 1.3-fold larger than the mean wild-type cell size represented a small fraction (±10%) of the lz>GFP+ population in wild-type larvae, they constituted the prevalent population in mlf or dnaj-1 mutant (respectively 49.6 and 37%) (Fig 7A). Interestingly, Notch protein level was maximum in the population of lz>GFP+ cells of mean cell size but lower in larger cells of wild-type larvae (Fig 7B), whereas it continued to increase in the larger cell populations of mlf or dnaj-1 larvae (Fig 7B–7D). Actually we observed a similar trend when we monitored Notch expression by in situ hybridization. In wild-type larvae, Notch transcripts were readily seen in small/medium lz>GFP+ cells but barely detectable in large lz>GFP+ cells (Fig 7E and 7F). In contrast, Notch transcripts continued to accumulate in large lz>GFP+ cells from mlf or dnaj-1 mutant larvae (Fig 7H and 7J). Hence, MLF/DnaJ-1 loss is associated with the accumulation of large crystal cells exhibiting aberrant maintenance of Notch expression. Since the Notch pathway is activated in a ligand-independent manner in Lz+ cells [30], a tight regulation of the level of Notch is particularly critical to control crystal cell growth and number. All together, our data suggest that in mlf or dnaj-1 mutant larvae, Notch expression fails to be turned down when lz>GFP+ cells reach a critical size, leading to the maintenance of a high level of Notch signaling and thus to increased crystal cell growth and survival.
We showed above that forcing the expression of Lz rescues the increase in crystal cell number and size caused by mlf or dnaj-1 loss. It is thus plausible that this RUNX transcription factor directly participates in down-regulation of Notch signaling. To explore this hypothesis, we asked whether a reduction in lz activity might cause an expansion of the Lz+ cell lineage associated with an over-activation of the Notch pathway. Accordingly, we introduced the lzr1 null allele into the lzGAL4 context. This hypomorphic allelic combination caused a decrease in Lz expression (Fig 8B) and resulted in an increase in lz>GFP+ cell number and size (Fig 8E and 8F). Interestingly, lzGAL4/Y hemizygous larvae displayed similar phenotypes (Fig 8C, 8E and 8F), indicating that this P{GAL4} insertion in lz alters its expression in the crystal cell lineage. As an alternate strategy, we interfered with Lz activity by expressing a fusion protein between Lz partner Brother (Drosophila CBFß homolog) and the non-muscular myosin heavy chain SMMHC [48]. This chimera mimics the CBFß-MYH11 fusion protein generated by the Inv(16) translocation in human AML and can sequester RUNX factors in the cytoplasm [1, 49]. Bro-SMMHC expression in lz>GFP+ cells titrated Lz from the nucleus and also caused an increase in lz>GFP+ cell number and size (Fig 8D–8F). Importantly, the expression of the Notch pathway reporters NRE-GFP and Klu-Cherry was strongly increased in lzGAL4/lzR1 mutant or upon Bro-SMHCC expression in the Lz+ blood cell lineage (Fig 8G and 8H). Moreover, knocking down Su(H) or over-expressing the Notch protein inhibitor Su(dx) was sufficient to prevent the rise in lz>GFP+ cell number and size of lzGAL4/Y hemyzigotes (S5 Fig). Thus, a reduction in lz activity causes similar defects as the mlf or dnaj-1 mutations and likely involves the overactivation of the Notch pathway.
Then we analyzed the relathionship between Lz and Notch levels. In Lz+ cells of increasing size, Lz levels continuously increased while Notch became less abundant (S7A Fig). This suggested that Lz level rises as crystal cells grow/mature and, in view of the above results, we surmised that this increase might participate in the down-regulation of the Notch receptor. Indeed, we found that the Notch receptor level was significantly augmented in lz>GFP+ cells of hypomorphic lzGAL4/Y hemyzigote mutant larvae, whereas it was reduced when Lz was over-expressed (Fig 9A–9E). In addition, the increase in Notch expression observed in lzGAL4/Y larvae was suppressed by forcing Lz expression. Moreover, in situ hybridization experiments revealed that, unlike in control larvae, Notch expression was not repressed in large lz>GFP+ cells in lzGAL4/Y larvae (S7 Fig). Therefore Notch might be a direct transcriptional target of Lz. By analyzing the expression of different GAL4 lines that cover potential Notch regulatory regions [50], we identified two lines that drive expression in circulating Lz+ blood cells (Fig 9F and S7 Fig). The regulatory elements carried by these two lines (GMR30A01 and GMR30C06) overlap on a 668bp DNA segment that contains two consensus binding sites for RUNX transcription factors conserved in other Drosophila species (S7A Fig), suggesting that Lz might directly regulate Notch transcription by targeting this region. We thus tested the effect of Lz dosage manipulation on the activity of this enhancer-GAL4 line. Strikingly, a hypomorphic lozenge mutation (lzg/Y) [51] or the expression of Bro-SMMHC caused an increase in the expression of this enhancer, whereas the over-expression of Lz resulted in its down-regulation (Fig 9G–9K). These findings strongly argue that Lz directly represses Notch expression.
All together, these results demonstrate that high levels of Lz are required to prevent the accumulation of over-grown lz>GFP+ cells as well as over-activation of the Notch pathway, and we propose that Lz-mediated repression of Notch transcription is critical during this process.
Members of the RUNX and MLF families have been implicated in the control of blood cell development in mammals and Drosophila and deregulation of their expression is associated with human hemopathies including leukemia [1, 9, 15, 52]. Our results establish the first link between the MLF/DnaJ-1 complex and the regulation of a RUNX transcription factor in vivo. In addition, our data show that the stabilization of Lz by the MLF/DnaJ-1 complex is critical to control Notch expression and signaling and thereby blood cell growth and survival. These findings pinpoint the specific function of the Hsp40 chaperone DnaJ-1 in hematopoiesis, reveal a potentially conserved mechanism of regulation of RUNX activity and highlight a new layer of control of Notch signaling at the transcriptional level.
In line with results published as this manuscript was in preparation [20], we found that MLF binds DnaJ-1 and Hsc70-4 and that these two proteins, like MLF, are required for Lz stable expression in Kc167 cells. In addition, our data show that MLF and DnaJ-1 bind to each other via evolutionarily conserved domains and also interact with Lz, suggesting that Lz is a direct target of a chaperone complex formed by MLF, DnaJ-1 and Hsc70-4. Of note, a systematic characterization of Hsp70 chaperone complexes in human cells identified MLF1 and MLF2 as potential partners of DnaJ-1 homologs, DNAJB1, B4 and B6 [53], a finding corroborated by Dyer et al. [20]. Therefore, the MLF/DnaJ-1/Hsc70 complex could play a conserved role in mammals, notably in the regulation of the stability of RUNX transcription factors. How MLF acts within this chaperone complex remains to be determined. In vivo, we demonstrate that dnaj-1 mutations lead to defects in crystal cell development strikingly similar to those observed in mlf mutant larvae and we show that these two genes act together to control Lz+ cells development by impinging on Lz activity. Our data suggest that in the absence of DnaJ-1, high levels of MLF lead to the accumulation of defective Lz protein whereas lower levels of MLF allow its degradation. We thus propose that MLF stabilizes Lz and, together with DnaJ-1, promotes its proper folding/conformation. In humans, DnaJB4 stabilizes wild-type E-cadherin but induces the degradation of mutant E-cadherin variants associated with hereditary diffuse gastric cancer [54]. Thus the fate of DnaJ client proteins is controlled at different levels and MLF might be an important regulator in this process.
In this work, we present the first null mutant for a gene of the DnaJB family in metazoans and our results demonstrate that a DnaJ protein is required in vivo to control hematopoiesis. There are 16 DnaJB and in total 49 DnaJ encoding genes in mammals and the expansion of this family has likely played an important role in the diversification of their functions [55, 56]. DnaJB9 overexpression was found to increase hematopoietic stem cell repopulation capacity [57] and Hsp70 inhibitors have anti-leukemic activity [58], but the participation of other DnaJ proteins in hematopoiesis or leukemia has not been explored. Actually DnaJ’s molecular mechanism of action has been fairly well studied but we have limited insights as to their role in vivo. Interestingly though, both DnaJ-1 and MLF suppress polyglutamine protein aggregation and cytotoxicity in Drosophila models of neurodegenerative diseases [17, 23, 24, 59–63, 64], and this function is conserved in mammals [24, 25, 65, 66]. It is tempting to speculate that MLF and DnaJB proteins act together in this process as well as in leukemogenesis. Thus a better characterization of their mechanism of action may help develop new therapeutic approaches for these diseases.
As shown here, mlf or dnaj-1 mutant larvae harbor more crystal cells than wild-type larvae. This rise in Lz+ cell number is not due to an increased induction of crystal cell fate as we could rescue this defect by re-expressing DnaJ-1 or Lz with the lz-GAL4 driver, which turns on after crystal cell induction, and it was also observed in lz hypomorph mutants, which again suggests a post-lz / cell fate choice process. Moreover mlf or dnaj-1 mutant larvae display a higher fraction of the largest lz>GFP+ cell population, which could correspond to the more mature crystal cells [31, 40]. It is thus tempting to speculate that mlf or dnaj-1 loss promotes the survival of fully differentiated crystal cells. Our RNAseq data demonstrate that mlf is critical for expression of crystal cell associated genes, but we observed both up-regulation and down-regulation of crystal cell differentiation markers in mlf or dnaj-1 mutant Lz+ cells. Also these changes did not appear to correlate with crystal cell maturation status since we found alterations in gene expression in the mutants both in small and large Lz+ cells. In addition our transcriptome did not reveal a particular bias toward decreased expression for “plasmatocyte” markers in Lz+ cells from mlf- mutant larvae. Thus, it appears that MLF and DnaJ-1 loss leads to the accumulation of mis-differentiated crystal cells.
Our data support a model whereby MLF and DnaJ-1 act together to promote Lz accumulation, which in turn represses Notch transcription and signaling pathway to control crystal cell size and number (Fig 10). Indeed, we observe an abnormal maintenance of Notch expression in the larger Lz+ cells as well as an over-activation of the Notch pathway in the crystal cell lineage of mlf and dnaj-1 mutants or when we interfere with Lz activity. Moreover our data as well as previously published experiments show that Notch activation promotes crystal cell growth and survival [30, 31, 40]. Importantly too the increase in Lz+ cell number and size observed in mlf or dnaJ-1 mutant is suppressed when Notch dosage is decreased. Yet, some of the mis-differentiation phenotypes in the mlf or dnaj-1 mutants might be independent of Notch since changes in crystal cell markers expression seem to appear before alterations in Notch are apparent. At the molecular level, our results suggest that Lz directly represses Notch transcription as we identified a Lz-responsive Notch cis-regulatory element that contains conserved RUNX binding sites. The activation of the Notch pathway in circulating Lz+ cells is ligand-independent and mediated through stabilization of the Notch receptor in endocytic vesicles [30, 45]. Hence a tight control of Notch expression is of particular importance to keep in check the Notch pathway and prevent the abnormal development of the Lz+ blood cell lineage. Notably, Notch transcription was shown to be directly activated by Notch signaling [67]. Such an auto-activation loop might rapidly go awry in a context in which Notch pathway activation is independent of ligand binding. By promoting the accumulation of Lz during crystal cell maturation, MLF and DnaJ-1 thus provide an effective cell-autonomous mechanism to inhibit Notch signaling. Further experiments will now be required to establish how Lz represses Notch transcription. RUNX factors can act as transcriptional repressors by recruiting co-repressor such as members of the Groucho family [68]. Whether MLF and DnaJ-1 directly contribute to Lz-induced-repression in addition to regulating its stability is an open question. MLF and DnaJ-1 were recently found to bind and regulate a common set of genes in cell culture [20]. They may thus provide a favorable chromatin environment for Lz binding or be recruited with Lz and/or favor a conformation change in Lz that allows its interaction with co-repressors. The scarcity of lz>GFP+ cells precludes a biochemical characterization of Lz, MLF and DnaJ-1 mode of action notably at the chromatin level but further genetic studies should help decipher their mode of action. While the post-translational control of Notch has been extensively studied, its transcriptional regulation seems largely overlooked [69]. Our findings indicate that this is nonetheless an alternative entry point to control the activity of this pathway. Given the importance of RUNX transcription factor and Notch signaling in hematopoiesis and blood cell malignancies [1, 2], it will be of particular interest to further study whether RUNX factors can regulate Notch expression and signaling during these processes in mammals.
In conclusion, our study shows that MLF and DnaJ-1 act together to regulate RUNX transcription factor activity, which in turn controls Notch signaling during hematopoiesis in vivo. We anticipate that the extraordinary genetic toolbox available in Drosophila will help shed further light on the mechanism of action of these evolutionarily conserved proteins and will bring valuable insights into the control of protein homeostasis by MLF and DnaJ-1 during normal or pathological situations.
The following Drosophila melanogaster lines were used: mlfΔC1, UAS-mlf [17], UAS-ds-mlf (National Institute of Genetics), UAS-lz, lzGAL4,UAS-mCD8-GFP, lzg, lzr1, N55e11, UAS-dsSu(H), P{EPgy2}DnaJ-1EY04359, UAS-dnaj-1, Def(3L)BSC884, vas-Cas9, UAS-GFPnls, NRE-GFP, GMR30C06, GMR30A01, UAS-dsSu(H) (Bloomington Drosophila Stock Center), Bc-GFP [70], Klu-mCherry [31] UAS-Bro-SMMHC [48], UAS-DnaJ-1ΔJ [61], UAS-dsSu(H), UAS-Su(H)-VP16 [46], UAS-Su(dx) [71]. To generate dnaj-1 deficient flies, we designed two guide RNA targeting dnaj-1 locus (S4 Fig) and the corresponding DNA oligonucleotides (g2: GTCGACCACAACGCGCCGGATCAA; g3: GTCGCATCACAGTCACGCTTTCCT) were cloned in pCFD3 (Addgene). vas-cas9 females were crossed to P{EPgy2}DnaJ-1EY04359 males and the resulting embryos were injected using standard procedures with both pCFD3-g2 and pCFD3-g3 plasmids (500ng/ul). Deletion of the P{EPgy2}EY04359 transposon, as revealed by loss of the w+ marker, was screened for at the F2 generation, and deletion of dnaj-1 locus was assessed by PCR and sequencing.
All crosses were conducted at 25°C on standard food medium as described in [72].
For each sample, four third instar larvae were bled (or 5.103 Kc167 cells were dispensed) in 1ml of PBS in 24-well-plate containing a glass coverslip. Unless mentioned otherwise, only female larvae were used. The hemocytes were centrifuged for 2 min at 900g, fixed for 20 min with 4% paraformaldehyde in PBS and washed twice in PBS. For immunostainings: cells were permeabilized in PBS-0.3% Triton (PBST) and blocked in PBST- 1% Bovine Serum Albumin (BSA). The cells were incubated with primary antibodies at 4°C over night in PBST-BSA, washed in PBST, incubated for 2h at room temperature with corresponding Alexa Fluor-labeled secondary antibodies (Molecular Probes), washed in PBST and mounted in Vectashield medium (Eurobio-Vector) following incubation with Topro3 (ThermoFisher). The following antibodies were used: anti-Lz, anti-Notch intracellular domain, anti-Notch extracellular domain (Developmental Studies Hybridoma Bank, DSHB), anti-MLF [73], anti-PPO1 [74], anti-GFP (Fisher Scientific), anti-HA (Sigma).
For in situ hybridizations: after fixation, the cells were washed and permeabilized in PBS- 0.1% Tween 20 (PBSTw), pre-incubated for 1h at 65°C in HB buffer (50% formamide, 2x SSC, 1 mg/ml Torula RNA, 0.05 mg/ml Heparin, 2% Roche blocking reagent, 0.1% CHAPS, 5 mM EDTA, 0.1% Tween 20) and incubated over-night with anti-sense DIG-labeled RNA probes (against CG6733, CG7860, Jafrac, Notch and Oscillin) diluted in HB. The samples were washed in HB for 1h at 65°C, in 50% HB- 50% PBSTw for 30 min at 65°C and three times in PBSTw for 20 min at room temperature. Then the cells were incubated for 30 min in PBSTw- 1% BSA before being incubated with anti-DIG antibody conjugated to alkaline phosphatase (Roche, 1/2000 in PBSTw) for 3h. After 4 washes in PBSTw, in situ hybridization signal was revealed with FastRed (Roche). The cells were then processed for immunostaining against GFP as described above, incubated in Topro3, washed in PBS and mounted in Vectashield medium for analysis.
Experiments were performed using at least biological triplicates. Samples were imaged with laser scanning confocal microscopes (Leica) and images were analyzed with ImageJ. Cell size and protein expression levels were measured on maximal intensity projections of Z-sections through the whole cell on a minimum of 25 cells per genotype.
The following previously described plasmids were used: pAc-Lz-V5, 4xPPO2-Firefly luciferase (originally named 4xPO45-Fluc, [37]), pAc-MLF [17]. We generated the following Drosophila expression plasmids for C-terminally tagged or N-terminally tagged proteins using standard cloning techniques: pAc-Lz-EGFP, pAc-MLF-EGFP, pMT-MLF-V5-His, pAc-DnaJ-J1-EGFP, pAc-Hsc70-4-EGFP, pAc-3xHA-DnaJ-1 (2–334), pAc-3xHA-DnaJ-1 (P32S), pAc-3xHA-DnaJ-1 (58–334), pAc-3xHA-DnaJ-1 (2–156), pAc-3xHA-DnaJ-1 (2–191), pAc-3xHA-DnaJ-1 (2–269), pAc-3xHA-DnaJ-1 (157–334), pAc-3xHA-MLF (2–309), pAc-3xHA-MLF (2–147), pAc-3xHA-MLF (2–202), pAc-3xHA-MLF (202–309), pAc-3xHA-MLF (148–309), pAc-3xHA-MLF (96–309), pAc-3xHA-MLF (96–202). DnaJ-1 and MLF cDNA were also cloned into pBlueScript II to generate pBS-DnaJ-1 and pBS-MLF and in pGEX-2T to generate pGEX-DnaJ-1 and pGEX-MLF. All constructs were verified by sequencing.
Drosophila Kc167 cells were grown at 25°C in Schneider medium (Invitrogen) supplemented with 10% fetal bovine serum (FBS) and 50 μg/ml of penicillin/streptomycin (Invitrogen). For RNAi experiments, double stranded RNA duplexes (dsRNA) corresponding to 400-600bp exonic regions were produced using T7 promoter-containing primers and MEGAscript T7 transcription kit (Ambion). After an annealing step, dsRNA probes were purified using the RNeasy cleanup protocol (Qiagen). Independent dsRNA targeting different regions of dnaj-1 were produced. The sequences of the T7-containing primers used to generate the dsRNA are available on request. Cells were seeded at 2x106/ml on dsRNA (16 μg/well for 6-well-plate, 8 μg for 12-well-plate and 1 μg for 96-well-plate) and incubated in Schneider medium without FBS for 40 min before being transferred to 5% FBS containing medium. 24h later, cells were transfected with the plasmids of interest using Effectene (Qiagen) and they were collected 72h later for subsequent analyses.
For luciferase assays, 50 ng of 4xPPO2-Firefly luciferase reporter plasmid, were contransfected with 20 ng of pAc-Renilla luciferase plasmid, 10 ng of pAc-Lz-V5 and/or 10 ng of pAc expression plasmid for the protein of interest in 96 well-plate. Firefly and Renilla luciferases activities were measured 72h after transfection using Promega Dual luciferase reporter assay. Three biological replicates were performed for each transfection assay.
For RT-qPCR, RNAs were prepared from Kc167 cells using RNeasy kit (Qiagen) with an additional on-column DNAse treatment step. 1 μg of total RNA was used for reverse transcription using Superscript II and random primers (Invitrogen). 10 μl of a 1/300 dilution of cDNA was used as template for real time PCR using HOT Pol Evagreen qPCR mix (Bio-rad). The sequences of the primers used to assess the expression of dnaj-1, mlf, lz, PPO2, Renilla luciferase and rp49 are available upon request. All experiments were performed using biological triplicates or quadruplicates.
pET-3c-Lz, pBS-MLF and pBS-DnaJ-1 plasmids were used as template to produce 35S-methionine-labeled proteins in vitro using Rabbit Reticulocyte Lysate coupled transcription-translation system (Promega). pGEX-2T, pGEX-MLF and pGEX-DnaJ-1 were used to produce GST, GST-MLF and GST-DnaJ-1 in Escherichia coli BL21. Equivalent amounts of GST purified proteins immobilized on Gluthation-Sepharose beads were used to pull down Lz, MLF or DnaJ-1. Proteins were incubated for 2h at 4°C in buffer A (20 mM Tris–HCl, pH 8.0, 150 mM NaCl, 10 mM KCl, 1 mM EDTA, 0.1mg/ml BSA, 1 mM DTT, 0.05% NP40). After extensive washing in buffer buffer B (20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 1 mM EDTA, 1mM DTT, 0.05% NP40), bound proteins were eluted in SDS-loading buffer, separated by SDS–PAGE and visualized by autoradiography.
Kc167 cells were collected, washed in PBS and incubated for 30 min in IP buffer (150 mM NaCl, 0.5% NP40, 50 mM Tris-HCl, pH8.0, 1mM EGTA) supplemented with protease inhibitor cocktail (Roche). The extracts were cleared by centrifugation at 13.000g for 15 min at 4°C and subjected to SDS-PAGE (50 μg of proteins par lane) or immunoprecipitation (1 mg per point). For immunoprecipitation, proteins were preadsorbed with 100 μl of sepharose beads slurry for 1h at 4°C before being incubated with 20 μl of anti-GFP (Chromotek), anti-V5 (Sigma-Aldrich) or anti-HA (Covance) antibody coupled to sepharose beads, or with 10 μl of rabbit anti-MLF [19] or rabbit IgG (SantaCruz) in the presence of 20 μl of protein A sepharose beads (Sigma), for 4h at 4°C. The beads were spun down and washed in IP buffer and immunoprecipitated proteins were processed for SDS-PAGE and Western Blot analyses. Western blots were performed using standard techniques and the blots were developed by photoluminescence procedure using Lumi-LightPLUS Western Blotting Substrate (Roche) and Amersham HyperfilmTM ECL (GE Healthcare) or Chemidoc Touch Imaging System (BioRad). The following antibodies were used for Western blots: anti-V5 (Invitrogen), anti-HA (BioLegend), anti-GFP, anti-tubulin (Sigma-Aldrich), anti-Renilla luciferase (MBL), and anti-MLF [19].
Stable Kc167 cells carrying an inducible expression vector for MLF were obtained by cotransfecting pMT-MLF-V5-His and pCoBlast (Thermo Fisher Scientific) expression plasmids and selecting individual clones with 25μg/ml blasticidin. For affinity purification, MLF-inducible or parental Kc167 cells were seeded at 106/ml and cultivated for 24h in the presence of 50 mM CuSO4 to induce MLF expression. 20 mg of proteins extracted in IP buffer were then incubated on 200 μl of anti-V5 coupled sepharose beads (Sigma-Aldrich) or 400 μl of anti-V5 coupled magnetic beads (MBL). After several washes in IP buffer, affinity purified proteins were eluted in Laemmli buffer, reduced in 30 mM DTT and alkylated with 90 mM Iodoacetamide before being loaded on 12% SDS-PAGE. The single band of proteins was cut and digested overnight at 37°C with 1 μg of Trypsin (Promega) in 50 mM NH4CO3. Digested peptides were extracted from the gel by incubating 15 min at 37°C in 50 mM NH4CO3 and twice for 15 min at 37°C in 5% formic acid/acetonitrile (1:1). The dried peptide extracts were dissolved in 17 μl of 2% acetonitrile, 0.05% trifluoroacetic acid and the peptide mixtures were analyzed by nanoLC-MS/MS using an Ultimate3000-RS system (Dionex) coupled to an LTQ-Orbitrap Velos mass spectrometer (Thermo Fisher Scientific). 5 μl of each peptide extract were loaded on a 300 μm ID x 5 mm PepMap C18 precolumn (LC Packings, Dionex,) at 20 μl/min in 5% acetonitrile, 0.05% trifluoroacetic acid. After 5 minutes desalting, peptides were online separated on a 75 μm ID x 50 cm C18 Reprosil C18 column. The flow rate was set at 300 nl/min. Peptides were eluted using a 0 to 50% linear gradient of solvent B (solvent A: 0.2% formic acid in 5% acetonitrile, solvent B: 0.2% formic acid in 80% acetonitrile) for 80 min at 300nl/min. The LTQ Orbitrap was operated in data-dependent acquisition mode with the XCalibur software (version 2.0 SR2, Thermo Fisher Scientific), on the 350–1800 m/z mass range with the resolution set to a value of 60 000. The twenty most intense ions per survey scan were selected for CID-MS/MS fragmentation and the resulting fragments were analyzed in the linear ion trap (parallel mode). A 60 s dynamic exclusion window was used to prevent repetitive selection of the same peptide. The Mascot Daemon software (version 2.2.0, Matrix Science, London, UK) was used for protein identification against a non-redundant SwissProt database. Mascot results were parsed with Mascot File Parsing and Quantification (MFPaQ) version 4.0 [75]. Quantification of proteins was performed using the label-free module of the MFPaQ software, where a protein abundance index based on the average of peak area values for the three most intense tryptic peptides of the protein was calculated [76]. Triplicate injections were performed.
RNAseq experiments were performed using independent biological triplicates. For each sample, around 150 third instar larvae of control (lz-GAL4,UAS-mCD8GFP/+) or mlf mutant (lz-GAL4,UAS-mCD8GFP/+, mlf∂C1/mlf∂C1) genotypes were bled in ice-cold PBS. The hemocytes were centrifuged through a 40 μm mesh at 1000 rpm for 1 min and lz>GFP+ cells were collected by FACS (FacsAria II) under a pressure of 20 psi. A fraction of the collected cells were used to control GFP+ cell purification specificity by examination under an epifluorescent microscope after fixation and mounting in Vectashield medium with DAPI. RNAs were extracted from sorted cells using Arcturus PicoPure RNA kit (Applied Biosystems). RNA samples were run on Agilent Bioanalyzer to assess RNA integrity and concentration. The NuGEN Ovation RNASeq system with Ribo-SPIA technology was used to prepare the cDNA according to the manufacturer instruction. Library preparation was performed using the Illumina TruSeq RNASeq library preparation kit. The resulting libraries were sequenced using a 1x50-bp on Illumina HiSeq 2500. Initial sequence data QC was done using FASTQC. Reads were filtered and trimmed to remove adapter-derived or low quality bases using Trimmomatic and checked again with FASTQC. Illumina reads were aligned to Drosophila reference genome (BDGP R5/dm3) with TopHat and Bowtie2. Read counts were generated for each annotated gene using HTSeq-Count. RPKM (Reads Per Kilobase of exon per Megabase of library size) values were calculated using Cufflinks. Read normalization, variance estimation and pair-wise differential expression analysis with multiple testing correction was conducted using the R Bioconductor DESeq2 package. Heatmaps and hierarchical clustering were generated with R Bioconductor. The RNAseq data were deposited on GEO under the accession number GSE93823.
|
10.1371/journal.pcbi.1000295 | Rapid Sampling of Molecular Motions with Prior Information
Constraints | Proteins are active, flexible machines that perform a range of different
functions. Innovative experimental approaches may now provide limited partial
information about conformational changes along motion pathways of proteins.
There is therefore a need for computational approaches that can efficiently
incorporate prior information into motion prediction schemes. In this paper, we
present PathRover, a general setup designed for the integration
of prior information into the motion planning algorithm of rapidly exploring
random trees (RRT). Each suggested motion pathway comprises a sequence of
low-energy clash-free conformations that satisfy an arbitrary number of prior
information constraints. These constraints can be derived from experimental data
or from expert intuition about the motion. The incorporation of prior
information is very straightforward and significantly narrows down the vast
search in the typically high-dimensional conformational space, leading to
dramatic reduction in running time. To allow the use of state-of-the-art energy
functions and conformational sampling, we have integrated this framework into
Rosetta, an accurate protocol for diverse types of structural modeling. The
suggested framework can serve as an effective complementary tool for molecular
dynamics, Normal Mode Analysis, and other prevalent techniques for predicting
motion in proteins. We applied our framework to three different model systems.
We show that a limited set of experimentally motivated constraints may
effectively bias the simulations toward diverse predicates in an outright
fashion, from distance constraints to enforcement of loop closure. In
particular, our analysis sheds light on mechanisms of protein domain swapping
and on the role of different residues in the motion.
| Incorporating external knowledge into computational frameworks is a challenge of
prime importance in many fields of biological research. In this study, we show
how computational power can be harnessed to make use of limited external
information and to more effectively simulate the molecular motion of proteins.
While experimentally solved protein structures restrict our knowledge to static
molecular “snapshots”, a vast number of proteins are
flexible entities that constantly change shape. Protein motion is therefore
intrinsically related to protein function. State-of-the-art experimental
approaches are still limited in the information that they provide about protein
motion. Therefore, we suggest here a very general computational framework that
can take into account diverse external constraints and include experimental
information or expert intuition. We explore in detail several biological systems
of prime interest, including domain swapping and substrate binding, and show how
limited partial information enhances the accuracy of predictions. Suggested
motion pathways form detailed lab-testable hypotheses and can be of great
interest to both experimentalists and theoreticians.
| Mechanistic understanding of protein motions intrigued structural biologists,
bio-informaticians and physicists to explore molecular motions for the last five
decades. In two seminal breakthroughs in 1960 [1],[2], the structures of
Haemoglobin and Myoglobin were solved and consequently, for the first time,
mechanistic structural insights into the motion of a protein were deduced from its
snap-shot image. This finding paved the way to a by-now classical model for
cooperativity in binding of allosteric proteins [3]. Nowadays, hundreds of
proteins with known multiple conformations, together with their suggested molecular
motion, are recorded in databases such as MolMovDB [4]. This number increases
with the influx of solved structures from the Protein Data Bank [5]. An
inherent flexibility is characteristic of fundamental protein functions such as
catalysis, signal transduction and allosteric regulation. Elucidating motion of
protein structures is essential for understanding their function, and in particular,
for understanding control mechanisms that prevent or allow protein motions.
Understanding the relation between protein sequence and protein motion can allow
de-novo design of dynamic proteins, enhance our knowledge about
transition states and provide putative conformations for targeting drugs. Accurate
prediction of protein motion can also help address other computational challenges.
For instance, Normal Mode Analysis (NMA) motion predictions [6] can be used for efficient
introduction of localized flexibility into docking procedures [7],[8].
Experimental knowledge of macro-molecular motions has been discouragingly limited
to this day by the fact that high-resolution structures solved by X-ray
crystallography are merely the outmost stable conformations of proteins, in a
sense a snap shot of a dynamic entity. While high resolution experimental data
of molecular motion are still beyond reach, innovative breakthroughs in
time-resolved optical spectroscopy, single molecule Förster resonance
energy transfer (FRET), small-angle X-ray scattering (SAXS) [9],
as well as advances in NMR spectroscopy such as residual dipolar coupling
methods and paramagnetic relaxation enhancements [10]–[13] now provide
increasingly detailed experimental data on molecular motion, e.g., distance and
angle constraints or measurements of rotational motion [14].
In spite, and perhaps due to the limited amount of experimental information,
computational techniques like molecular dynamics (MD) simulations [15],[16] have been used
extensively for the last three decades to simulate macro-molecular motion.
Unfortunately, standard MD simulations are computationally intensive, and
moreover, they often remain trapped in repetitive cycles of Brownian motion
throughout the simulation, without being able to cross significant energy
barriers. Therefore, they are often limited to pico-to-nano second timescales of
motion [17], whereas events like enzyme catalysis [18], protein folding [19] and protein
recognition [20] may require more time. As researchers often
possess intuition and explicit partial knowledge about the nature of a motion or
target conformations, biasing techniques were devised in steered MD simulations
[21]. Such methods incorporate prior knowledge or
expert intuition about the system and compromise the intended purity of MD
simulations as a physical simulation. Nonetheless, they still rely to the most
part on an approximation of physical forces, and guarantee that some plausible
assumptions are satisfied. Subsequent motion trajectories were shown useful for
designing experiments and deriving mechanistic insights into protein motion.
Complementary coarse-grained methods such as Normal-Mode Analysis and
Gö models [6],[22],[23] (reviewed in [10]) provide quick impressions about protein
conformational changes when given a native conformation, but do not aim at the
very fine details of the motion.
In recent years, a novel approach for sampling motion pathways, rooted in
algorithmic robotics motion planning, has been applied to large-scale molecular
motion prediction. This approach suggests an efficient alternative to slow
step-by-step simulations of Newton equations. Instead, a sequence of clash-free
conformations is generated by sampling the topology of the conformational space.
This sequence is a fine discretization of continuous motion. In their original
context, motion planning techniques like probabilistic road-maps (PRM) [24],
Rapidly-exploring Random Trees (RRT) [25],[26]
and similar methods [27]–[29] (all reviewed in [26],[30]) have been used to
plan the motion of objects with many degrees of freedom (dofs)
among obstacles in a constrained environment [31]. (Usually, these
objects are referred to as “robots”, but can be any moving
object, such as digital avatars, manufactured parts, or molecules in the context
of this study). For simplicity, we collectively refer to this family of
techniques as motion planning sampling techniques.
In molecular biology, motion planning techniques were used to predict motion
pathways for molecules while considering a large numbers of
dofs [32]–[39], and contributed to
our understanding of molecular kinetics in applications such as energy landscape
exploration, protein and nucleic acids folding pathways and ligand binding [32],
[34]–[36],[40].
Their performance has been compared to molecular dynamics [36] and integrated with
Normal Mode Analysis [38]. In several cases, they were shown to capture
known conformational intermediates and other experimental indicators [33],
[37]–[39].
Motion planning techniques are optimized for finding complete motion pathways.
They record the history and approximate the topology of the sampled search space
in a tree or a graph data structure, the “road-map”.
Molecular motions are extracted from paths or “roads” in the
graph, where nodes stand for feasible (low-energy) conformations and edges
connect close-by conformations. Therefore, paths in this graph are sequences of
clash-free conformations. This also adds a whole new dimension of memory to the
sampling process and the resulting search in conformation space is shown to be
less prone to futile repetitive sampling [25].
Motion planning techniques are very fast – it takes between minutes and
hours to generate a full motion pathway of relatively large time-scales with
dozens of dofs and hundreds of amino-acids [38],[39], compared to weeks
to months in MD simulations of motions with shorter time-scales. Hence, in
contrast to MD simulations, sampling based methods are fast enough to generate a
very large number of alternative pathways, whereas in an MD simulation it is
often hard to decide if the pathway is representative or just the outcome of
specific random start conditions. As the application of motion planning
techniques to molecular motion is relatively new, further research is required
in order to validate and calibrate its use. The external incorporation of
experimental measurements into sampling-based simulations can increase the
credibility of predictions, and turn them into a fair complement to
ab-initio simulations.
In addition, as the dimensionality of the search space increases, it is
advantageous to exploit prior information about the nature of the motion to
direct the search. A common practice in sampling methods of single conformations
like MC is to bias the energy function itself towards known constraints [41]. In
the context of sampling-based motion planning, it is common to explicitly bias
the sampling to include the target conformation (e.g. [42],[43]).
Another common bias is towards narrow passages in the space of configurations
[26],[44]. In order to
avoid getting stuck due to over-bias, biased sampling is often restricted to a
fraction of the tree growth iterations. Kalisiak and Panne [45] terminated RRT
branches that lead to immediate collisions, by sensing the local environment
on-the-fly in order to save running time. Zucker et al.
[46]
used various features of the workspace environment (the Cartesian representation
of the world) to bias the sampling of motion planning algorithms, by introducing
ad-hoc relations between robotic dofs and workspace features,
and using a grid discretization of the workspace.
Here, we present PathRover, a comprehensive and generalized framework for
efficiently sampling and generating motion pathways that satisfy constraints of
prior information with the RRT algorithm [25]. PathRover
generates low-energy, clash-free motion pathways that are biased towards
external constraints. This is in analogy to similar approaches for finding a
single optimal structure (but not a motion pathway) under a set of experimental
constraints [47],[48]. Our approach
follows the notion that the combination of a number of partial constraints can
significantly limit the number of feasible solutions. We rely on a generalized
RRT formalism that allows efficient, flexible and straightforward integration of
prior information into the basic RRT algorithm. Partial information is
incorporated through a branch-termination scheme where the
growth of undesired pathways from the RRT tree is terminated (see Figure 1 for a toy example
that illustrates the effect of constraints on RRT motion sampling). To our
knowledge, this is the first thorough generalized attempt to incorporate diverse
types of prior biological information into the RRT algorithm in biological
context.
We examine how limited geometric constraints can guide different types of motion
towards a correct conformation. We deal with 8 to 198 backbone torsions, and
model flexibility for all side-chain rotamers. We are motivated by the progress
in experimental methods for extracting transient and non-transient distance
constraints [10], e.g. using “experimental
rulers” such as FRET and site-directed spin labeling experiments, or
dynamic experimental measurements of the relative orientation of secondary
structures [14] (Table 1).
PathRover is integrated into the Rosetta molecular modeling framework [49], an
accurate protocol for a range of different structural modeling tasks (e.g.,
[50]–[53]). Thus, PathRover
is equipped with state-of-the-art energy functions, sampling and optimization
protocols. All generated motion pathways are guaranteed to form a sequence of
clash-free low-energy conformations, and to satisfy the input constraints.
We mainly focus on domain swapping of two molecular model systems, the
CesT and the Cyanovirin-N proteins. Domain
swapping occurs in multi-domain proteins, when a domain from one chain packs
against the complementary domain in an identical chain [54], forming a
pseudo-monomer (Figure 2 and Figure S1).
The pseudo-monomer resembles the native structure, and the
interface between the swapped domains is native-like. Domain swapping can lead
to undesired effects of aggregation, such as the formation of amyloidal fibrils
[55]. Investigation of domain unpacking and repacking
may improve our understanding of the general mechanism of oligomerization [56].
Domain swapping is an interesting target for motion simulations [57]–[60]. It requires
the unpacking of domains in the original chain, and the subsequent repacking to
another chain. The main structural changes between swapped conformations are
usually restricted to a few hinge residues that connect the two domains [54]. This
may allow for some simplifying assumptions about the degrees of freedom that are
involved in the motion. Since the structure of each domain is identical in
different conformations, we may assume they remain rigid during the motion. We
examine the validity of this simplifying assumption and experiment with various
choices of dofs. A clever choice of dofs may
reduce the running time, but may introduce additional bias to the motion. We
compare restricted runs where only a subset of torsion angles is allowed to
rotate, to free runs where all degrees of freedom are mobile.
We note that the implications of the domain swapping examples are far reaching,
since a large set of conformational motions is presumed to involve hinge motions
with similar characteristics [61]. As an example, we consider the substrate
binding motion of the Ribose Binding Protein.
PathRover supports full-atom simulations in which the output conformations
contain the coordinates of all side-chains and hydrogen atoms. These
conformations can be used to formulate precise, lab-testable hypotheses (e.g.,
suggest mutations that may interfere with the motion), which are of substantial
interest to both experimentalists and theoreticians. In the following sections,
we provide a detailed analysis of these models.
The conformation space is described in terms of internal coordinates. Backbone
torsion angles uniquely define the conformation of a protein, since the
side-chain torsion angles are optimized on the fly for each given backbone
conformation. Bond lengths and angles are fixed, assuming that changes in
torsion angles can in general compensate milder changes observed in bond angles
and length.
We are interested in finding a collision-free, low-energy motion pathway that
starts from a given initial conformation, and is consistent with partial
information about the motion or the target conformation. We have formulated
diverse types of predicates to constrain the sampling of motion pathways
according to prior information. Here, we focus on partial information motivated
by experiments, comparative analysis and expert intuition. For example,
comparative analysis of biological databases can provide partial information
from homologue structures, or from alternative conformations of the native
protein, and distance constrains can be extracted from time-resolved
spectroscopy. Table 1
includes a list of examples for predicates that are motivated by existing or yet
to improve experimental methods for assessing transient conformations.
Importantly, different types of partial information can be combined into a joint
predicate. We note that distance constraints and additional constraints have
been previously used in Rosetta to direct Monte-Carlo with Minimization
sampling, although in a different algorithmic and biological context (see Discussion).
As the combinatorial search space grows exponentially with the number of
dofs, it is also beneficial to restrict the choice of
flexible torsion angles. An automated, accurate choice of mobile
dofs is a challenging aspect of motion prediction, and in this
step, prior information can be most useful (see [38] for an attempt
in this direction). In this work, we have combined information from several
sources for restricting the number of dofs, such as: (1)
careful inspection of structures, (2) relevant literature, (3) computational
tools for detecting hinge regions like Normal Mode Analysis (NMA) [22], and
(4) comparison of structural changes in alternative (native or homologue)
conformations. When both a source conformation and a target conformation are
available, we used the FlexProt flexible alignment tool [62] to extract fixed
regions of the protein, and defined the dofs by the regions
in-between. These were used to manually restrict the allowed
dofs in the examined model systems (Table 2). The effect of the choice of mobile
degrees of freedom is examined in detail in the Results section.
The Rapidly-exploring Random Tree (RRT) algorithm is a general framework for
rapid exploration of a conformation space (referred to as
“configuration space” in robotics) in a highly constrained
environment. It was first presented in algorithmic robotics, where it was used
to plan the motion of moving objects among obstacles [25]. RRT produces a
tree of conformations and records the topology of the search space. Nodes stand
for feasible (low-energy) conformations, edges connect close-by conformations,
and paths are sequences of feasible conformations. It was shown that the RRT
tends to grow towards unexplored regions at progressively increasing resolution
[25].
In order to characterize the predicted motion in our simulations, we have
examined what portions of proteins remained rigid during the motion and what
residues served as hinge residues. We note that inspection of
φ/ψ values is not necessarily suitable for this purpose,
since small backbone perturbations can result in large scale motions and
vice versa. In Figure S2, we describe our protocol for
detecting hinge residues in simulated motion. In brief, we rely on structural
comparison of different pairs of conformation in the simulated motion. Rigid
portions of the protein are detected by the FlexProt [62] flexible
structural alignment algorithm (Figure S2), and the hinge residues are
defined as the regions that connect the rigid parts. We score each residue for
how often it serves as a hinge in different alignments throughout the simulation
(Figure
S2). The structural alignment is performed at different resolutions of
RMSD, using a resolution parameter ρ. Low resolution hinges are involved
in strong hinge motions, and high-resolution hinges are involved in milder ones.
In the first part of this section, we examine the usage of various geometric
constraints and a combination of constraints to bias the motion during simulations.
We also show how the energy function prevents over-bias by the input constraints. In
the second part, we deal with another form of partial information – the
choice of degrees of freedom that are allowed to move during the simulations. We
examine the robustness of simulations to different choices of degrees of freedom,
and analyze in full-atom detail the domain-swapping motion of inspected model
systems.
We now examine in detail the importance of different degrees of freedom for
another model system of domain swapping motion: Cyanovirin-N is
an anti-viral fusion inhibitor protein that binds to viral sugars, and is
trialed for preventing sexual transmission of HIV. It comprises two repeat
domains of 30% sequence identity. The domain swapped dimer has higher
anti-viral affinity than the monomer [72], and it was shown
that the two forms can exist in solution, with a high energy transition barrier
between them. In addition, it has been reported that certain mutations can
affect the energy barrier and stabilize alternative conformations [73]. We examined here how two repeat domains of a
single chain can unpack from the tightly-intertwined monomeric conformation to
an extended domain-swapped conformation. The conformational transition during
swapping is substantial, as the swapped conformations deviate by 14 Å
RMSD.
In all simulations, we started from the monomer conformation (pdb-id 2ezm
[74]),
and for biasing the motion towards the swapped conformation (pdb-id 1l5e
[73]), we minimized the RMSD distance towards it. The
difference between the following simulations is the sets of degrees of freedom
that are allowed to rotate during the motion (Table 2).
In this work we showed how different types of partial information can be incorporated
into the Rapidly-exploring Random Tree (RRT) algorithm. We present PathRover as a
comprehensive framework, implemented within the Rosetta modeling infrastructure. In
structural biology, partial information constraints are widely used in predictions
of static minimal-energy conformations [47],[48],[76] and
in MD simulations. The novelty in this work is the systematic introduction and the
integration of partial information to sampling-based motion planning of molecules.
In this sense, sampling based methods like RRT pose a natural framework for
integrating prior biological information. From the perspective of algorithmic
robotics, partial information is employed through a branch-termination scheme which
is somewhat different from explicitly biasing the sampling of new conformations,
used in previous works [25],[42],[43],[46]. This allows for the use of very general
features, whereas biased sampling may require ad-hoc computations
of a biased distribution functions that differ between various types of information.
We incorporated partial information into simulations of three different systems:
CesT type III secretion chaperone, Ribose Binding
Protein (RBP), and Cyanovirin-N anti-viral protein.
Our analysis demonstrates how partial information constraints limit the search in
the vast space of possible motion pathways. These constraints are motivated by
existing and novel experimental methods for measuring constraints over transient
conformations, or by expert intuition. In turn, computational observations allow for
further subsequent validation by introducing detailed predictions of the motion that
can be validated by experimental methods. We showed that the energy function
prevents an over-bias by the partial information constraints, in case our prior
information is inexact. PathRover simulations allowed us to assess the contribution
of different residues to motion. Apparently, modest motions in specific regions may
facilitate large-scale motions. The results from different simulations produced
consistent patterns, and may therefore justify partial restriction of motion to
improve running times. In particular, restricted and free simulations resulted in
similar patterns of motion.
An important aspect of PathRover is its full embedding into the Rosetta modeling
framework. Rosetta has repeatedly demonstrated an exceptional ability to produce
high-quality results for a variety of different modeling tasks in the field of
protein modeling, docking, protein design and other modeling challenges at
atomic-level detail (e.g., [50]–[53]). The incorporation
into Rosetta provides well-calibrated energy functions (both for centroid and
full-atom simulations), efficient energy calculations, and a battery of established
conformational sampling protocols. It also allows extension to additional predicates
of partial information that were previously implemented in Rosetta, such as NMR
coupling measurements and docking interface constraints. These have been used to
guide and filter Rosetta Monte-Carlo searches, and will here be incorporated into
RRT-based motion prediction.
Previous applications of the RRT algorithm have mainly been based on geometric
considerations of clash avoidance or Van der Waals terms of established force
fields. In some cases, more sophisticated terms were employed [32],[34]. Here we introduce
the established Rosetta full-atom energy function into sampling based methods.
Hence, we are able to generate motion pathways for complex movements that are at
the same time energetically favorable and that abide by possibly known
constraints about the motion. The full-atom energy function of Rosetta (we used
here score12 [50]) includes physical terms such as van der
Waals potential and solvation terms, as well as statistical knowledge based
terms like the Ramachandran score, rotamer likelihood, statistical hydrogen
bonding term and a simplified electrostatic score [49]. In some cases we
observed that the repulsive energy term dominates the motion pathway: in a
cluttered environment, clash avoidance is indeed probably the main contribution.
Naturally, however, additional energy terms will affect the details of the
motion pathways, such as solvation effects and electrostatic interactions [77]. We
note that the statistical terms in Rosetta have straightforward interpretation
in terms of physical properties. For instance, the Ramachandran score and
rotamer likelihoods reflect steric hindrance in disallowed regions. The hydrogen
bonding term was also shown to correlate remarkably with quantum-mechanical
calculation [64]. While the original Rosetta energy function
was optimized for native conformations, we postulate that it can also be used
for the generation of clash-free, reasonable motion paths, which also account
for other physical principles. Comparison to other common force-fields like
CHARMM [78] and Amber [79] will provide
additional credibility to PathRover simulations. In principle, PathRover is not
restricted to any energy scoring function, as the energy scoring is a
“black-box” in the implementation of the algorithm. As
molecular mechanics energy functions are currently being added to the Rosetta
modeling framework, we intend to compare different energy scoring functions in
future work.
This study proposes PathRover as a general and flexible setup where molecular
systems can be explored, and constraints can be incorporated in a general and
straightforward manner. Partial information can improve the performance of
sampling based algorithms, by narrowing down the search in the vast
conformational space of proteins. This is demonstrated in the present study on a
number of molecular motions of specific interest. Future work will concentrate
on refining protocols for additional systems and types of motions.
Beneficial crosstalk between experimental procedures and in
silico simulations will ultimately optimize the wide integration of
partial information into fast sampling-based algorithms–and forward
our general understanding of protein motion and function.
|
10.1371/journal.pgen.1001325 | The SUVR4 Histone Lysine Methyltransferase Binds Ubiquitin and Converts H3K9me1 to H3K9me3 on Transposon Chromatin in Arabidopsis | Chromatin structure and gene expression are regulated by posttranslational modifications (PTMs) on the N-terminal tails of histones. Mono-, di-, or trimethylation of lysine residues by histone lysine methyltransferases (HKMTases) can have activating or repressive functions depending on the position and context of the modified lysine. In Arabidopsis, trimethylation of lysine 9 on histone H3 (H3K9me3) is mainly associated with euchromatin and transcribed genes, although low levels of this mark are also detected at transposons and repeat sequences. Besides the evolutionarily conserved SET domain which is responsible for enzyme activity, most HKMTases also contain additional domains which enable them to respond to other PTMs or cellular signals. Here we show that the N-terminal WIYLD domain of the Arabidopsis SUVR4 HKMTase binds ubiquitin and that the SUVR4 product specificity shifts from di- to trimethylation in the presence of free ubiquitin, enabling conversion of H3K9me1 to H3K9me3 in vitro. Chromatin immunoprecipitation and immunocytological analysis showed that SUVR4 in vivo specifically converts H3K9me1 to H3K9me3 at transposons and pseudogenes and has a locus-specific repressive effect on the expression of such elements. Bisulfite sequencing indicates that this repression involves both DNA methylation–dependent and –independent mechanisms. Transcribed genes with high endogenous levels of H3K4me3, H3K9me3, and H2Bub1, but low H3K9me1, are generally unaffected by SUVR4 activity. Our results imply that SUVR4 is involved in the epigenetic defense mechanism by trimethylating H3K9 to suppress potentially harmful transposon activity.
| The characteristics of the diverse cell types in multicellular organisms result from differential gene expression that is dependent on the level of DNA packaging. Genes that are essential for the function of the cell are expressed; while unessential genes, and DNA elements (transposons or “jumping genes”) that can move from one position to another within a genome and potentially cause deleterious mutations, are repressed. The mechanisms evolved in eukaryotes to avoid unwanted gene expression and transposon movement include DNA methylation and specific combinations of post translational modifications (PTMs) of the histones that package DNA. Here we show that the SUVR4 enzyme binds the signaling protein ubiquitin and that ubiquitin enables the enzyme to trimethylate lysine 9 (H3K9me3) of histone H3. In contrast to other reports demonstrating an activating role on expressed genes, we show that H3K9me3 has a locus-specific repressive effect on the expression of transposons. The specificity is maintained by the communication with other PTMs on transposons and euchromatic genes, which has a stimulating or repressing effect on enzyme activity, respectively. Our results demonstrate how repression of transcription can be restricted to specific targets and demonstrate that this repression involves a context-dependent read-out of different PTMs.
| In eukaryotes, gene expression and chromatin structure is specified by the combinatorial pattern of posttranslational modifications (PTMs) on the histone tails, which include phosphorylation, acetylation, methylation, SUMOylation and ubiquitination [1], [2]. These PTMs are interdependent, thus providing regulatory cross-talk, and established at the histone tails in a coordinated manner by different classes of highly specific chromatin modifying enzymes.
The combination of PTMs constitutes the so-called histone code, and their downstream effect on chromatin organization and gene expression is mediated by nonhistone effector proteins that contain domains that bind or “read” this code in order to specify epigenetic function. Such domains show specificity for particular modified residues (e.g. acetylation or methylation of lysine) in the context of its surrounding amino acid sequence, and for the state of the modification (e.g. H3K9me1 vs H3K9me3) [1], [3]. For example, domains belonging to the Royal Superfamily, including the chromodomain, Tudor domain and MBT domain and members of the PHD finger family, bind methylated lysine residues on the histone tails [4]. More specifically, the PHD finger of the ORC1 protein in Arabidopsis binds H3K4me3, but not H3K4me1 or H3K4me2 at target genes, and this mediates H4K20 trimethylation and activates transcription [5].
Lysine ubiquitination of histones and other target proteins is a three step process involving Ub (ubiquitin)-activating (E1), Ub-conjugating (E2) and Ub-ligating (E3) enzymes, eventually leading to monoubiquitination, multi-monoubiquitination or polyubiquitination [6], [7]. Ubiquitin binding domains (UBDs) represent a new class of motifs that enable proteins to bind non-covalently to the PTM ubiquitin. More than twenty families have been identified to date, and they differ in structure and the type of ubiquitin modification they recognize [6], [7]. Poly-Ub chains linked via the K48 residue of ubiquitin are largely recognized by UBDs of receptors that target proteins for proteosomal degradation, while monoubiquitin is recognized by UBDs of proteins involved in processes like DNA repair, regulation of protein activity, chromatin remodeling and transcription [6]–[8].
The cross-talk between H2B monoubiquitination (H2Bub1) and histone methylation has been extensively studied and is highly conserved from yeast to human. These studies show that monoubiquitination of H2B recruits proteins that direct histone H3K4 di- and trimethylation but not monomethylation by activation of the Set1 histone lysine methyltransferase (HKMTase) of the COMPASS complex (reviewed in [9], [10]). In Arabidopsis, H2B monoubiquitination at K143 coincides with active transcription [11]–[13]. Deubiquitinating enzymes (DUBs) oppose the function of E3 ligases by deubiquitinating Ub-conjugated proteins. Increased H2Bub1 caused by a mutation in the DUB SUP32/UBP26, leads to reduced H3K9me2 and increased H3K4me3 at transposons that correlate with increased transcription [11]. A key function for DUBs is to generate a pool of free ubiquitin monomers from ubiquitin precursors synthesized from Ub-encoding genes, and from polyubiquitin chains and ubiquitin conjugates [14]. Free monomeric ubiquitin is required under stress conditions, and organisms defective in ubiquitin precursor proteins or DUBs are more sensitive to stress. In yeast, heat stress stimulates the production and activation of the Doa4 deubiquitinase which increases the supply of free monomeric ubiquitin by cleaving polyubiquitin [15].
HKMTases contain SET domains with specificities for different lysine residues on the histone tails, and may be involved in either gene activation or gene repression depending on which lysine residue is methylated [16]. In general, methylation of H3K9, H3K27 and H4K20 has been associated with heterochromatin and gene repression, while H3K4, H3K36 and H3K79 methylation has been related to euchromatin and gene activation [1]. The downstream effect of histone methylation also depends on the number of methyl groups at each lysine residue. Histones mono-, di-, or trimethylated at lysines are differently distributed within eu- and heterochromatin, each potentially indexing a specific biological outcome [17], [18]. For example, in Arabidopsis, H3K36 trimethylation, but not H3K36 monomethylation, shows a strong positive correlation with transcription of MADS box genes involved in flowering-time and flower development [19], [20].
Although lysine methylation to a large extent is conserved between eukaryotes, the distribution and biological outcome of the methylation may be different. H3K9me1, H3K9me2 and H3K27me2 are for instance predominantly found in the chromocenters of Arabidopsis but not in mouse chromocenters (reviewed in [21], [22]). Conversely, H3K9me3 and H4K20me3 that localize to heterochromatin in mouse are mainly associated with euchromatin in Arabidopsis. Additionally, recent results suggest that in contrast to other eukaryotes, H3K9me3 methylation correlates with gene transcription and might have a slight activating function in Arabidopsis [23], [24].
H3K9 methylation is carried out by proteins of the SU(VAR)3-9 subgroup which consists of 14 proteins in Arabidopsis; the SU(VAR) 3-9 HOMOLOGs SUVH1-SUVH9, and the more distantly related SU(VAR) 3-9 RELATED proteins SUVR1-5 [25]. In addition to the SET domain the SUVH proteins contain the YDG/SRA domain that has been shown to bind methylated DNA and might direct SUVH mediated H3K9me2 to heterochromatin or stimulate its activity [26]. Thus in Arabidopsis, the SUVH proteins link the epigenetic gene-silencing marks H3K9me2 and DNA-methylation and work as transcriptional repressors of transposons or inverted repeat sequences, for instance by directing CHG methylation via the CMT3 DNA methyltransferase (reviewed in [27]). In contrast to the SUVH proteins, the SUVR1, SUVR2 and SUVR4 proteins do not contain an YDG/SRA domain, but an N-terminal WIYLD domain of unknown function [28], suggesting another mode of action for these proteins. SUVR proteins associate with the nucleolus or euchromatin, and we have earlier shown that SUVR4 can dimethylate H3K9 when this position is monomethylated [28].
In the present study we show that the WIYLD domain of SUVR4 specifically binds ubiquitin, demonstrating a close connection between ubiquitin binding and histone H3K9 methylation. We have furthermore revealed that ubiquitin stimulates the enzyme activity of SUVR4 and converts SUVR4 from a strict dimethylase to a di/trimethylase in vitro. Chromatin Immunoprecipitation (ChIP) analysis of Arabidopsis lines with reduced or enhanced expression of SUVR4, demonstrate that SUVR4 localizes to both euchromatin and heterochromatin in vivo, but only converts H3K9me1 to H3K9me3 at transposons and pseudogenes. SUVR4 dependent H3K9 trimethylation correlates with locus specific transcriptional repression of transposable elements intercalated within euchromatin of the Arabidopsis genome.
To address the function of the SUVR4 WIYLD domain, a construct encompassing only this domain (Figure 1A) was used in a yeast two-hybrid screen to identify interacting proteins. One positive clone identified in this screen, contained the full-length coding sequence (CDS) of UBIQUITIN EXTENSION PROTEIN 1 (UBQ1, AT3G52590) (Figure 1B). The UBQ1 protein consists of an N-terminal ubiquitin moiety and the C-terminal ribosomal protein L40 [29]. These moieties were subcloned and tested separately for their interaction with SUVR4-WIYLD. Clones containing the ubiquitin moiety, but not clones containing the L40 moiety, supported growth on selective media when transformed into yeast cells and mated with cells containing SUVR4-WIYLD, suggesting that SUVR4 specifically interacts with ubiquitin (Figure 1B). This was confirmed in an in vitro pull-down experiment, where SUVR4-WIYLD pulled down full-length UBQ1 and ubiquitin but not L40 (Figure 1C).
To address whether the WIYLD domain binds ubiquitin in its unconjugated form and to identify residues directly involved in the interaction between WIYLD and ubiquitin, an NMR analysis was performed. The [1H,15N]-HSQC spectrum of 15N-isotopically labeled SUVR4-WIYLD is well-dispersed demonstrating that the protein domain is folded (Figure 1D). Upon titration of ubiquitin, chemical shift perturbations were observed for a number of residues including the six consecutive amino acids Y69TALVD74 of helix 3 (Figure 1D), indicating that they are involved in binding. Alignment of SUVR4-WIYLD with WIYLD domains in other proteins have earlier shown that many of these residues are highly conserved (Figure 1A and [28]).
SUVR4 binds and efficiently methylates calf thymus histone H3 as well as H3K9me1 peptides in vitro, but shows only weak activity against recombinant histones, arguing that SUVR4 cross-talks to premodified histones [28]. Since the WIYLD domain binds ubiquitin, and SUVR4 binds and methylates histones, we tested whether the WIYLD domain binds H2B monoubiquitinated on lysine 143 (H2Bub1), which is the only ubiquitination on core histones reported so far in Arabidopsis [11], [30]. In these experiments the WIYLD domain indeed was able to pull down H2Bub1, however, when R37 and D74 were mutated, the interaction was strongly reduced (Figure 1E). This supports the chemical shift perturbations shown by the NMR analysis, arguing that these residues are directly involved in ubiquitin binding. Interestingly, the invariant W61 residue that showed no shift in the NMR analysis, only weakly affected the WIYLD-ubiquitin interaction when mutated, confirming that this position is not crucial for ubiquitin binding.
As the WIYLD domain was able to bind ubiquitin (Figure 1D), we asked whether ubiquitin could stimulate SUVR4 enzyme activity, as previously shown for the deubiquitinase USP5 [31]. To this end, we compared the activity of a SUVR4 protein without the WIYLD domain to a full-length SUVR4 protein, both in fusion with the Maltose Binding Protein (MBP-SACSET and MBP-SUVR4, Figure 1A), with and without the addition of ubiquitin. In both cases the full-length protein showed higher enzymatic activity than the truncated SACSET fragment (Figure 2A, B), suggesting that the WIYLD domain has a positive effect on the catalytic activity of SUVR4 although the domain itself does not contain HMTase activity (Figure S1C). The difference in activity was more pronounced when ubiquitin was added to the reaction. With ubiquitin the full-length protein was stimulated 2-3 fold whereas the SACSET construct was only weakly affected, suggesting that most of the ubiquitin response is mediated through the WIYLD domain (Figure 2A, B). Addition of free ubiquitin only stimulates enzymatic activity of the SUVR4 protein on histone H3 but does not affect its specificity as no other core histones becomes methylated (Figure 2C).
Using H3K9me1 and H3K9me2 peptides we tested whether the increased SUVR4 enzyme activity after the addition of ubiquitin also affected the product specificity. As expected from previous results [28], H3K9me1 peptides were the preferred substrate as unmethylated peptides were only weakly methylated (Figure S1A), and no activity against H3K9me2 peptides was observed in the absence of ubiquitin. Methylation of H3K9me1 modified peptides was increased 2.5–3 fold when ubiquitin was added to the reaction (Figure 2D). Unexpectedly we also observed methylation of the H3K9me2 peptide in the presence of ubiquitin, suggesting that ubiquitin converted the SUVR4 protein to a histone H3K9 trimethylase (Figure 2D, Figure S1B). The activity on H3K9me2 peptides was however several folds lower than when H3K9me1 peptides were used. No activity was observed on H3K9me3 peptides either with or without ubiquitin, excluding the possibility that any other lysine of histone H3 1-21 was methylated by SUVR4, underscoring the specificity against H3K9 (Figure 2D).
The products from the enzyme reactions using peptide substrates were analyzed by peptide mass fingerprinting. After 3 hours incubation, the reactions containing SUVR4 only converted 40.9% of the H3K9me1 peptide to H3K9me2, while 0% was converted to H3K9me3 (Figure 2E, upper middle panel). In the reactions containing ubiquitin, 90.2% of the H3K9me1 peptide was converted to H3K9me2 while 3.5% was converted to H3K9me3 (Figure 2E, upper right panel). When H3K9me2 peptides were used as substrate, we did not see any conversion to H3K9me3 above background level in the absence of ubiquitin (3% background H3K9me3, versus 3.5% when SUVR4 was added to the reaction) (Figure 2E, lower middle panel), however when ubiquitin was present together with SUVR4, a 16.4% conversion from H3K9me2 to H3K9me3 was found (Figure 2E, lower right panel). This suggests that ubiquitin stimulates the catalytic activity of SUVR4 and alters the product specificity in that it converts SUVR4 from a strict dimethylase to a di/trimethylase.
As SUVR4 converts H3K9me1 to H3K9me2/3 in vitro, we asked how these modifications were affected by SUVR4 in vivo. Since no SUVR4 T-DNA knock-out insertion lines were available, knock-down RNAi lines for SUVR4 were established. We also generated GFP overexpression (OE) lines where SUVR4-GFP expression was driven by the strong constitutive 35S promoter, giving a uniform SUVR4-distribution in the nucleus in addition to accumulation in the nucleolus or in foci of unknown function (Figure S2). A weaker glucocorticoid-inducible construct has earlier been reported to give an almost exclusive nucleolar localization of SUVR4 [28]. We did not observe any phenotypes under the tested growing conditions for neither the SUVR4-GFP line, nor the SUVR4 RNAi line.
H3K9me1-3 display different nuclear distributions, with high H3K9me1/2 in chromocenters and pericentric heterochromatin, whereas H3K9me3 is distributed more uniformly in the nucleoplasm with highest concentration in euchromatin and at expressed genes [32]. Immunocytological analysis on seedling leaves using specific antibodies against H3K9me showed a strong reduction in H3K9me1 and a corresponding increase in H3K9me3 in nuclei with high SUVR4-GFP expression (Figure 3A). Nuclei from lines with a low SUVR4-GFP expression did not show this effect on H3K9me1 and H3K9me3 methylation, suggesting that the global changes in H3K9me1 and H3K9me3 correlated with SUVR4-GFP expression (Figure 3A).
To analyze this effect at individual genes, ChIP experiments were performed with the same antibodies as used for immunocytological analysis and an antibody specific for GFP, respectively. Different classes of transposon sequences were selected for ChIP analysis, as these sequences are likely targets of SUVR4 because of their high H3K9me1 level (Figure 3B and Table 1). These experiments confirmed that SUVR4 is associated with transposons and genes both in eu- and heterochromatin, but a significantly higher amount of SUVR4-GFP is found at euchromatic genes like TUB8 and ACTIN2 (Figure S3). However, only transposon and pseudogenes like AtSN1, AtGP1, AtMU1, AtCOPIA4 and MULE At2g15810 were affected by overexpression of SUVR4, resulting in a drastic increase in H3K9me3 and reduction of H3K9me1 (Figure 3B). We did not see any effect of SUVR4 OE for highly expressed genes like TUB8 or ACTIN2, or for the moderately expressed transposon At4g13120, all with an already low level of H3K9me1. Although having a dramatic effect on H3K9me3 at transposons, SUVR4 OE did not affect the distribution of the euchromatic mark H2Bub1 at any of the tested sequences (Figure S4A).
As the 35S driven SUVR4-GFP construct could lead to unspecific downstream effects due to ectopic and elevated SUVR4 expression, we complemented the OE data with ChIP analysis of two of the transposons in knock-down SUVR4 RNAi plants. The RNAi lines showed a 90% reduction of the SUVR4 expression level compared to wild type (Figure S5 A). In contrast to the OE line, there was an increase of H3K9me1 on AtSN1 and MULE At2g15810 (Figure 3C). Furthermore, there was a corresponding reduction of H3K9me3, suggesting that SUVR4 directs H3K9me3 methylation on transposons. The weak reduction of H3K9me3 could reflect the residual SUVR4 expression in the RNAi line and possibly redundancy with other H3K9me3 methyltransferases at these sequences. Together, these data suggest that although SUVR4 is localized in both eu- and heterochromatin, it is active only on target sequences with a high level of H3K9me1, where its activity increases H3K9me3 at the expense of the H3K9me1 level.
Recent studies suggest that in Arabidopsis H3K9me3 associates with euchromatin and transcriptional activation of genes [23], [24], [32]. In contrast, H3K9me1 is a mark mainly associated with repetitive sequences in chromocenters and pericentric heterochromatin in Arabidopsis [21]. The specific activity of SUVR4 on transposon chromatin although associated with both transposons and euchromatic genes (Figure 3, S3), made us speculate that the lack of SUVR4 activity on euchromatic genes was due to cross-talk to PTMs characteristic for euchromatin. We thus tested histone tail peptides that were mono- or trimethylated at H3K4 but devoid of H3K9me in an in vitro HKMTase assay (Figure 4). SUVR4 activity was not affected by monomethyl H3K4, whereas trimethyl H3K4 reduced SUVR4 activity significantly (Figure 4 A, B), arguing that chromatin associated with genes like TUB8 and ACTIN2, with a high level of this mark, might not be good substrate for SUVR4 activity.
To evaluate the effect of SUVR4 mediated H3K9me3 methylation on transposon transcription we investigated the expression of three of the ChIP-analyzed transposons, MULE At2g15810, AtIS112A (At4g04293) and AtCOPIA4, which all had a high level of H3K9me1 and were expressed in wild type plants (Figure 3B, C, Figure S5 B and Table 1). In the OE line, all the studied transposons showed significant reduction in expression compared to wild type (60%, 80% and 35%, respectively, Figure 5A), suggesting that SUVR4 acts as a repressor of these transposable elements. As a control, we used the At4g13120 transposable element of intermediate expression with a very low H3K9me1 level which is not a target of SUVR4 methylation (Figure 3, Figure 5A and Table 1). This transposon was also unaffected in its transcription level in SUVR4-GFP overexpression lines.
In the RNAi line we did not see a corresponding release of repression for the AtCOPIA4 and AtIS112A elements, however, the MULE At2g15810 element was induced 2.5 to 3- fold in the RNAi line compared to wild type (Figure 5A). Interestingly, the gene Cyp40 which is known to be regulated by MULE [33] showed the same expression response to SUVR4 as MULE At2g15810, although weaker (Figure 5A). The AtSN1 repeat interspersed within euchromatin, and the heterochromatin localized AtMU1 that are silent in wild type plants (Table 1 and Figure S5 B), were examined in both the RNAi and OE line but we did not detect any signal above the –RT control reaction, arguing that these transposons were not reactivated in any of the lines (data not shown).
H3K9me2 directed by SUVH proteins regulates non-CG methylation in Arabidopsis [34]. To determine if there was a similar correlation between DNA methylation and the H3K9me3 methylation directed by SUVR4, bisulfite sequencing was performed on two of the transposons that are targets of SUVR4 histone lysine methylation. We did not detect an effect of SUVR4 activity on DNA methylation of the MULE At2g15810 transposon for CG, CHG or CHH in neither SUVR4 OE nor SUVR4 RNAi lines (Figure 5B). This suggests that the repressive effect of H3K9me3 added by SUVR4 is not mediated by DNA methylation. In contrast, the AtSN1 transposon showed an increase in CHH methylation (Figure 5C) in the OE line. The CG and CHG methylation levels were unaffected. There was, however, no corresponding reduction of CHH methylation in the RNAi-line.
The ubiquitin binding properties of the SUVR4 WIYLD domain and the ubiquitin-enhanced H3K9me3 activity of SUVR4 in vitro led us to look for links between ubiquitin and H3K9 trimethylation in vivo. Interestingly, deubiquitination of H2BUb1 by the nuclear UBP26/SUP32 ubiquitin protease, is required for repression of transposons [11], which also are targets of SUVR4. Therefore we investigated the H3K9me levels in the ubp26-1/sup32 mutant (Figure S6). No effect was seen on highly expressed genes like TUB8 and ACTIN2 (Figure 6), and consistent with earlier findings [11], our ChIP analysis showed a reduction of H3K9me2 on transposons and repeat sequences (Figure 6A). Similarly, H3K9me3 was also reduced on transposons in the mutant compared to the wild type (Figure 6B). Although mutation in the UBP26/SUP32 gene has been reported to lead to a global accumulation of H2Bub1 [35], the H2Bub1 level on transposons was only weakly affected by the mutation (Figure 6C), and the level of free ubiquitin monomers in the nuclei of ubp26-1/sup32 was similar to the level in the wild type (Figure 6D).
We next tested the effect of global reduction of H2Bub1 on H3K9me3 level on transposon chromatin using the hub2-2 mutant. This mutant is defect in the HISTONE MONOUBIQUITINATION2 E3 ligase, which acts non-redundantly with HUB1 to monoubiquitinate histone H2B [13]. The hub2-2 mutant showed an almost complete lack of H2Bub1 at the TUB8 gene, while the effect was absent or negligible on the AtGP1 transposon. As reported for H3K9me2 [13], [36], the H3K9me3 level was not affected either on TUB8 or on transposon chromatin (Figure S7).
H3K9me3 has only recently been confirmed as a histone modification present in Arabidopsis, and its significance in gene regulation has only been indicative [23], [24]. The presented work identifies SUVR4 as the first histone H3K9me3 methyltransferase in Arabidopsis and demonstrates how it cross-talks to ubiquitin and chromatin modifications like H3K9me1 and H3K4me3 to repress transposon transcription.
Our experiments have identified the WIYLD domain of the SUVR4 HKMTase as a new ubiquitin interacting domain, demonstrating a direct link between ubiquitin binding and H3K9 methylation. Ubiquitin is extensively distributed in the eukaryotic proteome, and exists as free ubiquitin monomers, ubiquitin extension proteins, polyubiquitin, or ubiquitin conjugates [14]. The interactions with free ubiquitin, the ubiquitin moiety of the ubiquitin extension protein UBQ1 and the ubiquitin conjugate H2Bub1 (Figure 1), indicate that the SUVR4 WIYLD domain can target ubiquitin either in its free or conjugated form.
The interaction between the WIYLD domain of SUVR4 and ubiquitin is further supported by the WIYLD-dependent positive effect of ubiquitin on enzymatic activity (Figure 2). Free ubiquitin stimulated the HKMTase activity of the full-length SUVR4 protein without compromising the substrate specificity because no histones other than H3 were methylated (Figure 2C). However, the addition of free ubiquitin (Ub) converted the protein from a strict H3K9me2 to a H3K9me2/me3 methyltransferase (Figure 2D, 2E), suggesting that ubiquitin either in its free form or conjugated to other proteins like H2B can act as a signal for H3K9 trimethylation. We only observed 3% conversion of H3K9me1 to H3K9me3 after a 3 hour reaction time in our in vitro HKMTase assay while most of the H3K9me1 was converted to H3K9me2 (Figure 2E). In contrast, a massive shift from H3K9me1 to H3K9me3 was seen in vivo when over-expressing SUVR4 (Figure 3A, 3B). Together this implies the need for another component in addition to ubiquitin for SUVR4 to efficiently convert H3K9me1 to H3K9me3 in vitro, as shown for the murine ESET HKMTase [37]. In recombinant form in vitro ESET only catalyzes mono- and dimethylation of H3K9, but in complex with the transcriptional repressor mAM the enzyme generates H3K9me3.
Interestingly, the truncated SUVR4 SACSET protein showed a lower HKMTase activity compared to the full-length SUVR4 protein on core histones (Figure 2A), arguing that the N-terminal WIYLD domain is essential for normal activity of the C-terminal SET domain. Furthermore, the activity of the SUVR4 SACSET was only weakly enhanced by ubiquitin (Figure 2A, 2B), demonstrating that ubiquitin in its free form stimulates SUVR4 activity mainly through the WIYLD domain. Several enzymes that are involved in Ub pathways have shown to be regulated by ubiquitin. Recently, the activity of the mammalian deubiquitination enzyme ataxin-3 was shown to be enhanced by ubiquitination [38], and binding of free ubiquitin to the N-terminal ZnF-UBP domain of the deubiquitinase USP5 led to a conformational change that stimulated enzyme activity [31].
In Arabidopsis H3K9me3 methylation broadly marks 40% of all genes within euchromatin [39]. In addition a low but detectable level of H3K9me3 methylation is found in regions with silenced transposons and pseudogenes [24] (Figure 3 and Figure 6). Our ChIP results suggest that although associated with both eu- and heterochromatin, SUVR4 has no HKMTase activity on euchromatic genes, but specifically targets transposons and repeat sequences where it converts H3K9me1 to H3K9me3 (Figure 3B, 3C). This is perfectly in line with our in vitro HKMTase results, which show that SUVR4 preferably uses H3K9me1 as substrate (Figure 2D). Together the in vivo and in vitro data indicate that SUVR4 only methylates transposons with a high H3K9me1 level although the protein might also associate with regions with a low level of this modification (Figure S3).
SUVR4 methylates unmethylated H3 poorly, and the level of H3K9me1 decreases in the OE line (Figure S1A and Figure 3B). This suggests that SUVR4 does not itself monomethylate H3K9 in vivo. Both SUVH4 and SUVH6 are efficient monomethyl transferases in vitro [40], which together with SUVH5 control the deposition of the majority of H3K9me1 at transposons and repeat sequences [41]. As SUVR4 targets the same type of sequences, it is likely that SUVR4 uses the monomethylated histone substrates created by the SUVH proteins to trimethylate H3K9. In mammalian cells, the SUV39H1 HKMTase depends on a monomethylase as it preferably converts H3K9me1 of H3.1, but not H3K9me2 of H3.3, to H3K9me3. [42]. Similarly, SUVR4 is stimulated by H3K9me1, but is only active on H3K9me2 if ubiquitin is added to the in vitro reaction.
The SUVH2 HKMTase has a strong impact on centromeric and pericentromeric heterochromatinization and gene silencing and reduces the level of H3K9me3 when overexpressed [32]. In contrast, overexpression of SUVR4 leads to increased H3K9me3 levels, and no changes in heterochromatinization could be observed (Figure 3A). Pericentromeric regions contain high levels of H3K9me1 and H3K9me2 in plants, but also H3S10 phoshporylation during mitosis and meiosis II [22]. The cell cycle dependent H3S10ph modification generated by Aurora kinase 1 inhibits SUVR4 activity in vitro [43]. This and the uninterrupted regions of high levels of H3K9me2 associated with the many transposons and pseudogenes located in pericentromeric and centromeric heterochromatin [44], may contribute to repress SUVR4 activity in these regions in dividing cells. Alternatively, SUVR4 might be able to methylate histones in pericentric heterochromatin before H3S10ph is added as Aurora kinase 1 is active on methylated histones. Although pericentric heterochromatin most likely is not the preferred target of SUVR4 activity because of the high level of uninterrupted H3K9me2 [44], SUVR4 could potentially methylate transposons in these regions under certain conditions when ubiquitin levels are high, as demonstrated by the ability of SUVR4 to methylate H3K9me2 peptides when ubiquitin is added (Figure 2D, 2E, Figure S1B, and Figure 7B).
Mutation in the SUP32/UBP26 deubiquitinating enzyme that removes the ubiquitin conjugate from H2Bub1 has been reported to lead to reduction in H3K9me2 [11]. Using ChIP analysis we found low levels of H2Bub1 at all tested transposons, which were only weakly altered in the ubp26 mutant line (Figure 6C). A reduction of both H3K9me2 and H3K9me3 was, however, observed on the same sequences targeted by SUVR4 (Figure 3B, 3C and Figure 6A, 6B). We therefore suggest that SUVR4 and UBP26 act in the same pathway leading to repression of transposon activity, and speculate that the reduction of H3K9me3 in ubp26-1 mutant background can be due to reduced SUVR4 activity. Thus UBP26 can repress transposon transcription by lowering the H2Bub1 level at these sequences to maintain repressive H3 methylation as suggested by Sridhar et al. [11], and/or by maintaining a high local level of free ubiquitin which stimulates SUVR4-mediated H3K9me3 (Figure 7). Possibly UBP26/SUP32 can also cleave the ubiquitin extension protein UBQ1 initially found in our yeast two-hybrid screen to obtain free ubiquitin, as it has been shown to also be active on the human homologue CEP52 [11] which has 92% sequence identity with UBQ1. We did not however observe any reduction of free ubiquitin in the nuclear extracts of ubp26-1 mutants (Figure 6D) that might have affected SUVR4 activity, and there was no effect on H3K9me3 or H2Bub1 at transposon sequences in the hub2-2 line (Figure S7). Thus, HUB2 seems not to be involved in regulation of H2Bub1 or H3K9me2/3 or to be the counterpart of UBP26 on transposon chromatin. The minor reduction of H2Bub1 at transposons and the ability of UBP26/SUP32 to deubiquitinate the CEP52 in vitro, opens the possibility that UBP26 regulates SUVR4-dependent H3K9me2/3 by additional mechanisms, for instance transient changes in the levels or subnuclear distribution of free ubiquitin.
Highly transcribed euchromatic genes like ACTIN2 and TUB8 were unaffected by SUVR4, and the in vitro assay implies that SUVR4 activity is inhibited by H3K4me3 which is abundant in euchromatin (Figure 4). Furthermore, the in vivo data shows that the targets for SUVR4 activity have low levels of H3K4me3, H3K9me3 and H2Bub1 (Figure 3, Figure 6, and Figure S4). Intercalary heterochromatic sequences located within euchromatin are associated with intermediate amounts of opposing histone marks like H3K4me2 and H3K9me2 [33], [44], but have comparable levels of H3K9me1 as heterochromatin (Figure 3B, 3C). As depicted in the model in Figure 7, this suggests that SUVR4 cross-talks to other PTMs and preferably targets transposons outside pericentric and centromeric heterochromatin, with low H3S10ph, H3K9me2, H3K4me3 and H2Bub1 and high H3K9me1 in order to trimethylate H3K9.
For transposon sequences with a low or intermediate expression level in wild type plants, increase in H3K9me3 levels mediated by SUVR4 is associated with repression of transcription (Figure 3, Figure 5, and Figure 7). In the RNAi line only the MULE At2g15810 transposon, localized in euchromatin outside the typical pericentric heterochromatin or centromeric regions [33], showed relief of repression (Figure 5A), suggesting it to be a normal target of SUVR4 activity. However, AtIS112A, another transposon intercalated in euchromatin with an intermediate expression level, was only affected in the OE line. The heterochromatin localized AtMU1 and the euchromatin localized AtSN1, both silent in wild type plants, were also targets for SUVR4 methylation but showed no reactivation in the RNAi line. This suggests that SUVR4-directed H3K9me3 regulates transposon activity in a locus specific manner, where SUVR4 activity alone is sufficient for repression of MULE At2g15810, while it works redundantly with an unknown HKMTase at other elements like AtIS112A, AtMU1 and AtSN1. A similar regulation can be seen for the SUVH2 and SUVH9 SET domain proteins that act redundantly at some loci but independently at others [45]. Thus different transposons are regulated by different combinations of epigenetic marks (Table 1).
Genes in euchromatin have a much higher level of H3K9me3 than transposons, and in these regions this modification seems to correlate with activation of transcription and the deposition of other activating marks [23], [24]. This argues for a combinatorial readout where the context of other PTMs with which H3K9me3 appears decides the biological outcome (Figure 7). In contrast to genes, transposon and repeat sequences contain a high level of H3K9me1 and low levels of H3K4me3 and H2Bub1 (Figure 3B, 3C, and Figure S4) and in this context H3K9me3 may lead to repression of transcription.
H3K9me1 on transposon chromatin seems to be a prerequisite and the preferred substrate for SUVR4 activity, as the control transposon At4g13120, with very low H3K9me1, was not methylated or affected at the transcriptional level (Figure 5A). Several studies have reported the accumulation of H3K9me1 in heterochromatin (reviewed in [22]) but little is known about the function of this mark. Our data supports a model where H3K9me1 is associated with both pericentric and centromeric heterochromatin and transposons intercalated in euchromatin, but does not act as a repressive signal, but rather a template for other methyltransferases. This is supported by the observation that increased H3K9me1 level correlated with increased transcription in the SUVR4 RNAi line and inversely correlated with increased H3K9me3 and repression of transcription in the SUVR4-GFPOE line (Figure 3A–3C and Figure 5A).
The level of DNA methylation of the MULE At2g15810 transposon did not correlate with SUVR4 expression. At the AtSN1 transposon, however, increased H3K9me3 mediated by SUVR4 overexpression coincided with an increase of CHH while no effect was seen for CG methylation (Figure 5B, 5C). Pericentric H3K9me2 shows a strong correlation with CHG methylation but a weaker correlation with CG and CHH methylation [44], while transposons located outside pericentric or centromeric heterochromatin have shorter patches of H3K9me2 at lower levels. Together with the repressive effect of H3K9me2 on SUVR4 activity this argues that the main DNA methylation regulated by SUVR4 is CHH.
The DRM2 methylase is the main regulator of asymmetric CHH methylation, while CHROMOMETHYLASE3 (CMT3) is the main regulator of CHG methylation in Arabidopsis, but at some loci they work together [46], [47]. At dispersed repeats within euchromatin like AtSN1, DRM1, DRM2 and CMT3 act redundantly to maintain CHH and CHG methylation [48]. At such loci we suggest that the H3K9me3 methylation by SUVR4 might mark the underlying transposon sequence for CHH methylation by DRM2/CMT3 (Figure 7B). Interestingly, many transposon sequences contain both H3K27me3 and H3K9me3, a combination that CMT3 has been shown to bind in vitro (Table 1, [24], [30], [49], [50]). The redundant regulation of AtSN1 by CMT3 and DRM1 might thus explain the lack of reactivation and DNA methylation upon reduction of SUVR4 H3K9me3 methylation in the SUVR4 RNAi line.
Although a target of SUVR4-directed H3K9me3 and repression, the MULE transposon was not affected at the DNA methylation level (Figure 5B). In contrast to AtSN1, this transposon has been shown earlier to be activated only in mom1 mutants, and not in mutants with reduced non-CG methylation and kyp/suvh4 mutants (Table 1). MOM1 is a transcriptional repressor that regulates transcriptional gene silencing of loci outside centromeric and pericentromeric heterochromatin, with only small effects on epigenetic marks [33], [51], [52]. This suggests that non-CG methylation is not involved in silencing of MULE. The similar relief of silencing without any effect on DNA methylation between SUVR4 RNAi and mom1 makes it tempting to speculate that SUVR4 recruits MOM1 to its targets in order to repress transcription at this locus (Figure 7B). The intermediately expressed AtIS112A is repressed in SUVR4 OE lines but did not show any relief of expression in the RNAi line. As for AtSN1, this transposon is regulated by non-CG methylation, but also by MOM1. This argues that SUVR4 mediated repression might act via DNA methylation-independent mechanisms such as for MULE At2g15810, but also by DNA methylation-dependent mechanisms as seen for AtSN1, or possibly both as seen for AtIS112A.
DUBs are important to maintain ubiquitin homeostasis by recycling ubiquitin from free ubiquitin chains, ubiquitin conjugates and ubiquitin fusion proteins [14], [15]. UBP26 regulates H3K9me2 and H3K9me3 methylation as well as non-CG methylation at the same sequences as SUVR4 [11]. We hypothesize that UBP26 acts in concert with SUVR4 to trimethylate transposons with a high level of H3K9me1 and low level of H3K4me3 and H2Bub1 (Figure 7). The H3K9me3 methylation thus directs locus-specific methylation-dependent or -independent repression of transposon activity.
Arabidopsis plants, ecotype Columbia (Col), were grown under long day greenhouse conditions at 18°C. Transgenic Arabidopsis plants were generated by the floral dip method [53] using the Agrobacterium tumefaciens strain C58 pCV2260. Transgenic plants containing the pEG104 [54] or pART27 [55] vectors were selected on MS-2 medium (1x Murashige and Skoog salts, 0.05% 2-N-morpholino/ethanesulfonic acid, 2% sucrose, 0.8% agar) containing 10 µg/ml basta or 50 µg/ml kanamycin, respectively. For ChIP, RT-PCR and cytology experiments, Col wild type plants and non-segregating lines containing the respective T-DNA constructs were grown on MS-2 without antibiotic selection. The ubp26-mutant [11] and the hub2-2 [13] mutant lines have been described earlier.
RNA was isolated from approx. 100 mg of 14 day old seedlings using the Spectrum Plant Total RNA Kit with on-column DNase treatment (Sigma). cDNA synthesis and Real time RT-PCR experiments were performed as described previously [20] using gene specific primers (Table S1), except that 4 µg of total RNA was used to synthesize first strand cDNA with Superscript III Reverse Transcriptase and random primers (Invitrogen).
SUVR4-Full (At3g04380), SUVR4-SACSET, SUVR4-WIYLD, UBQ1, ubiquitin moiety of UBQ1 and L40 moiety of UBQ1 were PCR amplified from cDNA using gene specific attB gateway primers (Table S1) and Pfu DNA polymerase (Fermentas). The attB PCR products were recombined into the pDONR/Zeo vector using the Gateway BP Clonase II Enzyme Mix (Invitrogen) according to the manufacturer's instructions. The resulting pDONR/Zeo entry clones were recombined into destination vectors using the Gateway LR Clonase Enzyme Mix (Invitrogen). All constructs were verified by sequencing. The knock-down SUVR4 RNAi construct was made by cloning a unique fragment from the SUVR4 5′end as an inverted repeat on each side of an intron into the binary vector pART27. Cloning procedures are described in detail (Text S1).
Two-hybrid interactions were screened by mating the yeast strain Y187 carrying the pGBKT7-SUVR4-WIYLD bait construct with the strain AH109 carrying a cDNA library (Matchmaker library construction and screening kit, Clontech) at 30°C ON. The cDNA library was created from Columbia wt 14 day old seedlings and recombined into the pGADT7-Rec vector to create an AD-fusion library. Selective media for the nutritional reporter genes ADE2, HIS3 and MEL1 (QDO) containing 20 mg l-1 X-alpha-Gal, was used to identify positive two-hybrid interactions according to the suppliers suggestions. To confirm interaction with SUVR4-WIYLD, the pGADT7-UBQ1, pGADT7-ubiquitin and pGADT7-L40 were mated separately with the pGBKT7-SUVR4-WIYLD or the empty pGBKT7 vector (BD control). Diploid colonies were selected on SD –L/-T, and then streaked out on SD –L/-T/-H +3 AT medium selective for protein-protein interactions.
pHMGWA-SUVR4-Full and pHMGWA-SUVR4-SACSET constructs were transformed into E. coli BL21-Star DE3 and grown at 150 rpm, 37°C in LB-medium with 1% Glucose and 100 µg/ml ampicillin. At an OD600 0.6–0.8, the cells were induced with 1 mM IPTG overnight at 20°C. The cells were lysed with Express and then resuspended in pre-cooled lysis Buffer: 20 mM Tris–HCl, pH 7.5, 400 mM NaCl, 100 mM KCl, 1 mM EDTA, 1 mM DTT, 0.05% Triton X-100 and Protease inhibitor. After centrifugation (15,000 rpm), the supernatant containing recombinant protein was filtered through 0.45 µm filters and prepared for affinity chromatography.
Recombinant proteins SUVR4-Full and SUVR4-SACSET were purified by Ni-NTA affinity chromatography using HisTrap FF 5 ml (GE Healthcare) column in the ÄKTA purifier. Binding buffer or Buffer A and Elution Buffer or Buffer B in the purification step were as follows, Buffer A: 20mM Tris–HCl, pH 7.5, 500mM NaCl, 1 mM EDTA, 1 mM DTT, 20 mM Imidazole and Buffer B: 20 mM Tris–HCl, pH 7.5, 500 mM NaCl, 1 mM EDTA, 1 mM DTT, 500 mM Imidazole.
HKMTase assays were essentially performed as described in [28]. Twenty µg of MBP-SUVR4 protein was incubated in reaction buffer (50 mM Tris pH 8.5, 20 mM KCl, 20 mM MgCl2, 10 mM β-mercaptoethanol and 250 mM sucrose) with 7.5 µl µCi 14C S-adenosyl methionine (SAM) (Amersham/Perkin Elmer) or 100 µM unlabelled SAM (New England Biolabs) as methyl donor. Twenty µg of core histones from calf thymus (Roche), or 5 µg histone H3 peptides were used as substrate. Reactions were incubated at 30°C for 3 hours, and each experiment was repeated at least 4 times. Core histones from calf thymus (Roche), unmodified histone H3 peptide (#12-403, Millipore), monomethyl-histone H3 (Lys9) peptide (#12-569, Millipore), dimethyl-Histone H3 (Lys9) Peptide (#12-430, Millipore), Trimethyl-Histone H3 (Lys9) Peptide (#12-568, Millipore), Trimethyl-Histone H3 (Lys4) Peptide (#12-564, Millipore), monomethyl histone H3 (Lys 4) peptide (gift from Thomas Jenuwein) and ubiquitin (U6253, Sigma) were used in the assays.
Recombinant proteins were expressed in BL21 cells, lysed in 1 X PBS with 0.1 mg/ml lysozyme, 0.2–1% Triton X-100 and protease inhibitor cocktail (Roche), and immobilized on glutathione sepharose beads (Amersham). 3 µg of GST-S4WIYLD was incubated with MBP protein lysates at 4°C for 2.5 hours or 10 µg of GST-SUVR4-WIYLD with 20 µg of precleared core histones (Roche) at 4°C for 3 hours, following a series of washes. Pull-down reactions were run on SDS-PAGE gels, blotted onto a PVDF membrane (Machery Nagel) and probed with either anti-MBP (1∶10000, New England Biolabs, #E8030S) or anti-H2Bub1 (1∶1000, MediMabs, MM-0029). Detection of primary antibody was performed with peroxidase-conjugated secondary antibody; goat anti-rabbit HRP for pulldown of MBP-proteins (1∶10000, Thermo Scientific, PA1-74361) and anti-mouse HRP for pull-down of core histones (1∶10000, Abcam, ab6728) using the ECL kit (GE HealthCare, RPN2135).
Reverse phase (C18) nano online liquid chromatographic MS/MS analyses of proteolytic peptides from HKMTase reactions using unlabelled SAM were performed using a HPLC system as described [56].
Uniformly 15N- or 15N,13C-labeled SUVR4-WIYLD (residues 1-89) was expressed as a GST-fusion (pGEX4T3) in minimal media containing 15NH4Cl and 13C-glucose as the sole nitrogen and carbon sources, respectively, after induction at 18°C for 18 hours. Protein was purified by glutathione sepharose affinity and size-exclusion chromatography and thrombin digestion to remove the affinity tag. NMR samples contained 0.5 mM protein in PBS at pH 7.4, 5 mM d10-DTT and 10% D2O. All spectra were acquired at 25°C on a 500MHz or 600MHz Bruker spectrometer.
For each experiment 2-3 g of fifteen day old seedlings was crosslinked in 1% formaldehyde under vacuum until the tissue was translucent. Chromatin immunoprecipitation was done as described in [57]. The antibodies used for immunoprecipitation were anti-H2Bub1 (#MM-0029, Medimabs), anti-H3K9me1 (#07-450, Millipore), anti-H3K9me2 (#07-212, Millipore) anti-H3K9me3 (#07-442, Millipore), anti-H3K4me3 (#07-473, Millipore) and anti-GFP (#ab290-50, Abcam). Immunoprecipitated chromatin was eluted in a total of 250 µl elution buffer (1% SDS, 0.1 M NaHCO3) and after reversion of crosslinking, DNA was extracted using the Qiaquick PCR purification kit (Qiagen) and eluted in 100 µl elution buffer. 5 µl of a 4 X dilution was used as a template for real-time PCR in a Lightcycler (Roche). Typically a program of: 1 cycle 95°C 10 min, 45 cycles of 95°C 20 s, 52° 30 s and 72°C 30 s was used to amplify target sequences with gene specific primers (Table S1). PCR was performed on ChIP DNA isolated from two independent experiments, each quantified two separate times.
Nuclear protein extracts were isolated from a chromatin preparation as described [57]. The protein lysate obtained after sonication was separated on a 10-20% SDS-PAGE (Invitrogen, catalog no. EC6625BOX) and transferred to a PVDF membrane (Machery Nagel). Nuclear protein levels were determined using the following antibodies; anti-ubiquitin (1∶4000, Millipore, 07-375), anti-H2Bub1 (1∶1000, MediMabs, MM-0029) and anti-PBA1 (1∶1000, abcam, ab98999).
Leaves from 14 day old seedlings were chopped in 4% formaldehyde on slides, covered with coverslips and flash frozen in liquid N2. The coverslips were removed from the slides when the material was still frozen, and then the slides were washed three times 5 minutes in 1 X PBS. The material was then blocked for 30 min at 37°C in blocking solution (1% BSA in PBS), and incubated with primary antibody (anti H3K9me1, 1∶200; antiH3K9me3, 1∶100) diluted in blocking solution for one hour at 37°C. After a series of washes in PBS, the slides were incubated with goat-anti rabbit Alexa 555 (Invitrogen) secondary antibody (1∶200). Before microscopy the slides were washed in PBS and counterstained in DAPI and inspected with a Zeiss Axiovision2 microscope equipped with epifluorescence attachment. All images were captured using the same exposure times and at 100X magnification.
2 µg of genomic DNA, prepared from leaf material using the Invisorb Spin Plant Kit (INVITEK Berlin), was restricted with ApaI and PstI and used in the bisulfite reaction with the EpiTect Bisulfite Kit (Quiagene Hilden). Bisulfite treated DNA was used as template in a PCR with specific primers. The PCR-Fragments are ligated into pGEMT-vector (Promega) and transformed in DH5alpha cells. Plasmid DNA from several colonies was sequenced with the ABI Prism 310.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.